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

## TL;DR

This paper introduces a new recommendation system model that improves personalization and handles rare items and users better than existing methods.

## Contribution

A novel dual-tower framework with contrastive learning and personalized interaction mechanisms for recommendation systems.

## Key findings

- CL-EPIDTN outperforms six state-of-the-art models on Amazon Books and TmallData datasets.
- The model achieves higher Hit Rate@10 and Recall@50 metrics, showing better performance in sparse data scenarios.
- The proposed framework effectively captures dynamic user preferences and strengthens user–item feature dependencies.

## Abstract

Dual-tower retrieval models have become a prevalent solution in large-scale recommendation systems due to their scalability and deployment efficiency. However, they face critical limitations including insufficient modeling of user behavior sequences, lack of personalized inter-tower interactions, and poor representation learning for long-tail content. To address these issues, we propose a novel framework called Contrastive Learning-Enhanced Personalized Interaction Dual Tower Network (CL-EPIDTN). This model integrates a multi-layer Transformer to capture dynamic user preference shifts, and introduces a dual-path personalized enhancement mechanism to strengthen user–item feature dependencies. Additionally, a contrastive learning strategy is employed to enhance the representation learning of long-tail items and low-activity users under sparse data conditions. Extensive experiments on two public datasets (Amazon Books and TmallData) demonstrate the effectiveness of our method. CL-EPIDTN achieves the best performance across multiple metrics, with Hit Rate@10 of 0.0351 and Recall@50 of 0.1123 on Amazon Books, and Hit Rate@10 of 0.0901 and Recall@50 of 0.1599 on TmallData, outperforming six state-of-the-art baselines. These results highlight the potential of CL-EPIDTN for both academic research and practical deployment in real-world recommender systems, particularly in handling personalization and data sparsity challenges.