Re-considering neighborhood-based collaborative filtering parameters in the context of new data
Adele E. Howe, Ryan D. Forbes · 2008
The Movielens dataset and the Herlocker et al. study of 1999 have been very influential in collaborative filtering. Yet, the age of both invites re-examining their applicability. We use Netflix challenge data to re-visit the prior results. In particular, we re-evaluate the parameters of Herlocker et al.'s method on two critical factors: measuring similarity between users and normalizing the ratings of the users. We find that normalization plays a significant role and that Pearson Correlation is not necessarily the best similarity metric.