Communications of the ACM
Communications of the ACM
Adaptive Web Sites
Designing a complex Web site so that it readily yields its information is a difficult task. The designer must anticipate the users' needs and structure the site accordingly. Yet users may have vastly differing views of the site's information, their needs may change over time, and their usage patterns may violate the designer's initial expectations. As a result, Web sites are all too often fossils cast in HTML, while user navigation is idiosyncratic and evolving.
Understanding user needs requires understanding how users view the data available and how they actually use the site. For a complex site this can be difficult since user tests are expensive and time-consuming, and the site's server logs contain massive amounts of data. We propose a Web management assistant: a system that can process massive amounts of data about site usage and suggest useful adaptations to the Web master. The use of such assistants is one way to develop adaptive Web sitessites that semiautomatically improve their organization and presentation by learning from visitor access patterns.
For example, imagine a site devoted to information about automobiles. The Web master initially decides to organize the site by manufacturer; each auto company will be represented by a dedicated page, containing links to each of their models. However, many visitors to this site intend to comparison shop; a visitor wanting to compare minivans, for example, would have to go to each manufacturer and look up their minivan offerings. Our Web management assistant will observe this common behavior pattern and suggest the Web master add a "minivans" page containing a link to each minivan directly. Even better, the assistant could suggest a complete change in viewcreating a way of browsing the information at the site by car model instead of by manufacturer for all models, not just minivans.
While adaptive Web sites are potentially valuable, their feasibility is still unproven: can nontrivial adaptations be automated? Will adaptive Web sites run amok, yielding chaos rather than improvement? What is an appropriate division of labor between the automated system and the human Web master? To investigate these issues empirically, we consider a case study: we analyze the problem of automatic index page synthesis based on visitor access patterns. An index page is a page consisting of links to a set of pages that cover a particular topic (for example, minivans).
Here, we consider the basic design decisions that underlie any kind of adaptive Web site and examine several kinds of adaptive Web sites by way of illustration. To demonstrate the potential power of adaptive Web sites, we consider the index page synthesis case study in more depth. We introduce the IndexFinder page synthesis system and describe its application to a live Web site. More detail about our algorithms and experiments can be found in [6].
Varieties of adaptive Web sites. Work on adaptive Web sites has taken a variety of approaches, from providing automated "tour guides," to making personalized recommendations, to charting the "footprints" left by visitors to a site. All approaches, however, must address the following questions in one way or another.
What is the division of labor between the automated assistant and the human Web master? The program could be fully manualsimply carrying out changes the Web master specifiesor fully automatedallowed to make changes to the Web site on its own. Many programs will be semi-automated, requiring Web master approval to make limited kinds of alterations. For example, a program may be restricted to non-destructive adaptations that may add information but not destroy existing structure; examples include adding pages (but not removing them) and adding or highlighting links.
Customization or transformation? Customization is adapting the site's presentation to the needs of each individual visitor, based on information about those individuals. Any changes impact only that single user, effectively creating numerous versions of the site, one per user. Customization will be familiar to many readers; many sites allow the creation of personalized home pages. Transformation is improving the site's structure based on interactions with all visitors. A list of "most popular news stories" is a simple example of a global transformation; the site factors in the interests of past visitors to present links likely to appeal to future visitors, even new ones.
What aspects of the Web site are open to change? Manual customization at sites like Yahoo! or Excite allows the user to specify values for a small number of parameters (the stocks in the user's portfolio, the weather report to receive, and so forth). However, other aspects of the pages and the links between pages are etched in stone. More generally, adaptations could add or remove links, highlight links, change text or formatting, and even create completely new Web pages.
Are adaptations based on the content of pages or people's navigational choices? A site that uses content-based adaptation organizes and presents pages based on their contentwhat the pages say and what they are about. Computer programs are still unable to reliably determine the meaning of a Web page or other human utterance; content-based approaches (using keywords, for example) are, necessarily, approximate. On the other hand, access-based adaptation uses the ways previous visitors have interacted with the site to guide how it should be modified. Access information (derived from Web server logs, for instance) might contain common patterns of user behavior that suggest valuable adaptations. Naturally, content-based and access-based adaptations are complementary and may be used together (see the sidebar for examples of different approaches currently found on the Web.)
