Research map / Image generation / Fast diffusion sampling

Bias Mitigation in Graph Diffusion Models

ICLR 2025 · Image generation › Fast diffusion sampling · approach: Diffusion theory and design space · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes S++, a training-free sampling correction for graph diffusion models that uses Langevin sampling to start reverse sampling from the forward maximum-perturbation distribution and a score-difference correction against a pseudo score network to mitigate reverse-starting and exposure biases.

Problem

Graph diffusion models truncate the forward process so the maximum perturbation distribution deviates from the standard Gaussian, yet reverse sampling starts from the standard Gaussian (reverse-starting bias), and they also suffer from exposure bias, degrading generation quality.

Approach

Align the reverse starting point with the actual forward maximum-perturbation distribution via Langevin sampling using the score at time T, and correct scores with a score difference between the original network and a pseudo network trained on the model's own generated samples, without modifying architectures.

Key results

• On Community-small with GDSS-OC, S++ reduces average MMD from 0.064 to 0.029 (Deg/Clus/Orbit all improved).
• On QM9 with GDSS-OC, S++ improves FCD from 4.584 to 1.661 and NSPDK MMD from 0.016 to 0.001 while reducing sampling time from 0.73e2 to 5.10 (units as reported).
• MOOD-S++ improves novel hit ratio over MOOD on all five docking targets (e.g., parp1 7.017% to 8.286%).

Measured effects reported in the paper

Surprising findings

• Langevin-based reverse-starting alignment combined with correction reduced sampling time substantially on QM9 (GDSS-OC-S++ sampling time reported as 5.10 vs 0.73e2), which the authors attribute to shortened sampling chains.
• Some S++ variants show decreased validity on ZINC250k (GDSS-OC validity 41.84% vs S++ 59.50% is an improvement, but HGDM-WC-S++ and GDSS-WC-S++ show mixed validity changes across tables).

Limitations

• Performance sensitive to hyperparameters (lambda, omega, step size eta, cut-off time t_c) tuned per model and dataset, as shown in appendix tables.
• Requires training an additional pseudo score network on generated samples, which adds cost.
• Some reported results are mixed across tables (e.g., inconsistent FCD and validity numbers for the same models between main and appendix tables).

Experimental setup

Datasets: Community-small Enzymes Grid QM9 ZINC250k BRENDA

Benchmarks: Community-small Enzymes Grid QM9 ZINC250k Novel hit ratio on parp1/fa7/5ht1b/braf/jak2 docking tasks

Metrics: MMD degree MMD clustering MMD orbit validity w/o correction NSPDK MMD FCD sampling time novel hit ratio novel top 5% docking score

Baselines: GDSS HGDM GSDM MOOD GraphRNN GraphAF GraphDF GraphVAE EDP-GNN MoFlow GraphEBM DeepGMG

Builds on: GDSS Langevin sampling classifier-free guidance DPM-Solver-style noise schedule analysis (DPM-Fixes, DPM-Leak) ADM-ES TS-DPM

Closest papers by meaning

• Elucidating the Exposure Bias in Diffusion Models (ICLR 2024 · similarity 0.89)
• Pard: Permutation-Invariant Autoregressive Diffusion for Graph Generation (NeurIPS 2024 · similarity 0.88)
• When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors (NeurIPS 2026 · similarity 0.87)
• Simple and Critical Iterative Denoising: A Recasting of Discrete Diffusion in Graph Generation (ICML 2025 · similarity 0.87)
• Is Noise Conditioning Necessary? A Unified Theory of Unconditional Graph Diffusion Models (NeurIPS 2025 · similarity 0.87)
• Generative Modelling of Structurally Constrained Graphs (NeurIPS 2024 · similarity 0.86)