Pith. sign in

REVIEW 5 major objections 7 minor 50 references

Treating each federated client as a causal environment lets shared prototype codebooks undo environment-specific confounding in spatio-temporal forecasts.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Federated aggregation of aligned environmental prototype codebooks acts as implicit back-door adjustment, with de-confounding error bounded linearly by average client confounding strength.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid federated STF engineering with a useful client-as-environment framing; the de-confounding “bound” is mostly an identity and the semi-synthetic check is stacked. the 5 major comments →

arxiv 2607.24218 v1 pith:YMCDWXJW submitted 2026-07-27 cs.LG cs.AI

Every Client Is an Environment: Federated De-confounding for Spatio-Temporal Forecasting

Qingxiang Liu , Anqi Liang , Heng Wang , Yuxuan Liang This is my paper
classification cs.LG cs.AI
keywords spatio-temporal forecastingfederated learningcausal inferencede-confoundingback-door adjustmentprototype codebookenvironmental regimes
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Federated spatio-temporal forecasting usually treats differences across sensors or cities as a nuisance to be fixed by personalization. This paper argues the opposite: those differences are complementary views of one underlying system under different environments, and federation can exploit them to remove confounding. It builds a structural causal model in which environment opens a back-door path from history to future, shows why a single client cannot close that path (missing strata and skewed weights), and proposes Atlas, which learns a global prototype codebook of environmental regimes. After clients align and average their codebooks, the residual de-confounding error is linearly controlled by the average—not the worst—client confounding strength. Experiments on traffic and air-quality data show the resulting forecasts beat other federated methods, transfer to held-out nodes, stay compact to communicate, and yield interpretable environmental assignments.

Core claim

When federated clients are treated as distinct causal environments and their local environmental prototype codebooks are aligned then aggregated, federation itself performs an implicit back-door adjustment: the de-confounding bias of the resulting representation is bounded by the weighted average of client-specific prototype deviations, not by any single client’s confounding strength.

What carries the argument

The global prototype codebook Q together with Hungarian alignment and soft assignment: each client maps its latent z_k onto shared environmental prototypes; the server matches and averages the codebooks so that the mixture approximates the global environmental distribution P(E).

Load-bearing premise

Under any fixed environmental condition, every client is assumed to share the same forecasting mechanism; only how often each condition appears can differ across clients.

What would settle it

If clients that experience identical measured environments still produce systematically different history-to-future conditionals (for example after matching weather and time-of-day), then averaging aligned codebooks should fail to reduce confounding leakage relative to purely local prototypes, contradicting the linear bound and the semi-synthetic cancellation result.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Cross-client diversity becomes a resource for causal identification rather than only a personalization problem.
  • Only a compact codebook need be exchanged, so communication cost stays far below full-model federated averaging while still approaching centralized accuracy.
  • A codebook pretrained on a small subset of nodes can be frozen and reused by unseen nodes with only local adaptation.
  • Prototype assignment weights supply an interpretable trace of which latent environmental regimes are active at each place and time.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same client-as-environment view could be tried on other federated sequential tasks (energy load, epidemic curves) where sites share a mechanism but differ in exposure.
  • If continuous or high-dimensional environment shifts dominate, the fixed discrete codebook may need to become dynamic or hierarchical, as the authors themselves flag.
  • Privacy analyses that treat uploaded prototypes as the sole leakage channel become especially relevant, because the rest of each model never leaves the client.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 7 minor

Summary. The paper proposes Atlas, a federated spatio-temporal forecasting framework that treats each client as a distinct causal environment. Instead of averaging model parameters, clients locally train encoders/decoders and a small prototype codebook intended to capture environmental regimes; the server aligns codebooks across clients via Hungarian matching and aggregates them. A causal analysis (§4.1–4.2) argues that local back-door adjustment fails because clients observe biased environmental marginals, and that federated aggregation of codebooks acts as an implicit back-door adjustment; Proposition 1 claims the resulting representation error is linearly controlled by the federation-averaged confounding strength. Experiments on five benchmarks (METR-LA, PEMS-BAY, PEMS03/04, KnowAir) show Atlas outperforming federated baselines and approaching centralized models on two traffic datasets, with ablations, held-out-client transfer, noise robustness, efficiency analysis, and prototype interpretability.

