Abstract: Existing recommendation systems cannot distinguish social influence from potential interest well, and ignore the graph structure characteristics and changes of social networks. In view of the above deficiencies, a social recommendation system based on graph attention adversarial network (GAASR) is proposed. Social influence and potential interest were separated by adversarial network. Hadamard projection method was used to obtain the values of context weight. The graph attention network was used to learn the potential vector of social embedding and capture the social structure of users more accurately. In order to verify the performance of the recommendation system, three recommendation system data sets were used for analysis experiments. The experimental results show that GAASR is better than currently popular recommendation methods, which can effectively improve the recommendation accuracy.