Multi-modal recommendation algorithm fusing visual and textual features.

Area of Science:

• Artificial Intelligence
• Computer Science

Background:

• Recommender systems face challenges with data sparsity and cold starts due to limited user-item interaction data.
• Multi-modal features (image, text) improve recommendations but existing methods overlook user interaction sequences and use simplistic feature aggregation.

Purpose of the Study:

• To propose the Fusing Visual and Textual Features (FVTF) algorithm to enhance recommender systems.
• To address limitations in current multi-modal recommendation approaches by considering user interaction sequences and adaptive feature importance.

Main Methods:

• Designed a user history visual preference extraction module using Query-Key-Value attention for visual feature-based interest modeling.
• Developed a feature fusion and interaction module employing multi-head bit-wise attention for adaptive mining of feature combinations and higher-order representations.

Main Results:

• The FVTF algorithm demonstrated superior performance compared to benchmark recommendation algorithms.
• Experiments conducted on the Movielens-1M dataset validated the effectiveness of the proposed FVTF approach.

Conclusions:

• The FVTF algorithm effectively models user interests by integrating visual and textual features through advanced attention mechanisms.
• The proposed method offers a significant improvement in recommendation accuracy by overcoming limitations of existing multi-modal approaches.