Research on Improved Methods for LightGCN Based on Diffusion Models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommendation systems have become crucial tools for mitigating information overload and enhancing user experience. Among them, LightGCN has achieved remarkable results in both efficiency and performance by simplifying graph convolution operations. However, in scenarios with sparse training data, LightGCN tends to produce homogeneous recommendation results, failing to reflect users’ genuine and diverse interest preferences. This study proposes an improved method for LightGCN based on the diffusion model, named Diffusion_LightGCN, which aims to enhance the performance of the LightGCN recommendation system by generating high-quality pseudo-user embeddings. Different from existing diffusion-based recommendation methods that use diffusion models to directly generate items, item distributions, or pseudo-users (e.g., DiffuRec, C-DiffRec, and CDiff4Rec), Diffusion_LightGCN generates pseudo-user embeddings that serve as auxiliary training signals for graph-based collaborative filtering, thereby organically coupling the generative capability of diffusion models with the propagation mechanism of graph neural networks. To address the issue of embedding quality degradation in traditional LightGCN under sparse data, this study combines the generative capability of diffusion models with the efficient propagation mechanism of graph neural networks. By integrating original embeddings with pseudo-embeddings through an embedding dynamic fusion mechanism and optimizing the loss function, the training effectiveness of LightGCN is ultimately improved. The results demonstrate that this method outperforms LightGCN in sparse data scenarios.