acceptodds
Under review as a conference paper at ICLR 2027

Dynamic Hypergraph Source Identification via Topology-aware Mamba and Diffusion Supervision

Abstract

Source identification in dynamic hypergraphs is challenging because evolving hyperedges continuously alter the available propagation paths, making diffusion dynamics strongly dependent on structural evolution. Existing methods mainly learn diffusion patterns from snapshot sequences, but often overlook dynamic higher-order interactions. Therefore, we propose a framework for dynamic hypergraph source identification via Topology-guided Mamba and Diffusion Supervision (TMDS). To capture the hyperedge evolution, the hyperedge evolution modeling module first characterizes node-level structural evolution from participation scale, neighbor composition, and group organization. These transition signals are then incorporated into a dynamic topology-aware Mamba, where retention and propagation gates adapt temporal inference to evolving higher-order structures. Furthermore, a forward diffusion supervision mechanism reconstructs the diffusion trajectory from the predicted source distribution and feeds the reconstruction discrepancy back to the inference model. Experiments on four diverse datasets demonstrate consistent improvements over representative baselines and verify the effectiveness of the proposed approach.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
source identification, social networks, information propagation

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_G81wOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the source identification problem in the more complex setting of hypergraphs. Building on prior work, it introduces two enhancements. First, it explicitly incorporates expert knowledge to model dynamical changes in graph structure, thereby facilitating information flow within a Mamba architecture. Second, it incorporates consistency in the diffusion process as an additional constraint. Experiments demonstrate the validiy of the design.

Critical strengths
  1. The paper is clearly written and well structured. Its coherent organization makes it easy to follow the arguments and locate relevant information.
  2. The method is well designed. It explicitly computes an interpretable signal of structural change and uses it to control how much historical state is retained at each step. Since this signal is decoupled from the backbone network, it could, in principle, be incorporated into other sequence-based source identification models, offering a useful and transferable design principle.
  3. The experimental design is sound. In particular, the ablation study provides convincing evidence for the effectiveness of the proposed design choices.
Critical weaknesses
  1. The paper does not clearly introduce classical approaches to source identification, instead moving directly to data-driven methods. This makes it difficult for readers to understand the underlying mathematical principles. The authors should also conduct a more thorough literature review to determine whether approaches grounded in statistics or mathematical physics have been developed for source identification on hypergraphs and, where applicable, discuss and compare these methods with the proposed approach.
  2. Given the stochastic nature of the diffusion process, the experimental results may exhibit substantial variability. The evaluation should therefore assess statistical robustness, for example by reporting variability across repeated runs. The paper should also clarify the underlying stochastic process: are the diffusion trajectories obtained from real-world observations or simulated according to predefined rules? If they are simulated, the authors should specify the generative mechanism and consider additional validation using real-world diffusion data.
  3. According to the calculation of F-score and ACC, we can obtain the following derivatives. Whether the ACC should be equal to F-score or not?

Substituting $ F_1^+ $ and $ F_1^-$ into F1,weightedF_{1,\text{weighted}} gives

F1,weighted=1N[(TP+FN)⋅2TP2TP+FP+FN+(TN+FP)⋅2TN2TN+FP+FN]F_{1,\text{weighted}} = \frac{1}{N}\left[(TP+FN)\cdot\frac{2TP}{2TP+FP+FN} + (TN+FP)\cdot\frac{2TN}{2TN+FP+FN}\right]

After transformation, we can get

F1,weighted=1N[TP⋅2(TP+FN)2TP+FP+FN+TN⋅2(TN+FP)2TN+FP+FN].F_{1,\text{weighted}} = \frac{1}{N} [ TP\cdot \frac{2(TP+FN)}{2TP+FP+FN} + TN\cdot \frac{2(TN+FP)}{2TN+FP+FN} ] .

Because of the number of predicted source set matching the number of ground-truth sources mentioned in B.1, so we can get

FP=FNFP = FN

After substituting into F1,weightedF_{1,\text{weighted}}, we can obtain

F1,weighted=1N[TP+TN]=ACCF_{1,\text{weighted}} = \frac{1}{N} [ TP + TN ] = ACC


So the F-score should be equal to ACC. But this is not in this paper. Please clarify this.

