FUMMER: A fine-grained self-supervised momentum distillation framework for multimodal recommendation

The considerable semantic information contained in multimodal data is increasingly appreciated by industry and academia. To effectively leverage multimodal information, existing multimodal recommendation methods mainly build multimodal auxiliary graphs to improve the representation of users and items. However, the weak value density of multimodal data inevitably leads to serious noise issues, making it difficult to effectively exploit valuable information from the multimodal contents. To address this issue, we propose a novel Fine-grained Self-supervised Mom entum Distillation Framework (FUMMER) for multimodal recommendations. Specifically, we propose a Transformer-based Fine-grained Feature Extractor (TFFE) and a Momentum Distillation (MoD) structure that incorporates intra- and inter-modal contrastive learning to fully pre-train TFFE for fine-grained feature extraction. Additionally, we design a structure-aware fine-grained contrastive learning module to fully exploit the self-supervised signals from fine-grained structural features. Extensive experiments on three real-world datasets show that our method outperforms state-of-the-art multimodal recommendation methods. Further experiments verify that the fine-grained feature extraction method we propose can serve as a pre-trained model, enhancing the performance of recommendation methods effectively by learning the fine-grained feature representations of items. The code is publicly available at https://github.com/BIAOBIAO12138/FUMMER.