Notes on Collaborative Filteringg

• based on Ch. 9 in Rajaraman and Ullman: Mining Massive Data Sets. • CF – major alternative to content-based approach to RS. • more popular, because ... • assumes no meta-data or “features” on items or users available; nor requires any features to be easily extractable. • Idea: Treat the column of each item as a representation of the item; how is it liked by various users. • Similarly, the profile of each user is just his/her rating pattern on the items s/he has rated. • Two approaches, depending on whether we want to work with users or with items: user-user or item-item. • Orthogonally, two approaches, depending on whether we use the ratings matrix as is, to make our predictions, or want to build models of user behavior – memory-based vs. model-based. • In all cases, need notion of user-user similarity or item-item similarity. • cosine distance or Jaccard distance are two possibilities.