General factorization framework for context-aware recommendations

General feature based solutions are emerging in the field of recommender systems with the increased need to incorporate multiple sources of information into a single model. The common property of these approaches is that they use a fixed factorization model class that can be extended to handle arbitrary number of context dimensions; the model is then learnt using different optimization techniques. In this paper we propose a general framework in which arbitrary linear feature models can be learnt efficiently. Moreover, both the factorization model and the model members are unrestricted in the framework, thus it is more flexible than state-of-the-art general feature based solutions. The framework allows for both implicit feedback based item ranking and rating prediction from explicit feedback. The paper focuses on the implicit feedback based recommendation problems, due to its expansive use in practical systems compared to explicit feedback. Using the flexibility of the framework we (1) evaluate various factorization models using 5 implicit feedback data sets and injecting contextual information into the model; (2) identify models that can solve the implicit feedback based context-aware recommendation task better than previously proposed model classes. Advantages and drawbacks of various models and learning strategies are also discussed briefly.