Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

GCAL: Adapting Graph Models to Evolving Domain Shifts

T0 review · 2 major / 6 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 A pretrained graph model can be continually adapted to unlabeled out-of-distribution graphs without catastrophic forgetting by replaying compact generated memory graphs.

desk verdict The assembled method is solid and the ablations are honest, but the 'substantially outperforms existing methods' claim is unverified because the closest continual graph-memory baselines are cited yet absent from the comparison. read the letter →

arxiv 2505.16860 v1 pith:PTZFYV5P submitted 2025-05-22 cs.LG cs.AI

Ziyue Qiao , Qianyi Cai , Hao Dong , Jiawei Gu , Pengyang Wang , Meng Xiao , Xiao Luo , Hui Xiong This is my paper
classification cs.LGcs.AI
keywords continuallearninggraphdomainadaptationtest-timecatastrophicforgettingmemoryreplaycondensationinformationbottleneckneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a pretrained graph neural network can keep adapting to an unlabeled stream of out-of-distribution graphs—each new graph drawn from a different distribution—without falling into catastrophic forgetting. The proposed GCAL method does this with a bilevel loop: an adaptation phase applies an information-maximization objective to the new graph while replaying small synthetic memories of earlier graphs, and a memory-generation phase learns a compact variational memory graph for the current graph, guided by a lower bound derived from information-bottleneck theory. On four evolving graph datasets spanning regional and temporal shifts, GCAL reports higher average performance and better retention than existing one-step domain adaptation and continual test-time adaptation methods, with every ablated component contributing to the result. If the central claim is right, graph models can be reused sustainably on changing unlabeled graph data without needing labels for replay or full retraining.

What carries the argument

The load-bearing mechanism is the variational memory graph generator, a GNN whose top-$k$ selector reduces the current graph to $K$ latent node distributions; node features are sampled by reparameterization and edges are built from Gumbel-softmax Bernoulli weights. The generator is trained with three losses aligned to the lower bound in Theorem 3.1: a gradient-matching condensation loss that makes the memory graph mimic the current graph's training signal, a KL regularization loss pushing the latent distribution toward a Bernoulli-Gaussian prior, and a generation loss comparing summed hidden representations of the memory graph and the original graph. These losses are nested in a bilevel optimization: the model parameters adapted in the replay phase define the classifier used to supervise the memory graph's condensation loss.

What would settle it

Compute the exact negative log-likelihood of $\hat{G}_t$ under the generator's Bernoulli and Gaussian distributions and compare it with the squared-discrepancy generation loss across training steps; if the rankings disagree, Equation 14 is not implementing the third term of Equation 5, and the Theorem 3.1 bound no longer applies to the objective actually optimized.

Watch

Extended reading notes

Core claim

The paper's central claim is that catastrophic forgetting during graph domain adaptation is preventable without labels, provided the model stores and replays compact generated memory graphs. Each arriving graph $G_t$ is condensed by a variational generator $g(\Phi)$ into a much smaller graph $\hat{G}_t$ with $K \ll N_t$ nodes, trained toward the information-bottleneck goal $\max I(\hat{G}_t; \hat{Y}_t) - \beta I(\hat{G}_t; G_t)$ via a derived lower bound. The adaptation phase then minimizes the same information-maximization loss on the new graph and on all stored memories, so one unsupervised objective both adapts to the new domain and reinforces old ones. In the reported experiments, GCAL achieves the best average performance and average forgetting on Twitch-Explicit, Facebook-100, Elliptic, and OGB-Arxiv, and its ablations show that the regularization loss, generation loss, and exponential moving average updates each contribute positively.

Load-bearing premise

The entire forgetting-prevention claim rests on assuming that the generation loss—a squared difference between summed hidden node representations of the memory graph and the original graph—faithfully implements the log-probability term of the theoretical lower bound; if that proxy drops information that future replay needs, the retention guarantee can break.

