DAKRec: Generative Denoising and Adaptive Embedding for Knowledge Graph-Based Recommendation

The exponential growth of data has revolutionized recommendation systems (RS), highlighting their critical role in diverse applications. Although the use of neural networks and graph neural networks (GNNs) has improved collaborative filtering (CF), challenges like data sparsity and noisy connections in knowledge graphs (KGs) still exist. In response to these difficulties, we set up DAKRec, a novel knowledge graphbased recommendation model that integrates adaptive embedding transformations and a generative denoising approach. DAKRec employs Adaptive Relation-aware Embedding Transformation (ARET) to capture the varying importance of relations within the KG, enhancing user-item representations. Additionally, it introduces a Dynamic Noise Scaling Strategy (DNSS) during the diffusion process, dynamically adjusting noise levels based on training progress to balance exploration and convergence. Empirical studies on real-world datasets including Amazon Books and Yelp2018, show that DAKRec consistently outperforms state-of-the-art methods. The findings from the ablation trial provide more evidence of the efficacy of ARET and DNSS, with significant performance gains attributed to each. These results highlight DAKRec's robustness and adaptability across datasets with varying sparsity and noise levels. Moreover, the newly introduced user-item level contrastive learning module enhances embedding representations in sparse interaction scenarios, further showcasing the model's robustness and generalizability. By seamlessly integrating diffusion processes with collaborative filtering and knowledge graph information, DAKRec provides an innovative and efficient solution to long-standing RS challenges, achieving superior performance and robustness.