Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Causal Adjacency Learning for Spatiotemporal Prediction Over Graphs

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Learning causal adjacency improves out-of-distribution graph forecasting.

desk verdict A sensible SyPI-to-adjacency extension, but the experiments don't isolate causality from pruning, so the central claim is overreach. read the letter →

arxiv 2411.16142 v1 pith:PF3XJQ7S submitted 2024-11-25 cs.LG stat.ML

Zhaobin Mo , Qingyuan Liu , Baohua Yan , Longxiang Zhang , Xuan Di This is my paper · ORCID
classification cs.LGstat.ML
keywords causaladjacencylearningspatiotemporalpredictiongraphneuralnetworksout-of-distributiongeneralizationconditionalindependencetestSyPIhumanmobilityforecastingCOVID-19
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the adjacency matrix fed into a spatiotemporal graph predictor should encode causal relations among nodes rather than memorized similarities. It introduces Causal Adjacency Learning (CAL), a procedure that runs temporal conditional independence tests on the training time series and filters the results with the SyPI algorithm to build a causal adjacency matrix, which is then used in a downstream graph convolutional network. On real-world records of weekly visits to 172 ZIP codes during the COVID-19 pandemic, CAL's matrix improves mobility forecasts on a 16-week out-of-distribution test window that includes the February 2022 Omicron surge, compared with distance-, correlation-, and attention-based matrices. The finding matters because conventional adjacency matrices are fitted to the training distribution and tend to fail when the test distribution shifts; a transferable causal matrix would make graph forecasting robust to shocks without re-estimating relations at test time.

What carries the argument

The load-bearing object is the causal adjacency matrix $A_{CAU}$, built by an upstream Causal Adjacency Learning module. For each node, CAL first restricts candidate causes to the $M$ neighbors with highest Pearson correlation (spatial pre-selection), then applies a kernel-based conditional independence test to the lagged time series of each candidate against the target, conditioning on all other candidates and the target's own past. The SyPI algorithm filters the resulting dependencies to keep only those that survive two thresholded tests, removing spurious relations. This matrix is normalized into a graph Laplacian and used in a $K$-order Chebyshev polynomial graph convolution in the downstream spatiotemporal GCN; the claim is that this causal structure, being invariant across distributions, transfers to out-of-distribution test data where correlation and attention structures do not.

What would settle it

Re-estimate the causal adjacency matrix on the 16-week out-of-distribution test window alone and compare it with the matrix learned from the first 50 weeks. If the two matrices differ substantially, or if a GCN using the test-window matrix beats the fixed training matrix on the test data, then the assumption that causal relations are constant across the distribution shift is violated and the transferability claim fails.

Watch

Extended reading notes

Core claim

The central claim is that causal relations among nodes in a spatiotemporal graph can be discovered once, encoded in an adjacency matrix, and transferred to a downstream prediction task, where they improve out-of-distribution performance even though the downstream model never performs causal inference. CAL tests, for each target node, whether a candidate neighbor's lagged time series is conditionally independent of the target given all other candidate series, using a kernel-based conditional independence test; the SyPI algorithm then removes false positives and negatives, and a correlation-based pre-selection step keeps the computation at $O(Mn)$ instead of $O(n^2)$. When the learned matrix $A_{CAU}$ is used in a graph convolutional network with Chebyshev spectral filters and a temporal convolution layer, it yields lower RMSE and MAE than distance, correlation, and attention matrices on all prediction horizons from one to four weeks ahead, with an average RMSE reduction over the four-week horizon of 24.71% relative to the distance matrix. The paper also reports that the matrix is sparser than the correlation and attention baselines and that row and column sums give an interpretable picture of regional influence, with Manhattan exerting the largest outward influence and Staten Island the most influenced.

Load-bearing premise

The causal relations among regions learned from the first 50 weeks of training data remain unchanged during the February 2022 Omicron surge, so the fixed adjacency matrix still describes how regions influence each other on the out-of-distribution test window.

