acceptodds
Under review as a conference paper at ICLR 2027

Global Shortcut Contamination in End-to-End Graph Learning for Multivariate Forecasting

Abstract

Graph neural networks for multivariate time-series forecasting increasingly learn graph structures end-to-end, jointly optimizing latent relational structures and forecasting models from data. Although such learned graphs are often treated as variable dependency structures, the end-to-end forecasting objective does not inherently support their structural correctness. As a result, the learned graph should be viewed as a prediction-optimized structure, which it may fail to capture invariant local dependencies. Here we analyze why and when this mismatch arises. We show that, in the presence of persistent global predictive modes, end-to-end graph learning favors train-specific shortcut structures aligned with global components, and we theoretically characterize when such shortcut-contaminated graphs are preferred. Based on this analysis, we propose a simple projection-based decoupling strategy that removes dominant global components before graph learning in an architecture-agnostic manner. Experiments on controlled synthetic benchmarks and real traffic benchmarks under time-of-day shifts show that our proposed projection method substantially improves robustness for unstable graph learners and often reduces out-of-distribution (OOD) forecasting error, while its benefits for already stable architectures are shift-dependent. Beyond OOD robustness, our results highlight the importance of separating predictive usefulness from structural correctness when interpreting learned graphs in end-to-end forecasting models.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Network, Time series analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_LVmPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the issue of "global shortcut contamination" that can arise in end-to-end graph structure learning for multivariate time series forecasting. The authors propose a "projection-based decoupling" approach, which removes dominant shared components via projection prior to graph learning. They introduce two implementations—MeanProj and PCA—and use a "last-observation persistence" strategy to reconstruct the removed coordinates. Experimental results demonstrate that this projection technique can significantly reduce out-of-distribution (OOD) errors in certain models and scenarios; however, its effectiveness is highly dependent on the specific model architecture and the chosen projection direction.

Critical strengths
  1. The theoretical analysis centers on specific failure modes.
  2. The paper repeatedly emphasizes that the "learned adjacency" should be interpreted as a "predictive computation graph."
Critical weaknesses
  1. Eqs. (2)–(5) serve more as an "existence proof" rather than a general characterization of practical end-to-end (E2E) graph learners.
  2. The paper lacks a rigorous link between the risk analysis based on binary support and the actual differentiable graph generator; consequently, the theoretical constructs do not perfectly align with the empirical models.
  3. While the theoretical results demonstrate that "certain shortcut edges can lead to lower risk," they do not prove that practical, non-linear E2E training will necessarily—or with high probability—learn these specific edges.
  4. The "projection intervention" affects not only the graph learning process but also the forecasting task itself.
  5. The synthetic experiments are conducted on a small scale: N=8N=8, with a limited autoregressive (AR) order and a simple rank-one factor structure.
  6. The real-world experiments are limited to only two traffic forecasting datasets (METR-LA and PEMS-BAY). Although the title refers to "multivariate forecasting," the paper lacks validation on multivariate time series (MTS) from other domains, such as energy, weather, finance, or industrial applications.
  7. The real-world experiments aggregate results from only three random seeds, and the tables lack standard deviations (std), confidence intervals (CI), or significance tests. In some cases, the observed performance gains are marginal.
  8. I will update my final score based on the authors' rebuttal.
Questions for authors

Q1. How can it be demonstrated that the gains truly stem from "graph decontamination" rather than from the projection process itself?

Q2. How does this approach compare to common methods for removing common components?

Q3. Does the actual "traffic shift" truly correspond to the "loading rotation" described in Theorem 2?

Q4. Is the graph obtained after projection genuinely "more correct"?

Q5. What is the impact of this method on in-distribution (ID) performance and on other domains?

1 comment

Thank you for the detailed assessment. We added controlled completion, structure, rank, rotation, graph-utility, and cross-domain evaluations, while clarifying what the theory does—and does not—guarantee.

1. Binary support vs. nonlinear training

Addressed in: Sec. 2; Appendices A.3, A.8–A.9, G.6.

Theorem 1 gives an exact population-risk gap and strictness conditions in a specified one-step model, not merely an existence example. Since the conditional mean is linear under the Gaussian assumptions, the same gap holds for potentially nonlinear predictors with the same observed-coordinate access (Corollary 1).

For an actual feasible graph-related parameter direction dd, we prove

DϕR[d]=−2E⟨Y−F,DϕF[d]⟩D_\phi R[d] = -2 \mathbb{E}\langle Y - F, D_\phi F[d] \rangle

(Proposition 3), a conditional local descent statement, not a convergence guarantee.

