NLGCL+: Naturally Existing Neighbour Layers Graph Contrastive Learning with Adaptive Sample Weighting for Multimodal Recommendation

While Graph Neural Networks (GNNs) have demonstrated remarkable success in multimodal recommendation systems by capturing high-order user-item relationships, their performance is often hindered by the inherent data sparsity in real-world scenarios. Although Graph Contrastive Learning (GCL) has emerged as a promising solution to enhance data representations, most existing methods rely on computationally intensive augmentation strategies, which risk introducing semantically irrelevant noise. To address these limitations, we propose NLGCL, a novel and efficient contrastive learning framework that leverages the intrinsic structural properties of GNNs by constructing positive contrastive views from naturally related neighboring layers, thereby eliminating the need for external data augmentations and their associated computational overhead. With the proliferation of multimedia data, multimodal recommendation has become increasingly prevalent. However, directly applying NLGCL to multimodal settings overlooks the rich information embedded in modality-specific features. To this end, we further extend NLGCL to NLGCL+, a tailored framework specifically designed for multimodal recommendation. NLGCL+ uniquely integrates multimodal information to perform adaptive sample weighting, enabling more discriminative and fine-grained representation learning. Designed as a plug-and-play module, NLGCL+ can be seamlessly integrated into existing multimodal recommendation models to enhance their accuracy. Comprehensive experiments on widely-used multimodal recommendation datasets demonstrate that NLGCL+ significantly outperforms state-of-the-art baselines in both effectiveness and efficiency. Code can be found in https://github.com/Jinfeng-Xu/NLGCL-Plus.