Collaborative Filtering for Recommender Systems

Collaborative filtering (CF) predicts user preferences in item selection based on the known user ratings of items. As one of the most common approach to recommender systems, CF has been proved to be effective for solving the information overload problem. CF can be divided into two main branches: memory-based and model-based. Most of the present researches improve the accuracy of Memory-based algorithms only by improving the similarity measures. But few researches focused on the prediction score models which we believe are more important than the similarity measures. The most well-known algorithm to model-based is the matrix factorization. Compared to the memory-based algorithms, matrix factorization algorithm generally has higher accuracy. However, the matrix factorization may fall into local optimum in the learning process which leads to inadequate learning. CF approaches are usually designed to provide products to potential customers. Therefore the accuracy of the methods is crucial. In this paper, we propose various solutions to make a quality recommendation. First, we proposed a new prediction score model for the Memory-based method. Second, we proposed a differential model that considers the adjustment process after the training process in the existing matrix factorization methods. Third, a novel hybrid collaborative filtering is outlined to avoid or compensate for the shortcomings of matrix factorization and neighbor-based methods. In the end, we performed the experiments on Movie Lens datasets and the results confirmed the effectiveness of our methods.