acceptodds
Withdrawn from ICLR 2027

Inferring Higher-Order Interactions from Dynamics via Hodge-based message passing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Higher-order networks represent group interactions that shape collective dynamics. The joint influence of pairwise and higher-order interactions on node evolution makes inferring these structures from observed node states challenging. In this work, we propose SCHOD (Simplicial Complex Inference with Hodge-based Dynamic Modeling), a framework for jointly inferring pairwise and 2-simplex interactions without prior knowledge of the underlying dynamical equations. The Hodge -Laplacian captures lower- and higher-order relations among edges in a simplicial complex. SCHOD builds on this property by incorporating learnable structural probabilities into its lower and upper components, enabling edge-to-edge message passing for structure inference. To account for node-level variation in the contributions of pairwise and -simplex interactions, an adaptive gating mechanism modulates their aggregated messages for state prediction. Experiments on synthetic networks demonstrate the effectiveness of SCHOD in recovering higher-order interactions across different network types.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Network Structure Inference, Higher-Order Interactions, Simplicial Complexes, Hodge Laplacian

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7bWcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the inverse problem of recovering latent higher-order interaction structure from observed node states. Existing methods either focus only on pairwise graphs or require prior knowledge of the governing dynamical equations. The paper aims to infer both pairwise edges and 2-simplices from time-series data in a model-free way.

Critical strengths

The paper infers both pairwise and 2-simplex interactions. It performs well against two baseline methods on three types of synthetic networks.

Critical weaknesses

Although honestly recognized by the authors, SCHOD only work for very small networks because of the cost of enumerating candidate 2-simplices. Moreover, experiments are only made on synthetic networks. The results are promising, but the experimental setup is limited to fully convince me about the advantage of SCHOD. This pertains to the breadth of synthetic networks with varying parameters (including size), to the types of dynamics, and the baselines tested against.

Questions for authors
  1. Is the limitation for very small networks for all methods or only for SCHOD?
  2. What about comparing against the ARNI method?
  3. How were the parameters for the synthetic networks chosen, especially the density for ER, the m for BA, and the rewiring probability for WS. We would need more experiments to show that SCHOD still outperforms for different structures of the synthetic networks.
  4. What is the restriction about testing on real-world networks?
Reviewer_ujQ1OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a framework that considers nodes, edges, and triangles to learn the underlying higher-order interactions between nodes in a dynamic graph. The main contributions of this work are the construction of learnable and (pseudo-)differentiable probabilities for edges and triangles that are used to reconstruct both the state of the nodes over time as well as the higher-order interactions. The specific inductive bias imposed by the authors constrains the learned probabilities via a Hodge-based message passing that has desirable properties. Hence, the authors are able to recover the true 2-simplicial complexes (triangles) on synthetic graphs contrarily to the other baselines considered.

Critical strengths

In my view, the paper uses simplicial complexes in message passing in a way that is novel and interesting. Works in the past have generally extracted a priori such kind of information and then used it to augment node features, but the use here of the Hodge representation is clever especially given the stated goal of the paper, which is to discover the underlying higher-order interaction that generated the dynamic graph. The probabilities "naturally" align to something meaningful thanks to the upper and lower representations that are computed from first principles. This has determined my score on Informativeness.

The second strength, although with some improvements, is the presentation. I believe the paper does a really good job at introducing the readers to the concept of simplicial complexes and the Hodge representation. I learned something by reading this paper. This determines my score on Presentation.

Critical weaknesses

I like this paper, but I believe that the bar, in terms of experimental validation, should be higher for a conference like ICLR. The main weaknesses of the paper are:

  • no "real" dataset: as much as I am supportive of the use of synthetic data, a core ML paper about a new methodology should demonstrate, to some extent, how it affects some application domain. It would be nice to see, for example, if the approach can recover higher-order interactions from true epidemic data, or if there is an advantage on specific temporal tasks where we assume higher-order interactions are present (maybe traffic networks?).
  • model hyper-parameters have been fixed: this is a missed opportunity, because a proper model selection would have showed how different hyper-parameter choices (the lambdas, the embedding size, etc.) affect the performances on a validation set. Instead, the analysis the authors present is about the learning rate, which is mildly interesting given the specific contribution.
  • no model selection for the baselines: I could not find information about performing model selection on the baselines as well as the proposed approach (related to the previous point). When evaluating on new datasets, it is standard practice, for a fair evaluation, to perform model selection on all models tested.

This justifies my soundness score, not because there are huge mistakes but rather because the paper feels not mature enough at this stage, and it requires a bit more time polishing and enriching the evaluation both quantitatively and qualitatively.

