PEARL: A dual-layer graph learning for multimodal recommendation

Multimodal recommendation has increasingly become a mainstream information service technology, which transcends traditional user–item interaction-based recommender systems by integrating the multimodal features of items. Although existing works have made notable progress by focusing on user–item interaction graph structures and self-supervised learning to enhance multimodal representation learning , they still exhibit the following two limitations: (1) Performing graph convolution operations on a fixed interaction graph introduces misleading noisy signals caused by user’s imbalanced attention on various modalities. (2) Lacking exploration of inherent self-supervised signals in multimodal attributes fails to mitigate the distribution bias introduced during data augmentation . To address these issues, we propose a novel method named P urified-int E raction and A ffinity g R aph L earning ( PEARL ) for multimodal recommendation, which utilizes a dual-layer graph learning to model the user preference. Specifically, to eliminate misleading noisy signals, we designed a graph purification strategy that constructs purified modality-specific interaction graphs, thereby removing noisy edges from the raw interaction graph. Then, the user and item affinity graphs are constructed based on the user co-occurrences and the item multimodal features, respectively, which are then utilized for message passing to mine the implicit self-supervised signals between similar users or items. After that, we propose an augmentation-free contrastive learning task in the fusion module to improve the quality of ID embeddings and multimodal features, ultimately generating the final representations of users and items. Comprehensive experimental results on three publicly available datasets verify the superiority of PEARL against diverse state-of-the-art baselines, where the improvements are especially noticeable in large-scale datasets. Our experimental code and data utilized in this work can be accessed at https://github.com/Yuzhuo-Dang/PEARL . • Enhancing multimodal recommendation through a dual-layer graph learning framework. • Eliminating the noise caused by the user’s imbalanced attention for the first time. • Mining the intra-modal self-supervised signals through the affinity graph learning. • Introducing an augmentation-free contrastive learning task based on modal attributes. • Sufficient experimental results on three datasets verify the superiority of PEARL.