Multi-Intent Learning for Multi-Modal Recommendation

Recommendation Systems play an increasingly important role in social platforms, providing users with personalized content services. With the explosive growth of multi-modal contents, traditional recommendation systems based solely on interaction data can no longer meet user needs. In recent years, many researchers have attempted to integrate multi-modal features into recommendation system. Meanwhile, user intent determines whether users will interact with items or not. Intent modeling is crucial for capturing user preferences and explaining recommendation results. Recommendation system based on interaction data model user intents at a coarse-grained level, because the information contained in user IDs and item IDs is too limited. In contrast, multi-modal features can reflect finer-grained user preferences from a semantic perspective. To address this issue, we propose a Multi-Intent Learning method for Multi-Modal Recommendation (MIMM). Specifically, we design a Modality-Specific Intent Graph Convolution Network to disentangle user and item intents from different modalities and filter out self-supervision signals at the semantic level. Finally, a cross-modality intent contrastive learning task is introduced to align user intents across different modality. Experiments on real-world dataset demonstrate the effectiveness of our MIMM.