acceptodds
Under review as a conference paper at ICLR 2027

Composing Priors for Temporal Graph Learning

14 views

Abstract

Temporal graph models often perform well on individual datasets yet struggle to generalize the same architecture across multiple datasets or domains. We argue that this transfer gap stems from a common architectural choice: jointly learning temporal dynamics and structural propagation entangles domain-specific spatio-temporal dependencies. We show that temporal and structural priors can instead be effectively composed before graph-specific learning. Building on this insight, we introduce a framework that decouples temporal and structural representation learning and equips a GNN with three complementary, domain-agnostic priors: foundation-model temporal embeddings, local temporal-context summaries, and supra-Laplacian positional encodings of the evolving topology. Rather than asking a recurrent architecture to rediscover temporal dynamics and structural propagation for each domain, our framework learns how reusable representations interact over the graph. Across multiple real-world temporal graph benchmarks, it achieves competitive or state-of-the-art supervised performance. On the MiNT zero-shot benchmark, our framework consistently mitigates negative transfer and outperforms strong temporal graph baselines across 20 unseen networks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural network, temporal graph

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yQ8COpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MOSAIC, a temporal graph learning framework that decouples temporal dynamics learning from structural graph learning. MOSAIC composes three complementary priors: temporal embeddings extracted from frozen time-series foundation models, local temporal-context summaries, and supra-Laplacian positional encodings that capture graph topology. These representations are integrated through a GNN with residual connections and trajectory-aware layer aggregation, allowing spatial relationships to be learned without backpropagation through time. The framework is evaluated on supervised node forecasting and classification tasks, where it achieves competitive or state-of-the-art performance across multiple benchmarks. The authors further investigate zero-shot cross-network transfer on the MiNT benchmark, demonstrating improved average AUC and reduced negative transfer across 20 unseen networks compared with recurrent temporal graph baselines.

Critical strengths
  1. The paper presents an interesting and well-motivated perspective on temporal graph learning by decoupling temporal representation extraction from structural graph learning. The proposed composition of pretrained temporal embeddings, local temporal statistics, and supra-Laplacian positional encodings provides a modular approach to incorporating complementary inductive biases, offering potential advantages in training efficiency and transferability across different graph domains.
  2. The experimental evaluation covers a diverse range of temporal graph learning tasks, including node-level forecasting, node classification, and zero-shot cross-network transfer. The supervised experiments demonstrate competitive performance against various representative baselines, while the evaluation on 20 unseen MiNT networks provides additional evidence of the framework's potential for cross-network generalization.
  3. The paper provides relatively comprehensive analyses of the proposed framework, including computational analysis and sensitivity to different pretrained time-series foundation models and supra-Laplacian configurations.
Critical weaknesses
  1. Insufficient validation of prior complementarity in zero-shot transfer. A central contribution of the paper is to utilize three complementary priors to improve cross-network generalization. However, the MiNT benchmark consists of featureless interaction graphs with constant node features. Under this setting, the temporal foundation-model embeddings and local temporal summaries receive essentially constant input signals, while the supra-Laplacian positional encodings contribute most. This raises the question of whether the observed transfer gains primarily originate from structural representations rather than the proposed composition of complementary priors.
  2. Insufficient characterization of dataset-dependent transfer behavior. Substantial performance discrepancies between single-network supervised training and zero-shot transfer on several datasets are observed, yet these observations are not adequately investigated. For example, MOSAIC achieves an AUC of only 0.098 under supervised training on BAG, compared with 0.972 under zero-shot transfer, whereas SDEX exhibits the opposite behavior, dropping from 0.954 to 0.454. Although the authors interpret it as evidence of substantial positive transfer, detailed dataset-specific evolving pattern analysis is expected.
  3. Limited supervised performance on forecasting benchmarks. Although MOSAIC achieves strong results on several node-level prediction tasks, its performance is less convincing on traffic forecasting benchmarks. Simply attribute them to architectural trade-offs and inductive biases is not enough to prove whether the performance gaps arise from the decoupling of temporal and structural learning.
Questions for authors
  1. Could authors further investigate the contributions of the three priors to zero-shot transfer on MiNT? An ablation study in the zero-shot setting would help clarify whether the observed improvements arise from all the proposed priors.
  2. Could authors provide further explanations for the substantial discrepancies between supervised and zero-shot performance on certain MiNT networks, particularly BAG and SDEX?
  3. Please provide more empirical evidence to clarify why architectural trade-offs and inductive biases are the major causes of MOSAIC's suboptimal performance in forecasting tasks.
3 comments
AuthorsOpenReview ↗

Response to W1 and Q1

We thank you for your careful evaluation. We are encouraged by your recognition that the paper presents an interesting and well-motivated perspective on temporal graph learning, that the composition of priors provides a modular approach to incorporating complementary inductive biases, and that it offers relatively comprehensive analyses of the proposed framework. We address each point below, following our summary.

Overall summary of our responses: we include (i) an ablation on MiNT, complemented by new zero-shot traffic experiments where all priors are informative and a further ablation in this setting, (ii) a per-network analysis of the MiNT results showcasing the motivation behind strong positive/negative transfer, and (iii) an analysis on the performance limitation on traffic forecasting.

