acceptodds
Under review as a conference paper at ICLR 2027

TE-MDG: Transfer Entropy for Mutation-aware Dynamic Graph Representation Learning

Abstract

Dynamic graph representation learning aims to capture the temporal patterns of evolving graphs, which is critical for time-aware applications like social networks and recommendation systems. Most of existing methods typically model long-term dynamics by aggregating the nodes' historical behaviors. However, in real-world dynamic graphs, node features and their interactions often evolve with two types of mutations: Rational Mutations that are consistent with stable evolution patterns and Occasional Mutations that are incidental and irrelevant. Recent advances fail to distinguish different types of mutations due to the absence of a clear standard, causing the meaningful information in mutations to be weakened and hard to extract. To this end, we propose a **T**ransfer **E**ntropy for **M**utation-aware **D**ynamic **G**raph Representation Learning framework (**TE-MDG**), which selectively incorporates different types of mutations into node representations. Specifically, TE-MDG consists of two key components: (i) Evolution Pattern Decoupling, which learns a set of temporal prototypes that capture the stable evolution patterns and serve as a guiding standard for distinguishing rational mutations from occasional ones; and (ii) Mutation Filtering, which uses transfer entropy to measure and strengthen the information gain from temporal prototypes to mutations, thereby encouraging the model to retain rational mutations while filtering out occasional ones. Extensive experiments on multiple real-world benchmark datasets demonstrate that TE-MDG is more effective than state-of-the-art baselines. The code and datasets are anonymously available at https://anonymous.4open.science/r/TE_MDG.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Dynamic Graph, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_J24MOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies continuous-time dynamic graph representation learning, focusing on distinguishing rational mutations that follow stable evolution patterns from occasional mutations caused by incidental factors. This paper proposes TE-MDG, which learns global temporal prototypes to represent stable evolution and adds a transfer-entropy regularizer to increase the information these prototypes provide about future local representations. The key idea is to use the prototypes as a reference that retains predictive behavioral changes and downweights incidental ones. On TGB and DyGLib, the TE-MDG report improvements over several baselines on temporal link prediction and node-level ranking.

Critical strengths

The paper identifies a meaningful gap in dynamic graph representation learning: not all changes in historical interactions are equally informative for future behavior, and distinguishing persistent trends from incidental interactions is a useful direction. Its evolution-pattern decoupling module offers a straightforward way to inject global, reusable temporal patterns into local sequence representations, and the learned prototypes combine naturally with patch-level features inside existing temporal encoders. The evaluation covers temporal link prediction and node-level ranking on TGB and DyGLib, with consistent gains across the reported metrics on UCI, Enron, and tgbn-genre.

Critical weaknesses

Equation (2) makes the prototype representation a deterministic function of the current local state, so the conditional mutual information in Equation (6) is identically zero. This also contradicts Equation (8), which requires a rational mutation to remain dependent on the prototype after conditioning on that same state. The mini-batch shuffling strategy does not appear to sample from the conditional distribution required for the stated transfer entropy and may instead estimate ordinary mutual information. Consequently, the proposed objective does not theoretically justify the claimed distinction between rational and occasional mutations.

Mutation types are assigned by inspecting PCA plots, with no pre-specified rule or controlled injection of the two changes, so downstream accuracy does not show that rational mutations are preserved and occasional ones are filtered. The transfer-entropy term also contributes little: removing it changes average TGB AUC by only 0.45, and on tgbl-wiki the variant without it has higher AP than the full model. Confidence in the comparison is further limited by the gap from published TGB results. On tgbl-coin, published DyGFormer and GraphMixer reach about 75 MRR, while Table 1 reports 26.56 and 35.80. The unspecified subsampling rate and the w/o-S rows that reproduce DyGFormer, including the standard deviations, remain to be clarified.

Questions for authors

Equation (2) makes the prototype representation a deterministic function of the current local state. How can the conditional mutual information in Equation (6) be nonzero, and how can the conditional dependence asserted in Equation (8) hold?

What rule identifies rational and occasional mutations independently of the embedding geometry shown in Figure 1?

What evidence shows that the learned prototypes represent stable evolution patterns shared across nodes and time, rather than only a recombination of the current patch?

Why are the tgbl-coin test MRRs in Table 1, 26.56 for DyGFormer and 35.80 for GraphMixer, substantially below the published results of about 75 under the official TGB protocol?

Why do the w/o-S results on tgbl-wiki, tgbl-review, and UCI match the DyGFormer row in Table 1, including the standard deviations?

8 comments
AuthorsOpenReview ↗

Response to Reviewer J24M (1/6):

Dear J24M,

We sincerely appreciate the time and effort you have devoted to reviewing our work. We are particularly grateful for your insightful and constructive comments, which have helped us further refine and strengthen our study.


W1&Q1: Theoretical Derivation of Non-zero Conditional Mutual Information and Clarification on Mutation Dependence

We sincerely appreciate your insightful comment. We formalize the stochastic nature of the global prototypes and their relationship with local states, and then present a theoretical derivation and a concrete example to demonstrate the conditions under which the conditional mutual information is nonzero.

1. Theoretical Derivation of Non-Zero Conditional Mutual Information.

The global prototypes SS in Eq. (2) are random prototypes jointly optimized during training. They represent shared evolution knowledge learned from historical interactions across nodes, and inherently capture cross-node factors that extend beyond the current local state. Consequently, the prototype representation is fully expressed as p~m=f(hmctx,S)\tilde{p}_m = f(h_m^{\mathrm{ctx}}, S).

Across the joint probability distributions of different dynamic graph environments, varying historical interactions and learning processes naturally produce varying global prototypes. Thus, we model SS as a random variable. Because these shared prototypes contain information about future states extending beyond the current local state, p~m\tilde{p}_m provides additional predictive information given hmctxh_m^{\mathrm{ctx}}. This analysis follows Schreiber's definition of information transfer between stochastic processes [1] and the framework for multivariate information transfer proposed by Lizier et al. [2].

