Research On Deep Social Collaborative Filtering

With the rapid development of the Internet, the data size that people contact grows explosively. As an effective solution to the problem of "information overload", the recommendation system plays an important role in people's life increasingly. Among them, without the effects from the content of the recommended items, and easy to realize the personalized recommendation, collaborative filtering has become the most popular and successful recommendation algorithm. But there are some inherent problems, such as data sparsity, scalability, cold start and so on, which seriously affect the quality of the recommendation.In order to alleviate the aforementioned problems, people began to deal them with auxiliary information in the rating information. With the rapid development of social media, there are a lot of social relations, which is very easy to be obtained. Usually, people are more willing to accept the views from their acquaintances(such as partners, friends, relatives, colleagues, etc.). This means that there is a potential social network among users. In recent years, more and more researchers have committed themselves to integrate social information into the recommendation system, and put forward the social recommender systems. How to effectively use the ratings and social information is the main problem to be solved. Most of the existing social recommendation systems use the social relations information among users to linearly constrain the user's interest preferences through the techniques of regularization, integration, matrix decomposition and so on, so that the target user's interest preferences are similar to the interest preferences of friends, thus to improve the recommended quality.People still ask for higher quality of the recommendation systems. The non-linear relationship between rating information and social relations information is also worth exploring. In this paper, DSCF model which is a new social collaborative filtering, based on matrix decomposition, is proposed. This model uses SDAE(Stacked Denoising Autoencoders) effectively to bridge the rating information and trust information and to make the two information flows in the two-way flow, to fully exploit the user's true preference, so as to improve the quality of recommendation. And the probability generating process of DSCF model is given. At the same time, in our model, it also explains how the spread of trust affects the user's true preference with nonlinearity. On three different real datasets, it is proved that the proposed model is better than the existing collaborative filtering model, which significantly improves the quality of recommendation.