Abstract:Recommender systems become increasingly significant in solving the information explosion problem. Two typical kinds of techniques treat the recommendation problem as either a rating prediction or a ranking prediction one. In contrast, we propose a two-step framework that considers recommendation as a simulation of users’ behaviors to generate ratings. The first step is to predict the probability that a user rates an item, and the second step is to predict rating values. After that, the predicted results from both steps are combined to compute the expectations of users’ ratings on items, which are used to generate recommendations. Based on this framework, we propose a hybrid approach which uses topic model in the first step and matrix factorization in the second to solve the recommendation problem. Experiments with MovieLens and EachMovie datasets demonstrate the effectiveness of the proposed framework and the recommendation approach.

A hybrid approach of topic model and matrix factorization based on two-step recommendation framework