《Learning Recommender Systems with Adaptive Regularization》.pdf

Learning Recommender Systems with Adaptive Regularization Steffen Rendle Social Network Analysis University of Konstanz 78457 Konstanz, Germany steffen.rendle@uni-konstanz.de ABSTRACT 1. INTRODUCTION Many factorization models like matrix or tensor factoriza- Recommender systems are an important tool with many tion have been proposed for the important application of applications e.g. online-shopping, video rental, personalized recommender systems. The success of such factorization web sites, etc. Recently factorization models became very models depends largely on the choice of good values for popular due to their success in several challenges including the regularization parameters. Without a careful selection the Netflix prize1. Most of the research in this field focuses they result in poor prediction quality as they either under- on new and improved models, among them are matrix fac- fit or overfit the data. Regularization values are typically torization [16, 12, 18], probabilistic latent semantic analysis determined by an expensive search that requires learning (PLSA) [6], SVD++ [7], SoRec [11], time-variant matrix fac- the model parameters several times: once for each tuple of torization [8] for collaborative filtering or Tucker decomposi- candidate values for the regularization parameters. In this tion