Collaborative filtering on skewed datasets

Many real life datasets have skewed distributions of events when the probability of observing few events far exceeds the others. In this paper, we observed that in skewed datasets the state of the art collaborative filtering methods perform worse than a simple probabilistic model. Our test bench includes a real ad click stream dataset which is naturally skewed. The same conclusion is obtained even from the popular movie rating dataset when we pose a binary prediction problem of whether a user will give maximum rating to a movie or not.