Significance. If the empirical results hold up, this is a useful contribution to federated spatio-temporal forecasting: the communication cost (a 10×64 codebook per round) is roughly two orders of magnitude below FedAvg-style methods while consistently beating federated baselines, and the transferability result (Table 3) — a codebook pretrained on 5% of clients transfers to held-out clients at near full-federation quality — is a falsifiable, practically relevant finding. The ablations A.1–A.6, noise-robustness study, and hyperparameter sensitivity (Appendix B) are above the usual standard for this venue. The conceptual reframing of clients as causal environments is appealing and likely to influence subsequent work. However, the theoretical contribution as currently written does not deliver what the abstract claims (see Major Comment 1), so the significance rests primarily on the empirical method rather than on the 'implicit back-door adjustment' theory.

major comments (5)
  1. [§4.2 / Appendix A, Eqs. (11)–(15), and Eq. (4)] Proposition 1 is presented as a guarantee, but the Appendix A proof is definitional. Eq. (11) introduces ΔQ_k as the residual Π_k Q_k − Q* for a stipulated optimal codebook Q*, so Eq. (15), ||Q − Q*|| = ||Σ_k w_k ΔQ_k||, holds identically for any aggregate Q and any choice of Q*. Nothing in the proof shows that (a) local training concentrates Π_k Q_k near a common Q*, (b) the deviations are small, or (c) the averaged deviation is smaller than individual ones — by the triangle inequality ||Σ w_k ΔQ_k|| ≤ Σ w_k ||ΔQ_k||, so federation only reduces error if deviations cancel across clients, a condition never stated or established. The same pattern appears in Eq. (4), where P(E) := Σ_k w_k P_k(E) is asserted with ≜: recovery of the global environmental distribution is assumed, not derived. As written, the 'federated de-confounding bound' provides no de-confounding guarantee. A repair would n
  2. [§4.3 and Figure 3] The semi-synthetic validation engineers the cancellation that Proposition 1 needs. With c_k = c̄ + δ·r_k and r_k ∈ {±1} balanced so Σ r_k = 0, the identity Σ w_k c_k = c̄ holds by construction, so Fig. 3 can only demonstrate that federation recovers the *average* confounder — and indeed Fig. 3(a) shows Atlas's leakage ℓ(p) growing linearly with c̄, i.e., the federated model retains exactly the mean confounding strength and de-confounds only in the engineered case c̄ = 0. Real environmental confounders (weather fronts, rush-hour structure) are positively correlated across nearby clients, so the realistic regime is same-sign deviations with c̄ ≠ 0 — precisely the regime in which the validation shows residual leakage. The experiment should be extended to same-sign and partially correlated sign patterns, and the claim in the abstract ('linearly controlled by the averaged confounding strength
  3. [§4.1, Definition 4.1] The shared-mechanism assumption P(Y_k | X_k, E=e) common to all clients is load-bearing for Eq. (4) and Lemma 1, yet it is exactly what is doubtful in the motivating applications: road geometry, sensor calibration, and unmodeled spatial structure can leave site-specific mechanisms after conditioning on E. If mechanisms differ, aligned-codebook averaging does not recover the global adjustment τ. The paper should (i) state explicitly what falsifies this assumption, (ii) discuss its plausibility for the five benchmarks (e.g., METR-LA sensors on different road types), and ideally (iii) include a semi-synthetic variant with heterogeneous mechanisms to characterize the degradation. Currently the assumption is stated once in Definition 4.1 and never revisited, including in the limitations discussion of §6.