Regarding W1 and Q1

Thank you for this observation. Since MiNT contains featureless D-TDGs, Φs\Phi_s and Φf\Phi_f receive constant inputs, and Φp\Phi_p is the only prior that varies across nodes and snapshots. Ablating Φf\Phi_f or Φs\Phi_s on MiNT would thus only remove a constant input. Therefore, we ablate the choice of the structural prior. We replace the supra-Laplacian Φp\Phi_p with Laplacian Positional Encodings (LapPE) [1], computed on each snapshot independently. We pre-train MOSAIC with LapPE on the 64 MiNT networks and show in the table below the zero-shot test AUC$\uparrow$ on the 20 unseen networks, using the same hyperparameters of the full run (Appendix B.8 of the revision).

VariantAUC$\uparrow$
MOSAIC (full)0.744
with LapPE Φp\Phi_p0.676

Replacing the supra-Laplacian Φp\Phi_p with a per-snapshot LapPE lowers the AUC from 0.744 to 0.676, showing that coupling consecutive snapshots matters on evolving topologies. However, given that this comparison does not include a scenario where we pre-train MOSAIC exploiting all three priors, we also conduct a zero-shot traffic forecasting experiment where all are used. Specifically, we pre-train MOSAIC in a leave-one-out fashion on PeMS-D{5,6,10,11,12} [2] and three of PeMS0{3,4,7,8}, and evaluate on the left-out dataset. In this experiment, Φf\Phi_f is Chronos-Bolt small [3] over wf=864w_f=864 steps. We compare with graph-free baselines (TSFMs) and a recent spatiotemporal foundation model engineered for traffic prediction (OpenCity [5]). We report the performance (Masked MAE$\downarrow$) in the table below, in Table 2 and Section 4.3 of the revision, where †^\dagger denotes that the pre-training corpus of the model contains the evaluation dataset, or traffic data from the same district:

ModelPeMS03PeMS04PeMS07PeMS08
Chronos-Bolt small [3]17.4923.5226.0219.12
Chronos-2 [4]15.12$^\dagger$20.50$^\dagger$22.22$^\dagger$16.27$^\dagger$
OpenCity-Plus [5]16.1619.86$^\dagger$22.44$^\dagger$15.05$^\dagger$
MOSAIC zero-shot14.8820.2621.9815.64

MOSAIC outperforms all competitors where the setting is fully zero-shot, and is competitive with or outperforms them even where they have seen the target dataset during pre-training.

To fully answer your very important request, we ablate all priors as we did in Table 4, pre-training on PeMS-D{5,6,10,11,12} and PeMS0{3,4,8}, and evaluating zero-shot on PeMS07 (Masked MAE$\downarrow$, zero-shot column of Table 4 of the revision):

VariantPeMS07 Masked MAE$\downarrow$
MOSAIC zero-shot (full)21.98
Without Φf\Phi_f29.19
Without Φs\Phi_s116.24
Without Φp\Phi_p21.71
Only Φf\Phi_f78.47
Only Φs\Phi_s27.90
Only Φp\Phi_p177.86

As in Table 4, Φf\Phi_f and Φs\Phi_s are the most important priors in this setting: removing Φf\Phi_f raises the error from 21.98 to 29.19, and removing Φs\Phi_s to 116.24, since without the window statistics the model cannot recover the scale of an unseen road network. In contrast to the supervised setting, Φp\Phi_p does not contribute here (21.71 without it): the road networks have a static topology, so the pre-training corpus contains only 8 distinct graphs, and their positional encodings carry little information to transfer.

Putting all together, (i) on MiNT, where the graphs are featureless and the topology evolves, the supra-Laplacian form of Φp\Phi_p outperforms a per-snapshot positional encoding, and (ii) in zero-shot traffic forecasting, the model relies on the TSFM embeddings and on the local window. Together with Table 4, this isolates the contribution of each prior in the supervised and in the zero-shot setting. We thank the reviewer for allowing us to run these further experiments, which made the characterization of our results stronger and positioned our contribution better.

AuthorsOpenReview ↗

Response to W2, Q2, W3, Q3 and References

Regarding W2 and Q2

Thank you for raising this important question, which motivated a deeper investigation of the substantial discrepancies observed on BAG and SDEX. We analyzed three representative cases: strong positive transfer (BAG), near-zero positive transfer (HOICHI), and strong negative transfer (SDEX). The full dataset-specific analysis is provided in Appendix B.1.2 of the revision.

Our findings suggest that these discrepancies are not necessarily attributable to MOSAIC or training instability. In particular, similar effects occur across different architectures: pre-training improves the performance of all three models on BAG, while it also leads to negative transfer for HTGN on SDEX. Moreover, supervised performance is positively correlated across networks, suggesting that part of the difficulty is shared across models and stems from the underlying data.

We further find that these discrepancies are related to changes in the relationship between network activity and future growth across training and test periods. Although this relationship is predictive across most networks, its strength can vary substantially over time. For instance, it becomes much stronger on BAG, potentially explaining why zero-shot transfer substantially outperforms single-network supervised training. Conversely, this relationship nearly disappears on SDEX, providing a possible explanation for the substantial performance degradation under zero-shot transfer. To support this analysis, Figure 3 of the revised paper illustrates the evolution of network activity over time, which is particularly relevant because network activity directly defines the prediction targets.

