Pith. sign in

REVIEW 3 major objections 4 minor 76 references

Invariant Link Selector for Spatial-Temporal Out-of-Distribution Problem

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that selecting invariant links in a temporal graph, via an information-bottleneck objective, provably bounds the train-to-test generalization gap in link prediction.

desk verdict A promising idea for temporal-graph invariant link selection that is currently unsupported by its own theory: the variational bound contradicts the model, and the error-bound proof drops non-vanishing terms. read the letter →

arxiv 2505.24178 v1 pith:55O2RYIF submitted 2025-05-30 cs.LG cs.AI

Katherine Tieu , Dongqi Fu , Jun Wu , Jingrui He This is my paper
classification cs.LGcs.AI
keywords out-of-distributiongeneralizationtemporalgraphsinvariantlearninginformationbottlenecklinkpredictiondynamicgraphneuralnetworksmutualvariationalbounds
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 takes on out-of-distribution generalization for temporal link prediction: given historical snapshots of a graph and a query link at a future time, can we predict whether the link appears when the test environment differs from training? It proposes OOD-Linker, a learnable invariant link selector that extracts a sparse subgraph of links most informative for the label while compressing away the rest, guided by an Information Bottleneck objective. The central claim is Theorem 1, which bounds the gap between expected loss on the training distribution and the test distribution by the mutual information of the extracted subgraph, the KL divergence between the two distributions, and a sample-size term. If this bound is valid, it is the first generalization guarantee for dynamic-graph out-of-distribution learning, and it gives a principled reason to make predictions from invariant links rather than from the full computational graph.

What carries the argument

The load-bearing object is the invariant link selector $p_{\phi_2}$, a neural network defined over each query link's L-hop computational subgraph. For each historical link $(a,b,t)$ it computes a soft selection probability using neighborhood aggregation over both previous invariant links and current neighbors, with a temperature parameter controlling how close to hard 0/1 the selection is. This selector is paired with $q_{\phi_3}$, a conditional prior over selected links given earlier selections, and $p_{\phi_1}$, a task head that predicts the query link from the selected links. The variational identity that carries the argument is the upper bound $I(\{e\}^T_1; \{G\}^T_1) \leq \sum_{t=1}^T D_{\mathrm{KL}}(p_{\phi_2}(e_t|G_t,\{e\}^{t-1}_1) \| q_{\phi_3}(e_t|\{e\}^{t-1}_1))$, which turns the information-bottleneck objective into the concrete loss used for training and feeds directly into Theorem 1.

What would settle it

Train OOD-Linker on a temporal graph and estimate, from the model's own selection probabilities, the conditional mutual information $I(e_t; \{G\}^{t-1}_1 | G_t, \{e\}^{t-1}_1)$ for a range of time steps; a measurably nonzero estimate would falsify the proof's decomposition in Eq. 24, meaning the KL loss is not the claimed upper bound and Theorem 1's guarantee does not apply to the trained model.

Watch

Extended reading notes

Core claim

The paper's discovery is a method, OOD-Linker, whose invariant link selection is designed so that the domain-shift gap in a temporal link prediction task is controlled by a mutual-information term. The selection objective maximizes $I(\{e\}^T_1; Y_{T+1})$ while constraining $I(\{e\}^T_1; \{G\}^T_1)$, and because direct optimization is intractable, the paper derives variational upper bounds that make the objective trainable: minimizing the negative log-likelihood of the link predictor plus a sum of KL divergences between the selector distribution and a prior over previously selected links. The proof of Theorem 1 then decomposes the train-test gap into a sequential generalization term and an information-theoretic term, yielding, with probability at least $1-\delta$, the bound $$\left|\mathbb{E}_\mu[\ell(f(\tilde{G}), Y)] - \mathbb{E}_\nu[\ell(f(\tilde{G}'), Y')]\right| \leq O\left(\frac{1}{N}\sum_{i=1}^N \sqrt{2\$sigma^{2}$ I(\$\varphi$(\tilde{G}_i), \tilde{G}_i)} + D_{\mathrm{KL}}(\mu\|\nu) + \sqrt{\frac{\log(1/\delta)}{N}}\right).$$ If the paper is right, the model's ability to compress the computational graph into invariant links is not just a heuristic; it is the quantity that governs worst-case performance under distribution shift.

Load-bearing premise

The bound assumes that the set of edges selected at time t reveals nothing about the past graphs once the current snapshot and earlier selected edges are known; when that conditional independence fails, the KL-divergence sum is no longer a proven upper bound on the information-bottleneck term.

Editorial extensions

If this is right

  • If Theorem 1 is correct, any temporal link prediction model that extracts a subgraph with small mutual information with the raw graph while preserving label information inherits a controlled train-test gap, growing at most as the square root of that mutual information.
  • The variational loss gives a practical training recipe: fit the link predictor and the selector together by minimizing the negative log-likelihood of the label plus a per-timestep KL divergence, and the same loss can be reused for fine-tuning on tasks such as citation and merchandise recommendation.
  • In the paper's experiments, selecting invariant links rather than using all links in the computational graph gives lower validation loss under edge-attribute shift, which supports the claim that the selection mechanism itself, rather than the backbone architecture, drives robustness.
  • The theorem justifies adding a mutual-information penalty to the loss instead of changing the architecture, because the bound's dominant data-dependent term is exactly the mutual information between what the model uses and the raw computational graph.