Editorial extensions

If this is right

  • A pretrained GNN can be updated online on unlabeled graph streams while retaining past-domain performance, so continual domain adaptation needs neither labels nor full retraining.
  • Because each memory graph uses only a small fraction of the original nodes, replay storage stays compact even as the number of domains grows.
  • The consistent gains over single-step adaptation baselines imply that memory replay, not stronger domain alignment, is the critical missing ingredient in graph continual adaptation.
  • Removing any of the regularization loss, generation loss, or EMA updates lowers average performance on every dataset, so the full training recipe is needed for the reported result.
  • The same framework works for regional shifts and temporal shifts, suggesting the mechanism is not tied to a particular kind of domain drift.

Reading between the lines

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

  • Because the generation term is implemented as a summed-representation discrepancy rather than the exact log-probability of the generator, the information-bottleneck derivation is better read as motivation for the loss design than as a tight guarantee on what is optimized; a stricter guarantee would require bounding the discrepancy by the log-likelihood.
  • The memory graphs are learned from the current model's soft pseudo-labels, so noisy early-domain predictions could be baked into the stored memories; an extension would be to weight replay by prediction confidence.
  • As the number of domains grows, the stored memory pool grows linearly with the domain count, so at very long streams the compact-per-graph memories may still dominate the budget; testing larger domain counts would reveal the practical ceiling.
  • The generated-graph visualizations suggest memories preserve coarse structure while dropping density, so a natural experiment is to check whether structure-aware fidelity measures predict replay value better than node-feature fidelity does.
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

2 major / 6 minor

Summary. The paper proposes GCAL, a method for unsupervised continual adaptation of graph models to a sequence of evolving out-of-distribution graphs. At each step, GCAL performs test-time adaptation by information maximization on the new graph while replaying previously generated memory graphs, and then generates a small synthetic memory graph for the current graph using a variational GNN with a top-k selector, Gumbel-softmax edge sampling, and three losses derived from an information-bottleneck-style lower bound. The method is evaluated on Twitch-explicit, Facebook-100, Elliptic, and OGB-Arxiv, where it reports higher Average Performance (AP) and better Average Forgetting (AF) than eight baselines, with ablations, hyperparameter sensitivity, and backbone analyses. The central claim is that GCAL substantially outperforms existing methods in both adaptability and knowledge retention.

Significance. If the empirical claim holds, GCAL addresses a genuine and underexplored problem: unsupervised continual adaptation of graph models under evolving distribution shifts, where most prior graph domain adaptation methods are single-step and supervised continual graph learning methods do not apply. The paper has several strengths: it releases code, reports five-run means and standard deviations, provides ablations for each loss component, tests multiple GNN backbones, and includes a proof of the proposed lower bound. The variational memory generation module is a sensible design and the replay-without-labels idea is practically motivated. However, the significance is currently tempered by two gaps: the closest continual graph-memory baselines cited in the paper are absent from the comparison table, and the implemented generation loss is not actually connected to the derived lower bound. Both need to be addressed before the stated 'substantially outperforms' claim can be accepted.

