*This preview shows
page 1. Sign up
to
view the full content.*

**Unformatted text preview: **6
Association Analysis:
Basic Concepts and
Algorithms
Many business enterprises accumulate large quantities of data from their dayto-day operations. For example, huge amounts of customer purchase data are
collected daily at the checkout counters of grocery stores. Table 6.1 illustrates
an example of such data, commonly known as market basket transactions.
Each row in this table corresponds to a transaction, which contains a unique
identiﬁer labeled T ID and a set of items bought by a given customer. Retailers are interested in analyzing the data to learn about the purchasing behavior
of their customers. Such valuable information can be used to support a variety of business-related applications such as marketing promotions, inventory
management, and customer relationship management.
This chapter presents a methodology known as association analysis,
which is useful for discovering interesting relationships hidden in large data
sets. The uncovered relationships can be represented in the form of associaTable 6.1. An example of market basket transactions.
T ID
1
2
3
4
5 Items
{Bread, Milk}
{Bread, Diapers, Beer, Eggs}
{Milk, Diapers, Beer, Cola}
{Bread, Milk, Diapers, Beer}
{Bread, Milk, Diapers, Cola} 328 Chapter 6 Association Analysis tion rules or sets of frequent items. For example, the following rule can be
extracted from the data set shown in Table 6.1:
{Diapers} −→ {Beer}.
The rule suggests that a strong relationship exists between the sale of diapers
and beer because many customers who buy diapers also buy beer. Retailers
can use this type of rules to help them identify new opportunities for crossselling their products to the customers.
Besides market basket data, association analysis is also applicable to other
application domains such as bioinformatics, medical diagnosis, Web mining,
and scientiﬁc data analysis. In the analysis of Earth science data, for example,
the association patterns may reveal interesting connections among the ocean,
land, and atmospheric processes. Such information may help Earth scientists
develop a better understanding of how the diﬀerent elements of the Earth
system interact with each other. Even though the techniques presented here
are generally applicable to a wider variety of data sets, for illustrative purposes,
our discussion will focus mainly on market basket data.
There are two key issues that need to be addressed when applying association analysis to market basket data. First, discovering patterns from a large
transaction data set can be computationally expensive. Second, some of the
discovered patterns are potentially spurious because they may happen simply
by chance. The remainder of this chapter is organized around these two issues. The ﬁrst part of the chapter is devoted to explaining the basic concepts
of association analysis and the algorithms used to eﬃciently mine such patterns. The second part of the chapter deals with the issue of evaluating the
discovered patterns in order to prevent the generation of spurious results. 6.1 Problem Deﬁnition This section reviews the basic terminology used in association analysis and
presents a formal description of the task.
Binary Representation Market basket data can be represented in a binary
format as shown in Table 6.2, where each row corresponds to a transaction
and each column corresponds to an item. An item can be treated as a binary
variable whose value is one if the item is present in a transaction and zero
otherwise. Because the presence of an item in a transaction is often considered
more important than its absence, an item is an asymmetric binary variable. 6.1 Problem Deﬁnition 329 Table 6.2. A binary 0/1 representation of market basket data.
TID
1
2
3
4
5 Bread
1
1
0
1
1 Milk
1
0
1
1
1 Diapers
0
1
1
1
1 Beer
0
1
1
1
0 Eggs
0
1
0
0
0 Cola
0
0
1
0
1 This representation is perhaps a very simplistic view of real market basket data
because it ignores certain important aspects of the data such as the quantity
of items sold or the price paid to purchase them. Methods for handling such
non-binary data will be explained in Chapter 7.
Itemset and Support Count Let I = {i1 ,i2 ,. . .,id } be the set of all items
in a market basket data and T = {t1 , t2 , . . . , tN } be the set of all transactions.
Each transaction ti contains a subset of items chosen from I . In association
analysis, a collection of zero or more items is termed an itemset. If an itemset
contains k items, it is called a k -itemset. For instance, {Beer, Diapers, Milk}
is an example of a 3-itemset. The null (or empty) set is an itemset that does
not contain any items.
The transaction width is deﬁned as the number of items present in a transaction. A transaction tj is said to contain an itemset X if X is a subset of
tj . For example, the second transaction shown in Table 6.2 contains the itemset {Bread, Diapers} but not {Bread, Milk}. An important property of an
itemset is its support count, which refers to the number of transactions that
contain a particular itemset. Mathematically, the support count, σ (X ), for an
itemset X can be stated as follows:
σ (X ) = {ti |X ⊆ ti , ti ∈ T } ,
where the symbol | · | denote the number of elements in a set. In the data set
shown in Table 6.2, the support count for {Beer, Diapers, Milk} is equal to
two because there are only two transactions that contain all three items.
Association Rule An association rule is an implication expression of the
form X −→ Y , where X and Y are disjoint itemsets, i.e., X ∩ Y = ∅. The
strength of an association rule can be measured in terms of its support and
conﬁdence. Support determines how often a rule is applicable to a given 330 Chapter 6 Association Analysis data set, while conﬁdence determines how frequently items in Y appear in
transactions that contain X . The formal deﬁnitions of these metrics are
Support, s(X −→ Y ) =
Conﬁdence, c(X −→ Y ) = σ (X ∪ Y )
;
N
σ (X ∪ Y )
.
σ (X ) (6.1)
(6.2) Example 6.1. Consider the rule {Milk, Diapers} −→ {Beer}. Since the
support count for {Milk, Diapers, Beer} is 2 and the total number of transactions is 5, the rule’s support is 2/5 = 0.4. The rule’s conﬁdence is obtained
by dividing the support count for {Milk, Diapers, Beer} by the support count
for {Milk, Diapers}. Since there are 3 transactions that contain milk and diapers, the conﬁdence for this rule is 2/3 = 0.67.
Why Use Support and Conﬁdence? Support is an important measure
because a rule that has very low support may occur simply by chance. A
low support rule is also likely to be uninteresting from a business perspective
because it may not be proﬁtable to promote items that customers seldom buy
together (with the exception of the situation described in Section 6.8). For
these reasons, support is often used to eliminate uninteresting rules. As will
be shown in Section 6.2.1, support also has a desirable property that can be
exploited for the eﬃcient discovery of association rules.
Conﬁdence, on the other hand, measures the reliability of the inference
made by a rule. For a given rule X −→ Y , the higher the conﬁdence, the more
likely it is for Y to be present in transactions that contain X . Conﬁdence also
provides an estimate of the conditional probability of Y given X .
Association analysis results should be interpreted with caution. The inference made by an association rule does not necessarily imply causality. Instead,
it suggests a strong co-occurrence relationship between items in the antecedent
and consequent of the rule. Causality, on the other hand, requires knowledge
about the causal and eﬀect attributes in the data and typically involves relationships occurring over time (e.g., ozone depletion leads to global warming).
Formulation of Association Rule Mining Problem
rule mining problem can be formally stated as follows: The association Deﬁnition 6.1 (Association Rule Discovery). Given a set of transactions
T , ﬁnd all the rules having support ≥ minsup and conﬁdence ≥ minconf ,
where minsup and minconf are the corresponding support and conﬁdence
thresholds. 6.1 Problem Deﬁnition 331 A brute-force approach for mining association rules is to compute the support and conﬁdence for every possible rule. This approach is prohibitively
expensive because there are exponentially many rules that can be extracted
from a data set. More speciﬁcally, the total number of possible rules extracted
from a data set that contains d items is
R = 3d − 2d+1 + 1. (6.3) The proof for this equation is left as an exercise to the readers (see Exercise 5
on page 405). Even for the small data set shown in Table 6.1, this approach
requires us to compute the support and conﬁdence for 36 − 27 + 1 = 602 rules.
More than 80% of the rules are discarded after applying minsup = 20% and
minconf = 50%, thus making most of the computations become wasted. To
avoid performing needless computations, it would be useful to prune the rules
early without having to compute their support and conﬁdence values.
An initial step toward improving the performance of association rule mining algorithms is to decouple the support and conﬁdence requirements. From
Equation 6.2, notice that the support of a rule X −→ Y depends only on
the support of its corresponding itemset, X ∪ Y . For example, the following
rules have identical support because they involve items from the same itemset,
{Beer, Diapers, Milk}:
{Beer, Diapers} −→ {Milk},
{Diapers, Milk} −→ {Beer},
{Milk} −→ {Beer,Diapers}, {Beer, Milk} −→ {Diapers},
{Beer} −→ {Diapers, Milk},
{Diapers} −→ {Beer,Milk}. If the itemset is infrequent, then all six candidate rules can be pruned immediately without our having to compute their conﬁdence values.
Therefore, a common strategy adopted by many association rule mining
algorithms is to decompose the problem into two major subtasks:
1. Frequent Itemset Generation, whose objective is to ﬁnd all the itemsets that satisfy the minsup threshold. These itemsets are called frequent
itemsets.
2. Rule Generation, whose objective is to extract all the high-conﬁdence
rules from the frequent itemsets found in the previous step. These rules
are called strong rules.
The computational requirements for frequent itemset generation are generally more expensive than those of rule generation. Eﬃcient techniques for
generating frequent itemsets and association rules are discussed in Sections 6.2
and 6.3, respectively. 332 Chapter 6 Association Analysis
null a b c d e ab ac ad ae bc bd be cd ce de abc abd abe acd ace ade bcd bce bde cde abcd abce abde acde bcde abcde Figure 6.1. An itemset lattice. 6.2 Frequent Itemset Generation A lattice structure can be used to enumerate the list of all possible itemsets.
Figure 6.1 shows an itemset lattice for I = {a, b, c, d, e}. In general, a data set
that contains k items can potentially generate up to 2k − 1 frequent itemsets,
excluding the null set. Because k can be very large in many practical applications, the search space of itemsets that need to be explored is exponentially
large.
A brute-force approach for ﬁnding frequent itemsets is to determine the
support count for every candidate itemset in the lattice structure. To do
this, we need to compare each candidate against every transaction, an operation that is shown in Figure 6.2. If the candidate is contained in a transaction,
its support count will be incremented. For example, the support for {Bread,
Milk} is incremented three times because the itemset is contained in transactions 1, 4, and 5. Such an approach can be very expensive because it requires
O(N M w) comparisons, where N is the number of transactions, M = 2k − 1 is
the number of candidate itemsets, and w is the maximum transaction width. 6.2 Frequent Itemset Generation 333
Candidates TID
1
2
N
3
4
5 Transactions
Items
Bread, Milk
Bread, Diapers, Beer, Eggs
Milk, Diapers, Beer, Coke
Bread, Milk, Diapers, Beer
Bread, Milk, Diapers, Coke M Figure 6.2. Counting the support of candidate itemsets. There are several ways to reduce the computational complexity of frequent
itemset generation.
1. Reduce the number of candidate itemsets (M ). The Apriori principle, described in the next section, is an eﬀective way to eliminate some
of the candidate itemsets without counting their support values.
2. Reduce the number of comparisons. Instead of matching each candidate itemset against every transaction, we can reduce the number of
comparisons by using more advanced data structures, either to store the
candidate itemsets or to compress the data set. We will discuss these
strategies in Sections 6.2.4 and 6.6. 6.2.1 The Apriori Principle This section describes how the support measure helps to reduce the number
of candidate itemsets explored during frequent itemset generation. The use of
support for pruning candidate itemsets is guided by the following principle.
Theorem 6.1 (Apriori Principle). If an itemset is frequent, then all of its
subsets must also be frequent.
To illustrate the idea behind the Apriori principle, consider the itemset
lattice shown in Figure 6.3. Suppose {c, d, e} is a frequent itemset. Clearly,
any transaction that contains {c, d, e} must also contain its subsets, {c, d},
{c, e}, {d, e}, {c}, {d}, and {e}. As a result, if {c, d, e} is frequent, then
all subsets of {c, d, e} (i.e., the shaded itemsets in this ﬁgure) must also be
frequent. 334 Chapter 6 Association Analysis
null b a c d e ab ac ad ae bc bd be cd ce de abc abd abe acd ace ade bcd bce bde cde abcd abce abde acde bcde Frequent
Itemset
abcde Figure 6.3. An illustration of the Apriori principle. If {c, d, e} is frequent, then all subsets of this
itemset are frequent. Conversely, if an itemset such as {a, b} is infrequent, then all of its supersets
must be infrequent too. As illustrated in Figure 6.4, the entire subgraph
containing the supersets of {a, b} can be pruned immediately once {a, b} is
found to be infrequent. This strategy of trimming the exponential search
space based on the support measure is known as support-based pruning.
Such a pruning strategy is made possible by a key property of the support
measure, namely, that the support for an itemset never exceeds the support
for its subsets. This property is also known as the anti-monotone property
of the support measure.
Deﬁnition 6.2 (Monotonicity Property). Let I be a set of items, and
J = 2I be the power set of I . A measure f is monotone (or upward closed) if
∀X, Y ∈ J : (X ⊆ Y ) −→ f (X ) ≤ f (Y ), 6.2 Frequent Itemset Generation 335 null Infrequent
Itemset a b c d e ab ac ad ae bc bd be cd ce de abc abd abe acd ace ade bcd bce bde cde abcd abce abde acde bcde Pruned
Supersets
abcde Figure 6.4. An illustration of support-based pruning. If {a, b} is infrequent, then all supersets of {a, b}
are infrequent. which means that if X is a subset of Y , then f (X ) must not exceed f (Y ). On
the other hand, f is anti-monotone (or downward closed) if
∀X, Y ∈ J : (X ⊆ Y ) −→ f (Y ) ≤ f (X ),
which means that if X is a subset of Y , then f (Y ) must not exceed f (X ).
Any measure that possesses an anti-monotone property can be incorporated directly into the mining algorithm to eﬀectively prune the exponential
search space of candidate itemsets, as will be shown in the next section. 6.2.2 Frequent Itemset Generation in the Apriori Algorithm Apriori is the ﬁrst association rule mining algorithm that pioneered the use
of support-based pruning to systematically control the exponential growth of
candidate itemsets. Figure 6.5 provides a high-level illustration of the frequent
itemset generation part of the Apriori algorithm for the transactions shown in 336 Chapter 6 Association Analysis Candidate
1-Itemsets
Item
Count
Beer
3
4
Bread
Cola
2
Diapers
4
4
Milk
Eggs
1 Itemsets removed
because of low
support Minimum support count = 3 Candidate
2-Itemsets
Itemset
Count
{Beer, Bread}
2
{Beer, Diapers}
3
{Beer, Milk}
2
{Bread, Diapers}
3
{Bread, Milk}
3
{Diapers, Milk}
3
Candidate
3-Itemsets
Itemset
Count
{Bread, Diapers, Milk}
3 Figure 6.5. Illustration of frequent itemset generation using the Apriori algorithm. Table 6.1. We assume that the support threshold is 60%, which is equivalent
to a minimum support count equal to 3.
Initially, every item is considered as a candidate 1-itemset. After counting their supports, the candidate itemsets {Cola} and {Eggs} are discarded
because they appear in fewer than three transactions. In the next iteration,
candidate 2-itemsets are generated using only the frequent 1-itemsets because
the Apriori principle ensures that all supersets of the infrequent 1-itemsets
must be infrequent. Because there are only four frequent 1-itemsets, the number of candidate 2-itemsets generated by the algorithm is 4 = 6. Two
2
of these six candidates, {Beer, Bread} and {Beer, Milk}, are subsequently
found to be infrequent after computing their support values. The remaining four candidates are frequent, and thus will be used to generate candidate
3-itemsets. Without support-based pruning, there are 6 = 20 candidate
3
3-itemsets that can be formed using the six items given in this example. With
the Apriori principle, we only need to keep candidate 3-itemsets whose subsets
are frequent. The only candidate that has this property is {Bread, Diapers,
Milk}.
The eﬀectiveness of the Apriori pruning strategy can be shown by counting the number of candidate itemsets generated. A brute-force strategy of 6.2 Frequent Itemset Generation 337 enumerating all itemsets (up to size 3) as candidates will produce
6
6
6
+
+
1
2
3 = 6 + 15 + 20 = 41 candidates. With the Apriori principle, this number decreases to
6
4
+
+ 1 = 6 + 6 + 1 = 13
1
2
candidates, which represents a 68% reduction in the number of candidate
itemsets even in this simple example.
The pseudocode for the frequent itemset generation part of the Apriori
algorithm is shown in Algorithm 6.1. Let Ck denote the set of candidate
k -itemsets and Fk denote the set of frequent k -itemsets:
• The algorithm initially makes a single pass over the data set to determine
the support of each item. Upon completion of this step, the set of all
frequent 1-itemsets, F1 , will be known (steps 1 and 2).
• Next, the algorithm will iteratively generate new candidate k -itemsets
using the frequent (k − 1)-itemsets found in the previous iteration (step
5). Candidate generation is implemented using a function called apriorigen, which is described in Section 6.2.3.
Algorithm 6.1 Frequent itemset generation of the Apriori algorithm.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14: k = 1.
Fk = { i | i ∈ I ∧ σ ({i}) ≥ N × minsup}. {Find all frequent 1-itemsets}
repeat
k = k + 1.
Ck = apriori-gen(Fk−1 ). {Generate candidate itemsets}
for each transaction t ∈ T do
Ct = subset(Ck , t). {Identify all candidates that belong to t}
for each candidate itemset c ∈ Ct do
σ (c) = σ (c) + 1. {Increment support count}
end for
end for
Fk = { c | c ∈ Ck ∧ σ (c) ≥ N × minsup}. {Extract the frequent k -itemsets}
until Fk = ∅
Result = Fk . 338 Chapter 6 Association Analysis • To count the support of the candidates, the algorithm needs to make an
additional pass over the data set (steps 6–10). The subset function is
used to determine all the candidate itemsets in Ck that are contained in
each transaction t. The implementation of this function is described in
Section 6.2.4.
• After counting their supports, the algorithm eliminates all candidate
itemsets whose support counts are less than minsup (step 12).
• The algorithm terminates when there are no new frequent itemsets generated, i.e., Fk = ∅ (step 13).
The frequent itemset generation part of the Apriori algorithm has two important characteristics. First, it is a level-wise algorithm; i.e., it traverses the
itemset lattice one level at a time, from frequent 1-itemsets to the maximum
size of frequent itemsets. Second, it employs a generate-and-test strategy
for ﬁnding frequent itemsets. At each iteration, new candidate itemsets are
generated from the frequent itemsets found in the previous iteration. The
support for each candidate is then counted and tested against the minsup
threshold. The total number of iterations needed by the algorithm is kmax + 1,
where kmax is the maximum size of the frequent itemsets. 6.2.3 Candidate Generation and Pruning The apriori-gen function shown in Step 5 of Algorithm 6.1 generates candidate
itemsets by performing the following two operations:
1. Candidate Generation. This operation generates new candidate k itemsets based on the frequent (k − 1)-itemsets found in the previous
iteration.
2. Candidate Pruning. This operation eliminates some of the candidate
k -itemsets using the support-based pruning strategy.
To illustrate the candidate pruning operation, consider a candidate k -itemset,
X = {i1 , i2 , . . . , ik }. The algorithm must determine whether all of its proper
subsets, X − {ij } (∀j = 1, 2, . . . , k), are frequent. If one of them is infrequent, then X is immediately pruned. This approach can eﬀectively reduce
the number of candidate itemsets considered during support counting. The
complexity of this operation is O(k ) for each candidate k -itemset. However,
as will be shown later, we do not have to examine all k subsets of a given
candidate itemset. If m of the k subsets were used to generate a candidate,
we only need to check the remaining k − m subsets during candidate pruning. 6.2 Frequent Itemset Generation 339 In principle, there are many ways to generate candidate itemsets. The following is a list of requirements for an eﬀective candidate generation procedure:
1. It should avoid generating too many unnecessary candidates. A candidate itemset is unnecessary if at least one of its subsets is infrequent.
Such a candidate is guaranteed to be infrequent according to the antimonotone property of support.
2. It must ensure that the candidate set is complete, i.e., no frequent itemsets are left out by the candidate generation procedure. To ensure completeness, the set of candidate itemsets must subsume the set of all frequent itemsets, i.e., ∀k : Fk ⊆ Ck .
3. It should not generate the same candidate itemset more than once. For
example, the candidate itemset {a, b, c, d} can be generated in many
ways—by merging {a, b, c} with {d}, {b, d} with {a, c}, {c} with {a, b, d},
etc. Generation of duplicate candidates leads to wasted computations
and thus should be avoided for eﬃciency reasons.
Next, we will brieﬂy describe several candidate generation procedures, including the one used by the apriori-gen function.
Brute-Force Method The brute-force method considers every k -itemset as
a potential candidate and then applies the candidate pruning step to remove
any unnecessary candidates (see Figure 6.6). The number of candidate itemd
sets generated at level k is equal to k , where d is the total number of items.
Although candidate generation is rather trivial, candidate pruning becomes
extremely expensive because a large number of itemsets must be examined.