JVP/autograd matched numerically. At the smallest step, train loss decreased in 4/5 seeds for both tested directions, but OOD loss increased in all 30 direction/step/seed tests. Shared embeddings also affect forecasting weights; these are not graph-only interventions or observed training trajectories.

Q1. Source of gains

Addressed in: Secs. 3.2 and 4.5; Appendices G.2–G.3.

In a five-seed AGCRN synthetic experiment (N=32N = 32 observed variables, one node each), Original, Original with matched persistence, and projected-hard have mean OOD MSE 1.34422, 1.09065, 1.06519, respectively.

Matched output replacement reproduces 90.9% of the Original-to-projected-hard gain arithmetically; yet projected-hard improves on matched completion by 0.02546±0.001810.02546 \pm 0.00181 (5/5 paired seeds). This remainder includes other training differences; it is not identified as a graph-only effect.

With graph source varied under common forecasting conditions, the MeanProj-derived graph gives a smaller synthetic OOD benefit of 0.00336±0.001640.00336 \pm 0.00164; traffic fixed-graph results are mixed.

Q2. Common-component methods

Addressed in: Sec. 5.

Mean removal/PCA are not claimed as new. Unlike factor-adjusted network approaches such as FNETS [1], we study why prediction-optimal graph access diverges from local support and quantify loading-shift risk.

We retain the forecasting backbone and use last-observation completion. No numerical superiority over FNETS is claimed.

Q3. Traffic shift vs. loading rotation

Addressed in: Secs. 2 and 4.5; Appendix G.5.

Thank you for this question. We have not identified traffic time-of-day/weekend shifts as loading-only rotations.

A separate controlled rotation experiment with constant Bayes MSE tests Theorem 2: train-population per-variable excess MSE follows

0.1265625sin⁡2θ.0.1265625 \sin^2\theta.

Full OLS excess rises from 0.00592 at 0∘0^\circ to 0.13314 at 90∘90^\circ, while local-mask OLS stays near 0.0804, reversing their ordering at larger angles (five seeds).

Fixed MeanProj/PCA do not outperform Full OLS under these rotations, highlighting a limitation of fixed training projectors.

Q4. Structural correctness

Addressed in: Sec. 4.5; Appendix G.4.

With N=64N = 64 variables and two generating factors, five-seed structural AUPRC increases from 0.0601±0.02260.0601 \pm 0.0226 (Original) to 0.6278±0.05470.6278 \pm 0.0547 (PCA rank 2); OOD MSE decreases from 1.5612±0.09721.5612 \pm 0.0972 to 1.0818±0.02081.0818 \pm 0.0208.

Both metrics improve over rank-one PCA in 5/5 paired seeds. We score symmetrized learned adjacency against undirected generating support, not causal edges. The generator does not satisfy all exact-projection assumptions; this is empirical ranking evidence, not guaranteed graph recovery.

Q5. ID, other domains, and reliability

Addressed in: Secs. 4.4–4.5; Appendices B, G.1, G.4, G.7.

Rank-two PCA improves synthetic ID MSE 0.27818→0.269950.27818 \to 0.26995 (5/5 seeds).

We added Electricity (321 variables, MTGNN, five seeds). Chronological held-out test MSE (in 10610^6 units) is 3.025±0.6263.025 \pm 0.626 Original, 2.948±0.1802.948 \pm 0.180 MeanProj, 9.671±0.0239.671 \pm 0.023 PCA.

MeanProj improves MSE in only 3/5 seeds and worsens MAE; PCA worsens MSE in 5/5. The original traffic results remain three seeds, with ID/SD in existing tables; new experiments use five and report paired differences.

Thus, neither universal cross-domain improvement nor statistical significance of marginal effects is claimed.

Reference: [1] Barigozzi, Cho, and Owens (2024). “FNETS: Factor-Adjusted Network Estimation and Forecasting for High-Dimensional Time Series.” Journal of Business & Economic Statistics, 42(3), 890–902.

Reviewer_QoyvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This work analyzes how end-to-end graph learning may favor train-specific shortcut structures rather than capture invariant local dependencies. This authors evaluates a projection-based decoupling strategy on controlled synthetic benchmarks and real traffic benchmarks under time-of-day shifts, showing architecture- and shift-dependent improvements in robustness.

Critical strengths

The paper provides a clear theoretical analysis of shortcut contamination and evaluates a simple, architecture-agnostic intervention on synthetic and real-world traffic data.

