Multi-Frequency Filter Based Denoising Diffusion Model for Knowledge-Aware Recommendation

Knowledge Graph (KG) introduces fruitful domain information of structure and semantic to solve data sparsity. However, the presence of irrelevant entities as noise results in the inability to predict user true preference. Recently, diffusion models have been broadly applied to remove noise. In addition, filtering algorithms are current effective methods for filtering out noise signals. To fundamentally eliminate noise in KG, we propose a remarkable multi-frequency filter based denoising diffusion model (FDiff) for knowledge-aware recommendation. In our framework, we devise a diffusion with multi-frequency filter to generate a denoised KG. Specifically, the reverse process incorporates a multi-frequency filter, which divides frequencies with different weights to eliminate irrelevant signals. Meanwhile, the top k optimal values are returned to represent the knowledge data. For prediction results, the aggregation strategy consists of knowledge aggregation and collaborative knowledge graph aggregation to obtain user personalized preferences. Furthermore, we utilize a contrastive learning method to optimize the recommendation task. Extensive experiments are performed on three realistic datasets. The results consistently prove that our FDiff surpasses other competitive baselines.