Given that the amount of computations needed for each candidate is O(k ),
d
d
the overall complexity of this method is O
= O d · 2d−1 .
k=1 k × k
Fk−1 × F1 Method An alternative method for candidate generation is to
extend each frequent (k − 1)-itemset with other frequent items. Figure 6.7
illustrates how a frequent 2-itemset such as {Beer, Diapers} can be augmented with a frequent item such as Bread to produce a candidate 3-itemset
{Beer, Diapers, Bread}. This method will produce O(|Fk−1 | × |F1 |) candidate k -itemsets, where |Fj | is the number of frequent j -itemsets. The overall
complexity of this step is O( k k |Fk−1 ||F1 |).
The procedure is complete because every frequent k -itemset is composed
of a frequent (k − 1)-itemset and a frequent 1-itemset. Therefore, all frequent
k -itemsets are part of the candidate k -itemsets generated by this procedure. 340 Chapter 6 Association Analysis
Candidate Generation Items
Item
Beer
Bread
Cola
Diapers
Milk
Eggs Itemset
{Beer, Bread, Cola}
{Beer, Bread, Diapers}
{Beer, Bread, Milk}
{Beer, Bread, Eggs}
{Beer, Cola, Diapers}
{Beer, Cola, Milk}
{Beer, Cola, Eggs}
{Beer, Diapers, Milk}
{Beer, Diapers, Eggs}
{Beer, Milk, Eggs}
{Bread, Cola, Diapers}
{Bread, Cola, Milk}
{Bread, Cola, Eggs}
{Bread, Diapers, Milk}
{Bread, Diapers, Eggs}
{Bread, Milk, Eggs}
{Cola, Diapers, Milk}
{Cola, Diapers, Eggs}
{Cola, Milk, Eggs}
{Diapers, Milk, Eggs} Candidate
Pruning
Itemset
{Bread, Diapers, Milk} Figure 6.6. A brute-force method for generating candidate 3-itemsets.
Frequent
2-itemset
Itemset
{Beer, Diapers}
{Bread, Diapers}
{Bread, Milk}
{Diapers, Milk}
Candidate Generation Frequent
1-itemset
Item
Beer
Bread
Diapers
Milk Itemset
{Beer, Diapers, Bread}
{Beer, Diapers, Milk}
{Bread, Diapers, Milk}
{Bread, Milk, Beer} Candidate
Pruning
Itemset
{Bread, Diapers, Milk} Figure 6.7. Generating and pruning candidate k -itemsets by merging a frequent (k − 1)-itemset with a
frequent item. Note that some of the candidates are unnecessary because their subsets are infrequent. This approach, however, does not prevent the same candidate itemset from
being generated more than once. For instance, {Bread, Diapers, Milk} can
be generated by merging {Bread, Diapers} with {Milk}, {Bread, Milk} with
{Diapers}, or {Diapers, Milk} with {Bread}. One way to avoid generating 6.2 Frequent Itemset Generation 341 duplicate candidates is by ensuring that the items in each frequent itemset are
kept sorted in their lexicographic order. Each frequent (k −1)-itemset X is then
extended with frequent items that are lexicographically larger than the items in
X . For example, the itemset {Bread, Diapers} can be augmented with {Milk}
since Milk is lexicographically larger than Bread and Diapers. However, we
should not augment {Diapers, Milk} with {Bread} nor {Bread, Milk} with
{Diapers} because they violate the lexicographic ordering condition.
While this procedure is a substantial improvement over the brute-force
method, it can still produce a large number of unnecessary candidates. For
example, the candidate itemset obtained by merging {Beer, Diapers} with
{Milk} is unnecessary because one of its subsets, {Beer, Milk}, is infrequent.
There are several heuristics available to reduce the number of unnecessary
candidates. For example, note that, for every candidate k -itemset that survives
the pruning step, every item in the candidate must be contained in at least
k − 1 of the frequent (k − 1)-itemsets. Otherwise, the candidate is guaranteed
to be infrequent. For example, {Beer, Diapers, Milk} is a viable candidate
3-itemset only if every item in the candidate, including Beer, is contained in
at least two frequent 2-itemsets. Since there is only one frequent 2-itemset
containing Beer, all candidate itemsets involving Beer must be infrequent.
Fk−1 × Fk−1 Method The candidate generation procedure in the apriori-gen
function merges a pair of frequent (k − 1)-itemsets only if their ﬁrst k − 2 items
are identical. Let A = {a1 , a2 , . . . , ak−1 } and B = {b1 , b2 , . . . , bk−1 } be a pair
of frequent (k − 1)-itemsets. A and B are merged if they satisfy the following
conditions:
ai = bi (for i = 1, 2, . . . , k − 2) and ak−1 = bk−1 .
In Figure 6.8, the frequent itemsets {Bread, Diapers} and {Bread, Milk} are
merged to form a candidate 3-itemset {Bread, Diapers, Milk}. The algorithm
does not have to merge {Beer, Diapers} with {Diapers, Milk} because the
ﬁrst item in both itemsets is diﬀerent. Indeed, if {Beer, Diapers, Milk} is a
viable candidate, it would have been obtained by merging {Beer, Diapers}
with {Beer, Milk} instead. This example illustrates both the completeness of
the candidate generation procedure and the advantages of using lexicographic
ordering to prevent duplicate candidates. However, because each candidate is
obtained by merging a pair of frequent (k −1)-itemsets, an additional candidate
pruning step is needed to ensure that the remaining k − 2 subsets of the
candidate are frequent. 342 Chapter 6 Association Analysis Frequent
2-itemset
Itemset
{Beer, Diapers}
{Bread, Diapers}
{Bread, Milk}
{Diapers, Milk} Frequent
2-itemset Candidate
Generation Candidate
Pruning Itemset
{Bread, Diapers, Milk} Itemset
{Bread, Diapers, Milk} Itemset
{Beer, Diapers}
{Bread, Diapers}
{Bread, Milk}
{Diapers, Milk} Figure 6.8. Generating and pruning candidate k -itemsets by merging pairs of frequent (k −1)-itemsets. 6.2.4 Support Counting Support counting is the process of determining the frequency of occurrence
for every candidate itemset that survives the candidate pruning step of the
apriori-gen function. Support counting is implemented in steps 6 through 11
of Algorithm 6.1. One approach for doing this is to compare each transaction
against every candidate itemset (see Figure 6.2) and to update the support
counts of candidates contained in the transaction. This approach is computationally expensive, especially when the numbers of transactions and candidate
itemsets are large.
An alternative approach is to enumerate the itemsets contained in each
transaction and use them to update the support counts of their respective candidate itemsets. To illustrate, consider a transaction t that contains ﬁve items,
{1, 2, 3, 5, 6}. There are 5 = 10 itemsets of size 3 contained in this transac3
tion. Some of the itemsets may correspond to the candidate 3-itemsets under
investigation, in which case, their support counts are incremented. Other
subsets of t that do not correspond to any candidates can be ignored.
Figure 6.9 shows a systematic way for enumerating the 3-itemsets contained
in t. Assuming that each itemset keeps its items in increasing lexicographic
order, an itemset can be enumerated by specifying the smallest item ﬁrst,
followed by the larger items. For instance, given t = {1, 2, 3, 5, 6}, all the 3itemsets contained in t must begin with item 1, 2, or 3. It is not possible to
construct a 3-itemset that begins with items 5 or 6 because there are only two 6.2 Frequent Itemset Generation 343 Transaction, t
12356
Level 1
12356 2356 3 56 Level 2
12 3 5 6 13 5 6 123
125
126 135
136 Level 3 15 6 156 23 5 6 235
236 25 6 256 35 6 356 Subsets of 3 items Figure 6.9. Enumerating subsets of three items from a transaction t. items in t whose labels are greater than or equal to 5. The number of ways to
specify the ﬁrst item of a 3-itemset contained in t is illustrated by the Level
1 preﬁx structures depicted in Figure 6.9. For instance, 1 2 3 5 6 represents
a 3-itemset that begins with item 1, followed by two more items chosen from
the set {2, 3, 5, 6}.
After ﬁxing the ﬁrst item, the preﬁx structures at Level 2 represent the
number of ways to select the second item. For example, 1 2 3 5 6 corresponds
to itemsets that begin with preﬁx (1 2) and are followed by items 3, 5, or 6.
Finally, the preﬁx structures at Level 3 represent the complete set of 3-itemsets
contained in t. For example, the 3-itemsets that begin with preﬁx {1 2} are
{1, 2, 3}, {1, 2, 5}, and {1, 2, 6}, while those that begin with preﬁx {2 3} are
{2, 3, 5} and {2, 3, 6}.
The preﬁx structures shown in Figure 6.9 demonstrate how itemsets contained in a transaction can be systematically enumerated, i.e., by specifying
their items one by one, from the leftmost item to the rightmost item. We
still have to determine whether each enumerated 3-itemset corresponds to an
existing candidate itemset. If it matches one of the candidates, then the support count of the corresponding candidate is incremented. In the next section,
we illustrate how this matching operation can be performed eﬃciently using a
hash tree structure. 344 Chapter 6 Association Analysis
Hash Tree Leaf nodes
{Beer, Bread}
containing
candidate {Beer, Diapers}
{Beer, Milk}
2-itemsets {Bread, Diapers}
{Diapers, Milk}
{Bread, Milk} Transactions
TID
1
2
3
4
5 Items
Bread, Milk
Bread, Diapers, Beer, Eggs
Milk, Diapers, Beer, Cola
Bread, Milk, Diapers, Beer
Bread, Milk, Diapers, Cola Figure 6.10. Counting the support of itemsets using hash structure. Support Counting Using a Hash Tree
In the Apriori algorithm, candidate itemsets are partitioned into diﬀerent
buckets and stored in a hash tree. During support counting, itemsets contained
in each transaction are also hashed into their appropriate buckets. That way,
instead of comparing each itemset in the transaction with every candidate
itemset, it is matched only against candidate itemsets that belong to the same
bucket, as shown in Figure 6.10.
Figure 6.11 shows an example of a hash tree structure. Each internal node
of the tree uses the following hash function, h(p) = p mod 3, to determine
which branch of the current node should be followed next. For example, items
1, 4, and 7 are hashed to the same branch (i.e., the leftmost branch) because
they have the same remainder after dividing the number by 3. All candidate
itemsets are stored at the leaf nodes of the hash tree. The hash tree shown in
Figure 6.11 contains 15 candidate 3-itemsets, distributed across 9 leaf nodes.
Consider a transaction, t = {1, 2, 3, 5, 6}. To update the support counts
of the candidate itemsets, the hash tree must be traversed in such a way
that all the leaf nodes containing candidate 3-itemsets belonging to t must be
visited at least once. Recall that the 3-itemsets contained in t must begin with
items 1, 2, or 3, as indicated by the Level 1 preﬁx structures shown in Figure
6.9. Therefore, at the root node of the hash tree, the items 1, 2, and 3 of the
transaction are hashed separately. Item 1 is hashed to the left child of the root
node, item 2 is hashed to the middle child, and item 3 is hashed to the right
child. At the next level of the tree, the transaction is hashed on the second 6.2 Frequent Itemset Generation 345 Hash Function 1,4,7 3,6,9
2,5,8 1+ 2356 Transaction 2+ 356 12356 3+ 56 Candidate Hash Tree 234
567
145 136 345 356 367 357 368 689
124 125 457 458 159 Figure 6.11. Hashing a transaction at the root node of a hash tree. item listed in the Level 2 structures shown in Figure 6.9. For example, after
hashing on item 1 at the root node, items 2, 3, and 5 of the transaction are
hashed. Items 2 and 5 are hashed to the middle child, while item 3 is hashed
to the right child, as shown in Figure 6.12. This process continues until the
leaf nodes of the hash tree are reached. The candidate itemsets stored at the
visited leaf nodes are compared against the transaction. If a candidate is a
subset of the transaction, its support count is incremented. In this example, 5
out of the 9 leaf nodes are visited and 9 out of the 15 itemsets are compared
against the transaction. 6.2.5 Computational Complexity The computational complexity of the Apriori algorithm can be aﬀected by the
following factors.
Support Threshold Lowering the support threshold often results in more
itemsets being declared as frequent. This has an adverse eﬀect on the com- 346 Chapter 6 Association Analysis
1+ 2356 Transaction 2+ 356 12356 3+ 56 Candidate Hash Tree
12+ 356 13+ 56 234
567 15+ 6
145 136 345 356 367 357 368 689
124 125 457 458 159 Figure 6.12. Subset operation on the leftmost subtree of the root of a candidate hash tree. putational complexity of the algorithm because more candidate itemsets must
be generated and counted, as shown in Figure 6.13. The maximum size of
frequent itemsets also tends to increase with lower support thresholds. As the
maximum size of the frequent itemsets increases, the algorithm will need to
make more passes over the data set.
Number of Items (Dimensionality) As the number of items increases,
more space will be needed to store the support counts of items. If the number of
frequent items also grows with the dimensionality of the data, the computation
and I/O costs will increase because of the larger number of candidate itemsets
generated by the algorithm.
Number of Transactions Since the Apriori algorithm makes repeated
passes over the data set, its run time increases with a larger number of transactions.
Average Transaction Width For dense data sets, the average transaction
width can be very large. This aﬀects the complexity of the Apriori algorithm in
two ways. First, the maximum size of frequent itemsets tends to increase as the 6.2 Frequent Itemset Generation 347 ×105 4 Support = 0.1%
Support = 0.2%
Support = 0.5% Number of Candidate Itemsets 3.5
3 2.5
2 1.5
1 0.5
0
0 5 10
Size of Itemset 15 20 (a) Number of candidate itemsets. ×105 4 Support = 0.1%
Support = 0.2%
Support = 0.5% Number of Frequent Itemsets 3.5
3
2.5
2
1.5
1
0.5
0 0 5 10
Size of Itemset 15 20 (b) Number of frequent itemsets. Figure 6.13. Effect of support threshold on the number of candidate and frequent itemsets. average transaction width increases. As a result, more candidate itemsets must
be examined during candidate generation and support counting, as illustrated
in Figure 6.14. Second, as the transaction width increases, more itemsets 348 Chapter 6 Association Analysis
×105 10 Width = 5
Width = 10
Width = 15 9 Number of Candidate Itemsets 8
7
6
5
4
3
2
1
0
0 5 10 15 20 25 Size of Itemset (a) Number of candidate itemsets. 10 ×105
Width = 5
Width = 10
Width = 15 9 Number of Frequent Itemsets 8
7
6
5
4
3
2
1
0
0 5 10
15
Size of Itemset 20 25 (b) Number of Frequent Itemsets. Figure 6.14. Effect of average transaction width on the number of candidate and frequent itemsets. are contained in the transaction. This will increase the number of hash tree
traversals performed during support counting.
A detailed analysis of the time complexity for the Apriori algorithm is
presented next. 6.3 Rule Generation 349 Generation of frequent 1-itemsets For each transaction, we need to update the support count for every item present in the transaction. Assuming
that w is the average transaction width, this operation requires O(N w) time,
where N is the total number of transactions.
Candidate generation To generate candidate k -itemsets, pairs of frequent
(k − 1)-itemsets are merged to determine whether they have at least k − 2
items in common. Each merging operation requires at most k − 2 equality
comparisons. In the best-case scenario, every merging step produces a viable
candidate k -itemset. In the worst-case scenario, the algorithm must merge every pair of frequent (k − 1)-itemsets found in the previous iteration. Therefore,
the overall cost of merging frequent itemsets is
w w (k − 2)|Ck | < Cost of merging <
k=2 (k − 2)|Fk−1 |2 .
k=2 A hash tree is also constructed during candidate generation to store the candidate itemsets. Because the maximum depth of the tree is k , the cost for
w
populating the hash tree with candidate itemsets is O
k=2 k |Ck | . During
candidate pruning, we need to verify that the k − 2 subsets of every candidate
k -itemset are frequent. Since the cost for looking up a candidate in a hash
w
tree is O(k ), the candidate pruning step requires O
k=2 k (k − 2)|Ck | time.
t
Support counting Each transaction of length |t| produces |k| itemsets of
size k . This is also the eﬀective number of hash tree traversals performed for
each transaction. The cost for support counting is O N k w αk , where w
k
is the maximum transaction width and αk is the cost for updating the support
count of a candidate k -itemset in the hash tree. 6.3 Rule Generation This section describes how to extract association rules eﬃciently from a given
frequent itemset. Each frequent k -itemset, Y , can produce up to 2k − 2 association rules, ignoring rules that have empty antecedents or consequents (∅ −→ Y
or Y −→ ∅). An association rule can be extracted by partitioning the itemset
Y into two non-empty subsets, X and Y − X , such that X −→ Y − X satisﬁes
the conﬁdence threshold. Note that all such rules must have already met the
support threshold because they are generated from a frequent itemset. 350 Chapter 6 Association Analysis Example 6.2. Let X = {1, 2, 3} be a frequent itemset. There are six candidate association rules that can be generated from X : {1, 2} −→ {3}, {1, 3} −→
{2}, {2, 3} −→ {1}, {1} −→ {2, 3}, {2} −→ {1, 3}, and {3} −→ {1, 2}. As
each of their support is identical to the support for X , the rules must satisfy
the support threshold.
Computing the conﬁdence of an association rule does not require additional
scans of the transaction data set. Consider the rule {1, 2} −→ {3}, which is
generated from the frequent itemset X = {1, 2, 3}. The conﬁdence for this rule
is σ ({1, 2, 3})/σ ({1, 2}). Because {1, 2, 3} is frequent, the anti-monotone property of support ensures that {1, 2} must be frequent, too. Since the support
counts for both itemsets were already found during frequent itemset generation, there is no need to read the entire data set again. 6.3.1 Conﬁdence-Based Pruning Unlike the support measure, conﬁdence does not have any monotone property.
For example, the conﬁdence for X −→ Y can be larger, smaller, or equal to the
˜
˜
˜
˜
conﬁdence for another rule X −→ Y , where X ⊆ X and Y ⊆ Y (see Exercise
3 on page 405). Nevertheless, if we compare rules generated from the same
frequent itemset Y , the following theorem holds for the conﬁdence measure.
Theorem 6.2. If a rule X −→ Y − X does not satisfy the conﬁdence threshold,
then any rule X −→ Y − X , where X is a subset of X , must not satisfy the
conﬁdence threshold as well.
To prove this theorem, consider the following two rules: X −→ Y − X and
X −→ Y − X , where X ⊂ X . The conﬁdence of the rules are σ (Y )/σ (X ) and
σ (Y )/σ (X ), respectively. Since X is a subset of X , σ (X ) ≥ σ (X ). Therefore,
the former rule cannot have a higher conﬁdence than the latter rule. 6.3.2 Rule Generation in Apriori Algorithm The Apriori algorithm uses a level-wise approach for generating association
rules, where each level corresponds to the number of items that belong to the
rule consequent. Initially, all the high-conﬁdence rules that have only one item
in the rule consequent are extracted. These rules are then used to generate
new candidate rules. For example, if {acd} −→ {b} and {abd} −→ {c} are
high-conﬁdence rules, then the candidate rule {ad} −→ {bc} is generated by
merging the consequents of both rules. Figure 6.15 shows a lattice structure
for the association rules generated from the frequent itemset {a, b, c, d}. If any
node in the lattice has low conﬁdence, then according to Theorem 6.2, the 6.3 Rule Generation 351 Low-Confidence
Rule
abcd=>{ } bcd=>a cd=>ab bd=>ac d=>abc acd=>b bc=>ad c=>abd abd=>c ad=>bc abc=>d ac=>bd b=>acd ab=>cd a=>bcd Pruned
Rules Figure 6.15. Pruning of association rules using the conﬁdence measure. entire subgraph spanned by the node can be pruned immediately. Suppose
the conﬁdence for {bcd} −→ {a} is low. All the rules containing item a in
its consequent, including {cd} −→ {ab}, {bd} −→ {ac}, {bc} −→ {ad}, and
{d} −→ {abc} can be discarded.
A pseudocode for the rule generation step is shown in Algorithms 6.2 and
6.3. Note the similarity between the ap-genrules procedure given in Algorithm 6.3 and the frequent itemset generation procedure given in Algorithm
6.1. The only diﬀerence is that, in rule generation, we do not have to make
additional passes over the data set to compute the conﬁdence of the candidate
rules. Instead, we determine the conﬁdence of each rule by using the support
counts computed during frequent itemset generation.
Algorithm 6.2 Rule generation of the Apriori algorithm.
1: for each frequent k -itemset fk , k ≥ 2 do
2:
H1 = {i | i ∈ fk }
{1-item consequents of the rule.}
3:
call ap-genrules(fk , H1 .)
4: end for 352 Chapter 6 Association Analysis Algorithm 6.3 Procedure ap-genrules(fk , Hm ).