Specifically, let H=hmctxH = h_m^{\mathrm{ctx}}, Y=hm+1ctxY = h_{m+1}^{\mathrm{ctx}}, and P=p~mP = \tilde{p}_m. Since P=f(H,S)P = f(H,S), the chain rule of conditional mutual information gives:

= I(Y;S\mid H)-I(Y;S\mid H,P).$$ Therefore, the prototype representation retains a portion of the conditional predictive information about future states contained in the global prototypes. Formally, this is expressed as: $$I(Y;S\mid H)>I(Y;S\mid H,P),$$ which directly yields $I(Y;P\mid H)>0$. We further construct a concrete example to demonstrate this condition, consistent with the softmax-based prototype aggregation mechanism in Eq. (2). Let $B\sim\operatorname{Bernoulli}(1/2)$ and $S=(B,B+1)$, which gives $P=B+\sigma(H/\tau)$. Suppose the future state satisfies $Y=B\oplus O$, where $O\sim\operatorname{Bernoulli}(q)$ is independent noise with $0<q<1/2$. We then obtain: $$I(Y;P\mid H)=1-H_2(q)>0.$$ This example confirms that, **under the joint distribution with random prototypes , the prototype aggregation mechanism in Eq. (2) yields strictly positive conditional mutual information in Eq. (6)**. **2. Establishment of the Conditional Dependence Relationships in Eq. (8)** Eq. (8) characterizes $\delta_{\mathrm{rat}}$ and $\delta_{\mathrm{oca}}$ as two latent information components of local mutations. These components are defined by their distinct informational roles in predicting future states, extending beyond mere similarities to historical interaction patterns. Specifically, **Rational Mutations** refer to change components that are related to the evolution patterns encoded by the shared prototypes and provide predictive information about future states. In contrast, **Occasional Mutations** refer to perturbation components that do not contain such shared predictive information. Rational mutations do not need to follow existing local historical trends, as newly emerging behavior patterns may also carry useful information about future evolution. Therefore, Eq. (8) precisely formulates the distinct conditional dependence relationships between these two components and the prototype representation: $$\delta_{\mathrm{rat}} \not\perp \tilde{p}_m \mid h_m^{\mathrm{ctx}}, \qquad \delta_{\mathrm{oca}} \perp \tilde{p}_m \mid h_m^{\mathrm{ctx}}.$$ In the constructive example above, let $\delta_{\mathrm{rat}}=B$ and $\delta_{\mathrm{oca}}=O$. We can simultaneously obtain: $$I(\delta_{\mathrm{rat}};P\mid H)=1\ \mathrm{bit}>0, \qquad \delta_{\mathrm{oca}}\perp P\mid H.$$ This theoretically establishes that **the nonzero condition of transfer entropy and the conditional dependence relationships of the two mutation components can hold simultaneously under the same probabilistic model**. Thank you again for your valuable suggestion, which helps us improve the theoretical explanation of our method. **We have added the complete proof to Appendix B.1.** [1] Schreiber T. Measuring information transfer[J]. Physical review letters, 2000, 85(2): 461. [2] Lizier J, Rubinov M. Multivariate construction of effective computational networks from observational data[J]. 2012.
AuthorsOpenReview ↗

Response to Reviewer J24M (2/6):

W2&Q2: Rules for Distinguishing Rational and Occasional Mutations

Thank you for this constructive suggestion. While we initially visualize the difference between rational and occasional mutations in the representation space via the PCA in Fig. 1, following your suggestion, we further broaden our verification dimensions to provide more objective and direct empirical evidence. Specifically, we distinguish these two types of mutations directly from the perspectives of raw interaction behaviors and independent predictive utility.

First, we verify the objective existence of mutations in raw data. By partitioning historical interactions into consecutive windows and applying Jensen–Shannon Divergence (JSD) with permutation tests, we find that 44.95% of adjacent window pairs exhibit significant distributional shifts. This confirms that local mutations are real behavioral changes rather than artifacts of the learned embeddings.

Building on this, our core rule distinguishes mutations based on their future predictive utility. We analyze 2,828 significant mutations by constructing three predictors: a base predictor that relies solely on historical interactions, an enhanced predictor that incorporates the true mutation, and a Sham control predictor that adds an equal number of resampled historical interactions to control for sample size. Based on the predictive gain in subsequent probe windows and its Bootstrap confidence intervals, we classify the mutations into Predictive, Non-predictive, and Uncertain categories, and subsequently evaluate their future predictive utility in an independent verification window. Specifically, Mean Net Gain measures the average improvement in future prediction, while Positive-Gain Rate quantifies the proportion of mutations yielding positive predictive gains in the independent verification window.

Table 1. Predictive Utility of Significant Local Mutations on UCI

Mutation TypeNumber of MutationsMutation Proportion (%)Mutation Magnitude (Mean JSD, nats)Future Predictive Utility (Mean Net Gain, nats/interaction) ↑Positive-Gain Rate(%) ↑
Uncertain87931.10.4830.42837.7
Non-predictive37413.20.4780.24323.3
Predictive1,57555.70.4671.44470.5

As shown in table 1, the Predictive and Non-predictive groups have highly similar mutation magnitudes (Mean JSD: 0.467 vs. 0.478). However, their future predictive utilities are substantially different. The Predictive group achieves a mean net gain of 1.444 nats/interaction with a 70.5% positive-gain rate, whereas the Non-predictive group yields only 0.243 nats/interaction with a 23.3% positive-gain rate. The difference in net gain is 1.201 (95% CI: [1.053, 1.365]). These results provide direct empirical support for our definitions. Rational mutations correspond to the Predictive group, as they contain information that continuously improves future predictions despite potentially deviating from historical trends. Occasional mutations correspond to the Non-predictive group, which exhibit similar magnitudes but lack future predictive value. This proves that predictive utility serves as a rigorous, independent rule to distinguish the mutations, fundamentally justifying the necessity of selective mutation modeling in TE-MDG.

Thank you again for your time in understanding our paper and for the insightful questions. We have added the detailed experimental designs and results to Appendices C.4.1 and C.4.2, respectively.

AuthorsOpenReview ↗

Response to Reviewer J24M (3/6):

W2: Selective Mutation Modeling Capability of TE-MDG

We greatly appreciate your insightful comment. As we have demonstrated in Section 4.3, there is a significant positive correlation between the TE gain and the model's predictive capability. We also theoretically discuss the selective modeling capability of TE-MDG for the two types of mutations in Appendix A.3. To provide more direct evidence for the selective mutation modeling capability of TE-MDG, we conduct a Counterfactual Mutation Replacement experiment.