Questions for authors

Please refer to weakness. I will consider to raise my score if the rebuttal can address my concerns.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We sincerely thank the reviewer for the positive assessment of the paper’s clarity, methodological design, and experimental evaluation, as well as for recognizing the interpretability of the proposed topology-aware mechanism. In response to the reviewer’s comments, we further strengthen the discussion of classical source-identification methods, statistical robustness and diffusion generation, and the definition of the evaluation metrics, as detailed below.


Q1: The paper does not clearly introduce classical approaches to source identification, instead moving directly to data-driven methods. This makes it difficult for readers to understand the underlying mathematical principles. The authors should also conduct a more thorough literature review to determine whether approaches grounded in statistics or mathematical physics have been developed for source identification on hypergraphs and, where applicable, discuss and compare these methods with the proposed approach.

A1: Thank you for suggesting a fuller discussion and evaluation of classical approaches. To address this omission, we conduct a more thorough literature review and additional performance comparisons.

## Literature Review of Classical Methods:

(1.1) Beyond data-driven approaches, we introduce representative classical methods for source localization on hypergraphs. Specifically,

  • DMP [1] infers sources through dynamic message passing derived from an explicit SIR process on factor-graph representations of hypergraphs;
  • HDMPN [2] performs likelihood-based inference under stochastic SI dynamics;
  • Hypertree [3] derives an approximate maximum-likelihood estimator for SI contagion processes on structurally constrained hypertrees.

(1.2) These methods provide a complementary perspective to data-driven approaches, and we will incorporate this discussion into the revised Section 2.

## Empirical Comparison with Classical Baselines:

(1.3) Following your suggestion, we reproduced DMP, HDMPN, and Hypertree, and perform repeated runs under different diffusion realizations across four datasets. The results are reported in the table below.

Table A: Performance comparison of classical baselines and TMDS across four datasets.

MethodHouseCoraNTUTest-Bills
DMP [1]0.7517 ± 0.00540.7439 ± 0.00590.7832 ± 0.00510.7374 ± 0.0056
HDMPN [2]0.7518 ± 0.00580.7422 ± 0.00620.7768 ± 0.00550.7374 ± 0.0059
Hypertree [3]0.7530 ± 0.00510.8066 ± 0.00490.8137 ± 0.00470.7374 ± 0.0053
TMDS0.9400 ± 0.00380.9170 ± 0.00400.9490 ± 0.00350.9160 ± 0.0045

TMDS consistently outperforms these statistic-based approaches across all datasets. These results suggest that explicitly modeling evolving higher-order topology and multiple observations is beneficial in the evaluated setting.

## References:

[1] Yu X, Nie Y, Li W, et al. Source inference for misinformation spreading on hypergraphs[J]. Chaos, Solitons & Fractals, 2024, 187: 115457.

[2] Ke Q, Masuda N, Jin Z, et al. Source detection in epidemic dynamics on hypergraphs using a dynamic message passing algorithm[J]. Chaos: An Interdisciplinary Journal of Nonlinear Science, 2026, 36(1).

[3] Spencer S, Varshney L R. Social Bubbles and Superspreaders: Source Identification for Contagion Processes on Hypertrees[C]. 2021 IEEE Statistical Signal Processing Workshop (SSP). IEEE, 2021: 471-475.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Q2: Given the stochastic nature of the diffusion process, the experimental results may exhibit substantial variability. The evaluation should therefore assess statistical robustness, for example by reporting variability across repeated runs. The paper should also clarify the underlying stochastic process: are the diffusion trajectories obtained from real-world observations or simulated according to predefined rules? If they are simulated, the authors should specify the generative mechanism and consider additional validation using real-world diffusion data.

A2: Thank you for recommending a clearer account of variability and the diffusion-generation process. Following your suggestion, we reported mean+std across repeated runs, clarified the stochastic diffusion-generation process, and discussed the limitation regarding real-world diffusion data.