Reading between the lines

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

  • Editorial inference: the structure of the bound suggests a model-agnostic diagnostic; a practitioner could compute the mutual information between what any temporal graph encoder actually uses and the raw graph, and a small value on held-out environments would predict OOD robustness regardless of architecture.
  • Editorial inference: a stress test that follows from the proof is to build a temporal graph in which the truly invariant edges at time t depend on a hidden variable that shifts at test time, then compare the realized generalization gap to the bound; this would show whether the KL proxy tracks the true mutual information or only an approximation.
  • Editorial inference: the same selector could be applied to node classification or forecasting on temporal graphs, because the theorem only assumes a Lipschitz, sub-Gaussian loss and does not otherwise use link-prediction structure.
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

3 major / 4 minor

Summary. The manuscript proposes OOD-Linker, an invariant link selector for temporal link prediction under distribution shift. The method uses an information-bottleneck objective with variational bounds, a selector network that chooses invariant links based on past selections, and a link predictor; the central theoretical claim is Theorem 1, an OOD error bound in terms of mutual information I(phi(G_i), G_i), the KL divergence between training and test distributions, and a 1/sqrt(N) sample term. Experiments on COLLAB, ACT, and Aminer compare OOD-Linker against static, dynamic, and OOD baselines, with the best results on the edge-shift settings and competitive but mixed results on node-shift settings.

Significance. If valid, the paper would contribute a principled invariant-selection framework for temporal link prediction with a provable OOD guarantee, a claim of being first in dynamic-graph OOD generalization. The paper also provides a complexity analysis, a pseudocode training procedure, and a code URL. However, the theoretical core is not established as written: the variational bound relies on an assumption contradicted by the model, the main-text loss is inconsistent with the stated objective, and the proof of Theorem 1 contains an invalid empirical-average rewriting and silently discards terms that do not vanish with N. Since the advertised guarantee is the main contribution, the manuscript's central claim is not supported.

major comments (3)
  1. [Appendix A.2, Eq. (24)-(28)] The variational upper bound for I({e}_1^T; {G}_1^T) assumes that e_t is a noisy function of G_t alone, so that I(e_t; {G}_1^{t-1} | G_t, {e}_1^{t-1}) = 0 and the mutual information decomposes into a sum of per-time conditional mutual informations. This is contradicted by the paper's own selector network in Eq. (7), where p_phi2(e_t | G_t, {e}_1^{t-1}) explicitly aggregates over previous invariant links via the first summation over (w,t') with t' < t. Consequently, the chain-rule decomposition in Eq. (28) is not established, and the KL-sum used in the loss is not shown to be an upper bound on beta I({e}_1^T; {G}_1^T). Since Theorem 1 is stated in terms of I(phi(G_i), G_i), the link between the trained model and the error bound is broken.
  2. [Sec. 3.4, Eq. (15)] The main-text loss in Eq. (15) omits the beta coefficient that appears in the objective Eq. (6) and in the appendix loss Eq. (29), and the second summation over KL divergences also lacks the 1/N normalization present in the first term. As written, Eq. (15) does not match the information-bottleneck objective the paper claims to minimize. This is a load-bearing inconsistency because the theoretical results are about that objective.
  3. [Appendix B, Eq. (32)] The proof of Theorem 1 does not establish the stated bound. In Eq. (32), the empirical average (1/N)∑_{i=1}^N ℓ(f(̃G_i), Y_i) is rewritten as (1/N)∑_{i=1}^N ∑_{i=1}^T ℓ(f(̃G_i^t), Y_i), but ̃G_i^t is never defined and N query-link draws do not produce NT independent samples. The subsequent application of Corollary 2 of [16] introduces a sequential Rademacher term Rseq_T(ℓ∘F) that does not vanish with N, and the inequalities that follow insert factors (log T)^3 and sqrt((log T)^3/T) before the final equality to O(sqrt(log(1/δ)/N)). Those T-dependent terms and the M sqrt(log(1/δ)) term are silently discarded; big-O notation cannot eliminate quantities independent of N. Thus the advertised (N, δ) dependence does not follow from the displayed calculation.
minor comments (4)
  1. [Sec. 3.3.2, Eq. (11)-(12)] Eq. (11) defines h_{a,N,ϕ_2} and Eq. (12) uses weights W^{(3)}_{ϕ_2}, W^{(2)}_{ϕ_2}, W^{(1)}_{ϕ_1}; these appear to be typographical errors for ϕ_3 and W^{(3)}_{ϕ_3}, W^{(2)}_{ϕ_3}, W^{(1)}_{ϕ_3}, respectively, and should be corrected for reproducibility.
  2. [Appendix B, Eq. (32)] The notation ̃G_i^t used in the double sum is never defined; the paper defines computational graphs ̃G_i for query links but not per-timestamp graphs with this tilde notation.
  3. [Sec. 5.4, Figures 1-2] The text states that there is a 'substantial gap' between OOD-Linker and the all-links baseline on COLLAB, but the figures are not included in the text provided; please ensure the figures are legible and the claimed gap is visible or quantified.
  4. [Abstract and Appendix D.4] The abstract and Section D.4 say the code will be released upon publication, while the abstract's footnote claims the code is available at a GitHub URL; please make this consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 1 is a conditional information-bottleneck bound proved from external tools, not a restatement of the fitted training loss.

