|CSE Home||AI Home||About CSE||Search||Contact Info|
Web Site Personalizers
Most web sites today are designed with a "one-size-fits-all" philosophy. The web site designer "specs out" the needs of the web visitor, and then creates the content accordingly. However, one size frequently does not fit all. Visitors may use the site in ways unanticipated by the site designer. For example, a site that contains a broad array of information, such as a software support site, may be structured with as deep hierarchcy according to product, version, operating system, and error condition. However, visitors may follow only a few paths through this site, and the deep hierarchy is needlessly burdensome. In addition, a visitor's web client may be constrained in same way, perhaps by a small screen size or slow network connection. Content designed for large, color displays and fast networks is seldom useful for such constrained devices.
To break the mold of one-size-fits-all sites, we propose building web site personalizers -- intermediaries between servers and visitors that automatically adapt and customize content for each visitor. A personalizer can:
How Proteus works
We view personalization as a two step process. In the first step, the personalizer builds a model of each visitor. The model includes information about navigational browsing behavior as well as content interests. This model could also include "out-of-band" information, such as visitors' geographic location or demographics. In the second step, the personalizer transforms the site to maximize the expected utility for a given visitor. The expected utility of a personalized web site is a measure of how much benefit the visitor will receive by browsing the site; the personalizer computes this value based on the visitor model derived in the first step.
Proteus builds the visitor models by mining the web usage logs from the site. One part of the visitor model estimates the probability the visitor will view each page given where he or she has gone in the past. We have experimented with a variety of predictive models with Proteus, including Markov models and mixtures of models. The other part of the model tries to predict the visitor's interest in a particular block of content (paragraph of text) on a page. Proteus builds a word vector model of the visitor, which is a table that associates a score with every word the visitor has ever seen before. Essentially, the more often the word has appeared, the higher its score. To evaluate the visitor's interest in a block of text, Proteus compares the words in the text with the visitor's model -- the more high-scoring matches, the higher the interest.
Proteus uses the visitor model to estimate the expected utility of a personalized web site for the given visitor. Intuitively, expected utility is the sum of the value the visitor receives by viewing each page in the site, in isolation, discounted by the difficulty of reaching each page. That is, more distant pages are more heavily discounted. Expected utility is a bit more complex than this, and its calculation requires some subtlety; take a look at our WWW10 paper for more details.
To actually personalize the web content, Proteus uses a hill-climbing search through the space of web sites. Each state in the space is a complete, personalized web site. The search operators Proteus uses can add links to the page requested by the visitor, and can elide content from that page (Proteus limits its changes to affect only the requested page to reduce the complexity of its search). Expected utility based on the visitor model provides the evaluation function, and Proteus continues the search either until it reaches a maxima or until some time limit expires.
Last updated: April 13th, 2001
Department of Computer Science & Engineering|
University of Washington
Seattle, WA 98195-2350
(206) 543-1695 voice, (206) 543-2969 FAX
[comments to Corin Anderson]