  4. [Table 1 and §5.2 claims relative to centralized baselines] The claim that Atlas 'narrows the gap to centralized methods' is supported on METR-LA (MAE 3.30 vs. 3.05) and PEMS-BAY, where the quoted 3–7% gaps come from, but it fails elsewhere. On KnowAir, the centralized CaST achieves RMSE 16.38 versus Atlas's 23.72 — a ~45% gap — and on PEMS03 the centralized FaST (RMSE 25.08) beats Atlas (26.32). Group (i) baselines should either be discussed per-dataset in §5.2 with an explanation of the KnowAir anomaly (which may also indicate an evaluation-protocol inconsistency worth checking), or the summary claims should be scoped to the datasets where they hold.
  5. [§4.2, Lemma 1] Lemma 1 asserts that the soft prototype assignment A(z; Q_k) is a 'parameter-level approximation' of the local back-door adjustment b_k, but no formal or even semi-formal link is given between a convex combination of learned codebook vectors (Eq. 5) and the stratification sum Σ_e P(Y|X,e)P_k(e). The lemma is doing real work — it is the only bridge between the causal analysis of §4.1 and the architecture of §4.4 — and is currently a statement of intent rather than a lemma. At minimum, the conditions under which prototype mixtures correspond to stratum mixtures should be sketched, and the text should acknowledge what is assumed.
minor comments (7)
  1. [Throughout] Typos and language: 'distangle' (§1), 'casual environments' for 'causal environments' (contribution bullet), 'i,e,' (Definition 4.1), double colon in 'Limitations and future works::' (§6). Several uses of 'promises' where 'provides' or 'ensures' is intended.
  2. [Page headers/footer] Leftover ACM boilerplate: 'Conference acronym ’XX, June 03–05, 2018, Woodstock, NY' headers and the 'Received 20 February 2007...' footer should be removed or updated before camera-ready.
  3. [Figure 3] Figure 3 axis labels (|ps|/||p||) are not self-explanatory; define ℓ(p) in the caption. The y-axis range in Fig. 3(b) (0–0.08) makes Atlas's flatness visually stronger than the absolute scale warrants; consider a shared scale with (a) or state absolute values.
  4. [Table 1] The color/bold/underline encoding is stated in the caption but the 'yellow' highlight for Group (i) and the underline-vs-bold distinction for Groups (ii)–(iii) are hard to distinguish in grayscale; consider asterisks/daggers instead.
  5. [§5.5.2, final paragraph] The robustness results (noise added to the codebook does not much degrade MAE) do not by themselves support the claim that Atlas 'promises the incorporation of diverse differential privacy approaches'; robustness to perturbation is not a DP guarantee, and codebook leakage (e.g., prototype inversion) is not analyzed. Soften or add a leakage discussion.
  6. [§3, problem formulation] §3 states 'the spatial correlation can be implicitly learned by aggregation in the server,' but the server only aggregates codebooks (Eq. 8) — no spatial module exists server-side. Clarify what is meant, since omitting the topology graph is a notable modeling choice relative to the GNN-based baselines.
  7. [References] References [16] and [17] are duplicates (both DCRNN, ICLR 2018).

