Time-Sensitive Collaborative Filtering via Item-Item Similarities

• Activity Number: 491
• Type: Invited
• Date/Time: Wednesday, August 3, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Rahul Mazumder*+ and Deepak Agarwal and Liang Zhang
• Companies: Stanford University and Yahoo Research and Yahoo Research
• Keywords: collaborative filtering ; covariance regularization ; sparsity (L1 regularization) ; large scale optimization ; expectation maximization ; web / internet applications

Abstract

Consider a generic collaborative system problem with users rating (clicking) items over time --- with potentially a large fraction of the user-item matrix unobserved. We will address how to come up with predictions/ recommendations for the unobserved entries in the matrix by incorporating item-item similarity in a model-based fashion. This novel proposal takes us well beyond the typical low-rank matrix factorization models used in collaborative filtering. Various notions of item-item similarities can be used via a prior regularization eg the sparse inverse covariance regularization accounting for interpretable conditional independencies among the items resulting in a sparse item-item similarity graph and sophisticated variants thereof. We propose computationally tractable, theoretically justified, novel algorithms that enable us to learn these complex models for large-scale recommender systems, with potentially billions of variables. We will also discuss how adopt our framework to account for dynamic user /item visit patterns and time sensitive user-item interactions, as arsing in Yahoo! front page and News applications.