Specifically, we replace the true mutation (Original) with resampled historical interactions (Sham), while strictly controlling for the number of interactions, timestamps, and future validation samples. We then measure two metrics: Counterfactual Predictive Gain (measuring how much the model leverages the true mutation for prediction) and Counterfactual Representation Distance (measuring the model's representational sensitivity to the mutation). If TE-MDG selectively models mutations, retaining a rational mutation should yield a high predictive gain and distinct representation response, whereas replacing a non-predictive occasional mutation should have a negligible impact.

To further verify whether TE gain reflects the predictive utility differences of local mutations, we utilize the MINE statistics network to compute the information association difference between the true temporal transfer and the negative sampling control, which is then aggregated into per-mutation TE score. Subsequently, we compare the score distributions across different mutation types, employ AUROC to quantify the discriminative power between rational and occasional mutations, and compare these with JSD and prototype assignment weights to evaluate the representational capability of TE scores regarding mutation predictive utility. The results are summarized in Table 2.

Table 2. Selective Mutation Modeling and TE Score Analysis on UCI

Evaluation MetricPredictiveUncertainNon-predictive
Counterfactual Predictive Gain ↑0.1035(95% CI [0.0805, 0.1328])0.0314(95% CI [0.0103, 0.0617])0.0040(95% CI [-0.0197, 0.0332])
Counterfactual Representation Distance ↑0.18190.06720.0472
TE Score ↑0.34490.24230.2235
As shown in Table 2, the Predictive group achieves a Counterfactual Predictive Gain of 0.1035, compared to merely 0.0040 for the Non-predictive group. This indicates that removing rational mutations significantly impairs TE-MDG's predictive performance, while replacing occasional mutations causes no stable average performance drop. Furthermore, the Counterfactual Representation Distance for the Predictive group (0.1819) is substantially larger than that of the Non-predictive group (0.0472), demonstrating that the model is highly sensitive to changes in rational mutations. Additionally, the Counterfactual Predictive Gain strongly correlates with the independent predictive utility metric established in our previous analysis (Spearman ρ = 0.718).

Furthermore, the TE Score of the Predictive group (0.3449) is higher than that of the Non-predictive group (0.2235). When distinguishing between Predictive and Non-predictive mutations, the AUROC of the TE Score reaches 0.612 , surpassing the performance of both JSD and Prototype Weight, which stand at approximately 0.499. This indicates that relying solely on the magnitude of local mutations or prototype matching weights struggles to effectively distinguish mutations with different predictive values. In contrast, the TE Score captures additional information highly relevant to future predictive utility, demonstrating a superior discriminative capability. This further supports the critical role of TE gain in identifying rational mutations. Collectively, these results confirm that the internally learned TE scores successfully reflect the predictive utility differences of local mutations. (Detailed analysis is provided in Appendix C.5.2 and Figure 8).

Once again, we sincerely appreciate your insightful comments and constructive suggestions. We have added the detailed experimental designs, evaluation metrics, and result analyses to Appendix C.5.

Reviewer_P1P5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies node representation for continuous-time dynamic graphs (CTDGs). It argues that node behavior in real dynamic graphs exhibits two types of "mutations": "rational mutations" consistent with stable evolution, and occasional "accidental mutations"; existing methods lack a criterion to distinguish them and encode both indiscriminately. TE-MDG has two components: (i) evolution-pattern decoupling — learning K global temporal prototypes S to capture stable evolution patterns, serving as the "reference standard" for distinguishing mutations; (ii) mutation filtering — using Transfer Entropy to measure the information gain from prototypes to local patterns (estimated via MINE) and maximizing this gain as a regularization term, thereby retaining rational mutations and suppressing accidental ones. Experiments compare against TGN/TGAT/JODIE/CAWN/GraphMixer/DyGFormer/DIDA on 7 datasets (TGB: tgbl-wiki/review/coin + DyGLib: CanParl/Enron/UCI/MOOC), reporting link prediction (AUC/AP/MRR) and node classification (NDCG@10).

Critical strengths

Clear motivation: distinguishing rational vs. accidental mutations in dynamic graphs is a real problem, and the introduction and preliminary analysis (Fig.1 PCA) are intuitive. Sufficient experiments: 7 datasets including MRR, 5 seeds, ablations (w/o TE, w/o S), and TE training-dynamics visualization (Fig. 3).

Critical weaknesses

(1)Missing recent strong dynamic-graph baselines. The paper cites DGIB (Yuan 2024) but does not compare; (2)The mutation concept is unobservable and operationally fuzzy. Lacking quantitative validation that mutations truly exist and are confounded by existing methods.Can you provide evidence that local patterns with high TE gain indeed correspond to "rational mutations" and low-gain ones to "accidental mutations"?

Questions for authors

(1)Missing recent strong dynamic-graph baselines. The paper cites DGIB (Yuan 2024) but does not compare; (2)The mutation concept is unobservable and operationally fuzzy. Lacking quantitative validation that mutations truly exist and are confounded by existing methods.Can you provide evidence that local patterns with high TE gain indeed correspond to "rational mutations" and low-gain ones to "accidental mutations"?

4 comments
AuthorsOpenReview ↗

Response to Reviewer P1P5 (1/3):

Dear P1P5,

We sincerely appreciate the time and effort you have devoted to reviewing our manuscript. We are particularly grateful for your recognition of our work and your insightful and constructive suggestions, which have been invaluable in helping us further refine and strengthen our study.


W1&Q1: Comparison with Recent Dynamic Graph Baselines

We sincerely appreciate you for this valuable suggestion regarding the inclusion of recent competitive dynamic graph baselines. To address this concern, we have incorporated **four additional methods covering different approaches to dynamic graph representation learning.

  • DyG-Mamba[1] employs continuous state-space modeling with irregular time intervals to capture long-term temporal dependencies.
  • DGIB[2] applies the information bottleneck principle to learn robust and informative dynamic graph representations.
  • TAMI-GraphMixer[3] enhances GraphMixer with log-time encoding and link-history aggregation to address temporal interaction heterogeneity.
  • TPNet[4] introduces temporal walk matrices and random feature propagation to efficiently capture structural and temporal dependencies.

We evaluate these methods on four DyGLib datasets, including CanParl, Enron, MOOC, and UCI. The AUC and AP results are presented in Tables 1 and 2.

