acceptodds
Under review as a conference paper at ICLR 2027

Continuous-Time Dynamic Graph Condensation for Temporal Edge Classification

Abstract

Graph condensation has emerged as a promising technique for distilling large-scale graphs into compact yet informative ones. However, existing approaches primarily focus on static graphs and largely ignore continuous-time dynamic graphs (CTDGs), where graph evolution is represented as asynchronous event streams. Condensing CTDGs is particularly challenging because both structural evolution and temporal dynamics must be preserved under stringent compression constraints. To address this challenge, we propose Continuous-Time Dynamic Graph Condensation (CTDGC), the first graph condensation framework for event-stream continuous-time dynamic graphs and temporal edge classification. CTDGC adopts a learnable selection-based condensation paradigm that assigns each event an importance score and identifies informative events for condensation. To preserve knowledge from the original event stream, CTDGC first trains a lightweight relay model and then optimizes the condensed graph through feature-level and task-level preservation objectives, retaining both temporal representations and task-relevant semantic information. Furthermore, CTDGC employs class-wise top- event selection to construct condensed graphs. Since the top- operation is non-differentiable, a Straight-Through Estimator (STE) is introduced to enable end-to-end optimization. This design naturally satisfies arbitrary compression budgets without requiring additional sparsity-inducing constraints. In extensive experiments on multiple benchmarks, CTDGC consistently outperforms compared baselines under various compression ratios and generalizes well across diverse temporal graph learning architectures. Moreover, on several datasets, models trained on CTDGC-condensed graphs achieve performance competitive with those trained on the original full event streams while using only a small fraction of the training events. Our code will be shared publicly upon paper acceptance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph representation learning, data distillation, graph condensation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Lwn3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper explores the task of continuous-time dynamic graph condensation tailored for temporal edge classification. Instead of generating synthetic events, the approach relies on learning importance scores for original training events, picking a compact subset through a class-wise top-k mechanism and a straight-through estimator. To steer this selection, a frozen temporal relay model is employed to optimize two main goals: aligning the class-wise representation statistics between the original and condensed histories, and maintaining label prediction accuracy. Across five datasets and various downstream models, experiments show that the selected subsets generally surpass baseline event-selection methods, particularly at lower retention ratios.

Critical strengths

This paper tackles a meaningful problem with unique temporal challenges. Condensing an event stream is inherently trickier than filtering independent training examples because dropping a single event alters the historical context of everything that follows. Exploring reusable, compact training streams for temporal edge classification is therefore a worthwhile pursuit that nicely pushes graph condensation beyond static and snapshot-based setups.

The approach is conceptually coherent and backed by encouraging empirical results. Aligning the representations of selected events across original and condensed histories is an intuitive way to handle these temporal dependencies. The experiments span five datasets, multiple retention ratios, and several downstream architectures, offering solid evidence that the method produces genuinely useful training subsets. The cross-architecture evaluation is a particular highlight, as it alleviates concerns that the selected events might only be overfitted to the relay model.

Critical weaknesses

The current analysis does not fully isolate the specific contribution of temporal-context preservation. The proposed losses might simply favor events that are easy for the frozen relay model to classify or whose representations remain stable when history is dropped, rather than capturing broader temporal dependencies.

While the ablations confirm that both objectives help, they do not distinguish your temporal mechanism from standard supervised subset selection using the same relay model.

Crucial details regarding the optimization and evaluation setup also remain unclear. Although the straight-through estimator mask is defined, the manuscript does not explain how it integrates into temporal neighborhood construction and aggregation, or how unselected events receive gradients to re-enter the subset.

It is ambiguous whether downstream evaluation can access discarded training events as historical context.

Questions for authors

Could you walk us through how the straight-through estimator mask is actually integrated into temporal neighborhood sampling and aggregation?

What kind of historical information is accessible during downstream training, validation, and testing?

How can we be sure how much performance gain stems directly from accounting for shifts in temporal context?

What does the overall preprocessing overhead look like?

Could you clarify the discrepancy in the Enron results at the 2.5% retention ratio? Table 2 lists 59.18±0.7359.18 \pm 0.73, while the complete model in Table 4 reports 61.16±1.3361.16 \pm 1.33. Were these generated using different experimental settings?

Reviewer_etrGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CTDGC, a selection-based graph-condensation method for continuous-time dynamic graphs in temporal edge classification. It trains a frozen lightweight relay model on the original event stream, learns an importance score per training event, selects class-wise top-k events using a straight-through estimator, and optimizes feature-statistic and classification losses computed on the selected stream.

Critical strengths

The paper addresses a practically relevant problem and extends research on graph compression. CTDGC tackles this problem by contrasting the representations of selected events in their original and compressed temporal contexts, rather than treating events as independent samples. Preserving the original timestamps and event attributes also makes the resulting subset compatible with different temporal graph models. The experimental results are promising: retaining only 2.5% of the training events yields a weighted F1 score of 59.18, compared with 60.12 when training on the full dataset.

Critical weaknesses

Section 3.4 defines the straight-through mask as m = (m−w)—_detach +w, but Equations (1)–(3) do not show how this mask enters the computation. If the implementation constructs the reduced event list and temporal neighborhoods through discrete indexing, defining the mask alone does not make these operations differentiable. Conversely, if the mask weights event features or aggregation coefficients, its placement and normalization jointly determine both the forward-pass semantics and the surrogate gradients.

