acceptodds
Under review as a conference paper at ICLR 2027

Learning on Dynamic Hypergraphs from Incidence Streams

Abstract

Hypergraphs model higher-order interactions, and in many systems their hyperedges persist: nodes join over time, carry attributes such as roles and may leave. Dynamic hypergraph learning addresses separate tasks on transient hyperedges, while persistent hyperedges are reduced to their star expansion for temporal graph networks. Neither line characterizes what an evaluation protocol reveals beyond the history or what a model must retain. A unified formulation of all tasks would make both questions precise. We represent a dynamic hypergraph as a stream of incidence records and factorize its likelihood, so that every task is a conditional of one factorization. Protocol leakage is then the conditional mutual information an exposure adds to the history, equal to the Bayes log-risk it removes. For persistent processes with exponential kernels and Markov roles (Dirichlet or known role kernels), a finite state of node, hyperedge, pair and incidence statistics suffices. First-hop readouts of the star expansion cannot represent its pair terms. The incidence state machine (ISM) maintains this state in one pass and reads it by attention pooling. On persistent streams from public datasets, the ISM performs best overall across join, role and departure prediction and outperforms memory-based and first-hop temporal graph networks on joining.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
dynamic hypergraphs, temporal hypergraph learning, hypergraph incidence

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_F2b8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies dynamic hypergraphs made up of a timeline of records. The main concern the paper addresses is protocol leakage (information beyond the historical record that is unfairly provided to a model that inflates scores) and what state a model should keep. The stream likelihood is factorized where each task is condition of it and leakage is defined as conditional mutual information. The paper claims that under a specific model of group dynamics, a small set of counters, including pair-terms, is sufficient. The authors' proposal is the incidence state machine (ISM), which they evaluate on a new benchmark from Stack Exchange. dump-time information is found by the leakage audit to inflate scores. ISM is reported the best performer, compared to temporal graph networks.

Critical strengths
  1. The audit is useful and has practical impact. It found that data dumps add features that provide after-the-fact information, like the final number of answers in a thread, which raises role-prediction macro-F1 by up to 0.19. This difference is larger than the difference between any two models. It also found that a smaller leak is detected when processing same day events in the order the file present.

  2. The evaluation is rigorous. The benchmark was carefully built and evaluated with known leaks, stratified negatives, using multiple random seeds, careful treatment of time dependence, and honest disclosures in the appendix. These make the benchmark and leakage reports credible.

Critical weaknesses

Although I see potential in the paper's directions, my concern is that the paper as it stands has major flaws whose possible fixes would produce a substantially different paper. In light of this, I'd argue for rejection in its current form. My main concerns are:

  1. By construction, the theory is true. Proposition 2 shows sufficiency only for a process constructed from the stored statistics, ie, persistent processes with exponential kernels, Markov roles and parameters known in advance. This matters because we estimate the parameters in practice, and the sufficiency result won't hold as stated. Fixing it with a new theorem that holds broadly would require a major rework of the paper. Reframing the claims would produce a narrower paper, different from the submission. The paper presents theory based on textbook standard definitions, such as results from conditional mutual/usable information, Pinsker's inequality. I believe the paper would be stronger if it demotes its core from theory to a benchmark scaffolded by information theory.

  2. Performance claims don't match results. The abstract claims "best overall" and "outperforms" but the authors' own results in the appendix show a different picture. In D.1 on aligned update schedules, ISM is 014-.039 below TPNet in one direction and at most .008 ahead in the other. The authors conclusion "parity with TPNet on T-JOIN, not a margin." goes against the abstract.

  3. A single data source. That matters, because the pair-terms the authors investigate only account for 0.003 MRR on joining and no more than 0.030 AT on leaving. A dataset where pair-terms matter is needed to support the theory. Novelty is also a concern. The paper should compare its proposal with prior work like DyNAM-i (not cited) and RHEM (cited without further comparison), which already model group events from relational statistics.

Questions for authors

The weaknesses discussed are larger than missing details. Addressing them would require a new theorem, a new platform, or a reframed paper. So, I don't have questions about the paper's details. I'm open to reconsider my assessment if I misunderstood any assumptions of Proposition 2 or the results in D.1.