Overall, these findings suggest that the contrasting behaviors on BAG and SDEX reflect, at least in part, network-specific temporal variability rather than systematic instability of MOSAIC. They also highlight that pre-training can provide substantial transfer gains, but does not guarantee positive transfer on every network.

Regarding W3 and Q3

Thank you for pointing out this issue. We show below that this gap comes mainly from the short context we provide to Φf\Phi_f, rather than from our proposed decoupling of temporal and structural learning. Φf\Phi_f is the embedding of a frozen pre-trained TSFM that processes its input in patches, so it needs a window long enough to contain meaningful patches in order to produce informative embeddings. Making the temporal component trainable would let the model adapt to short windows, but it contradicts the claims of our paper.

Moreover, as stated in [6], the time-then-graph paradigm is advantageous for expressivity compared to time-and-graph, so decoupling time from graph does not in itself restrict what can be represented. Instead, MOSAIC can exploit longer contexts at a flat training cost, since Φf\Phi_f is precomputed (empirical runtime and memory analysis support this, see Appendix B.6). To support this discussion, we compare the performance of MOSAIC and baselines with its long context version, MOSAIC-LC, where Φf\Phi_f is Chronos-Bolt small [3], since MOMENT reads at most 512 steps, and we set wf=864w_f = 864 and ws=12w_s = 12, showing the results in the table below (Masked MAE$\downarrow$):

VariantPeMS03PeMS04PeMS07PeMS08
Best baseline (Table 2)14.2218.2019.3213.58
MOSAIC14.6819.0819.6814.09
MOSAIC-LC (long context)14.3618.4919.2513.53

The gap between MOSAIC and the best baseline of Table 2 is reduced from a mean of 0.55 to a mean of 0.08 over PeMS03, 04, 07, 08, when using a longer context. We added this analysis in Appendix B.9 (Table 17) of the revised paper. Thank you.

We thank you once again, and hope that our clarifications and additional experiments will lead to a reconsideration of your score.


References:

[1] Dwivedi and Bresson. A generalization of transformer networks to graphs. AAAI Workshop on Deep Learning on Graphs, 2021.

[2] Chen et al. Freeway performance measurement system: mining loop detector data. Transportation Research Record, 2001.

[3] Ansari et al. Chronos: learning the language of time series. TMLR, 2024.

[4] Ansari et al. Chronos-2: from univariate to universal forecasting. Technical report, 2025.

[5] Li et al. Open spatio-temporal foundation models for traffic prediction. ACM TIST, 2025.

[6] Gao and Ribeiro. On the equivalence between temporal and static equivariant graph representations. ICML, 2022.

AuthorsOpenReview ↗

Did our responses address your concerns?

Dear Reviewer yQ8C,

Thank you again for your careful review and valuable suggestions.

We hope that our responses have helped clarify the contribution of MOSAIC, the role of each prior in zero-shot transfer, the dataset-dependent transfer on MiNT networks such as BAG and SDEX, and the role of the temporal context in its traffic forecasting performance.

We would appreciate knowing whether our responses have addressed your main concerns, and we welcome any further questions or suggestions.

Best regards,

The Authors

Reviewer_HSLUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MOSAIC, a discrete-time dynamic graph learning framework that separates temporal representation extraction from structural graph learning. It combines three complementary priors: foundation-model temporal embeddings, local temporal-context summaries, and supra-Laplacian positional encodings of the evolving topology. These representations are treated as fixed inputs to a static GNN, which learns their interactions across the graph without requiring backpropagation through time. Experiments show competitive performance across multiple supervised temporal graph benchmarks and consistent reductions in negative transfer across 20 unseen networks on the MiNT zero-shot benchmark.

Critical strengths

S1. MOSAIC separates temporal representation extraction from graph learning and combines three complementary priors in a modular framework.

S2. Experiments show competitive performance across supervised benchmarks and zero-shot transfer to 20 unseen networks, with ablations supporting the contributions of the three priors.

Critical weaknesses

W1. MOSAIC primarily integrates established techniques, such as the time-then-graph paradigm, supra-Laplacian positional encodings, and historical activation pooling. While their integration is well motivated, the methodological contribution appears incremental, with limited algorithmic novelty beyond combining these existing components.

W2. The discussion and evaluation of recent DTDG methods are incomplete. Relevant approaches include LAMP [1], DyAug [2], and PromptDyG [3]. In particular, LAMP is cited but not evaluated despite sharing the Wikipedia Math, Twitter Tennis, and ChickenPox benchmarks.

[1] Adaptive Memory Retention in Dynamic Graphs. ICML, 2026.

[2] Rationalizing and Augmenting Dynamic Graph Neural Networks. ICLR, 2025.

[3] PromptDyG: Test-Time Prompt Adaptation on Dynamic Graphs. ICML, 2026.

