An efficient co-Attention Neural Network for Social Recommendation
Munan Li, Kenji Tei, Yoshiaki Fukazawa · IEEE/WIC/ACM International Conference on Web Intelligence · 2019
The recent boom in social networking services has prompted the research of recommendation systems. The basic assumption behind these works was that "users’ preference is similar to or influenced by their friends". Although many studies have attempted to use social relations to enhance recommendation system, they neglected that the heterogeneous nature of online social networks and the variations in users' trust of friends according to different items. As a natural symmetry between the latent preference vector of the user and her friends, we propose a new social recommendation method called ScAN (short for “co-Attention Neural Network for Social Recommendation”). ScAN is based on a co-attention neural network, which learns the influence value between the user and her friends from the historical data of the interaction between the user/her friends and an item. When the user interacts with different items, different attention weights are assigned to the user and her friends respectively, and the user's new latent preference feature is obtained through an aggregation strategy. To enhance the recommendation performance, a network embedding technique is utilized as a pre-training strategy to extract the users’ embedding and to incorporate the extracted factors into the neural network model. By conducting extensive experiments on three different real-world datasets, we demonstrate that our proposed method ScAN achieves a superior performance for all datasets compare with state-of-the-art baseline methods in social recommendation task.