Bias-Aware Multi-Modal Graph Neural Network for Recommendation Systems

Existing multimodal recommendation systems generally consider information from multiple modalities, using pre-trained networks for initial feature extraction, followed by the application of graph neural networks (GNNs) for further feature propagation across nodes. The extracted features from each modality are then fused to generate the final recommendation score. However, this fusion approach often overlooks the global graph information and the impact of imbalanced modality information on the fusion process. While some fusion methods consider the heterogeneous information between modalities, they tend to neglect the internal variations within each modality. To address this, this paper proposes a Multimodal Dispersive Bias Alignment Network (MDBA), which integrates multimodal graph neural networks. The model separately considers the feature information of nodes at both the modality-level and within each modality to compute cross-modal bias and internal dispersion. By quantifying the importance of different modalities and intra-modality features through cross-modal bias and internal dispersion, the model achieves more fine-grained control over the fusion process.