Table 1. AUC (%) Comparison with Recent Dynamic Graph Baselines

ModelCanParlEnronMOOCUCI
DyG-Mamba74.31±2.4789.18±2.9782.15±1.6994.04±0.09
DGIB58.10±0.1274.55±0.0757.05±4.3265.31±7.36
TAMI-GraphMixer78.59±0.9490.83±0.15OOM93.12±0.55
TPNet85.27±0.1992.54±0.4788.87±0.6792.94±0.65
TE-MDG76.37±5.1692.91±0.0784.89±0.6396.54±0.24

Table 2. AP (%) Comparison with Recent Dynamic Graph Baselines

ModelCanParlEnronMOOCUCI
DyG-Mamba69.00±1.3189.32±2.2883.14±1.1095.42±0.10
DGIB58.69±0.1574.05±0.0954.73±2.6169.14±7.11
TAMI-GraphMixer68.98±0.8089.09±0.32OOM94.72±0.52
TPNet79.07±0.2991.65±0.2887.36±0.7894.35±0.70
TE-MDG71.43±3.1192.42±0.0484.80±0.5497.65±0.12

Note: OOM denotes an out-of-memory failure.

The results demonstrate that TE-MDG remains competitive with these recent approaches. Specifically, TE-MDG achieves the highest AUC and AP on both Enron and UCI, outperforming all four additional baselines. Meanwhile, TPNet achieves the best performance on CanParl and MOOC, reflecting the varying strengths of different temporal modeling strategies across datasets. Overall, TE-MDG achieves the best performance on two of the four datasets under both metrics, supporting the effectiveness of our approach in learning informative dynamic graph representations.

Due to the limited time available during the rebuttal period, we have prioritized additional comparisons on the four DyGLib datasets. We plan to further extend the evaluation of these recent baselines to the TGB datasets in the subsequent revision, providing a more comprehensive assessment across different dynamic graph benchmarks.

We sincerely thank you again for your constructive suggestion, which has helped us strengthen the experimental evaluation and provide a more comprehensive comparison with recent advances in dynamic graph learning. We have incorporated these additional comparisons into Appendix C.7.

References

[1]Li, Dongyuan, et al. "Dyg-mamba: Continuous state space modeling on dynamic graphs." Advances in Neural Information Processing Systems 38 (2026): 129101-129130.

[2]Yuan, Haonan, et al. "Dynamic graph information bottleneck." Proceedings of the ACM Web Conference 2024. 2024.

[3]Yu, Zhongyi, et al. "TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction." Advances in Neural Information Processing Systems 38 (2026): 117623-117657.

[4]Lu, Xiaodong, et al. "Improving temporal link prediction via temporal walk matrix projection." Advances in Neural Information Processing Systems 37 (2024): 141153-141182.

AuthorsOpenReview ↗

Response to Reviewer P1P5 (2/3):

W2&Q2: Local Mutation Existence Detection

We sincerely appreciate for your important comment regarding the definition of mutations and their empirical evidence. In Section 2, we have visualized local dynamic graph representations using PCA and observed noticeable mutations in the local representations. Following your suggestion, we further conduct statistical significance tests on raw interaction distributions to directly validate the existence of local mutations. We also perform an additional analysis of mutation types based on their future predictive utility to characterize the different roles of local mutations.

1. Empirical Validation of the Existence of Local Mutations

To distinguish genuine behavioral changes from artifacts potentially introduced by representation learning or dimensionality reduction, we directly analyze raw dynamic graph interactions on the UCI dataset, without relying on representations learned by TE-MDG or other models. Specifically, we partition each node's historical interactions into consecutive windows and use Jensen–Shannon Divergence (JSD) [5] to quantify changes in interaction-partner distributions between adjacent windows. We then apply permutation tests with Benjamini–Hochberg false discovery rate (FDR) correction to identify local changes that significantly exceed random fluctuations.

Experimental results show that, among 7,409 pairs of adjacent interaction windows from 459 valid nodes in UCI, 3,330 pairs (44.95%) exhibit statistically significant distributional shifts, distributed across different temporal periods. For example, Node 1647 exhibits a JSD of 0.627 in its 17th window, substantially exceeding the permutation-based null mean of 0.218 (FDR-adjusted q=0.0109q=0.0109), with a new-neighbor interaction ratio of 100%.

These findings demonstrate that local mutations are observable behavioral changes in raw dynamic graph interactions that can be independently detected, quantified, and statistically validated. We have included the detailed experimental design and results in Appendices C.4.1 and C.4.2, with the corresponding visualization presented in Figure 5.

2. Mutation Classification Based on Future Predictive Utility

We further examine the future predictive utility of 2,828 significant mutations with sufficient subsequent interactions, using predictors independent of TE-MDG representations. Specifically, we compare a baseline predictor using only historical interactions, an augmented predictor incorporating the observed mutation, and a Sham control predictor incorporating an equal number of interactions resampled from historical records. Based on predictive gains and their bootstrap confidence intervals in a subsequent probing window, we classify mutations into three categories, Predictive, Non-predictive, and Uncertain, and evaluate their future predictive utility in an independent validation window.

Table 3. Predictive Utility of Significant Local Mutations on UCI

Mutation TypeNumber of MutationsMutation Proportion (%)Mutation Magnitude (Mean JSD, nats)Future Predictive Utility (Mean Net Gain, nats/interaction) ↑Positive-Gain Rate (%) ↑
Uncertain87931.10.4830.42837.7
Non-predictive37413.20.4780.24323.3
Predictive1,57555.70.4671.44470.5

As shown in Table 3, predictive and Non-predictive mutations have comparable mean JSD values (0.467 vs. 0.478), but markedly different future predictive utilities. Their validation gains are 1.444 and 0.243 nats/interaction, a difference of 1.201 (95% CI: [1.053, 1.365]). Their positive-gain rates are 70.5% and 23.3%, respectively.

These results demonstrate that mutation magnitude alone does not determine future predictive utility. Some local mutations contain informative signals that improve future interaction prediction, while others exhibit similarly significant distributional changes but contribute little predictive value. This provides empirical support for distinguishing predictive rational mutations from non-predictive occasional mutations and further highlights the importance of selectively modeling local mutations.

