acceptodds
Withdrawn from ICLR 2027

Stochastic Message Passing for Epidemics Modeling on Social Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The behavior of stochastic social phenomena such as epidemics or crowd movements is often unknown and rarely easy to model. This work presents a framework for stochastic graph neural networks (SGNN) training designed to learn complex stochastic behavior from graph data. Unlike standard graph neural networks, the proposed architecture introduces stochastic message propagation through a continuous relaxation of Bernoulli transmission events. Repeated rollouts from the same graph initial condition generate multiple possible trajectories, from which node-level marginal state distributions are estimated. The model evolves recurrently over time where at each time step, stochastic messages are propagated along eligible edges, aggregated at the receiving nodes, and combined with the current node states through a multilayer perceptron. We evaluate the model on synthetic graphs and real social networks using the susceptible-infected (SI) epidemic process as a case study. The evaluation examines both the model's ability to reproduce the temporal evolution of node states and its ability to recover the underlying transmission probability. We further analyze the sensitivity of parameter recovery and predictive performance of two factors: the number of model simulations and the initial value of the learnable transmission probability. This work represents the first step towards graph neural networks that can act as stochastic simulations of graph-based phenomena.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Stochastic message passing, Epidemic modeling, Social networks

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_qPSSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces SGNN, a stochastic graph neural net for learning discrete processes such as epidemic spreading on graphs. They focus on the SI (susceptible/infected) model. The key idea is to use the reparametrization trick and Gumbel-softmax to get differentiable probability for the S/I states. The rest of the model is similar to MPNN. They use the model recurrently to simulate epidemic spreading trajectories. They want to learn the parameters of the spreading process, including the probability of infection via each edge. Their loss is basically the mean difference ∣pˉu(t)(c;ω)−pu(t)(c)∣|\bar{p}_u^{(t)}(c;\omega) - p_u^{(t)}(c)| over time and states S/I between the actual marginal probability pˉu(t)(c)\bar{p}_u^{(t)}(c) of node uu being in state c∈{S,I}c\in \{S,I\} at time tt, and the average of the simulation probabilities pˉu(t)(c;ω)\bar{p}_u^{(t)}(c;\omega) with model parameters ω\omega. They test their model on simulated SI processes on a few synthetic and real graphs.

Critical strengths
  1. Using Gumbel-softmax for discrete processes on graphs makes sense.
  2. Their method is a reasonable first step, but with important shortcomings (see below).
  3. The small SI experiments are informative.
Critical weaknesses
  1. No comparison with any baseline methods. I'd expect even GNN or MPNN to perform reasonably well SI and get the average probabilities right.
  2. SI is a very basic spreading model. In particular, since S(t)+I(t)=NS(t)+I(t)=N, it has basically one variable. The mean-field version (not on graphs) is fully integrable and becomes logistic growth. The graph version is a tad more involved, but still very simple. More complex processes like SIR, or SEIR would be the real challenge.
  3. I expect the average over simulations pˉu(t)\bar{p}_u^{(t)} to have a critical flaw: its value depends heavily on patient zero location in modular graphs. You can see this the Sandia and coauthorship datasets. I believe your very high error (>10%) is because these real graphs have clusters and are modular, unlike the ER, BA and WS synthetic graphs. The clusters introduce bottlenecks in the spreading process. Therefore, which cluster patient zero belongs to becomes critical in determining p(t)p^{(t)} at finite times. This translates to large variability among simulations and the mean pˉ\bar{p} becomes uninformative.
  4. Key related work on GNN models for discrete processes on graphs is not discussed: Neural Relational Inference (Kipf et al., 2018), GTS ("DISCRETE GRAPH STRUCTURE LEARNING FOR FORE-CASTING MULTIPLE TIME SERIES" Shang et al., 2021), NodeFormer (Wu et al., 2022)...
Questions for authors
  1. How does your method compare to other methods I mentioned above?
  2. Did you run vanilla MPNN to see if it works on this data?
Reviewer_ELvyOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a recurrent MPNN with edge-level stochastic gates. Each gate is a straight-through Gumbel-Softmax Bernoulli with learnable α=σ(a). Edge eligibility is p_v(I)·p_u(S), and an MLP updates node states with a softmax. Several rollouts are averaged into node marginals, which are fitted with an L1 loss to empirical SI marginals from 10³ simulations (shown to equal the TV distance). Experiments use three synthetic graphs (200–300 nodes) and two real ones (86 and 379 nodes), with α∈{0.2, 0.6} and sensitivity to R_m and α₀.