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14: k = |fk | {size of frequent itemset.}
m = |Hm | {size of rule consequent.}
if k > m + 1 then
Hm+1 = apriori-gen(Hm ).
for each hm+1 ∈ Hm+1 do
conf = σ (fk )/σ (fk − hm+1 ).
if conf ≥ minconf then
output the rule (fk − hm+1 ) −→ hm+1 .
else
delete hm+1 from Hm+1 .
end if
end for
call ap-genrules(fk , Hm+1 .)
end if 6.3.3 An Example: Congressional Voting Records This section demonstrates the results of applying association analysis to the
voting records of members of the United States House of Representatives. The
data is obtained from the 1984 Congressional Voting Records Database, which
is available at the UCI machine learning data repository. Each transaction
contains information about the party aﬃliation for a representative along with
his or her voting record on 16 key issues. There are 435 transactions and 34
items in the data set. The set of items are listed in Table 6.3.
The Apriori algorithm is then applied to the data set with minsup = 30%
and minconf = 90%. Some of the high-conﬁdence rules extracted by the
algorithm are shown in Table 6.4. The ﬁrst two rules suggest that most of the
members who voted yes for aid to El Salvador and no for budget resolution and
MX missile are Republicans; while those who voted no for aid to El Salvador
and yes for budget resolution and MX missile are Democrats. These highconﬁdence rules show the key issues that divide members from both political
parties. If minconf is reduced, we may ﬁnd rules that contain issues that cut
across the party lines. For example, with minconf = 40%, the rules suggest
that corporation cutbacks is an issue that receives almost equal number of
votes from both parties—52.3% of the members who voted no are Republicans,
while the remaining 47.7% of them who voted no are Democrats. 6.4 Compact Representation of Frequent Itemsets 353 Table 6.3. List of binary attributes from the 1984 United States Congressional Voting Records. Source:
The UCI machine learning repository.
1. Republican
2. Democrat
3. handicapped-infants = yes
4. handicapped-infants = no
5. water project cost sharing = yes
6. water project cost sharing = no
7. budget-resolution = yes
8. budget-resolution = no
9. physician fee freeze = yes
10. physician fee freeze = no
11. aid to El Salvador = yes
12. aid to El Salvador = no
13. religious groups in schools = yes
14. religious groups in schools = no
15. anti-satellite test ban = yes
16. anti-satellite test ban = no
17. aid to Nicaragua = yes 18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34. aid to Nicaragua = no
MX-missile = yes
MX-missile = no
immigration = yes
immigration = no
synfuel corporation cutback = yes
synfuel corporation cutback = no
education spending = yes
education spending = no
right-to-sue = yes
right-to-sue = no
crime = yes
crime = no
duty-free-exports = yes
duty-free-exports = no
export administration act = yes
export administration act = no Table 6.4. Association rules extracted from the 1984 United States Congressional Voting Records.
Association Rule
{budget resolution = no, MX-missile=no, aid to El Salvador = yes }
−→ {Republican}
{budget resolution = yes, MX-missile=yes, aid to El Salvador = no }
−→ {Democrat}
{crime = yes, right-to-sue = yes, physician fee freeze = yes}
−→ {Republican}
{crime = no, right-to-sue = no, physician fee freeze = no}
−→ {Democrat} 6.4 Conﬁdence
91.0%
97.5%
93.5%
100% Compact Representation of Frequent Itemsets In practice, the number of frequent itemsets produced from a transaction data
set can be very large. It is useful to identify a small representative set of
itemsets from which all other frequent itemsets can be derived. Two such
representations are presented in this section in the form of maximal and closed
frequent itemsets. 354 Chapter 6 Association Analysis
null Maximal Frequent
Itemset
a b c d e ab ac ad ae bc bd be cd ce de abc abd abe acd ace ade bcd bce bde cde abcd abce abde acde bcde
Frequent abcde
Infrequent Frequent
Itemset
Border Figure 6.16. Maximal frequent itemset. 6.4.1 Maximal Frequent Itemsets Deﬁnition 6.3 (Maximal Frequent Itemset). A maximal frequent itemset is deﬁned as a frequent itemset for which none of its immediate supersets
are frequent.
To illustrate this concept, consider the itemset lattice shown in Figure
6.16. The itemsets in the lattice are divided into two groups: those that are
frequent and those that are infrequent. A frequent itemset border, which is
represented by a dashed line, is also illustrated in the diagram. Every itemset
located above the border is frequent, while those located below the border (the
shaded nodes) are infrequent. Among the itemsets residing near the border,
{a, d}, {a, c, e}, and {b, c, d, e} are considered to be maximal frequent itemsets
because their immediate supersets are infrequent. An itemset such as {a, d}
is maximal frequent because all of its immediate supersets, {a, b, d}, {a, c, d},
and {a, d, e}, are infrequent. In contrast, {a, c} is non-maximal because one
of its immediate supersets, {a, c, e}, is frequent.
Maximal frequent itemsets eﬀectively provide a compact representation of
frequent itemsets. In other words, they form the smallest set of itemsets from 6.4 Compact Representation of Frequent Itemsets 355 which all frequent itemsets can be derived. For example, the frequent itemsets
shown in Figure 6.16 can be divided into two groups:
• Frequent itemsets that begin with item a and that may contain items c,
d, or e. This group includes itemsets such as {a}, {a, c}, {a, d}, {a, e},
and {a, c, e}.
• Frequent itemsets that begin with items b, c, d, or e. This group includes
itemsets such as {b}, {b, c}, {c, d},{b, c, d, e}, etc.
Frequent itemsets that belong in the ﬁrst group are subsets of either {a, c, e}
or {a, d}, while those that belong in the second group are subsets of {b, c, d, e}.
Hence, the maximal frequent itemsets {a, c, e}, {a, d}, and {b, c, d, e} provide
a compact representation of the frequent itemsets shown in Figure 6.16.
Maximal frequent itemsets provide a valuable representation for data sets
that can produce very long, frequent itemsets, as there are exponentially many
frequent itemsets in such data. Nevertheless, this approach is practical only
if an eﬃcient algorithm exists to explicitly ﬁnd the maximal frequent itemsets
without having to enumerate all their subsets. We brieﬂy describe one such
approach in Section 6.5.
Despite providing a compact representation, maximal frequent itemsets do
not contain the support information of their subsets. For example, the support
of the maximal frequent itemsets {a, c, e}, {a, d}, and {b,c,d,e} do not provide
any hint about the support of their subsets. An additional pass over the data
set is therefore needed to determine the support counts of the non-maximal
frequent itemsets. In some cases, it might be desirable to have a minimal
representation of frequent itemsets that preserves the support information.
We illustrate such a representation in the next section. 6.4.2 Closed Frequent Itemsets Closed itemsets provide a minimal representation of itemsets without losing
their support information. A formal deﬁnition of a closed itemset is presented
below.
Deﬁnition 6.4 (Closed Itemset). An itemset X is closed if none of its
immediate supersets has exactly the same support count as X .
Put another way, X is not closed if at least one of its immediate supersets
has the same support count as X . Examples of closed itemsets are shown in
Figure 6.17. To better illustrate the support count of each itemset, we have
associated each node (itemset) in the lattice with a list of its corresponding 356 Chapter 6
TID Items 1 abc 2 Association Analysis abcd 3 acde 5 null bce 4 minsup = 40% de 1,2
ab 1,2
abc 1,2,4
a 1,2,4
ac 2,4
ad 2 1,2,3
b 1,2,3 4
ae 2,4
abd 1,2,3,4
c abe 2
bc 3 bd 4
acd 2,4,5
d 4
ace 2 2,4
be 2
ade 3,4,5
e cd 3,4 ce 3
bcd 4,5 de 4
bce bde cde 4
abcd Closed Frequent Itemset abce abde acde bcde abcde Figure 6.17. An example of the closed frequent itemsets (with minimum support count equal to 40%). transaction IDs. For example, since the node {b, c} is associated with transaction IDs 1, 2, and 3, its support count is equal to three. From the transactions
given in this diagram, notice that every transaction that contains b also contains c. Consequently, the support for {b} is identical to {b, c} and {b} should
not be considered a closed itemset. Similarly, since c occurs in every transaction that contains both a and d, the itemset {a, d} is not closed. On the other
hand, {b, c} is a closed itemset because it does not have the same support
count as any of its supersets.
Deﬁnition 6.5 (Closed Frequent Itemset). An itemset is a closed frequent itemset if it is closed and its support is greater than or equal to minsup.
In the previous example, assuming that the support threshold is 40%, {b,c}
is a closed frequent itemset because its support is 60%. The rest of the closed
frequent itemsets are indicated by the shaded nodes.
Algorithms are available to explicitly extract closed frequent itemsets from
a given data set. Interested readers may refer to the bibliographic notes at the
end of this chapter for further discussions of these algorithms. We can use the
closed frequent itemsets to determine the support counts for the non-closed 6.4 Compact Representation of Frequent Itemsets 357 Algorithm 6.4 Support counting using closed frequent itemsets.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11: Let C denote the set of closed frequent itemsets
Let kmax denote the maximum size of closed frequent itemsets
Fkmax = {f |f ∈ C, |f | = kmax }
{Find all frequent itemsets of size kmax .}
for k = kmax − 1 downto 1 do
Fk = {f |f ⊂ Fk+1 , |f | = k }
{Find all frequent itemsets of size k .}
for each f ∈ Fk do
if f ∈ C then
/
f.support = max{f .support|f ∈ Fk+1 , f ⊂ f }
end if
end for
end for frequent itemsets. For example, consider the frequent itemset {a, d} shown
in Figure 6.17. Because the itemset is not closed, its support count must be
identical to one of its immediate supersets. The key is to determine which
superset (among {a, b, d}, {a, c, d}, or {a, d, e}) has exactly the same support
count as {a, d}. The Apriori principle states that any transaction that contains
the superset of {a, d} must also contain {a, d}. However, any transaction that
contains {a, d} does not have to contain the supersets of {a, d}. For this
reason, the support for {a, d} must be equal to the largest support among its
supersets. Since {a, c, d} has a larger support than both {a, b, d} and {a, d, e},
the support for {a, d} must be identical to the support for {a, c, d}. Using this
methodology, an algorithm can be developed to compute the support for the
non-closed frequent itemsets. The pseudocode for this algorithm is shown in
Algorithm 6.4. The algorithm proceeds in a speciﬁc-to-general fashion, i.e.,
from the largest to the smallest frequent itemsets. This is because, in order
to ﬁnd the support for a non-closed frequent itemset, the support for all of its
supersets must be known.
To illustrate the advantage of using closed frequent itemsets, consider the
data set shown in Table 6.5, which contains ten transactions and ﬁfteen items.
The items can be divided into three groups: (1) Group A, which contains
items a1 through a5 ; (2) Group B , which contains items b1 through b5 ; and
(3) Group C , which contains items c1 through c5 . Note that items within each
group are perfectly associated with each other and they do not appear with
items from another group. Assuming the support threshold is 20%, the total
number of frequent itemsets is 3 × (25 − 1) = 93. However, there are only three
closed frequent itemsets in the data: ({a1 , a2 , a3 , a4 , a5 }, {b1 , b2 , b3 , b4 , b5 }, and
{c1 , c2 , c3 , c4 , c5 }). It is often suﬃcient to present only the closed frequent
itemsets to the analysts instead of the entire set of frequent itemsets. 358 Chapter 6 Association Analysis Table 6.5. A transaction data set for mining closed itemsets.
TID
1
2
3
4
5
6
7
8
9
10 a1
1
1
1
0
0
0
0
0
0
0 a2
1
1
1
0
0
0
0
0
0
0 a3
1
1
1
0
0
0
0
0
0
0 a4
1
1
1
0
0
0
0
0
0
0 a5
1
1
1
0
0
0
0
0
0
0 b1
0
0
0
1
1
1
0
0
0
0 b2
0
0
0
1
1
1
0
0
0
0 b3
0
0
0
1
1
1
0
0
0
0 b4
0
0
0
1
1
1
0
0
0
0 b5
0
0
0
1
1
1
0
0
0
0 c1
0
0
0
0
0
0
1
1
1
1 c2
0
0
0
0
0
0
1
1
1
1 c3
0
0
0
0
0
0
1
1
1
1 c4
0
0
0
0
0
0
1
1
1
1 c5
0
0
0
0
0
0
1
1
1
1 Frequent
Itemsets
Closed
Frequent
Itemsets
Maximal
Frequent
Itemsets Figure 6.18. Relationships among frequent, maximal frequent, and closed frequent itemsets. Closed frequent itemsets are useful for removing some of the redundant
association rules. An association rule X − Y is redundant if there exists
another rule X −→ Y , where X is a subset of X and Y is a subset of Y , such
that the support and conﬁdence for both rules are identical. In the example
shown in Figure 6.17, {b} is not a closed frequent itemset while {b, c} is closed.
The association rule {b} −→ {d, e} is therefore redundant because it has the
same support and conﬁdence as {b, c} −→ {d, e}. Such redundant rules are
not generated if closed frequent itemsets are used for rule generation.
Finally, note that all maximal frequent itemsets are closed because none
of the maximal frequent itemsets can have the same support count as their
immediate supersets. The relationships among frequent, maximal frequent,
and closed frequent itemsets are shown in Figure 6.18. 6.5 6.5 Alternative Methods for Generating Frequent Itemsets 359 Alternative Methods for Generating Frequent
Itemsets Apriori is one of the earliest algorithms to have successfully addressed the
combinatorial explosion of frequent itemset generation. It achieves this by applying the Apriori principle to prune the exponential search space. Despite its
signiﬁcant performance improvement, the algorithm still incurs considerable
I/O overhead since it requires making several passes over the transaction data
set. In addition, as noted in Section 6.2.5, the performance of the Apriori
algorithm may degrade signiﬁcantly for dense data sets because of the increasing width of transactions. Several alternative methods have been developed
to overcome these limitations and improve upon the eﬃciency of the Apriori
algorithm. The following is a high-level description of these methods.
Traversal of Itemset Lattice A search for frequent itemsets can be conceptually viewed as a traversal on the itemset lattice shown in Figure 6.1.
The search strategy employed by an algorithm dictates how the lattice structure is traversed during the frequent itemset generation process. Some search
strategies are better than others, depending on the conﬁguration of frequent
itemsets in the lattice. An overview of these strategies is presented next.
• General-to-Speciﬁc versus Speciﬁc-to-General: The Apriori algorithm uses a general-to-speciﬁc search strategy, where pairs of frequent
(k − 1)-itemsets are merged to obtain candidate k -itemsets. This generalto-speciﬁc search strategy is eﬀective, provided the maximum length of
a frequent itemset is not too long. The conﬁguration of frequent itemsets that works best with this strategy is shown in Figure 6.19(a), where
the darker nodes represent infrequent itemsets. Alternatively, a speciﬁcto-general search strategy looks for more speciﬁc frequent itemsets ﬁrst,
before ﬁnding the more general frequent itemsets. This strategy is useful to discover maximal frequent itemsets in dense transactions, where
the frequent itemset border is located near the bottom of the lattice,
as shown in Figure 6.19(b). The Apriori principle can be applied to
prune all subsets of maximal frequent itemsets. Speciﬁcally, if a candidate k -itemset is maximal frequent, we do not have to examine any of its
subsets of size k − 1. However, if the candidate k -itemset is infrequent,
we need to check all of its k − 1 subsets in the next iteration. Another
approach is to combine both general-to-speciﬁc and speciﬁc-to-general
search strategies. This bidirectional approach requires more space to 360 Chapter 6 Association Analysis Frequent
Itemset
Border null {a1,a2,...,an}
(a) General-to-specific Frequent
Itemset
Border null {a1,a2,...,an} Frequent
Itemset
Border (b) Specific-to-general null {a1,a2,...,an}
(c) Bidirectional Figure 6.19. General-to-speciﬁc, speciﬁc-to-general, and bidirectional search. store the candidate itemsets, but it can help to rapidly identify the frequent itemset border, given the conﬁguration shown in Figure 6.19(c).
• Equivalence Classes: Another way to envision the traversal is to ﬁrst
partition the lattice into disjoint groups of nodes (or equivalence classes).
A frequent itemset generation algorithm searches for frequent itemsets
within a particular equivalence class ﬁrst before moving to another equivalence class. As an example, the level-wise strategy used in the Apriori
algorithm can be considered to be partitioning the lattice on the basis
of itemset sizes; i.e., the algorithm discovers all frequent 1-itemsets ﬁrst
before proceeding to larger-sized itemsets. Equivalence classes can also
be deﬁned according to the preﬁx or suﬃx labels of an itemset. In this
case, two itemsets belong to the same equivalence class if they share
a common preﬁx or suﬃx of length k . In the preﬁx-based approach,
the algorithm can search for frequent itemsets starting with the preﬁx
a before looking for those starting with preﬁxes b, c, and so on. Both
preﬁx-based and suﬃx-based equivalence classes can be demonstrated
using the tree-like structure shown in Figure 6.20.
• Breadth-First versus Depth-First: The Apriori algorithm traverses
the lattice in a breadth-ﬁrst manner, as shown in Figure 6.21(a). It ﬁrst
discovers all the frequent 1-itemsets, followed by the frequent 2-itemsets,
and so on, until no new frequent itemsets are generated. The itemset 6.5 Alternative Methods for Generating Frequent Itemsets 361
null null b a ab ac ad abc abd acd c bc bd bcd d cd b a ab ac abc c bc ad abd d bd acd abcd cd bcd abcd (a) Prefix tree. (b) Suffix tree. Figure 6.20. Equivalence classes based on the preﬁx and sufﬁx labels of itemsets. (a) Breadth first (b) Depth first Figure 6.21. Breadth-ﬁrst and depth-ﬁrst traversals. lattice can also be traversed in a depth-ﬁrst manner, as shown in Figures
6.21(b) and 6.22. The algorithm can start from, say, node a in Figure
6.22, and count its support to determine whether it is frequent. If so, the
algorithm progressively expands the next level of nodes, i.e., ab, abc, and
so on, until an infrequent node is reached, say, abcd. It then backtracks
to another branch, say, abce, and continues the search from there.
The depth-ﬁrst approach is often used by algorithms designed to ﬁnd
maximal frequent itemsets. This approach allows the frequent itemset
border to be detected more quickly than using a breadth-ﬁrst approach.
Once a maximal frequent itemset is found, substantial pruning can be 362 Chapter 6 Association Analysis
null a b c d
e ab ac ad
ae abc abd bc cd
be acd
abe bd ce de bcd
ace bce ade bde cde abcd
abce abde acde bcde abcde Figure 6.22. Generating candidate itemsets using the depth-ﬁrst approach. performed on its subsets. For example, if the node bcde shown in Figure
6.22 is maximal frequent, then the algorithm does not have to visit the
subtrees rooted at bd, be, c, d, and e because they will not contain any
maximal frequent itemsets. However, if abc is maximal frequent, only the
nodes such as ac and bc are not maximal frequent (but the subtrees of
ac and bc may still contain maximal frequent itemsets). The depth-ﬁrst
approach also allows a diﬀerent kind of pruning based on the support
of itemsets. For example, suppose the support for {a, b, c} is identical
to the support for {a, b}. The subtrees rooted at abd and abe can be
skipped because they are guaranteed not to have any maximal frequent
itemsets. The proof of this is left as an exercise to the readers.
Representation of Transaction Data Set There are many ways to represent a transaction data set. The choice of representation can aﬀect the I/O
costs incurred when computing the support of candidate itemsets. Figure 6.23
shows two diﬀerent ways of representing market basket transactions. The representation on the left is called a horizontal data layout, which is adopted
by many association rule mining algorithms, including Apriori. Another possibility is to store the list of transaction identiﬁers (TID-list) associated with
each item. Such a representation is known as the vertical data layout. The
support for each candidate itemset is obtained by intersecting the TID-lists of
its subset items. The length of the TID-lists shrinks as we progress to larger 6.6
Horizontal
Data Layout
TID
1
2
3
4
5
6
7
8
9
10 Items
a,b,e
b,c,d
c,e
a,c,d
a,b,c,d
a,e
a,b
a,b,c
a,c,d
b FP-Growth Algorithm 363 Vertical Data Layout
a
1
4
5
6
7
8
9 b
1
2
5
7
8
10 c
2
3
4
8
9 d
2
4
5
9 e
1
3
6 Figure 6.23. Horizontal and vertical data format. sized itemsets. However, one problem with this approach is that the initial
set of TID-lists may be too large to ﬁt into main memory, thus requiring
more sophisticated techniques to compress the TID-lists. We describe another
eﬀective approach to represent the data in the next section. 6.6 FP-Growth Algorithm This section presents an alternative algorithm called FP-growth that takes
a radically diﬀerent approach to discovering frequent itemsets. The algorithm
does not subscribe to the generate-and-test paradigm of Apriori. Instead, it
encodes the data set using a compact data structure called an FP-tree and
extracts frequent itemsets directly from this structure. The details of this
approach are presented next. 6.6.1 FP-Tree Representation An FP-tree is a compressed representation of the input data. It is constructed
by reading the data set one transaction at a time and mapping each transaction
onto a path in the FP-tree. As diﬀerent transactions can have several items
in common, their paths may overlap. The more the paths overlap with one
another, the more compression we can achieve using the FP-tree structure. If
the size of the FP-tree is small enough to ﬁt into main memory, this will allow
us to extract frequent itemsets directly from the structure in memory instead
of making repeated passes over the data stored on disk. 364 Chapter 6 Association Analysis Transaction
Data Set
TID
Items
{a,b}
1
2
{b,c,d}
{a,c,d,e}
3
4
{a,d,e}
5
{a,b,c}
{a,b,c,d}
6
7
{a}
8
{a,b,c}
9
{a,b,d}
10
{b,c,e} null
null a:1 b:1 a:1
c:1 b:1
b:1 d:1
(i) After reading TID=1 (ii) After reading TID=2
null
a:2 b:1 b:1 c:1 c:1
d:1 d:1 e:1
(iii) After reading TID=3
null
a:8 b:2 b:5 c:2
c:1 c:3
d:1
d:1 d:1
d:1 d:1
e:1 e:1 e:1 (iv) After reading TID=10 Figure 6.24. Construction of an FP-tree. Figure 6.24 shows a data set that contains ten transactions and ﬁve items.