Questions for authors

Q1. Several minor errors should be corrected: “benfit” on page 9 should read “benefit,” and the citation in “even on busy roads Chen et al. (2001)” on page 6 should use the parenthetical form “(Chen et al., 2001).” Additionally, Appendix B.6.2 states that DCRNN and AGCRN run out of memory at W=96, whereas Table 16 marks DCRNN and GraphWaveNet as OOM and reports results for AGCRN.

Q2. Section 3.3 states that the method incurs “a total memory cost of O(nLd) per step that depends on none of T, wfw_f, or wpw_p.” However, Table 16 reports that MOSAIC’s peak GPU memory increases from 2.3 GB to 9.0 GB as the input window grows from 12 to 96. Could the authors explain this discrepancy?

7 comments
AuthorsOpenReview ↗

Weaknesses

We thank you for your careful and precise evaluation. We are encouraged by your recognition that MOSAIC separates temporal representation extraction from graph learning and combines three complementary priors in a modular framework, that the experiments show competitive performance across supervised benchmarks and zero-shot transfer to 20 unseen networks, and that their integration is well motivated.

In short: As requested by the reviewer, we include: (i) LAMP, a very recent state-of-the-art temporal GNN, in Table 1, (ii) a discussion of DyAug and PromptDyG in related work, and (iii) the corrections you pointed out in Q1 and a revised memory analysis. Overall, during the rebuttal we ran 6 new experiments and 3 new analyses, all listed in our general comment. We hope that you find our responses satisfactory, and that you will consider revising your score.

Regarding Weaknesses

W1 We thank the reviewer for this insightful observation and agree that MOSAIC builds upon existing methods, such as the time-then-graph paradigm [1], whose original contributions we duly acknowledge. However, combining established components into a novel architecture is itself a valuable contribution, as demonstrated by prior works such as Chronos [5], DCRNN [6], Graph WaveNet [7], and GraphGPS [8].

MOSAIC aims to show that carefully composing existing methods can yield an effective and efficient architecture. We demonstrate that its three priors are empirically complementary, as each of them is necessary: for every prior, there is a setting where removing it degrades performance (Table 4). Thus, MOSAIC achieves competitive or state-of-the-art performance in both supervised and zero-shot settings. Beyond the compositional framework itself, our contribution is a step towards foundational architectures for temporal graphs, whose heterogeneity in observation regimes, tasks, and distributions makes their design particularly challenging. We have clarified this aspect in the revised paper and sincerely thank you for raising it.

W2 Thank you for these suggestions.

(a) We add a very recent state-of-the-art model, LAMP [2], to Table 1, which we report here for your convenience:

ModelWiki MathTennisChickenPox
Best baseline0.542$\pm$0.0080.300$\pm$0.0030.907$\pm$0.007
LAMP0.444$\pm$0.0030.304$\pm$0.0010.748$\pm$0.006
MOSAIC0.440$\pm$0.0020.262$\pm$0.0020.745$\pm$0.003

LAMP lowers the previous best MSE by about 18% on Wiki Math and ChickenPox, and MOSAIC still outperforms it on all three datasets.

(b) We differentiate our contribution from DyAug [3] and PromptDyG [4] as follows. They are methods that improve dynamic GNN architectures, which learn temporal dynamics, for link prediction at different stages. DyAug is a training-time method that augments the training data of the backbone; PromptDyG is an inference-time method that freezes the backbone and updates a small prompt online, without labels, as new snapshots arrive. In contrast, MOSAIC acts on the representation of time itself: temporal information is precomputed once as priors, and the trained model is a static GNN with no temporal module, augmentation or test-time update, even zero-shot. Thus the three act on different parts of the pipeline: the training data, the inference-time parameters, and the representation of time. Moreover, unlike DyAug and PromptDyG, which target link prediction, MOSAIC is not restricted to this task: as demonstrated by our experiments, it supports a broader range of downstream tasks, including node-level classification, regression and forecasting, as well as graph-level classification, through appropriate readout training. We added a discussion of both in related work (Section 2, Transfer Learning on Dynamic Interaction Networks).

AuthorsOpenReview ↗

Questions and References

Regarding Questions

Q1 Thank you for the careful reading. We removed the sentence on page 9 that contained the typo, and corrected the citation format on page 6. We also fixed Appendix B.6.2, and we apologize for the confusion in the runtime analysis.

Q2 Thank you for this question, and we apologize for the lack of clarity in Section 3.3 and Appendix B.6.2. The claim concerns the trainable GNN, whose memory per training step is O(nLd)O(nLd) and independent of T,wf,wpT,w_f,w_p, as these windows are consumed during pre-processing: each sample holds one embedding Φf\Phi_f and one encoding Φp\Phi_p per node. Memory still depends on wsw_s, which contributes to the input dimension of the projection, ws+2k+dTSFMw_s+2k+d_{\text{TSFM}}, where kk is the number of retained structural eigenvectors and eigenvalues, and dTSFMd_{\text{TSFM}} is the TSFM embedding dimension. The growth in Table 16 came from our benchmark, which set ws=wf=wp=Ww_s=w_f=w_p=W and allocated Φf\Phi_f and Φp\Phi_p for every step of the window, although the model reads only the last one. We measured again the peak memory (GB) with one Φf\Phi_f and one Φp\Phi_p per sample, as in training:

WW12244896
Table 16 (submission)2.33.35.29.0
One prior per sample, ws=Ww_s=W1.361.381.421.50
One prior per sample, ws=12w_s=121.361.361.361.36

That is, memory is constant in wfw_f and wpw_p and grows only slightly with wsw_s, while the runtime is constant since we do not perform BPTT. We revised Section 3.3, Appendix B.6.2 and Table 16 accordingly.