major comments (2)
  1. [§4.2.1, Table 2] The central empirical claim that GCAL 'substantially outperforms existing methods' is not yet supported because Table 2 omits the closest competitors cited in the manuscript itself: CaT (Liu et al., 2023b), PUMA (Liu et al., 2023c), and Qiao et al. (ICLR 2025). These are memory-replay and graph-condensation methods for continual graph learning discussed in §D.1 and §D.3, and Qiao et al. (2025), by the same research group, appears to target the same unsupervised continuous reuse setting. None of them appears in Table 2 or Figure 3. If PUMA and CaT require supervised labels, the paper should state that explicitly and, where feasible, adapt them to the unsupervised protocol; Qiao et al. (2025) in particular should be compared directly or its exclusion justified in detail. Without these comparisons, the abstract and §4.2.1 overstate the state-of-the-art claim.
  2. [§3.2.1 (Theorem 3.1, Eq. 5) and §3.2.5 (Eq. 14)] The implemented generation loss is not the objective derived in Theorem 3.1. Equation 14 minimizes the squared L2 distance between summed hidden representations of the memory graph and the original graph, whereas the theorem's final term is E[log Pg(Ĝt|Gt, Zt)], a log-likelihood under the variational generator. No derivation shows that Eq. 14 is proportional to, or bounds, that log-likelihood. Moreover, the two β-dependent terms in Eq. 5 cancel algebraically: for a fixed generator distribution Pg, -βE[KL(Pg(Ĝt|Gt,Zt)||Q(Ĝt))] + βE[log Pg(Ĝt|Gt,Zt)] equals βE[log Q(Ĝt)], a prior cross-entropy that contains no term encouraging preservation of information from Gt. Thus the lower bound in Theorem 3.1 does not justify LGen as an information-preservation objective, and the claim that the generator is 'guided by a theoretical lower bound' is not established. The Table 3 ablation shows removing LGen costs only 0.3–0.9 points, so this is not the primary empirical weakness, but the theory-practice gap is load-bearing for the method's rationale and should be fixed by either deriving a bound for Eq. 14 or reformulating the theoretical claim.
minor comments (6)
  1. [§4.2.2 and Figure 4 caption] The text says GCAL 'predominantly displays lighter shades' compared to CoTTA in Figure 4, while the caption states that darker shades signify better performance; if lighter is worse, the sentence contradicts the claimed advantage, and if the figure or the sentence is wrong, the visual comparison needs to be corrected.
  2. [Figure 5] The horizontal axis is labeled 'Prompt Ratio' while the text and the surrounding discussion refer to the synthetic-node ratio K/Nt; the axis label should match the terminology used in §4.2.4.
  3. [Table 2] The caption mentions 'Full' as the upper bound, but no Full row appears in the table; the row should either be added or the caption should be adjusted.
  4. [§3.2.5, Eq. (14)] The notation \bu_i(Θ) and u_i(Θ) is not defined in the text, and the comparison of a sum over K generated-node representations with a sum over Nt original-node representations deserves a short explanation of how the two sets are aligned or pooled.
  5. [Appendix B and §4.3] There are several editorial slips: 'Detialed' in the appendix title, 'A void Forgetting' in §2, and 'mode parameters' where 'model parameters' is meant; these should be corrected.
  6. [§4.2.1, Table 2] For EERM and GTrans, AF is listed as N/A because they train new parameters at each graph, but AP is still reported; the paper should clarify how AP is computed for methods that cannot evaluate the final model on previous domains.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the IB lower bound is a self-contained variational derivation and the empirical claims are judged on external labeled metrics against external baselines.

full rationale

The derivation chain in GCAL is not circular. Theorem 3.1 (Eq. 5) is obtained by the standard variational decomposition of mutual information: dropping the nonnegative entropy term H(Yhat) lower-bounds I(Ghat;Yhat), and dropping the nonnegative KL(P(Ghat)||Q(Ghat)) lower-bounds -I(Ghat;Gt,Zt), while H(Ghat|Gt)>=0 lower-bounds I(Ghat;Zt|Gt). Each inequality uses only nonnegativity of entropy/KL, not the claim being proved, so the bound is an independent mathematical statement rather than an equation that presupposes its own conclusion. The three implemented losses are surrogates for the three bound terms: L_MGL is a gradient-matching approximation to E[log Pf(Yhat|Ghat)], L_Reg is the closed-form KL term, and L_Gen is a squared-distance proxy for E[log Pg(Ghat|Gt,Zt)]. The L_Gen proxy is a heuristic approximation rather than an equivalence, but this is a correctness/implementation gap, not a circular reduction; moreover the ablation (Table 3) shows removing LGen costs only 0.3-0.9 points, so the forgetting-prevention claim does not reduce to that proxy. The central empirical claim is evaluated with labeled AP/AF metrics on held-out test data against external methods (Tent, CoTTA, EATA, EERM, etc.), so the optimization objective (entropy minimization plus memory replay) is not the same quantity being measured. Self-citations such as Qiao et al. (ICLR 2025) appear in related work but are not used to justify the IB derivation or the benchmark comparisons. The absence of PUMA, CaT, and the authors' own ICLR 2025 method from Table 2 is a legitimate concern about the strength of the 'substantially outperforms' claim, but it is an empirical comparison gap, not a circularity of the derivation. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and no known result is merely relabeled. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method relies on several untested modeling assumptions: the information bottleneck objective is taken as the right memory-generation criterion, pseudo-labels from information maximization stand in for true labels, the Gaussian/Bernoulli generator is assumed expressive enough, and the variational prior is chosen for tractability. The hyperparameters β, λ1, λ2, K, q, and τ are tuned per dataset, not fixed by the theory.

