From Uniform to Learned Graph Priors: Diffusion for Structure Discovery

The article introduces Diff-prior, a novel diffusion-parameterized adaptive prior that enhances neural relational inference (NRI) methods by calibrating latent graph distributions rather than generating graphs. This approach addresses the limitations of traditional uniform priors by refining edge posteriors through a learnable denoising calibration process, leading to improved structural inference and more decisive edge posteriors across various NRI architectures. The results from standard benchmarks demonstrate the effectiveness of Diff-prior, which is significant for practitioners aiming to achieve more reliable graph structure discovery in dynamic systems.