acceptodds
Desk rejected at ICLR 2027

Gated Graph Neural Networks for Learning Hidden Independent Cascade Dynamics

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Information and infectious diseases spread through social networks, but the spreading probabilities driving them are hard to estimate without per-node activation times. Applications seldom supply these, and inference must instead proceed through indirect and noisy proxies for the terminal infection states. We study this inverse problem on a fixed, known graph under the Hidden Independent Cascade (HIC) model, with one spreading probability per node rather than a single global rate, so the number of unknowns scales with the number of nodes. Seed sets and observation parameters are known, while activation times and terminal infection states are latent, and the observed-data likelihood requires marginalizing over every spreading outcome. We propose a simulation-based amortized estimator that recovers the full node-level parameter vector without reconstructing individual latent cascades. Repeated seed-conditioned symptom observations are summarized as Symptom-Aware Cascade Features (SACF), which combine empirical symptom statistics with neighborhood and structural information. SACF are mapped to parameters by SAGE-HC, a permutation-equivariant gated graph neural network whose learned gates attenuate neighborhood messages corrupted by false positives and false negatives, and training on simulated HIC realizations yields a reusable inverse map. As a benchmark under the same hidden observations, we extend the Dynamic Message Passing learning framework to the HIC emission model. On synthetic and empirical graphs the two methods separate by topology. DMP is highly accurate on trees, whereas SAGE-HC is substantially better on heterogeneous, loopy graphs under noisy terminal symptoms.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Hidden Information Cascades, Network Diffusion, Likelihood free Inference, Latent State Inference

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_o42AOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

NA

Critical strengths

This paper should be desk rejected as the funding information is leaked in the manuscript.

Critical weaknesses

This paper should be desk rejected as the funding information is leaked in the manuscript.

Questions for authors

NA

Reviewer_j13rOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript studies how to infer node-level spreading probabilities in the Hidden Independent Cascade (HIC) model using only known sets and noisy final observations, without knowing the actual spreading process or final infection states. The authors propose SACF to extract features from the observations and SAGE-HC, a gated graph neural network trained on simulated cascades to estimate the parameters. The experiments show that SAGE-HC works better on heterogeneous graphs under noisy observations.

Critical strengths

The heterogeneous HIC inverse problem with noisy terminal-only observations is realistic and genuinely hard (likelihood requires marginalizing over all spreading outcomes).

Critical weaknesses

DMP is accurate on trees and degrades on loopy graphs is textbook knowledge in the message-passing literature (e.g., Karrer & Newman, 2010, which the paper itself cites). It is also reasonable to expect GNNs to benefit from information from neighboring nodes. So I think the main experimental findings are largely expected from previous work. I am not sure what new insights these results provide beyond what we already know.

I find the non-monotonic results more interesting, but they are not well explained. For example, BA(2) performs better than BA(4), even though BA(4) has more connections. I also noticed that in Fig. 6, adding more edges eventually reduces accuracy. This seems inconsistent with the manuscript explanation that more structural redundancy helps the model deal with observation noise. If more connections are helpful, why does accuracy eventually decrease even without noise?

Questions for authors
  1. Can you compare graphs with different numbers of loops but similar densities (or vice versa)? If without this experiment, how can we tell whether the non-monotonic trend comes from loops or simply from graph density?

  2. In Fig. 6, why does accuracy eventually decrease as more edges are added, even without noise? How does this fit with the claim that structural redundancy helps the model handle noisy observations?

  3. Can you show an ablation of SACF components and compare the gated model with a standard GNN without gating?

  4. Does SAGE-HC need to be trained separately for each graph? If so, how much does simulation and training cost for a new graph compared with DMP inference? Can the trained model be reused on other graphs?

  5. What happens if Θ_obs is not known at test time?

Reviewer_q3jxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In the hidden independent cascade model, the authors consider the inverse problem of learning the weights of the edges from noisy information about the cascades, whether we only get the terminal state of the nodes (noise-free), or whether we get a noisy version of it. They introduce a set of features and use a gated GNN to learn the parameters.

Note that, contrary to the usual inverse problem:

  • the entire graph is known
  • the weights of edges are all the same when they start from the same node
  • the noise distribution is entirely known

This makes it a particularly easy version of an inverse problem.

Critical strengths
  • Their method seems to work in the specific setting mentioned
  • Their noisy setting is interesting
Critical weaknesses
  • the inverse problem is much easier than it usually is
  • in particular, the estimator for a tree is straightforward and should be included as a baseline
  • in the noise-free setting the likelihood of having a cascade of size exactly one when the seed is u is (1−pu)du(1-p_u)^{d_u}, which is a decent estimator when the degree is not too big (the average degree of all the graphs in their evaluation is below 10 and often below 5), so with 50 cascades for each seed this is an needed baseline, especially since their precision is 0.1. In their noisy setting, the probability of having a cascade of size 1 and all the neighbors showing -1 (which is informative and tells us the node was not infected) is (1−pu)du⋅η−du(1-p_u)^{d_u}\cdot \eta_-^{d_u}, which may be too small, but is still an exact estimator of the edge weight and should be included as baseline.
  • 50 cascades starting from each node is an unrealistic setting for any real applications
  • there is no information about the distribution of the pup_u, which impacts the number of cascades needed to learn the parameters
  • in general, the baselines are severely lacking. There is no ablations on the features to see if they're all useful. There are no other models using the features to show that using the gated GNN is needed. There is no gated GNN without using the features to show they are needed.

In summary, I believe the setting lacks sufficient complexity, the method is not justified compared to simpler versions, and the baselines are not strong enough as comparisons.

Missed SOTA:

  1. Learning the Graph of Epidemic Cascades: this is the seminal paper for this inverse problem
  2. Learning graphs from noisy epidemic cascades: The entire noise-free setting is tackled and solved (for an unknown graph and more weights freedom) in the equivalent of their noise-free setting. For any graph, including loopy ones.
Questions for authors

I'll be happy to change my mind if you show me I've misunderstood something critical.

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

Related papers

Discussion (0)

Sign in to comment.