Editorial extensions

If this is right

  • A single upstream causal learning pass is enough: the learned adjacency matrix can be reused by any downstream STPG model without re-running causal discovery.
  • Forecasts on the out-of-distribution window improve across all horizons; average RMSE over the four-week horizon falls by 24.71% versus the distance matrix and by 10.98% versus the correlation matrix.
  • The sparser causal matrix (9.67% fewer edges than correlation, 15.98% fewer than attention) reduces computing cost while retaining predictive signal, suggesting many correlation-based edges are noise.
  • The method surfaces interpretable structure: regional influence patterns, such as Manhattan's dominant outgoing influence, can be read directly from the matrix.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves untested whether the learned causal structure remains stable over longer horizons or across further shocks; re-estimating $A_{CAU}$ on rolling windows and measuring drift would be a direct stress test.
  • The same upstream/downstream recipe could extend to other STPG domains with sudden shifts, such as traffic incidents, disease spread, or energy demand, but only the mobility case is demonstrated here.
  • Because Pearson correlation is used for spatial pre-selection, candidates are pre-filtered by a linear criterion; a nonlinear pre-selection step might admit different edges and change which causal relations survive SyPI.
  • The downstream model is a deliberately simple GCN to isolate the effect of the adjacency matrix; testing the same matrix in more expressive predictors would show whether the transferable benefit persists across model families.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes Causal Adjacency Learning (CAL), an upstream module that discovers causal relations among nodes in spatiotemporal graph data using a kernel-based conditional independence test followed by the SyPI algorithm, and encodes the results into an adjacency matrix for a downstream spatiotemporal GCN. The method is evaluated on SafeGraph weekly mobility data for 172 ZIP codes in the New York area over 90 weeks, with the last 16 weeks (including the February 2022 Omicron surge) treated as out-of-distribution test data. The paper reports that the CAL adjacency matrix improves RMSE/MAE compared with distance-, correlation-, and attention-based adjacency matrices, and it claims that the learned matrix captures causal relations that transfer to the downstream task.

Significance. If fully supported, the paper would make a useful contribution: a transferable, interpretable causal adjacency matrix that improves out-of-distribution spatiotemporal prediction is a valuable module for traffic and mobility forecasting. The idea of extending temporal causal feature selection to spatiotemporal graphs is reasonable and timely. The authors also provide a clear algorithmic description and a geospatial visualization that is suggestive of interpretability. However, the current experimental design does not isolate the causal mechanism from sparsity or edge-selection effects, and no ground-truth validation of the causal edges is provided. The significance is therefore not yet established beyond a promising initial study.

major comments (5)
  1. [Section V.B.2, Table II] The attention baseline is not an independent attention-derived graph: it is forced to keep the same edges as the SyPI adjacency matrix and only varies the weights. Thus the comparison to Attention in Table II does not test whether attention-based edge selection is inferior to causal edge selection; it only compares edge weights on a fixed edge set. Please construct an attention baseline with its own top-k edge selection and also match edge counts across all baselines.
  2. [Section V.C.1, Table II] There is no control that randomizes or non-causally selects edges while preserving the sparsity and edge count of CAL. Because the CAL adjacency differs from every baseline in edge-selection rule, edge count, and edge weights, the reported 10–50% RMSE improvements cannot be attributed to causal content; they may arise from the particular edge support or from increased sparsity. Add at least a random-edge control with the same number of edges and a correlation-based baseline with edge count matched to CAL.
  3. [Section IV.A.3 and Section V.B] The spatial pre-selection size M is never reported in the experiment section, although it is a tunable hyperparameter that determines the candidate edge set and hence the final sparsity of the causal adjacency matrix. Without reporting M and the resulting edge count for each method, the comparison in Table II and the sparsity discussion in Section V.C.1 are not reproducible.
  4. [Section II.B and Section V.B.3] The stability assumption that causal relations are invariant across distributions is adopted without testing. The matrix is learned from the first 50 weeks and applied to the last 16 weeks containing the February 2022 Omicron surge; no experiment checks whether the recovered edges would be stable across windows or whether a matrix estimated on the OOD period would differ. Please validate on synthetic data with known ground truth and/or report rolling-window stability results.
  5. [Section V.C.1, Table II] The text says “Given different random seeds,” but Table II reports only point estimates without standard deviations, confidence intervals, or significance tests. Since some gains are small (e.g., at T+3, CAL is not the best method), the reported improvements may be within run-to-run noise. Report mean ± standard deviation over seeds and run a paired significance test.
minor comments (6)
  1. [Section I] In the paper-organization paragraph, “real-wi experiments and results” is a typo; it should read “real-world experiments and results.”
  2. [Section V.B] The sentence mentioning “images from the dataset and vehicle speed data” appears to be a leftover from another dataset description; the SafeGraph dataset used here contains mobility counts and contextual variables, not images or vehicle speed data.
  3. [Algorithm 2] The notation in Algorithm 2 is inconsistent: “min lags” and “w” are used interchangeably, and the conditioning set “Si” inside the loop is not explicitly defined for each i. Please clarify the algorithm steps and variable definitions.
  4. [Section III.A, Eq. (2)] The comma after the display equation is misplaced, and the time-series length T used in the correlation calculation should be defined explicitly.
  5. [Sections I and II.B] The claim of being “the first to consider temporal dimensions” in STPG is too strong, since [12] and [13] operate on temporal chunks; please temper the claim to “first to consider within-temporal causal structure.”
  6. [Section V.D, Fig. 3] Please clarify what the color scale and the row/column aggregations represent (e.g., sum of causal weights versus number of causal edges).