## Mean and Standard Deviation Results:

(2.1) All experiments are repeated over five random seeds 0,1,2,3,4\\{0,1,2,3,4\\}. We report the mean and standard deviation for all methods and metrics across repeated runs in the revised version.
(2.2) Due to space limitations, we present here a representative subset, comparing the F1 results of TMDS with the strongest learning-based baseline, SDMamba, across various datasets at β=0.15\beta=0.15.

Table A: Performance of SDMamba and TMDS across four datasets over five random seeds.

MethodHouseCoraNTUTest-Bills
SDMamba0.918 ± 0.0060.868 ± 0.0060.928 ± 0.0050.881 ± 0.005
TMDS0.940 ± 0.0040.917 ± 0.0040.949 ± 0.0040.916 ± 0.005

The small standard deviations indicate the stability of TMDS performance under repeated runs over five random seeds.

## Data Generation for Diffusion Trajectories:

(2.3) The diffusion trajectories are generated by an SI-like stochastic process on empirical hypergraph structures.

(2.4) Real-world dynamic dataset. Test-Bills provides a real-world dynamic hypergraph sequence with six temporally ordered snapshots.

  • To generate the diffusion trajectories, source nodes are randomly initialized according to a fixed source ratio, and SI-like stochastic diffusion is then simulated over the observed dynamic structures. At each diffusion step, susceptible nodes exposed through the current hyperedges become infected according to the propagation rate β\beta, while infected nodes remain infected. Diffusion noise further perturbs the propagation probability to introduce stochastic uncertainty.

(2.5) Real-world static datasets. House, Cora, and NTU provide static empirical hypergraphs.

  • To construct dynamic topology sequences, we randomly apply hyperedge addition, deletion, or reorganization between consecutive snapshots while keeping the node set fixed. The amount of structural change is controlled by the topology-evolution intensity ρ\rho.
  • The same SI-like diffusion process is then simulated over the evolving structures.

(2.6) Infection states are sampled at fixed propagation intervals to form the observed snapshot sequence. The default interval is 0.10.1, with additional evaluations over 0.050.05–0.200.20.

## Limitation for Real-World Diffusion Trajectories:

(2.7) We agree that evaluation on observed real-world diffusion trajectories would further strengthen the empirical evidence. Existing source-identification studies commonly rely on simulated diffusion processes on network structures, while complete diffusion trajectories with verified source labels remain limited. We clarify this limitation in the revised version and will use real-world diffusion data in future studies.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Q3: According to the calculation of F-score and ACC, we can obtain the following derivatives. Whether the ACC should be equal to F-score or not?

A3: Thank you for drawing our attention to this inconsistency. We apologize for the inconsistent metric definition in the Appendix. We will correct the metric description in the main text and replace the class-weighted definition in Appendix B.1 with the positive-class F1 definition used in our experiments.

## Positive-Class F1 in Our Experiments:

(3.1) Our implementation and all reported F1 results use the positive-class F1F_1, with source nodes treated as the positive class: F1=2TP2TP+FP+FN.F_1 = \frac{2TP}{2TP+FP+FN}.

(3.2) Accuracy is computed as ACC=TP+TNTP+TN+FP+FN.\mathrm{ACC}=\frac{TP+TN}{TP+TN+FP+FN}.

(3.3) Since ACC additionally depends on true negatives, whereas positive-class F1F_1 does not, the two metrics are not generally equal.

(3.4) We use positive-class F1F_1 because source identification primarily concerns the accurate recovery of source nodes. Thus, positive-class F1F_1 is utilized, which focuses specifically on the precision and recall of the source class and does not include true negatives.


We again thank the reviewer for the valuable suggestions, which have helped us improve the literature review, experimental robustness, and clarity of the evaluation protocol. We hope these responses help address your concerns and would welcome any further questions or suggestions.

Thank you for the detailed response and the additional experiments (classical baselines, mean ± std over five seeds, and the clarified diffusion and metric protocol); I have raised my rating.