We thank you once again, and hope that our clarifications will lead to a reconsideration of your score.


References:

[1] Gao and Ribeiro. On the equivalence between temporal and static equivariant graph representations. ICML, 2022.

[2] De Castelli et al. Adaptive memory retention in dynamic graphs. ICML, 2026.

[3] Zhang et al. Rationalizing and augmenting dynamic graph neural networks. ICLR, 2025.

[4] Ai et al. PromptDyG: Test-time prompt adaptation on dynamic graphs. ICML, 2026.

[5] Ansari et al. Chronos: Learning the language of time series. TMLR, 2024.

[6] Li et al. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. ICLR, 2018.

[7] Wu et al. Graph WaveNet for deep spatial-temporal graph modeling. IJCAI, 2019.

[8] Rampášek et al. Recipe for a general, powerful, scalable graph transformer. NeurIPS, 2022.

AuthorsOpenReview ↗

Did our responses address your concerns?

Dear Reviewer HSLU,

Thank you again for your careful review and valuable suggestions.

We hope that our responses have helped clarify the contribution of MOSAIC, its comparison with LAMP, its relationship to DyAug and PromptDyG, and the memory cost of its training.

We would appreciate knowing whether our responses have addressed your main concerns, and we welcome any further questions or suggestions.

Best regards,

The Authors

Reviewer_nVGFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes MOSAIC, a "time-then-graph" framework for discrete-time dynamic graphs. Node representations are built from three precomputed, parameter-free or frozen priors: (i) embeddings of each node's time series from a frozen time-series foundation model (MOMENT); (ii) local temporal summaries (the raw window or an adaptive average pooling of it); (iii) supra-Laplacian positional encodings of a sliding window of snapshots. These are projected, combined with a PE-conditioned gate, and processed by a residual GCN whose layer outputs are aggregated with a signed "historical activation pooling", followed by node-wise MHSA. Because temporal processing is done once in preprocessing, training needs no BPTT, and its cost does not grow with the input window. The paper argues that decoupling temporal and structural learning improves transfer across graphs. It evaluates supervised node regression, node classification, and zero-shot transfer on unseen token networks.

Critical strengths
  1. Simplicity and efficiency. The method is conceptually simple, modular, and cheap to train. Table 16 convincingly shows flat training time and modest memory growth in the input window, compared with recurrent spatio-temporal models that run out of memory at W = 96. A BPTT-free pipeline that is competitive with specialized models is practically relevant, and the ablations (Table 4) and sensitivity studies (Tables 13-14) are reasonably thorough.

  2. Strong results on several standard benchmarks. On Twitter Tennis, ChickenPox and Wikipedia Math the gains over the reported baselines are large, with low variance over 10 runs. On MiNT, the zero-shot variant obtains the best average AUC and rank among the compared zero-shot models.

Critical weaknesses
  1. The central claim about composing temporal and structural priors for transfer is not supported by the transfer experiment.
  • On MiNT the node features are constant (X = 1 for every node and snapshot, Sec. 4.3). Both temporal priors therefore receive a constant input and carry no information. The only informative input is the supra-Laplacian PE. The zero-shot experiment thus says nothing about the transferability of TSFM-based temporal priors, which is the paper's headline motivation ("temporal and structural priors can be composed before graph-specific learning").
  • Pretraining uses shuffled snapshots, "i.e., without enforcing temporal order", which makes it even less clear what temporal knowledge is being transferred.
  • The negative-transfer analysis (Table 6) measures each model's gain relative to its own single-network version.
  • The transfer setting is also a single domain. The abstract's claim about generalization "across multiple datasets or domains" is not tested.
  1. Novelty and fairness of the supervised comparisons.
  • Each component is taken from prior work: time-then-graph (Gao & Ribeiro, 2022), supra-Laplacian PE (Karmim et al., 2024; Galron et al., 2026a), historical activation pooling (Galron et al., 2026b), and off-the-shelf TSFMs. The contribution is mainly their combination, which places a higher bar on the empirical evidence.
  • The baselines in Table 1 are mostly older (2018–2023). Recent strong and simple baselines are missing: e.g. STID/STAEformer-style models for traffic, and, crucially, a graph-free baseline in which the same TSFM forecasts each node independently. Without the latter, it is unclear how much the graph component contributes at all.
Questions for authors

Clarifications:

  1. On MiNT, with X=1X = 1, what is the input to the TSFM and to Φs\Phi_s?
  2. Which input window wf/wsw_f/w_s was used for each PeMS result in Table 2?
  3. What is the performance of a graph-free baseline?
  4. Why does supervised MOSAIC fail on several MiNT networks? Is this a training instability, and does it affect the interpretation of the transfer gains in Table 6?
  5. How were hyperparameters selected for each dataset (validation split?), and was a comparable tuning budget available to the baselines whose numbers are copied?

