Discussion Notes

(courtesy Saverio Perugini)

purely collaborative
mines URLs
role specialization
uses IR metrics - precision and recall - acceptable here
results verified against FAQs
results could be used to create FAQs in domains where they currently do not exist and/or improve upon existing FAQs

discovers a `static' social network; entities are assumed to not move stations during the time under consideration
lifetime important - in this case it is many years
social network modeling can be very fickle; need to know when two entities being referred to in different ways are actually the same and when to give importance to distinctions
no weight associated with links
only one type of link
strong theoretical basis from bibliometrics
very error prone
Auto-refresh tricks in AltaVista and Excite
see fifth reference for one of the first untouted works on recommender systems

a hybrid of both content and collaborative-based filtering
content features are orthogonal
people's tastes are not necessarily orthogonal - need to mine types, which introduces excess and defect, which is what we recommend on
cluster by ratings, not tastes

overlap in bookmark folders
how do they get the bookmarks?
similarity relationship is directed and not necessarily reciprocal - sets Siteseer apart from the other recommender systems in this CACM special issue on recommender systems
backflip system allow users to share bookmarks
not to be confused with Citeseer

clusters users - one way of handling sparsity (another way coming in next class)
recommend only within the bucket / cluster
author went on to start NetPerceptions - personalization

mostly collaborative - reflects the fact that this issue was born out of a collaborative filtering workshop
no talk of any detailed domain-specific modeling
bedside reading, not very technical