MCA-TAN: Cross-Modal Graph Neural Recommendation with Temporal Behavior and Social Network Fusion

Abstract

This paper addresses the problem of how to integrate user behavior sequences and social network information to improve personalized recommendation. A recommendation model based on cross-modal graph neural networks is proposed. The method introduces a Modality-Coordinated Aggregation (MCA) mechanism to jointly model the user's behavior graph and social graph. This enables the model to capture the structural complementarity and semantic interaction between the two modalities. A Temporal-Aware Node (TAN) representation module is also introduced to fully extract temporal evolution patterns from behavior sequences, enhancing the model's ability to capture changes in user interests over time. The model first extracts high-order representations separately from the behavior graph and the social graph. Then, a modality gating mechanism is used to perform dynamic weighted fusion, generating a unified user representation vector. This fused representation is combined with the target item embedding and input into a matching function to produce the final recommendation results. Experiments on public datasets show that the proposed method outperforms several representative recommendation models across multiple mainstream evaluation metrics. It demonstrates notable advantages in precision, coverage, and ranking quality. Furthermore, ablation studies confirm the contribution of each module to performance gains. Cold-start analysis and modality weight sensitivity experiments further reveal the model's stability and adaptability. The experimental results strongly support the value of joint modeling of behavior and social graphs in addressing complex recommendation tasks. They also highlight the modeling capacity of graph neural architectures in multimodal recommendation scenarios.