Spectral Graph Conditional Exchangeability (SGCE)

• SGCE is a framework defining conditional exchangeability in graph time series by separating structured low-frequency trends from nearly exchangeable high-frequency components.
• The approach underpins SCALE, which employs spectral graph wavelets, adaptive gating, and low-frequency conditioning for conformal prediction.
• Empirical evaluations on traffic datasets show that SCALE achieves valid coverage with significantly sharper prediction intervals than traditional methods.

Searching arXiv for the specified paper and topic to ground the article in the cited source. Spectral Graph Conditional Exchangeability (SGCE) is a concept introduced for conformal prediction in graph-structured multivariate time series, where the standard exchangeability assumption is often violated by cross-node coupling. In the formulation of Guo et al., global trends and cross-node dependence are represented in low-frequency graph spectral components, while high-frequency components are treated as nearly exchangeable once conditioned on the low-frequency part. On that basis, the paper proposes Spectral Conformal prediction via wAveLEt transform (SCALE), which performs conformalization in the spectral domain using graph wavelets and adaptive gating over a low-frequency embedding (Guo et al., 6 May 2026).

1. Problem setting and spectral preliminaries

The setting is a graph-structured multivariate time series on an undirected graph $G=(V,E)$ with $N$ nodes, adjacency $A\in\mathbb{R}^{N\times N}$, and degree matrix $D=\operatorname{diag}(A\mathbf{1})$ (Guo et al., 6 May 2026). At each time $t$, the observed signal is a snapshot $x_t\in\mathbb{R}^N$. Over a window of length $W$, the history is written as

$$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$

The spectral construction uses the normalized graph Laplacian

$L = I - D^{-1/2} A D^{-1/2},$

with eigendecomposition

$$L = U\Lambda U^\top,\qquad U\in\mathbb{R}^{N\times N}\ \text{orthonormal},\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_N),\qquad 0=\lambda_1\le \cdots \le \lambda_N\le 2.$$

The decomposition underlying SGCE is based on the spectral graph wavelet transform (SGWT) of Hammond et al. (2011). Given $N$0 band-pass kernels $N$1 and a low-pass kernel $N$2, any signal $N$3 is mapped to

$N$4

with decomposition

$N$5

For a chosen cutoff scale $N$6, the low-frequency and high-frequency parts are defined as

$N$7

This construction is central because the non-exchangeability of raw graph time series is attributed to structured low-frequency coupling, whereas the residual high-frequency component is the object on which conformal prediction is applied.

2. Formal definition of SGCE

SGCE is defined on paired low- and high-frequency spectral components. Let $N$8 be $N$9 i.i.d. examples of low/high spectral components, for example residuals produced by a forecasting model (Guo et al., 6 May 2026). The residual process satisfies Spectral Graph Conditional Exchangeability if, for any permutation $A\in\mathbb{R}^{N\times N}$0 of $A\in\mathbb{R}^{N\times N}$1,

$A\in\mathbb{R}^{N\times N}$2

Equivalently, conditional on the realized low-frequency components, the high-frequency components are jointly exchangeable. This definition does not assert exchangeability of the original graph-structured time series itself. Rather, it relocates the exchangeability requirement to a conditional statement in the spectral domain. In the paper’s formulation, this is the mechanism by which global trends are preserved while conformal prediction remains applicable (Guo et al., 6 May 2026).

3. Assumptions and spectral intuition

Two assumptions underpin SGCE in the paper.

First, there is spectral concentration of global trends. The dominant cross-node correlations and “global” modes of the graph signal are assumed to lie in the first $A\in\mathbb{R}^{N\times N}$3 eigenvectors associated with small eigenvalues, and therefore to be concentrated in the low-frequency component $A\in\mathbb{R}^{N\times N}$4 (Guo et al., 6 May 2026). Empirically, the paper measures the “correlation intensity” of a wavelet band $A\in\mathbb{R}^{N\times N}$5 via average $A\in\mathbb{R}^{N\times N}$6 and reports that it decays sharply for small $A\in\mathbb{R}^{N\times N}$7, corresponding to high frequencies.