Circularity Check

2 steps flagged

Proposition 1's equality ||Q−Q*||=||Σ w_k ΔQ_k|| is algebraic once ΔQ_k is defined as the residual; the bound is Lipschitz-of-average-residual, not an independent de-confounding guarantee.

specific steps
  1. self definitional [Appendix A / Proposition 1, Eqs. (11)–(15)]
    "After prototype alignment, each client-specific codebook can be decomposed into the optimized global codebook Q* and a client-dependent environmental shift ΔQ_k: Π_k Q_k = Q* + ΔQ_k. ... Q = Q* + Σ_k w_k ΔQ_k. Therefore, ||Q − Q*|| = ||Σ_k w_k ΔQ_k||. ... ||ẑ − z*|| ≤ L_A ||Q − Q*||. ... ||ẑ − z*|| ≤ O(||Σ_k w_k ΔQ_k||)."

    ΔQ_k is introduced as the residual of the aligned local codebook from a stipulated optimum Q*. The equality ||Q − Q*|| = ||Σ w_k ΔQ_k|| then holds for every choice of Q* by rearranging the weighted average; it does not constrain local training, prove cancellation of client deviations, or show that the average residual is smaller than individuals (triangle inequality only upper-bounds it). The proposition therefore restates that representation error tracks the averaged residual under Lipschitz continuity of A, rather than deriving a non-definitional de-confounding rate.

  2. self definitional [Section 4.2, Eq. (4)]
    "τ(X_k)= Σ_e P(Y_k|X_k,e)P(e) ≜ Σ_e P(Y_k|X_k,e) Σ_k w_k P_k(e) = Σ_k w_k b_k(X_k)."

    The step marked ≜ replaces the global environment law P(e) by the weighted mixture of local laws. Under that substitution, aggregating local back-door estimates equals τ by algebra. The claim that federation recovers the global environmental distribution is therefore assumed by equating P(E) to Σ w_k P_k(E), not established from an external identification argument.

full rationale

The paper's central theoretical claim (federated de-confounding error linearly controlled by averaged confounding strength) rests on a short Appendix A argument that is mostly definitional algebra plus a standard continuity step, not a circular fit-to-prediction loop. Defining ΔQ_k := Π_k Q_k − Q* makes ||Q − Q*|| = ||Σ_k w_k ΔQ_k|| an identity for any stipulated Q*; the only non-tautological content is local Lipschitz continuity of the soft-assignment map A(·;Q), which transfers codebook residual to representation residual. Eq. (4) similarly inserts P(E) ≜ Σ_k w_k P_k(E) by definition of the client mixture. That weakens the rhetoric of a derived bound and of federation recovering a true global environment law, but it does not make the forecasting results or the semi-synthetic leakage plots self-fulfilling: Table 1 uses ordinary supervised losses on chronological held-out splits, and §4.3 injects confounding strength c̄ independently of the evaluation metric ℓ(p). There is no load-bearing self-citation chain, uniqueness import, or ansatz smuggled from prior author work. Score 3 reflects one mild self-definitional theoretical step without collapsing the empirical claims.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The central claim rests on standard do-calculus back-door form, a strong shared-mechanism assumption across clients given E, the modeling choice that a finite prototype codebook approximates environmental strata, and several hand-set architectural knobs. The bound further absorbs a Lipschitz constant of the assignment map and treats post-alignment prototype deviation as the manifestation of confounding strength.

free parameters (5)
  • Number of prototypes I = 10
    Discrete stratum count; default I=10 chosen and swept in Appendix B; directly controls capacity of the environmental codebook and communication size.
  • Prototype dimension D = 64
    Width of each prototype / latent vector; default 64, swept in sensitivity; affects both expressivity and upload cost I×D.
  • Assignment temperature α = 0.1
    Softmax temperature in Eq. (5); default 0.1 selected as optimum in sensitivity; controls soft vs sharp regime mixing.
  • Encoder depth (ResMLP layers) = 3
    Hand-chosen backbone depth (default 3); sensitivity shows diminishing returns beyond 3.
  • Federated rounds R and local epochs = R=200, 1 local epoch
    Training schedule defaults (R=200, 1 local epoch, Adam lr=0.001, batch 256) set by authors; not derived.
axioms (6)
  • domain assumption Back-door adjustment formula P(Y|do(X))=Σ_e P(Y|X,e)P(e) applies to the federated STF SCM with latent environment E on the path X←E→Y.
    Invoked in §3–§4.1 as the target causal estimand; standard Pearl calculus but assumes E is the relevant confounder class and no other open back-doors.
  • domain assumption Under the same environmental stratum, clients share a common conditional P(Y_k|X_k,E=e); only P_k(E) differs.
    Stated in Definition 4.1 / §4.1; required for local estimates b_k to be reweightings of one shared stratum model.
  • ad hoc to paper Soft prototype assignment A(z;Q) is a parameter-level approximation of local back-door adjustment b_k (Lemma 1).
    Bridges causal adjustment and representation learning; asserted rather than derived from a generative model of E.
  • standard math The assignment operator is locally Lipschitz in the codebook, so representation error tracks ||Q−Q*||.
    Used in the proof of Proposition 1 (Appendix A), citing continuity of cosine+softmax; standard analysis fact.
  • domain assumption Hungarian alignment on cosine costs recovers a coherent global indexing of exchangeable prototype slots across clients.
    §4.2 Eqs. (6)–(8); necessary so slot-wise averaging does not mix unrelated regimes.
  • domain assumption Spatial topology can be omitted and spatial correlation recovered implicitly via server aggregation.
    Stated in §3 Formulation; simplifies the federated setup relative to graph-based STF.
invented entities (2)
  • Global environmental prototype codebook Q no independent evidence
    purpose: Finite set of vectors standing for shared latent environmental regimes; the only object aligned and averaged across clients.
    Core modeling construct of Atlas; not an observed physical object. Interpretability panels give qualitative external handles (meteorology) but no independent measurement protocol for ‘true’ regimes.
  • Federated de-confounding bias / bound of Proposition 1 no independent evidence
    purpose: Quantifies residual representation error after codebook aggregation as linear in average prototype deviation.
    Paper-specific theoretical object; supported by a short continuity proof and a semi-synthetic leakage plot, not by external theorem libraries or formal verification.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Every Client Is an Environment: Federated De-confounding for Spatio-Temporal Forecasting." pith.science (2026). https://pith.science/paper/YMCDWXJW

