Mike Perkowitz


6910 Roosevelt Way NE, #124
Seattle, WA 98115

m a i l (at) p e r k o w i t z (dot) n e t
http://www.perkowitz.net



Research Statement.

The focus of my research has been applications of machine learning and data mining to user interfaces, particularly on the World Wide Web. The deep problems of artificial intelligence lead to fascinating research; and if we are still far away from true intelligence, AI has many more immediate applications. My own interest is in improving human-computer interaction; I would like to make computers easier to use by making them somewhat more intelligent. Although I focus on the "more intelligent" rather than the "easier to use" aspect, I believe that computers and the Internet could and should be useable by and accessible to everyone.

My research, then, has tended to focus on making computers more responsive or more helpful in their interactions with people, from interacting with information resources on our behalf to improving the presentation of a web site based on how people use it. Techniques from machine learning and data mining, I believe, are especially applicable to this goal. Especially in web applications, there is great potential to learn from the way users interact with software.

As an undergraduate, I studied cognitive science at Brown University, with emphasis on linguistics and artificial intelligence. My honors thesis -- advised by Leslie Kaelbling and Mark Johnson -- was on reinforcement learning for simple agents in a simulated non-Markovian domain. While there I also did some work with Eugene Charniak and several other students on statistical language processing. When I came to the University of Washington, I began working with Oren Etzioni.



The Internet Learning Agent.

The Internet Learning Agent (ILA) learns how to understand Internet information resources (such as phone books) by interacting with them. Starting with a small amount of knowledge, ILA can bootstrap itself to understanding a number of information sources. We published a paper on ILA in IJCAI '95: Category Translation: Learning to Understand Information on the Internet [1]. My work on ILA was also my masters thesis in 1995. ILA was a part of the Internet Softbot, an intelligent agent for the internet; users would tell the softbot what to do (e.g., finding someone's phone number), and the softbot would figure out how to do it (e.g., searching online phonebooks). ILA's ability to learn how to use new information sources would make the softbot capable of expanding its abilities automatically.

Automatically learning to interact with information resources is a complex problem and ILA tackled only part of it -- translating the output of such a resource into familiar concepts. The ShopBot, in a sense, addressed another portion of this problem -- discovering information resources and learning how to communicate with them. The authors of ILA and of the ShopBot (myself, Robert Doorenbos, Oren Etzioni, and Dan Weld) wrote an article giving an overview of this problem and discussing ILA and the ShopBot in that context. Learning to Understand Information on the Internet: an Example-Based Approach[2] appeared in the Journal of Intelligent Internet Systems in 1997.



Web Browsing and Sharing.

Our next project was Clio, an approach to allowing a person to browse her own web history through a dialogue with an intelligent assistant. The user can say things like "Show me that page about cars I was looking at last week" and the system will find the page, possibly narrowing down the choices by asking questions such as "Do you mean the one about sports cars or the one about sedans?". Unfortunately, though we had many ideas and several implementations, we never published any findings.

I spent the summer of 1996 as an intern at Microsoft Research with the User Interfaces Group. While there, I worked on an idea for categorizing and sharing one's web history through interaction with characters. Characters had distinct personalities and kept track of their own favorite web pages and could make recommendations to users. Multiple users shared access to a collection of characters, allowing page references to be shared among users filtered through the characters. I presented this work at the Lifelike Computer Characters conference in 1996.



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 a priori expectations. As a result, many web sites are difficult to use and difficult to design and maintain.

As a solution, we propose Adaptive Web Sites: web sites that improve their structure and presentation by learning from interactions with users. Examples include personalized recommendations, intelligent "tour guides", and automatic customization. Our own system observes the behavior of visitors to a site to find topics of interest and creates new web pages on those topics. For example, the system might notice that many visitors to an automobile information site view information on various minivans, one at a time; the system would create a new page with information on minivans to faciliate this comparison shopping.

We have published a number of papers on this subject, including a challenge paper in IJCAI '97[3] as well as an article[5] for the AI Journal. This problem domain has led us to new algorithms for statistical cluster mining and conceptual cluster mining, as well as useful applications for web sites. Some links and downloadable data are available at UW CSE.



References.

[1]

M. Perkowitz and O. Etzioni "Category Translation: Learning to Understand Information on the Internet." Proceedings of IJCAI95. pp. 930-936.

[2]

M. Perkowitz, R. Doorenbos, O. Etzioni, and D. Weld. "Learning to Understand Information on the Internet: an Example-Based Approach." Journal of Intelligent Information Systems. Vol. 8, num. 2, 1997. pp. 133-153.

[3]

M. Perkowitz and O. Etzioni "Adaptive Web Sites: an AI Challenge." Proceedings of IJCAI97.

[4]

M. Perkowitz, and O. Etzioni. "Adaptive Web Sites: Concept and Case Study." Communications of the ACM. To appear.

[5]

M. Perkowitz, and O. Etzioni. "Towards Adaptive Web Sites: Conceptual Framework and Case Study." Artificial Intelligence Journal. To appear.