Index Page Synthesis
Our goal has been to explore the possibilities of adaptive Web sites with a test case. We have chosen to investigate the automatic synthesis of new index pages; an index page is a page consisting of links to a set of pages that cover a particular topic (for example, minivans). Index pages are central to site organization. If we are able to generate index pages valued and adopted by the human Web master, we can have a significant impact on the site's organizational structure over time. Although index page synthesis is an ambitious and challenging goal, it is a non-destructive transformation; the Web master need not worry that the site design will be scrambled or the site's data lost.
We define the index page synthesis problem: given a Web site and a visitor access log, create new index pages containing collections of links to related but currently unlinked pages. An access log contains one entry for each page requested of the Web server (see Figure 1). Each request lists at least the origin (IP address) of the request, the URL requested, and the time of the request. Related but unlinked pages are pages that share a common topic but are not currently linked at the site; two pages are considered linked if there exists a link from one to the other or if there exists a page that links to both of them.
The problem of synthesizing a new index page can be decomposed into several subproblems:
- What is the content (that is, set of hyperlinks) of the index page?
- Does it have a coherent topic? What should its title be?
- How are the hyperlinks on the page ordered?
- How are the hyperlinks labeled?
- Is the page consistent with the site's overall graphical style?
- Is it appropriate to add the page to the site? If so, where?
IndexFinder does most of this work automatically, leaving to the Web master the questions of whether and where the page should be added to the site, and how it should be titled. The first two subproblems have been the focus of our research. The third and fourth are adequately handled with simple rules; hyperlinks are labeled with the title of the target page and ordered alphabetically. Naturally, future work may explore other options. When IndexFinder generates a candidate page, the contents are presented to the Web master. IndexFinder also generates a conceptual description of the page's topic, which is presented in the form of a "rule" defining what appears on the page. Figure 2 shows how IndexFinder and the human Web master interact. IndexFinder processes the site's access logs and generates a candidate index page. The rule shown on the candidate index page specifies that the page contains audio samples of drum machines. The Web master is prompted to accept or reject the candidate page and to provide a title based on the description. The Web master also specifies where in the site the page should be linked. If the page is accepted, IndexFinder creates a final version of the page, using the supplied title and a graphic template for the site. Finally, the page is linked into the site.
While adaptive Web sites are potentially valuable, their feasibility is still unproven.
The main source of information we rely on is the site's Web server log, which records the pages visited by a user at the site. We also rely on the visit-coherence assumption: the pages a user visits during one session at the site are likely to be closely related. One visitor may be interested in minivans, and view several pages on that topic; another may be looking for inexpensive sedans and view all the pertinent pages. We do not assume that all pages in a single visit are related. After all, the information we glean from individual visits is noisy; visitors may pursue multiple distinct interests in a single visit, follow the wrong links, or simply get distracted. However, if a large number of visitors continue to visit and revisit the same set of pages, that provides strong evidence that the pages in the set are related. Thus, we accumulate statistics over many visits by numerous visitors and search for overall trends. Figure 3 depicts the inner workings of IndexFinder.
To generate useful index pages, it is not enough to detect frequently co-occurring pages. Valuable index pages typically correspond to a topic or concept that is intuitive to visitors; they cannot be composed of an arbitrary set of links. More formally, the set of links comprising an index page must be both:
- Pure: containing no inappropriate links; and
- Complete: containing all links that accord with the topic.
For example, if the topic of the index page is "minivans," the page should contain no links to information about sports cars or trucks, and there should be no pages at the site about minivans that are not included in the index page.