In summary, statistical significance tests on raw interaction distributions directly validate the existence of local mutations, while independent predictive utility analysis further demonstrates that these mutations differ substantially in their future predictive value.

Thank you again for your constructive suggestions, which have helped us refine the operational definition of mutations and extend our preliminary observations in representation space to statistical validation based on raw interaction data, thereby strengthening the empirical foundation of our research motivation. We have incorporated the detailed experimental design and results into Appendix C.4.

References

[5]Lin, J. (1991). Divergence measures based on the Shannon entropy. IEEE Transactions on Information Theory, 37(1), 145–151.

AuthorsOpenReview ↗

Response to Reviewer P1P5 (3/3):

W2&Q2: Evidence for the Association Between Transfer Entropy Gains and Mutations (W2,Q2)

We sincerely thank you for this important comment regarding the relationship between transfer entropy gains and mutation types. To address your concern, we conduct two additional experiments examining the association between TE Score and mutation predictive utility and the selective modeling behavior of TE-MDG. We first classify mutations as Predictive, Non-predictive, or Uncertain using a future predictive utility test independent of TE-MDG, ensuring that mutation labels are determined without relying on the model's TE scores.

1. Relationship Between TE Gains and Mutation Types

To directly address your main concern, we examine whether TE Score reflects the independently assessed predictive utility of local mutations. Specifically, we derive mutation-level TE scores from the MINE statistics network by comparing observed temporal transitions with negative-sampling controls. We then compare the scores across mutation types and evaluate their discriminative ability using AUROC, with mutation magnitude (JSD) and prototype assignment weight as baselines.

Table 4. TE Score and Selective Mutation Modeling Analysis on UCI

Mutation TypeTE Score ↑Counterfactual Predictive Gain ↑Representation Distance
Predictive0.34490.1035 (95% CI [0.0805, 0.1328])0.1819
Uncertain0.24230.0314 (95% CI [0.0103, 0.0617])0.0672
Non-predictive0.22350.0040 (95% CI [-0.0197, 0.0332])0.0472

As shown in Table 4, TE Score exhibits a positive association with the future predictive utility of local mutations. The Predictive group achieves a mean TE Score of 0.3449, compared with 0.2235 for the Non-predictive group, yielding a difference of 0.1213 (95% CI: [0.0005, 0.2248]).

Furthermore, TE Score achieves an AUROC of 0.612 (95% CI: [0.513, 0.707]) in distinguishing Predictive from Non-predictive mutations, outperforming JSD and Prototype Weight, both of which achieve approximately 0.499. These results indicate that local mutation magnitude and prototype assignment weight alone provide limited discrimination of future predictive utility, while TE Score captures additional information associated with future evolution and exhibits better discriminative ability.

Therefore, mutations with higher TE scores tend to exhibit greater future predictive utility, while those with lower scores are more likely to have limited predictive value. This statistical relationship is consistent with our operational distinction between rational and occasional mutations, providing empirical support for TE-guided mutation differentiation. Detailed results are provided in Appendix C.5.2 and Figure 8.

2. Selective Modeling of Different Mutation Types by TE-MDG

We further conduct a Counterfactual Mutation Replacement experiment to examine whether TE-MDG responds differently to mutations with distinct predictive utilities. Specifically, we replace observed mutations with historically resampled interactions while preserving interaction counts, timestamps, and future validation samples, and measure the resulting changes in prediction performance and node representations.

As shown in Table 4, the Predictive group yield substantially higher Counterfactual Predictive Gain than Non-predictive group (0.1035 vs. 0.0040), with a difference of 0.0994 (95% CI: [0.0646, 0.1371]). Their representation distances are 0.1819 and 0.0472, respectively. Moreover, counterfactual gains correlate strongly with independently assessed predictive utility (Spearman ρ = 0.718).

These findings demonstrate that TE-MDG exhibits stronger predictive and representational responses to mutations with greater future predictive utility, supporting its selective mutation modeling capability. Detailed results are provided in Appendix C.5.1 and Figure 7.

In summary, TE Score is positively associated with independently determined mutation predictive utility, while counterfactual experiments further confirm that TE-MDG exhibits selective responses to different mutation types. Together, these findings provide empirical support for TE-guided selective mutation modeling.

Thank you again for your constructive suggestions, which has helped us strengthen the empirical validation of our mutation distinction mechanism. We have incorporated the detailed experimental design and results into Appendix C.5.


We sincerely thank you again for your recognition of our work and your insightful and constructive suggestions. We have addressed all your concerns and incorporated the corresponding revisions into the revised manuscript, with the changes highlighted in blue for ease of reference. As some of your suggestions were also raised by other reviewers, the corresponding revisions are highlighted in yellow. Should you have any additional comments or suggestions, we would greatly appreciate your feedback.

Dear P1P5,

We thank you again for your valuable comments and suggestions, which have helped us strengthen our work. As the discussion period concludes in one day, we sincerely welcome your feedback regarding whether our rebuttal adequately addresses your concerns. Should any concerns remain, we are willing to provide further clarification.

Reviewer_24G6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of representation learning in temporal networks while incorporating the impact of rational mutations which are stable and somewhat predictable and occasional mutations, which are unpredictable. The authors propose an architecture that uses global prototypes to encode stable patterns of network evolution and a training scheme involving a transfer entropy based regularisation that potentially directs the embeddings towards a space that tries to preserve the predictable components of the network generating process and filter out occasional mutations which can be thought of as random noise. They demonstrate through comparisons on standard temporal graph representation learning benchmarks, that their approach beats a set of SOTA approaches in predicting unseen edges. In the appendix, they provide more details about the motivation behind their transfer entropy based regularisation and further experimental studies.

Critical strengths
  1. The paper studies an important and naturally motivated problem. They tries to give a solution that is intuitive.
  2. The experiments are extensive and demonstrates that their method is competitive with existing methods in this direction.
  3. The introduction is quite well written and the reader will at once get the sense of the problem under study.
Critical weaknesses
  1. The explanation behind the design of the transfer-entropy-based regularisation is somewhat confusing. It becomes more confusing in light of Appendix B.1 where a random SS is assumed, whereas in the algorithm a set s1,…,sKs_1,\ldots,s_K which are deterministic functions of the embeddings is used.

