P1: c-143Braynov-2
Braynov2 WL040/Bidgoli-Vol III-Ch-05 July 11, 2003 11:43 Char Count= 0
58 PERSONALIZATION ANDCUSTOMIZATIONTECHNOLOGIESDirect approaches to preference elicitation, however,
are generally not feasible, due to the exponential number
of comparisons and due to the complexity of the ques-
tions to be asked. In applications such as recommender
systems, product configuration, and adaptive Web stores,
users cannot be expected to have the patience to provide
detailed preference relations or utility functions. In ad-
dition, users cannot always be expected to provide accu-
rate and complete information about their preferences. In
general, direct preference elicitation requires a significant
level of introspection and imposes a significant cognitive
overload on users. Instead of using direct preference elic-
itation, Chajewska, Getoor, Norman, and Shahar (1998)
have used classification to identify a user’s utility function.
The authors partition users’ utility functions into clusters
with very similar utility functions in each cluster. Every
cluster is characterized by a single reference utility func-
tion called prototype. Then for every new user the system
finds the cluster to which he is most likely to belong, and
uses the prototype utility function for that cluster to pre-
dict his preferences.Reasoning with Conditional Preferences
Most of the research on preference modeling focuses on
additive utility functions in which every attribute of the
function is independent of other attributes. Such utility
functions allow a very elegant representation in the form
of a weighted sum of attribute utilities. For example, if a
user’s preferences for product qualityQare independent
of his preferences for product brandB, then the user’s
utility could be represented as a weighted sum,U(Q,B)=w 1 U(Q)+w 2 U(B)whereU(Q,B) is the overall product utility,U(Q) is the util-
ity for product quality, andU(B) is the utility for brand.
Additive utility functions are easy to elicit and evaluate.
A problem arises, however, if preferences over product
attributes are not independent. For example, a prefer-
ence for brand usually depends on the quality level and
vice versa. To overcome this problem, Boutilier, Brafman,
Hoos, and Poole (1999) proposed a model of conditional
preferences. Preferences are represented as a network
called a conditional preference network (CP-network)
that specifies the dependence between attributes. In CP-
networks a user can specify a conditional preference in
the form:If product quality is medium, then I prefer Brand 1 to
Brand 2.Boutilier et al. (1999) have also formulated and stud-
ied the product configuration problem, i.e, what is the bestproduct configuration that satisfies a set of customer pref-
erences. The authors have proposed several algorithms for
finding an optimal product configuration for a given set
of constraints. In Domshlak, Brafman, & Shimony (2001)
CP-networks are used for Web-page personalization. The
optimal presentation of a Web page is determined by tak-
ing into account the preferences of the Web designer, the
layout constraints, and the viewer interaction with the
browser. For example, the preferences of the Web page de-
signer are represented by a CP-network and constrained
optimization techniques are used to determine the opti-
mal Web page configuration.Preference-Based Queries
Preference modeling plays an increasingly important role
in e-commerce applications where the users face the prob-
lem of information overload, i.e., how to choose from
among thousands and even millions of products and prod-
uct descriptions. Preference queries (Chomicki, 2002)
have been proposed as a tool to help a user formulate the
most appropriate queries and receive the results ranked
according to his preferences. Preference queries are based
on a preference operator (called winnow) which picks
from a given relation the set of the most preferred tuples,
according to a given preference formula.
For example, consider the instance of the book relation
shown in Figure 2 (Chomicki, 2002). A user’s preferences
could be formulated as follows: the user prefers a book
to another if and only if their ISBNs are the same and
the price of the first book is lower. In this case, because
the second book is preferred to the first and to the third,
and there are no preferences defined between the last two
books, the winnow operator will return the second, the
fourth, and the fifth book. Chomicki (2002) has studied
the properties of the winnow operator and has proposed
several algorithms for evaluating preference queries.APPLICATIONS
This section discusses how personalization techniques
can be used to tailor the content, products, and inter-
actions to customers’ needs. Some typical personaliza-
tion applications are explained, such as adaptive Web
sites, recommender systems, adaptive Web stores, and
customer relationship management.Adaptive Web Sites
Web sites have been traditionally designed to fit the needs
of a generic user. Adaptive Web sites, which customize
content and interface to suit individual users or groups of
users, provide a more effective way to interact with these
users. An adaptive Web site can semiautomatically im-
prove its organization and presentation by learning fromISBN Vendor Price
0679726691 BooksForLess 14.75
0679726691 LowestPrices 13.50
0679726691 QualityBooks 18.80
0062059041 BooksForLess 7.30
0374164770 LowestPrices 21.88Figure 2: An instance of a book relation.