P1: c-143Braynov-2
Braynov2 WL040/Bidgoli-Vol III-Ch-05 July 11, 2003 11:43 Char Count= 0
56 PERSONALIZATION ANDCUSTOMIZATIONTECHNOLOGIESthe last page before backtracking occurs. The algorithm
first converts server log data into a set of maximal for-
ward references and then determines frequent traversal
patterns, i.e., maximal forward references that occur fre-
quently.
Another approach is taken in the Web Utilization Miner
(WUM) (Spiliopoulou & Faulstich, 1999). The authors of
WUM not only identify sequences of frequently accessed
pages but also find less frequent paths having structural
or statistical properties of interest. In WUM the path fol-
lowed by a user is called a trail. Because many users can
display similar navigation patterns, users’ trails are aggre-
gated into a tree by merging trails sharing a prefix. The ag-
gregate tree can be subsequently used for predicting user’s
behavior.
Markov models have also been used for modeling users’
browsing behavior (Deshpande & Karypis, 2001). Markov
models predict the Web page accessed by a user given the
sequence of Web pages previously visited. Such models
have proved to display high predictive accuracy. On the
other hand, they have high state-space complexity, which
significantly limits the scope of their applicability.
Mobasher, Dai, Luo, and Nakagawa (2002) proposed an
aggregate profiling algorithm based on clustering transac-
tions (PACT). User’s transactions are represented as mul-
tidimensional space vectors of page views. The vectors
are grouped into clusters, each of them representing a set
of users with similar navigational patterns. Subsequently,
every cluster is associated with a single point (the cluster’s
centroid) representing an aggregate profile of all users in
that cluster. A new user activity is matched against aggre-
gate profiles and items are recommended based on the
degree of matching.INTELLIGENT AGENTS FOR
PERSONALIZATION
Intelligent agent technology provides a useful mechanism
for personalization and customization. The term intelli-
gent agent has been used in different meaning by differ-
ent authors. By agent we refer here to a software program
acting on behalf of its user and having the following prop-
erties (Weiss, 1999):Autonomy: agents operate without the direct intervention
of their user.
Social ability: agents interact with other agents (including
humans) via agent-communication language.
Reactivity: agents perceive and adapt to a dynamically
changing environment.
Proactiveness: agents do not simply act in response
to their environment; they are able to exhibit goal-
directed behavior.
Rationality: every agent has a representation of its user’s
preferences and tries to satisfy them in the best possible
way.Intelligent agents come in different types. Internet
agentshelp the user collect, manipulate, and analyze infor-
mation. Some of them are embedded within an Internet
browser and help the user navigate through a Web site.Interface agentsact like personal assistants collaborating
with the user. Interface agents monitor, observe, and learn
from the user’s actions. Some interface agents can assume
the form of a synthetic character; other can model users’
emotions or chat with the user, using natural language.
Collaborative agentsact as a team to achieve some com-
mon goal.Mobile agentscan roam the Internet and inter-
act with foreign hosts and other agents on behalf of their
users.
In other words, intelligent agents can be viewed as
proxies of human users, capable of learning and reasoning
about users’ preferences. WebWatcher (Joachims, Freitag,
& Mitchell, 1997) was one of the first software agents to
assist users browsing the Web. It guides users through a
Web site by trying to predict the paths they will follow
based on the navigation history of previous users of the
Web site. WebWatcher believes that a particular hyperlink
is likely to be followed by a user if like-minded visitors pre-
viously followed it. WebWatcher suggests a link based on
current user, user’s interest, and a Web page. The user’s in-
terest is represented by a set of keywords and a hyperlink
is represented by a feature vector. When a new user enters
a Web page, WebWatcher compares the current user’s in-
terest with the descriptions of the links on the page and
suggests the links correlated with the user interest. Web-
Watcher also uses reinforcement learning to learn how
to navigate through a Web site. The learning is based on
positive reinforcement WebWatcher receives whenever it
chooses a link that fits the user’s interests.
Letizia (Lieberman, 1997) is another software agent for
client-side personalization. It learns a model of its user by
observing his or her behavior. The user model consists of
a list of weighted keywords representing the user’s inter-
ests. Letizia explores the Web ahead of its user and recom-
mends potential links of interest. It records every choice
of the user on a Web page and takes the act of viewing
a Web page as a positive feedback (evidence of interest).
Letizia tries to incrementally learn the user’s interest in a
page by observing the choices he or she makes. This is an
example of unobtrusive personalization, which does not
require explicit user interaction: the user is not asked to
explicitly rank or evaluate Web pages.LOCATION-BASED PERSONALIZATION
The information used for personalization may range from
a history of past purchases and browsing behavior to
explicitly provided user preferences. The rapid growth
of wireless networks and mobile commerce provide new
opportunities for personalization by offering more user-
specific information such as geographic location, date,
time, and travel direction. Handheld devices, for exam-
ple, allow customers to receive personalized content and
recommendations on the move, at home, and at work.
One of the most promising technologies is location-
based services (LBS), which allows business to identify a
user’s location and offer context-dependent services. LBS
holds the potential to significantly improve CRM, wireless
marketing, and emergency services.
In October 2000, Ericsson, Motorola, and Nokia
founded the location interoperability forum (LIF) es-
tablished to provide location-based services. The forum