full rationale

The paper's central claim, Theorem 1, is stated for an arbitrary selector phi and predictor f, and its proof invokes external results [16], [25], [28], and [37]. The right-hand side contains the information-bottleneck mutual information I(phi(G_i), G_i), which is the quantity that the variational loss in Eq. 15 is designed to control; however, the theorem is a conditional generalization bound rather than a restatement of the loss. No equation substitutes the fitted KL loss for the bound, and no term is set equal to the training objective by construction. The variational upper bound of Appendix A.2 and the proof steps of Appendix B raise validity concerns (an unjustified Markov assumption contradicting Eq. 7, and an undefined N-by-T double sum with dropped T-dependent terms), but these are internal proof gaps, not circular reductions. The paper's self-citations occur in background and related-work contexts, while the load-bearing theoretical steps rely on external references, so no self-citation chain forces the result.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on two hand-chosen hyperparameters (beta, tau), the standard invariant-learning premise that a stable subgraph determines the label, and a set of analytical assumptions. The most auditable cost is the ad hoc independence assumption in Appendix A.2, which is contradicted by the paper's own equations and is necessary for the KL upper bound to follow. No new physical or latent entities are introduced; the invariant link selector is a learned module, not a new postulated object.

free parameters (2)
  • beta (IB trade-off coefficient) = not reported
    Appears in the objective Eq. 6 and in the appendix loss Eq. 29, but is omitted from the main-text loss Eq. 15. It controls the compression-preservation trade-off and is chosen by hand or tuned per dataset, with no reported value in the hyperparameter table.
  • tau (temperature of the selector sigmoid) = 1.0
    Appendix D.4 sets tau = 1.0 for all datasets. It controls how hard the link-selection probabilities are, and it is fixed by hand rather than learned or justified.
assumptions (5)
  • domain assumption The label Y of a query link is determined by an invariant subgraph of its computational graph; variant parts are spurious, as in the cow-versus-camel example of [1].
    Section 3.2 opens with 'Suppose the label of a link is determined by an invariant subgraph'; this is the standard invariant-learning premise, not proven for temporal graphs.
  • ad hoc to paper The selected edge set e_t is a noisy function of G_t alone, so that cross mutual information terms vanish in the decomposition of I({e}^T_1; {G}^T_1).
    Appendix A.2 (Eq. 24 and the paragraph before it) assumes e_t = f(G_t, epsilon) to drop I(e_t; {G}^{t-1}_1 | G_t, {e}^{t-1}_1) and I({e}^{t-1}_1; G_t | {G}^{t-1}_1). This contradicts Eq. 7, where p_phi2 conditions on previous invariant links {e}^{t-1}_1. Without this assumption the KL sum is not an upper bound on the stated MI objective.
  • standard math The loss function is alpha-Lipschitz and sigma-sub-Gaussian.
    Assumed in Theorem 1 to apply concentration and Rademacher complexity arguments; a standard boundedness condition, not specific to this paper.
  • domain assumption The transfer bound of Liu et al. [25] applies directly to OOD-Linker's adaptive selection function phi on temporal computational graphs.
    Appendix B states 'we can obtain the upper bound by directly apply results from [25]' without verifying that [25]'s conditions hold for sequential, history-dependent link selection; this is the key imported result of the theorem.
  • domain assumption The L-hop neighborhood of the query link over all historical timestamps is a sufficient scope for invariant selection.
    Section 3.3.1 defines the computational subgraph as the L-hop neighborhood at all times before the query; the paper proposes to learn L adaptively, but the sufficiency of this scope for the label is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Invariant Link Selector for Spatial-Temporal Out-of-Distribution Problem." pith.science (2026). https://pith.science/paper/55O2RYIF

@misc{pith2026250524178,
  author       = {Pith},
  title        = {Pith review of: Invariant Link Selector for Spatial-Temporal Out-of-Distribution Problem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55O2RYIF}},
  note         = {Machine review of arXiv:2505.24178}
}
read the original abstract

In the era of foundation models, Out-of- Distribution (OOD) problems, i.e., the data discrepancy between the training environments and testing environments, hinder AI generalization. Further, relational data like graphs disobeying the Independent and Identically Distributed (IID) condition makes the problem more challenging, especially much harder when it is associated with time. Motivated by this, to realize the robust invariant learning over temporal graphs, we want to investigate what components in temporal graphs are most invariant and representative with respect to labels. With the Information Bottleneck (IB) method, we propose an error-bounded Invariant Link Selector that can distinguish invariant components and variant components during the training process to make the deep learning model generalizable for different testing scenarios. Besides deriving a series of rigorous generalizable optimization functions, we also equip the training with task-specific loss functions, e.g., temporal link prediction, to make pretrained models solve real-world application tasks like citation recommendation and merchandise recommendation, as demonstrated in our experiments with state-of-the-art (SOTA) methods. Our code is available at https://github.com/kthrn22/OOD-Linker.