free parameters (5)
  • β (information bottleneck trade-off) = not reported
    Balances compression versus task relevance in Eq. 3 and Theorem 3.1; value is not listed in the paper and is presumably tuned per dataset.
  • λ1, λ2 (loss weights) = not reported
    Combine regularization and generation losses in Eq. 15; tuned per dataset, values deferred to code.
  • K (memory graph size) = 0.01 to 0.13 of original nodes depending on dataset
    Chosen per dataset; the hyperparameter experiments in Figure 5 scan K, confirming it is data-dependent.
  • q (Bernoulli prior for edges) = not reported
    Prior probability for each edge in the memory graph (Eq. 13); chosen by hand, value not stated.
  • τ (Gumbel temperature) = not reported
    Controls binarization of edges in Eq. 8; set by hand, value not stated.
assumptions (4)
  • domain assumption Graph information bottleneck objective (Eq. 3) is the correct criterion for memory graph generation.
    The whole generator is designed to optimize this objective; the paper does not justify why compressing I(Ghat;Gt) while maximizing I(Ghat;Yhat) yields memories useful for future adaptation.
  • domain assumption Information maximization provides reliable pseudo-labels for both adaptation and memory condensation.
    Eq. 1 is used as the training signal in Eq. 9 and Eq. 10 although no label and no calibration analysis is provided.
  • domain assumption The variational generator parameterization (Gaussian node features, Bernoulli edges) can express informative memory graphs.
    Eq. 6-8 assume this flexibility without capacity or expressiveness arguments.
  • domain assumption The variational prior Q(Ghat) = Bernoulli(q) x N(0,I) is a suitable reference distribution.
    Chosen for tractability in Eq. 12-13; no evidence it matches properties of real graphs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GCAL: Adapting Graph Models to Evolving Domain Shifts." pith.science (2026). https://pith.science/paper/PTZFYV5P

@misc{pith2026250516860,
  author       = {Pith},
  title        = {Pith review of: GCAL: Adapting Graph Models to Evolving Domain Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PTZFYV5P}},
  note         = {Machine review of arXiv:2505.16860}
}
read the original abstract

This paper addresses the challenge of graph domain adaptation on evolving, multiple out-of-distribution (OOD) graphs. Conventional graph domain adaptation methods are confined to single-step adaptation, making them ineffective in handling continuous domain shifts and prone to catastrophic forgetting. This paper introduces the Graph Continual Adaptive Learning (GCAL) method, designed to enhance model sustainability and adaptability across various graph domains. GCAL employs a bilevel optimization strategy. The "adapt" phase uses an information maximization approach to fine-tune the model with new graph domains while re-adapting past memories to mitigate forgetting. Concurrently, the "generate memory" phase, guided by a theoretical lower bound derived from information bottleneck theory, involves a variational memory graph generation module to condense original graphs into memories. Extensive experimental evaluations demonstrate that GCAL substantially outperforms existing methods in terms of adaptability and knowledge retention.

Figures

Figures reproduced from arXiv: 2505.16860 by the authors.