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity; the CAL adjacency is learned on training data and tested on held-out OOD weeks, with only minor non-load-bearing self-citations.

full rationale

The proposed derivation is not circular by construction. The CAL adjacency matrix is produced by the SyPI algorithm (cited to external work [26]) on the first 50 weeks of training data, then fixed as a constant input to a GCN that is trained on the first 74 weeks and evaluated on the last 16 held-out weeks (Section V.B.3). The learned edges are not derived from the OOD test labels, so the reported RMSE improvements in Table II are an empirical outcome rather than a fitted quantity renamed as a prediction. The self-citations ([14], [17], [18], [19]) appear only in background and motivation; none is invoked as a theorem that forces the result, and the core causal-selection machinery is external ([26]). The lack of a sparsity-matched baseline and the untested stability of causal relations across the Omicron surge are experimental-design and assumption concerns, not construction-level circularity. Accordingly, the circularity score is low (2), reflecting minor self-reference rather than a circular derivation.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The contribution is a representation (causal adjacency matrix) built from existing causal discovery machinery. The central claim depends on unverified invariance and pre-selection assumptions, and on hyperparameters that are partly unreported.

free parameters (7)
  • M (spatial pre-selection size)
    Number of top-correlation candidate causes kept per node in Section IV.A.3; the paper says it is a tunable hyperparameter but never reports the value used in experiments.
  • threshold1 = 0.1
    First-stage p-value threshold in SyPI (Algorithm 2), set in Section V.B.1.
  • threshold2 = 0.08
    Second-stage p-value threshold in SyPI (Algorithm 2), set in Section V.B.1.
  • KCIT kernel sigma = 10
    Gaussian kernel scale for the conditional independence test, Section V.B.1.
  • KCIT kernel width sigma_x = 0.8
    Feature kernel width for KCIT, Section V.B.1.
  • SyPI time series length = 50 weeks
    Length of the historical window used for causal discovery, Section V.B.1.
  • min_lag = 1
    Minimum time lag between candidate cause and target in SyPI, Section V.B.1.
assumptions (4)
  • domain assumption Causal relations between nodes remain invariant across the training distribution and the OOD test window.
    Invoked in Section II.B and implicitly in Section V.B.3 when transferring the adjacency matrix learned on the first 50 weeks to the last 16 weeks.
  • domain assumption The SyPI algorithm's assumptions hold for this dataset (e.g., its treatment of latent common causes and causal sufficiency).
    Adopted without verification in Section IV.A.2; the paper cites [26] but does not state or test the assumptions on SafeGraph data.
  • ad hoc to paper Restricting candidate causes to the M most-correlated nodes does not exclude true causal parents.
    Added in Section IV.A.3 to reduce computational cost; a true cause with low linear correlation could be dropped, invalidating the causal adjacency.
  • domain assumption Weekly temporal aggregation preserves the causal timescales of mobility.
    Section V.A aggregates SafeGraph to weeks and Section V.B.1 sets min lag to 1, assuming causality operates at weekly resolution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Adjacency Learning for Spatiotemporal Prediction Over Graphs." pith.science (2026). https://pith.science/paper/PF3XJQ7S

@misc{pith2026241116142,
  author       = {Pith},
  title        = {Pith review of: Causal Adjacency Learning for Spatiotemporal Prediction Over Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PF3XJQ7S}},
  note         = {Machine review of arXiv:2411.16142}
}
read the original abstract

Spatiotemporal prediction over graphs (STPG) is crucial for transportation systems. In existing STPG models, an adjacency matrix is an important component that captures the relations among nodes over graphs. However, most studies calculate the adjacency matrix by directly memorizing the data, such as distance- and correlation-based matrices. These adjacency matrices do not consider potential pattern shift for the test data, and may result in suboptimal performance if the test data has a different distribution from the training one. This issue is known as the Out-of-Distribution generalization problem. To address this issue, in this paper we propose a Causal Adjacency Learning (CAL) method to discover causal relations over graphs. The learned causal adjacency matrix is evaluated on a downstream spatiotemporal prediction task using real-world graph data. Results demonstrate that our proposed adjacency matrix can capture the causal relations, and using our learned adjacency matrix can enhance prediction performance on the OOD test data, even though causal learning is not conducted in the downstream task.