Critical weaknesses
  1. The experiments do not clearly separate the effects of projection from those of persistence completion.
  2. The projection-placement ablation changes the training inputs, validation inputs, and output handling simultaneously, making it difficult to determine which change causes the performance differences.
  3. The paper contains only two figures, making the proposed method and experimental findings difficult to understand intuitively.
Questions for authors
  1. Can the authors clarify the novelty of the proposed projection-based decoupling strategy beyond removing dominant global components before graph learning?
  2. Can the authors provide a more controlled ablation study to separately assess the effects of projection and persistence completion?
2 comments

Thank you for highlighting the need to separate projection from persistence completion. We added a matched-output control and clarified the original ablation (Secs. 3.2 and 4.5; Appendix G.2).

1. Projection vs. completion

Addressed in: Eq. (25); Appendix G.2, Tables 16–17.

Let P=I−QP = I - Q, Q=qq⊤Q = q q^\top, f0f_0 be Original, and hh be trained on projected inputs/targets. We compare:

  • Original: A=f0(X)A = f_0(X).
  • Matched completion: B=Pf0(X)+QxtB = P f_0(X) + Q x_t.
  • Proposed soft output: C=h(PX)+QxtC = h(PX) + Q x_t.
  • Diagnostic hard output: D=Ph(PX)+QxtD = P h(PX) + Q x_t.

BB and DD use identical completion.

For AGCRN on synthetic data with N=32N = 32 observed variables (one node per variable), five-seed MeanProj OOD MSE is:

ConditionOOD MSE
Original (AA)1.34422±0.098041.34422 \pm 0.09804
Matched (BB)1.09065±0.006891.09065 \pm 0.00689
Proposed (CC)1.06558±0.005681.06558 \pm 0.00568
Hard (DD)1.06519±0.005841.06519 \pm 0.00584

Output replacement alone reproduces 90.9% of the mean AA-to-DD OOD improvement (an arithmetic fraction, not causal mediation).

Crucially, DD still outperforms matched BB by 0.02546±0.001810.02546 \pm 0.00181 OOD MSE, in 5/5 paired seeds. The ID difference is 0.00967±0.001660.00967 \pm 0.00166 (5/5). PCA gives a similar matched OOD difference of 0.02532±0.001760.02532 \pm 0.00176 (5/5).

Thus, completion explains much, but not all, of the gain. The remainder includes projected inputs/targets, fitting, and checkpoint selection; we do not attribute it wholly to graph changes.

2. Placement ablation

Addressed in: Sec. 4.4; Appendix B.2; fixed-graph results in Appendix G.3.

The earlier Train+Eval versus Eval-only comparison changes training, validation, and output handling simultaneously. We now explicitly describe it as a pipeline comparison, not an isolated placement effect.

The new control addresses the completion confound. Additionally, varying only graph source under common synthetic forecasting conditions yields a modest MeanProj-derived graph OOD advantage of 0.00336±0.001640.00336 \pm 0.00164 over the raw-trained graph (five paired seeds). Traffic fixed-graph results are mixed, so this is not a general graph-mediated explanation.

3. Novelty

Addressed in: Sec. 5.

Our contribution is the exact support-risk advantage and its strictness conditions (Theorem 1), plus the loading-shift excess-risk analysis (Theorem 2), not PCA or mean removal. We distinguish these aims from factor-adjusted network estimation such as FNETS [1].

In the controlled two-factor setting, removing two PCs improves structural ranking and forecasting more than rank-one removal (Appendix G.4). This does not establish automatic nuisance-subspace identification.

4. Visual clarity

Addressed in: Fig. 2; Eq. (25); Appendix G.5.

Figure 2 now shows projected inputs and targets, the training loss, and the separate causal-completion branch. Eq. (25) defines all four output controls together, and a rotation plot is provided in Appendix G.5.

Reference: [1] Barigozzi, Cho, and Owens (2024). “FNETS: Factor-Adjusted Network Estimation and Forecasting for High-Dimensional Time Series.” Journal of Business & Economic Statistics, 42(3), 890–902.

Thank you for responding carefully to the review comments and supplementing the experiments, which resolved most of my doubts. The authors did not avoid unfavorable results, but actively acknowledged the limited contribution of graph structure and inconsistent performance of real traffic data, and correspondingly narrowed their claims. This attitude of candidly discussing the limitations of the experiment deserves some recognition, and the newly added matched-completion and fixed-graph experiments also enhance the argument to a certain extent.

However, Projection Placement Ablation still fails to fully isolate the influence of different factors, and the contribution of graph structure itself to performance improvement is still not clear enough. In addition, there are still some shortcomings in the visualization and overall presentation of the paper, which may require readers to spend more time understanding it. As the original readability issue has not been fully resolved, I will keep the presentation score at 1. Considering the author's supplementary work deserves recognition, I am willing to raise the overall rating to 3.