Figures

Figures reproduced from arXiv: 2505.24178 by the authors.

Figure 1
Figure 1. Comparison of link prediction task loss on the training and edge OOD settings of ACT. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Comparison of link prediction task loss on the training and edge OOD settings of COLLAB. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 62 canonical work pages

  1. [16]

    Discrepancy-based theory and algorithms for forecasting non-stationary time series.Ann

    Vitaly Kuznetsov and Mehryar Mohri. Discrepancy-based theory and algorithms for forecasting non-stationary time series.Ann. Math. Artif. Intell., 88(4):367–399, 2020

    work page 2020

  2. [1]

    Invariance principle meets information bottleneck for out- of-distribution generalization

    Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets information bottleneck for out- of-distribution generalization. In Marc’Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, ed- itors, Advances in Neural Info...

    work page 2021

  3. [2]

    Alemi, Ian Fischer, Joshua V

    Alexander A. Alemi, Ian Fischer, Joshua V. Dil- lon, and Kevin Murphy. Deep variational informa- tion bottleneck. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Pro- ceedings. OpenReview.net, 2017

    work page 2017

  4. [3]

    Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fa- had Shahbaz Khan

    Muhammad Awais, Muzammal Naseer, Salman H. Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fa- had Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook.CoRR, abs/2307.13721, 2023

    arXiv 2023

  5. [4]

    Pagerank bandits for link prediction

    Yikun Ban, Jiaru Zou, Zihao Li, Yunzhe Qi, Dongqi Fu, Jian Kang, Hanghang Tong, and Jin- grui He. Pagerank bandits for link prediction. In NeurIPS, 2024

    work page 2024

  6. [5]

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

    Yongqiang Chen, Yonggang Zhang, Yatao Bian, Han Yang, Kaili Ma, Binghui Xie, Tongliang Liu, Bo Han, and James Cheng. Learning causally invariant representations for out-of-distribution generalization on graphs. In Sanmi Koyejo, S. Mo- hamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Informa- tion Processing Systems 35: A...

    work page 2022

  7. [6]

    Torvik, and Jingrui He

    Dongqi Fu, Liri Fang, Ross Maciejewski, Vetle I. Torvik, and Jingrui He. Meta-learned metrics over multi-evolution temporal graphs. In Aidong Zhang and Huzefa Rangwala, editors,KDD, 2022

    work page 2022

  8. [7]

    SDG: A simplified and dynamic graph neural network

    Dongqi Fu and Jingrui He. SDG: A simplified and dynamic graph neural network. In Fernando Diaz, Chirag Shah, Torsten Suel, Pablo Castells, Rosie Jones, and Tetsuya Sakai, editors,SIGIR, 2021

    work page 2021

Show all 76 references
  1. [8]

    Vcr-graphormer: A mini-batch graph transformer via virtual con- nections

    Dongqi Fu, Zhigang Hua, Yan Xie, Jin Fang, Si Zhang, Kaan Sancak, Hao Wu, Andrey Male- vich, Jingrui He, and Bo Long. Vcr-graphormer: A mini-batch graph transformer via virtual con- nections. In ICLR, 2024

    2024

  2. [9]

    When graph neural network meets causality: Opportuni- ties, methodologies and an outlook, 2024

    Wenzhao Jiang, Hao Liu, and Hui Xiong. When graph neural network meets causality: Opportuni- ties, methodologies and an outlook, 2024

    2024

  3. [10]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Variational graph auto-encoders. CoRR, abs/1611.07308, 2016

    2016 arXiv

  4. [11]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi- supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Pro- ceedings. OpenReview.net, 2017

    2017

  5. [12]

    Courville

    David Krueger, Ethan Caballero, Jörn-Henrik Ja- cobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Rémi Le Priol, and Aaron C. Courville. Out-of-distribution generalization via risk extrap- olation (rex). In Marina Meila and Tong Zhang, editors, Proceedings of the 38th Interna...

    2021

  6. [13]

    Kullback-leibler divergence, 1951

    Solomon Kullback. Kullback-leibler divergence, 1951

    1951

  7. [14]

    Predicting dynamic embedding trajectory in tem- poral interaction networks

    Srijan Kumar, Xikun Zhang, and Jure Leskovec. Predicting dynamic embedding trajectory in tem- poral interaction networks. In Ankur Teredesai, Vipin Kumar, Ying Li, Rómer Rosales, Evimaria Terzi, and George Karypis, editors,Proceedings of the 25th ACM SIGKDD International Confe...