Reviewer_ib7COpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TMDS for source identification in dynamic hypergraphs, where evolving hyperedges reshape propagation paths. It combines three components: (1) node-level topology-change descriptors (participation scale, neighbor composition, group organization) computed across adjacent snapshots; (2) a reverse-time Mamba with retention/propagation gates derived from those descriptors; and (3) forward diffusion supervision that reconstructs the cascade from the predicted source and penalizes trajectory mismatch. Evaluated on four datasets against six baselines, it reports the best accuracy, F1, and AUC, backed by extensive ablations.

Critical strengths
  1. Well-chosen problem. Source identification under evolving higher-order structure is genuinely understudied, and the motivating example is intuitive.
  2. Clean mechanism design. The three-level topology descriptors are interpretable, and the gates respond monotonically to change magnitude with measured saturation below 5%.
Critical weaknesses
  1. Unclear challenge–method mapping. The introduction poses three challenges, but the method section never explicitly states which component addresses which challenge or why the proposed mechanism resolves it — the reader must reconstruct the intended mapping themselves. This makes the technical narrative difficult to follow and leaves it unverifiable whether each stated challenge is actually addressed, which directly weakens the paper's claimed contributions.

  2. Unclear why hypergraphs are needed, limiting real-world applicability. The paper never establishes that higher-order structure is necessary — only one of the three descriptors truly requires hyperedges, and TMDS is never ablated on a pairwise conversion. Moreover, in real-world applications hyperedges are rarely observed directly and must be heuristically constructed, so a method that depends on known higher-order relations has limited practical applicability.

  3. Experiments are not convincing. The largest dataset has only 2,012 nodes, observation sequences are just T = 3 snapshots, and three of four datasets rely on random hyperedge perturbations rather than real dynamics — too small and too synthetic to substantiate the paper's claims.

Questions for authors
  1. Which method component addresses which of the three stated challenges, and why does each mechanism resolve its challenge? The mapping is currently left for the reader to infer.

  2. What happens when TMDS is run on the weighted 2-section graph conversion used for the graph baselines? This would directly test whether higher-order structure is necessary for the reported gains.

  3. In real-world applications hyperedges are rarely observed directly — how were they obtained for Test-Bills, and how sensitive is the method to heuristically constructed hyperedges?

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/4)

We sincerely thank the reviewer for the careful evaluation and for recognizing the importance of source identification under evolving higher-order structures, as well as the interpretability of the proposed topology-transition modeling and gating mechanisms. We have carefully addressed the concerns regarding the challenge–method mapping, the necessity and availability of higher-order structures, and the scale and realism of the experimental evaluation, as detailed below.


W1&Q1: Unclear challenge–method mapping. The introduction poses three challenges, but the method section never explicitly states which component addresses which challenge or why the proposed mechanism resolves it — the reader must reconstruct the intended mapping themselves. This makes the technical narrative difficult to follow and leaves it unverifiable whether each stated challenge is actually addressed, which directly weakens the paper's claimed contributions. Which method component addresses which of the three stated challenges, and why does each mechanism resolve its challenge? The mapping is currently left for the reader to infer.

A1: Thank you for pointing out that the connection between the challenges and the proposed components was not sufficiently clear. We appreciate the opportunity to clarify this mapping.

## Explicit Challenge–Method Mapping

(1.1) In TMDS, the first two components form a sequential mechanism, the third component provides trajectory-level supervision:

  • HEM module determines what structural changes occurred;
  • DTM module determines how reverse temporal inference should respond to those changes;
  • FDS module evaluates whether the inferred sources can explain the observed cascade.

Table A: Mapping between the identified challenges and the corresponding TMDS components.

ChallengeCorresponding component
1. Lack of topology-evolution modelingHyperedge Evolution Modeling (HEM)
2. Limited topology sensitivity in temporal modelsDynamic Topology-aware Mamba (DTM)
3. Propagation reconstruction gapForward Diffusion Supervision (FDS)

## How to Address Its Corresponding Challenge:

(1.2) HEM addresses challenge 1: Challenge 1 concerns the lack of explicit modeling of topology evolution across snapshots. HEM compares adjacent hypergraph snapshots and constructs the node-level transition descriptor (q_{i,t}) from three complementary aspects: participation scale, neighbor composition, and group organization. These descriptors characterize changes in group participation, interaction partners, and hyperedge organization, including splitting and merging behaviors. Therefore, structural evolution is explicitly quantified across snapshots rather than remaining implicit in independently encoded representations.

