Abstract: Aimed at the noisy problem of explicit social relationships and the problem that most social recommendation algorithms ignore the dynamic changes between friends, a social recommendation algorithm (AGCN) fused to generate adversarial graph convolutional networks is proposed. The algorithm built the user's potential friend relationship based on rating information and explicit social relationships, and used a streamlined and efficient graph convolutional neural network to learn the structural features of the information to obtain the deep-level features of users and products. It used a generative confrontation network to dynamically build trusted friends with the same preferences as the user, punished false friends, and achieved dynamic changes in friends. The results on the Filmtrust and Ciao datasets show that compared with the BPR, SBPR, CUNE-BPR, and LightGCN algorithms, this algorithm achieves the best recommendation performance for both ordinary users and cold-start users.