DICES: Diffusion-Based Contrastive Learning with Knowledge Graphs for Recommendation

Hao Dong, Haochen Liang, Jing Yu, Keke Gai

Abstract: The effectiveness of Knowledge Graphs (KGs) in enhancing recommendation systems has been recognized. However, the effectiveness of KG-enhanced recommendations is often hampered by issues of entity sparsity and noise. To address these challenges, we propose a Diffusion-based Contrastive Learning with Knowledge Graphs for Recommendation (DICES). Our method combines diffusion models with multi-level contrastive learning approaches, aiming to enhance the performance of existing recommendation systems. By utilizing diffusion models, we ensure that the generated augmented samples are context-aware, thereby increasing the robustness of contrastive learning. Additionally, we introduce a multi-level contrastive learning approach to improve recommendation accuracy. Finally, we design a joint training framework to optimize both the recommendation task and the multi-level contrastive learning tasks, further enhancing the overall effectiveness of the recommendation system. Extensive experiments on multiple benchmark datasets demonstrate that our DICES framework significantly outperforms existing state-of-the-art methods in scenarios with sparse user-item interactions and noisy KG data.