P1: c-143Braynov-2
Braynov2 WL040/Bidgoli-Vol III-Ch-05 July 11, 2003 11:43 Char Count= 0
APPLICATIONS 59visitors’ access patterns (Perkowitz & Etzioni, 2000). An
adaptive Web site may automatically create new pages,
remove or add new links, highlight or rearrange links,
reformat content, etc. In general, Web site adaptiveness
could be classified as being either individual- or group-
oriented. An individually adaptive Web site consists of a
large number of versions—one for each individual user.
Group-oriented adaptiveness targets groups of users and
requires a smaller number of Web site versions—one for
each group. For example, a Web site may have one view
for corporate users and another view for individual users,
or one view for domestic visitors and another view for
international visitors.
Depending on the way it is performed, adaptation can
be classified as content-based or as access-based. Content-
based adaptation involves presenting and organizing Web
pages according to their content. Access-based adaptation
is based on the way visitors interact with a Web site. It
involves tracking users’ activity and personalizing content
and layout according to users’ navigation patterns.
The problem of designing personalized Web sites is
complicated by several factors. First, different users may
have different goals and needs. Second, the same user may
have different goals at different times. Third, a Web site
may be used in a way different from the designer’s expec-
tations. It is still not clear what kind of adaptation can be
automated and what is the appropriate tradeoff between
user-controlled and automatically guided navigation.
The idea of adaptive Web sites was popularized by
Perkowitz and Etzioni (2000). They proposed the Page-
Gather algorithm, which automatically generates index
pages that facilitate navigation on a Web site. An index
page is a page consisting of links to a set of existing but cur-
rently unlinked pages that cover a particular topic of inter-
est. In order to find a collection of related pages on a Web
site, the PageGather algorithm employs cluster mining.
The algorithm takes the Web server access log, computes
the co-occurrence frequencies between pages, and creates
a similarity matrix. The matrix is then transformed into
a similarity graph, and maximal cliques are found. Each
maximal clique represents a set of pages that tend to be
visited together.Recommender Systems
Recommender systems (Schafer, Konstan, & Riedel,
2001) have been used in B-to-C e-commerce sites to make
product recommendations and to provide customers with
information that helps them decide which product to buy.
Recommender systems provide a solution to the prob-
lem of how to choose a product in the presence of an
overwhelming amount of information. Many e-commerce
sites offer millions of products, and therefore, choosing
a particular product requires processing large amounts
of information, making a consumer choice difficult and
tedious.
Recommender systems contribute to the success of e-
commerce sites in three major ways (Schafer et al., 2001).
First, they help to improve cross-sell. Cross-sell is usu-
ally improved by recommending additional products for
the customer to buy. For example, by looking at the prod-
ucts in the customer’s shopping cart during the checkoutprocess, a system could recommend additional compli-
mentary products. Second, recommender systems could
help convert occasional visitors into buyers. By providing
a recommendation, a retailer could deliver customized in-
formation, increase the amount of time spent on a Web
site, and finally, increase the customer’s willingness to
buy. Third, recommender systems help build loyalty and
improve customer retention. Personalized recommenda-
tions create a relationship between a customer and a Web
site. The site has to invest additional resources into learn-
ing customers’ preferences and needs; on their part, cus-
tomers have to spend time teaching a Web site what their
preferences are. Switching to a competitor’s Web site be-
comes time-consuming and inefficient for customers who
have to start again the whole process of building person-
alized profiles. In addition, customers tend to return to
Web sites that best match their needs.
Recommender systems use different methods for sug-
gesting products (Schafer et al., 2001). One of the
most common methods is item-to-item correlation. This
method relates one product to another using purchase
history, attribute correlation, etc. CDNOW, for instance,
suggests a group of artists with styles similar to that of
the artist that the customer likes. Another recommen-
dation method is user-to-user correlation. This method
recommends products to a customer based on a correla-
tion between that customer and other customers visiting
the same Web site. A typical example is “Customers who
bought” in Amazon.com. When a customer is buying or
browsing a selected product, this method returns a list
of products purchased by customers of the selected prod-
uct. Other recommendation methods include statistical
summaries. For example, the Purchase Circles feature of
Amazon.com allows customers to view the “top 10” list
for a given geographic region, company, educational in-
stitution, government, or other organization.
CDNOW enables customers to set up their own mu-
sic stores, based on albums and artists they like. Cus-
tomers can rate albums and provide feedback on albums
they have purchased. The system keeps track of previously
purchased albums and predicts six albums the customer
might like based on what is already owned.
A recent study by Haubl and Murray (2001) shows that
recommendation algorithms have the potential to ma-
nipulate and influence user preferences in a systematic
fashion. The authors performed a controlled experiment
showing that the inclusion of a product feature in a rec-
ommendation renders the feature more prominent in cus-
tomers’ purchase decisions.Adaptive Web Stores
Adaptive Web stores are a special type of adaptive Web
sites that can use a customer profile to suggest items best
fitting the customer’s needs. The main difference between
adaptive Web stores and adaptive Web sites is that Ardis-
sono and Goy (2001) describe a prototype toolkit (SETA)
for creating adaptive Web stores. In contrast to other
adaptive Web stores, SETA adapts not only item selec-
tion, but also layout and content selection to the prefer-
ences and expertise of customers. SETA chooses a catalog
presentation style (item description, background colors,