@misc{pith2026260724218,
  author       = {Pith},
  title        = {Pith review of: Every Client Is an Environment: Federated De-confounding for Spatio-Temporal Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YMCDWXJW}},
  note         = {Machine review of arXiv:2607.24218}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Federated learning has emerged as a promising paradigm for spatio-temporal forecasting (STF), enabling collaborative model training without sharing raw observations. Existing federated STF methods primarily regard cross-client heterogeneity as an optimization challenge and mitigate it through personalized approaches. However, such heterogeneity fundamentally stems from diverse \emph{environmental conditions}, and these methods capture environment-specific forecasting patterns, hardly generalizing under environmental shifts. Our key insight is that the environmental diversity across federated clients should be exploited, as they provide \emph{complementary observations of the same underlying spatio-temporal system}. Based on this insight, we propose \method, a novel federated de-confounding framework that \textbf{treats clients as distinct causal environments}. \method leverages the client heterogeneity as distributed environmental evidence and learns a global prototype codebook to capture shared environmental regimes. We further derive a theoretical federated de-confounding bound that is linearly controlled by the averaged confounding strength. Extensive experiments demonstrate that \method consistently outperforms federated baselines, while providing transferable, interpretable, and communication-efficient environmental representations.

Figures

Figures reproduced from arXiv: 2607.24218 by Anqi Liang, Heng Wang, Qingxiang Liu, Yuxuan Liang.

Figure 1
Figure 1. Figure 1: Clients are treated as environments, and federated [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Causal analysis of federated STF, with (a) client [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical validation of Proposition 1. The feder￾ated de-confounding error is approximately linearly with averaged confounding strengths. Proposition 1 (Federated De-confounding Bound). After federated codebook aggregation, the federated de-confounding bias is bounded by ∥𝑧ˆ − 𝑧 ∗ ∥ ≤ 𝑂 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The overall implementation of Atlas. Each component is designed exactly according to federated de￾confounding mechanism. clients with 2000 samples for 60 federated rounds. We conduct 5 trials and report the mean and standard deviation values. In Fed optimum setting, we exhaustively search for the optimized shared 𝑝 under the same model architecture. In local setting, each client trains its specific prototy… view at source ↗
Figure 5
Figure 5. Figure 5: Trade-off between local updates and uploads. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Prototype interpretation on the KnowAir dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Hyper-parameter sensitivity on METR-LA. We in￾vestigate the number of the prototypes 𝐼, hidden dimension 𝐷, assignment temperature 𝛼, and encoder depth. The starred markers denote the default configuration adopted through￾out the paper (𝐼 = 10, 𝐷 = 64, 𝛼 = 0.1, and 3 ResMLP layers) [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 7 linked inside Pith