    2019

  8. [15]

    Time series prediction and online learning

    Vitaly Kuznetsov and Mehryar Mohri. Time series prediction and online learning. In Vitaly Feldman, Alexander Rakhlin, and Ohad Shamir, editors, Proceedings of the 29th Conference on Learning Theory, COLT 2016, New York, USA, June 23-26, 2016, volume 49 of JMLR Workshop and Con...

    2016

  9. [17]

    Disen- tangled graph self-supervised learning for out-of- distribution generalization

    Haoyang Li, Xin Wang, Zeyang Zhang, Haibo Chen, Ziwei Zhang, and Wenwu Zhu. Disen- tangled graph self-supervised learning for out-of- distribution generalization. In Forty-first Inter- national Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenR...

    2024

  10. [18]

    Out-of-distribution generalization on graphs: A survey.CoRR, abs/2202.07987, 2022

    Haoyang Li, Xin Wang, Ziwei Zhang, and Wenwu Zhu. Out-of-distribution generalization on graphs: A survey.CoRR, abs/2202.07987, 2022

    2022 arXiv

  11. [19]

    Learning invariant graph representations for out-of-distribution generalization

    Haoyang Li, Ziwei Zhang, Xin Wang, and Wenwu Zhu. Learning invariant graph representations for out-of-distribution generalization. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Bel- grave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems 35: An-...

    2022

  12. [20]

    Apex2: Adaptive and extreme summarization for personalized knowledge graphs.CoRR, 2024

    Zihao Li, Dongqi Fu, Mengting Ai, and Jingrui He. Apex2: Adaptive and extreme summarization for personalized knowledge graphs.CoRR, 2024

    2024

  13. [21]

    Everything evolves in personalized pagerank

    Zihao Li, Dongqi Fu, and Jingrui He. Everything evolves in personalized pagerank. InWWW, 2023

    2023

  14. [22]

    Backtime: Backdoor attacks onmultivariatetimeseriesforecasting

    Xiao Lin, Zhining Liu, Dongqi Fu, Ruizhong Qiu, and Hanghang Tong. Backtime: Backdoor attacks onmultivariatetimeseriesforecasting. In NeurIPS, 2024

    2024

  15. [23]

    Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre- train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Comput. Surv., 55(9):195:1–195:35, 2023

    2023

  16. [24]

    Beyond generaliza- tion: A survey of out-of-distribution adaptation on graphs

    Shuhan Liu and Kaize Ding. Beyond generaliza- tion: A survey of out-of-distribution adaptation on graphs. CoRR, abs/2402.11153, 2024

    2024 arXiv

  17. [25]

    An information-theoretic framework for out-of-distribution generalization

    Wenliang Liu, Guanding Yu, Lele Wang, and Ren- jie Liao. An information-theoretic framework for out-of-distribution generalization. In IEEE In- ternational Symposium on Information Theory, ISIT 2024, Athens, Greece, July 7-12, 2024, pages 2670–2675. IEEE, 2024

    2024

  18. [26]

    Disentangled graph convolu- tional networks

    Jianxin Ma, Peng Cui, Kun Kuang, Xin Wang, and Wenwu Zhu. Disentangled graph convolu- tional networks. In Kamalika Chaudhuri and Rus- lan Salakhutdinov, editors,Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, Califor...

    2019

  19. [27]

    Some methods for classification and analysis of multivariate observations

    J Macqueen. Some methods for classification and analysis of multivariate observations. InProceed- ings of 5-th Berkeley Symposium on Mathematical Statistics and Probability/University of California Press, 1967

    1967

  20. [28]

    Exploiting smoothness in sta- tisticallearning, sequentialprediction, andstochas- tic optimization

    Mehrdad Mahdavi. Exploiting smoothness in sta- tisticallearning, sequentialprediction, andstochas- tic optimization. CoRR, abs/1407.5908, 2014

    2014 arXiv

  21. [29]

    Position: Graph foun- dation models are already here

    Haitao Mao, Zhikai Chen, Wenzhuo Tang, Jianan Zhao, Yao Ma, Tong Zhao, Neil Shah, Mikhail Galkin, and Jiliang Tang. Position: Graph foun- dation models are already here. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. ...

    2024

  22. [30]

    Interpretable and generalizable graph learning via stochastic atten- tion mechanism

    Siqi Miao, Mia Liu, and Pan Li. Interpretable and generalizable graph learning via stochastic atten- tion mechanism. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvári, Gang Niu, and Sivan Sabato, editors,International Confer- ence on Machine Learning, ICML 202...

    2022

  23. [31]

    Corrado, and Jeffrey Dean

    Tomás Mikolov, Ilya Sutskever, Kai Chen, Gre- gory S. Corrado, and Jeffrey Dean. Distributed representations of words and phrases and their compositionality. In Christopher J. C. Burges, Léon Bottou, Zoubin Ghahramani, and Kilian Q. Weinberger, editors,Advances in Neural Infor...

