P1: c-143Braynov-2
Braynov2 WL040/Bidgoli-Vol III-Ch-05 July 11, 2003 11:43 Char Count= 0
PREFERENCEMODELING 57aims at providing a common device location standard
to achieve global interoperability between location-based
services regardless of their underlying technologies. At
present, there is a wide range of location identification
technologies. One method, for example, involves the cell
ID number, which identifies a cellular device to a network.
Other methods (Deitel, Deitel, Nieto, & Steinbuhler, 2002)
are triangulation (used by several satellites in GPS), cell
of origin (a cellular phone is located by a nearby tower),
angle of arrival (several towers measure the angles from
which a cellular phone’s signals are received), and ob-
served time difference (the travel time between a cellular
phone and multiple towers is measured to determine the
phone’s location).
Location-based services allow content providers to of-
fer personalized services based on customers’ geographic
position. Mobile users can receive local weather reports,
news, travel reports, traffic information, maps, hotels,
restaurant information, etc. For example, Go2 Systems
(www.go2online.com) provides a mobile Yellow Pages di-
rectory based on users’ location. The directory allows
users to get directions to various nearby services such
as entertainment, real estate, finance, recreation, govern-
ment, and travel.
One problem with LBS is that the small screen and
the limited capabilities of wireless devices can reduce the
level of personalization. For example, it is often imprac-
tical to use Web-based registration forms or question-
naires for explicit personalization. Billsus, Brunk, Evans,
Gladish, and Pazzani (2002) report that only 2–5% of
wireless users customize their interfaces, due to technical
problems or poor content management. Another problem
is complex navigation and the structure of WAP (wire-
less application protocol) sites. Each WAP site consists
of multiple decks, each of them containing one or more
cards. Hypertext links can be made between cards in the
same or in different decks. As a result, users must make
too many selections and move through too many cards in
order to achieve their goals. In addition, the limited pro-
cessing power and slow network access for many mobile
devices lead to extended response times.
To improve Web site navigation for wireless devices,
Anderson, Domingos, and Weld (2001) proposed a person-
alization algorithm, MinPath, which automatically sug-
gests useful shortcut links in real time. MinPath finds
shortcuts by using a learned model of Web visitor behav-
ior to estimate the savings of shortcut links and to suggest
only the few best links. The algorithm takes into account
the value a visitor receives from viewing a page and the
effort required to reach that page.PREFERENCE MODELING
User preferences play an important role in user model-
ing, personalization, and customization. According to the
decision-theoretic tradition, human preferences are mod-
eled as a binary relationRover a set of possible alterna-
tives such as products and services, information content,
layout, and interaction style. A preference relationRholds
between two alternatives (or choice options)XandY(i.e.,
R(X,Y)) ifXis more preferred toY. That is, the user will
choseXwhen he or she faces a choice betweenXandY.The indifference betweenXandYcould be represented as
not(R(X,Y)) and not(R(Y,X)). It has been proved
(Fishburn, 1970) that, under certain conditions, a prefer-
ence relation can be represented by an order-preserving
numeric functionUcalled a utility function. In other
words, alternativeXis preferred to alternativeYif and
only if the utility of Xis greater than the utility of
Y,(U(X)>U(Y)). Knowing a user’s utility function al-
lows a system to offer its customers those products, ser-
vices, or information that maximize the customers’ utility.
The problem of preference modeling has been approached
relatively recently in the computer science community.
Usually users’ preferences are represented based on ad
hoc approaches such as attribute–value pairs and asso-
ciation rules. It is still an open question how to repre-
sent user preferences in a computationally tractable way
and how to reason with incomplete or inaccurate pref-
erences. In this section we discuss three important pref-
erence problems: how to elicit user’s preferences, how to
reason with conditional preferences, and how to take ad-
vantage of users’ preferences in personalizing access to
databases.Preference Elicitation
The process of preference elicitation consists of finding a
user’s preferences for a well-defined set of choices (for ex-
ample, products). In general, preference elicitation could
be performed by interviewing or observing user’s behavior.
Various methods for preference elicitation have been pro-
posed. Most of them assume that consumer preferences
are additive functions over different attributes or decision
objectives. That is, a user multiattribute utility function is
a weighted sum of single-attribute utility functions,U(x 1 ,...,xn)=∑
wiUi(xi),whereU(x 1 ,...,xn) is the user utility function,Ui(xi) are
single-attribute utility functions, andwiare their weights.
The intuition behind this assumption is that it may be
natural to think in terms of utility for each attribute and
then combine these utilities into an overall multiattribute
utility function.
The analytic hierarchy process (AHP) (Saaty, 1980)
is a common method for discovering attribute weights
wi. AHP is carried out in two steps. In the first step, an
attribute hierarchy is set up. In the second step, the user
is asked to compare attributes sharing a parent in the hier-
archy. Pairwise attribute comparisons determine the rel-
ative importance of each attribute with respect to the
attribute on the level immediately above it. The strength
of the comparison is measured on a ratio scale. The com-
parisons are used to build a reciprocal matrix, which is
subsequently used to derive the relative weightswifor the
overall utility function.
Another method for preference elicitation is multiat-
tribute conjoint analysis (Luce, 1977). Attribute values are
usually discretized and every combination of discrete at-
tribute levels is ranked by the user. The rank is then used
as its utility value. The coefficientswiin the overall util-
ity function are derived using statistical methods such as
regression analysis.