  1. [1]

    Manoj Ghuhan Arivazhagan, Vinay Aggarwal, Aaditya Kumar Singh, and Sunav Choudhary. 2019. Federated Learning with Personalization Layers.arXiv preprint arXiv:1912.00818(2019)

    Pith/arXiv arXiv 2019

  2. [2]

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. 2019. Invariant Risk Minimization.arXiv:1907.02893(2019)

    Pith/arXiv arXiv 2019

  3. [3]

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.arXiv preprint arXiv:1803.01271(2018)

    Pith/arXiv arXiv 2018

  4. [4]

    Lingxiao Cao, Bin Wang, Guiyuan Jiang, Yanwei Yu, and Junyu Dong. 2025. Spatiotemporal-aware trend-seasonality decomposition network for traffic flow forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 11463–11471

    2025

  5. [5]

    Yifan Duan, Jian Zhao, Junyuan Mao, Hao Wu, Jingyu Xu, Shilong Wang, Caoyuan Ma, Kai Wang, Kun Wang, and Xuelong Li. 2024. Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model.arXiv preprint arXiv:2409.19608(2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Yuchen Fang, Hao Miao, Yuxuan Liang, Liwei Deng, Yue Cui, Ximu Zeng, Yuyang Xia, Yan Zhao, Torben Bach Pedersen, Christian S Jensen, et al. 2026. Unraveling spatio-temporal foundation models via the pipeline lens: A comprehensive review. IEEE Transactions on Knowledge and Data Engineering(2026)

    2026

  7. [7]

    Alex Graves. 2013. Generating sequences with recurrent neural networks.arXiv preprint arXiv:1308.0850(2013)

    Pith/arXiv arXiv 2013

  8. [8]

    Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. 2019. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 922–929

    2019

  9. [9]

    Xianjie Guo, Kui Yu, Lizhen Cui, Han Yu, and Xiaoxiao Li. 2025. Federated causally invariant feature learning. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 16978–16986

    2025

  10. [10]

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. 1991. Adaptive mixtures of local experts.Neural computation3, 1 (1991), 79–87

    1991

  11. [11]

    Jiahao Ji, Wentao Zhang, Jingyuan Wang, and Chao Huang. 2025. Seeing the unseen: Learning basis confounder representations for robust traffic prediction. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 577–588

    2025

  12. [12]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907(2016)

    Pith/arXiv arXiv 2016

  13. [13]

    Harold W. Kuhn. 1955. The Hungarian method for the assignment problem. Naval Research Logistics Quarterly2, 1-2 (1955), 83–97

    1955

  14. [14]

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. 2020. Federated Optimization in Heterogeneous Networks. In MLSys

    2020

  15. [15]

    Ying Li, Xingwei Wang, Rongfei Zeng, Praveen Kumar Donta, Ilir Murturi, Min Huang, and Schahram Dustdar. 2025. Federated domain generalization: A survey. Proc. IEEE(2025)

    2025

  16. [17]

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. InICLR

    2018

  17. [18]

    Yuxuan Liang, Yutong Xia, Songyu Ke, Yiwei Wang, Qingsong Wen, Junbo Zhang, Yu Zheng, and Roger Zimmermann. 2023. Airformer: Predicting nationwide air quality in china with transformers. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 14329–14337

    2023

  18. [19]

    Qingxiang Liu, Sheng Sun, Yuxuan Liang, Jingjing Xue, and Min Liu. 2025. Per- sonalized Federated Learning for Spatio-Temporal Forecasting: A Dual Semantic Alignment-Based Contrastive Approach. InAAAI. 12192–12200

    2025

  19. [20]

    Yuxuan Liu, Wenchao Xu, Haozhao Wang, Zhiming He, Zhaofeng Shi, Chongyang Xu, Peichao Wang, and Boyuan Zhang. 2026. Causality-inspired federated learn- ing for dynamic spatio-temporal graphs. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. 23926–23934