(1.3) DTM addresses challenge 2: Challenge 2 concerns the limited sensitivity of temporal models to evolving propagation paths. DTM incorporates the topology-transition descriptor qi,tq_{i,t} into the reverse-time state update through retention and propagation gates. When structural variation is strong, the retention gate reduces the contribution of historical states, while the propagation gate increases the contribution of hypergraph messages under the current topology; when the topology is relatively stable, more historical information is preserved. This design makes temporal inference directly responsive to structural evolution and changing higher-order propagation paths.

(1.4) FDS addresses challenge 3: Challenge 3 concerns the lack of propagation-level validation for inferred sources. FDS uses the predicted source probabilities to initialize forward diffusion over the observed dynamic hypergraph sequence and reconstructs the propagation trajectory. The discrepancy between the reconstructed and observed states is incorporated into the training objective as trajectory-level supervision. Thus, source predictions are optimized not only against source labels but also according to how well they explain the observed diffusion process.

AuthorsOpenReview ↗

Rebuttal by Authors (2/4)

W2&Q2: Unclear why hypergraphs are needed, limiting real-world applicability. The paper never establishes that higher-order structure is necessary — only one of the three descriptors truly requires hyperedges, and TMDS is never ablated on a pairwise conversion. Moreover, in real-world applications hyperedges are rarely observed directly and must be heuristically constructed, so a method that depends on known higher-order relations has limited practical applicability. What happens when TMDS is run on the weighted 2-section graph conversion used for the graph baselines? This would directly test whether higher-order structure is necessary for the reported gains.

A2: Thank you for asking us to clarify the role of higher-order structure and its practical availability. Our intended claim is specific to the scenarios targeted by TMDS; we do not claim that hypergraphs are necessary for all source-identification problems. The choice of representation should depend on the underlying diffusion process. Below, we explain the motivation for higher-order modeling, compare TMDS on the original hypergraph and its weighted 2-section representation, and discuss the practical limitation.

## Why Higher-order Structure Matters:

(2.1) Scenario perspective:

  • Group interactions are inherent in many diffusion scenarios. Examples include discussion threads, co-sponsorship groups, and multi-user communication events, where a single event reaches multiple nodes at the same step. Converting group interactions into pairwise edges preserves only node-to-node co-occurrence, so different group organizations may induce nearly identical pairwise connectivity while implying very different propagation processes. Hence, preserving higher-order structure is particularly important in such scenarios. Faithful modeling requires a representation that retains the group structure itself, which is exactly provided by hypergraph formulation in TMDS.

(2.2) Method perspective:

  • The framework of TMDS is designed around the hypergraph structure. Its core modules, including hypergraph aggregation and propagation modeling, operate directly on node–hyperedge–node interactions rather than on pairwise adjacency. The hypergraph is therefore not a preprocessing choice but the structural basis of the whole pipeline.

  • The propagation features are defined at the hyperedge level. The participation scale and neighbor-composition signals in TMDS describe how nodes are reorganized among hyperedges over time, that is, how the group structure itself evolves. Such changes can be partially characterized from pairwise relations, yet the hyperedge-level formulation describes them explicitly and avoids lossy projection.

## Empirical Evaluation on Weighted 2-section Graph:

(2.3) We conduct additional ablation on the same weighted 2-section graph conversion used for the graph-based baselines. In this variant, the hypergraph structure is replaced by its weighted pairwise projection. The experimental results are shown below.

Table A: Performance comparison of TMDS using weighted 2-section graphs and original hypergraphs.

StructureNTU ACCNTU F1NTU AUCTest-Bills ACCTest-Bills F1Test-Bills AUC
Weighted 2-section graph0.8440.8950.9020.8320.8860.896
Original hypergraph0.9190.9490.9670.8590.9160.923

These results provide empirical evidence that preserving higher-order relations contributes to the reported gains of TMDS.

