Pith. sign in

REVIEW 4 major objections 5 minor 23 references

diffIRM: A Diffusion-Augmented Invariant Risk Minimization Framework for Spatiotemporal Prediction over Graphs

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that diffIRM makes spatiotemporal graph predictors generalize to distribution shifts by combining a causal mask generator and a diffusion-based environment augmentor with an invariance penalty, and proves conditional…

desk verdict A promising STPG framework whose central identifiability theorem rests on an unproven condition that the mask-generator training objective arguably discourages; the empirical gains are real but under-supported. read the letter →

arxiv 2501.00305 v1 pith:HTBHM2SB submitted 2024-12-31 cs.LG

Zhaobin Mo , Haotian Xiang , Xuan Di This is my paper · ORCID
classification cs.LG
keywords out-of-distributiongeneralizationspatiotemporalpredictionovergraphsgraphneuralnetworksinvariantriskminimizationcausalfeatureidentifiabilitydiffusionmodelsdataaugmentationtraffic
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

The paper claims that out-of-distribution generalization in spatiotemporal prediction over graphs can be improved by combining two principles that graph-OOD methods usually keep apart: assuming invariant causal features exist, and diversifying training environments. diffIRM does this with a min-max game in which a causal mask generator identifies which node features are causal, while a graph-based diffusion model augments only the leftover environment features. The augmented data define an invariance penalty that regularizes the prediction model. A theorem states that, when the game reaches equilibrium with a correct causal mask, the loss exhibits causal feature identifiability: minimizing it drives the model onto invariant features. On SafeGraph, PeMS04, and PeMS08, diffIRM reports lower MAE and RMSE than ERM, IRM, REx, and InvRat, and its learned masks highlight features such as COVID-19 case counts.

What carries the argument

The central object is the min-max game between the causal mask generator $T_\phi$, an MLP that outputs a soft causal mask $M_{\mathrm{cau}}$, and the environment augmentor $G_\psi$, a GCN-based denoising diffusion model that outputs a stochastic environment mask $M_{\mathrm{env}}$. Augmented features are formed as $\tilde{X} = X \odot M_{\mathrm{cau}} + \hat{X} \odot (1 - M_{\mathrm{cau}})$, so causal coordinates are preserved while environment coordinates are diversified. The augmented data produce an invariance penalty $r(\theta)$ in Eq. 4.6, which compares the prediction loss of the shared model against environment-specific predictors; in implementation this is replaced with a first-order approximation. The full loss in Eq. 4.7 is what Theorem 1 proves to exhibit causal feature identifiability, meaning the loss attained when the model uses only causal features is strictly below the loss for any other feature subset.

What would settle it

Run diffIRM on a synthetic structural causal model with known causal features, such as $X_1 \to Y$ and $Y \to X_2$ with environment-dependent noise variance, train until convergence over many seeds, and inspect the learned mask; if the mask assigns non-negligible weight to $X_2$ or fails to cover $X_1$, then Condition 1 fails and the claimed identifiability does not hold in that regime.

Watch

Extended reading notes

Core claim

The central discovery is that the loss function in Eq. 4.7, a sum of an augmentation loss and an invariance penalty, has a causal-feature-identifiability property: under Condition 1, training the prediction model with this loss leads it to identify invariant features rather than spurious ones. Condition 1 requires that at equilibrium of the min-max game the causal mask generator outputs the mask $[\mathbb{1}_{d_v},\mathbb{0}_{d_s}]$, marking causal coordinates and unmarking environment coordinates. The paper proves Theorem 1 that this condition is sufficient for identifiability, given four assumptions about loss lower bounds and penalties, and proves a proposition that the condition is also necessary. Using data from a known structural causal model, the paper shows that the learned regressor approximates the true causal relationship; on real mobility and traffic data, diffIRM outperforms the compared baselines, works with different spatiotemporal backbones, and produces interpretable causal masks.

Load-bearing premise

The load-bearing premise is that the min-max game between the causal mask generator and the diffusion augmentor reaches an equilibrium in which the mask exactly marks causal features and leaves environment features unmarked; the paper supports this condition only with a heuristic argument, not a proof.

