A recommendation-oriented diffusion model (S-Diff) that runs the forward/reverse diffusion process in the graph spectral domain and schedules noise anisotropically by graph frequency. Instead of corrupting a user's item-interaction vector with isotropic Gaussian noise (which collapses the signal-to-noise ratio and ignores preferences shared across similar users), the method builds an item-item graph Laplacian from the interaction matrix, maps each user vector into the spectral domain via the Graph Fourier Transform, and parameterizes the per-component noise scale by the corresponding graph eigenvalue. High-frequency components are aggressively noised while low-frequency components -- which encode global, collective user preferences -- are preserved, keeping the effective signal-to-noise ratio high. Motivated by an analogy between graph-convolution over-smoothing / graph heat diffusion and the forward corruption of diffusion models, a conditional denoising network conditioned on the user's history reverses the process, and the Graph Inverse Fourier Transform maps the recovered spectral signal back to item scores for top-K recommendation.
S-Diff is a diffusion model for collaborative-filtering recommendation that runs the diffusion process in the graph spectral domain rather than the raw item space. Standard denoising diffusion probab…