ST-GFM: Spectral-Temporal Graph Forecasting with Measured Perturbation Bounds and Masked Spectral Reconstruction
Authors: Aashish Bohra, Vivek Vijay
Organizations: Department of Computer Science and Engineering, Indian Institute of Technology Jodhpur, Jodhpur, Rajasthan, India · Department of Mathematics, Indian Institute of Technology Jodhpur, Jodhpur, Rajasthan, India
Graph forecasters for sensor networks are trained in one regime and deployed in others, where sensor coupling and signal frequency content both change. We split this spectral shift into an operator channel and a frequency channel, each with a dimensionless measure. ST-GFM (Spectral-Temporal Graph Forecasting Model) filters every input step with Chebyshev and bidirectional diffusion filters on one learned, non-symmetric operator and alternates them with per-node temporal attention. For the operator channel we prove Frobenius-norm perturbation bounds for both graph branches, show that the Chebyshev bound is attained to first order, and compute its constant from the trained weights and the measured operator norm: 17.9 on METR-LA and 21.5 on PEMS-BAY. Under random edge dropout the trained layer moves by 2.0% to 3.1% of this bound. Masked Spectral Reconstruction (MSR), a self-supervised loss active in the first 40 epochs, matches the log power spectrum of the encoder output to that of the input. For small log-spectral residuals on unclamped bins it is proportional to an Itakura-Saito divergence to leading order. Removing it raises the 60-minute MAE by 0.30 on METR-LA, close to the 0.33 lost when all three self-supervised losses go. Under one shared evaluation pipeline, the 60-minute MAE of ST-GFM is 3.5% below a retrained STAEformer on METR-LA and 13.0% below it on PEMS-BAY. Its 60-minute MAE is 3.0980.017 on METR-LA (five seeds) and 1.5760.006 on PEMS-BAY (three seeds), the lowest PEMS-BAY MAE among twelve published baselines.
Figures & tables
Figure 1: Spectral shift between two regimes of the same freeway sensor network. The left and centre panels are schematic: sensor coupling above, the power spectrum of a typical speed series below. Spectral shift reaches a deployed model through an operator channel and a frequency channel, which the paper analyses and tests separately.
Method
Cheb.
Learned
SSL
Freq.
Bound
filter
graph
pretext
DCRNN [ 36 ]
✗
✗
✗
✗
✗
STGCN [ 59 ]
✓
✗
✗
✗
✗
ASTGCN [ 19 ]
✓
✗
✗
✗
✗
Graph WaveNet [ 56 ]
✗
✓
✗
✗
✗
MTGNN [ 55 ]
✗
✓
✗
✗
✗
Table 1: Positioning of ST-GFM. “Cheb. Filter” Means a Chebyshev Polynomial Graph Filter. CITRUS Uses Continuous Heat-Kernel Filters on a Product Graph. “Bound” Means a Stated Bound on How the Output Moves When the Graph Operator Is Perturbed.
Figure 2: Overview of ST-GFM. Row 1: time-preserving decomposition, embeddings and the learned anchor operator. Row 2: DCSE layers, which share one operator, alternate with per-node temporal attention. Row 3: the bounds of Section 5 and the forecasting head with its persistence term. Row 4: self-supervised branches, active for the first 40 epochs.
Figure 3: One DCSE layer, applied to the N×d state at one time step. Both graph branches read the anchor operator L~θ . The Chebyshev branch applies a polynomial filter of orders 0 to 2, and the diffusion branch takes two propagation steps in each direction. Their outputs are concatenated, mixed by a linear map and added to the input before layer normalisation. The red tags mark the result in Section 5 that bounds each branch.
∥Θ0∥2
∥Θ1∥2
∥Θ2∥2
ccheb(1,0)
METR-LA
Seed 13
4.133
3.005
2.958
14.84
Seed 42
4.004
2.866
2.730
13.79
Seed 101
3.535
3.153
2.900
14.75
Seed 202
3.344
3.170
2.971
15.05
Seed 303
3.608
2.977
2.766
14.04
Table 2: Spectral Norms of the Trained Chebyshev Coefficients in the First DCSE Layer of Every Run, and the Unit-Norm Constant ccheb(1,0)=∥Θ1∥2+4∥Θ2∥2 of ( 14 ), Computed From Unrounded Norms.
METR-LA
PEMS-BAY
∥\Ltθ∥2
1.30
1.35
ccheb(ρ,0)
17.93
21.48
Learned-graph gate α
0.45
0.45
a=∥A^∥2
1.54
1.55
m=mini(A^1)i
1.38
1.33
4a/m
4.46
4.66
Table 3: Operator and Branch Constants in the First DCSE Layer, Mean Over Runs. ccheb(ρ,0)=∥Θ1∥2+4ρ∥Θ2∥2 and cdiff=w1+(4a/m)w2 With ws=∥Wf,s∥2+∥Wb,s∥2 , a=∥A^∥2 and m=mini(A^1)i , as in Lemma 1 for Kdiff=2 .