Figure 1
Figure 1. (a) The challenge of continual adaptation of graph mod￾els on evolving OOD graph sequences. (b) Empirical evaluations of the SOTA graph adaptation method across four OOD graph datasets in a continual adaptation setting. tricate connections, have been pivotal in advancing data min￾ing and knowledge discovery. Classic graph models such as Graph Convolutional Networks (GCNs) (Kipf & Welling, 2016) and Graph Attention N… view at source ↗
Figure 2
Figure 2. The GCAL framework involves several steps: Starting with the graph model f(Θt−1), the current graph Gt, and the accumulated memory graph pool G = {Gbi} t−1 i=1, GCAL first applies the Adaptation with Memory Replay method using loss LAMR for model adaptation. Next, a Variational Memory Generator creates a new memory graph Gbt for Gt, which is refined using the memory graph learning loss LMGL, the regularization losse… view at source ↗
Figure 3
Figure 3. Dynamics of the average performance during continual adaptation on evolving OOD graphs. ages past insights for improved adaptation but also enriches the training process, boosting the model’s adaptability and robustness across evolving domains. DANN achieves rela￾tively high performance among the baselines as it uses the source graph to guide the adaptation, while others only use the target graph data. Notably, our … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Performance matrices of GCAL and CoTTA in different datasets [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The model performance with different synthetic ratios. with K nodes for memory replay, where K is much less than the number of nodes in the initial graph. Thus, we set K as 0.01, 0.03, 0.05, 0.07, 0.09 of the num￾ber of nodes for the datasets Facebook-100 and OGB￾Arxiv…
Figure 6
Figure 6. Figure 6: Visualized comparison of the original graphs (the first line) and generative graphs (the second line) of GCAL in the Twitch dataset. reflects the dynamic and challenging nature of financial transactions. Our training strategy begins by pretraining the model on selected…
Figure 7
Figure 7. Figure 7: Performance matrices of GCAL and CoTTA in different datasets. GCN GSAGE GAT GIN 50 52 54 56 58 60 Average AUC Test GCAL (a) Twitch-explicit GCN GSAGE GIN 50.0 50.5 51.0 51.5 52.0 52.5 53.0 53.5 54.0 Average ACC Test GCAL (b) Facebook-100 GCN GSAGE GAT GIN 50 55 60 65 7…
Figure 8
Figure 8. Figure 8: The model performance with different GNN backbones. of the Facebook and Ogbn-arxiv datasets in [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Cross-Resolution Semantic Learning for Graph Domain Adaptation

    cs.LG 2026-07 conditional novelty 6.0 of 10

    CReSL improves graph domain adaptation by learning cross-resolution source-to-target routing and grafting target representations toward source class prototypes.

Reference graph

Works this paper leans on

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

  1. [1]

    Multimodal continual graph learning with neural architecture search

    Cai, J., Wang, X., Guan, C., Tang, Y ., Xu, J., Zhong, B., and Zhu, W. Multimodal continual graph learning with neural architecture search. In Proceedings of the ACM Web Conference 2022, pp. 1292–1300,

    work page 2022

  2. [3]

    K., Roy, D

    Dziugaite, G. K., Roy, D. M., and Ghahramani, Z. Training generative neural networks via maximum mean discrep- ancy optimization. arXiv preprint arXiv:1505.03906 ,

    arXiv

  3. [6]

    reflects the dynamic and challenging nature of financial transactions

    Visualized comparison of the original graphs (the first line) and generative graphs (the second line) of GCAL in the Twitch dataset. reflects the dynamic and challenging nature of financial transactions. Our training strategy begins by pretraining the model on selected graphs from each dataset, then continually adapting to the remaining unlabeled graph da...

    work page 2017

  4. [8]

    Using our framework demonstrates remarkable enhancements, showing the effectiveness of our proposed techniques

    We compare the results of incorporating these GNN backbones within our framework versus utilizing them individually as standalone models. Using our framework demonstrates remarkable enhancements, showing the effectiveness of our proposed techniques. Lastly, it is important to note that the consistent use of different backbones significantly enhances resul...

    work page 2022

  5. [9]