    2026

  20. [21]

    Yi Liu, James J. Q. Yu, Jiawen Kang, Dusit Niyato, and Shuyu Zhang. 2020. Privacy- preserving Traffic Flow Prediction: A Federated Learning Approach.IEEE Internet of Things Journal7, 8 (2020), 7751–7763

    2020

  21. [22]

    Jiaming Ma, Binwu Wang, Pengkun Wang, Zhengyang Zhou, Xu Wang, and Yang Wang. 2025. Robust spatio-temporal centralized interaction for ood learning. In Forty-second International Conference on Machine Learning

    2025

  22. [23]

    Yingchi Mao, Hongliang Zhou, Ling Chen, Rongzhi Qi, Zhende Sun, Yi Rong, Xiaoming He, Mingkai Chen, Shahid Mumtaz, Valerio Frascolla, et al. 2025. A survey on spatio-temporal prediction: From transformers to foundation models. Comput. Surveys58, 4 (2025), 1–36

    2025

  23. [24]

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. InAISTATS

    2017

  24. [25]

    Chuizheng Meng, Sirisha Rambhatla, and Yan Liu. 2021. Cross-Node Federated Graph Neural Network for Spatio-Temporal Data Modeling. InKDD

    2021

  25. [26]

    2004.Introductory Lectures on Convex Optimization: A Basic Course

    Yurii Nesterov. 2004.Introductory Lectures on Convex Optimization: A Basic Course. Springer

    2004

  26. [27]

    2009.Causality

    Judea Pearl. 2009.Causality. Cambridge university press

    2009

  27. [28]

    Jonas Peters, Peter Bühlmann, and Nicolai Meinshausen. 2016. Causal Inference by using Invariant Prediction: Identification and Confidence Intervals.J. R. Stat. Soc. B78, 5 (2016), 947–1012

    2016

  28. [29]

    Zhuang Qi, Sijin Zhou, Lei Meng, Han Hu, Han Yu, and Xiangxu Meng. 2025. Federated deconfounding and debiasing learning for out-of-distribution general- ization.arXiv preprint arXiv:2505.04979(2025)

    Pith/arXiv arXiv 2025

  29. [30]

    Zezhi Shao, Fei Wang, Yongjun Xu, Wei Wei, Chengqing Yu, Zhao Zhang, Di Yao, Tao Sun, Guangyin Jin, Xin Cao, et al. 2024. Exploring progress in multi- variate time series forecasting: Comprehensive benchmarking and heterogeneity analysis.IEEE Transactions on Knowledge and Data Engineering37, 1 (2024), 291–305

    2024

  30. [31]

    Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, and Yongjun Xu. 2022. Spatial- Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting. InCIKM

    2022

  31. [32]

    Chao Song, Youfang Lin, Shengnan Guo, and Huaiyu Wan. 2020. Spatial- temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 914–921

    2020

  32. [33]

    Yongduo Sui, Xiang Wang, Jiancan Wu, Min Lin, Xiangnan He, and Tat-Seng Chua

  33. [34]

    Xueyang Tang, Song Guo, Jingcai Guo, Jie Zhang, and Yue Yu. 2024. Causally moti- vated personalized federated invariant learning with shortcut-averse information- theoretic regularization. InForty-first International Conference on Machine Learn- ing

    2024

  34. [35]

    Xueyang Tang, Song Guo, Jie Zhang, and Jingcai Guo. 2024. Learning personal- ized causally invariant representations for heterogeneous federated clients. In International Conference on Learning Representations, Vol. 2024. 10016–10037

    2024

  35. [36]

    Binwu Wang, Jiaming Ma, Pengkun Wang, Xu Wang, Yudong Zhang, Zhengyang Zhou, and Yang Wang. 2024. STONE: A Spatio-temporal OOD Learning Frame- work Kills Both Spatial and Temporal Shifts. InKDD. Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Liu et al

    2024

  36. [37]