Questions for authors
  • Could the authors please make more practical examples about how higher-order interactions shape collective dynamics in the paper?
  • Detaching the term of Eq 4 from the computational graph is a bit unusual. How did the authors arrived at this choice?
  • Eq 13: what's the rationale behind the inductive bias of summation?
  • line 291: how are the posterior distribution computed? Should p symbols be bold?
  • why fixing lambda_sparse and not putting it into model selection? 0.4 may not be the best value for all datasets
  • if I understand correctly, the two-stage training is just a "trick" to reduce the edges before computing triangles? This is the heart of the authors' method and it feels a bit weak as a process
  • I could not understand how many trajectories per graph the authors generate, could the authors please clarify?
  • Sec 5.1: could the authors please explain a bit better where the higher-order dependencies emerge from?

Some comments:

  • line 50: the authors cannot imply they propose a general structure for inferring relationships, as the authors have to restrict to a two stage procedures and consider at most triangles
  • more intuition about equation 4 would be appreciated
  • Figure 1 can be improved by somehow plugging the probabilistic framework
  • line 223: relation type is confusing, please keep the language consistent across the text
Reviewer_Gw3HOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a new method to infer pairwise and triadic interactions from node dynamics. Input data takes the form of a (sample from a) trajectory in some Euclidean space Rd\mathbb{R}^d for each node: xit∈Rd:1≤i≤n,1≤t≤T\\{x_i^t \in \mathbb{R}^d: 1 \leq i \leq n, \\ 1 \leq t \leq T\\}. The goal is to infer a simplicial complex from these data. For that the paper proposes a method based on:

  • Finding the probability of appearance of edges and triangles as a learnable logit vector with 2 dimensions.

  • Using the learnt probabilities, construct a message for each pair of nodes.

  • Use a modified version of the incidence matrices (which turn out to be the incidence matrices of the weighted simplicial complexes, with weights given by the probabilities, but this is not said explicitly in the paper) to define the Hodge Laplacian over signals on edges.

  • Apply the up and down Laplacian separately to the signal given by the edge messages, to obtain two new up/down edge signals

  • Transform the up/down edge signals into two up/down node signals, by summing the corresponding up/down signals from incident edges.

  • Apply a gating mechanism to both up and down node signals.

  • Update the initial node signal with two MLPs applied to: the initial signal; the up/down gated value and the up/down node signal.

The authors show the empirical behavior of such a method in comparison to THIS, an EM-based method and an extension of NRI to infer pairwise and triadic interactions. The authors also present some ablations on the sensitivity of the method to the number of samples available and the underlying number of nodes, and to some hyperparameters of the method.

Critical strengths

I think the paper notices an existing gap in the literature in inferring higher order dynamics from dynamics, namely, that existing methods often rely on a priori knowledge about the network or the dynamical system, and when the don't, they require strong assumptions on the sampling that gives rise to the input data to successfully retrieve the higher-order interactions.

Critical weaknesses

However, after carefully reading the paper, I do not understand why Hodge theory is the answer to this limitation, or why it is the tool that helps overcome this limitation. In my view, the paper proposes a method that seems to work, but which is full of unjustified choices that remain unexplained. To name a few examples: why the logit for learning the probabilities needs to be bidimensional and not one dimensional? Why that choice of message passing to build the initial edge signals? Why do you need to modify the Hodge Laplacian introducing the learnt probabilities for the method to work? Why is the gated mechanism necessary? In the end, due to all these open questions, I am not sure that the improvements in performance reported are coming from the use of Hodge theory or from any other component of the architecture.

In sum, I feel the paper does not answer the following question: why is this architecture a suitable and desirable approach to infer higher-order interactions from dynamics?

Moreover, the paper entirely neglects the entire topological deep learning over simplicial complexes literature. In 2026, with AI assistants, I don't think it is acceptable not to exhaustively (or, at least, sufficiently) frame a work in the broader literature. Just to provide the authors with one possible resource among many possible ones: https://arxiv.org/pdf/2407.09468

In addition, and a reason in itself that motivates my decision to reject: there is no accompanying code for the proposed method, so there is no way for me to reproduce the experiments reported in the paper. The description of the implementation is not sufficient to fully reproduce the protocol and make sure that the results are comparable to those reported in the paper. There is no direct comparison of the parameter budget and computational cost of the methods contrasted in the results section, which I believe is also an important axis for the comparison.

Questions for authors
  • What are the Hodge Laplacians providing that previous methods lack? I need to see a clear reason why this is the answer to the limitations of previous methods.

  • Why do you need all the architectural choices (see examples above)? Are they all necessary?

  • What are the parameter budgets of the different methods compared in the result section? What are their runtimes and computational budgets?

  • Why do THIS and EM-based methods obtain exactly the same results in different runs?

  • Are there no assumptions whatsoever on the input data for your method to work? In case there are some assumptions, how do they compare to the ones from THIS and the EM-based methods?

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

Related papers

Discussion (0)

Sign in to comment.