    Revisit- ing batch normalization for practical domain adaptation

    Li, Y ., Wang, N., Shi, J., Liu, J., and Hou, X. Revisit- ing batch normalization for practical domain adaptation. arXiv preprint arXiv:1603.04779,

    arXiv

  6. [11]

    Structural re-weighting improves graph domain adap- tation

    Liu, S., Li, T., Feng, Y ., Tran, N., Zhao, H., Qiu, Q., and Li, P. Structural re-weighting improves graph domain adap- tation. In International Conference on Machine Learning, pp. 21778–21793. PMLR, 2023a. Liu, Y ., Qiu, R., and Huang, Z. Cat: Balanced continual graph learning with graph condensation. In 2023 IEEE International Conference on Data Mining ...

    arXiv 2023

  7. [12]

    and Savarese, S

    Sener, O. and Savarese, S. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489,

    arXiv

  8. [13]

    Single- view graph contrastive learning with soft neighborhood awareness

    Sun, Q., Chen, C., Qiao, Z., Zheng, X., and Wang, K. Single- view graph contrastive learning with soft neighborhood awareness. In AAAI 2025,

    work page 2025

Show all 28 references
  1. [14]

    All in one: Multi-task prompting for graph neural networks

    Sun, X., Cheng, H., Li, J., Liu, B., and Guan, J. All in one: Multi-task prompting for graph neural networks. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 2120–2131, 2023b. Sun, X., Zhang, J., Wu, X., Cheng, H., Xiong, Y ., and L...

    arXiv

  2. [15]

    Graph attention networks.arXiv preprint arXiv:1710.10903,

    Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y . Graph attention networks.arXiv preprint arXiv:1710.10903,

    arXiv

  3. [16]

    Does graph prompt work? a data operation perspective with theoretical analysis

    Wang, Q., Sun, X., and Cheng, H. Does graph prompt work? a data operation perspective with theoretical analysis. arXiv preprint arXiv:2410.01635, 2024b. Wang, X., Duan, M., Li, J., Ma, A., Xin, G., Xu, D., Li, Z., Liu, B., and Ma, Q. Marsgt: Multi-omics analysis for rare popul...

    arXiv

  4. [18]

    log P ( bGt|Gt, Zt) Q( bGt) # + KL(P ( bGt) ∥ Q( bGt)) ≥ −E bGt,Gt,Zt

    12 GCAL: Adapting Graph Models to Evolving Domain Shifts A. Proof of Theorem 3.1. In the section of Variational Memory Graph Generation, we give Theorem 1 to define a lower bound of the information bottleneck for generating memory graphs bGt from the original graph Gt: Theorem...

    2012

  5. [19]

    These networks vary greatly in size, density, and degree distribution

    Regional Shifts: The Facebook-100 dataset comprises 100 snapshots of Facebook friendship networks from 2005, each representing users from a specific American university. These networks vary greatly in size, density, and degree distribution. Additionally, the Twitch-Explicit da...

    2005

  6. [20]

    For baselines not originally designed for graphs, their architectures have been adapted to GCNs to ensure consistency in evaluation

    enhances adaptation efficiency through entropy minimization and employs a Fisher-based regularizer to maintain performance across domain shifts. For baselines not originally designed for graphs, their architectures have been adapted to GCNs to ensure consistency in evaluation....

    2011

  7. [23]

    and RieGrace (Sun et al., 2023a), which focus on maintaining essential parameters and structural topologies. Parameter isolation techniques allocate distinct parameters for new tasks to maintain those relevant to prior tasks (Niu et al.; Zhang et al., 2023a; 2022a), as seen in...

    2024

  8. [26]

    Continual Test-Time Adaptation (CTTA), a critical facet of Continual Domain Adaptation, addresses the unique demands of non-static domains

    leverage graph topology to improve adaptability, reducing discrepancies between source and target graphs via local and global consistencies and a graph domain discriminated loss, respectively. Continual Test-Time Adaptation (CTTA), a critical facet of Continual Domain Adaptati...