Critical strengths
  1. The goal is interesting: GNNs as learnable stochastic simulators with an interpretable mechanistic parameter.
  2. The construction is simple and clear, and α can be read directly from the model.
  3. The sensitivity analysis honestly reports where the method fails.
Critical weaknesses
  1. Neither parameter recovery nor the stochastic-simulator claim is well supported. The transmission probability and the MLP update are fitted jointly to marginals, so the MLP can absorb part of the transmission effect, which is consistent with the strong dependence on initialization: with α₀ = 0.9 and α = 0.2 the error on the synthetic graphs is about 0.7. Node states are propagated as soft probabilities rather than sampled states, so a rollout is not a valid SI realization, and the evaluation never goes beyond per-node marginals. In Figure 2, the deterministic model even matches the empirical marginals more closely than the stochastic one.
  2. No baselines are compared. Apart from a qualitative comparison with a deterministic model, both the predicted marginals and the recovered α are only compared with ground truth. Fitting α directly through the dynamic message-passing equations discussed in Section 2 would provide a natural reference for both tasks.
  3. The scope and experimental rigor are limited. The study covers only the SI model on small static graphs of 86 to 379 nodes, with a single initially infected node and two values of α. The train and test split is not defined, and results vary strongly across runs, for example 0.168 ± 0.315 on Erdős–Rényi with α = 0.2 and 100 simulations. Proposition 1, which states that the L1 loss equals the total variation distance for two-class distributions, is a direct identity rather than a substantive result.
Questions for authors
  1. How are the training and test sets defined?
  2. How does α recovery compare with fitting α through the dynamic message-passing equations?
Reviewer_xYiDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  1. This paper proposes a stochastic graph neural network for modeling epidemic dynamics on static social networks.
  2. The method introduces learnable stochastic transmission gates into message passing, uses repeated stochastic rollouts to estimate node-level state marginals, and jointly optimizes the transmission probability and neural state-update parameters.
  3. Experiments on synthetic and real-world graph topologies investigate predictive performance and sensitivity to simulation count and parameter initialization.
Critical strengths
  1. The paper studies a meaningful problem of learning stochastic dynamics on graphs, with potential applications beyond epidemic modeling.

  2. The proposed stochastic transmission mechanism is interpretable and naturally motivated by the underlying epidemic process.

Critical weaknesses

1. Insufficient discussion of closely related work and unclear methodological novelty.

The paper overlooks several closely related studies [1–3] on stochastic spreading processes. The authors should better clarify the methodological novelty of the proposed approach in relation to these existing works.

2. Lack of comparisons with relevant baselines.

The experimental evaluation lacks comparisons with existing methods for stochastic epidemic modeling. Without such comparisons, it is difficult to assess the effectiveness and empirical advantages of the proposed method.

[1] Neural Enhanced Dynamic Message Passing. AISTATS 2022.

[2] Reconstructing Parameters of Spreading Models from Partial Observations. NeurIPS 2016.

[3] Learning of networked spreading models from noisy and incomplete data. 2023

Questions for authors

Please See Weaknesses.

Reviewer_9ANjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a Stochastic Graph Neural Network (SGNN) to simulate disease spread over social networks. It embeds Bernoulli‑based stochastic propagation into graph message‑passing, using differentiable sampling to jointly learn the global transmission probability and node‑state update function. Experiments on three synthetic and two real‑world networks under the SI spreading setting assess infection prediction, transmission‑parameter estimation, and the effects of simulation numbers and parameter initialization.

Critical strengths

1.The paper introduces a stochastic message-passing mechanism into GNNs using differentiable sampling. It jointly learns the transmission probability and node-state update function, providing an interesting approach to modeling stochastic spreading processes.

Critical weaknesses

1.The necessity of SGNN is insufficiently demonstrated. Only the SI model with one global transmission probability is studied. A simpler baseline estimating this parameter with a standard SI simulator may achieve similar performance, but it is not compared.

2.Matching only node‑level infection probabilities may not guarantee the model learns the full stochastic spreading process. Moreover, deterministic GNN works better in some cases, and transmission‑parameter recovery is sensitive to initialization. Hence, the merits of stochastic message‑passing are not well validated.

Questions for authors

1.Since the paper only considers the SI model with a single global transmission probability, could the authors compare SGNN with a simpler approach that directly estimates this parameter and then uses a standard SI simulator?

2.The training objective only matches node-level infection probabilities. However, does matching these probabilities also ensure that SGNN can capture the overall spreading patterns and their variations?

3.In Section 5.1.1, the authors state that α = 0.1 is fixed, but Tables 1 and 2 report results for α = 0.2 and 0.6. Please clarify the experimental settings. Also, could they explain the large infection probability estimation error on the Coauthorship network when α = 0.6 and α₀ = 0.1 (Table 4)?

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

Related papers

Discussion (0)

Sign in to comment.