Unformatted text preview: ised learning over unrated book descriptions to improve supervised learning from
a smaller number of rated examples. A successful method
for doing this in text categorization is presented in 24 . Another approach is active learning, in which examples are ac2 Amazon has already made signi cantly more income from the
rst author based on recommendations provided by Libra than those
provided by its own recommender system; however, this is hardly a
rigorous, unbiased comparison. quired incrementally and the system attempts to use what
it has already learned to limit training by selecting only
the most informative new examples for the user to rate 9 .
Speci c techniques for applying this to text categorization
have been developed and shown to signi cantly reduce the
quantity of labeled examples required 18, 19 .
A slightly di erent approach is to advise users on easy
and productive strategies for selecting good training examples themselves. We have found that one e ective approach
is to rst provide a small number of highly rated examples
which are presumably easy for users to generate, running
the system to generate initial recommendations, reviewing
the top recommendations for obviously bad items, providing
low ratings for these examples, and retraining the system to
obtain new recommendations. We intend to conduct experiments on the existing data sets evaluating such strategies
for selecting training examples.
Studying additional ways of combining content-based and
collaborative recommending is particularly important. The
use of collaborative content in Libra was found to be useful, and if signi cant data bases of both user ratings and
item content are available, both of these sources of information could contribute to better recommendations 3, 4 .
One additional approach is to automatically add the related
books of each rated book as additional training examples
with the same or similar rating, thereby using collaborative information to expand the training examples available
for content-based recommending.
A list of additional topics for investigation include the
Allowing a user to initially provide keywords that are
of known interest or disinterest, and incorporating
this information into learned pro les by biasing the
parameter estimates for these words 25 .
Comparing di erent text-categorization algorithms: In
addition to more sophisticated Bayesian methods, neuralnetwork and case-based methods could be explored.
Combining content extracted from multiple sources:
For example, combining information about a title from
Amazon, BarnesAndNoble, on-line library catalogs, etc.
Using full-text as content: Utilizing complete on-line
text for a book, instead of abstracted summaries and
reviews, as the content description. 5 CONCLUSIONS
Unlike collaborative ltering, content-based recommending
holds the promise of being able to e ectively recommend unrated items and to provide quality recommendations to users
with unique, individual tastes. Libra...
View Full Document
- Fall '13
- Machine Learning, Spearman's rank correlation coefficient, Recommender system, training examples