The IndexFinder algorithm takes as input a Web server log and a conceptual description of each page at a site. Descriptions are represented as attribute/value pairs (filetype=image, price=cheap). Data about how often pages occur together in user visits is extracted from the Web server logs; IndexFinder then applies a novel statistical cluster mining technique [6] to the data and produces candidate clusters as output. In statistical cluster mining, objects (that is, Web pages) are grouped together based solely on a similarity measure (how often they co-occur in user visits); the algorithm makes no guarantee that the clusters are pure or complete with respect to some topic. Next, each candidate cluster (represented as a set of positive examples of a concept) is fed to a concept learning algorithm, which then produces the simple concept that most closely describes what the set of pages have in common. The concept is expressed as a conjunction of logical attributes; in Figure 2, for example, these attributes are (1) the instruments are all drum machines; and (2) the files are all audio samples. Pages in the original candidate cluster that do not match the concept description are discarded; similarly, other pages at the site that do fit the description are added. This new complete and pure cluster is presented to the Web master, who decides whether or not it should be added to the site as a new index page.
Deploying IndexFinder
To test IndexFinder in practice we deployed it on a Web site and compared user response to the pages IndexFinder generated against the pages generated by other algorithms and against human-authored index pages. We found that IndexFinder outperforms previous algorithms and its pages even compare quite favorably to human-authored index pages.
Our experiments draw on data collected from the Music Machines Web site (see machines.hyperreal. org), which is devoted to information about many kinds of electronic musical instruments. Music Machines contains approximately 2,500 distinct pages, including HTML pages, plain text, images, and audio samples. The site receives approximately 10,000 hits per day from roughly 1,200 distinct visitors. Each page at the site was described by 13 attributes such as the type of instrument, the cost, the country of manufacture, the file type, and other domain-specific attributes. All pages were hand-tagged by the Web master (one of the authors).
We evaluate IndexFinder by comparing the quality of the index pages it synthesizes with the quality of two other sets of index pages:
- Human-authored: We took four index pages from the Music Machines site and included them in our experiment.
- COBWEB: We applied COBWEB [3], the leading conceptual clustering algorithm, to the task of index page synthesis. Like IndexFinder, COBWEB produces only clusters that are pure and complete with respect to some concept.
Note that, unlike IndexFinder and COBWEB, statistical clustering techniques such as [7, 8] and standard data mining algorithms such as [1] do not yield complete and pure index pages of the sort we want to provide to the Web master, so we do not include them in this comparison. In separate experiments, we have shown that PageGather (the statistical cluster mining component of IndexFinder) outperforms standard methods [6].
To generate useful index pages, it is not enough to detect frequently co-occurring pages.
Both COBWEB and IndexFinder generate a set of clusters, of which the top 10 are chosen. Each cluster is converted into a Web page containing a link to each page in the cluster. Link text is generated from the title of the linked-to page (or the filename if no title is available). Each cluster is presented to the Web master, along with its logical description, for naming. For example, presented with the rule "filetype=image, price=cheap, instrument-type=synth," the Web master might name the page "images of cheap synthesizers."
In this way, we generated a large set of index pages; next, we added in the human-authored index pages. To conduct our experiment, five of these index pages are randomly chosen and made accessible on the Web site each day. The index pages appear in the site's navigation bar, which appears on all site pages (see machines.hyperreal.org).
We compared index pages with respect to impact and benefit. Impact measures the number of people who use the automatically generated index pages; benefit measures the number of links these users follow on each index page. Over the course of each day, we recorded how many users visited each index page, and how many links they clicked on. We computed the number of visitors who clicked on at least one link, the number who clicked on at least two links, and so on. We accumulated these daily counts over the course of a month, and calculated an average daily impact and benefit for each index page. We then averaged these impact and benefit statistics over all index pages generated by a particular algorithm. For each algorithm, we plotted a line of the number of pages viewed (benefit) vs. the number of people who viewed that many pages (impact). If no users viewed more than, say, five pages from any index page for an algorithm, then the line for that algorithm will stop at five. The results of our experiment are shown in Figure 4. Note that impact is measured on a log scale, as it drops off exponentially with increasing benefit. IndexFinder performs substantially better than COBWEB but not as well as the human-authored pages. While IndexFinder and the human clusters seem to be converging, COBWEB drops off more quickly.
