A Dynamic Collaborative Recommendation Method Based on Multimodal Fusion

Traditional Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) struggle with semantic comprehension and long-distance dependency capture in recommendation systems. To address this, we propose MBTRec, a multimodal recommendation model based on the Transformer encoder. It employs an innovative bidirectional tower-type attention mechanism (Bi Towernet) for modal fusion, ensuring the independent contribution of each modality while optimizing interaction and feature representation. MBTRec integrates forgetting functions and StreamLDA techniques to capture users’ dynamic interest topics and uses Deep Canonical Correlation Analysis (DCCA) to explore the correlation between topics and multimodal information. Through dense incremental dynamic time windows, MBTRec captures users’ latest preferences and leverages the Transformer model to predict recommendation outcomes.