This preview shows page 1. Sign up to view the full content.
Unformatted text preview: both the
structure and the parameters of graphic models can be learned directly from databases
(Buntine 1996; Heckerman 1996). Modelevaluation criteria are typically Bayesian in form,
and parameter estimation can be a mixture of
closedform estimates and iterative methods
depending on whether a variable is directly
observed or hidden. Model search can consist
of greedy hillclimbing methods over various
graph structures. Prior knowledge, such as a
partial ordering of the variables based on
causal relations, can be useful in terms of reducing the model search space. Although still
primarily in the research phase, graphic model
induction methods are of particular interest to
KDD because the graphic form of the model
lends itself easily to human interpretation. Relational Learning Models
Although decision trees and rules have a representation restricted to propositional logic, relational learning (also known as inductive logic
programming) uses the more ﬂexible pattern
language of ﬁrstorder logic. A relational learner can easily ﬁnd formulas such as X = Y. Most
research to date on modelevaluation methods
for relational learning is logical in nature. The
extra representational power of relational
models comes at the price of signiﬁcant computational demands in terms of search. See
Dzeroski (1996) for a more detailed discussion. Discussion
Given the broad spectrum of datamining
methods and algorithms, our overview is in 48 AI MAGAZINE evitably limited in scope; many datamining
techniques, particularly specialized methods
for particular types of data and domains, were
not mentioned speciﬁcally. We believe the
general discussion on datamining tasks and
components has general relevance to a variety of methods. For example, consider timeseries prediction, which traditionally has
been cast as a predictive regression task (autoregressive models, and so on). Recently,
more general models have been developed for
timeseries applications, such as nonlinear basis functions, examplebased models, and kernel methods. Furthermore, there has been
signiﬁcant interest in descriptive graphic and
local data modeling of time series rather than
purely predictive modeling (Weigend and
Gershenfeld 1993). Thus, although different
algorithms and applications might appear different on the surface, it is not uncommon to
ﬁnd that they share many common components. Understanding data mining and model
induction at this component level clariﬁes
the behavior of any datamining algorithm
and makes it easier for the user to understand
its overall contribution and applicability to
the KDD process.
An important point is that each technique
typically suits some problems better than
others. For example, decision tree classiﬁers
can be useful for ﬁnding structure in highdimensional spaces and in problems with
mixed continuous and categorical data (because tree methods do not require distance
metrics). However, classiﬁcation trees might
not be suitable for problems where the true
decision boundaries between classes are described by a secondorder polynomial (for example). Thus, there is no univ...
View
Full
Document
This document was uploaded on 02/15/2014.
 Spring '14

Click to edit the document details