Editorial extensions

If this is right

  • If the causal-feature identifiability result holds in practice, diffIRM predictors should remain accurate when test environments shift away from the training distribution, because they rely only on features whose relation to the label is stable.
  • The learned causal mask becomes an interpretable diagnostic, showing which input features, such as confirmed COVID-19 case counts, drive predictions under distribution shift.
  • Because the framework is GNN-agnostic, the same augmentation and invariance recipe can be attached to any spatiotemporal backbone, as the paper demonstrates with ASTGCN, STGCN, and LSTM.
  • Since diffIRM generates environments rather than assuming them known, it can apply the environment-diversity principle to graph tasks where environment labels are unavailable, including spatiotemporal prediction.
  • The invariance penalty uses a moderate number of generated environments; the paper's ablation suggests five environments suffice, so the method can avoid the overhead of many expensive augmentation samples.

Reading between the lines

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

  • A natural extension, not tested in the paper, is to prove or empirically verify Condition 1 on a known structural causal model: if the min-max game does not reliably converge to the perfect causal mask, the theorem's conclusion would not apply to the trained model.
  • The paper keeps the graph topology fixed and augments only node features; an editorially suggested follow-up is to test whether edge-level augmentation of the adjacency matrix also helps environment diversity, since graph OOD is often driven by structural shifts.
  • The large performance gain on SafeGraph relative to the PEMS datasets suggests the method's value grows with the severity of distribution shift; a testable extension would be a benchmark where the spurious feature-label relationship flips sign in the test segment.
  • Because the invariance penalty is implemented by a first-order approximation, one could investigate whether the exact penalty, trained with all $K$ environment-specific predictors, changes the identifiability behavior observed in the reported experiments.
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

4 major / 5 minor

Summary. The paper proposes diffIRM, a framework for spatiotemporal prediction over graphs under distribution shift. The method combines a causal mask generator that is supposed to identify invariant (causal) features with a diffusion-based environment augmentor that diversifies spurious features, and then uses the augmented environments to define an invariance penalty for training the prediction model. The authors claim that training with their loss (Eq. 4.7) achieves causal feature identifiability (Theorem 1) and that the method outperforms ERM, IRM, REx, and InvRat on SafeGraph, PeMS04, and PeMS08, while also producing interpretable causal masks.

Significance. If the identifiability claim were established, the paper would make a useful contribution to OOD generalization for spatiotemporal graphs by explicitly combining invariance and environment-diversity principles, and the proposed diffusion-based augmentation with a learned causal mask is a sensible architectural idea. The empirical results on SafeGraph in particular are dramatic, and the interpretability visualizations (Fig. 10) are suggestive. However, the central theoretical result is load-bearing: the abstract and Sec. 4.3 claim that diffIRM identifies invariant features, yet the proof is conditional on Condition 1, which is essentially the same claim. The paper also ships a concrete implementation and algorithm, but the implementation does not appear to match the loss analyzed in the theorem. The empirical gains may be real, but the stated causal-identifiability and interpretability claims are not supported by the present evidence.

