Recommendation from raw data with adaptive compound Poisson factorization

Abstract

Count data are often used in recommender sys-tems: they are widespread (song play counts,product purchases, clicks on web pages) andcan reveal user preference without any explicitrating from the user. Such data are known to besparse, over-dispersed and bursty, which makestheir direct use in recommender systems chal-lenging, often leading to pre-processing stepssuch as binarization. The aim of this paper isto build recommender systems from these rawdata, by means of the recently proposed com-pound Poisson Factorization (cPF). The papercontributions are three-fold: we present a uni-fied framework for discrete data (dcPF), lead-ing to an adaptive and scalable algorithm; weshow that our framework achieves a trade-offbetween Poisson Factorization (PF) applied toraw and binarized data; we study four specificinstances that are relevant to recommendationand exhibit new links with combinatorics. Ex-periments with three different datasets showthat dcPF is able to effectively adjust to over-dispersion, leading to better recommendationscores when compared with PF on either rawor binarized data.