Dataset
Signal
Sensors
Steps
Period
Split (%)
METR-LA
speed
207
34,272
03/2012–06/2012
70/10/20
PEMS-BAY
speed
325
52,116
01/2017–06/2017
70/10/20
PEMS04
flow
307
16,992
01/2018–02/2018
60/20/20
PEMS08
flow
170
17,856
07/2016–08/2016
60/20/20
Table 4: Datasets
Figure 4: 60-minute MAE on METR-LA and PEMS-BAY for every compared model that reports both datasets. Lower left is better. The shaded region holds the models with a higher MAE than ST-GFM on both datasets. Gray dots are published values from the original papers. The open pentagon is STAEformer retrained under our pipeline (Table 8 ).
H=3 (15 min)
H=6 (30 min)
H=12 (60 min)
Method
MAE
RMSE
MAPE (%)
MAE
RMSE
MAPE (%)
MAE
RMSE
MAPE (%)
DCRNN [ 36 ]
2.77
5.38
7.30
3.15
6.45
8.80
3.60
7.59
10.50
Graph WaveNet [ 56 ]
2.69
5.15
6.90
3.07
6.22
8.37
3.53
7.37
10.01
MTGNN [ 55 ]
2.69
5.18
6.86
3.05
6.17
8.19
3.49
7.23
9.87
STEP [ 46 ]
2.61
4.98
6.60
2.96
5.97
7.96
3.37
6.99
9.61
D 2 STGNN [ 47 ]
2.56
4.88
6.48
2.90
5.89
7.78
3.35
7.03
9.40
Table 5: Forecasting Results on METR-LA. ST-GFM Entries Are Five-Seed Means With 95% Seed Intervals. Bold Marks the Best and Underline the Second Best in Each Column. Baseline Numbers Are From the Original Papers.
H=3 (15 min)
H=6 (30 min)
H=12 (60 min)
Method
MAE
RMSE
MAPE (%)
MAE
RMSE
MAPE (%)
MAE
RMSE
MAPE (%)
DCRNN [ 36 ]
1.38
2.95
2.90
1.74
3.97
3.90
2.07
4.74
4.90
Graph WaveNet [ 56 ]
1.30
2.74
2.73
1.63
3.70
3.67
1.95
4.52
4.63
MTGNN [ 55 ]
1.32
2.79
2.77
1.65
3.74
3.69
1.94
4.49
4.53
STEP [ 46 ]
1.26
2.73
2.59
1.55
3.58
3.43
1.79
4.20
4.18
D 2 STGNN [ 47 ]
1.24
2.60
2.58
1.55
3.52
3.49
1.85
4.30
4.37
Table 6: Forecasting Results on PEMS-BAY. ST-GFM Entries Are Three-Seed Means With 95% Seed Intervals. Formatting as in Table 5 .
Figure 5: Accuracy by horizon against seven baselines, including TITAN and T-Graphormer. (a) and (b) MAE of ST-GFM and the baselines, with 95% seed intervals on ST-GFM (five seeds on METR-LA, three on PEMS-BAY, smaller than the marker on PEMS-BAY). (c) and (d) Relative MAE reduction of ST-GFM over each baseline, computed from unrounded seed means and the values in Tables 5 and 6 . Baselines are single published numbers.
PEMS04
PEMS08
Method
MAE
RMSE
MAPE (%)
MAE
RMSE
MAPE (%)
PDFormer [ 24 ]
18.32
29.97
12.10
13.58
23.51
9.05
STAEformer [ 37 ]
18.22
30.18
11.98
13.46
23.25
8.88
ST-GFM (ours)
18.03
29.85
11.79
13.18
22.91
8.73
STD-MAE [ 17 ]
17.80
29.25
11.97
13.44
22.47
8.76
Table 7: Average Over All 12 Horizons on the Flow Datasets PEMS04 and PEMS08. Bold Marks the Best and Underline the Second Best Among Models With a 12-Step Input. STD-MAE Pre-Trains on 864 to 2016 Steps and Is Listed Below the Rule as a Long-History Reference. Baseline Numbers Are From the Original Papers. ST-GFM Is a Single Run (Seed 42).
METR-LA
PEMS-BAY
Method
15 min
30 min
60 min
15 min
30 min
60 min
Persistence
3.557
4.153
5.145
1.301
1.649
2.173
Historical average
4.193
4.193
4.193
2.631
2.631
2.630
Graph WaveNet
–
–
3.258
–
–
–
STAEformer
2.55
2.86
3.21
1.26
1.56
1.81
ST-GFM (ours)
2.554
2.800
3.098
1.098
1.325
1.576
Table 8: MAE Under Our Pipeline, With the Same Test Windows, Code and Mask for Every Model. Graph WaveNet Is the Reduced Model of Section 6 . ST-GFM Is the Mean Over Seeds. Bold Marks the Lowest MAE per Column, Including Ties at the Reported Precision.