Suggestions: 6. A cross-domain transfer experiment in which node features are actually time-varying (e.g. pretraining on some traffic/epidemic/web datasets and testing on others) would directly test the paper's main hypothesis.

4 comments
AuthorsOpenReview ↗

Response to W1 and W2.1

We thank you for your careful evaluation. We are encouraged by your recognition that the method is conceptually simple, modular, and cheap to train, that a BPTT-free pipeline that is competitive with specialized models is practically relevant, and that the zero-shot variant obtains the best average AUC and rank among the compared zero-shot models. We answer each point below.

In short: we include (i) new zero-shot traffic experiments where all priors are informative (W2.2), (ii) more baselines, (iii) an explicit discussion of the transfer claims, and (iv) a discussion where we better position our contribution. Overall, during the rebuttal we ran 6 new experiments and 3 new analyses, all listed in our general comment.

Regarding Weaknesses

W1.1 Thank you for this observation. Since the MiNT benchmark is featureless by construction (X=1X=1), MOSAIC processes the input without creating new node-level temporal features. Thus, Φf\Phi_f and Φs\Phi_s process a constant window of ones. We want to clarify that since each prior is independently fed as input, MOSAIC needs no architectural modifications or different hyperparameters in this scenario, which highlights the flexibility of our approach to work in different observation settings.

Regarding the transferability of TSFM-based temporal priors, we conducted a new experiment where we tested it, which is shown in our discussion to W2.2. In particular, we provided an experiment where we use MOSAIC for zero-shot traffic forecasting, thereby showing that the TSFM-based prior is a strong prior in a zero-shot setting.

W1.2 Thank you for pointing out this detail. Since with the three priors we encode a sliding window, each sample after the one-time extraction phase encodes recent temporal information about the signal and the graph. Thus, the temporal order is already encoded inside each sample and the trainable GNN keeps no temporal state across samples, so we can shuffle samples when training. We have clarified this in the revised version. Thank you.

W1.3 Thank you for this comment. Table 6 follows the protocol of MiNT [1], which compares "single models", i.e., "models that are trained on only a single network", with MiNT models pre-trained on multiple networks, showing that "MiNT models outperform single-network models". This experimental setting is common to other works [2,3,4], and is also used for pretrained GNNs [5]. In Table 10, we presented the per-network performance of supervised versus zero-shot models employed in the comparison, from which we extracted positive and negative transfer. Measuring transfer against the same model trained only on the target is the standard definition of negative transfer [4]. The absolute zero-shot AUC is also in Table 5, where MOSAIC-zs has the highest average AUC (0.744 vs. 0.727 for HTGN-zs and 0.662 for GC-LSTM-zs).

W1.4 We thank the reviewer for this insightful observation. We agree with you that our MiNT transfer evaluation is in-domain and therefore does not establish cross-domain generalization. Our intention in the first sentence of the abstract was to describe a limitation of existing temporal graph models, rather than to claim that our framework addresses cross-domain transfer. Nevertheless, we acknowledge that the wording could be interpreted otherwise. We have therefore clarified the scope of our claims in the revised abstract and introduction, distinguishing the supervised performance across multiple real-world benchmarks from zero-shot generalization to unseen networks in an in-domain setting. We also added an experiment in Q6 to further investigate this aspect. We sincerely appreciate the reviewer’s comment, which helped us clarify this important distinction.

W2.1 We thank the reviewer for this insightful observation and agree that MOSAIC builds upon existing methods, whose original contributions we duly acknowledge. However, combining established components into a novel architecture is itself a valuable contribution, as demonstrated by works such as Chronos [6], DCRNN [7], Graph WaveNet [8], and GraphGPS [9].

MOSAIC aims to show that carefully composing existing methods can yield an effective and efficient architecture. We demonstrate that its three priors are empirically complementary, as each of them is necessary: for every prior, there is a setting where removing it degrades performance (Table 4). Thus, MOSAIC achieves competitive or state-of-the-art performance in both supervised and zero-shot settings. Beyond the compositional framework itself, our contribution is a step towards foundational architectures for temporal graphs, whose heterogeneity in observation regimes, tasks, and distributions makes their design particularly challenging. We have clarified this aspect in the revised paper and sincerely thank you for raising it.

AuthorsOpenReview ↗

Response to W2.2, Q1, Q2, and Q3

W2.2 Thank you for these suggestions.

(a) We add a very recent state-of-the-art model, LAMP [10], to Table 1, which we report here for your convenience:

ModelWiki MathTennisChickenPox
Best baseline0.542$\pm$0.0080.300$\pm$0.0030.907$\pm$0.007
LAMP0.444$\pm$0.0030.304$\pm$0.0010.748$\pm$0.006
MOSAIC0.440$\pm$0.0020.262$\pm$0.0020.745$\pm$0.003

LAMP lowers the previous best MSE by about 18% on Wiki Math and ChickenPox, and MOSAIC still outperforms it on all three datasets.