Figures

Figures reproduced from arXiv: 2411.16142 by the authors.

Figure 1
Figure 1. Framework of the upstream CAL and the downstream spatiotemporal GCN for the problem of STPG. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. GCN prediction of future mobility based on different [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Row (left) and column (right) aggregation of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Causal Learning for Heterogeneous Subgroups Based on Nonlinear Causal Kernel Clustering

    cs.LG 2025-01 reject novelty 4.0 of 10

    A kernel clustering method with a u-centered sample mapping is proposed to discover heterogeneous subgroups with different causal structures.

Reference graph

Works this paper leans on

26 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [12]

    Dynamic graph neural networks under spatio-temporal distribution shift,

    Z. Zhang, X. Wang, Z. Zhang, H. Li, Z. Qin, and W. Zhu, “Dynamic graph neural networks under spatio-temporal distribution shift,” Ad- vances in neural information processing systems , vol. 35, pp. 6074– 6089, 2022

    2022

  2. [13]

    Deciphering spatio-temporal graph forecasting: A causal lens and treatment,

    Y . Xia, Y . Liang, H. Wen, X. Liu, K. Wang, Z. Zhou, and R. Zimmer- mann, “Deciphering spatio-temporal graph forecasting: A causal lens and treatment,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  3. [1]

    St- gin: An uncertainty quantification approach in traffic data imputation with spatio-temporal graph attention and bidirectional recurrent united neural networks,

    Z. Wang, D. Zhuang, Y . Li, J. Zhao, P. Sun, S. Wang, and Y . Hu, “St- gin: An uncertainty quantification approach in traffic data imputation with spatio-temporal graph attention and bidirectional recurrent united neural networks,” in 2023 IEEE 26th International Conference on Intelligent Transportation Systems (ITSC) . IEEE, 2023, pp. 1454– 1459

    work page 2023

  4. [2]

    St-mlp: A cascaded spatio-temporal linear framework with channel-independence strategy for traffic forecasting,

    Z. Wang, Y . Nie, P. Sun, N. H. Nguyen, J. Mulvey, and H. V . Poor, “St-mlp: A cascaded spatio-temporal linear framework with channel-independence strategy for traffic forecasting,” arXiv preprint arXiv:2308.07496, 2023

    arXiv 2023

  5. [3]

    Robust node classification on graphs: Jointly from bayesian label transition and topology-based label propa- gation,

    J. Zhuang and M. Al Hasan, “Robust node classification on graphs: Jointly from bayesian label transition and topology-based label propa- gation,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Mana gement , 2022, pp. 2795–2805

    work page 2022

  6. [4]

    How does bayesian noisy self- supervision defend graph convolutional networks?

    J. Zhuang and M. A. Hasan, “How does bayesian noisy self- supervision defend graph convolutional networks?” Neural Processing Letters, vol. 54, no. 4, pp. 2997–3018, 2022

    work page 2022

  7. [5]

    Defending graph convolutional networks against dynamic graph perturbations via bayesian self-supervision,

    J. Zhuang and M. Al Hasan, “Defending graph convolutional networks against dynamic graph perturbations via bayesian self-supervision,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 4, 2022, pp. 4405–4413

    work page 2022

  8. [6]

    Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,

    B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,” arXiv preprint arXiv:1709.04875, 2017

    arXiv 2017

Show all 26 references
  1. [7]

    Invariant graph neural network for out-of-distribution nodes,

    Z. Chen, Y . Gong, L. Yang, J. Zhang, W. Zhang, S. He, and X. Zhang, “Invariant graph neural network for out-of-distribution nodes,” in Proceedings of the 2023 15th International Conference on Machine Learning and Computing , 2023, pp. 192–196

    2023

  2. [8]

    Adaptive graph convolutional recurrent network for traffic forecasting,

    L. Bai, L. Yao, C. Li, X. Wang, and C. Wang, “Adaptive graph convolutional recurrent network for traffic forecasting,” Advances in neural information processing systems , vol. 33, pp. 17 804–17 815, 2020

    2020

  3. [9]

    Attention based spatial-temporal graph convolutional networks for traffic flow forecast- ing,

    S. Guo, Y . Lin, N. Feng, C. Song, and H. Wan, “Attention based spatial-temporal graph convolutional networks for traffic flow forecast- ing,” in Proceedings of the AAAI conference on artificial intelligence , vol. 33, no. 01, 2019, pp. 922–929

    2019

  4. [10]

    Event-aware multimodal mobility nowcasting,

    Z. Wang, R. Jiang, H. Xue, F. D. Salim, X. Song, and R. Shibasaki, “Event-aware multimodal mobility nowcasting,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36, no. 4, 2022, pp. 4228–4236

