ConDiff: Conditional graph diffusion model for recommendation

Currently, most existing graph diffusion models do not explicitly integrate key features of user collaboration signals and user–item (U–I) interaction graph in recommendation systems, limiting their ability to enhance recommendation performance. To alleviate this limitation, we propose a conditional graph diffusion model for recommendation, named ConDiff. Specifically, we introduce random Gaussian noise during the forward diffusion process to perturb the original graph structure. In the reverse generation process, we design an autoencoder for conditional graph generation, CGG-AE, which: (1) introduces personalized collaboration signals for each user online through logical operation; (2) utilizes user collaboration signals and U–I interaction information as conditional inputs to the diffusion model, obtain diffusion-collaboration and diffusion-interaction data in latent space through the encoder, and then use the decoder to reconstruct and generate higher-quality original U–I interaction information. Extensive experiments on three benchmark datasets demonstrate that ConDiff outperforms state-of-the-art models. Notably, on the Anime dataset, ConDiff improves Recall@10 and Recall@20 by 18.99% and 17.94%, reaching 0.2607 and 0.3721, respectively. The code is available at https://github.com/xl-wen/ConDiff.