Recommendation Algorithm Based on Federated Multi-modal Learning
Chenyuan Feng, Zhenyu Feng, Qing Wang · 2024
Recommendation algorithm is a crucial tool for personalized information filtering. However, traditional recommendation algorithms typically need the collection of raw user data in a centralized server for analysis. Deploying the recommendation algorithm in a federated learning (FL) manner is a promising solution, which can train a good-quality model on the premise that the server does not directly access the sensitive user data, in light of the growing attention to privacy security. Nevertheless, the majority of recommendation algorithm researches has only considered uni-modality learning. To fill the gap, this paper designs a federated multi-modal RS. Firstly, an adaptive multi-modal federated recommendation algorithm is proposed to extract features from the item image, the commodity introductory text, and user historical ratings. The algorithm also address the issue of missing modalities via a modal completion mechanism. Additionally, an attention mechanism is designed to distinguish the importance of various modal data on model training, wherein different weight ratios are assigned to various features to realize an adaptive multi-modal RS. Furthermore, one multi-modal dataset is established to verify the efficacy of the proposed algorithm. The simulation results show that our proposed measures contribute a reliable and high-quality recommendation model.