Second, there is approximate orthogonality or weak coupling in $A\in\mathbb{R}^{N\times N}$8. The high-frequency wavelet bands $A\in\mathbb{R}^{N\times N}$9 are described as exhibiting near-zero cross-node correlation, which makes $D=\operatorname{diag}(A\mathbf{1})$0 nearly exchangeable (Guo et al., 6 May 2026).

These assumptions delimit the intended scope of SGCE. The framework does not remove graph dependence globally; it separates dependence into a structured low-frequency component and a high-frequency component on which conditional exchangeability is plausible. A common misconception would be to treat SGCE as ordinary exchangeability after a graph transform. The definition is explicitly conditional, and the conditioning variable is the realized low-frequency component.

4. SCALE: spectral conformal prediction via wavelets

SCALE implements a conformal prediction layer designed to exploit SGCE (Guo et al., 6 May 2026). Its inputs are a pre-trained point forecaster $D=\operatorname{diag}(A\mathbf{1})$1, which predicts $D=\operatorname{diag}(A\mathbf{1})$2 from $D=\operatorname{diag}(A\mathbf{1})$3, and a calibration set of residuals

$D=\operatorname{diag}(A\mathbf{1})$4

The procedure consists of the following stages.

First, for each time $D=\operatorname{diag}(A\mathbf{1})$5, the residual snapshot $D=\operatorname{diag}(A\mathbf{1})$6 is decomposed by SGWT into $D=\operatorname{diag}(A\mathbf{1})$7 and $D=\operatorname{diag}(A\mathbf{1})$8 using cutoff scale $D=\operatorname{diag}(A\mathbf{1})$9. Over a window of length $t$0, this yields $t$1 and $t$2.

Second, the low-frequency sequence is mapped to a conditioning embedding. Specifically, $t$3, together with positional or time embeddings, is passed through a small STGNN/MLP to obtain

$t$4

Third, the high-frequency component is summarized by per-node temporal statistics without graph weights:

$t$5

Fourth, SCALE uses a gated quantile predictor. Two parallel feature maps are formed,

$t$6

both projected to quantile channels of size $t$7 for lower and upper quantiles. A gating map is then computed as

$t$8

and the conditional quantile output is

$t$9

The model is trained end-to-end with pinball loss at level $x_t\in\mathbb{R}^N$0 on calibration residuals.

Fifth, at test time, SCALE forms prediction sets by adding the point forecast and the low-frequency component back to the residual quantiles:

$x_t\in\mathbb{R}^N$1

The operational role of each component is explicit in the formulation: SGWT separates residual structure, the low-frequency encoder provides conditioning information, the high-frequency summary captures nodewise variability, and the gating mechanism combines both sources to produce conditional quantiles.

5. Coverage guarantees

The theoretical claims are stated for residuals $x_t\in\mathbb{R}^N$2 satisfying exact SGCE (Guo et al., 6 May 2026).

Theorem 5.1 (Finite-Sample Coverage under SGCE) states that if $x_t\in\mathbb{R}^N$3 is the prediction set produced by SCALE at miscoverage level $x_t\in\mathbb{R}^N$4, then for any $x_t\in\mathbb{R}^N$5 each snapshot,

$x_t\in\mathbb{R}^N$6

The proof sketch in the paper has three steps: the sequence of high-frequency residuals $x_t\in\mathbb{R}^N$7 is exchangeable conditional on $x_t\in\mathbb{R}^N$8; conformal prediction on $x_t\in\mathbb{R}^N$9 with a permutation-invariant score $W$0 yields coverage at least $W$1 for the test high-frequency residual; and reconstruction through

$W$2

is a deterministic shift that preserves coverage.

Theorem 5.2 (Approximate Coverage under Imperfect SGWT) addresses transform error and residual coupling. If the wavelet filters produce $W$3 with bounded leakage

$W$4

and the true $W$5 exhibits residual couplings of TV-distance $W$6 to an exchangeable law, then SCALE satisfies

$W$7

As $W$8 and $W$9, exact $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$0 coverage is recovered.