major comments (4)
  1. [Sec. 4.3, Theorem 1 and Condition 1] Theorem 1 is conditional on Condition 1, which states that at equilibrium the causal mask generator outputs the true mask M*=[1_{d_v},0_{d_s}], i.e., that the trained generator has already identified the causal features. This is precisely the conclusion the theorem is meant to establish, so the proof is circular. Appendix A.3 does not prove Condition 1; it offers only heuristic arguments about the optimal strategies for the augmentor and mask generator. The manuscript itself says "In the appendix, we will discuss how Condition 1 can be met," but the appendix does not deliver a proof. A theorem whose premise already assumes the identifiability result does not support the paper's central claim that diffIRM identifies invariant features.
  2. [Eqs. (4.1), (4.4); Algorithm 1, line 11; Appendix B] The proposed min-max game actively discourages Condition 1 from holding. For any fixed theta, setting M_cau to all ones makes tilde X equal X in Eq. (4.4) regardless of M_env, so the augmentation loss equals the original loss and the inner max is trivial. A well-fitted f_theta makes this loss small, so the all-ones mask is a natural global minimum of the inner game. The causal-ratio regularizer in Appendix B, with alpha=0.5, constrains only the fraction of ones in M_cau; it does not favor causal features over any other subset of the same size. Moreover, Algorithm 1 updates phi by descending only on the augmentation loss (line 11), not on the invariance penalty r(theta), so phi receives no gradient signal that would reward M*. Consequently, the premise of Theorem 1 is not merely unproven; the optimization dynamics point away from it.
  3. [Theorem 1 statement vs. Appendix A.1 proof] The theorem's quantitative conditions are inconsistent with the proof. Theorem 1 states lambda > (L*(Y)+2epsilon)/(delta*C - 4K*epsilon), but the derivation in Appendix A.1 yields lambda > (L*(Y)+2epsilon)/(delta*K*C - 4K*epsilon), with an extra factor K. Likewise, the theorem states epsilon < min{delta*C/(4K), gamma/(2+4*lambda*K)}, while the proof uses epsilon < delta*C/4. These discrepancies mean the stated bound does not follow from the proof. In addition, Assumptions 3 and 4 introduce constants delta and C that are never verified on any of the three datasets, so the numerical applicability of the theorem is unsupported even under Condition 1.
  4. [Eq. (4.6), Eq. (4.7), Appendix B] The theoretical result applies to the loss in Eq. (4.7) with the invariance penalty r(theta) defined in Eq. (4.6). However, the implementation replaces r(theta) with a first-order gradient approximation denoted r_hat in Appendix B, with no error analysis or proof that the approximation preserves the CFI property. The experiments and Algorithm 1 therefore train a different objective from the one analyzed in Theorem 1. This gap is not acknowledged in the main text, where Algorithm 1 is described as implementing Eq. (4.7).
minor comments (5)
  1. [Sec. 4.3.2] The section heading "Condition 1→CFI" appears to be a typo; the text correctly states that this section proves Condition 1 is necessary for CFI. The proposition statement in Sec. 4.3.2 also contains a typo: "Proposition 4.3.2 indices" should be "indicates."
  2. [Appendix A.2] The proof of the proposition introduces the feature set X_bar_v and the causal feature X_v^{k'} without precise definitions. In particular, the notation for the maximum over feature subsets is not well-formed, and it is unclear whether the maximum is over feature subsets that violate the invariance penalty. Please clarify the construction of X_bar_v.
  3. [Table 5 and Sec. 5.2] The reported improvements over baselines, especially on SafeGraph, are large, but no standard deviations, number of random seeds, or statistical significance tests are reported. Given that the differences between diffIRM and diffIRM− are sizable, error bars would help the reader assess stability.
  4. [Appendix B, r_hat] The formula for the approximated penalty r_hat has a malformed nested summation: it is written as a sum over k of an expectation of another sum over k divided by K, which reuses k in two nested summations and is dimensionally inconsistent. Please rewrite with distinct indices.
  5. [Algorithm 1, line 7] The update for M_{\psi,\phi}^{(k)} is written as "M_cau + M_env^(k) ⊕ 1-(M_cau)", which is ambiguous about operator precedence. A parenthesized expression such as "M_cau + M_env^(k) * (1 - M_cau)" would be clearer, and should match Eq. (4.4).

Circularity Check

1 steps flagged · score 7.0 of 10

The theoretical identifiability claim (Theorem 1) depends on Condition 1, which already asserts that the mask generator identifies causal features; Appendix A.3 gives only a heuristic, so the central claim reduces to its own premise.

  1. self definitional [Sec. 4.3, Condition 1 and Theorem 1 (Sec. 4.3.1); Appendix A.3]
    "Condition 1. (CFI Condition) When the min-max game (Eq. 4.1) between the mask generator Tφ and the environment augmentor Gψ reaches its equilibrium, the causal mask generator is able to identify causal features Xcau. ... Our proof will focus on proving that Condition 1 is both the sufficient and necessary condition for diffIRM to achieve CFI. ... Theorem 1. If λ> ... we conclude that L∗cau < L∗env+. Thus, training the prediction model fθ with Eq. 4.7 leads to the identification of invariant features."

    The advertised result—that training fθ with Eq. 4.7 identifies invariant features—is proven only under Condition 1, which already asserts that the trained causal mask generator is able to identify causal features Xcau. The paper itself calls Condition 1 the 'CFI Condition' and states that it is both sufficient and necessary for CFI, so the identifiability conclusion is equivalent to this unproven premise. Appendix A.3 does not prove Condition 1; it only argues heuristically that 'to lower the objective function, the causal mask generator is trained to find causal features Xcau.' Thus the hard part of the claim—showing that the min-max equilibrium yields the true causal mask—is assumed rather than derived, and Theorem 1 restates that assumption as a theoretical result.