The structures of the FP-tree after reading the ﬁrst three transactions are also
depicted in the diagram. Each node in the tree contains the label of an item
along with a counter that shows the number of transactions mapped onto the
given path. Initially, the FP-tree contains only the root node represented by
the null symbol. The FP-tree is subsequently extended in the following way:
1. The data set is scanned once to determine the support count of each
item. Infrequent items are discarded, while the frequent items are sorted
in decreasing support counts. For the data set shown in Figure 6.24, a
is the most frequent item, followed by b, c, d, and e. 6.6 FP-Growth Algorithm 365 2. The algorithm makes a second pass over the data to construct the FPtree. After reading the ﬁrst transaction, {a, b}, the nodes labeled as a
and b are created. A path is then formed from null → a → b to encode
the transaction. Every node along the path has a frequency count of 1.
3. After reading the second transaction, {b,c,d}, a new set of nodes is created for items b, c, and d. A path is then formed to represent the
transaction by connecting the nodes null → b → c → d. Every node
along this path also has a frequency count equal to one. Although the
ﬁrst two transactions have an item in common, which is b, their paths
are disjoint because the transactions do not share a common preﬁx.
4. The third transaction, {a,c,d,e}, shares a common preﬁx item (which
is a) with the ﬁrst transaction. As a result, the path for the third
transaction, null → a → c → d → e, overlaps with the path for the
ﬁrst transaction, null → a → b. Because of their overlapping path, the
frequency count for node a is incremented to two, while the frequency
counts for the newly created nodes, c, d, and e, are equal to one.
5. This process continues until every transaction has been mapped onto one
of the paths given in the FP-tree. The resulting FP-tree after reading
all the transactions is shown at the bottom of Figure 6.24.
The size of an FP-tree is typically smaller than the size of the uncompressed
data because many transactions in market basket data often share a few items
in common. In the best-case scenario, where all the transactions have the
same set of items, the FP-tree contains only a single branch of nodes. The
worst-case scenario happens when every transaction has a unique set of items.
As none of the transactions have any items in common, the size of the FP-tree
is eﬀectively the same as the size of the original data. However, the physical
storage requirement for the FP-tree is higher because it requires additional
space to store pointers between nodes and counters for each item.
The size of an FP-tree also depends on how the items are ordered. If
the ordering scheme in the preceding example is reversed, i.e., from lowest
to highest support item, the resulting FP-tree is shown in Figure 6.25. The
tree appears to be denser because the branching factor at the root node has
increased from 2 to 5 and the number of nodes containing the high support
items such as a and b has increased from 3 to 12. Nevertheless, ordering
by decreasing support counts does not always lead to the smallest tree. For
example, suppose we augment the data set given in Figure 6.24 with 100
transactions that contain {e}, 80 transactions that contain {d}, 60 transactions 366 Chapter 6 Association Analysis
null d:3 e:3 b:1 c:2 d:2
c:1 c:1 a:1 a:1
b:1 b:2 b:2 b:1
a:1 c:2 a:1
a:1 a:2 a:1 Figure 6.25. An FP-tree representation for the data set shown in Figure 6.24 with a different item
ordering scheme. that contain {c}, and 40 transactions that contain {b}. Item e is now most
frequent, followed by d, c, b, and a. With the augmented transactions, ordering
by decreasing support counts will result in an FP-tree similar to Figure 6.25,
while a scheme based on increasing support counts produces a smaller FP-tree
similar to Figure 6.24(iv).
An FP-tree also contains a list of pointers connecting between nodes that
have the same items. These pointers, represented as dashed lines in Figures
6.24 and 6.25, help to facilitate the rapid access of individual items in the tree.
We explain how to use the FP-tree and its corresponding pointers for frequent
itemset generation in the next section. 6.6.2 Frequent Itemset Generation in FP-Growth Algorithm FP-growth is an algorithm that generates frequent itemsets from an FP-tree
by exploring the tree in a bottom-up fashion. Given the example tree shown in
Figure 6.24, the algorithm looks for frequent itemsets ending in e ﬁrst, followed
by d, c, b, and ﬁnally, a. This bottom-up strategy for ﬁnding frequent itemsets ending with a particular item is equivalent to the suﬃx-based approach
described in Section 6.5. Since every transaction is mapped onto a path in the
FP-tree, we can derive the frequent itemsets ending with a particular item,
say, e, by examining only the paths containing node e. These paths can be
accessed rapidly using the pointers associated with node e. The extracted
paths are shown in Figure 6.26(a). The details on how to process the paths to
obtain frequent itemsets will be explained later. 6.6 FP-Growth Algorithm 367
null null
a:8
a:8 b:2
b:5 c:1 b:2 c:2
d:1 c:3 c:2
c:1 d:1
d:1 e:1
e:1
e:1
(a) Paths containing node e null b:2
c:2
c:1 c:3
(c) Paths containing node c d:1 d:1 null a:8
b:2 b:5 d:1 (b) Paths containing node d null a:8 d:1 a:8 b:5
(d) Paths containing node b (e) Paths containing node a Figure 6.26. Decomposing the frequent itemset generation problem into multiple subproblems, where
each subproblem involves ﬁnding frequent itemsets ending in e, d, c, b, and a. Table 6.6. The list of frequent itemsets ordered by their corresponding sufﬁxes. Suﬃx
e
d
c
b
a Frequent Itemsets
{e}, {d,e}, {a,d,e}, {c,e},{a,e}
{d}, {c,d}, {b,c,d}, {a,c,d}, {b,d}, {a,b,d}, {a,d}
{c}, {b,c}, {a,b,c}, {a,c}
{b}, {a,b}
{a} After ﬁnding the frequent itemsets ending in e, the algorithm proceeds to
look for frequent itemsets ending in d by processing the paths associated with
node d. The corresponding paths are shown in Figure 6.26(b). This process
continues until all the paths associated with nodes c, b, and ﬁnally a, are
processed. The paths for these items are shown in Figures 6.26(c), (d), and
(e), while their corresponding frequent itemsets are summarized in Table 6.6.
FP-growth ﬁnds all the frequent itemsets ending with a particular suﬃx
by employing a divide-and-conquer strategy to split the problem into smaller
subproblems. For example, suppose we are interested in ﬁnding all frequent 368 Chapter 6 Association Analysis
null null a:8 b:2 c:1 a:2 c:2
c:1 d:1 c:1 d:1
e:1 e:1 e:1 (a) Prefix paths ending in e
null d:1 d:1 (b) Conditional FP-tree for e
null a:2
c:1
a:2 d:1 d:1 (c) Prefix paths ending in de (d) Conditional FP-tree for de
null null
a:2
c:1 c:1 (e) Prefix paths ending in ce a:2
(f) Prefix paths ending in ae Figure 6.27. Example of applying the FP-growth algorithm to ﬁnd frequent itemsets ending in e. itemsets ending in e. To do this, we must ﬁrst check whether the itemset
{e} itself is frequent. If it is frequent, we consider the subproblem of ﬁnding
frequent itemsets ending in de, followed by ce, be, and ae. In turn, each
of these subproblems are further decomposed into smaller subproblems. By
merging the solutions obtained from the subproblems, all the frequent itemsets
ending in e can be found. This divide-and-conquer approach is the key strategy
employed by the FP-growth algorithm.
For a more concrete example on how to solve the subproblems, consider
the task of ﬁnding frequent itemsets ending with e.
1. The ﬁrst step is to gather all the paths containing node e. These initial
paths are called preﬁx paths and are shown in Figure 6.27(a).
2. From the preﬁx paths shown in Figure 6.27(a), the support count for e is
obtained by adding the support counts associated with node e. Assuming
that the minimum support count is 2, {e} is declared a frequent itemset
because its support count is 3. 6.6 FP-Growth Algorithm 369 3. Because {e} is frequent, the algorithm has to solve the subproblems of
ﬁnding frequent itemsets ending in de, ce, be, and ae. Before solving
these subproblems, it must ﬁrst convert the preﬁx paths into a conditional FP-tree, which is structurally similar to an FP-tree, except
it is used to ﬁnd frequent itemsets ending with a particular suﬃx. A
conditional FP-tree is obtained in the following way:
(a) First, the support counts along the preﬁx paths must be updated
because some of the counts include transactions that do not contain
item e. For example, the rightmost path shown in Figure 6.27(a),
null −→ b:2 −→ c:2 −→ e:1, includes a transaction {b, c} that
does not contain item e. The counts along the preﬁx path must
therefore be adjusted to 1 to reﬂect the actual number of transactions containing {b, c, e}.
(b) The preﬁx paths are truncated by removing the nodes for e. These
nodes can be removed because the support counts along the preﬁx
paths have been updated to reﬂect only transactions that contain e
and the subproblems of ﬁnding frequent itemsets ending in de, ce,
be, and ae no longer need information about node e.
(c) After updating the support counts along the preﬁx paths, some
of the items may no longer be frequent. For example, the node b
appears only once and has a support count equal to 1, which means
that there is only one transaction that contains both b and e. Item b
can be safely ignored from subsequent analysis because all itemsets
ending in be must be infrequent.
The conditional FP-tree for e is shown in Figure 6.27(b). The tree looks
diﬀerent than the original preﬁx paths because the frequency counts have
been updated and the nodes b and e have been eliminated.
4. FP-growth uses the conditional FP-tree for e to solve the subproblems of
ﬁnding frequent itemsets ending in de, ce, and ae. To ﬁnd the frequent
itemsets ending in de, the preﬁx paths for d are gathered from the conditional FP-tree for e (Figure 6.27(c)). By adding the frequency counts
associated with node d, we obtain the support count for {d, e}. Since
the support count is equal to 2, {d, e} is declared a frequent itemset.
Next, the algorithm constructs the conditional FP-tree for de using the
approach described in step 3. After updating the support counts and
removing the infrequent item c, the conditional FP-tree for de is shown
in Figure 6.27(d). Since the conditional FP-tree contains only one item, 370 Chapter 6 Association Analysis a, whose support is equal to minsup, the algorithm extracts the frequent itemset {a, d, e} and moves on to the next subproblem, which is
to generate frequent itemsets ending in ce. After processing the preﬁx
paths for c, only {c, e} is found to be frequent. The algorithm proceeds
to solve the next subprogram and found {a, e} to be the only frequent
itemset remaining.
This example illustrates the divide-and-conquer approach used in the FPgrowth algorithm. At each recursive step, a conditional FP-tree is constructed
by updating the frequency counts along the preﬁx paths and removing all
infrequent items. Because the subproblems are disjoint, FP-growth will not
generate any duplicate itemsets. In addition, the counts associated with the
nodes allow the algorithm to perform support counting while generating the
common suﬃx itemsets.
FP-growth is an interesting algorithm because it illustrates how a compact
representation of the transaction data set helps to eﬃciently generate frequent
itemsets. In addition, for certain transaction data sets, FP-growth outperforms
the standard Apriori algorithm by several orders of magnitude. The run-time
performance of FP-growth depends on the compaction factor of the data
set. If the resulting conditional FP-trees are very bushy (in the worst case, a
full preﬁx tree), then the performance of the algorithm degrades signiﬁcantly
because it has to generate a large number of subproblems and merge the results
returned by each subproblem. 6.7 Evaluation of Association Patterns Association analysis algorithms have the potential to generate a large number
of patterns. For example, although the data set shown in Table 6.1 contains
only six items, it can produce up to hundreds of association rules at certain
support and conﬁdence thresholds. As the size and dimensionality of real
commercial databases can be very large, we could easily end up with thousands
or even millions of patterns, many of which might not be interesting. Sifting
through the patterns to identify the most interesting ones is not a trivial task
because “one person’s trash might be another person’s treasure.” It is therefore
important to establish a set of well-accepted criteria for evaluating the quality
of association patterns.
The ﬁrst set of criteria can be established through statistical arguments.
Patterns that involve a set of mutually independent items or cover very few
transactions are considered uninteresting because they may capture spurious
relationships in the data. Such patterns can be eliminated by applying an 6.7 Evaluation of Association Patterns 371 objective interestingness measure that uses statistics derived from data
to determine whether a pattern is interesting. Examples of objective interestingness measures include support, conﬁdence, and correlation.
The second set of criteria can be established through subjective arguments.
A pattern is considered subjectively uninteresting unless it reveals unexpected
information about the data or provides useful knowledge that can lead to
proﬁtable actions. For example, the rule {Butter} −→ {Bread} may not be
interesting, despite having high support and conﬁdence values, because the
relationship represented by the rule may seem rather obvious. On the other
hand, the rule {Diapers} −→ {Beer} is interesting because the relationship is
quite unexpected and may suggest a new cross-selling opportunity for retailers.
Incorporating subjective knowledge into pattern evaluation is a diﬃcult task
because it requires a considerable amount of prior information from the domain
experts.
The following are some of the approaches for incorporating subjective
knowledge into the pattern discovery task.
Visualization This approach requires a user-friendly environment to keep
the human user in the loop. It also allows the domain experts to interact with
the data mining system by interpreting and verifying the discovered patterns.
Template-based approach This approach allows the users to constrain
the type of patterns extracted by the mining algorithm. Instead of reporting
all the extracted rules, only rules that satisfy a user-speciﬁed template are
returned to the users.
Subjective interestingness measure A subjective measure can be deﬁned
based on domain information such as concept hierarchy (to be discussed in
Section 7.3) or proﬁt margin of items. The measure can then be used to ﬁlter
patterns that are obvious and non-actionable.
Readers interested in subjective interestingness measures may refer to resources listed in the bibliography at the end of this chapter. 6.7.1 Objective Measures of Interestingness An objective measure is a data-driven approach for evaluating the quality
of association patterns. It is domain-independent and requires minimal input from the users, other than to specify a threshold for ﬁltering low-quality
patterns. An objective measure is usually computed based on the frequency 372 Chapter 6 Association Analysis Table 6.7. A 2-way contingency table for variables A and B .
B B A f11 f10 f1+ A f01 f00 f0+ f+1 f+0 N counts tabulated in a contingency table. Table 6.7 shows an example of a
contingency table for a pair of binary variables, A and B . We use the notation
A (B ) to indicate that A (B ) is absent from a transaction. Each entry fij in
this 2 × 2 table denotes a frequency count. For example, f11 is the number of
times A and B appear together in the same transaction, while f01 is the number of transactions that contain B but not A. The row sum f1+ represents
the support count for A, while the column sum f+1 represents the support
count for B . Finally, even though our discussion focuses mainly on asymmetric binary variables, note that contingency tables are also applicable to other
attribute types such as symmetric binary, nominal, and ordinal variables.
Limitations of the Support-Conﬁdence Framework Existing association rule mining formulation relies on the support and conﬁdence measures to
eliminate uninteresting patterns. The drawback of support was previously described in Section 6.8, in which many potentially interesting patterns involving
low support items might be eliminated by the support threshold. The drawback of conﬁdence is more subtle and is best demonstrated with the following
example.
Example 6.3. Suppose we are interested in analyzing the relationship between people who drink tea and coﬀee. We may gather information about the
beverage preferences among a group of people and summarize their responses
into a table such as the one shown in Table 6.8.
Table 6.8. Beverage preferences among a group of 1000 people.
Cof f ee Cof f ee T ea 150 50 200 T ea 650 150 800 800 200 1000 6.7 Evaluation of Association Patterns 373 The information given in this table can be used to evaluate the association
rule {T ea} −→ {Cof f ee}. At ﬁrst glance, it may appear that people who drink
tea also tend to drink coﬀee because the rule’s support (15%) and conﬁdence
(75%) values are reasonably high. This argument would have been acceptable
except that the fraction of people who drink coﬀee, regardless of whether they
drink tea, is 80%, while the fraction of tea drinkers who drink coﬀee is only
75%. Thus knowing that a person is a tea drinker actually decreases her
probability of being a coﬀee drinker from 80% to 75%! The rule {T ea} −→
{Cof f ee} is therefore misleading despite its high conﬁdence value.
The pitfall of conﬁdence can be traced to the fact that the measure ignores
the support of the itemset in the rule consequent. Indeed, if the support of
coﬀee drinkers is taken into account, we would not be surprised to ﬁnd that
many of the people who drink tea also drink coﬀee. What is more surprising is
that the fraction of tea drinkers who drink coﬀee is actually less than the overall
fraction of people who drink coﬀee, which points to an inverse relationship
between tea drinkers and coﬀee drinkers.
Because of the limitations in the support-conﬁdence framework, various
objective measures have been used to evaluate the quality of association patterns. Below, we provide a brief description of these measures and explain
some of their strengths and limitations.
Interest Factor The tea-coﬀee example shows that high-conﬁdence rules
can sometimes be misleading because the conﬁdence measure ignores the support of the itemset appearing in the rule consequent. One way to address this
problem is by applying a metric known as lift:
Lif t = c(A −→ B )
,
s(B ) (6.4) which computes the ratio between the rule’s conﬁdence and the support of
the itemset in the rule consequent. For binary variables, lift is equivalent to
another objective measure called interest factor, which is deﬁned as follows:
I (A, B ) = N f11
s(A, B )
=
.
s(A) × s(B )
f1+ f+1 (6.5) Interest factor compares the frequency of a pattern against a baseline frequency computed under the statistical independence assumption. The baseline
frequency for a pair of mutually independent variables is
f11
f1+ f+1
=
×
,
N
N
N or equivalently, f11 = f1+ f+1
.
N (6.6) 374 Chapter 6 Association Analysis Table 6.9. Contingency tables for the word pairs ({p,q } and {r,s}.
p p r q 880 50 930 q 50 20 70 930 70 1000 r s 20 50 70 s 50 880 930 70 930 1000 This equation follows from the standard approach of using simple fractions
as estimates for probabilities. The fraction f11 /N is an estimate for the joint
probability P (A, B ), while f1+ /N and f+1 /N are the estimates for P (A) and
P (B ), respectively. If A and B are statistically independent, then P (A, B ) =
P (A) × P (B ), thus leading to the formula shown in Equation 6.6. Using
Equations 6.5 and 6.6, we can interpret the measure as follows:
⎧
⎨ = 1, if A and B are independent;
> 1, if A and B are positively correlated;
I (A, B )
⎩
< 1, if A and B are negatively correlated. (6.7) 0.15
For the tea-coﬀee example shown in Table 6.8, I = 0.2×0.8 = 0.9375, thus suggesting a slight negative correlation between tea drinkers and coﬀee drinkers. Limitations of Interest Factor We illustrate the limitation of interest
factor with an example from the text mining domain. In the text domain, it
is reasonable to assume that the association between a pair of words depends
on the number of documents that contain both words. For example, because
of their stronger association, we expect the words data and mining to appear
together more frequently than the words compiler and mining in a collection
of computer science articles.
Table 6.9 shows the frequency of occurrences between two pairs of words,
{p, q } and {r, s}. Using the formula given in Equation 6.5, the interest factor
for {p, q } is 1.02 and for {r, s} is 4.08. These results are somewhat troubling
for the following reasons. Although p and q appear together in 88% of the
documents, their interest factor is close to 1, which is the value when p and q
are statistically independent. On the other hand, the interest factor for {r, s}
is higher than {p, q } even though r and s seldom appear together in the same
document. Conﬁdence is perhaps the better choice in this situation because it
considers the association between p and q (94.6%) to be much stronger than
that between r and s (28.6%). 6.7 Evaluation of Association Patterns 375 Correlation Analysis Correlation analysis is a statistical-based technique
for analyzing relationships between a pair of variables. For continuous variables, correlation is deﬁned using Pearson’s correlation coeﬃcient (see Equation 2.10 on page 77). For binary variables, correlation can be measured using
the φ-coeﬃcient, which is deﬁned as
φ= f11 f00 − f01 f10
.
f1+ f+1 f0+ f+0 (6.8) The value of correlation ranges from −1 (perfect negative correlation) to +1
(perfect positive correlation). If the variables are statistically independent,
then φ = 0. For example, the correlation between the tea and coﬀee drinkers
given in Table 6.8 is −0.0625.