  2. Table 1 somewhat produces a misleading picture of the utility of the method in light of Table 10. To give a honest comparison, some comparators from Table 10 should be included in Table 1 with clear statement that the benefit of using this method is dependent on datasets.

  3. The authors do not provide a recipe by which an user can compute some descriptive statistic from the networks and decide which method to use. Table 10 clearly establishes that in all circumstances the method proposed in this paper is not the best. To sum up, there is less clarity on limitations and workarounds.

  4. The absence of a simulation that clearly shows the filtering of occasional mutations which baseline methods fail to achieve. The authors make some attempt to address this in Appendix C.4 and C.5 but a model based simulation will be much more enlightening. Also, it is not clear to me that how much gain comes from using the prototypes and how much comes from the TE regularisation.

Questions for authors
  1. I am somewhat confused by one issue. My understanding is that the objective of the transfer-entropy-based regularisation is to use the prototypes to inject additional predictive information into the embeddings. Since these prototypes are supposed to capture stable patterns, the authors believe that this will eventually filter out the impact of occasional mutations. However, this filtering property is neither established nor clearly explained, which somewhat obscures the main objective of the paper. I think this should go to the main text itself rather rather lots of ablation studies.

  2. Section 3.1 describes the prototypes s1,…,sKs_1,\ldots,s_K as shared learnable parameters across nodes and time. For a fixed learned dictionary, however, the prototype representation in Equation (2) is a deterministic function of the current local representation, making the conditional mutual information in Equation (6) identically zero. Appendix B.1 instead treats the dictionary as random across graph environments. Could the authors clarify how this stochastic interpretation corresponds to the actual training procedure and how the proposed estimator captures this randomness when the dictionary is shared across the training examples?

  3. I have difficulty following the derivation in Appendix B.3. In particular, Equation (43) does not appear to follow from the stated conditional independence assumptions. The correct chain rule is

I\big((\delta_{\mathrm{rat}}^{(m)},\delta_{\mathrm{oca}}^{(m)});\widetilde p_m\mid h_m^{\mathrm{ctx}}\big)=I\big(\delta_{\mathrm{rat}}^{(m)};\widetilde p_m\mid h_m^{\mathrm{ctx}}\big)+I\big(\delta_{\mathrm{oca}}^{(m)};\widetilde p_m\mid h_m^{\mathrm{ctx}},\delta_{\mathrm{rat}}^{(m)}\big). \end{aligned}$$ The assumptions $\delta_{\mathrm{rat}}^{(m)}\perp\delta_{\mathrm{oca}}^{(m)}\mid h_m^{\mathrm{ctx}}$ and $\delta_{\mathrm{oca}}^{(m)}\perp\widetilde p_m\mid h_m^{\mathrm{ctx}}$ do not imply that the final term is zero. For a simple counterexample, take $h_m^{\mathrm{ctx}}=0$, let $\delta_{\mathrm{rat}}^{(m)}$ and $\delta_{\mathrm{oca}}^{(m)}$ be independent Bernoulli variables with parameters $1/2$ and $1/4$, respectively, and set $\widetilde p_m=\delta_{\mathrm{rat}}^{(m)}\oplus\delta_{\mathrm{oca}}^{(m)}$. All the stated independence assumptions hold, but the left-hand side of Equation (43) equals one bit, whereas its claimed right-hand side equals $1-H_2(1/4)\approx0.189$ bits. Moreover, Equation (44) replaces an upper bound with an approximate equality without controlling the approximation error. Could the authors clarify these steps and specify the additional assumptions needed to justify the derivation? This is not a fatal thing, mostly a mathematical nuance, but clarification will be beneficial. 4. I think the ablation study in Table 2 deserves a more careful discussion in relation to the SOTA comparisons in Table 1. In particular, it would be useful to compare the model without transfer-entropy regularization (w/o TE) directly against the SOTA methods. This would help us understand how much of the performance improvement comes from the proposed regularization and how much comes from the underlying encoder architecture and the learned prototypes. Just put a table and discuss. The results are already there. 5. The other thing that struck me is the stark contrast between how the method in the paper compares with other methods in Table 1 and Table 10. In Table 1, all numbers look very rosy compared to the baselines, but in Table 10 we get a mixed picture which is perfectly reasonable and typically expected. Therefore, if a reader just goes by Table 1 which is the main paper, I think they can be mislead. I think the authors should do something about this as well. 6. Is the architecture used in Section 3.1 standard in this line of literature? Can there be a bit more description of the history behind the particular design adopted? Also, while citing papers for dynamic networks in the first line, I would not cite only papers that considers representation learning. Finally, the abstract can be written better with more insights and less jargon. In light of Table10, I would not make this statement in the conclusions: "Extensive experiments on multiple benchmarks demonstrate that TE-MDG consistently outperforms state-of-the-art baselines on link prediction and node classification."
9 comments
AuthorsOpenReview ↗

Response to Reviewer 24G6 (1/6):

Dear 24G6,

We sincerely appreciate your time and effort in reviewing our manuscript and your positive recognition of our work. Your insightful comments and constructive suggestions have been extremely valuable in helping us strengthen the theoretical foundations and improve the experimental validation.


Q1: Theoretical and Empirical Justification of Prototype-guided Mutation Filtering

Thank you for this insightful comment.

1. Theoretical justification. As detailed in our response to Q3 and Appendix B.3, under the stated conditional independence assumptions, transfer entropy (TE) characterizes the prototype-related information in rational mutations that is preserved in subsequent states, whereas occasional mutations do not directly contribute additional prototype-related information. Therefore, maximizing TE encourages the model to retain predictive information associated with stable evolution patterns, providing a theoretical basis for selective mutation modeling.

2. Empirical verification. We provide two complementary analyses to support this mechanism:

  • Global prototype analysis (Appendix C.6): We verify that the learned prototypes represent stable evolutionary patterns that can be reused across nodes and time, providing a meaningful reference for interpreting local mutations.
  • Selective mutation modeling (Appendix C.5): Through counterfactual mutation replacement, we show that TE-MDG is more responsive to mutations with predictive utility and less sensitive to non-predictive perturbations.

These findings are further supported by the ablation results in Table 2, which demonstrate the contributions of global prototypes and TE regularization to predictive performance. Together, these analyses explain why global prototypes provide a meaningful reference for distinguishing mutations, how TE encourages the preservation of predictive information, and how this selectivity manifests in practical representation learning.

