RKGR: Robust Knowledge Graph-Based Recommendation Method Using Debiased Contrastive Learning

Noisy and long-tailed user behavior in real scenarios is a common challenge for recommendation systems, especially in knowledge graph-based systems with hierarchical levels. In this paper, we propose a robust knowledge graph-based recommendation method (RKGR) using debiased contrastive learning, which can effectively resist noise even across the original and subgraphs of the user-item interaction graph. To improve the system’s capacity to resist disturbances in nodes and edges with different link modes (such as user-user, user-item, and entity-entity), we integrate projection heads with shuffling batch normalization in a multi-perspective manner. Additionally, we propose a uniform Gaussian random noise-based approach for false negative filtering through debiased contrastive learning. This approach addresses negative sampling bias while improving the alignment and uniformity of the representation space. To further enhance RKGR’s robustness, we introduce an adversarial learning strategy for complex knowledge graphs. Extensive experiments on three benchmark datasets demonstrate that RKGR outperforms existing state-of-the-art methods in terms of recommendation accuracy and robustness to noise, making it an effective solution for real-world recommendation systems with noisy and long-tailed data. We release the codes and datasets at https://github.com/yuzilan/RKGR.