Contrastive learning-enhanced personalized interaction dual tower network for recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Dual-tower retrieval models are efficient for large-scale recommendations but struggle with user behavior, personalization, and long-tail content.
• Existing models lack dynamic user preference tracking and personalized feature interactions.
• Poor representation learning affects performance with sparse data and niche items.

Purpose of the Study:

• To propose a novel framework, Contrastive Learning-Enhanced Personalized Interaction Dual Tower Network (CL-EPIDTN), to overcome limitations of current dual-tower models.
• To enhance the modeling of user behavior sequences and personalized inter-tower interactions.
• To improve representation learning for long-tail content and low-activity users.

Main Methods:

• Integrated a multi-layer Transformer to capture dynamic user preferences.
• Introduced a dual-path personalized enhancement mechanism for user-item feature dependencies.
• Employed a contrastive learning strategy to address data sparsity and improve representation learning.

Main Results:

• CL-EPIDTN demonstrated superior performance on Amazon Books and TmallData datasets compared to six state-of-the-art baselines.
• Achieved a Hit Rate@10 of 0.0351 and Recall@50 of 0.1123 on Amazon Books.
• Achieved a Hit Rate@10 of 0.0901 and Recall@50 of 0.1599 on TmallData.

Conclusions:

• CL-EPIDTN effectively addresses personalization and data sparsity challenges in recommender systems.
• The proposed framework shows significant potential for both academic research and practical applications.
• The integration of Transformer and contrastive learning enhances the scalability and efficiency of dual-tower models.