Thank you again for your constructive suggestion, which has helped us clarify this central mechanism of TE-MDG. Following your suggestion, we have added a concise explanation of the mutation filtering mechanism to Section 3.2 of the revised manuscript, with explicit references to Appendices B.3, C.5, and C.6 for detailed theoretical derivations and empirical evidence.


W1&Q2: Clarification on the Stochasticity of Prototype S

Thank you for this insightful question. We clarify the relationship between prototype sharing, stochasticity, and practical training as follows.

1. Parameter sharing and stochasticity are compatible. The global prototype dictionary SS is learned from historical interactions across nodes and time, rather than being a predefined constant. Formally,

S=A(D,ξ),S=\mathcal{A}(\mathcal{D},\xi),

where D\mathcal{D} denotes the training graph, ξ\xi represents training randomness, and A\mathcal{A} is the learning algorithm. While all samples share the same dictionary at a given iteration, different training data and stochastic optimization processes induce different realizations of SS.

2. The theoretical formulations are consistent. Since p~m=f(hmctx,S)\tilde{p}_m=f(h_m^{\mathrm{ctx}},S), we acknowledge that conditioning additionally on SS yields zero conditional mutual information. However, Eq. (6) does not condition on a particular realization of SS. As derived in Appendix B.1,

I(hm+1ctx;p~m∣hmctx)=I(hm+1ctx;S∣hmctx)−I(hm+1ctx;S∣hmctx,p~m).I(h_{m+1}^{\mathrm{ctx}};\tilde{p}_m\mid h_m^{\mathrm{ctx}})=I(h_{m+1}^{\mathrm{ctx}};S\mid h_m^{\mathrm{ctx}})-I(h_{m+1}^{\mathrm{ctx}};S\mid h_m^{\mathrm{ctx}},\tilde{p}_m).

This quantity can be strictly positive when the learned prototypes preserve additional predictive information, as demonstrated by the constructive example in Appendix B.1.

3. Connection to practical training. During mini-batch training, the shared prototype dictionary SS is continuously updated through stochastic optimization, yielding successive realizations that incorporate evolutionary information from different nodes and temporal contexts. The corresponding prototype representations and subsequent local states are used to jointly optimize the MINE statistics network and the global prototypes. In this way, the stochasticity of the learned dictionary is reflected in the evolving prototype representations throughout training. The theoretical analysis in Appendix B.1 explains how these data-dependent prototypes can provide additional predictive information beyond the current local context, while the joint optimization encourages the model to capture and strengthen such predictive dependencies.

We sincerely thank you again for this valuable suggestion. We have added the corresponding clarification to Appendix B.1.

AuthorsOpenReview ↗

Response to Reviewer 24G6 (2/6):

Q3: Clarification of Eq.(43) and Eq.(44)

Thank you for identifying this important mathematical nuance and providing constructive suggestions. We clarify the conditional independence assumption required for Eq. (43) and provide an exact decomposition for Eq. (44).

1. Conditional Independence Assumptions in Eq. (43).

In our theoretical framework, occasional mutations arise from incidental factors unrelated to rational mutations and global evolution patterns. Therefore, given the current local state, we assume that occasional mutations are jointly independent of rational mutations and global prototype representations:

\perp \left( \delta_{\mathrm{rat}}^{(m)},\tilde{p}_m \right) \mid h_m^{\mathrm{ctx}}.$$ This assumption directly implies $$I(\delta_{\mathrm{oca}}^{(m)}; \tilde{p}_m \mid h_m^{\mathrm{ctx}}, \delta_{\mathrm{rat}}^{(m)})=0.$$ thereby establishing Eq. (43) through the chain rule of conditional mutual information. **2. Approximation in Eq. (44) and Its Error Analysis.** The global prototype provides additional predictive information about the next local state beyond the current state by capturing evolutionary patterns associated with rational mutations. In Eq. (40), we assume that once the current state and both types of mutations are given, the prototype provides no further predictive information about the next state. This yields the following conditional independence relationship $$h_{m+1}^{\mathrm{ctx}} \perp \tilde{p}_m \mid h_m^{\mathrm{ctx}}, \delta_{\mathrm{rat}}^{(m)}, \delta_{\mathrm{oca}}^{(m)}.$$ Combining this relationship with the joint conditional independence assumption above and marginalizing out occasional mutations yields $$h_{m+1}^{\mathrm{ctx}} \perp \tilde{p}_m \mid h_m^{\mathrm{ctx}}, \delta_{\mathrm{rat}}^{(m)}.$$ Accordingly, by the chain rule of conditional mutual information, we replace the original approximation in Eq. (44) with the following exact decomposition: $$I(h_{m+1}^{\mathrm{ctx}}; \tilde{p}_m \mid h_m^{\mathrm{ctx}})=I(\delta_{\mathrm{rat}}^{(m)}; \tilde{p}_m \mid h_m^{\mathrm{ctx}})-\varepsilon_m,$$ where $$\varepsilon_m=I(\delta_{\mathrm{rat}}^{(m)}; \tilde{p}_m \mid h_m^{\mathrm{ctx}}, h_{m+1}^{\mathrm{ctx}})\geq 0.$$ The residual $\varepsilon_m$ quantifies prototype-related information in rational mutations that is not retained in the subsequent state. If $\varepsilon_m\leq\epsilon$, the approximation error is bounded by $\epsilon$, and equality holds when $\varepsilon_m=0$. We sincerely thank you again for your careful analysis and constructive suggestions. **We have incorporated these theoretical clarifications into Appendix B.3.**
AuthorsOpenReview ↗

Response to Reviewer 24G6 (3/6):

Q4: Comparison with SOTA without TE regularization

Thank you for this valuable suggestion. To better disentangle the contributions of the underlying encoder, global prototypes, and TE regularization, we directly compare the w/o TE variant with the strongest SOTA baselines in Table 1. Best SOTA denotes the strongest baseline for each dataset and metric in Table 1 of the manuscript, while Δ TE denotes the performance improvement obtained by incorporating TE regularization.

Table 1: Comparison with SOTA without TE regularization (AUC / AP, %).

