Recommendation model based on generative adversarial network and social reconstruction

Abstract

Social relations can effectively alleviate the data sparsity problem in recommendation, but how to make effective use of social relations is a difficulty. However, the existing social recommendation models have two deficiencies. First, these models assume that social relations are applicable to various interaction scenarios, which does not match the reality. Second, it is believed that close friends in social space also have similar interests in interactive space and then indiscriminately adopt friends' opinions. To solve the above problems, this paper proposes a recommendation model based on generative adversarial network and social reconstruction (SRGAN). We propose a new adversarial framework to learn interactive data distribution. On the one hand, the generator selects friends who are similar to the user's personal preferences and considers the influence of friends on users from multiple angles to get their opinions. On the other hand, friends' opinions and users' personal preferences are distinguished by the discriminator. Then, the social reconstruction module is introduced to reconstruct the social network and constantly optimize the social relations of users, so that the social neighborhood can assist the recommendation effectively. Finally, the validity of our model is verified by experimental comparison with multiple social recommendation models on four datasets.

Keywords

• recommendation algorithm
• social recommendation
• generative adversarial network
• dynamic reconfiguration
• graph neural network