full rationale

The paper's core theoretical claim is conditional: if the min-max game reaches an equilibrium in which the causal mask generator outputs the true mask M*, then Eq. 4.7 ranks the causal feature subset below any other subset; and the Proposition makes Condition 1 necessary as well. Hence the entire theoretical content of the CFI result is equivalent to Condition 1. Condition 1, however, is never proven. Appendix A.3 is explicitly a 'Discussion' and gives intuitive arguments rather than a proof of existence, uniqueness, or convergence to the true mask. The situation is exacerbated by the objective's incentives: for fixed θ, the all-ones mask makes X~ = X in Eq. 4.4, leaving the augmentor inert and the inner max equal to the original loss; the causal-feature-ratio regularizer in Appendix B only constrains the fraction of masked entries, not which features are causal; and Algorithm 1 updates φ by descending only on the augmentation loss (line 11), not on the invariance penalty, so φ receives no direct gradient signal rewarding M*. This makes the derivation question-begging rather than self-contained. The empirical comparisons against ERM/IRM/REx/InvRat and the ablations are externally falsifiable and are not circular; that independent empirical content keeps the score below 8-9. There is no load-bearing self-citation chain here, and the motivating numerical example is a sanity check rather than a proof.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The theoretical result is purchased almost entirely by assumptions. Condition 1 states that the causal mask generator is correct at equilibrium, which is the very identifiability property to be established. Assumptions 2-4 introduce gamma, delta, and C as constants that are never verified for the experimental datasets. Hyperparameters lambda, K, and alpha are selected by hand or validation. No new physical entities are introduced.

free parameters (3)
  • lambda (invariance penalty weight) = tuned on validation set, value not reported
    Controls balance between augmentation loss and invariance penalty in Eq. 4.7; no sensitivity analysis is provided beyond the K ablation.
  • K (number of augmented environments) = 5
    Selected from ablation in Table 6; 3 and 7 environments perform worse, and 9 gives no clear gain.
  • alpha (causal feature ratio regularization) = 0.5
    Set as an initial estimate of the causal feature ratio in Appendix B to stabilize causal mask training.
assumptions (6)
  • ad hoc to paper Condition 1: at equilibrium, the causal mask generator produces the correct causal mask Mcau = [1_dv, 0_ds]
    This is the load-bearing premise of Theorem 1. It is only discussed heuristically in Appendix A.3 and is not proven.
  • standard math Assumption 1: neural networks can approximate the optimal loss arbitrarily well
    Used at the start of Appendix A.1 to bound the empirical loss around the optimal loss.
  • domain assumption Assumption 2: any added distinct feature strictly decreases optimal loss by at least gamma >= 0
    Assumed in the proof that including more features always helps predict Y; gamma is not verified on STPG data.
  • ad hoc to paper Assumption 3: invariance violations persist under addition of features with scaling delta > 0
    Technical assumption used to propagate the penalty lower bound; no empirical check is provided.
  • ad hoc to paper Assumption 4: extra augmentation of the same feature increases optimal loss by at least C > 0
    Assumes augmentations only add environment diversity and never reveal label information; central to deriving the lambda threshold.
  • domain assumption Augmented mask data X⊙M are valid environments with the same causal structure
    The invariance penalty in Eq. 4.6 treats each of the K augmented datasets as a distinct environment; this is assumed without empirical validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of diffIRM: A Diffusion-Augmented Invariant Risk Minimization Framework for Spatiotemporal Prediction over Graphs." pith.science (2026). https://pith.science/paper/HTBHM2SB

@misc{pith2026250100305,
  author       = {Pith},
  title        = {Pith review of: diffIRM: A Diffusion-Augmented Invariant Risk Minimization Framework for Spatiotemporal Prediction over Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HTBHM2SB}},
  note         = {Machine review of arXiv:2501.00305}
}
read the original abstract

Spatiotemporal prediction over graphs (STPG) is challenging, because real-world data suffers from the Out-of-Distribution (OOD) generalization problem, where test data follow different distributions from training ones. To address this issue, Invariant Risk Minimization (IRM) has emerged as a promising approach for learning invariant representations across different environments. However, IRM and its variants are originally designed for Euclidean data like images, and may not generalize well to graph-structure data such as spatiotemporal graphs due to spatial correlations in graphs. To overcome the challenge posed by graph-structure data, the existing graph OOD methods adhere to the principles of invariance existence, or environment diversity. However, there is little research that combines both principles in the STPG problem. A combination of the two is crucial for efficiently distinguishing between invariant features and spurious ones. In this study, we fill in this research gap and propose a diffusion-augmented invariant risk minimization (diffIRM) framework that combines these two principles for the STPG problem. Our diffIRM contains two processes: i) data augmentation and ii) invariant learning. In the data augmentation process, a causal mask generator identifies causal features and a graph-based diffusion model acts as an environment augmentor to generate augmented spatiotemporal graph data. In the invariant learning process, an invariance penalty is designed using the augmented data, and then serves as a regularizer for training the spatiotemporal prediction model. The real-world experiment uses three human mobility datasets, i.e. SafeGraph, PeMS04, and PeMS08. Our proposed diffIRM outperforms baselines.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 16 canonical work pages