Limitations of Correlation Analysis The drawback of using correlation
can be seen from the word association example given in Table 6.9. Although
the words p and q appear together more often than r and s, their φ-coeﬃcients
are identical, i.e., φ(p, q ) = φ(r, s) = 0.232. This is because the φ-coeﬃcient
gives equal importance to both co-presence and co-absence of items in a transaction. It is therefore more suitable for analyzing symmetric binary variables.
Another limitation of this measure is that it does not remain invariant when
there are proportional changes to the sample size. This issue will be discussed
in greater detail when we describe the properties of objective measures on page
377.
IS Measure IS is an alternative measure that has been proposed for handling asymmetric binary variables. The measure is deﬁned as follows:
IS (A, B ) = I (A, B ) × s(A, B ) = s(A, B )
s(A)s(B ) . (6.9) Note that IS is large when the interest factor and support of the pattern
are large. For example, the value of IS for the word pairs {p, q } and {r, s}
shown in Table 6.9 are 0.946 and 0.286, respectively. Contrary to the results
given by interest factor and the φ-coeﬃcient, the IS measure suggests that
the association between {p, q } is stronger than {r, s}, which agrees with what
we expect from word associations in documents.
It is possible to show that IS is mathematically equivalent to the cosine
measure for binary variables (see Equation 2.7 on page 75). In this regard, we 376 Chapter 6 Association Analysis Table 6.10. Example of a contingency table for items p and q .
q q p 800 100 900 p 100 0 100 900 100 1000 consider A and B as a pair of bit vectors, A • B = s(A, B ) the dot product
between the vectors, and |A| = s(A) the magnitude of vector A. Therefore:
s(A, B ) IS (A, B ) = s(A) × s(B ) = A•B
= cosine(A, B).
|A| × |B| (6.10) The IS measure can also be expressed as the geometric mean between the
conﬁdence of association rules extracted from a pair of binary variables: IS (A, B ) = s(A, B ) s(A, B )
×
=
s(A)
s(B ) c(A → B ) × c(B → A). (6.11) Because the geometric mean between any two numbers is always closer to the
smaller number, the IS value of an itemset {p, q } is low whenever one of its
rules, p −→ q or q −→ p, has low conﬁdence.
Limitations of IS Measure
sets, A and B , is
ISindep (A, B ) = The IS value for a pair of independent item- s(A, B )
s(A) × s(B ) = s(A) × s(B )
s(A) × s(B ) = s(A) × s(B ). Since the value depends on s(A) and s(B ), IS shares a similar problem as
the conﬁdence measure—that the value of the measure can be quite large,
even for uncorrelated and negatively correlated patterns. For example, despite
the large IS value between items p and q given in Table 6.10 (0.889), it is
still less than the expected value when the items are statistically independent
(ISindep = 0.9). 6.7 Evaluation of Association Patterns 377 Alternative Objective Interestingness Measures
Besides the measures we have described so far, there are other alternative measures proposed for analyzing relationships between pairs of binary variables.
These measures can be divided into two categories, symmetric and asymmetric measures. A measure M is symmetric if M (A −→ B ) = M (B −→ A).
For example, interest factor is a symmetric measure because its value is identical for the rules A −→ B and B −→ A. In contrast, conﬁdence is an
asymmetric measure since the conﬁdence for A −→ B and B −→ A may not
be the same. Symmetric measures are generally used for evaluating itemsets,
while asymmetric measures are more suitable for analyzing association rules.
Tables 6.11 and 6.12 provide the deﬁnitions for some of these measures in
terms of the frequency counts of a 2 × 2 contingency table.
Consistency among Objective Measures
Given the wide variety of measures available, it is reasonable to question
whether the measures can produce similar ordering results when applied to
a set of association patterns. If the measures are consistent, then we can
choose any one of them as our evaluation metric. Otherwise, it is important
to understand what their diﬀerences are in order to determine which measure
is more suitable for analyzing certain types of patterns.
Table 6.11. Examples of symmetric objective measures for the itemset {A, B }.
Measure (Symbol) Deﬁnition Correlation (φ) N
√ f11 −f1+ f+1 Odds ratio (α)
Kappa (κ) f1+ f+1 f0+ f+0 f11 f00 f10 f01 N f11 +N f00 −f1+ f+1 −f0+ f+0
N 2 −f1+ f+1 −f0+ f+0 Interest (I ) N f11 Cosine (IS ) f11 f1+ f+1
f1+ f+1
f1+ f+1
N2 Piatetsky-Shapiro (P S ) f11
N Collective strength (S ) f11 +f00
f1+ f+1 +f0+ f+0 Jaccard (ζ ) f11 f1+ + f+1 − f11 All-conﬁdence (h) min f11 f11
f1+ , f+1 − × N −f1+ f+1 −f0+ f+0
N −f11 −f00 378 Chapter 6 Association Analysis Table 6.12. Examples of asymmetric objective measures for the rule A −→ B .
Measure (Symbol) Deﬁnition Goodman-Kruskal (λ)
Mutual Information (M )
J-Measure (J )
Gini index (G) j maxk fjk − maxk f+k N − maxk f+k fij
N fij
i+
− i fN
i
j N log fi+ f+j
f11
N f11
f10
N f10
N log f1+ f+1 + N log f1+ f+0
f1+
f+1 2
f11 2
f10 2
N × ( f1+ ) + ( f1+ ) ] − ( N )
f01
f00
0+
+0
+ fN × [( f0+ )2 + ( f0+ )2 ] − ( fN )2 Laplace (L) f11 + 1
f1+ f+0 fi+
N f1+ + 2 Conviction (V ) log N f10 Certainty factor (F )
Added Value (AV ) f+1
f11
f1+ − N
f+1
f11
f1+ − N 1− f+1
N Table 6.13. Example of contingency tables.
Example
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10 f11
8123
8330
3954
2886
1500
4000
9481
4000
7450
61 f10
83
2
3080
1363
2000
2000
298
2000
2483
2483 f01
424
622
5
1320
500
1000
127
2000
4
4 f00
1370
1046
2961
4431
6000
3000
94
2000
63
7452 Suppose the symmetric and asymmetric measures are applied to rank the
ten contingency tables shown in Table 6.13. These contingency tables are chosen to illustrate the diﬀerences among the existing measures. The ordering
produced by these measures are shown in Tables 6.14 and 6.15, respectively
(with 1 as the most interesting and 10 as the least interesting table). Although
some of the measures appear to be consistent with each other, there are certain
measures that produce quite diﬀerent ordering results. For example, the rankings given by the φ-coeﬃcient agree with those provided by κ and collective
strength, but are somewhat diﬀerent than the rankings produced by interest 6.7 Evaluation of Association Patterns 379 Table 6.14. Rankings of contingency tables using the symmetric measures given in Table 6.11.
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10 φ
1
2
3
4
5
6
7
8
9
10 α
3
1
2
8
7
9
6
10
4
5 κ
1
2
4
3
6
5
7
8
9
10 I
6
7
4
3
2
5
9
8
10
1 IS
2
3
5
7
9
6
1
8
4
10 PS
2
5
1
3
6
4
8
7
9
10 S
1
2
3
4
6
5
7
8
9
10 ζ
2
3
6
7
9
5
1
8
4
10 h
2
3
8
5
9
7
1
7
4
10 Table 6.15. Rankings of contingency tables using the asymmetric measures given in Table 6.12.
E1
E2
E3
E4
E5
E6
E7
E8
E9
E10 λ
1
2
5
4
9
3
7
8
6
10 M
1
2
3
6
7
8
5
9
4
10 J
1
2
5
3
4
6
9
7
10
8 G
1
3
2
4
6
5
8
7
9
10 L
4
5
2
9
8
7
3
10
1
6 V
2
1
6
3
5
4
7
8
9
10 F
2
1
6
3
5
4
7
8
9
10 AV
5
6
4
1
2
3
9
7
10
8 factor and odds ratio. Furthermore, a contingency table such as E10 is ranked
lowest according to the φ-coeﬃcient, but highest according to interest factor.
Properties of Objective Measures
The results shown in Table 6.14 suggest that a signiﬁcant number of the measures provide conﬂicting information about the quality of a pattern. To understand their diﬀerences, we need to examine the properties of these measures.
Inversion Property Consider the bit vectors shown in Figure 6.28. The
0/1 bit in each column vector indicates whether a transaction (row) contains
a particular item (column). For example, the vector A indicates that item a 380 Chapter 6 Association Analysis
A B C D E F 1
0
0
0
0
0
0
0
0
1 0
0
0
0
1
0
0
0
0
0 0
1
1
1
1
1
1
1
1
0 1
1
1
1
0
1
1
1
1
1 0
1
1
1
1
1
1
1
1
0 0
0
0
0
1
0
0
0
0
0 (a) (b) (c) Figure 6.28. Effect of the inversion operation. The vectors C and E are inversions of vector A, while
the vector D is an inversion of vectors B and F . belongs to the ﬁrst and last transactions, whereas the vector B indicates that
item b is contained only in the ﬁfth transaction. The vectors C and E are in
fact related to the vector A—their bits have been inverted from 0’s (absence)
to 1’s (presence), and vice versa. Similarly, D is related to vectors B and F by
inverting their bits. The process of ﬂipping a bit vector is called inversion.
If a measure is invariant under the inversion operation, then its value for the
vector pair (C, D) should be identical to its value for (A, B). The inversion
property of a measure can be tested as follows.
Deﬁnition 6.6 (Inversion Property). An objective measure M is invariant
under the inversion operation if its value remains the same when exchanging
the frequency counts f11 with f00 and f10 with f01 .
Among the measures that remain invariant under this operation include
the φ-coeﬃcient, odds ratio, κ, and collective strength. These measures may
not be suitable for analyzing asymmetric binary data. For example, the φcoeﬃcient between C and D is identical to the φ-coeﬃcient between A and
B, even though items c and d appear together more frequently than a and b.
Furthermore, the φ-coeﬃcient between C and D is less than that between E
and F even though items e and f appear together only once! We had previously
raised this issue when discussing the limitations of the φ-coeﬃcient on page
375. For asymmetric binary data, measures that do not remain invariant under
the inversion operation are preferred. Some of the non-invariant measures
include interest factor, IS , P S , and the Jaccard coeﬃcient. 6.7 Evaluation of Association Patterns 381 Null Addition Property Suppose we are interested in analyzing the relationship between a pair of words, such as data and mining, in a set of
documents. If a collection of articles about ice ﬁshing is added to the data set,
should the association between data and mining be aﬀected? This process of
adding unrelated data (in this case, documents) to a given data set is known
as the null addition operation.
Deﬁnition 6.7 (Null Addition Property). An objective measure M is
invariant under the null addition operation if it is not aﬀected by increasing
f00 , while all other frequencies in the contingency table stay the same.
For applications such as document analysis or market basket analysis, the
measure is expected to remain invariant under the null addition operation.
Otherwise, the relationship between words may disappear simply by adding
enough documents that do not contain both words! Examples of measures
that satisfy this property include cosine (IS ) and Jaccard (ξ ) measures, while
those that violate this property include interest factor, P S , odds ratio, and
the φ-coeﬃcient.
Scaling Property Table 6.16 shows the contingency tables for gender and
the grades achieved by students enrolled in a particular course in 1993 and
2004. The data in these tables showed that the number of male students has
doubled since 1993, while the number of female students has increased by a
factor of 3. However, the male students in 2004 are not performing any better
than those in 1993 because the ratio of male students who achieve a high
grade to those who achieve a low grade is still the same, i.e., 3:4. Similarly,
the female students in 2004 are performing no better than those in 1993. The
association between grade and gender is expected to remain unchanged despite
changes in the sampling distribution.
Table 6.16. The grade-gender example.
High
Low Male
30
40
70 Female
20
10
30 50
50
100 (a) Sample data from 1993. High
Low Male
60
80
140 Female
60
30
90 120
110
230 (b) Sample data from 2004. 382 Chapter 6 Association Analysis
Table 6.17. Properties of symmetric measures. Symbol
φ
α
κ
I
IS
PS
S
ζ
h
s Measure
φ-coeﬃcient
odds ratio
Cohen’s
Interest
Cosine
Piatetsky-Shapiro’s
Collective strength
Jaccard
All-conﬁdence
Support Inversion
Yes
Yes
Yes
No
No
Yes
Yes
No
No
No Null Addition
No
No
No
No
Yes
No
No
Yes
No
No Scaling
No
Yes
No
No
No
No
No
No
No
No Deﬁnition 6.8 (Scaling Invariance Property). An objective measure M
is invariant under the row/column scaling operation if M (T ) = M (T ), where
T is a contingency table with frequency counts [f11 ; f10 ; f01 ; f00 ], T is a
contingency table with scaled frequency counts [k1 k3 f11 ; k2 k3 f10 ; k1 k4 f01 ;
k2 k4 f00 ], and k1 , k2 , k3 , k4 are positive constants.
From Table 6.17, notice that only the odds ratio (α) is invariant under
the row and column scaling operations. All other measures such as the φcoeﬃcient, κ, IS , interest factor, and collective strength (S ) change their values when the rows and columns of the contingency table are rescaled. Although
we do not discuss the properties of asymmetric measures (such as conﬁdence,
J-measure, Gini index, and conviction), it is clear that such measures do not
preserve their values under inversion and row/column scaling operations, but
are invariant under the null addition operation. 6.7.2 Measures beyond Pairs of Binary Variables The measures shown in Tables 6.11 and 6.12 are deﬁned for pairs of binary variables (e.g., 2-itemsets or association rules). However, many of them, such as
support and all-conﬁdence, are also applicable to larger-sized itemsets. Other
measures, such as interest factor, IS , P S , and Jaccard coeﬃcient, can be extended to more than two variables using the frequency tables tabulated in a
multidimensional contingency table. An example of a three-dimensional contingency table for a, b, and c is shown in Table 6.18. Each entry fijk in this
table represents the number of transactions that contain a particular combination of items a, b, and c. For example, f101 is the number of transactions
that contain a and c, but not b. On the other hand, a marginal frequency 6.7 Evaluation of Association Patterns 383 Table 6.18. Example of a three-dimensional contingency table. c b b a f111 f101 a f011
f+11 c b b f1+1 a f110 f100 f1+0 f001 f0+1 a f010 f000 f0+0 f+01 f++1 f+10 f+00 f++0 such as f1+1 is the number of transactions that contain a and c, irrespective
of whether b is present in the transaction.
Given a k -itemset {i1 , i2 , . . . , ik }, the condition for statistical independence
can be stated as follows:
fi1 i2 ...ik = fi1 +...+ × f+i2 ...+ × . . . × f++...ik
.
N k −1 (6.12) With this deﬁnition, we can extend objective measures such as interest factor
and P S , which are based on deviations from statistical independence, to more
than two variables:
I=
PS = N k−1 × fi1 i2 ...ik
fi1 +...+ × f+i2 ...+ × . . . × f++...ik
fi +...+ × f+i2 ...+ × . . . × f++...ik
fi1 i2 ...ik
−1
N
Nk Another approach is to deﬁne the objective measure as the maximum, minimum, or average value for the associations between pairs of items in a pattern. For example, given a k -itemset X = {i1 , i2 , . . . , ik }, we may deﬁne the
φ-coeﬃcient for X as the average φ-coeﬃcient between every pair of items
(ip , iq ) in X . However, because the measure considers only pairwise associations, it may not capture all the underlying relationships within a pattern.
Analysis of multidimensional contingency tables is more complicated because of the presence of partial associations in the data. For example, some
associations may appear or disappear when conditioned upon the value of certain variables. This problem is known as Simpson’s paradox and is described
in the next section. More sophisticated statistical techniques are available to
analyze such relationships, e.g., loglinear models, but these techniques are
beyond the scope of this book. 384 Chapter 6 Association Analysis Table 6.19. A two-way contingency table between the sale of high-deﬁnition television and exercise
machine.
Buy
Buy Exercise Machine
HDTV
Yes
No
Yes
99
81
180
No
54
66
120
153
147
300
Table 6.20. Example of a three-way contingency table.
Customer
Group
College Students
Working Adult 6.7.3 Buy
HDTV
Yes
No
Yes
No Buy Exercise
Yes
1
4
98
50 Machine
No
9
30
72
36 Total
10
34
170
86 Simpson’s Paradox It is important to exercise caution when interpreting the association between
variables because the observed relationship may be inﬂuenced by the presence
of other confounding factors, i.e., hidden variables that are not included in
the analysis. In some cases, the hidden variables may cause the observed
relationship between a pair of variables to disappear or reverse its direction, a
phenomenon that is known as Simpson’s paradox. We illustrate the nature of
this paradox with the following example.
Consider the relationship between the sale of high-deﬁnition television
(HDTV) and exercise machine, as shown in Table 6.19. The rule {HDTV=Yes}
−→ {Exercise machine=Yes} has a conﬁdence of 99/180 = 55% and the rule
{HDTV=No} −→ {Exercise machine=Yes} has a conﬁdence of 54/120 = 45%.
Together, these rules suggest that customers who buy high-deﬁnition televisions are more likely to buy exercise machines than those who do not buy
high-deﬁnition televisions.
However, a deeper analysis reveals that the sales of these items depend
on whether the customer is a college student or a working adult. Table 6.20
summarizes the relationship between the sale of HDTVs and exercise machines
among college students and working adults. Notice that the support counts
given in the table for college students and working adults sum up to the frequencies shown in Table 6.19. Furthermore, there are more working adults 6.7 Evaluation of Association Patterns 385 than college students who buy these items. For college students:
c {HDTV=Yes} −→ {Exercise machine=Yes} = 1/10 = 10%, c {HDTV=No} −→ {Exercise machine=Yes} = 4/34 = 11.8%, while for working adults:
c {HDTV=Yes} −→ {Exercise machine=Yes} = 98/170 = 57.7%, c {HDTV=No} −→ {Exercise machine=Yes} = 50/86 = 58.1%. The rules suggest that, for each group, customers who do not buy highdeﬁnition televisions are more likely to buy exercise machines, which contradict
the previous conclusion when data from the two customer groups are pooled
together. Even if alternative measures such as correlation, odds ratio, or
interest are applied, we still ﬁnd that the sale of HDTV and exercise machine
is positively correlated in the combined data but is negatively correlated in
the stratiﬁed data (see Exercise 20 on page 414). The reversal in the direction
of association is known as Simpson’s paradox.
The paradox can be explained in the following way. Notice that most
customers who buy HDTVs are working adults. Working adults are also the
largest group of customers who buy exercise machines. Because nearly 85% of
the customers are working adults, the observed relationship between HDTV
and exercise machine turns out to be stronger in the combined data than
what it would have been if the data is stratiﬁed. This can also be illustrated
mathematically as follows. Suppose
a/b < c/d and p/q < r/s,
where a/b and p/q may represent the conﬁdence of the rule A −→ B in two
diﬀerent strata, while c/d and r/s may represent the conﬁdence of the rule
A −→ B in the two strata. When the data is pooled together, the conﬁdence
values of the rules in the combined data are (a + p)/(b + q ) and (c + r)/(d + s),
respectively. Simpson’s paradox occurs when
c+r
a+p
>
,
b+q
d+s
thus leading to the wrong conclusion about the relationship between the variables. The lesson here is that proper stratiﬁcation is needed to avoid generating spurious patterns resulting from Simpson’s paradox. For example, market 386 Chapter 6 Association Analysis
×10 4 5
4.5
4
3.5 Support 3
2.5
2
1.5
1
0.5
0 0 500 1000
1500
Items sorted by support 2000 2500 Figure 6.29. Support distribution of items in the census data set. basket data from a major supermarket chain should be stratiﬁed according to
store locations, while medical records from various patients should be stratiﬁed
according to confounding factors such as age and gender. 6.8 Eﬀect of Skewed Support Distribution The performances of many association analysis algorithms are inﬂuenced by
properties of their input data. For example, the computational complexity of
the Apriori algorithm depends on properties such as the number of items in
the data and average transaction width. This section examines another important property that has signiﬁcant inﬂuence on the performance of association
analysis algorithms as well as the quality of extracted patterns. More speciﬁcally, we focus on data sets with skewed support distributions, where most of
the items have relatively low to moderate frequencies, but a small number of
them have very high frequencies.
An example of a real data set that exhibits such a distribution is shown in
Figure 6.29. The data, taken from the PUMS (Public Use Microdata Sample)
census data, contains 49,046 records and 2113 asymmetric binary variables.
We shall treat the asymmetric binary variables as items and records as transactions in the remainder of this section. While more than 80% of the items
have support less than 1%, a handful of them have support greater than 90%. 6.8 Eﬀect of Skewed Support Distribution 387 Table 6.21. Grouping the items in the census data set based on their support values.
Group
Support
Number of Items G1
< 1%
1735 G2
1% − 90%
358 G3
> 90%
20 To illustrate the eﬀect of skewed support distribution on frequent itemset mining, we divide the items into three groups, G1 , G2 , and G3 , according to their
support levels. The number of items that belong to each group is shown in
Table 6.21.