DatasetBest SOTAw/o TETE-MDGΔ TE
tgbl-wiki98.75 / 98.9797.17 / 97.5597.15 / 97.64−0.02 / +0.09
tgbl-review82.31 / 90.3882.24 / 90.8583.55 / 93.10+1.31 / +2.25
tgbl-coin96.08 / 97.2896.88 / 98.6896.94 / 98.70+0.06 / +0.02
CanParl85.27 / 79.0766.73 / 62.5176.37 / 71.43+9.64 / +8.92
Enron92.67 / 92.0489.34 / 89.2792.91 / 92.42+3.57 / +3.15
MOOC88.87 / 87.3681.74 / 82.3584.89 / 84.80+3.15 / +2.45
UCI94.94 / 96.0794.03 / 95.5496.54 / 97.65+2.51 / +2.11

Table 2: Average performance and additional gains from TE regularization.

Dataset GroupMetricBest SOTAw/o TETE-MDGΔ TE
TGBAUC92.3892.1092.55+0.45
AP95.5495.6996.48+0.79
DyGLibAUC90.4482.9687.68+4.72
AP88.6482.4286.58+4.16
1. Effectiveness of the encoder and global prototypes. On the TGB datasets, w/o TE achieves competitive performance against the strongest baselines. Notably, on tgbl-coin, w/o TE outperforms Best SOTA by 0.80 and 1.40 in AUC and AP, respectively. On tgbl-review, it also achieves competitive AUC and higher AP. These results demonstrate that the encoder combined with global prototypes can already provide effective dynamic graph representations, although its benefits vary across datasets.

2. Additional contribution of TE regularization. On the DyGLib datasets, w/o TE underperforms the strongest baselines, while incorporating TE substantially improves its performance. In particular, TE enables the full model to surpass the strongest baselines on Enron and UCI. On average, TE improves AUC and AP by 4.72 and 4.16 percentage points on DyGLib, respectively, compared with 0.45 and 0.79 percentage points on TGB. These results indicate that TE regularization plays an important role in enhancing the model's utilization of global evolutionary patterns.

Overall, these comparisons reveal the complementary roles of the two components: global prototypes provide a representation space for stable evolutionary patterns, while TE regularization further encourages the preservation of predictive evolutionary information. Although their relative contributions vary across datasets, their integration improves the effectiveness of dynamic graph representation learning. Their applicability and limitations are further discussed in Appendix C.8.

Thank you again for your constructive suggestion.

Comments on the paper

2 comments

Dear reviewers,

Thank you again for your time and thoughtful review. As the discussion period ends in two days, we would sincerely appreciate your feedback on whether our rebuttal has addressed your concerns. Should any concerns remain, we are happy to provide further clarification.

AuthorsOpenReview ↗

General Response to Area Chairs

Dear Area Chair,

We sincerely thank you for your time and efforts in overseeing the review process. To facilitate the next stage of evaluation, we have prepared this summary to outline our work and to provide an overarching response to the series of questions raised by the reviewers.


Summary of Our Work

a. Broad Impact and Relevance

  • We propose TE-MDG, a novel mutation-aware dynamic graph representation learning framework that distinguishes Rational Mutations, which contain useful predictive information, from Occasional Mutations, which arise from incidental disturbances.
  • TE-MDG learns global shared evolution patterns to guide the modeling of local mutations, enabling the selective preservation of predictive information. This provides a new perspective on long-term dynamic graph modeling and supports downstream tasks such as link prediction and node classification.

b. Technical Merit

  • Theoretical soundness: We provide rigorous theoretical derivations and mathematical proofs to support the effectiveness of global prototypes and TE-based mutation filtering, establishing a solid theoretical foundation for TE-MDG.
  • Extensive empirical validation: Beyond the original benchmark evaluations, we conduct additional experiments on mutation existence, independent predictive utility, selective mutation modeling, prototype reuse and temporal consistency, and robustness to incidental perturbations. Additional baseline comparisons and ablation analyses further demonstrate the competitiveness of TE-MDG while clarifying its dataset-dependent benefits and limitations.

Addressed Concerns

We have carefully addressed the reviewers' concerns through additional theoretical analyses, supplementary experiments, and detailed clarifications. The progress is summarized below.

  • We addressed Reviewer J24M's concerns regarding supplementary theoretical proofs (W1, Q1), mutation identification and selective modeling (W2, Q2), cross-node and temporal prototype consistency (Q3), evaluation discrepancies (W2, Q4), ablation correctness (W2, Q5), and TE estimation and dataset-dependent contributions (W1, W2). We conducted supplementary experiments and theoretical explanations to address these concerns.

  • We addressed Reviewer P1P5's concerns regarding recent SOTA comparisons (W1, Q1), mutation existence and identification (W2, Q2), and TE-guided selective mutation modeling (W2, Q2) through additional baseline evaluations, statistical analyses, and counterfactual experiments.

After we addressed the other two reviewers' concerns, Reviewer 24G6 (additional reviewer) evaluated our revised manuscript and raised further questions.

  • We addressed Reviewer 24G6's further concerns regarding the mutation filtering mechanism and its validation (Q1), prototype stochasticity and practical training (W1, Q2), conditional independence assumptions and theoretical derivations (Q3), mutation filtering compared with baselines and component-wise contributions (W4), broader SOTA comparisons, dataset-dependent performance, applicability, and limitations (W2, W3, Q5), w/o TE versus SOTA comparisons (Q4), and architectural background, references, and manuscript presentation (Q6).

We sincerely thank all three reviewers for their insightful comments and constructive suggestions, which have been invaluable in improving our work. Overall, these revisions have strengthened the theoretical rigor, empirical validation, and evaluation completeness of our work, while providing a clearer understanding of TE-MDG's contributions, applicability, and limitations.

Reviewer Feedback

During the rebuttal and discussion period, Reviewer 24G6 provided positive follow-up feedback, stating that "My concerns have been mostly addressed." We sincerely appreciate the reviewer's recognition of our responses and the constructive suggestions that have helped us further improve our work.


We sincerely thank you again for your time and consideration throughout the review process. We hope that the above summary demonstrates how we have constructively responded to every point raised by reviewers. All corresponding revisions have been incorporated into the revised manuscript, with the changes highlighted for ease of reference. We greatly appreciate the opportunity to strengthen our work and hope this summary will facilitate the next stage of discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.