Lessons and Future Work
We began this work believing that by computing statistics over the visitor access log for Web sites, we would be able to identify valuable adaptations and present them to the site's Web master. We've learned that to mine this information effectively, we needed to combine the statistical patterns gleaned from the log file with logical descriptions of the contents of each Web page. This observation led us to move from our statistical cluster-mining algorithm, called PageGather, to IndexFinder, which fuses statistical and logical information to synthesize index pages. Logical descriptions of Web pages could be derived from XML annotations or from a database of the kind used in Strudel [5].
In future work we plan to move toward a higher degree of automation where IndexFinder suggests possible names for the pages it creates (based on their logical description) as well as appropriate places to insert the pages into the Web site (based on visitor access patterns). We also plan to explore intermediate points on the spectrum between customization and transformation. Specifically, we plan to identify classes of visitors to the site (for example, prospective students visiting a university's Web site) and target specific adaptations at these classes of visitors.
More generally, the IndexFinder case study demonstrates the potential of automatic, and semi-automatic, adaptation to improve Web sites for populations of visitors. Beyond the manual customizations so common in today's portal Web sites (Yahoo! is a good example), we are beginning to glimpse the potential of data mining of visitor access logs to continually transform and refine Web sites.
References
1. Agrawal, R. and Srikant, R. Fast algorithms for mining association rules. In Proceedings of the 20th VLDB Conference, 1994.
2. Fink, J., Kobsa, A., and Nill, A. User-oriented Adaptivity and Adaptability in the AVANTI Project. In Designing for the Web: Empirical Studies. Microsoft Usability Group, Redmond, WA, 1996.
3. Fisher, D. Iterative optimization and simplification of hierarchical clusterings. Journal of Artificial Intelligence Research 4, 1996.
4. Good, N., Schafer, J., Konstan, J., Borchers, A., Sarwar, B., Herlocker, J., and Riedl, J. Combining collaborative filtering with personal agents for better recommendations. In Proceedings of the Sixteenth National Conference on Artificial Intelligence, 1999.
5. Levy, A., Florescu, D., Suciu, D., Kang, J., and Fernandez, M. Catching the boat with Strudel: Experiences with a Web site management system. In Proceedings of the ACM SIGMOD Conference on Management of Data, 1998.
6. Perkowitz, M. and Etzioni, O. Towards adaptive web sites: Conceptual framework and case study. Artificial Intelligence, To appear, 2000.
7. Rocchio, J. Document retrieval systems optimization and evaluation. Ph.D. dissertation, Harvard University, 1966.
8. Voorhees, E.M. Implementing agglomerative hierarchical clustering algorithms for use in document retrieval. Information Processing & Management 22 (1986).
9. Wexelblat, A. and Maes, P. Footprints: History-rich Web browsing. In Proceedings of the Conference on Computer-assisted Information Retrieval (RIAO), 1997, 7584.
Figures
Figure 1. Typical user access logs, these from a computer science Web site. Each entry corresponds to a single request to the server and includes originating machine, time, and URL requested. Note the series of accesses from each of two users (one from SFSU, one from UMN).
Figure 2. IndexFinder suggests candidate index pages to the Web master. IndexFinder generates a candidate index page. The human Web master acepts or rejects the candidate page. If the page is accepted, IndexFinder creates a final version of the page and adds it to the Web site; the Web master specifies the page's title and where in the site it should be linked. Rejected pages are discarded.
Figure 3. The inner workings of IndexFinder. IndexFinder consists of three basic modules. The log processing module takes the Web server access logs and computes how often pages co-occur in user visits. The cluster mining module takes this information and a graph of the Web site and finds clusters of frequently co-occurring pages. The conceptual clustering module uses conceptual descriptions of the site's pages to convert these clusters into coherent concepts. These are output as candidate index pages and presented to the Web master.
Figure 4. Results from our case study. Impact measures the number of people who use the automatically generated index pages; benefit measures the number of links these users follow on each index page. IndexFinder performs substantially better than COBWEB but not as well as the human-authored pages.
©2000 ACM 0002-0782/00/0800 $5.00
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.
The Digital Library is published by the Association for Computing Machinery. Copyright © 2000 ACM, Inc.
No entries found