## Limited Availability of Higher-order Relations:

(2.4) We agree that reliable hyperedges are not directly observable in all real-world applications. In settings such as discussion threads or co-sponsorship events, hyperedges can be obtained from observed group activities. We will clarify this setting and the corresponding scope of our study in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal by Authors (3/4)

W3: Experiments are not convincing. The largest dataset has only 2,012 nodes, observation sequences are just T = 3 snapshots, and three of four datasets rely on random hyperedge perturbations rather than real dynamics — too small and too synthetic to substantiate the paper's claims.

A3: Thank you for highlighting the need for evaluation on larger networks and longer observation sequences. In response, we extended the evaluation to a larger real-world dynamic hypergraph and to longer observation sequences in the revised manuscript.

## Evaluation on Large-Scale Real-World Dataset:

(3.1) To address the concern about dataset scale, we use the public threads-Ask-Ubuntu dataset:

  • A real-world temporal higher-order interaction dataset containing 125,602 users and 192,947 timestamped hyperedges.
  • Each hyperedge consists of users participating in the same discussion thread, and the timestamps directly provide the temporal evolution of the hypergraph.
  • No random hyperedge perturbations are required to construct the dynamic topology.

(3.2) We report the performance comparison on the threads-Ask-Ubuntu dataset below.

Table A: Performance comparison on the large-scale real-world threads-Ask-Ubuntu dataset.

MethodAsk-Ubuntu: ACCAsk-Ubuntu: F1Ask-Ubuntu: AUC
GIN-SD0.81280.84270.8735
IVGD0.83620.86190.8946
SL-VAE0.82350.85280.8861
HHAN0.85540.88960.9180
HyperDet0.87310.90350.9312
SDMamba0.88760.92180.9441
TMDS0.91250.94510.9666

TMDS maintains strong performance on the threads-Ask-Ubuntu dataset, providing evidence of its applicability to larger-scale hypergraphs with observed temporal topology under simulated diffusion.

## Evaluation on Longer Sequences:

(3.3) To address the concern about observation length, we further evaluate longer snapshot sequences with T∈5,10,15T \in \\{5,10,15\\} on the Cora and NTU datasets. The experimental results are reported below.

Table B: Performance of TMDS on Cora and NTU under different observation sequence lengths.

TTCora: ACCCora: F1Cora: AUCNTU: ACCNTU: F1NTU: AUC
50.85870.90850.89280.91880.94870.9676
100.84670.90080.87820.91840.94850.9659
150.84280.89840.88550.91640.94720.9658

TMDS achieves comparable performance across snapshot sequences with T∈5,10,15T \in \\{5,10,15\\}, suggesting its effectiveness under longer observation sequences.

Reviewer_6NmZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces TMDS (Topology-guided Mamba and Diffusion Supervision), which is an approach for the task of source detection in dynamic hypergraphs. Existing source identification approaches consider a static representation or snapshot based representation without considering the structural evolution of the hypergraphs. The proposed approach tries to solve this limitation. TMDS consists of three components- a hypergraph evolution process, a topology-aware temporal backbone based on Mamba, and FDS (Forward Diffusion Supervision). TMDS is assessed using four different datasets.

Critical strengths
  1. Moving beyond pairwise graphs to dynamic hypergraphs is a significant and realistic direction for network science and diffusion modeling, as real-world social groups constantly evolve.

  2. The integration of Mamba (selective SSMs) with topology-guided gating is a clean, well-motivated technical contribution. Rather than treating temporal dynamics as a black box, the gates explicitly adapt memory decay and message passing to structural variations.

Critical weaknesses

Eq. 13 is a plain linear time invariant recurrence with fixed A,B and C. The main component of the proposed method is the topology aware mamba. In mamba, state space parameters are input dependent and decides what to remember. However, the parameters in eq. 13 are not input dependent. Instead, topology guided gating in eq. 16 controls how much information to retain from previous timestamp. This is a gated linear RNN, not mamba. In ablation (Table 9), GRU performs better than standard mamba which shows that gains of TMDS is coming from hypergraph aggregation and topology gating, not from ssm. Also, sequence length is only 3 for House, Cora and NTU datasets where SSMs are effective in modeling long sequences.