  1. [1]

    Systematicgeneralisationwithgroupinvariantpredictions, in: International Conference on Learning Representations

    Ahmed,F.,Bengio,Y.,VanSeijen,H.,Courville,A.,2020. Systematicgeneralisationwithgroupinvariantpredictions, in: International Conference on Learning Representations. Ahuja, K., Shanmugam, K., Varshney, K., Dhurandhar, A.,

    work page 2020

  2. [2]

    As the relation between𝑌 and causal features𝑋cau is invariant across environments, including causal features can better lower the loss function

    On the other hand, to lower the objective function, the causal mask generator is trained to find causal features𝑋cau. As the relation between𝑌 and causal features𝑋cau is invariant across environments, including causal features can better lower the loss function. Thus the best strategy of the causal generator is to only exclude the environment features. Ap...

    work page 2019

  3. [3]

    Accident Analysis & Prevention 122, 239–254

    A spatiotemporal deep learning approach for citywide short-term crash risk prediction with multi-source data. Accident Analysis & Prevention 122, 239–254. Buffelli,D.,Liò,P.,Vandin,F.,2022. Sizeshiftreg:aregularizationmethodforimprovingsize-generalizationingraph neural networks. Advances in Neural Information Processing Systems 35, 31871–31885. Chang, S.,...

    work page 2022

  4. [4]

    Attention is all you need

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser,Ł.,Polosukhin,I.,2017. Attention is all you need. Advances in neural information processing systems

    work page 2017

  5. [6]

    Advancesinneuralinformationprocessingsystems33, 22092–22103

    Graph random neuralnetworksforsemi-supervisedlearningongraphs. Advancesinneuralinformationprocessingsystems33, 22092–22103. Guo,K.,Hu,Y.,Qian,Z.,Liu,H.,Zhang,K.,Sun,Y.,Gao,J.,Yin,B.,2020a. Optimizedgraphconvolutionrecurrent neural network for traffic prediction. IEEE Transactions on Intelligent Transportation Systems 22, 1138–1149. Guo, K., Hu, Y., Qian, ...

    work page 2020

  6. [8]

    Transportation Science 55, 553–573

    Graph signal processing techniques for analyzing aviation disruptions. Transportation Science 55, 553–573. Li, S., Wang, X., Zhang, A., Wu, Y., He, X., Chua, T.S., 2022c. Let invariant rationale discovery inspire graph con- trastive learning, in: International conference on machine learning, PMLR. pp. 13052–13065. Li, W., Yang, C., Jabari, S.E., 2022d. No...

    work page 2022

  7. [9]

    Distributionally Robust Semi-Supervised Learning Over Graphs

    Distributionally robust semi-supervised learning over graphs. arXiv preprint arXiv:2110.10582 . 27 Shengyu, Z., Yunze, T., Kun, K., Fuli, F., Jiezhong, Q., Jin, Y., Zhou, Z., Hongxia, Y., Zhongfei, Z., Fei, W.,

    work page Pith review arXiv

  8. [11]

    arXiv preprint arXiv:2211.02843

    Adversarial causal augmentation for graph covariate shift. arXiv preprint arXiv:2211.02843 . Sun, L.,

    arXiv

Show all 23 references
  1. [12]

    arXiv preprint arXiv:2206.04119

    Diffusion probabilistic modeling of protein backbones in 3d for the motif-scaffolding problem. arXiv preprint arXiv:2206.04119 . Vapnik, V.,

    arXiv

  2. [14]

    12191–12199

    Causalgnn: Causal-based graph neural networks for spatio-temporal epidemic forecasting, in: Proceedings of the AAAI conference on artificial intelligence, pp. 12191–12199. Wang,Y.,Zhao,Y.,Gan,S.,Li,K.,Chen,Y.,Lai,J.,2023. Optimizationofchargingstationsintegratedwithdynamic tra...