Choosing the right support threshold for mining this data set can be quite
tricky. If we set the threshold too high (e.g., 20%), then we may miss many
interesting patterns involving the low support items from G1 . In market basket analysis, such low support items may correspond to expensive products
(such as jewelry) that are seldom bought by customers, but whose patterns
are still interesting to retailers. Conversely, when the threshold is set too
low, it becomes diﬃcult to ﬁnd the association patterns due to the following
reasons. First, the computational and memory requirements of existing association analysis algorithms increase considerably with low support thresholds.
Second, the number of extracted patterns also increases substantially with low
support thresholds. Third, we may extract many spurious patterns that relate
a high-frequency item such as milk to a low-frequency item such as caviar.
Such patterns, which are called cross-support patterns, are likely to be spurious because their correlations tend to be weak. For example, at a support
threshold equal to 0.05%, there are 18,847 frequent pairs involving items from
G1 and G3 . Out of these, 93% of them are cross-support patterns; i.e., the patterns contain items from both G1 and G3 . The maximum correlation obtained
from the cross-support patterns is 0.029, which is much lower than the maximum correlation obtained from frequent patterns involving items from the
same group (which is as high as 1.0). Similar statement can be made about
many other interestingness measures discussed in the previous section. This
example shows that a large number of weakly correlated cross-support patterns can be generated when the support threshold is suﬃciently low. Before
presenting a methodology for eliminating such patterns, we formally deﬁne the
concept of cross-support patterns. 388 Chapter 6 Association Analysis Deﬁnition 6.9 (Cross-Support Pattern). A cross-support pattern is an
itemset X = {i1 , i2 , . . . , ik } whose support ratio
r(X ) = min s(i1 ), s(i2 ), . . . , s(ik )
,
max s(i1 ), s(i2 ), . . . , s(ik ) (6.13) is less than a user-speciﬁed threshold hc .
Example 6.4. Suppose the support for milk is 70%, while the support for
sugar is 10% and caviar is 0.04%. Given hc = 0.01, the frequent itemset
{milk, sugar, caviar} is a cross-support pattern because its support ratio is
r= min 0.7, 0.1, 0.0004
0.0004
= 0.00058 < 0.01.
=
0.7
max 0.7, 0.1, 0.0004 Existing measures such as support and conﬁdence may not be suﬃcient
to eliminate cross-support patterns, as illustrated by the data set shown in
Figure 6.30. Assuming that hc = 0.3, the itemsets {p, q }, {p, r}, and {p, q, r}
are cross-support patterns because their support ratios, which are equal to
0.2, are less than the threshold hc . Although we can apply a high support
threshold, say, 20%, to eliminate the cross-support patterns, this may come
at the expense of discarding other interesting patterns such as the strongly
correlated itemset, {q, r} that has support equal to 16.7%.
Conﬁdence pruning also does not help because the conﬁdence of the rules
extracted from cross-support patterns can be very high. For example, the
conﬁdence for {q } −→ {p} is 80% even though {p, q } is a cross-support pattern. The fact that the cross-support pattern can produce a high-conﬁdence
rule should not come as a surprise because one of its items (p) appears very
frequently in the data. Therefore, p is expected to appear in many of the
transactions that contain q . Meanwhile, the rule {q } −→ {r} also has high
conﬁdence even though {q, r} is not a cross-support pattern. This example
demonstrates the diﬃculty of using the conﬁdence measure to distinguish between rules extracted from cross-support and non-cross-support patterns.
Returning to the previous example, notice that the rule {p} −→ {q } has
very low conﬁdence because most of the transactions that contain p do not
contain q . In contrast, the rule {r} −→ {q }, which is derived from the pattern
{q, r}, has very high conﬁdence. This observation suggests that cross-support
patterns can be detected by examining the lowest conﬁdence rule that can be
extracted from a given itemset. The proof of this statement can be understood
as follows. 6.8 Eﬀect of Skewed Support Distribution 389
p
0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0 q
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0 r
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0 Figure 6.30. A transaction data set containing three items, p, q , and r, where p is a high support item
and q and r are low support items. 1. Recall the following anti-monotone property of conﬁdence:
conf ({i1 i2 } −→ {i3 , i4 , . . . , ik }) ≤ conf ({i1 i2 i3 } −→ {i4 , i5 , . . . , ik }).
This property suggests that conﬁdence never increases as we shift more
items from the left- to the right-hand side of an association rule. Because
of this property, the lowest conﬁdence rule extracted from a frequent
itemset contains only one item on its left-hand side. We denote the set
of all rules with only one item on its left-hand side as R1 .
2. Given a frequent itemset {i1 , i2 , . . . , ik }, the rule
{ij } −→ {i1 , i2 , . . . , ij −1 , ij +1 , . . . , ik }
has the lowest conﬁdence in R1 if s(ij ) = max s(i1 ), s(i2 ), . . . , s(ik ) .
This follows directly from the deﬁnition of conﬁdence as the ratio between the rule’s support and the support of the rule antecedent. 390 Chapter 6 Association Analysis 3. Summarizing the previous points, the lowest conﬁdence attainable from
a frequent itemset {i1 , i2 , . . . , ik } is
s({i1 , i2 , . . . , ik })
.
max s(i1 ), s(i2 ), . . . , s(ik )
This expression is also known as the h-conﬁdence or all-conﬁdence
measure. Because of the anti-monotone property of support, the numerator of the h-conﬁdence measure is bounded by the minimum support
of any item that appears in the frequent itemset. In other words, the
h-conﬁdence of an itemset X = {i1 , i2 , . . . , ik } must not exceed the following expression:
h-conﬁdence(X ) ≤ min s(i1 ), s(i2 ), . . . , s(ik )
.
max s(i1 ), s(i2 ), . . . , s(ik ) Note the equivalence between the upper bound of h-conﬁdence and the
support ratio (r) given in Equation 6.13. Because the support ratio for
a cross-support pattern is always less than hc , the h-conﬁdence of the
pattern is also guaranteed to be less than hc .
Therefore, cross-support patterns can be eliminated by ensuring that the
h-conﬁdence values for the patterns exceed hc . As a ﬁnal note, it is worth
mentioning that the advantages of using h-conﬁdence go beyond eliminating
cross-support patterns. The measure is also anti-monotone, i.e.,
h-conﬁdence({i1 , i2 , . . . , ik }) ≥ h-conﬁdence({i1 , i2 , . . . , ik+1 }),
and thus can be incorporated directly into the mining algorithm. Furthermore,
h-conﬁdence ensures that the items contained in an itemset are strongly associated with each other. For example, suppose the h-conﬁdence of an itemset
X is 80%. If one of the items in X is present in a transaction, there is at least
an 80% chance that the rest of the items in X also belong to the same transaction. Such strongly associated patterns are called hyperclique patterns. 6.9 Bibliographic Notes The association rule mining task was ﬁrst introduced by Agrawal et al. in
[228, 229] to discover interesting relationships among items in market basket 6.9 Bibliographic Notes 391 transactions. Since its inception, extensive studies have been conducted to
address the various conceptual, implementation, and application issues pertaining to the association analysis task. A summary of the various research
activities in this area is shown in Figure 6.31.
Conceptual Issues
Research in conceptual issues is focused primarily on (1) developing a framework to describe the theoretical underpinnings of association analysis, (2) extending the formulation to handle new types of patterns, and (3) extending the
formulation to incorporate attribute types beyond asymmetric binary data.
Following the pioneering work by Agrawal et al., there has been a vast
amount of research on developing a theory for the association analysis problem.
In [254], Gunopoulos et al. showed a relation between the problem of ﬁnding
maximal frequent itemsets and the hypergraph transversal problem. An upper
bound on the complexity of association analysis task was also derived. Zaki et
al. [334, 336] and Pasquier et al. [294] have applied formal concept analysis to
study the frequent itemset generation problem. The work by Zaki et al. have
subsequently led them to introduce the notion of closed frequent itemsets [336].
Friedman et al. have studied the association analysis problem in the context
of bump hunting in multidimensional space [252]. More speciﬁcally, they
consider frequent itemset generation as the task of ﬁnding high probability
density regions in multidimensional space.
Over the years, new types of patterns have been deﬁned, such as proﬁle
association rules [225], cyclic association rules [290], fuzzy association rules
[273], exception rules [316], negative association rules [238, 304], weighted
association rules [240, 300], dependence rules [308], peculiar rules[340], intertransaction association rules [250, 323], and partial classiﬁcation rules [231,
285]. Other types of patterns include closed itemsets [294, 336], maximal
itemsets [234], hyperclique patterns [330], support envelopes [314], emerging
patterns [246], and contrast sets [233]. Association analysis has also been
successfully applied to sequential [230, 312], spatial [266], and graph-based
[268, 274, 293, 331, 335] data. The concept of cross-support pattern was ﬁrst
introduced by Hui et al. in [330]. An eﬃcient algorithm (called Hyperclique
Miner) that automatically eliminates cross-support patterns was also proposed
by the authors.
Substantial research has been conducted to extend the original association
rule formulation to nominal [311], ordinal [281], interval [284], and ratio [253,
255, 311, 325, 339] attributes. One of the key issues is how to deﬁne the support
measure for these attributes. A methodology was proposed by Steinbach et 392 Chapter 6 Association Analysis al. [315] to extend the traditional notion of support to more general patterns
and attribute types. Research Issues in Mining
Association Patterns Conceptual
Issues Theroretical
Formulation Implementation
Issues Type of
Patterns Data Type Pattern
Discovery -optimization
-SQL support
-OLAP
-multi-database Constraints Postprocessing -Web analysis
-text analysis
-bioinformatics
-Earth Science -negative
-dependence
-causal
-weighted
-spatial and colocation patterns
-temporal (cyclic,
sequential)
-fuzzy
-exception rules Itemsets
-closed
-maximal
-emerging
patterns
-hyperclique
patterns
-support
envelope Other
Structures Computational
model -subtrees
-subgraphs -serial or parallel
-online or batch Interestingness Algorithm and
Data Structure
-Apriori
-DIC
-tree-projeciton
-FP-tree
-H-mine
-Partition
-Sampling-based
-CHARM Measure -objective
-subjective Figure 6.31. A summary of the various research activities in association analysis. Method
-ranking
-filtering
-summarizing Other data
mining
problems
-classification
-regression
-clustering
-recommender
systems Bibliographic Notes 393 Visualization
Rules Domains -item taxonomy
-templatebased
-multiple
support 6.9 -binary
-numeric
-nominal
-ordinal
-mixed -lattice theory
-bounds on
itemset
enumeration Database
issues Application
Issues 394 Chapter 6 Association Analysis Implementation Issues
Research activities in this area revolve around (1) integrating the mining capability into existing database technology, (2) developing eﬃcient and scalable
mining algorithms, (3) handling user-speciﬁed or domain-speciﬁc constraints,
and (4) post-processing the extracted patterns.
There are several advantages to integrating association analysis into existing database technology. First, it can make use of the indexing and query
processing capabilities of the database system. Second, it can also exploit the
DBMS support for scalability, check-pointing, and parallelization [301]. The
SETM algorithm developed by Houtsma et al. [265] was one of the earliest
algorithms to support association rule discovery via SQL queries. Since then,
numerous methods have been developed to provide capabilities for mining association rules in database systems. For example, the DMQL [258] and M-SQL
[267] query languages extend the basic SQL with new operators for mining association rules. The Mine Rule operator [283] is an expressive SQL operator
that can handle both clustered attributes and item hierarchies. Tsur et al.
[322] developed a generate-and-test approach called query ﬂocks for mining
association rules. A distributed OLAP-based infrastructure was developed by
Chen et al. [241] for mining multilevel association rules.
Dunkel and Soparkar [248] investigated the time and storage complexity
of the Apriori algorithm. The FP-growth algorithm was developed by Han et
al. in [259]. Other algorithms for mining frequent itemsets include the DHP
(dynamic hashing and pruning) algorithm proposed by Park et al. [292] and
the Partition algorithm developed by Savasere et al [303]. A sampling-based
frequent itemset generation algorithm was proposed by Toivonen [320]. The
algorithm requires only a single pass over the data, but it can produce more
candidate itemsets than necessary. The Dynamic Itemset Counting (DIC)
algorithm [239] makes only 1.5 passes over the data and generates less candidate itemsets than the sampling-based algorithm. Other notable algorithms
include the tree-projection algorithm [223] and H-Mine [295]. Survey articles
on frequent itemset generation algorithms can be found in [226, 262]. A repository of data sets and algorithms is available at the Frequent Itemset Mining
Implementations (FIMI) repository (http://ﬁmi.cs.helsinki.ﬁ). Parallel algorithms for mining association patterns have been developed by various authors
[224, 256, 287, 306, 337]. A survey of such algorithms can be found in [333].
Online and incremental versions of association rule mining algorithms had also
been proposed by Hidber [260] and Cheung et al. [242].
Srikant et al. [313] have considered the problem of mining association rules
in the presence of boolean constraints such as the following: 6.9 Bibliographic Notes 395 (Cookies ∧ Milk) ∨ (descendents(Cookies) ∧ ¬ancestors(Wheat Bread))
Given such a constraint, the algorithm looks for rules that contain both cookies and milk, or rules that contain the descendent items of cookies but not
ancestor items of wheat bread. Singh et al. [310] and Ng et al. [288] had also
developed alternative techniques for constrained-based association rule mining. Constraints can also be imposed on the support for diﬀerent itemsets.
This problem was investigated by Wang et al. [324], Liu et al. in [279], and
Seno et al. [305].
One potential problem with association analysis is the large number of
patterns that can be generated by current algorithms. To overcome this problem, methods to rank, summarize, and ﬁlter patterns have been developed.
Toivonen et al. [321] proposed the idea of eliminating redundant rules using
structural rule covers and to group the remaining rules using clustering.
Liu et al. [280] applied the statistical chi-square test to prune spurious patterns
and summarized the remaining patterns using a subset of the patterns called
direction setting rules. The use of objective measures to ﬁlter patterns
has been investigated by many authors, including Brin et al. [238], Bayardo
and Agrawal [235], Aggarwal and Yu [227], and DuMouchel and Pregibon[247].
The properties for many of these measures were analyzed by Piatetsky-Shapiro
[297], Kamber and Singhal [270], Hilderman and Hamilton [261], and Tan et
al. [318]. The grade-gender example used to highlight the importance of the
row and column scaling invariance property was heavily inﬂuenced by the
discussion given in [286] by Mosteller. Meanwhile, the tea-coﬀee example illustrating the limitation of conﬁdence was motivated by an example given in
[238] by Brin et al. Because of the limitation of conﬁdence, Brin et al. [238]
had proposed the idea of using interest factor as a measure of interestingness. The all-conﬁdence measure was proposed by Omiecinski [289]. Xiong
et al. [330] introduced the cross-support property and showed that the allconﬁdence measure can be used to eliminate cross-support patterns. A key
diﬃculty in using alternative objective measures besides support is their lack
of a monotonicity property, which makes it diﬃcult to incorporate the measures directly into the mining algorithms. Xiong et al. [328] have proposed
an eﬃcient method for mining correlations by introducing an upper bound
function to the φ-coeﬃcient. Although the measure is non-monotone, it has
an upper bound expression that can be exploited for the eﬃcient mining of
strongly correlated itempairs.
Fabris and Freitas [249] have proposed a method for discovering interesting associations by detecting the occurrences of Simpson’s paradox [309].
Megiddo and Srikant [282] described an approach for validating the extracted 396 Chapter 6 Association Analysis patterns using hypothesis testing methods. A resampling-based technique was
also developed to avoid generating spurious patterns because of the multiple
comparison problem. Bolton et al. [237] have applied the Benjamini-Hochberg
[236] and Bonferroni correction methods to adjust the p-values of discovered
patterns in market basket data. Alternative methods for handling the multiple
comparison problem were suggested by Webb [326] and Zhang et al. [338].
Application of subjective measures to association analysis has been investigated by many authors. Silberschatz and Tuzhilin [307] presented two principles in which a rule can be considered interesting from a subjective point of
view. The concept of unexpected condition rules was introduced by Liu et al.
in [277]. Cooley et al. [243] analyzed the idea of combining soft belief sets
using the Dempster-Shafer theory and applied this approach to identify contradictory and novel association patterns in Web data. Alternative approaches
include using Bayesian networks [269] and neighborhood-based information
[245] to identify subjectively interesting patterns.
Visualization also helps the user to quickly grasp the underlying structure of the discovered patterns. Many commercial data mining tools display
the complete set of rules (which satisfy both support and conﬁdence threshold criteria) as a two-dimensional plot, with each axis corresponding to the
antecedent or consequent itemsets of the rule. Hofmann et al. [263] proposed
using Mosaic plots and Double Decker plots to visualize association rules. This
approach can visualize not only a particular rule, but also the overall contingency table between itemsets in the antecedent and consequent parts of the
rule. Nevertheless, this technique assumes that the rule consequent consists of
only a single attribute.
Application Issues
Association analysis has been applied to a variety of application domains such
as Web mining [296, 317], document analysis [264], telecommunication alarm
diagnosis [271], network intrusion detection [232, 244, 275], and bioinformatics
[302, 327]. Applications of association and correlation pattern analysis to
Earth Science studies have been investigated in [298, 299, 319].
Association patterns have also been applied to other learning problems
such as classiﬁcation [276, 278], regression [291], and clustering [257, 329, 332].
A comparison between classiﬁcation and association rule mining was made
by Freitas in his position paper [251]. The use of association patterns for
clustering has been studied by many authors including Han et al.[257], Kosters
et al. [272], Yang et al. [332] and Xiong et al. [329]. Bibliography 397 Bibliography
[223] R. C. Agarwal, C. C. Aggarwal, and V. V. V. Prasad. A Tree Projection Algorithm
for Generation of Frequent Itemsets. Journal of Parallel and Distributed Computing
(Special Issue on High Performance Data Mining), 61(3):350–371, 2001.
[224] R. C. Agarwal and J. C. Shafer. Parallel Mining of Association Rules. IEEE Transactions on Knowledge and Data Engineering, 8(6):962–969, March 1998.
[225] C. C. Aggarwal, Z. Sun, and P. S. Yu. Online Generation of Proﬁle Association Rules.
In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 129–
133, New York, NY, August 1996.
[226] C. C. Aggarwal and P. S. Yu. Mining Large Itemsets for Association Rules. Data
Engineering Bulletin, 21(1):23–31, March 1998.
[227] C. C. Aggarwal and P. S. Yu. Mining Associations with the Collective Strength
Approach. IEEE Trans. on Knowledge and Data Engineering, 13(6):863–873, January/February 2001.
[228] R. Agrawal, T. Imielinski, and A. Swami. Database mining: A performance perspective. IEEE Transactions on Knowledge and Data Engineering, 5:914–925, 1993.
[229] R. Agrawal, T. Imielinski, and A. Swami. Mining association rules between sets of
items in large databases. In Proc. ACM SIGMOD Intl. Conf. Management of Data,
pages 207–216, Washington, DC, 1993.
[230] R. Agrawal and R. Srikant. Mining Sequential Patterns. In Proc. of Intl. Conf. on
Data Engineering, pages 3–14, Taipei, Taiwan, 1995.
[231] K. Ali, S. Manganaris, and R. Srikant. Partial Classiﬁcation using Association Rules.
In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 115–
118, Newport Beach, CA, August 1997.
[232] D. Barbar´, J. Couto, S. Jajodia, and N. Wu. ADAM: A Testbed for Exploring the
a
Use of Data Mining in Intrusion Detection. SIGMOD Record, 30(4):15–24, 2001.
[233] S. D. Bay and M. Pazzani. Detecting Group Diﬀerences: Mining Contrast Sets. Data
Mining and Knowledge Discovery, 5(3):213–246, 2001.
[234] R. Bayardo. Eﬃciently Mining Long Patterns from Databases. In Proc. of 1998 ACMSIGMOD Intl. Conf. on Management of Data, pages 85–93, Seattle, WA, June 1998.
[235] R. Bayardo and R. Agrawal. Mining the Most Interesting Rules. In Proc. of the 5th
Intl. Conf. on Knowledge Discovery and Data Mining, pages 145–153, San Diego, CA,
August 1999.
[236] Y. Benjamini and Y. Hochberg. Controlling the False Discovery Rate: A Practical
and Powerful Approach to Multiple Testing. Journal Royal Statistical Society B, 57
(1):289–300, 1995.
[237] R. J. Bolton, D. J. Hand, and N. M. Adams. Determining Hit Rate in Pattern Search.
In Proc. of the ESF Exploratory Workshop on Pattern Detection and Discovery in
Data Mining, pages 36–48, London, UK, September 2002.
[238] S. Brin, R. Motwani, and C. Silverstein. Beyond market baskets: Generalizing association rules to correlations. In Proc. ACM SIGMOD Intl. Conf. Management of Data,
pages 265–276, Tucson, AZ, 1997.
[239] S. Brin, R. Motwani, J. Ullman, and S. Tsur. Dynamic Itemset Counting and Implication Rules for market basket data. In Proc. of 1997 ACM-SIGMOD Intl. Conf. on
Management of Data, pages 255–264, Tucson, AZ, June 1997.
