@article{LIU2025, 
author = {Huiting LIU and Shaoxiong LIU and Jiale WANG and Peng ZHAO},
title = {Deep Reinforcement Learning Recommendation Model Based on Multi-Interest Contrast},
year = {2025},
journal = {Journal of South China University of Technology (Natural Science Edition)},
volume = {53},
number = {9},
pages = {11-21},
keywords = {multi-interest, reinforcement learning, contrastive learning, heterogeneous information network},
url = {https://www.sciopen.com/article/10.12141/j.issn.1000-565X.240088},
doi = {10.12141/j.issn.1000-565X.240088},
abstract = {Deep Reinforcement Learning (DRL) is widely applied in recommender systems to dynamically model user interests and maximize cumulative user benefits. However, the sparsity of user feedback has become a significant challenge for DRL-based recommendation algorithms. Contrastive learning, as a self-supervised learning method, enhances user interest representation by constructing multiple perspectives, thereby alleviating the issue of sparse user feedback. Existing contrastive learning methods typically rely on heuristic-based augmentation strategies, which often lead to the loss of key information and fail to fully utilize heterogeneous interaction data. To address these issues, this paper proposed a multi-interest oriented contrastive deep reinforcement learning recommendation (MOCIR) model. The model consists of two key modules: a contrastive representation module and a policy network module. The contrastive representation module utilizes a Heterogeneous Information Network (HIN) to model the user’s local interests from different aspects while capturing their global interests based on raw interaction data. It then treats the global and local interests of the same user as positive pairs and those of different users as negative pairs for contrastive learning, effectively enhancing user interest representation. The policy network module aggregates user state representations and generates recommendations. The two modules are trained using an alternating update mechanism. Experimental results on three benchmark datasets show that the proposed model outperforms several DRL-based models in recommendation performance, effectively addressing the problem of sparse user feedback in recommendations.}
}