Questions for authors
  • How is the state space selection done in eq. 13?
  • Can TMDS model long range temporal evolution? The gating function in eq. 15 follows exponential decay. What's the impact of this for long temporal hypergraph?
  • In table 1, is there a reason not to report mean$\pm$ std?
  • How would TMDS perform against GRU + hypergraph aggregation + topology gating?
5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We sincerely thank the reviewer for the careful evaluation and for recognizing the significance of dynamic hypergraph source identification and the technical contribution of combining selective SSMs with topology-guided gating. We have carefully considered the concerns regarding the Mamba formulation, long-range temporal modeling, statistical reporting, and temporal-backbone comparison, and address them point by point below.


W1&Q1: Eq. 13 is a plain linear time invariant recurrence with fixed A,B and C. The main component of the proposed method is the topology aware mamba. In mamba, state space parameters are input dependent and decides what to remember. However, the parameters in eq. 13 are not input dependent. Instead, topology guided gating in eq. 16 controls how much information to retain from previous timestamp. This is a gated linear RNN, not mamba. How is the state space selection done in eq. 13?

A1: Thank you for pointing out the ambiguity in Eq. 13. We apologize for the oversimplified fixed LTI formulation, which did not fully reflect the selective mechanism used in TMDS. In our implementation, we employ the standard input-dependent selection of Mamba, while the topology-guided gates provide additional modulation according to hypergraph evolution.

## State Space Selection in Eq. 13:

(1.1) In TMDS, state space selection follows the standard Mamba formulation.

(1.2) Given the current snapshot representation xtx_t, the selective parameters are generated as Δt=fΔ(xt),Bt=fB(xt),Ct=fC(xt).\Delta_t=f_{\Delta}(x_t), \qquad B_t=f_B(x_t), \qquad C_t=f_C(x_t).

These input-dependent parameters determine the discretized state transition

Aˉt=exp⁡(ΔtA),Bˉt=Discretize⁡(A,Bt,Δt),\bar{A}_t=\exp(\Delta_t A), \qquad \bar{B}_t=\operatorname{Discretize}(A,B_t,\Delta_t),

followed by the selective state update

ht=Aˉtht−1+Bˉtxt,yt=Ctht.h_t=\bar{A}_t h_{t-1}+\bar{B}_t x_t, \qquad y_t=C_t h_t.

(1.3) Thus, the state space update in TMDS is input-dependent, consistent with the selective mechanism of Mamba, rather than the fixed recurrence suggested by the original presentation of Eq. 13.

## The Role of Topology-Guided Gating:

(1.4) The topology-guided gates serve a different role from the state space selection: the transition descriptor qi,tq_{i,t} generates the retention gate gi,tRg_{i,t}^{R} and propagation gate gi,tPg_{i,t}^{P}, which further adapt the update to structural evolution.

hi,t=Aˉt(gi,tRhi,t−1)+Bˉtxi,t+(1+gi,tP)mi,t.h_{i,t} = \bar{A}_t\left(g_{i,t}^{R}h_{i,t-1}\right) + \bar{B}_t x_{i,t} + \left(1+g_{i,t}^{P}\right)m_{i,t}.

(1.5) TMDS combines the input-dependent state-space selection of Mamba with topology-dependent regulation based on structural evolution.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Q2: Can TMDS model long range temporal evolution? The gating function in eq. 15 follows exponential decay. What's the impact of this for long temporal hypergraph?

A2: Thank you for asking about the effect of the retention gate on longer sequences. The exponential decay form in Eq. 15 is determined by topology variation rather than temporal distance. Therefore, the effect of the gate on long-range information depends on the accumulated structural evolution. We explain this mechanism below and provide longer-sequence experiments to further examine its behavior.

## Impact for Long Temporal Hypergraph:

(2.1) The exponential form in Eq. 15 is defined as gi,tR=exp⁡(−wr⊤qi,t),g_{i,t}^{R}=\exp\left(-w_r^\top q_{i,t}\right), where qi,tq_{i,t} measures the structural change between two adjacent hypergraph snapshots.

