Abstract: In recent years, multimodal recommendation has shown outstanding performance by integrating user behavior and content information. When combined with Graph Neural Network (GNN), it exhibits stronger information propagation capability in modeling complex interactions. To further improve representation ability and recommendation quality, a Greedy Shapley Game (GSG) for Multimodal Recommendation algorithm Greedy Shapley Game for multimodal Recommendation (GSGRec) was proposed. The algorithm was designed based on a local-global GNN, in which global information and multimodal fusion were optimized through competitive and cooperative mechanisms. First, a Game-based Personalized PageRank (GPPR) module was introduced. In this module multi-hop connections at different scales within the graph network were utilized, and a ranking-based selection mechanism was applied to stimulate competition among global information of varying granularity. A multi-scale weighted fusion strategy was subsequently adopted to enhance high-contribution scales while suppressing noisy ones, resulting in a more accurate and robust global semantic representation. Next, an auxiliary reward loss function for multimodal data was defined through the GSG mechanism, leading to competitive relationships among modalities. Meanwhile, the Mutual Information (MI) between modalities was minimized to promote modality complementarity, enabling the model to adaptively identify key modalities and avoid redundant interference, thus producing more expressive and discriminative multimodal representations. Experiments on the Amazon-Baby, Sports, and Clothing datasets demonstrate that the proposed model outperforms existing multimodal GNN-based recommendation methods in terms of Recall and Normalized Discounted Cumulative Gain (NDCG), achieving improvements of 5.90%, 3.63%, and 2.36% in Recall@20 with the MIG-GT (Modality-Independent Graph neural networks with Global Transformers) algorithm. This result further indicates that the model has stronger ranking capability and generalization performance in multimodal recommendation tasks.

Key words: Keywords: multimodal recommendation, Greedy Shapley Game (GSG), Game-based Personalized PageRank (GPPR), Graph Neural Network (GNN)

摘要： 近年来，多模态推荐因融合用户行为与内容信息表现突出，与图神经网络(GNN)结合后在建模复杂交互中更具信息传播能力。为进一步提升表征能力与推荐质量，提出一种贪婪Shapley博弈(GSG)的多模态推荐算法(Greedy Shapley Game for multimodal Recommendation, GSGRec)，该算法基于局部-全局GNN，采用竞争与协作的方式优化全局信息和多模态融合。首先，提出博弈式PPR(Game-based Personalized PageRank, GPPR)模块，基于图网络中不同尺度的多跳连接，利用排序筛选机制激发不同粒度全局信息的竞争，并通过多尺度的加权融合强化高贡献尺度、抑制噪声尺度，从而得到更精确且鲁棒的全局语义表征。其次，通过GSG定义多模态的辅助奖励损失函数，使不同模态间产生权重竞争关系，同时通过最小化模态间互信息(Mutual Information,MI)促进模态之间互补，使模型能够自适应识别关键模态并避免冗余干扰，从而获得更具表达力和判别力的多模态表示。在Amazon-Baby、Sports和Clothing这3个数据集上的实验结果表明，本文模型在召回(Recall)和归一化折损累计增益(NDCG)指标上普遍优于现有的多模态GNN推荐方法，与MIG-GT(Modality-Independent Graph neural networks with Global Transformers)算法相比，在Recall@20上3个数据集性能分别提升了5.90%、3.63%和2.36%。这进一步表明，所提模型在多模态推荐任务中具备更强的排序能力与泛化性能。

关键词: 多模态推荐, 贪婪Shapley博弈, 博弈式PPR, 图神经网络