In the field of recommendation systems, to address the problem of poor embedding representation quality and limited model generalization ability due to data sparsity, this paper proposes a sequence-attention-based graph convolutional model. In existing methods, there is still considerable room for improvement in recommendation effectiveness due to the neglect of high-order relationships between users and items and the capture of dynamic behaviors. The ACGCN model, through its designed MLP interest learning module, is able to better capture the sequential features in user behavior and dynamically adjust the weights of positive and negative samples through an attention mechanism, thereby enhancing the ability to capture user interests. In addition, the model enhances the understanding of user behavior through supervised contrastive learning, improving the generalization ability of the recommendation system. Validation on three public datasets shows that the ACGCN model outperforms current mainstream models in both Recall@20 and NDCG@20 metrics, demonstrating the effectiveness and performance advantages of this method.