  24. [32]

    Generalized out-of-distribution detection and beyond in vi- sion language model era: A survey

    Atsuyuki Miyai, Jingkang Yang, Jingyang Zhang, Yifei Ming, Yueqian Lin, Qing Yu, Go Irie, Shafiq Joty, Yixuan Li, Hai Li, Ziwei Liu, Toshihiko Yamasaki, and Kiyoharu Aizawa. Generalized out-of-distribution detection and beyond in vi- sion language model era: A survey. CoRR, ab...

    2024 arXiv

  25. [33]

    Schardl, and Charles E

    Aldo Pareja, Giacomo Domeniconi, Jie Chen, TengfeiMa, ToyotaroSuzumura, HirokiKanezashi, Tim Kaler, Tao B. Schardl, and Charles E. Leiser- son. Evolvegcn: Evolving graph convolutional net- works for dynamic graphs. InThe Thirty-Fourth Katherine Tieu, Dongqi F u, Jun W u, Jingr...

    2020

  26. [34]

    Causal inference in statistics: An overview

    Judea Pearl. Causal inference in statistics: An overview. Statistics Surveys, 3(none):96 – 146, 2009

    2009

  27. [35]

    Graph retrieval-augmented generation: A survey.CoRR, abs/2408.08921, 2024

    Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Sil- iang Tang. Graph retrieval-augmented generation: A survey.CoRR, abs/2408.08921, 2024

    2024 arXiv

  28. [36]

    Graph neural bandits

    Yunzhe Qi, Yikun Ban, and Jingrui He. Graph neural bandits. InKDD, 2023

    2023

  29. [37]

    Online learning via sequential com- plexities

    Alexander Rakhlin, Karthik Sridharan, and Am- buj Tewari. Online learning via sequential com- plexities. J. Mach. Learn. Res., 16:155–186, 2015

    2015

  30. [38]

    Sheth, and Amitava Das

    Vipula Rawte, Amit P. Sheth, and Amitava Das. A survey of hallucination in large foundation models. CoRR, abs/2309.05922, 2023

    2023 arXiv

  31. [39]

    Hashimoto, and Percy Liang

    Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case gener- alization. CoRR, abs/1911.08731, 2019

    1911 arXiv

  32. [40]

    Dysat: Deep neural rep- resentation learning on dynamic graphs via self- attention networks

    Aravind Sankar, Yanhong Wu, Liang Gou, Wei Zhang, and Hao Yang. Dysat: Deep neural rep- resentation learning on dynamic graphs via self- attention networks. In James Caverlee, Xia (Ben) Hu, Mounia Lalmas, and Wei Wang, editors, WSDM ’20: The Thirteenth ACM International Confer...

    2020

  33. [41]

    Structured se- quence modeling with graph convolutional recur- rent networks

    Youngjoo Seo, Michaël Defferrard, Pierre Van- dergheynst, and Xavier Bresson. Structured se- quence modeling with graph convolutional recur- rent networks. In Long Cheng, Andrew Chi-Sing Leung, and Seiichi Ozawa, editors,Neural Infor- mation Processing - 25th International Con...

    2018

  34. [42]

    To- wards out-of-distribution generalization: A survey

    Zheyan Shen, Jiashuo Liu, Yue He, Xingxuan Zhang, Renzhe Xu, Han Yu, and Peng Cui. To- wards out-of-distribution generalization: A survey. CoRR, abs/2108.13624, 2021

    2021 arXiv

  35. [43]

    An overview of microsoft academic service (MAS) and applications

    Arnab Sinha, Zhihong Shen, Yang Song, Hao Ma, Darrin Eide, Bo-June Paul Hsu, and Kuansan Wang. An overview of microsoft academic service (MAS) and applications. In Aldo Gangemi, Ste- fano Leonardi, and Alessandro Panconesi, editors, Proceedings of the 24th International Confer...

    2015

  36. [44]

    Cross-domain collaboration recommendation

    Jie Tang, Sen Wu, Jimeng Sun, and Hang Su. Cross-domain collaboration recommendation. In Qiang Yang, Deepak Agarwal, and Jian Pei, edi- tors, The 18th ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining, KDD ’12, Beijing, China, August 12-16, 2012, pag...

    2012

  37. [45]

    Arnetminer: extrac- tion and mining of academic social networks

    Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang, and Zhong Su. Arnetminer: extrac- tion and mining of academic social networks. In Ying Li, Bing Liu, and Sunita Sarawagi, editors, Proceedings of the 14th ACM SIGKDD Interna- tional Conference on Knowledge Discovery and Dat...

    2008

  38. [46]

    Hamann, and Jingrui He

    Katherine Tieu, Dongqi Fu, Yada Zhu, Hendrik F. Hamann, and Jingrui He. Temporal graph neu- ral tangent kernel with graphon-guaranteed. In NeurIPS, 2024

    2024

  39. [47]

    Naftali Tishby, Fernando C. N. Pereira, and William Bialek. The information bottleneck method. CoRR, physics/0004057, 2000

    2000 arXiv

  40. [48]

    Llama: Open and efficient foundation lan- guage models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Ar- mand Joulin, Edouard Grave, and Guillaume Lam- ple. Llama: Open and efficient foundation lan-...

