Information Technology Reference
In-Depth Information
and the other for mobile users. A viable solution is adaptive Web sites (Perkowitz & Etzioni, 1997). An
adaptive Web site dynamically changes its contents or structure based on browsing activities.
Following the idea of adaptive Web sites, Anderson, Domingos, and Weld (2001) proposed shortcuts
to improve mobile Web navigation. A shortcut is a dynamic link that provides a shorter path with fewer
clicks for users to reach their desired pages. A shortcut to a destination page is dynamically created and
inserted into the next page a user is going to browse. If that destination page is the one in which the user
is interested, the user can access the page by following the shortcut. For example, assume a browsing
session consists of A-B-C-D-E-F-G, where each letter represents a page. After browsing pages A and
B, if a shortcut to G is created and inserted into page C, the user can follow the shortcut to reach G,
without going through intermediate pages D, E, and F. The critical question is how to find shortcuts
that are useful with only part of the session known. A shortcut C → H, for example, is useless in the
previous example.
In order to provide useful shortcuts, Web usage mining techniques are employed. User browsing
patterns are extracted from Web server logs. These patterns are built into prediction models that can be
used to predict user browsing behaviors. Given a partial session, such prediction models will compute
what other pages in which the user may be interested. These predictions are used to create and recom-
mend shortcuts for mobile users.
A critical component in this approach is the prediction model. The model should be as accurate as
possible with as little information about the session as possible. An accurate shortcut found earlier in
a session is more worthwhile than one found close to the end of a session. Moreover, the prediction
model should be easy to build and use. In their MINPATH algorithm, Anderson et al. (2001) used
Markov models, which proved to be accurate (Anderson et al., 2001). However, those models require
Web graphs. In this chapter, we propose to use a simpler prediction model, N-gram, for learning user
browsing patterns.
N-grams are well known and are widely used in speech and text processing applications. Researchers
have found that accuracy increases with N, the order of N-grams. For example, 4-grams are more accu-
rate than 3-grams, which is turn is more accurate than 2-grams. Though accuracy increases with higher
values of N, it requires a larger number of training sessions to have a well trained N-gram model.
An N-gram based prediction model for Web browsing patterns is proposed by Su et al. (2000). The
N-gram model has several advantages over other prediction models. It is simple, robust, and easy to
use. Besides, N-gram does not use a Web graph. In our study, the same N-gram model with a slightly
different lookup operation is used. Moreover, its effectiveness in improving mobile Web navigation is
examined.
In our approach, first, Web server logs are preprocessed to identify sessions. A session is conceptu-
ally a single visit. The sessions are then used to train an N-gram model. A revised version of MINPATH
algorithm, MINCOST, is proposed to find shortcuts. MINCOST uses a different function in calculating
the saving and ranking of shortcuts. Our approach has been implemented and evaluated against two real
data sets from NASA and EPA Web servers. Our experiments show that the N-gram prediction model
is as effective as more sophisticated models in recommending useful shortcuts.
The chapter is organized as follows. The second section discusses related work in Web usage mining,
adaptive Web sites, and MINPATH algorithm. Our approach is presented in Section 3. Experimental
results with two data sets are reported in the fourth section The fifth section concludes the chapter and
gives some future research direction.
Search WWH ::
Custom Search