Feature-Based Matrix Factorization

Recommender systems that recommend items according to users’potential inter-est have become more and more important due to the rapid growth of information in the Web. Collaborative filtering is one of the most promising techniques for recommender systems. In real world scenarios, many types of auxiliary information are available besides user rating matrix. One of the challenges of collaborative filtering is how to utilize them to improve prediction accuracy.In this paper, we introduce feature-based factorization, a general framework for collaborative filtering with auxiliary information. The feature-based setting allows us to build factorization models incorporating side information such as temporal dynam-ics, neighborhood relationship, and hierarchical information. Based on the model, we build an efficient toolkit to solve large-scale rate prediction and collaborative ranking problem. Using the framework, we build specific models on Yahoo! Music dataset utilizing time, taxonomical and neighborhood information available in the data. Our experiment results show our proposed method achieves the state of art result on the dataset.We further study the problem of feature learning. We propose to solve the problem via general functional matrix factorization, whose model includes conventional matrix factorization models as its special cases. Moreover, we propose a gradient boosting based algorithm to efficiently solve the optimization problem. Finally, we give two specific algorithms for efficient feature function construction for two specific tasks. Our method can construct more suitable feature functions by searching in an infinite functional space based on training data and thus can yield better prediction accuracy.