Research map / Graph learning / Graph self-supervised learning
Is Noise Conditioning Necessary? A Unified Theory of Unconditional Graph Diffusion Models
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.
Argues and empirically tests that graph diffusion models (GDSS, DiGress) can drop explicit timestep conditioning, backed by scaling theory (EFPC, ETDB, MDEP) for Bernoulli edge-flip and coupled Gaussian noise.
Problem
Graph diffusion models are assumed to need explicit noise-level/timestep conditioning; the paper asks whether denoisers can infer the noise level implicitly from corrupted graph structure, making conditioning unnecessary.
Approach
Proves via Bayesian posterior concentration that the number of potential edges M gives an inferable flip rate, so the one-step target deviation from omitting t is O(1/M) and multi-step error accumulates as O(T/M); then trains t-free GDSS and DiGress variants on QM9 and soc-Epinions1 to show parity or better quality with fewer parameters and faster epochs.
Key results
- On QM9, DiGress t-free (12.65M params) matches t-aware validity (99.99% vs 99.98%) with ~4% fewer parameters; GDSS t-free reaches 94.00% validity and 89.60% uniqueness vs 92.32% and 81.08% for t-aware.
- On soc-Epinions1 with N_max=50, DiGress t-free(warm) drops validity to 23.67% from 100% for t-aware, showing sensitivity of transformer-based discrete models.
- Synthetic experiments show posterior variance of flip rate scaling with slope -1.00 in |E| (R^2 0.999), matching the O(|E|^-1) theory.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove timestep conditioning | GDSS · QM9 | Validity (%): 92.32$\pm$0.19 -> 94.00$\pm$0.09 | Helped |
| Remove timestep conditioning | GDSS · QM9 | Unique (%): 81.08$\pm$0.31 -> 89.60$\pm$0.20 | Helped |
| Remove timestep conditioning | GDSS · QM9 | Params (M): 1.89 -> 1.79 | Helped |
| Remove timestep conditioning | GDSS · QM9 | Time: 9.56 -> 8.78 | Helped |
| Remove timestep conditioning | DiGress · QM9 | Validity (%): 99.98$\pm$0.01 -> 99.99$\pm$0.01 | No clear effect |
| Remove timestep conditioning | GDSS · soc-Epinions1 | Valid (%): 25.44$\pm$1.22 -> 33.36$\pm$1.43 | Helped |
| Remove timestep conditioning | GDSS · soc-Epinions1 | MMD Overall: 0.76$\pm$0.02 -> 0.72$\pm$0.00 | Helped |
| Remove timestep conditioning | DiGress · soc-Epinions1 | Valid (%): 100.00$\pm$0.00 -> 99.83$\pm$0.27 | Hurt |
| Remove timestep conditioning | DiGress · soc-Epinions1 | MMD Overall: 0.53$\pm$0.00 -> 0.49$\pm$0.00 | Helped |
| Warm-start from t-aware weights, then fine-tune without timestep embeddings (ablation) | DiGress t-free · soc-Epinions1 | Valid (%): 99.83$\pm$0.27 -> 23.67$\pm$2.08 | Hurt |
| Warm-start from t-aware weights, then fine-tune without timestep embeddings (ablation) | GDSS t-free · soc-Epinions1 | Valid (%): 33.36$\pm$1.43 -> 48.00$\pm$1.99 | Helped |
| Warm-start from t-aware weights, then fine-tune without timestep embeddings (ablation) | GDSS t-free · soc-Epinions1 | MMD Overall: 0.72$\pm$0.00 -> 0.66$\pm$0.00 | Helped |
Surprising findings
- DiGress t-free(warm) collapses to 23.67% validity on soc-Epinions1 at N_max=50, contrary to the claim that t-free models are broadly comparable.
- Stronger structure-feature coupling (gamma up to 0.99) improves unconditional reconstruction error by about 40%, which the theory does not predict.
Limitations
- Theory relies on strong assumptions (global Lipschitz denoiser with L near 1, O(1/M) model MSE) that are asserted rather than verified for trained networks.
- Small or sparse graphs and architectures like DiGress show loss of validity without time conditioning, limiting the claim to large graphs.
Experimental setup
Datasets: QM9 soc-Epinions1
Benchmarks: QM9 soc-Epinions1
Metrics: Validity Uniqueness Novelty MW mean MMD degree MMD clustering MMD triangle parameter count per-epoch time
Baselines: DiGress t-aware GDSS t-aware
Builds on: DiGress GDSS Bernstein-von Mises theorem
Closest papers by meaning
- Simple and Critical Iterative Denoising: A Recasting of Discrete Diffusion in Graph Generation (ICML 2025 · similarity 0.91)
- When Edge Independence Fails: Joint Graph Diffusion with Latent Sociability Priors (NeurIPS 2026 · similarity 0.90)
- Learn Beneficial Noise as Graph Augmentation (ICML 2025 · similarity 0.89)
- Generating Graphs via Spectral Diffusion (ICLR 2025 · similarity 0.88)
- Geometric Generative Modeling with Noise-Conditioned Graph Networks (ICML 2025 · similarity 0.87)
- Generative Modelling of Structurally Constrained Graphs (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.