Social Recommendation Based on Hierarchical Attention Network

Abstract

Social recommendation is based on the traditional collaborative filtering, mining users' potential preferences in the recommendation system by integrating social information, alleviating the data sparsity problem of traditional recommendation, and improving the accuracy of recommendation. Based on the excellent performance of deep learning in feature extraction, many recommendation systems use the network framework of deep learning for feature extraction, and combine attention network and deep learning to model social recommendation. From the perspective of social information, this paper selects the hierarchical attention network model to learn users' static behavior preferences, uses the multi-head structure in the self-attention mechanism to capture the different effects of different social friends on users' behavior preferences, and optimizes the modeling effect of attention network on users' social influence. The simple weight division of the influence of ordinary attention networks on social friends in the early stage of modeling can not meet the calculation of the influence value of friends in the current complex social relationships. The model based on hierarchical attention networks can more accurately learn the social influence of users' friends on their social networks. A number of experiments have been carried out on the proposed model on the data sets Epinions and Gowalla. First, the proposed model is compared with the existing classical cutting-edge recommendation methods. The experimental results verify that the proposed model is superior to the comparison experiment in the evaluation indicators of different data sets.