[240] C. H. Cai, A. Fu, C. H. Cheng, and W. W. Kwong. Mining Association Rules with
Weighted Items. In Proc. of IEEE Intl. Database Engineering and Applications Symp.,
pages 68–77, Cardiﬀ, Wales, 1998. 398 Chapter 6 Association Analysis [241] Q. Chen, U. Dayal, and M. Hsu. A Distributed OLAP infrastructure for E-Commerce.
In Proc. of the 4th IFCIS Intl. Conf. on Cooperative Information Systems, pages 209–
220, Edinburgh, Scotland, 1999.
[242] D. C. Cheung, S. D. Lee, and B. Kao. A General Incremental Technique for Maintaining
Discovered Association Rules. In Proc. of the 5th Intl. Conf. on Database Systems for
Advanced Applications, pages 185–194, Melbourne, Australia, 1997.
[243] R. Cooley, P. N. Tan, and J. Srivastava. Discovery of Interesting Usage Patterns
from Web Data. In M. Spiliopoulou and B. Masand, editors, Advances in Web Usage
Analysis and User Proﬁling, volume 1836, pages 163–182. Lecture Notes in Computer
Science, 2000.
[244] P. Dokas, L. Ert¨z, V. Kumar, A. Lazarevic, J. Srivastava, and P. N. Tan. Data Mining
o
for Network Intrusion Detection. In Proc. NSF Workshop on Next Generation Data
Mining, Baltimore, MD, 2002.
[245] G. Dong and J. Li. Interestingness of discovered association rules in terms of
neighborhood-based unexpectedness. In Proc. of the 2nd Paciﬁc-Asia Conf. on Knowledge Discovery and Data Mining, pages 72–86, Melbourne, Australia, April 1998.
[246] G. Dong and J. Li. Eﬃcient Mining of Emerging Patterns: Discovering Trends and
Diﬀerences. In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining,
pages 43–52, San Diego, CA, August 1999.
[247] W. DuMouchel and D. Pregibon. Empirical Bayes Screening for Multi-Item Associations. In Proc. of the 7th Intl. Conf. on Knowledge Discovery and Data Mining, pages
67–76, San Francisco, CA, August 2001.
[248] B. Dunkel and N. Soparkar. Data Organization and Access for Eﬃcient Data Mining.
In Proc. of the 15th Intl. Conf. on Data Engineering, pages 522–529, Sydney, Australia,
March 1999.
[249] C. C. Fabris and A. A. Freitas. Discovering surprising patterns by detecting occurrences
of Simpson’s paradox. In Proc. of the 19th SGES Intl. Conf. on Knowledge-Based
Systems and Applied Artiﬁcial Intelligence), pages 148–160, Cambridge, UK, December
1999.
[250] L. Feng, H. J. Lu, J. X. Yu, and J. Han. Mining inter-transaction associations with
templates. In Proc. of the 8th Intl. Conf. on Information and Knowledge Management,
pages 225–233, Kansas City, Missouri, Nov 1999.
[251] A. A. Freitas. Understanding the crucial diﬀerences between classiﬁcation and discovery of association rules—a position paper. SIGKDD Explorations, 2(1):65–69, 2000.
[252] J. H. Friedman and N. I. Fisher. Bump hunting in high-dimensional data. Statistics
and Computing, 9(2):123–143, April 1999.
[253] T. Fukuda, Y. Morimoto, S. Morishita, and T. Tokuyama. Mining Optimized Association Rules for Numeric Attributes. In Proc. of the 15th Symp. on Principles of
Database Systems, pages 182–191, Montreal, Canada, June 1996.
[254] D. Gunopulos, R. Khardon, H. Mannila, and H. Toivonen. Data Mining, Hypergraph
Transversals, and Machine Learning. In Proc. of the 16th Symp. on Principles of
Database Systems, pages 209–216, Tucson, AZ, May 1997.
[255] E.-H. Han, G. Karypis, and V. Kumar. Min-Apriori: An Algorithm for Finding Association Rules in Data with Continuous Attributes. http://www.cs.umn.edu/˜han,
1997.
[256] E.-H. Han, G. Karypis, and V. Kumar. Scalable Parallel Data Mining for Association
Rules. In Proc. of 1997 ACM-SIGMOD Intl. Conf. on Management of Data, pages
277–288, Tucson, AZ, May 1997. Bibliography 399
[257] E.-H. Han, G. Karypis, V. Kumar, and B. Mobasher. Clustering Based on Association
Rule Hypergraphs. In Proc. of the 1997 ACM SIGMOD Workshop on Research Issues
in Data Mining and Knowledge Discovery, Tucson, AZ, 1997.
[258] J. Han, Y. Fu, K. Koperski, W. Wang, and O. R. Za¨
ıane. DMQL: A data mining query
language for relational databases. In Proc. of the 1996 ACM SIGMOD Workshop on
Research Issues in Data Mining and Knowledge Discovery, Montreal, Canada, June
1996.
[259] J. Han, J. Pei, and Y. Yin. Mining Frequent Patterns without Candidate Generation.
In Proc. ACM-SIGMOD Int. Conf. on Management of Data (SIGMOD’00), pages
1–12, Dallas, TX, May 2000.
[260] C. Hidber. Online Association Rule Mining. In Proc. of 1999 ACM-SIGMOD Intl.
Conf. on Management of Data, pages 145–156, Philadelphia, PA, 1999.
[261] R. J. Hilderman and H. J. Hamilton. Knowledge Discovery and Measures of Interest.
Kluwer Academic Publishers, 2001.
[262] J. Hipp, U. Guntzer, and G. Nakhaeizadeh. Algorithms for Association Rule Mining—
A General Survey. SigKDD Explorations, 2(1):58–64, June 2000.
[263] H. Hofmann, A. P. J. M. Siebes, and A. F. X. Wilhelm. Visualizing Association Rules
with Interactive Mosaic Plots. In Proc. of the 6th Intl. Conf. on Knowledge Discovery
and Data Mining, pages 227–235, Boston, MA, August 2000.
[264] J. D. Holt and S. M. Chung. Eﬃcient Mining of Association Rules in Text Databases.
In Proc. of the 8th Intl. Conf. on Information and Knowledge Management, pages
234–242, Kansas City, Missouri, 1999.
[265] M. Houtsma and A. Swami. Set-oriented Mining for Association Rules in Relational
Databases. In Proc. of the 11th Intl. Conf. on Data Engineering, pages 25–33, Taipei,
Taiwan, 1995.
[266] Y. Huang, S. Shekhar, and H. Xiong. Discovering Co-location Patterns from Spatial
Datasets: A General Approach. IEEE Trans. on Knowledge and Data Engineering, 16
(12):1472–1485, December 2004.
[267] T. Imielinski, A. Virmani, and A. Abdulghani. DataMine: Application Programming
Interface and Query Language for Database Mining. In Proc. of the 2nd Intl. Conf.
on Knowledge Discovery and Data Mining, pages 256–262, Portland, Oregon, 1996.
[268] A. Inokuchi, T. Washio, and H. Motoda. An Apriori-based Algorithm for Mining
Frequent Substructures from Graph Data. In Proc. of the 4th European Conf. of Principles and Practice of Knowledge Discovery in Databases, pages 13–23, Lyon, France,
2000.
[269] S. Jaroszewicz and D. Simovici. Interestingness of Frequent Itemsets Using Bayesian
Networks as Background Knowledge. In Proc. of the 10th Intl. Conf. on Knowledge
Discovery and Data Mining, pages 178–186, Seattle, WA, August 2004.
[270] M. Kamber and R. Shinghal. Evaluating the Interestingness of Characteristic Rules. In
Proc. of the 2nd Intl. Conf. on Knowledge Discovery and Data Mining, pages 263–266,
Portland, Oregon, 1996.
[271] M. Klemettinen. A Knowledge Discovery Methodology for Telecommunication Network
Alarm Databases. PhD thesis, University of Helsinki, 1999.
[272] W. A. Kosters, E. Marchiori, and A. Oerlemans. Mining Clusters with Association
Rules. In The 3rd Symp. on Intelligent Data Analysis (IDA99), pages 39–50, Amsterdam, August 1999.
[273] C. M. Kuok, A. Fu, and M. H. Wong. Mining Fuzzy Association Rules in Databases.
ACM SIGMOD Record, 27(1):41–46, March 1998. 400 Chapter 6 Association Analysis [274] M. Kuramochi and G. Karypis. Frequent Subgraph Discovery. In Proc. of the 2001
IEEE Intl. Conf. on Data Mining, pages 313–320, San Jose, CA, November 2001.
[275] W. Lee, S. J. Stolfo, and K. W. Mok. Adaptive Intrusion Detection: A Data Mining
Approach. Artiﬁcial Intelligence Review, 14(6):533–567, 2000.
[276] W. Li, J. Han, and J. Pei. CMAR: Accurate and Eﬃcient Classiﬁcation Based on
Multiple Class-association Rules. In Proc. of the 2001 IEEE Intl. Conf. on Data
Mining, pages 369–376, San Jose, CA, 2001.
[277] B. Liu, W. Hsu, and S. Chen. Using General Impressions to Analyze Discovered
Classiﬁcation Rules. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data
Mining, pages 31–36, Newport Beach, CA, August 1997.
[278] B. Liu, W. Hsu, and Y. Ma. Integrating Classiﬁcation and Association Rule Mining.
In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 80–86,
New York, NY, August 1998.
[279] B. Liu, W. Hsu, and Y. Ma. Mining association rules with multiple minimum supports.
In Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 125–
134, San Diego, CA, August 1999.
[280] B. Liu, W. Hsu, and Y. Ma. Pruning and Summarizing the Discovered Associations. In
Proc. of the 5th Intl. Conf. on Knowledge Discovery and Data Mining, pages 125–134,
San Diego, CA, August 1999.
[281] A. Marcus, J. I. Maletic, and K.-I. Lin. Ordinal association rules for error identiﬁcation in data sets. In Proc. of the 10th Intl. Conf. on Information and Knowledge
Management, pages 589–591, Atlanta, GA, October 2001.
[282] N. Megiddo and R. Srikant. Discovering Predictive Association Rules. In Proc. of the
4th Intl. Conf. on Knowledge Discovery and Data Mining, pages 274–278, New York,
August 1998.
[283] R. Meo, G. Psaila, and S. Ceri. A New SQL-like Operator for Mining Association
Rules. In Proc. of the 22nd VLDB Conf., pages 122–133, Bombay, India, 1996.
[284] R. J. Miller and Y. Yang. Association Rules over Interval Data. In Proc. of 1997
ACM-SIGMOD Intl. Conf. on Management of Data, pages 452–461, Tucson, AZ, May
1997.
[285] Y. Morimoto, T. Fukuda, H. Matsuzawa, T. Tokuyama, and K. Yoda. Algorithms for
mining association rules for binary segmentations of huge categorical databases. In
Proc. of the 24th VLDB Conf., pages 380–391, New York, August 1998.
[286] F. Mosteller. Association and Estimation in Contingency Tables. Journal of the American Statistical Association, 63:1–28, 1968.
[287] A. Mueller. Fast sequential and parallel algorithms for association rule mining: A
comparison. Technical Report CS-TR-3515, University of Maryland, August 1995.
[288] R. T. Ng, L. V. S. Lakshmanan, J. Han, and A. Pang. Exploratory Mining and Pruning
Optimizations of Constrained Association Rules. In Proc. of 1998 ACM-SIGMOD Intl.
Conf. on Management of Data, pages 13–24, Seattle, WA, June 1998.
[289] E. Omiecinski. Alternative Interest Measures for Mining Associations in Databases.
IEEE Trans. on Knowledge and Data Engineering, 15(1):57–69, January/February
2003.
[290] B. Ozden, S. Ramaswamy, and A. Silberschatz. Cyclic Association Rules. In Proc. of
the 14th Intl. Conf. on Data Eng., pages 412–421, Orlando, FL, February 1998.
[291] A. Ozgur, P. N. Tan, and V. Kumar. RBA: An Integrated Framework for Regression
based on Association Rules. In Proc. of the SIAM Intl. Conf. on Data Mining, pages
210–221, Orlando, FL, April 2004. Bibliography 401
[292] J. S. Park, M.-S. Chen, and P. S. Yu. An eﬀective hash-based algorithm for mining
association rules. SIGMOD Record, 25(2):175–186, 1995.
[293] S. Parthasarathy and M. Coatney. Eﬃcient Discovery of Common Substructures in
Macromolecules. In Proc. of the 2002 IEEE Intl. Conf. on Data Mining, pages 362–369,
Maebashi City, Japan, December 2002.
[294] N. Pasquier, Y. Bastide, R. Taouil, and L. Lakhal. Discovering frequent closed itemsets
for association rules. In Proc. of the 7th Intl. Conf. on Database Theory (ICDT’99),
pages 398–416, Jerusalem, Israel, January 1999.
[295] J. Pei, J. Han, H. J. Lu, S. Nishio, and S. Tang. H-Mine: Hyper-Structure Mining of
Frequent Patterns in Large Databases. In Proc. of the 2001 IEEE Intl. Conf. on Data
Mining, pages 441–448, San Jose, CA, November 2001.
[296] J. Pei, J. Han, B. Mortazavi-Asl, and H. Zhu. Mining Access Patterns Eﬃciently from
Web Logs. In Proc. of the 4th Paciﬁc-Asia Conf. on Knowledge Discovery and Data
Mining, pages 396–407, Kyoto, Japan, April 2000.
[297] G. Piatetsky-Shapiro. Discovery, Analysis and Presentation of Strong Rules. In
G. Piatetsky-Shapiro and W. Frawley, editors, Knowledge Discovery in Databases,
pages 229–248. MIT Press, Cambridge, MA, 1991.
[298] C. Potter, S. Klooster, M. Steinbach, P. N. Tan, V. Kumar, S. Shekhar, and C. Carvalho. Understanding Global Teleconnections of Climate to Regional Model Estimates
of Amazon Ecosystem Carbon Fluxes. Global Change Biology, 10(5):693–703, 2004.
[299] C. Potter, S. Klooster, M. Steinbach, P. N. Tan, V. Kumar, S. Shekhar, R. Myneni,
and R. Nemani. Global Teleconnections of Ocean Climate to Terrestrial Carbon Flux.
J. Geophysical Research, 108(D17), 2003.
[300] G. D. Ramkumar, S. Ranka, and S. Tsur. Weighted Association Rules: Model and
Algorithm. http://www.cs.ucla.edu/˜czdemo/tsur/, 1997.
[301] S. Sarawagi, S. Thomas, and R. Agrawal. Integrating Mining with Relational Database
Systems: Alternatives and Implications. In Proc. of 1998 ACM-SIGMOD Intl. Conf.
on Management of Data, pages 343–354, Seattle, WA, 1998.
[302] K. Satou, G. Shibayama, T. Ono, Y. Yamamura, E. Furuichi, S. Kuhara, and T. Takagi.
Finding Association Rules on Heterogeneous Genome Data. In Proc. of the Paciﬁc
Symp. on Biocomputing, pages 397–408, Hawaii, January 1997.
[303] A. Savasere, E. Omiecinski, and S. Navathe. An eﬃcient algorithm for mining association rules in large databases. In Proc. of the 21st Int. Conf. on Very Large Databases
(VLDB‘95), pages 432–444, Zurich, Switzerland, September 1995.
[304] A. Savasere, E. Omiecinski, and S. Navathe. Mining for Strong Negative Associations
in a Large Database of Customer Transactions. In Proc. of the 14th Intl. Conf. on
Data Engineering, pages 494–502, Orlando, Florida, February 1998.
[305] M. Seno and G. Karypis. LPMiner: An Algorithm for Finding Frequent Itemsets Using
Length-Decreasing Support Constraint. In Proc. of the 2001 IEEE Intl. Conf. on Data
Mining, pages 505–512, San Jose, CA, November 2001.
[306] T. Shintani and M. Kitsuregawa. Hash based parallel algorithms for mining association
rules. In Proc of the 4th Intl. Conf. on Parallel and Distributed Info. Systems, pages
19–30, Miami Beach, FL, December 1996.
[307] A. Silberschatz and A. Tuzhilin. What makes patterns interesting in knowledge discovery systems. IEEE Trans. on Knowledge and Data Engineering, 8(6):970–974, 1996.
[308] C. Silverstein, S. Brin, and R. Motwani. Beyond market baskets: Generalizing association rules to dependence rules. Data Mining and Knowledge Discovery, 2(1):39–68,
1998. 402 Chapter 6 Association Analysis [309] E.-H. Simpson. The Interpretation of Interaction in Contingency Tables. Journal of
the Royal Statistical Society, B(13):238–241, 1951.
[310] L. Singh, B. Chen, R. Haight, and P. Scheuermann. An Algorithm for Constrained
Association Rule Mining in Semi-structured Data. In Proc. of the 3rd Paciﬁc-Asia
Conf. on Knowledge Discovery and Data Mining, pages 148–158, Beijing, China, April
1999.
[311] R. Srikant and R. Agrawal. Mining Quantitative Association Rules in Large Relational
Tables. In Proc. of 1996 ACM-SIGMOD Intl. Conf. on Management of Data, pages
1–12, Montreal, Canada, 1996.
[312] R. Srikant and R. Agrawal. Mining Sequential Patterns: Generalizations and Performance Improvements. In Proc. of the 5th Intl Conf. on Extending Database Technology
(EDBT’96), pages 18–32, Avignon, France, 1996.
[313] R. Srikant, Q. Vu, and R. Agrawal. Mining Association Rules with Item Constraints.
In Proc. of the 3rd Intl. Conf. on Knowledge Discovery and Data Mining, pages 67–73,
Newport Beach, CA, August 1997.
[314] M. Steinbach, P. N. Tan, and V. Kumar. Support Envelopes: A Technique for Exploring the Structure of Association Patterns. In Proc. of the 10th Intl. Conf. on
Knowledge Discovery and Data Mining, pages 296–305, Seattle, WA, August 2004.
[315] M. Steinbach, P. N. Tan, H. Xiong, and V. Kumar. Extending the Notion of Support.
In Proc. of the 10th Intl. Conf. on Knowledge Discovery and Data Mining, pages 689–
694, Seattle, WA, August 2004.
[316] E. Suzuki. Autonomous Discovery of Reliable Exception Rules. In Proc. of the 3rd
Intl. Conf. on Knowledge Discovery and Data Mining, pages 259–262, Newport Beach,
CA, August 1997.
[317] P. N. Tan and V. Kumar. Mining Association Patterns in Web Usage Data. In Proc.
of the Intl. Conf. on Advances in Infrastructure for e-Business, e-Education, e-Science
and e-Medicine on the Internet, L’Aquila, Italy, January 2002.
[318] P. N. Tan, V. Kumar, and J. Srivastava. Selecting the Right Interestingness Measure
for Association Patterns. In Proc. of the 8th Intl. Conf. on Knowledge Discovery and
Data Mining, pages 32–41, Edmonton, Canada, July 2002.
[319] P. N. Tan, M. Steinbach, V. Kumar, S. Klooster, C. Potter, and A. Torregrosa. Finding
Spatio-Temporal Patterns in Earth Science Data. In KDD 2001 Workshop on Temporal
Data Mining, San Francisco, CA, 2001.
[320] H. Toivonen. Sampling Large Databases for Association Rules. In Proc. of the 22nd
VLDB Conf., pages 134–145, Bombay, India, 1996.
[321] H. Toivonen, M. Klemettinen, P. Ronkainen, K. Hatonen, and H. Mannila. Pruning
and Grouping Discovered Association Rules. In ECML-95 Workshop on Statistics,
Machine Learning and Knowledge Discovery in Databases, pages 47 – 52, Heraklion,
Greece, April 1995.
[322] S. Tsur, J. Ullman, S. Abiteboul, C. Clifton, R. Motwani, S. Nestorov, and A. Rosenthal. Query Flocks: A Generalization of Association Rule Mining. In Proc. of 1998
ACM-SIGMOD Intl. Conf. on Management of Data, pages 1–12, Seattle, WA, June
1998.
[323] A. Tung, H. J. Lu, J. Han, and L. Feng. Breaking the Barrier of Transactions: Mining
Inter-Transaction Association Rules. In Proc. of the 5th Intl. Conf. on Knowledge
Discovery and Data Mining, pages 297–301, San Diego, CA, August 1999.
[324] K. Wang, Y. He, and J. Han. Mining Frequent Itemsets Using Support Constraints.
In Proc. of the 26th VLDB Conf., pages 43–52, Cairo, Egypt, September 2000. BIBLIOGRAPHY 403
[325] K. Wang, S. H. Tay, and B. Liu. Interestingness-Based Interval Merger for Numeric
Association Rules. In Proc. of the 4th Intl. Conf. on Knowledge Discovery and Data
Mining, pages 121–128, New York, NY, August 1998.
[326] G. I. Webb. Preliminary investigations into statistically valid exploratory rule discovery. In Proc. of the Australasian Data Mining Workshop (AusDM03), Canberra,
Australia, December 2003.
