220 likes | 488 Views
Ken Varnum Information Specialist Research Library & Information Services Ford Motor Company kvarnum@ford.com. Information Agents 14 October 2003. Tom Montgomery Technical Expert Infotronics & Systems Analytics Ford Motor Company tmontgo1@ford.com. Presentation Outline. Introduction
E N D
Ken Varnum Information Specialist Research Library & Information Services Ford Motor Company kvarnum@ford.com Information Agents14 October 2003 Tom Montgomery Technical Expert Infotronics & Systems Analytics Ford Motor Company tmontgo1@ford.com
Presentation Outline • Introduction • Intelligent Agents • Process • Monitoring & Tuning • Conclusions
Introduction • Intelligent agents developed and implemented by Thomas Montgomery, Bardia Madani, and Ken Varnum • Based on a collaboration with MIT that combined mathematical modeling and empirical validation • MIT: product recommendations (music, furniture) • Ford: information retrieval (automotive news)
About RLIS • Ford’s largest library • 9 MLS librarians • 3 Programmer/Developers • 2 Support staff • Branches in England (1) and Germany (2) • Serve Ford Motor Company’s global operations
World Automotive Information • Original abstracts of automotive news • Abstractors select abstracts for inclusion in one of 8 topical “Highlights” sent each week • Customers read the abstracts and click through to full text or document request
World Automotive Information • Inefficient use of abstractors’ time • “One size fits all” approach doesn’t work • Not scalable – becomes hard to add new topics
Intelligent Information Agents • Software analog to human agents • real estate agent, librarian, salesperson • Learn preferences over time
Intelligent Information Agents • Individual Recommendation Agents (not Collaborative Filtering) • Fine grained (users treated as individuals) • Driven by attributes of users and products, therefore can recommend new products
Intelligent Agents vs. Collaborative Filtering • CF: Items I interacted with are compared to Items other people interacted with • Assumes you are like others (requires others) • Requires interaction history prior to recommendation
Intelligent Agents vs. Collaborative Filtering • IA: Features of what I interacted with are compared to Features of new items • Assumes you are unique • Can recommend items with no interaction history
Intelligent Agents vs. Collaborative Filtering • Every document in WAI service is a “new product” • Customer’s interests evolve over time
Data Collection • We mine usage logs to learn about user preferences • Read full abstract • Order photocopy of full text • Click through to full text • Use of database • User doesn’t have to do anything
Agent Mechanism • Each document is turned into a mathematical vector of features: • Keywords - Author • Publication - Age (days old) • Agent compares: • Vectors of users’ previously-selected documents • Vectors of newly-published documents
User Advantages • No overhead from user perspective • No preference panels • No query refining • No document rating • Users unaware of process • Pro: Unobtrusiveness is good • Con: User actions can impact their content
Agent Interaction Feedback Train Agent Recommend
Intelligent Agent Data Flow User Feedback Click Thru Log WAI DB Individualized E-mail Extract Features (Click Thru) Train Intelligent Agent Priorities New Docs Recommend Extract Features from WAI
Percent of Content for Users who Received Personalized Content from Individual Agents Extra Content Distributed
Percent of Clickthroughs to Content by Users who Received Personalized Content from Individual Agents Extra Content Selected
Results • Use of agent improved usage • Technology proved to us it work • Is core technology in our next version of current awareness service
Privacy’s Two-Edge Sword • This works well in a closed environment • Corporate environment allows greater use of “personal” data • System can know a great deal about users • Perhaps less well on public Internet • Privacy concerns result in less data about users • Internet audiences often object to “invasive” observation of actions
Next Steps • Expand to larger subscription service • Allow users to edit their preferences • As an option • As a convenience • Ability for user to reset preferences
Thank You • Updated slides available atvarnum.org/agents.ppt • Questions?