MCLE: Multimodal collaborative learning and enhancement with similarity-based modal purification for pretrained...

Multimodal recommendation has become a crucial framework for leveraging various data modalities to boost recommendation performance. However, existing methods still face several limitations when exploiting multimodal features. Relying solely on limited raw features prevents full exploitation of the rich semantic information inherent in various modalities, while unfiltered noise in multimodal data can degrade recommendation quality. Moreover, naive modal feature fusion strategies struggle to address the significant differences in preferences for different modalities among various users. To address these limitations, we propose a Multimodal Collaborative Learning and Enhancement framework (MCLE). Specifically, we utilize a pretrained language model to augment user profiles and item attributes, and propose an adaptive fusion mechanism that combines the enhanced user profile features with interaction-derived user representations, rather than simply relying on either source alone. A Similarity-based Modal Purification (SMP) module is then designed to filter out modal feature noise using learnable reference vectors that evaluate feature quality through similarity computation, producing purified multimodal representations. On this basis, a Multimodal Adaptive Feature Aggregator (MAFA) employs difference-aware attention and dynamic modal gating to achieve personalized multimodal feature integration. Extensive experiments on three public datasets demonstrate the effectiveness of our approach.