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

This paper studies dynamic hypergraph learning by representing persistent hyperedges as streams of incidence records. It introduces a unified probabilistic factorization for joining, role prediction, and departure prediction, and formalizes protocol leakage through conditional mutual information. The authors further derive sufficient statistics for a family of persistent processes and propose the Incidence State Machine (ISM), which maintains node, hyperedge, pair, and incidence statistics using streaming updates and attention pooling. Experiments on five Stack Exchange datasets demonstrate the importance of avoiding information leakage and show competitive performance across three prediction tasks. However, the improvements over strong baselines are generally modest, leaving the practical significance of the proposed architecture unclear.

Critical strengths
  1. Unified problem formulation. Representing persistent hypergraphs as incidence streams provides a coherent formulation for multiple prediction tasks, including joining, role prediction, and departure prediction.

  2. Careful evaluation protocol. The analysis of information leakage from future attributes and records sharing timestamps identifies important evaluation pitfalls. The benchmark and leakage audit provide useful insights for evaluating dynamic hypergraph models.

Critical weaknesses
  1. Limited methodological novelty. The proposed ISM primarily combines manually designed temporal statistics, exponential filters, and standard neural components such as MLPs and attention pooling. Similarly, the probabilistic factorization and mutual information formulation build on established principles. The work introduces limited new mechanisms for learning dynamic hypergraph representations.

  2. Limited generality of theoretical results. The sufficiency guarantee is established for a specifically constructed process family whose predictive dependencies are explicitly encoded in the proposed statistics. Furthermore, the expressivity analysis considers restricted first hop readouts and bounded memory representations. These results do not establish that the proposed state is necessary or sufficient for general dynamic hypergraph learning.

  3. Marginal empirical improvements. The ISM achieves nearly identical joining performance to TPNet, with differences within 0.005 MRR across all five datasets. A simple GBDT using the same statistics also achieves comparable results, while removing member attention or temporal mixing has little effect on joining performance. These findings raise doubts about the added value of the proposed neural architecture.

  4. Unclear scalability and efficiency advantages. The method explicitly maintains statistics for previously interacting node pairs, with storage growing as more pairs are observed. Although the paper reports preprocessing and training times, it lacks a systematic comparison of memory consumption and computational costs against existing methods. The practical tradeoff between maintaining extensive pair statistics and improving prediction accuracy remains unclear.

Questions for authors
  1. What fundamentally new representation learning mechanism does the ISM introduce beyond manually constructed temporal statistics and standard neural readouts? How does it differ conceptually from applying existing predictors to carefully engineered features?

  2. To what extent do the sufficiency and expressivity results extend beyond the specifically defined process family and restricted baseline architectures? Can the authors clarify which conclusions remain valid for more general dynamic hypergraphs?

  3. Given the near parity with TPNet and GBDT, what evidence demonstrates the necessity of the proposed neural architecture? Can the authors further isolate its benefits from those provided by the underlying statistical features?

  4. How do memory consumption and update costs scale with the number of observed node pairs? Can the authors provide a comparison of accuracy, runtime, and memory usage against TPNet under comparable computational budgets?

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

This paper introduces a new formulation for modeling dynamic hypergraphs through incidence streams. It analyzes prototol leakage and proposes to use sufficient historical statistics for predicting hyperedge streams. It also introduces an incidence state machine which is further evaluated on joining, role, and departure prediction. The authors create and analyze a persistent-hypergraph benchmark alongside the proposed new analytic framework.

Critical strengths
  1. The idea of reformulating the existing hypergraph modeling frameworks is novel to me, though a lot of details need to be further motivated.
  2. The persistent-hypergraph benchmark and careful leakage audit are good analytical contributions.
Critical weaknesses
  1. I have primary concerns over this paper's readability. The wording and logic are extremely hard even for hypergraph experts to follow, and many terms lack explanation. I encourage the authors to motivate and define new concepts with more concrete examples, as well as polish figure 1 with more detailed explanations, e.g. what is "roles" in a hypergraph?
  2. I don't think the neural arch adds much over trees that use the same state. The joining performance is also close to TPNet as far as I can tell.
  3. I also feel that update schedules might affect comparisons. In your Appendix D.1, the TPNet comparison is reversed when the ISM state is stale. So I am not sure if the alignment provides a fully matched training & eval comarison.
Questions for authors
  1. Do the conclusions still hold after retraining both models with identical update schedules?
  2. Can you share more experiment in terms of how pair-state memory scale beyond the current datasets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.