METR-LA
PEMS-BAY
Variant
MAE
Δ
MAE
Δ
Full ST-GFM
3.098
–
1.576
–
w/o decomposition
3.29
+0.19
1.71
+0.13
w/o DCSE (GCN instead)
3.42
+0.32
1.79
+0.21
w/o MSR
3.40
+0.30
1.80
+0.22
w/o masked nodes
3.20
+0.10
1.63
+0.05
Table 9: Ablation at H=12 . Δ Is the MAE Increase Over the Full Model. The Full-Model Reference Is the Multi-Seed Mean of Tables 5 and 6 , and Each Variant Is a Single Run.
Figure 6: Self-supervised losses (a), (b) and validation MAE (c), (d) over training, for all eight runs. Thick lines are means, bands span the runs, thin lines are single runs and dots mark each run’s best epoch. Shading marks the SSL-only phase and the ramp-down of its weight. After epoch 40 the SSL losses are monitored, not optimised. Brackets count the averaged checkpoints that fall in epochs 30 to 50.
Figure 7: Test MAE of ST-GFM under random symmetric perturbations of the operator with Frobenius norm ε . Thick lines: mean over runs, bands: one standard deviation, thin lines: individual runs. One perturbation draw per run and level.
ε
Graph WaveNet
ST-GFM
Gap
0.00
3.258
3.076
0.182
0.05
3.259
3.076
0.183
0.10
3.259
3.076
0.183
0.20
3.259
3.077
0.182
0.30
3.260
3.079
0.181
0.50
3.261
3.080
0.181
Table 10: Matched Adjacency Perturbations on METR-LA at H=12 (Seed-101 Checkpoints, Mean of Two Draws). The Graph WaveNet Entries Are From the Reduced Model Retrained Under Our Preprocessing and Evaluation.
Figure 8: Edge dropout on the static adjacency, mean over runs and three draws (values in Supplementary Table S3). (a) Observed relative change of the Chebyshev-branch output in the first DCSE layer and the bound of ( 14 ) at the realised perturbation and the measured operator norm, on a log scale. Shading spans the gap on METR-LA. (b) Observed change as a percentage of the bound. (c) Relative increase of the 60-minute forecast MAE, with the absolute MAE at p=0 and p=0.5 annotated.
Figure 9: Spectral shift δfreq of the test windows by hour of the forecast origin. Lines are medians and bands interquartile ranges, which show spread, not confidence intervals. A weekend hour holds about 72 windows on METR-LA and 120 on PEMS-BAY, a weekday hour about 216 and 312.
H=3
H=6
H=12
Method
Full
Weekend
Δ (%)
Full
Weekend
Δ (%)
Full
Weekend
Δ (%)
METR-LA
Persistence (last reading)
3.557
3.337
−6.2
4.153
3.971
−4.4
5.145
5.073
−1.4
Historical average (time of week)
4.193
3.496
−16.6
4.193
3.493
−16.7
4.193
3.494
−16.7
PEMS-BAY
Persistence (last reading)
1.301
0.887
−31.8
1.649
1.016
−38.4
2.173
1.169
−46.2
Table 11: MAE of Two Naive Forecasters on the Full Test Sets and on Weekend Windows (First Forecast Step on Saturday or Sunday), Scored With the Same Windows, Code and Mask as ST-GFM. Δ Is the Change From Full to Weekend.
Encoder
IC
R-IC
IR
SR
p
ST-GFM (METR-LA → FF49) #1
0.055
0.049
0.21
0.57
0.004
ST-GFM (METR-LA → FF49) #2
0.043
0.041
0.16
0.54
0.019
ST-GFM (PEMS-BAY → FF49) #1
0.052
0.051
0.20
0.60
0.006
ST-GFM (PEMS-BAY → FF49) #2
0.059
0.057
0.22
0.71
0.003
Random frozen backbone, #1
0.042
0.040
0.16
0.49
0.020
Random frozen backbone, #2
0.036
0.034
0.14
0.42
0.038
Table 12: Transfer to 43 Fama–French Industry Portfolios (172 Scored Months, January 2011 to April 2025). Runs #1 and #2 Use METR-LA Seeds 202 and 303 and PEMS-BAY Seeds 101 and 303. R-IC Is Rank IC, IR the IC Information Ratio, SR the Annualised Sharpe Ratio Net of 1 bp and p the One-Sided t -Test of Mean IC.
Figure 10: Transfer to the 43 Fama–French industry portfolios (Table 12 ). Blue bars use METR-LA encoders, green bars PEMS-BAY encoders and gray bars the randomly initialised frozen backbone. Dashed lines are group means.
Figure 11: Accuracy against model size on METR-LA at 60 minutes. Each label gives the width, the 60-minute MAE and the training time on one RTX 5060 Ti. Circles are single runs with seed 42, and Supplementary Table S4 lists their values. The star is the five-seed mean of the full model. The dashed line is the published STAEformer MAE and the dotted line STAEformer retrained under our pipeline.
Department of Computer Science and Engineering, Indian Institute of Technology Patna, India · Sopra Banking Software India as Senior Software Development Engineer