Discussion Notes

Jan 26

Main Topic: Targeting Dichotomy

• Other Possible Dimensions: Geography, Genre etc.
• Lattice structure induced by choices of targeting dimension

Experience with Personalization on Yahoo!

• Targeting: by user
• Good guidelines about large-scale Web Personalization
• Author Udi Manber: used to be a theory professor at University of Arizona; experience with building large scalable systems
• Tradeoff (dumb X smart)
• Personalization improved with customization of user interfaces
• People are lazy and do not like changes
• Interesting homegrown-built database architecture 
• Usability depending on perceived tradeoff effort X benefit

Adaptive Web Sites 

• Targeting: By topic
• Constructive synthesis of  "index pages" by mining patterns of user browsing
• Access log processing: obvious approach 
• Considered the best paper of the issue (Controversial!) 
• Necessity of metadata/descriptions about pages
• Possibly not scalable if pages are not automatically generated
• Conceptual clustering
• yields clusters that are easy to describe 
• sacrifices accuracy
• there is no "right clustering"
• they do not know topics a priori 
• Pure X Complete = Precision X Recall
• Good observations about destructive versus non-destructive approaches
• Limited to patterns of browsing, no consideration about searching

Syskill & Webert: Identifying interesting web sites

• Targeting: Per user Per topic
• Learning of user profiles from user evaluations to recommend Web pages by topic
• Gross comparisons between machine learning and IR algorithms for this task
• Adopts a very simplistic view of the problem
• Small number of users in experiments
• Claimed conclusions do not appear to follow from results