    2023 arXiv

  41. [49]

    STONE: A spatio-temporal OOD learning framework kills both spatial and tempo- ral shifts

    Binwu Wang, Jiaming Ma, Pengkun Wang, Xu Wang, Yudong Zhang, Zhengyang Zhou, and Yang Wang. STONE: A spatio-temporal OOD learning framework kills both spatial and tempo- ral shifts. In Ricardo Baeza-Yates and Francesco Bonchi, editors, Proceedings of the 30th ACM SIGKDD Confer...

    2024

  42. [50]

    Learning graph quan- tized tokenizers for transformers.CoRR, 2024

    Limei Wang, Kaveh Hassani, Si Zhang, Dongqi Fu, Baichuan Yuan, Weilin Cong, Zhigang Hua, Hao Wu, Ning Yao, and Bo Long. Learning graph quan- tized tokenizers for transformers.CoRR, 2024. Invariant Link Selector for Spatial-T emporal Out-of-Distribution Problem

    2024

  43. [51]

    Ainsworth

    Jun Wu, Jingrui He, and Elizabeth A. Ainsworth. Non-iid transfer learning on graphs. In Brian Williams, Yiling Chen, and Jennifer Neville, edi- tors, AAAI, 2023

    2023

  44. [52]

    Discovering invariant ra- tionales for graph neural networks

    Yingxin Wu, Xiang Wang, An Zhang, Xiangnan He, and Tat-Seng Chua. Discovering invariant ra- tionales for graph neural networks. InThe Tenth International Conference on Learning Represen- tations, ICLR 2022, Virtual Event, April 25-29,

    2022

  45. [53]

    Deciphering spatio-temporal graph fore- casting: A causal lens and treatment

    Yutong Xia, Yuxuan Liang, Haomin Wen, Xu Liu, Kun Wang, Zhengyang Zhou, and Roger Zimmer- mann. Deciphering spatio-temporal graph fore- casting: A causal lens and treatment. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Ad...

    2023

  46. [54]

    Inductive representation learning on temporal graphs

    Da Xu, Chuanwei Ruan, Evren Körpeoglu, Sushant Kumar, and Kannan Achan. Inductive representation learning on temporal graphs. In8th International Conference on Learning Represen- tations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020

    2020

  47. [55]

    Lan- guage models are graph learners.CoRR, 2024

    Zhe Xu, Kaveh Hassani, Si Zhang, Hanqing Zeng, Michihiro Yasunaga, Limei Wang, Dongqi Fu, Ning Yao, Bo Long, and Hanghang Tong. Lan- guage models are graph learners.CoRR, 2024

    2024

  48. [56]

    Dynamic knowledge graph alignment

    Yuchen Yan, Lihui Liu, Yikun Ban, Baoyu Jing, and Hanghang Tong. Dynamic knowledge graph alignment. In AAAI, 2021

    2021

  49. [57]

    Generalized out-of-distribution detection: A survey.CoRR, abs/2110.11334, 2021

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Zi- wei Liu. Generalized out-of-distribution detection: A survey.CoRR, abs/2110.11334, 2021

    2021 arXiv

  50. [58]

    Dynamic graph information bottleneck

    Haonan Yuan, Qingyun Sun, Xingcheng Fu, Cheng Ji, and Jianxin Li. Dynamic graph information bottleneck. In Tat-Seng Chua, Chong-Wah Ngo, Ravi Kumar, Hady W. Lauw, and Roy Ka-Wei Lee, editors,Proceedings of the ACM on Web Con- ference 2024, WWW 2024, Singapore, May 13-17, 2024,...

    2024

  51. [59]

    Environment-aware dynamic graph learning for out-of-distribution generalization

    Haonan Yuan, Qingyun Sun, Xingcheng Fu, Zi- wei Zhang, Cheng Ji, Hao Peng, and Jianxin Li. Environment-aware dynamic graph learning for out-of-distribution generalization. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Ad- ...

    2023

  52. [60]

    PARROT: position-aware regular- ized optimal transport for network alignment

    Zhichen Zeng, Si Zhang, Yinglong Xia, and Hang- hang Tong. PARROT: position-aware regular- ized optimal transport for network alignment. In WWW, 2023

    2023

  53. [61]

    Generative graph dictionary learning

    Zhichen Zeng, Ruike Zhu, Yinglong Xia, Hanqing Zeng, and Hanghang Tong. Generative graph dictionary learning. InICML, 2023

    2023

  54. [62]

    Dynamic graph neural networks under spatio-temporal distribu- tion shift

    Zeyang Zhang, Xin Wang, Ziwei Zhang, Haoyang Li, Zhou Qin, and Wenwu Zhu. Dynamic graph neural networks under spatio-temporal distribu- tion shift. In Sanmi Koyejo, S. Mohamed, A. Agar- wal, Danielle Belgrave, K. Cho, and A. Oh, edi- tors, Advances in Neural Information Proces...