Reviewer_AHMfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper mainly analyzes why and when the graph learned from time-series for their future values forecasting helps and not helps for the forecasting. Based on the finding, the paper further propose a decoupling strategy to remove global components to enhance locality impacts.

Critical strengths
  1. There are many mathematical derivation and supporting materials for the paper.
Critical weaknesses
  1. Writing needs to be much improved. It's not clear what is trying to be proved with Theorem 1 and Proposition 1. It seems that one objective is to prove Forecasting loss can prefer shortcut graphs, which by itself is not useful, and a more useful formulation is whether it always prefer shortcut graph or under what conditions and to what degree. But even if that's the objective, it's not clear what is the expected equation and how to get there. Also see the comments in the end.
  • Line 095, M should be defined first before using
  • The meanings of variables should be described, e.g., line 097, Si(M) seems to be the set of ancestor nodes to i.
  • what does unstricted mean in line 132?
  • fontsize is too small.
  • used baselies in line 357 should be cited.
Questions for authors
  1. Is the paper essentially trying to separate out the global vs. local impacts of the graph on time series? If so, this has been widely studied.
3 comments

Thank you for the detailed comments. Theorem 1 shows precisely when—and by how much—access to variables outside the true local support reduces forecasting risk through a persistent global factor. It therefore explains why a graph optimized for prediction need not faithfully represent local dependencies. We now state this interpretation before the technical result (Sec. 2, pp. 2–3).

1. Theorem 1: Conditions and magnitude

Addressed in: Theorem 1, Eq. (10), Appendix A.3.

We compare optimal population risks for the local mask M⋆M^\star and its shortcut augmentation MLM_L. The exact gap is

J(M⋆)−J(ML)=∑i:Li≠0bi2Var⁡(gt∣xRi,t),b=ρL−BL,\mathcal{J}(M^\star) - \mathcal{J}(M_L) = \sum_{i: L_i \ne 0} b_i^2 \operatorname{Var}(g_t \mid x_{R_i, t}), \quad b = \rho L - B L,

where Ri={i}∪Si(M⋆)R_i = \{i\} \cup S_i(M^\star). Under the Gaussian assumptions, it is strictly positive if and only if some node has Li≠0L_i \ne 0, bi≠0b_i \ne 0, and nonzero factor loading outside RiR_i.

Thus, the formula quantifies the gain through the predictive contrast bib_i and the factor uncertainty left by local observations. We agree that nested classes alone imply a nonnegative gap; our result supplies its exact value and strictness conditions. The revised text gives a three-step proof roadmap.

Because the Gaussian conditional mean is linear, the same Bayes-risk result holds for all square-integrable predictors with the same coordinate access (Corollary 1, Appendix A.8). This is not a nonlinear training-convergence theorem.

2. Proposition 1: Purpose

Addressed in: Sec. 2, p. 5; Appendix A.2.

For xt=zt+Lgtx_t = z_t + L g_t with L⊤zt=0L^\top z_t = 0 and ∥L∥=1\|L\| = 1, exact projection satisfies PLxt=ztP_L x_t = z_t. For estimated qq,

Pqxt=zt−qq⊤zt+PqLgt,P_q x_t = z_t - q q^\top z_t + P_q L g_t,

separating loss of local information from remaining global leakage. We renamed the result Latent-signal preservation under exact projection. It does not prove graph recovery.

3. Novelty relative to global/local separation

Addressed in: Secs. 1 and 5.

Mean removal, PCA, and factor adjustment are established; e.g., FNETS [1] combines factor adjustment with sparse VAR estimation. Our contribution is the support-dependent forecasting incentive and loading-shift risk analysis, not component removal itself. The revised related-work discussion makes this distinction explicit.

4. Additional empirical evidence

Addressed in: Sec. 4.5; Appendix G.4.

In a synthetic benchmark with N=64N = 64 observed variables (one node per variable) and two factors, rank-two PCA raises structural AUPRC from 0.0601±0.02260.0601 \pm 0.0226 to 0.6278±0.05470.6278 \pm 0.0547 and lowers OOD MSE from 1.5612±0.09721.5612 \pm 0.0972 to 1.0818±0.02081.0818 \pm 0.0208 (five seeds). This evaluates adjacency ranking against generating support, not causal recovery.

5. Presentation corrections

Addressed in: Secs. 2 and 4.1; Figs. 1–2.

