Enhancing Knowledge-Aware Recommendation with Contrastive Learning

Abstract

Knowledge graph serves as a side information, bringing diversity and interpretability to the recommendation. A well-developed recommender system can efficiently capture user and item characteristics, accurately reflecting user preferences. However, supervised signals with graph structure are extraordinarily sparse, and the collaborative and knowledge graphs contain irrelevant edges, exacerbating noise propagation and reducing the robustness of recommendations. To address the above issues, we propose a model for enhancing Knowledge-aware Recommendation with Contrastive Learning (KRCL), including two contrastive learning tasks and three functional modules. Specifically, we construct two views, using TransR and TATEC to optimize knowledge representations from distance and semantic aspects, respectively. After the item-side knowledge is augmented, we remove unreliable interaction edges from collaborative graph to reduce noise propagation. We then perform contrastive learning on the output node representations of different views through graph propagation. To further tap the latent interest of users, we consider users/items that exhibit similar representations as semantic neighbors, treating them as positive pairs in contrastive learning. The structural and semantic contrastive tasks are eventually integrated in a multi-task learning manner to jointly boost the recommendation performance. To validate the effectiveness of our method, we conduct extensive experiments on three benchmark datasets. Experimental results demonstrate that our KRCL significantly outperforms previous state-of-the-art baselines.