    2022

  5. [11]

    Spatiotemporal adaptive gated graph convolution network for urban traffic flow forecasting,

    B. Lu, X. Gan, H. Jin, L. Fu, and H. Zhang, “Spatiotemporal adaptive gated graph convolution network for urban traffic flow forecasting,” in Proceedings of the 29th ACM International conference on information & knowledge management , 2020, pp. 1025–1034

    2020

  6. [14]

    Infostgcan: An information-maximizing spatial- temporal graph convolutional attention network for heterogeneous human trajectory prediction,

    K. Ruan and X. Di, “Infostgcan: An information-maximizing spatial- temporal graph convolutional attention network for heterogeneous human trajectory prediction,” Computers, vol. 13, no. 6, p. 151, 2024

    2024

  7. [15]

    Spatial-temporal fusion graph neural networks for traffic flow forecasting,

    M. Li and Z. Zhu, “Spatial-temporal fusion graph neural networks for traffic flow forecasting,” in Proceedings of the AAAI conference on artificial intelligence, vol. 35, no. 5, 2021, pp. 4189–4196

    2021

  8. [16]

    Trafficgan: Network-scale deep traffic prediction with generative adversarial nets,

    Y . Zhang, S. Wang, B. Chen, J. Cao, and Z. Huang, “Trafficgan: Network-scale deep traffic prediction with generative adversarial nets,” IEEE Transactions on Intelligent Transportation Systems , vol. 22, no. 1, pp. 219–230, 2019

    2019

  9. [17]

    Cross-and context-aware attention based spatial-temporal graph convolutional networks for human mobility prediction,

    Z. Mo, H. Xiang, and X. Di, “Cross-and context-aware attention based spatial-temporal graph convolutional networks for human mobility prediction,” ACM Transactions on Spatial Algorithms and Systems , 2024

    2024

  10. [18]

    Pi-neugode: Physics-informed graph neural ordinary differential equations for spatiotemporal trajectory prediction,

    Z. Mo, Y . Fu, and X. Di, “Pi-neugode: Physics-informed graph neural ordinary differential equations for spatiotemporal trajectory prediction,” in Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems , 2024, pp. 1418–1426

    2024

  11. [19]

    Causal imitation learn- ing via inverse reinforcement learning,

    K. Ruan, J. Zhang, X. Di, and E. Bareinboim, “Causal imitation learn- ing via inverse reinforcement learning,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  12. [20]

    Learning causally invariant representations for out-of- distribution generalization on graphs,

    Y . Chen, Y . Zhang, Y . Bian, H. Yang, M. Kaili, B. Xie, T. Liu, B. Han, and J. Cheng, “Learning causally invariant representations for out-of- distribution generalization on graphs,” Advances in Neural Information Processing Systems, vol. 35, pp. 22 131–22 148, 2022

    2022

  13. [21]

    Debiasing graph neural networks via learning disentangled causal substructure,

    S. Fan, X. Wang, Y . Mo, C. Shi, and J. Tang, “Debiasing graph neural networks via learning disentangled causal substructure,” Advances in Neural Information Processing Systems , vol. 35, pp. 24 934–24 946, 2022

    2022

  14. [22]

    Model-powered conditional independence test,

    R. Sen, A. T. Suresh, K. Shanmugam, A. G. Dimakis, and S. Shakkot- tai, “Model-powered conditional independence test,” Advances in neural information processing systems , vol. 30, 2017

    2017

  15. [23]

    Conditional independence test for weights-of-evidence modeling,

    F. P. Agterberg and Q. Cheng, “Conditional independence test for weights-of-evidence modeling,” Natural Resources Research, vol. 11, pp. 249–255, 2002

    2002

  16. [24]

    A permutation- based kernel conditional independence test

    G. Doran, K. Muandet, K. Zhang, and B. Sch ¨olkopf, “A permutation- based kernel conditional independence test.” in UAI, 2014, pp. 132– 141

    2014

  17. [25]

    Kernel-based conditional independence test and application in causal discovery,

    K. Zhang, J. Peters, D. Janzing, and B. Sch ¨olkopf, “Kernel-based conditional independence test and application in causal discovery,” arXiv preprint arXiv:1202.3775 , 2012

    2012 arXiv

  18. [26]

    Necessary and sufficient conditions for causal feature selection in time series with latent common causes,

    A. A. Mastakouri, B. Sch ¨olkopf, and D. Janzing, “Necessary and sufficient conditions for causal feature selection in time series with latent common causes,” in International Conference on Machine Learning. PMLR, 2021, pp. 7502–7511

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.