We define MM before use, explain Mij=1M_{ij} = 1 as j→ij \to i and Si(M)S_i(M) as direct incoming neighbors, correct “unstricted” to “unrestricted,” cite baselines on first mention, and redraw the figures with clearer labels. The original derivations remain.

Reference: [1] Barigozzi, Cho, and Owens (2024). “FNETS: Factor-Adjusted Network Estimation and Forecasting for High-Dimensional Time Series.” Journal of Business & Economic Statistics, 42(3), 890–902.

I thank authors for revising their draft to address my comments. Since I don't have access to the revised draft. I will have to retain my previous rating.

Thank you for your follow-up. We have already submitted the revised version, so the downloadable PDF now reflects the latest version of our paper.

Comments on the paper

1 comment

Thank you for the follow-up. Our central point is that a graph can improve forecasts by collecting information about a shared signal, even when its additional edges do not represent local interactions. We understand that quantifying this incentive is different from explaining which graph a practical optimizer learns. Below, we summarize what the theory and additional experiments establish about this connection.

1. What the population assumptions allow us to explain.

Consider a sensor affected by both local dynamics and a persistent common factor. A distant sensor can help estimate that factor even without a direct local interaction. Theorem 1 quantifies the forecasting benefit of this additional access. Under its stated one-step Gaussian model, it gives the exact optimal-risk difference and the conditions for a strict benefit: the target is factor-exposed, local observations leave uncertainty about that factor, and the factor adds predictive information beyond the local dynamics.

The Gaussian assumption makes this difference analytically computable. Because the conditional mean is already linear, allowing nonlinear functions of exactly the same current-state coordinates does not remove the gap (Corollary 1). Thus, the result is not merely an artifact of choosing linear predictors. It still retains the population and information-access assumptions and does not show that a neural optimizer selects the augmented graph. The revised text explains the mechanism and proof steps directly after Theorem 1; detailed proofs are in Appendices A.3 and A.8.

2. A checkable connection to actual network updates.

Proposition 3 asks a different question: when a feasible graph-related parameter update changes the network's predictions, does that change reduce the current prediction error? It expresses the squared-loss derivative through the alignment between the residual and the first-order output change, computed by a Jacobian-vector product (JVP). Positive alignment implies a sufficiently small loss-decreasing step under local smoothness. This result requires differentiability and integrability, but not the Gaussian or orthogonal local/global model. It is a conditional statement in the implemented parameterization, not a guarantee that training chooses a particular direction.

We tested two shared-component-related directions at trained AGCRN checkpoints. At the smallest step, training loss decreased in 4/5 seeds for each direction, whereas OOD loss increased in all 30 tested direction/step/seed combinations. These results are consistent with a local train/OOD trade-off, not evidence of the optimizer's actual trajectory. Since the embedding also controls forecasting weights, the test is not graph-only. Appendix A.9 gives the conditions and proof; Appendix G.6 reports both favorable and unfavorable update results.

3. Evidence beyond exact local/global separation.

We also tested whether the structural effect appears without constructing data to satisfy exact orthogonality. The added AGCRN benchmark has 64 observed variables (one node per variable) and two common factors. Its local process is an unprojected sparse VAR, so exact separation by projection is not imposed.

Removing two training principal components improves structural AP from 0.0601±0.02260.0601 \pm 0.0226 to 0.6278±0.05470.6278 \pm 0.0547 and OOD MSE from 1.5612±0.09721.5612 \pm 0.0972 to 1.0818±0.02081.0818 \pm 0.0208 (mean±SD\text{mean} \pm \text{SD}, five seeds). Here, structural AP measures how well symmetrized learned adjacency ranks node pairs against the known undirected local support. Both metrics also improve over rank-one removal in all five paired runs. To control output reconstruction, we give Original the same last-observation forecast in the removed subspace. Against this matched baseline, the projected-trained model with hard output projection retains an OOD MSE advantage of 0.06577±0.012790.06577 \pm 0.01279 in all five seeds.

This provides empirical evidence in a nonlinear learner beyond the exact orthogonal-local setting, not a theorem without assumptions. The removal rank is supplied from the known generating rank; structural ranking is not causal recovery, and the remaining forecasting gain is not identified as graph-only. Protocols are described in Appendix G.1. Appendix G.4 reports forecasting and structural results (Tables 22–23), matched-completion differences (Table 24), and direct comparisons with rank-one removal (Table 25).

In summary, the paper provides an exact account of the forecasting incentive, a conditional test for feasible network updates, and empirical evidence outside the exact-separation setting. These strengthen the connection, while leaving a general explanation of the graphs selected by nonlinear training unresolved.

Could you clarify which particular assumption or missing step remains most consequential to your assessment?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.