Deep Probabilistic Matrix Factorization Framework for Online Collaborative Filtering

As living data growing and evolving rapidly, traditional machine learning algorithms are hard to update models when dealing with new training data. When new data arrives, traditional collaborative filtering methods have to train their model from scratch. It is expensive for them to retrain a model and update their parameters. Compared with traditional collaborative filtering, the online collaborative filtering is effective to update the models instantly when new data arrives. But the cold start and data sparsity remain major problems for online collaborative filtering. In this paper, we try to utilize the convolutional neural network to extract user/item features from user/item side information to address these problems. First, we proposed a deep bias probabilistic matrix factorization (DBPMF) model by utilizing the convolutional neural network to extract latent user/item features and adding the bias into probabilistic matrix factorization to track user rating behavior and item popularity. Second, we constrain user-specific and item-specific feature vectors to further improve the performance of the DBPMF. Third, we update two models by an online learning algorithm. The extensive experiments for three datasets (MovieLens100K, MovieLens1M, and HetRec2011) show that our methods have a better performance than baseline approaches.