    2022

  9. [27]

    and meta-networks in EcoTTA (Song et al., 2023), contribute to improved model normalization and adaptability. Despite these advancements, challenges such as noisy pseudo-labels and calibration issues persist, and a notable gap remains in unsupervised graph continual domain ada...

    2023

  10. [28]

    Among recent innovations, GCDM (Liu et al., 2022), introduce graph-specific distribution alignment to enhance condensation effectiveness

    utilize gradient alignment to synthesize representative samples that maintain the statistical properties of the original data, drawing from principles of traditional sampling (Sener & Savarese, 2017). Among recent innovations, GCDM (Liu et al., 2022), introduce graph-specific ...

    2017

  11. [2015]

    and adversarial techniques (Dan et al., 2024; Qiao et al., 2023; Tzeng et al., 2017; Zhang et al., 2018), form the basis of this field. In graph-based domain adaptation, a variety of methods have been proposed (Ding et al., 2018; Jin et al., 2022; Liu et al., 2023a; Ma et al.,...

    2024

  12. [2016]

    Gao, X., Chen, T., Zang, Y ., Zhang, W., Nguyen, Q. V . H., Zheng, K., and Yin, H. Graph condensation for inductive node representation learning. In 2024 IEEE 40th Inter- national Conference on Data Engineering (ICDE) , pp. 3056–3069. IEEE,

    2024

  13. [2017]

    A., and Jin, W

    Hashemi, M., Gong, S., Ni, J., Fan, W., Prakash, B. A., and Jin, W. A comprehensive survey on graph reduc- tion: Sparsification, coarsening, and condensation. arXiv preprint arXiv:2402.03358,

    arXiv

  14. [2018]

    Hierarchical prototype net- works for continual graph representation learning

    Zhang, X., Song, D., and Tao, D. Hierarchical prototype net- works for continual graph representation learning. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 45(4):4622–4636, 2022a. Zhang, X., Song, D., and Tao, D. Sparsified subgraph memory for continual g...

    2022 arXiv

  15. [2020]

    Praga: Prototype-aware graph adaptive aggregation for spatial multi-modal omics anal- ysis

    9 GCAL: Adapting Graph Models to Evolving Domain Shifts Huang, X., Ma, Z., Meng, D., Liu, Y ., Ruan, S., Sun, Q., Zheng, X., and Qiao, Z. Praga: Prototype-aware graph adaptive aggregation for spatial multi-modal omics anal- ysis. In AAAI 2025,

    2025

  16. [2021]

    Graph condensation via receptive field distribution matching

    Liu, M., Li, S., Chen, X., and Song, L. Graph condensation via receptive field distribution matching. arXiv preprint arXiv:2206.13697,

    arXiv

  17. [2022]

    Kipf, T. N. and Welling, M. Semi-supervised classifica- tion with graph convolutional networks. arXiv preprint arXiv:1609.02907,

    arXiv

  18. [2023]

    Adaptive path-memory network for tempo- ral knowledge graph reasoning

    Dong, H., Ning, Z., Wang, P., Qiao, Z., Wang, P., Zhou, Y ., and Fu, Y . Adaptive path-memory network for tempo- ral knowledge graph reasoning. In Proceedings of the Thirty-Second International Joint Conference on Artifi- cial Intelligence, pp. 2086–2094,

  19. [2024]

    Graph condensation for graph neural networks

    Jin, W., Zhao, L., Zhang, S., Liu, Y ., Tang, J., and Shah, N. Graph condensation for graph neural networks. arXiv preprint arXiv:2110.07580,

    arXiv

  20. [2025]

    archive and revisit representative data from past tasks to alleviate the critical issue of catastrophic forgetting, as exemplified by ER-GNN (Zhou & Cao, 2021), SSM (Zhang et al., 2022b), SEM-curvature (Zhang et al., 2023b), PDGNNs (Zhang et al., 2024), and CaT (Liu et al., 20...

    2021

Pith tools

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