[327] H. Xiong, X. He, C. Ding, Y. Zhang, V. Kumar, and S. R. Holbrook. Identiﬁcation
of Functional Modules in Protein Complexes via Hyperclique Pattern Discovery. In
Proc. of the Paciﬁc Symposium on Biocomputing, (PSB 2005), Maui, January 2005.
[328] H. Xiong, S. Shekhar, P. N. Tan, and V. Kumar. Exploiting a Support-based Upper
Bound of Pearson’s Correlation Coeﬃcient for Eﬃciently Identifying Strongly Correlated Pairs. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and Data Mining,
pages 334–343, Seattle, WA, August 2004.
[329] H. Xiong, M. Steinbach, P. N. Tan, and V. Kumar. HICAP: Hierarchial Clustering
with Pattern Preservation. In Proc. of the SIAM Intl. Conf. on Data Mining, pages
279–290, Orlando, FL, April 2004.
[330] H. Xiong, P. N. Tan, and V. Kumar. Mining Strong Aﬃnity Association Patterns in
Data Sets with Skewed Support Distribution. In Proc. of the 2003 IEEE Intl. Conf.
on Data Mining, pages 387–394, Melbourne, FL, 2003.
[331] X. Yan and J. Han. gSpan: Graph-based Substructure Pattern Mining. In Proc. of
the 2002 IEEE Intl. Conf. on Data Mining, pages 721–724, Maebashi City, Japan,
December 2002.
[332] C. Yang, U. M. Fayyad, and P. S. Bradley. Eﬃcient discovery of error-tolerant frequent
itemsets in high dimensions. In Proc. of the 7th Intl. Conf. on Knowledge Discovery
and Data Mining, pages 194–203, San Francisco, CA, August 2001.
[333] M. J. Zaki. Parallel and Distributed Association Mining: A Survey. IEEE Concurrency,
special issue on Parallel Mechanisms for Data Mining, 7(4):14–25, December 1999.
[334] M. J. Zaki. Generating Non-Redundant Association Rules. In Proc. of the 6th Intl.
Conf. on Knowledge Discovery and Data Mining, pages 34–43, Boston, MA, August
2000.
[335] M. J. Zaki. Eﬃciently mining frequent trees in a forest. In Proc. of the 8th Intl.
Conf. on Knowledge Discovery and Data Mining, pages 71–80, Edmonton, Canada,
July 2002.
[336] M. J. Zaki and M. Orihara. Theoretical foundations of association rules. In Proc. of
the 1998 ACM SIGMOD Workshop on Research Issues in Data Mining and Knowledge
Discovery, Seattle, WA, June 1998.
[337] M. J. Zaki, S. Parthasarathy, M. Ogihara, and W. Li. New Algorithms for Fast
Discovery of Association Rules. In Proc. of the 3rd Intl. Conf. on Knowledge Discovery
and Data Mining, pages 283–286, Newport Beach, CA, August 1997.
[338] H. Zhang, B. Padmanabhan, and A. Tuzhilin. On the Discovery of Signiﬁcant Statistical Quantitative Rules. In Proc. of the 10th Intl. Conf. on Knowledge Discovery and
Data Mining, pages 374–383, Seattle, WA, August 2004.
[339] Z. Zhang, Y. Lu, and B. Zhang. An Eﬀective Partioning-Combining Algorithm for
Discovering Quantitative Association Rules. In Proc. of the 1st Paciﬁc-Asia Conf. on
Knowledge Discovery and Data Mining, Singapore, 1997.
[340] N. Zhong, Y. Y. Yao, and S. Ohsuga. Peculiarity Oriented Multi-database Mining. In
Proc. of the 3rd European Conf. of Principles and Practice of Knowledge Discovery in
Databases, pages 136–146, Prague, Czech Republic, 1999. 404 Chapter 6 6.10 Association Analysis Exercises 1. For each of the following questions, provide an example of an association rule
from the market basket domain that satisﬁes the following conditions. Also,
describe whether such rules are subjectively interesting.
(a) A rule that has high support and high conﬁdence.
(b) A rule that has reasonably high support but low conﬁdence.
(c) A rule that has low support and low conﬁdence.
(d) A rule that has low support and high conﬁdence.
2. Consider the data set shown in Table 6.22.
Table 6.22. Example of market basket transactions.
Customer ID
1
1
2
2
3
3
4
4
5
5 Transaction ID
0001
0024
0012
0031
0015
0022
0029
0040
0033
0038 Items Bought
{a, d, e}
{a, b, c, e}
{a, b, d, e}
{a, c, d, e}
{b, c, e}
{b, d, e}
{c, d}
{a, b, c}
{a, d, e}
{a, b, e} (a) Compute the support for itemsets {e}, {b, d}, and {b, d, e} by treating
each transaction ID as a market basket.
(b) Use the results in part (a) to compute the conﬁdence for the association rules {b, d} −→ {e} and {e} −→ {b, d}. Is conﬁdence a symmetric
measure?
(c) Repeat part (a) by treating each customer ID as a market basket. Each
item should be treated as a binary variable (1 if an item appears in at
least one transaction bought by the customer, and 0 otherwise.)
(d) Use the results in part (c) to compute the conﬁdence for the association
rules {b, d} −→ {e} and {e} −→ {b, d}.
(e) Suppose s1 and c1 are the support and conﬁdence values of an association
rule r when treating each transaction ID as a market basket. Also, let s2
and c2 be the support and conﬁdence values of r when treating each customer ID as a market basket. Discuss whether there are any relationships
between s1 and s2 or c1 and c2 . 6.10
3. Exercises 405 (a) What is the conﬁdence for the rules ∅ −→ A and A −→ ∅?
(b) Let c1 , c2 , and c3 be the conﬁdence values of the rules {p} −→ {q },
{p} −→ {q, r}, and {p, r} −→ {q }, respectively. If we assume that c1 , c2 ,
and c3 have diﬀerent values, what are the possible relationships that may
exist among c1 , c2 , and c3 ? Which rule has the lowest conﬁdence?
(c) Repeat the analysis in part (b) assuming that the rules have identical
support. Which rule has the highest conﬁdence?
(d) Transitivity: Suppose the conﬁdence of the rules A −→ B and B −→ C
are larger than some threshold, minconf . Is it possible that A −→ C has
a conﬁdence less than minconf ? 4. For each of the following measures, determine whether it is monotone, antimonotone, or non-monotone (i.e., neither monotone nor anti-monotone).
Example: Support, s =
s(Y ) whenever X ⊂ Y . σ (X )
|T | is anti-monotone because s(X ) ≥ (a) A characteristic rule is a rule of the form {p} −→ {q1 , q2 , . . . , qn }, where
the rule antecedent contains only a single item. An itemset of size k can
produce up to k characteristic rules. Let ζ be the minimum conﬁdence of
all characteristic rules generated from a given itemset:
ζ ({p1 , p2 , . . . , pk }) = min c {p1 } −→ {p2 , p3 , . . . , pk } , . . .
c {pk } −→ {p1 , p3 . . . , pk−1 } Is ζ monotone, anti-monotone, or non-monotone?
(b) A discriminant rule is a rule of the form {p1 , p2 , . . . , pn } −→ {q }, where
the rule consequent contains only a single item. An itemset of size k can
produce up to k discriminant rules. Let η be the minimum conﬁdence of
all discriminant rules generated from a given itemset:
η ({p1 , p2 , . . . , pk }) = min c {p2 , p3 , . . . , pk } −→ {p1 } , . . .
c {p1 , p2 , . . . pk−1 } −→ {pk } Is η monotone, anti-monotone, or non-monotone?
(c) Repeat the analysis in parts (a) and (b) by replacing the min function
with a max function.
5. Prove Equation 6.3. (Hint: First, count the number of ways to create an itemset
that forms the left hand side of the rule. Next, for each size k itemset selected
for the left-hand side, count the number of ways to choose the remaining d − k
items to form the right-hand side of the rule.) 406 Chapter 6 Association Analysis Table 6.23. Market basket transactions.
Transaction ID Items Bought
1
{Milk, Beer, Diapers}
2
{Bread, Butter, Milk}
3
{Milk, Diapers, Cookies}
4
{Bread, Butter, Cookies}
5
{Beer, Cookies, Diapers}
6
{Milk, Diapers, Bread, Butter}
7
{Bread, Butter, Diapers}
8
{Beer, Diapers}
9
{Milk, Diapers, Bread, Butter}
10
{Beer, Cookies}
6. Consider the market basket transactions shown in Table 6.23.
(a) What is the maximum number of association rules that can be extracted
from this data (including rules that have zero support)?
(b) What is the maximum size of frequent itemsets that can be extracted
(assuming minsup > 0)?
(c) Write an expression for the maximum number of size-3 itemsets that can
be derived from this data set.
(d) Find an itemset (of size 2 or larger) that has the largest support.
(e) Find a pair of items, a and b, such that the rules {a} −→ {b} and {b} −→
{a} have the same conﬁdence.
7. Consider the following set of frequent 3-itemsets:
{1, 2, 3}, {1, 2, 4}, {1, 2, 5}, {1, 3, 4}, {1, 3, 5}, {2, 3, 4}, {2, 3, 5}, {3, 4, 5}.
Assume that there are only ﬁve items in the data set.
(a) List all candidate 4-itemsets obtained by a candidate generation procedure
using the Fk−1 × F1 merging strategy.
(b) List all candidate 4-itemsets obtained by the candidate generation procedure in Apriori.
(c) List all candidate 4-itemsets that survive the candidate pruning step of
the Apriori algorithm.
8. The Apriori algorithm uses a generate-and-count strategy for deriving frequent
itemsets. Candidate itemsets of size k + 1 are created by joining a pair of
frequent itemsets of size k (this is known as the candidate generation step). A
candidate is discarded if any one of its subsets is found to be infrequent during
the candidate pruning step. Suppose the Apriori algorithm is applied to the 6.10 Exercises 407 Table 6.24. Example of market basket transactions.
Transaction ID
1
2
3
4
5
6
7
8
9
10 Items Bought
{a, b, d, e}
{b, c, d}
{a, b, d, e}
{a, c, d, e}
{b, c, d, e}
{b, d, e}
{c, d}
{a, b, c}
{a, d, e}
{b, d} data set shown in Table 6.24 with minsup = 30%, i.e., any itemset occurring
in less than 3 transactions is considered to be infrequent.
(a) Draw an itemset lattice representing the data set given in Table 6.24.
Label each node in the lattice with the following letter(s):
• N: If the itemset is not considered to be a candidate itemset by
the Apriori algorithm. There are two reasons for an itemset not to
be considered as a candidate itemset: (1) it is not generated at all
during the candidate generation step, or (2) it is generated during
the candidate generation step but is subsequently removed during
the candidate pruning step because one of its subsets is found to be
infrequent.
• F: If the candidate itemset is found to be frequent by the Apriori
algorithm.
• I: If the candidate itemset is found to be infrequent after support
counting.
(b) What is the percentage of frequent itemsets (with respect to all itemsets
in the lattice)?
(c) What is the pruning ratio of the Apriori algorithm on this data set?
(Pruning ratio is deﬁned as the percentage of itemsets not considered
to be a candidate because (1) they are not generated during candidate
generation or (2) they are pruned during the candidate pruning step.)
(d) What is the false alarm rate (i.e, percentage of candidate itemsets that
are found to be infrequent after performing support counting)?
9. The Apriori algorithm uses a hash tree data structure to eﬃciently count the
support of candidate itemsets. Consider the hash tree for candidate 3-itemsets
shown in Figure 6.32. 408 Chapter 6 Association Analysis
1,4,7 1,4,7
L1
{145}
{178} 2,5,8 L3
{125}
{158}
{458} 1,4,7 3,6,9 3,6,9
2,5,8 L5
{168} 1,4,7 2,5,8
L2
{127}
{457} 3,6,9 2,5,8 3,6,9
L4
{459}
{456}
{789} L6
{246}
{278} L7
{258}
{289} L8
{568} 1,4,7
3,6,9
2,5,8
L9
L11
L12
{346}
{356}
{367}
{379}
{689}
{678} Figure 6.32. An example of a hash tree structure. (a) Given a transaction that contains items {1, 3, 4, 5, 8}, which of the hash
tree leaf nodes will be visited when ﬁnding the candidates of the transaction?
(b) Use the visited leaf nodes in part (b) to determine the candidate itemsets
that are contained in the transaction {1, 3, 4, 5, 8}.
10. Consider the following set of candidate 3-itemsets:
{1, 2, 3}, {1, 2, 6}, {1, 3, 4}, {2, 3, 4}, {2, 4, 5}, {3, 4, 6}, {4, 5, 6}
(a) Construct a hash tree for the above candidate 3-itemsets. Assume the
tree uses a hash function where all odd-numbered items are hashed to
the left child of a node, while the even-numbered items are hashed to the
right child. A candidate k -itemset is inserted into the tree by hashing on
each successive item in the candidate and then following the appropriate
branch of the tree according to the hash value. Once a leaf node is reached,
the candidate is inserted based on one of the following conditions:
Condition 1: If the depth of the leaf node is equal to k (the root is
assumed to be at depth 0), then the candidate is inserted regardless
of the number of itemsets already stored at the node.
Condition 2: If the depth of the leaf node is less than k , then the candidate can be inserted as long as the number of itemsets stored at the
node is less than maxsize. Assume maxsize = 2 for this question.
Condition 3: If the depth of the leaf node is less than k and the number
of itemsets stored at the node is equal to maxsize, then the leaf
node is converted into an internal node. New leaf nodes are created
as children of the old leaf node. Candidate itemsets previously stored 6.10 Exercises 409 null a b c d e ab ac ad ae bc bd be cd ce de abc abd abe acd ace ade bcd bce bde cde abcd abce abde acde bcde abcde Figure 6.33. An itemset lattice in the old leaf node are distributed to the children based on their hash
values. The new candidate is also hashed to its appropriate leaf node.
(b) How many leaf nodes are there in the candidate hash tree? How many
internal nodes are there?
(c) Consider a transaction that contains the following items: {1, 2, 3, 5, 6}.
Using the hash tree constructed in part (a), which leaf nodes will be
checked against the transaction? What are the candidate 3-itemsets contained in the transaction?
11. Given the lattice structure shown in Figure 6.33 and the transactions given in
Table 6.24, label each node with the following letter(s):
• M if the node is a maximal frequent itemset,
• C if it is a closed frequent itemset,
• N if it is frequent but neither maximal nor closed, and
• I if it is infrequent.
Assume that the support threshold is equal to 30%.
12. The original association rule mining formulation uses the support and conﬁdence measures to prune uninteresting rules. 410 Chapter 6 Association Analysis (a) Draw a contingency table for each of the following rules using the transactions shown in Table 6.25.
Table 6.25. Example of market basket transactions.
Transaction ID
1
2
3
4
5
6
7
8
9
10 Items Bought
{a, b, d, e}
{b, c, d}
{a, b, d, e}
{a, c, d, e}
{b, c, d, e}
{b, d, e}
{c, d}
{a, b, c}
{a, d, e}
{b, d} Rules: {b} −→ {c}, {a} −→ {d}, {b} −→ {d}, {e} −→ {c}, {c} −→ {a}.
(b) Use the contingency tables in part (a) to compute and rank the rules in
decreasing order according to the following measures.
i. Support.
ii. Conﬁdence.
iii. Interest(X −→ Y ) =
iv. IS(X −→ Y ) = √ P (X,Y )
P (X ) P (Y
P (X,Y ) P (X )P (Y ) v. Klosgen(X −→ Y ) =
P (X,Y )
P (X ) . ). . P (X, Y )×(P (Y |X )−P (Y )), where P (Y |X ) = vi. Odds ratio(X −→ Y ) = P (X,Y )P (X,Y )
.
P (X,Y )P (X,Y ) 13. Given the rankings you had obtained in Exercise 12, compute the correlation
between the rankings of conﬁdence and the other ﬁve measures. Which measure
is most highly correlated with conﬁdence? Which measure is least correlated
with conﬁdence?
14. Answer the following questions using the data sets shown in Figure 6.34. Note
that each data set contains 1000 items and 10,000 transactions. Dark cells
indicate the presence of items and white cells indicate the absence of items. We
will apply the Apriori algorithm to extract frequent itemsets with minsup =
10% (i.e., itemsets must be contained in at least 1000 transactions)?
(a) Which data set(s) will produce the most number of frequent itemsets? 6.10 Exercises 411 (b) Which data set(s) will produce the fewest number of frequent itemsets?
(c) Which data set(s) will produce the longest frequent itemset?
(d) Which data set(s) will produce frequent itemsets with highest maximum
support?
(e) Which data set(s) will produce frequent itemsets containing items with
wide-varying support levels (i.e., items with mixed support, ranging from
less than 20% to more than 70%).
15. (a) Prove that the φ coeﬃcient is equal to 1 if and only if f11 = f1+ = f+1 .
(b) Show that if A and B are independent, then P (A, B )×P (A, B ) = P (A, B )×
P (A, B ).
(c) Show that Yule’s Q and Y coeﬃcients
Q=
Y = f11 f00 − f10 f01
f11 f00 + f10 f01
√
√
f11 f00 − f10 f01
√
√
f11 f00 + f10 f01 are normalized versions of the odds ratio.
(d) Write a simpliﬁed expression for the value of each measure shown in Tables
6.11 and 6.12 when the variables are statistically independent.
16. Consider the interestingness measure, M =
rule A −→ B . P (B |A)−P (B )
,
1−P (B ) for an association (a) What is the range of this measure? When does the measure attain its
maximum and minimum values?
(b) How does M behave when P (A, B ) is increased while P (A) and P (B )
remain unchanged?
(c) How does M behave when P (A) is increased while P (A, B ) and P (B )
remain unchanged?
(d) How does M behave when P (B ) is increased while P (A, B ) and P (A)
remain unchanged?
(e) Is the measure symmetric under variable permutation?
(f) What is the value of the measure when A and B are statistically independent?
(g) Is the measure null-invariant?
(h) Does the measure remain invariant under row or column scaling operations?
(i) How does the measure behave under the inversion operation? 412 Chapter 6 Association Analysis Items Items 4000
6000 2000
Transactions Transactions 2000 4000
6000 8000 8000 200 400 600 800
(a) 200 400 600 800
(b) Items Items 2000 4000
6000 Transactions Transactions 2000 4000
6000
8000 8000 200 400 600 800
(c) 200 400 600 800
(d) Items Items 2000 4000
6000 Transactions Transactions 2000
10% are 1s
90% are 0s
(uniformly distributed) 6000
8000 8000 200 400 600 800
(e) 4000 200 400 600 800
(f) Figure 6.34. Figures for Exercise 14. 6.10 Exercises 413 17. Suppose we have market basket data consisting of 100 transactions and 20
items. If the support for item a is 25%, the support for item b is 90% and the
support for itemset {a, b} is 20%. Let the support and conﬁdence thresholds
be 10% and 60%, respectively.
(a) Compute the conﬁdence of the association rule {a} → {b}. Is the rule
interesting according to the conﬁdence measure?
(b) Compute the interest measure for the association pattern {a, b}. Describe
the nature of the relationship between item a and item b in terms of the
interest measure.
(c) What conclusions can you draw from the results of parts (a) and (b)?
(d) Prove that if the conﬁdence of the rule {a} −→ {b} is less than the support
of {b}, then:
i. c({a} −→ {b}) > c({a} −→ {b}),
ii. c({a} −→ {b}) > s({b}),
where c(·) denote the rule conﬁdence and s(·) denote the support of an
itemset.
18. Table 6.26 shows a 2 × 2 × 2 contingency table for the binary variables A and
B at diﬀerent values of the control variable C .
Table 6.26. A Contingency Table.
A
1
1 C=1 15 0 15 30 5 0 0 B 0 1 C=0 0 0 15 B (a) Compute the φ coeﬃcient for A and B when C = 0, C = 1, and C = 0 or
.
1. Note that φ({A, B }) = √ P (A,B )−P (A)P (B )
P (A)P (B )(1−P (A))(1−P (B )) (b) What conclusions can you draw from the above result?
19. Consider the contingency tables shown in Table 6.27.
(a) For table I, compute support, the interest measure, and the φ correlation coeﬃcient for the association pattern {A, B}. Also, compute the
conﬁdence of rules A → B and B → A. 414 Chapter 6 Association Analysis
Table 6.27. Contingency tables for Exercise 19.
B B A 9 1 A 1 89 (a) Table I. B B A 89 1 A 1 9 (b) Table II. (b) For table II, compute support, the interest measure, and the φ correlation coeﬃcient for the association pattern {A, B}. Also, compute the
conﬁdence of rules A → B and B → A.
(c) What conclusions can you draw from the results of (a) and (b)?
20. Consider the relationship between customers who buy high-deﬁnition televisions
and exercise machines as shown in Tables 6.19 and 6.20.
(a) Compute the odds ratios for both tables.
(b) Compute the φ-coeﬃcient for both tables.
(c) Compute the interest factor for both tables.
For each of the measures given above, describe how the direction of association
changes when data is pooled together instead of being stratiﬁed. ...

View
Full
Document