(b) We added STID [11] to Table 2 of the revised paper. To do so, we copied their reference results for PeMS0{4,7,8} and used their official implementation to train it on PeMS03, which was missing in their experiments. Furthermore, following the reviewer suggestion, we show the TSFM zero-shot performance, without graph propagation. In particular, Chronos-Bolt small, with the same context of 864 steps, is the TSFM we use as Φf\Phi_f in MOSAIC zero-shot, so it is the graph-free counterpart of our model. We also include OpenCity [12], a recent spatiotemporal foundation model engineered for traffic forecasting. To make the comparison fair, we show MOSAIC performances (Masked MAE$\downarrow$), when pretrained in a leave-one-out fashion on PeMS-D{5,6,10,11,12} [13] and three among PeMS0{3,4,7,8}, and evaluated zero-shot on the left-out dataset. These results are shown in the table below (†^\dagger means that the pre-training corpus of the model contains the dataset corresponding to the column, or traffic data from the same district) and in Table 2 of the revised version.

ModelPeMS03PeMS04PeMS07PeMS08
Chronos-Bolt small [6]17.4923.5226.0219.12
Chronos-2 [14]15.12$^\dagger$20.50$^\dagger$22.22$^\dagger$16.27$^\dagger$
OpenCity-Plus [12]16.1619.86$^\dagger$22.44$^\dagger$15.05$^\dagger$
MOSAIC zero-shot14.8820.2621.9815.64

In a completely zero-shot setting, MOSAIC has the best performance among the compared models, whereas when other baselines have seen the target dataset during pre-training, we are competitive with or outperform them. A more complete discussion of these results is now included in Section 4.3, and the complete setting in Appendix B.7 of the revised version. We are grateful to the reviewer, because with this new experiment we can better understand and position the contribution of our trainable part of the framework, attributing its gains over a graph-free baseline even in the zero-shot settings. For further insights, we kindly redirect the reviewer to the ablation shown in the zero-shot column of Table 4 and discussed in Appendix B.7 (presented also in the discussion of Weakness 1 of reviewer yQ8C), where we isolate each prior in a completely zero-shot setting.

Regarding Questions

Q1 Thank you for this question. The frozen TSFM receives the constant series {1,…,1}\{1,\dots,1\} of length wf=8w_f=8 (Table 9) and returns a constant embedding. Φs\Phi_s receives the all-ones window.

Q2 Thank you for this detail. We noticed an oversight on Table 9, since for all traffic forecasting experiments, we fixed ws=12w_s = 12, following the standard setting of the benchmark, which baselines share. The hyperparameter search selected wf=288w_f=288 for PeMS0{3,7,8} and wf=12w_f=12 for PeMS04. We sincerely thank the reviewer for this important question, as we added an important detail for reproducibility.

Q3 Thank you for this question. We kindly redirect the reviewer to our response W2.2(b).

AuthorsOpenReview ↗

Response to Q4, Q5, Q6 and References

Q4 Thank you for raising this question, which motivated a deeper investigation of this phenomenon. We analyzed three representative cases: strong positive transfer (BAG), near-zero positive transfer (HOICHI), and strong negative transfer (SDEX). The full analysis is provided in Appendix B.1.2.

Our findings suggest that these failures are not specific to MOSAIC or necessarily attributable to training instability. First, similar effects occur across different architectures: pre-training improves the performance of all three models on BAG, while it also leads to negative transfer for HTGN on SDEX. Moreover, their supervised performance is positively correlated across networks, suggesting that part of the difficulty is shared across models and stems from the data.

We further find that this difficulty is related to changes in the relationship between network activity and future growth across training and test periods. Although this relationship is predictive across most networks, its strength can vary substantially over time. For instance, it becomes much stronger on BAG but nearly disappears on SDEX. This provides a possible explanation for why models trained on historical observations can struggle on certain test periods, while also highlighting that pre-training does not guarantee positive transfer on every network.

Overall, these findings suggest that the observed failures reflect, at least in part, network-specific temporal variability rather than a systematic instability of MOSAIC. They also help contextualize the transfer gains reported in Table 6, which are aggregate empirical improvements, not a guarantee of positive transfer on every network of that domain.

Q5 We thank the reviewer for pointing out this detail. As stated in Section 4.1, we follow each compared work’s protocol: forecasting selects the best validation score on the temporal split over the Table 9 grid, the experiments of Table 3 comprise no hyperparameter search, and on MiNT we perform the search on the validation splits of the 64 pre-training networks. Thus, as reported in Appendix A.4, we explore 8, 16, and 4 hyperparameter combinations for the three main validated experiments, respectively. The results of the baselines are copied from their original works, where each model was tuned by its authors with a budget that, when reported, is comparable with ours (e.g., PDFormer [15] selects on validation among 16 configurations).

