DynKGCL: Contrastive learning for recommendation with dynamic dual-channel positive expansion and adaptive negative sampling

Abstract

Data sparsity and cold start problems remain critical challenges in the field of recommendation systems, significantly restricting the predictive accuracy and generalization capability of user preference modeling. To overcome this bottleneck, this paper proposes DynKGCL, a contrastive learning(CL)-based recommendation framework that integrates dynamic dual-channel positive expansion and adaptive negative sampling to enhance recommendation performance. In terms of positive sample generation, the proposed method dynamically adjusts the cross-user similarity propagation ratio based on the sparsity of user interactions and incorporates a dual-channel positive expansion mechanism along with knowledge graph(KG)-enhanced neighborhood mining to effectively expand the positive sample pool in sparse scenarios, thereby improving user-item representation learning. For negative sample selection, we employ a dataset-adaptive strategy, utilizing a hybrid negative sampling approach in relatively dense datasets and pure random sampling in sparse datasets to balance sample diversity and model generalization. Experimental results demonstrate that DynKGCL achieves state-of-the-art performance across multiple benchmark test sets. Theoretical analysis further confirms that the proposed method, through graph-enhanced representation learning and a unified optimization paradigm, effectively alleviates the data sparsity problem, significantly enhances the robustness and generalization ability of the recommendation system, and provides a reliable solution for personalized recommendations.