    2022

  55. [63]

    Out-of-distribution gener- alized dynamic graph neural network with dis- entangled intervention and invariance promotion

    Zeyang Zhang, Xin Wang, Ziwei Zhang, Haoyang Li, and Wenwu Zhu. Out-of-distribution gener- alized dynamic graph neural network with dis- entangled intervention and invariance promotion. CoRR, abs/2311.14255, 2023

    2023 arXiv

  56. [64]

    Spectral invariant learning for dy- namic graphs under distribution shifts

    Zeyang Zhang, Xin Wang, Ziwei Zhang, Zhou Qin, Weigao Wen, Hui Xue, Haoyang Li, and Wenwu Zhu. Spectral invariant learning for dy- namic graphs under distribution shifts. In Al- ice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, edi- tors, A...

    2023

  57. [65]

    Drgnn: Deep residual graph neural network with contrastive learning

    Lecheng Zheng, Dongqi Fu, Ross Maciejewski, and Jingrui He. Drgnn: Deep residual graph neural network with contrastive learning. InTMLR, 2024

    2024

  58. [66]

    Pyg-ssl: A graph self-supervised learning toolkit

    Lecheng Zheng, Baoyu Jing, Zihao Li, Zhichen Zeng, Tianxin Wei, Mengting Ai, Xinrui He, Lihui Liu, Dongqi Fu, Jiaxuan You, Hanghang Tong, and Jingrui He. Pyg-ssl: A graph self-supervised learning toolkit. CoRR, 2024

    2024

  59. [67]

    Mentorgnn: Deriving cur- riculum for pre-training gnns

    Dawei Zhou, Lecheng Zheng, Dongqi Fu, Jiawei Han, and Jingrui He. Mentorgnn: Deriving cur- riculum for pre-training gnns. In Mohammad Al Hasan and Li Xiong, editors,CIKM, 2022

    2022

  60. [68]

    Maintaining the status quo: Cap- turing invariant relations for OOD spatiotemporal learning

    Zhengyang Zhou, Qihe Huang, Kuo Yang, Kun Wang, Xu Wang, Yudong Zhang, Yuxuan Liang, Katherine Tieu, Dongqi F u, Jun W u, Jingrui He and Yang Wang. Maintaining the status quo: Cap- turing invariant relations for OOD spatiotemporal learning. In Ambuj K. Singh, Yizhou Sun, Leman...

    2023

  61. [71]

    [Yes] (b) An analysis of the properties and complexity (time, space, sample size) of any algorithm

    For all models and algorithms presented, check if you include: (a) A clear description of the mathematical set- ting, assumptions, algorithm, and/or model. [Yes] (b) An analysis of the properties and complexity (time, space, sample size) of any algorithm. [Yes] (c) (Optional) ...

  62. [72]

    [Yes] (b) Complete proofs of all theoretical results

    For any theoretical claim, check if you include: (a) Statements of the full set of assumptions of all theoretical results. [Yes] (b) Complete proofs of all theoretical results. [Yes] (c) Clear explanations of any assumptions. [Yes]

  63. [73]

    [Not Applicable] (b) All the training details (e.g., data splits, hy- perparameters, how they were chosen)

    For all figures and tables that present empirical results, check if you include: (a) The code, data, and instructions needed to re- produce the main experimental results (either in the supplemental material or as a URL). [Not Applicable] (b) All the training details (e.g., dat...

  64. [74]

    [Yes] (b) The license information of the assets, if appli- cable

    If you are using existing assets (e.g., code, data, models) or curating/releasing new assets, check if you include: (a) Citations of the creator If your work uses existing assets. [Yes] (b) The license information of the assets, if appli- cable. [Yes] (c) Newassetseitherinthes...

  65. [75]

    [Not Applicable] (b) Descriptions of potential participant risks, withlinkstoInstitutionalReviewBoard(IRB) approvals if applicable

    If you used crowdsourcing or conducted research with human subjects, check if you include: (a) The full text of instructions given to partici- pants and screenshots. [Not Applicable] (b) Descriptions of potential participant risks, withlinkstoInstitutionalReviewBoard(IRB) appr...

  66. [76]

    Data Mining

    = I({e}t−1 1 ; Gt, {G}t−1 1 ) = I({e}t−1 1 ; {G}t−1 1 ) + I({e}t−1 1 ; Gt | {G}t−1 1 ) = I({e}t−1 1 ; {G}t−1 1 ) (24) As et is a subset of Gt, so et could be regarded as the result of a noisy function ofGt, i.e et = f (Gt, ϵ), with some noise ϵ. So when Gt is observed, et beco...

    1990

  67. [2013]

    Proceedings of a meeting held December 5-8, 2013, Lake Tahoe, Nevada, United States, pages 3111–3119, 2013

    2013

  68. [2022]

    OpenReview.net, 2022

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.