Dynamic Multi-Graph Convolutional Collaborative Filtering Recommendation Method

To address the limitation that most existing Graph Convolutional Network (GCN)-based recommendation methods rely solely on a single user-item interaction graph, thereby failing to capture users’ complex behavioral interests and resulting in suboptimal personalized recommendation performance, this study proposes a Dynamic Multi-Graph Convolutional Collaborative Filtering method (DMGCCF). To enhance the ability to capture node relationships, the method introduces a node-degree-informed structural similarity graph and employs Graph Convolutional Networks (GCNs) along with weighted aggregation to obtain user and item embedding vectors from both the original adjacency matrix and that of the structural similarity graph. To further reduce noise from irrelevant neighbors, the method incorporates an adaptive neighbor-selection strategy based on a dynamic window mechanism, enabling the precise extraction of item communities and the generation of high-order item embeddings. This method is not a standalone GCN model but an auxiliary technique that can be integrated into GCNs and related architectures, offering broad applicability. Evaluations on multiple public datasets demonstrate that the proposed method significantly enhances the performance of GCN backbone models, achieving superior prediction accuracy and ranking quality compared with traditional approaches.