Q6 Thank you for this very important suggestion, which we appreciate a lot. We conducted a cross-domain experiment, where we pre-train on traffic forecasting graphs and evaluate via linear probing on chickenpox (epidemics) and wikipedia math (web activity). In the table below we present the results, which we added in Appendix B.10 of the revised version:

ModelChickenPoxWikipedia Math
LAMP (best baseline)0.748$\pm$0.0060.444$\pm$0.003
MOSAIC-transfer0.812$\pm$0.0090.479$\pm$0.005
MOSAIC0.745$\pm$0.0030.440$\pm$0.002

MOSAIC-transfer outperforms every baseline of Table 1 except LAMP, and is close to its fully supervised counterpart. These results show the potential of further extending this work to explore out-of-distribution and out-of-domain adaptation methods given the modularity of our framework. Thank you again.

We thank you once again, and hope that our clarifications and additional experiments will lead to a reconsideration of your score.


References:

[1] Shamsi et al. MiNT: Multi-network transfer benchmark for temporal graph learning. NeurIPS Datasets and Benchmarks Track, 2025.

[2] Rosenstein et al. To transfer or not to transfer. NeurIPS Workshop on Inductive Transfer, 2005.

[3] Pan and Yang. A survey on transfer learning. IEEE TKDE, 2010.

[4] Wang et al. Characterizing and avoiding negative transfer. CVPR, 2019.

[5] Hu et al. Strategies for pre-training graph neural networks. ICLR, 2020.

[6] Ansari et al. Chronos: Learning the language of time series. TMLR, 2024.

[7] Li et al. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. ICLR, 2018.

[8] Wu et al. Graph WaveNet for deep spatial-temporal graph modeling. IJCAI, 2019.

[9] Rampášek et al. Recipe for a general, powerful, scalable graph transformer. NeurIPS, 2022.

[10] De Castelli et al. Adaptive memory retention in dynamic graphs. ICML, 2026.

[11] Shao et al. Spatial-temporal identity: A simple yet effective baseline for multivariate time series forecasting. CIKM, 2022.

[12] Li et al. Open spatio-temporal foundation models for traffic prediction. ACM TIST, 2025.

[13] Chen et al. Freeway performance measurement system: Mining loop detector data. Transportation Research Record, 2001.

[14] Ansari et al. Chronos-2: From univariate to universal forecasting. arXiv, 2025.

[15] Jiang et al. PDFormer: Propagation delay-aware dynamic long-range transformer for traffic flow prediction. AAAI, 2023.

AuthorsOpenReview ↗

Did our responses address your concerns?

Dear Reviewer nVGF,

Thank you again for your careful review and valuable suggestions.

We hope that our responses have helped clarify the contribution of MOSAIC, the transferability of the TSFM-based temporal prior, the comparison with recent and graph-free baselines, and the scope of our transfer claims, including a first cross-domain experiment.

We would appreciate knowing whether our responses have addressed your main concerns, and we welcome any further questions or suggestions.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision and of the new experiments

We thank all the reviewers for their careful evaluations and for their suggestions, which helped us improve and clarify very relevant and important details of our paper. We uploaded a revised PDF in which all changes are highlighted in blue. During the rebuttal, we ran 6 new experiments and 3 new analyses: they add 5 baselines, 2 new tables, 2 new figures and 5 new appendix sections, and extend Tables 2, 4 and 16. Below, each change is listed with its location and the reviewers who asked for it.

Experiments

  • LAMP is added to the comparison (Table 1; asked by nVGF, HSLU)
  • STID and zero-shot traffic forecasting against Chronos-Bolt small, Chronos-2 and OpenCity-Plus (Table 2, Sec. 4.3, App. B.7; asked by nVGF, yQ8C)
  • Ablation of the priors in zero-shot on PeMS07 (Table 4, Sec. 4.2; asked by nVGF, HSLU, yQ8C)
  • Prior ablation on MiNT zero-shot (Sec. 4.3, App. B.8; asked by yQ8C)
  • Long-context temporal prior (App. B.9, Table 17; asked by yQ8C)
  • Cross-domain transfer by linear probing (App. B.10, Table 18; asked by nVGF)

Analyses

  • Dataset-dependent transfer on MiNT: BAG, HOICHI, SDEX (App. B.1.2, Figure 3; asked by nVGF, yQ8C)
  • Embeddings of Φf\Phi_f at a short and a long context (App. B.9, Figure 4; asked by yQ8C)
  • Training memory with one prior per sample (App. B.6.2, Table 16; asked by HSLU)

Claims and clarifications

  • The first sentence of the abstract refers to existing models (Abstract; asked by nVGF)
  • The original works behind each ingredient are credited, and the zero-shot claims are scoped to unseen networks of the same domain (Sec. 1; asked by nVGF, HSLU)
  • The memory claim states its dependence on wsw_s (Sec. 3.3; asked by HSLU)
  • The supervised (transductive) and zero-shot (inductive) settings are defined (Sec. 4; asked by nVGF, yQ8C)
  • The selected hyperparameters and wsw_s are stated (Table 9, App. A.4; asked by nVGF)
  • The sentence with a typo is removed and a citation format is corrected (Sec. 4.1, Sec. 4.3; asked by HSLU)

Discussions

  • DyAug and PromptDyG are discussed in related work (Sec. 2; asked by HSLU)
  • Out-of-domain and out-of-distribution methods are added to future work (Sec. 5; asked by nVGF)

We look forward to your feedback and hope that our responses help clarify our contributions and that you will consider revising your score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.