    2023 arXiv

  3. [15]

    1907–1913

    Graph wavenet for deep spatial-temporal graph modeling, in: Proceedings of the 28th International Joint Conference on Artificial Intelligence, pp. 1907–1913. Xia, Y., Liang, Y., Wen, H., Liu, X., Wang, K., Zhou, Z., Zimmermann, R.,

    1907

  4. [16]

    arXiv preprint arXiv:2309.13378

    Deciphering spatio-temporal graph forecasting: A causal lens and treatment. arXiv preprint arXiv:2309.13378 . 28 Xu, Y., Jaakkola, T.,

    arXiv

  5. [17]

    arXiv preprint arXiv:2110.09940

    Learning representations that support robust transfer of predictors. arXiv preprint arXiv:2110.09940 . Xue,J.,Yabe,T.,Tsubouchi,K.,Ma,J.,Ukkusuri,S.,2022.Multiwavecovid-19predictionfromsocialawarenessusing web search and mobility data, in: Proceedings of the 28th ACM SIGKDD Co...

    2022 arXiv

  6. [18]

    Transportation Research Part C: Emerging Technologies 107, 248–265

    A deep learning approach to real-time parking occupancy prediction in transportation networks incorporating multiple spatio-temporal data sources. Transportation Research Part C: Emerging Technologies 107, 248–265. You,Y.,Chen,T.,Sui,Y.,Chen,T.,Wang,Z.,Shen,Y.,2020.Graphcontra...

    2020

  7. [20]

    arXiv preprint arXiv:2206.09345

    Finding diverse and predictable subgraphs for graph domain generalization. arXiv preprint arXiv:2206.09345 . Zhang, G., Zhao, H., Yu, Y., Poupart, P.,

    arXiv

  8. [21]

    TransportationresearchpartC:emergingtechnologies105, 297–322

    Multistep speed prediction on traffic networks: A deep learning approachconsideringspatio-temporaldependencies. TransportationresearchpartC:emergingtechnologies105, 297–322. Zhang,Z.,Wang,X.,Zhang,Z.,Li,H.,Qin,Z.,Zhu,W.,2022. Dynamicgraphneuralnetworksunderspatio-temporal dist...

    2022

  9. [22]

    Advances in Neural Information Processing Systems 35, 20257–20272

    Ood link prediction generalization capabilities of message-passing gnns in larger test graphs. Advances in Neural Information Processing Systems 35, 20257–20272. Zhu,Q.,Ponomareva,N.,Han,J., Perozzi,B.,2021. Shift-robustgnns:Overcomingthelimitationsoflocalized graph training d...

    2021

  10. [2017]

    arXiv preprint arXiv:1709.04875

    Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875 . Yu, J., Liang, J., He, R.,

    arXiv

  11. [2019]

    arXiv preprint arXiv:1907.02893

    Invariant risk minimization. arXiv preprint arXiv:1907.02893 . 25 Bao, J., Liu, P., Ukkusuri, S.V.,

    1907 arXiv

  12. [2020]

    Invariant rationalization, in: International Conference on Machine Learning, PMLR. pp. 1448–1458. Chen,Y.,Zhang,Y.,Bian,Y.,Yang,H.,Kaili,M.,Xie,B.,Liu,T.,Han,B.,Cheng,J.,2022. Learningcausallyinvariant representations for out-of-distribution generalization on graphs. Advances ...

    2022

  13. [2021]

    IEEE Transactions on Intelligent Transportation Systems 23, 5704–5716

    A multi-task matrix factorized graph neural network for co-prediction of zone-based and od-based ride-hailing demand. IEEE Transactions on Intelligent Transportation Systems 23, 5704–5716. Feng,S.,Wei,S.,Zhang,J.,Li,Y.,Ke,J.,Chen,G.,Zheng,Y.,Yang,H.,2023. Amacro–microspatio-te...

    2023

  14. [2022]

    Li, H., Wang, X., Zhang, Z., Zhu, W., 2022a

    Dstagnn: Dynamic spatial-temporal aware graph neural networkfortrafficflowforecasting,in:InternationalConferenceonMachineLearning,PMLR.pp.11906–11917. Li, H., Wang, X., Zhang, Z., Zhu, W., 2022a. Out-of-distribution generalization on graphs: A survey. arXiv preprint arXiv:2202...

    arXiv

  15. [2023]

    Stable prediction on graphs with agnostic distribution shifts, in: The KDD’23 Workshop on Causal Discovery, Prediction and Decision, PMLR. pp. 49–74. Soppert,M.,Steinhardt,C.,Müller,C.,Gönsch,J.,2022. Differentiatedpricingofsharedmobilitysystemsconsidering network effects. Tra...

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.