Tables 2 and 4 report different results for the full CTDGC model on Enron at the same 2.5% compression ratio: 59.18 ± 0.73 in Table 2 versus 61.16 ± 1.33 in Table 4. If the experimental settings are identical, these results should be consistent. This unexplained discrepancy complicates the interpretation of the ablation results and makes it difficult to assess the reproducibility of the reported performance.

Questions for authors

How is the STE mask applied to temporal aggregation, and do unselected events receive gradients?

Why do Tables 2 and 4 report different results on Enron at the same 2.5% compression ratio?

Are all baselines subject to the same per-class event budget? Under this budget, how does CTDGC compare with random sampling that matches both class proportions and temporal bins?

Compared with training on the full dataset, what are the relay training time, condensation time, downstream training time, peak memory usage, and retained storage size? How many downstream training runs are needed for condensation to become computationally cost-effective?

Reviewer_XasgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CTDGC for condensing continuous-time dynamic graphs through learnable event selection. A frozen relay model guides class-wise top-k selection using feature-statistics matching and a supervised task loss, while a straight-through estimator is used to optimize event scores. Experiments evaluate the selected event streams on five datasets and across seven downstream architectures.

Critical strengths
  1. The paper studies a useful problem: reducing continuous-time event streams for temporal edge classification while retaining the original timestamps and interactions. This setting complements existing work on static and snapshot-based graph condensation.

  2. The empirical results are promising. CTDGC achieves the highest mean weighted F1 in 12 of the 15 settings in Table 2 and the highest average across downstream models on all four datasets in Table 3. The loss ablations also provide initial evidence that the two objectives are complementary.

Critical weaknesses
  1. The central optimization procedure is insufficiently specified. Section 3.4 introduces an STE mask, but Equations (1)-(3) do not show how this mask enters temporal aggregation, feature statistics, or the task loss. If events are removed through hard indexing before these computations, the STE expression alone does not establish a gradient path to the scores. The authors should provide the masked computation and pseudocode, explaining how unselected events receive gradients and can re-enter the selected set. This is essential for assessing and reproducing the method.

  2. The preservation objectives do not fully justify the claimed retention of the original event stream. In Equation (2), both sets of statistics are computed over the same selected events, under different graph contexts; the reference statistics are not computed over all original events in each class. A small loss therefore does not establish that the selected events represent the full class distribution. Equation (3) additionally encourages correct classification of the selected events, but does not establish their usefulness for training a new model. Comparisons with full-class distribution matching and a simple relay-confidence selector would help clarify what these objectives contribute.

  3. The claimed computational and storage benefits are not quantitatively validated. The experiments report predictive performance and retained event ratios, but no condensation time, downstream training time, peak memory, or actual storage. CTDGC first trains a relay model on the full stream and then optimizes event scores, so its preprocessing cost matters. End-to-end costs and the number of downstream training runs needed to recover this cost should be reported. Event-count reduction alone does not establish an equivalent reduction in total storage, especially if node features are retained.

  4. The paper has several weaknesses in the experimental validation and reporting:

(1) Table 2 reports 59.18 +/- 0.73 for CTDGC on Enron at 2.5% compression, whereas Table 4 reports 61.16 +/- 1.33 under the same stated setting. This discrepancy needs to be resolved before the ablation results can be interpreted reliably.

(2) The baseline adaptations are underspecified. Class-distribution preservation is stated for Random Sampling, but it is unclear whether Herding, K-Center, EL2N, and GraNd use the same class-wise budgets as CTDGC. Their feature extractors, temporal segmentation, and training budgets also need to be specified. Matched controls are needed to separate the contribution of score optimization from class allocation and relay features.

(3) Table 4 only removes the two losses. It does not isolate the temporal mechanism, the relay design, or the selection estimator. Comparisons without time decay and with a simple recency-based selector would help establish the benefit specific to continuous-time graphs. Table 3 also lacks full-stream results for each downstream architecture and variability estimates, limiting conclusions about performance retention across models.

  1. Important reproducibility and evaluation details are absent. The paper does not specify the values or selection procedures for the decay coefficient and loss weight, event-score initialization, optimization schedule, neighborhood size, feature construction, or the number of independent runs. It also does not state whether evaluation neighborhoods contain discarded training events. This distinction matters for interpreting the claim that models use only the condensed graph and for assessing the practical compression benefit.
Questions for authors
  1. Could the authors provide the explicit masked temporal aggregation and loss computations used for score optimization, including how gradients reach currently unselected events?

  2. Are discarded training events excluded from the historical neighborhoods used during validation and testing? Are previously observed validation or test interactions incorporated causally, and is the same protocol used for every method?

  3. How are classes with a zero or single-event selection budget handled when computing the means and standard deviations in Equation (2)? Since weighting each class by its original frequency gives larger classes more weight, what supports the claim that this weighting prevents minority classes from being overwhelmed? Please report retained class coverage and macro-F1 in addition to weighted F1.

  4. What explains the Enron discrepancy between Tables 2 and 4, and what are the measured end-to-end time and memory costs relative to training on the full stream?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.