(2.2) The impact of the exponential form is mainly determined by structural evolution, rather than by the temporal horizon itself.

  • When structural changes persist over many steps: earlier states are gradually downweighted to reduce the effect of propagation information associated with outdated structures.
  • When the structure changes only mildly: the retention gate remains close to 1, allowing historical information to be preserved over longer sequences.

(2.3) Overall, the decay factor in Eq. 15 depends on the degree of topology evolution rather than on the sequence length, which adaptively controls how much historical information is retained according to structural changes. This behavior is consistent with the intended design: when the underlying propagation structure changes substantially, historical information should contribute less to the current estimation.

## Empirical Evaluation on Long Sequences:

(2.4) To address this question, we further evaluate longer observation sequences with T∈5,10,15T \in \\{5,10,15\\} on the Cora and NTU datasets to examine whether TMDS remains effective over longer temporal hypergraph sequences.

Table A: Performance of TMDS on Cora and NTU under different observation sequence lengths

TTCora: ACCCora: F1Cora: AUCNTU: ACCNTU: F1NTU: AUC
50.85870.90850.89280.91880.94870.9676
100.84670.90080.87820.91840.94850.9659
150.84280.89840.88550.91640.94720.9658

TMDS maintains consistently strong performance across longer observation sequences on various datasets.


Q3: In Table 1, is there a reason not to report mean+std?

A3: Thank you for pointing out the missing standard deviations. We apologize for the omission. We omitted them to keep the table compact; the experimental protocol itself was unchanged.

(3.1) For completeness, we have added the standard deviations in the revised manuscript.

(3.2) All experiments are repeated over five random seeds 0,1,2,3,4\\{0,1,2,3,4\\}. We present a representative subset below, comparing the F1 results of TMDS with the strongest learning-based baseline, SDMamba, across various datasets at β=0.15\beta=0.15.

Table B: Performance of SDMamba and TMDS over five random seeds

MethodHouseCoraNTUTest-Bills
SDMamba0.918 ± 0.0060.868 ± 0.0060.928 ± 0.0050.881 ± 0.005
TMDS0.940 ± 0.0040.917 ± 0.0040.949 ± 0.0040.916 ± 0.005

The small standard deviations indicate limited variability across repeated runs, while TMDS consistently outperforms SDMamba on various datasets.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Q4: In ablation (Table 9), GRU performs better than standard mamba which shows that gains of TMDS is coming from hypergraph aggregation and topology gating, not from ssm. How would TMDS perform against GRU + hypergraph aggregation + topology gating?

A4: Thank you for suggesting this comparison. We agree that the original ablation did not clearly assess the contribution of the temporal backbone, and we apologize for this lack of clarity.

## TMDS vs. GRU under Topology-aware Mechanisms:

(4.1) We construct a GRU variant equipped with the same hypergraph aggregation and topology-guided gating as TMDS, while keeping the remaining components and training settings unchanged.

Table A: Ablation study between TMDS and the topology-aware GRU variant on Cora and Test-Bills

MethodDatasetF1AUCTime (s)
GRU + Hypergraph Aggregation + Topology GatingCora0.90240.9134383.06
TMDSCora0.91700.9190368.00
GRU + Hypergraph Aggregation + Topology GatingTest-Bills0.89920.9092623.59
TMDSTest-Bills0.91600.9230506.00

TMDS (Mamba + Hypergraph Aggregation + Topology Gating) consistently achieves better performance than the GRU variant on various datasets.

## Contribution of Temporal Backbone of TMDS:

(4.2) The experiments above indicate that the Mamba backbone still provides an additional benefit within the complete TMDS architecture. This also suggests that the performance of Standard Mamba in Table 9 may not fully represent the contribution of the state-space backbone within the complete TMDS architecture; its effect becomes more favorable when combined with the proposed topology-aware update mechanism.


We again thank the reviewer for the constructive comments, which have helped us clarify the Mamba formulation and further strengthen the temporal modeling and experimental evaluation. We hope these responses help address your concerns and would welcome any further questions or suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.