A Unified Uncertainty Representation for Graph Neural Networks via Doubly-Spectral Stochastic Expansion

What it claims

DSS-GNN gives every node embedding a polynomial-chaos expansion in one shared Gaussian latent and filters it spectrally along both graph frequency and chaos order. One deterministic forward pass then yields separate readouts for each task:

• a quadrature-averaged predictive distribution for prediction and calibration;
• a mean-logit energy score for OOD detection;
• in hybrid mode, a regularized residual for robustness to distribution shift.

There is no Monte-Carlo sampling and no post-hoc temperature scaling. The authors say the theory establishes representational capacity only. The calibration, OOD and shift claims are empirical. The paper is at NeurIPS 2026.

How it works

1. Lift. The input features are lifted to P+1 coefficient matrices H_n, one per chaos order. Each node therefore stores P+1 hidden vectors instead of one. With P=0, the layer reduces to ChebNet/GCN (Proposition 1).
2. Graph axis. Each coefficient passes through two Chebyshev filters on the rescaled Laplacian, one low-pass and one high-pass.
3. Chaos axis. Per-order gates α_n and β_n mix the two branches. The gates are scalars or low-order chaos expansions; in practice the gate order P_g ≤ 1.
4. Nonlinearity. The field is evaluated at S Gauss–Hermite nodes. The activation is applied there, and the result is projected back onto the Hermite basis. With S ≥ P+1, the filter and gate terms are integrated exactly.
5. Readouts. The final layer produces per-order logits Z_{i,n}. - Prediction and calibration use the predictive p̄ = Σ_s μ_s softmax(z^(s)). - OOD detection uses the energy −log Σ exp(Z_{i,0}), with graph propagation as in GNNSafe. The energy score is chosen because a uniform shift in logits changes the energy, while softmax entropy and mutual information cancel it out.
6. Training. The loss is cross-entropy on the mean logit only, plus λ_reg × chaos energy (Σ_{n≥1} ‖Z_{i,n}‖²). The higher-order coefficients are trained only implicitly, through the projection step. The paper shows this loss matches the quadrature-averaged loss to first order. It also bounds the gap between p̄ and softmax(Z_0) by the chaos covariance.

DSS-Hybrid sums a base encoder's logits with the DSS branch's logits:

z_base + γ_res · Z_0

• The base encoder is a 2-layer GCN, chosen for comparability with the baselines.
• γ_res is a learned scale.
• The GCN is warmed up before the DSS branch is switched on.

flowchart LR
  X[Node features] --> GCN[2-layer GCN base]
  X --> DSS[DSS branch: P+1 chaos coeffs]
  GCN -->|warm up first| SUM[base + γ_res · DSS]
  DSS --> SUM
  SUM --> Q[Quadrature-avg softmax: predict / calibrate]
  SUM --> E[Mean-logit energy: OOD score]

Evidence

The fetched text cuts off at the start of the experimental setup, so no result tables are visible. The setup that is visible:

• Calibration
• Standalone DSS-GNN against TFE-GNN and G-ΔUQ.
• 14 node-classification benchmarks, 10 splits.
• Metrics: accuracy and Brier score, with no post-hoc calibration.
• OOD detection
• DSS-Hybrid on the GNNSafe benchmark: nine node-OOD settings on Cora, Amazon-Photo and Coauthor-CS, plus cross-graph Twitch and Arxiv, 3 seeds.
• GNNSafe, GNNSafe++ and GPN are compared using the numbers reported by Wu et al.
• Graph-EBM, MC-dropout (M=20) and Deep Ensembles (M=5) are rerun on the same GCN backbone with the same score propagation.
• Distribution shift: 7 GOOD concept-shift benchmarks, trained under ERM only.

Other results:

• Performance plateaus empirically at chaos order P = 1–2, so the extra cost over a deterministic layer is roughly a factor of 2–3 in coefficient states, plus S quadrature evaluations.
• OOD detection is reported as insensitive to chaos order, which the authors say is consistent with the energy readout using only Z_0.

Limits

• The capacity theorem needs full-rank features (d_in ≥ N). Most real graphs do not satisfy this; the paper notes Cora has 1,433 features for 2,708 nodes.
• Remark 5 extends the result to targets whose coefficients lie in range(X), under a positivity condition. Appending a constant feature column satisfies that condition.
• The exponential truncation rate additionally requires targets with an entire Gaussian extension.
• The theorems do not guarantee calibration, OOD performance or robustness.
• The hybrid's shift robustness is described as "an inductive bias rather than a distribution-free guarantee". It is shown only under ERM, not against invariance-learning baselines.
• Each mode is only partly effective on the other's tasks. The abstract notes "documented exceptions" when each mode is evaluated on the other's tasks. You need to check Section 4.5 to choose a mode.
• Not on the page: code, weights or licence information.