A latent factor model of fusing social regularization term and item regularization term

Abstract

As the social media boom and evolve, the studies of recommender system are facing increasing challenges, especially in solving the sparse data issue. The “latent factor model” (LFM), as a typical recommendation algorithm, is widely used in various recommender systems. However, LFM shows low recommendation accuracy in sparse data situation. This paper proposes a modified LFM model that fuse the information from both users’ social relationship and items’ information. To illustrate the social connections among users, we introduce the social regularization term, which is calculated from the users’ social information, into the traditional LFM. To process the item information, we use the Latent Semantic Index (LSI), which effectively constrains the loss function in the matrix decomposition, to generate item regularization. The model parameters are optimized through gradient descent. In order to verify its effectiveness, we made experiments using the Last.fm dataset, KDDcup_Track1 dataset and crawling data from Douban respectively. The experimental result shows that, our approach can achieves higher accuracy than the traditional recommendation algorithm.