Taken together, these results distinguish between an exact regime, where finite-sample coverage follows from SGCE itself, and an approximate regime, where deviations from the ideal spectral decomposition and from exchangeability are reflected as additive coverage degradation terms.

6. Empirical evaluation

The empirical study uses four real-world traffic datasets: METR-LA with 207 sensors, PEMS04 with 307, PEMS07 with 883, and PEMS08 with 170 (Guo et al., 6 May 2026). The protocol is a split of 40% train, 40% calibration, and 20% test, with $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$1, $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$2, and $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$3.

The baselines are SCP, SeqCP, NexCP, EnbPI, HopCPT, CoREL, and ConForME. Evaluation uses empirical Coverage, average PI-Width, and Winkler score, where lower Winkler is sharper. The paper marks a result with “✔” when $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$4.

For the GRU+DiffConv backbone at $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$5 on METR-LA, the reported comparison is:

Method	Coverage	Width
SCP	0.947 (✔)	13.62
CoREL	0.951 (✔)	12.19
SCALE	0.948 (✔)	11.69

The paper further reports that, across the four datasets and all $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$6, SCALE always meets nominal coverage within $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$7 and achieves the smallest average width and Winkler (Guo et al., 6 May 2026). On METR-LA at $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$8, SCALE is reported as 14.4% narrower than EnbPI. Ablation results indicate that removing SGWT or removing low-frequency conditioning degrades coverage or widens intervals. In the multi-step setting, with $$X_{t-W+1:t}=[x_{t-W+1},\dots,x_t]\in\mathbb{R}^{N\times W}.$$9 up to 96, SCALE maintains approximately nominal coverage and narrower intervals than CoREL and ConForME at every horizon. A sensitivity study on METR-LA varying the number of wavelet scales $L = I - D^{-1/2} A D^{-1/2},$0 yields coverage in $L = I - D^{-1/2} A D^{-1/2},$1 for target $L = I - D^{-1/2} A D^{-1/2},$2, while width varies by less than $L = I - D^{-1/2} A D^{-1/2},$3.

These results are presented as evidence for the central design choice of the method: separating residuals spectrally and conformalizing only the high-frequency component while conditioning on the low-frequency component.

7. Interpretation and position within conformal prediction for graph time series

Within the paper’s framework, SGCE formalizes the claim that on a graph, global couplings concentrate in low spectral bands while high bands are approximately exchangeable once conditioned on the low bands (Guo et al., 6 May 2026). SCALE is the corresponding practical conformal layer, combining graph wavelets with gated quantile prediction.

The method is positioned against two families of alternatives listed in the experiments. One family applies conformal prediction directly to residuals or to temporally adapted residual windows, as in SCP, SeqCP, and NexCP. The other family uses ensemble, associative-memory, or graph-aware quantile predictors, including EnbPI, HopCPT, CoREL, and ConForME. The paper’s stated motivation is that direct application of conformal prediction can be unreliable when graph-induced cross-node coupling violates exchangeability.

A further misconception would be to interpret SGCE as merely a heuristic spectral denoising step. In the formulation given, the spectral decomposition is not auxiliary; it is the basis of both the conditional exchangeability definition and the coverage argument. Likewise, the low-frequency component is not discarded. It is preserved explicitly in the reconstruction

$L = I - D^{-1/2} A D^{-1/2},$4

and the prediction set construction adds back both the point forecast and the low-frequency component.

The paper’s summary is correspondingly narrow and technical: SGCE is the formal statement, SCALE is the implementation, finite-sample $L = I - D^{-1/2} A D^{-1/2},$5 coverage holds under SGCE, approximate coverage degrades gracefully under imperfect transforms, and the empirical results on traffic forecasting benchmarks show valid coverage with sharper intervals than prior split or adaptive conformal prediction methods (Guo et al., 6 May 2026).