Multi-Modal Semantics Enhanced Graph Convolutional Network for Session-Based Recommendation

Session-based recommendation (SBR) predicts the next item for anonymous users. While multi-modal data has been used to alleviate sparsity, existing methods overlook intra-modal semantic correlations and underemphasize the last interaction, which signals current preference. We propose a multi-modal semantics enhanced graph convolutional network (M2SEGCN). This model constructs modal graphs to capture intra-modal item semantics and employs a dynamic weighting mechanism that prioritizes the last interaction and sequential dependencies. Experiments on three real datasets show M2SEGCN outperforms twelve baselines. The datasets and codes are available at https://github.com/ZZY-GraphMiningLab/M2SEGCN.