    Kun Wang, Hao Wu, Yifan Duan, Guibin Zhang, Kai Wang, Xiaojiang Peng, Yu Zheng, Yuxuan Liang, and Yang Wang. 2024. NuwaDynamics: Discovering and Updating in Causal Spatio-Temporal Modeling. InICLR

    2024

  37. [38]

    Shuo Wang, Yanran Li, Jiang Zhang, Qingye Meng, Lingwei Meng, and Fei Gao

  38. [39]

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. 2019. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. InIJCAI

    2019

  39. [40]

    Yutong Xia, Yuxuan Liang, Haomin Wen, Xu Liu, Kun Wang, Zhengyang Zhou, and Roger Zimmermann. 2023. Deciphering Spatio-Temporal Graph Forecasting: A Causal Lens and Treatment. InNeurIPS

    2023

  40. [41]

    Kuo Yang, Yunhe Guo, Qihe Huang, Zhengyang Zhou, and Yang Wang. 2025. Revealing concept shift in spatio-temporal graphs via state learning. InProceed- ings of the Thirty-Fourth International Joint Conference on Artificial Intelligence. 3525–3533

    2025

  41. [42]

    Linghua Yang, Wantong Chen, Xiaoxi He, Shuyue Wei, Yi Xu, Zimu Zhou, and Yongxin Tong. 2024. FedGTP: exploiting inter-client spatial dependency in federated graph-based traffic prediction. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 6105–6116

    2024

  42. [43]

    Haonan Yuan, Qingyun Sun, Xingcheng Fu, Ziwei Zhang, Cheng Ji, Hao Peng, and Jianxin Li. 2023. Environment-Aware Dynamic Graph Learning for Out-of- Distribution Generalization. InNeurIPS

    2023

  43. [44]

    Weijia Zhang, Jindong Han, Zhao Xu, Hang Ni, Hao Liu, and Hui Xiong. 2024. Urban foundation models: A survey. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6633–6643

    2024

  44. [45]

    Yudong Zhang, Xu Wang, Pengkun Wang, Binwu Wang, Zhengyang Zhou, and Yang Wang. 2024. Modeling spatio-temporal mobility across data silos via person- alized federated learning.IEEE Transactions on Mobile Computing23, 12 (2024), 15289–15306

    2024

  45. [46]

    Yiji Zhao, Zihao Zhong, Ao Wang, Haomin Wen, Ming Jin, Yuxuan Liang, Huaiyu Wan, and Hao Wu. 2026. FaST: Efficient and Effective Long-Horizon Forecasting for Large-Scale Spatial-Temporal Graphs via Mixture-of-Experts. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    2026

  46. [47]

    Chuanpan Zheng, Xiaoliang Fan, Cheng Wang, and Jianzhong Qi. 2020. Gman: A graph multi-attention network for traffic prediction. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 1234–1241

    2020

  47. [48]

    Chengyang Zhou, Zijian Zhang, Chunxu Zhang, Hao Miao, Yulin Zhang, Kedi Lyu, and Juncheng Hu. 2026. Feddis: A causal disentanglement framework for federated traffic prediction. InProceedings of the ACM Web Conference 2026. 7541–7551

    2026

  48. [49]

    Zhengyang Zhou, Qihe Huang, Kuo Yang, Kun Wang, Xu Wang, Yudong Zhang, Yuxuan Liang, and Yang Wang. 2023. Maintaining the Status Quo: Capturing Invariant Relations for OOD Spatiotemporal Learning. InKDD. 3603–3614. A Proofs and Full Statements Proof. After prototype alignment, each client-specific codebook can be decomposed into the optimized global codeb...

    2023

  49. [2020]

    5-gnn: A domain knowledge enhanced graph neural network for pm2

    Pm2. 5-gnn: A domain knowledge enhanced graph neural network for pm2. 5 forecasting. InProceedings of the 28th international conference on advances in geographic information systems. 163–166

  50. [2022]

    In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining

    Causal attention for interpretable and generalizable graph classification. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 1696–1705

This paper was first reviewed by grok-4.5 on July 31, 2026.