LEARNING TO SURF MULTIAGENT SYSTEMS FOR ADAPTIVE WEB PAGE RECOMMENDATION A DISSERTATION SUBMITTED TO THE DEPARTMENT OF COMPUTER SCIENCE AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEG REE OF DOCTOR OF PHILOSOPHY Marko balabanovic March 1998
LEARNING TO SURF: MULTIAGENT SYSTEMS FOR ADAPTIVE WEB PAGE RECOMMENDATION a dissertation submitted to the department of computer science and the committee on graduate studies of stanford university in partial fulfillment of the requirements for the degree of doctor of philosophy Marko Balabanovic March 1998
O Co pyright 1998 by M arko b al All Rights reserved
c Copyright 1998 by Marko Balabanovic All Rights Reserved ii
I CEFIIEY THAT I HAVE READ THIS DISSERLATION AND THAT IN MY OPINION IT IS FUITY ADEQUATE, IN SCOPE AND IN QUAIITY, AS A DISSEFIATION FOR THE DEGREE OF DOCIOR OF PHIIOSOPHY YOAV SHOHAM (PHINCIPAL ADVISER I CEFIIEY THAT I HAVE HEAD THIS DISSERIATION AND THAT IN MY OPINION IT IS FUITY ADEQUATE, IN SCOPE AND IN QUAITY, AS A DISSEFIATION FOR THE DEGREE OF DOCIOR OF PHIIOSOPHY TEREY WINOGRAD I CEFIIEY THAT I HAVE HEAD THIS DISSERIATION AND THAT IN MY OPINION IT IS FUITY ADEQUATE, IN SCOPE AND IN QUAlITY, As A DISSEFIATION FOR THE DEGHEE OF DOCIOR O ILOSOPHY ROBERT BARRETT APPROWED FOR THE UNIVERSITY COMMITIEE ON GEADUATE STUDIES
I certify that I have read this dissertation and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy. Yoav Shoham (Principal Adviser) I certify that I have read this dissertation and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy. Terry Winograd I certify that I have read this dissertation and that in my opinion it is fully adequate, in scope and in quality, as a dissertation for the degree of Doctor of Philosophy. Robert Barrett Approved for the University Committee on Graduate Studies: iii
Iv
iv
A bstract Imagine a newspaper per Sonalized for your tastes. Instead of a selection of articles hosen for a general audience by a human editor, a Software agent picksitemsjust for you, covering your particular topicS of interest. Since there are no journ alists at its disposal, the agent searches the Web for appropriate articles. Over time, it uses yo feedback on recommen ded articles to build a mo del of your interests. This thesis investigates the design of "recommen der systems" which create Such per Son alized newspapers Two research issu es motivate this work and distinguish it from approach es usually taken by inform ation retrieval or m achinelearning researchers. First, a recommender rstem will have many userS, with overl apping interests. How can this be exploited nalized new sp sts of a small set of ar ticles Techniques for deciding on the relevance of individu al ar ticles are well known, but how is the com dosition of the set determined? One of the primary contributions of this research is an implemented architecture linking populations of adaptive Software agents. Common interests among its userS ed both to increa recommendations. A novel interface infers document preferences by monitoring user crag and drop actions, and affords control over the com position of sets of recommen dationS. Results are presented from a variety of experiments: user tests meaSuring learning performance, simulation stu Cies isolating particular tradeoffs, and uSability tests investigating inter action designs
Abstract Imagine a newspaper personalized for your tastes. Instead of a selection of articles chosen for a general audience byahuman editor, a software agent picks items just for you, covering your particular topics of interest. Since there are no journalists at its disposal, the agent searches the Web for appropriate articles. Over time, it uses your feedback on recommended articles to build a model of your interests. This thesis investigates the design of \recommender systems" which create such personalized newspapers. Two research issues motivate this work and distinguish it from approaches usually taken by information retrieval or machine learning researchers. First, a recommender system will have many users, with overlapping interests. How can this be exploited? Second, each edition of a personalized newspaper consists of a small set of articles. Techniques for deciding on the relevance of individual articles are well known, but how is the composition of the set determined? One of the primary contributions of this research is an implemented architecture linking populations of adaptive software agents. Common interests among its users are used both to increase eciency and scalability, and to improve the quality of recommendations. A novel interface infers document preferences by monitoring user drag-and-drop actions, and a ords control over the composition of sets of recommendations. Results are presented from a variety of experiments: user tests measuring learning performance, simulation studies isolating particular tradeo s, and usability tests investigating interaction designs. v