acceptodds
Under review as a conference paper at ICLR 2027

Generative–Discriminative Source Localization from Noisy Diffusion Observations

Abstract

Graph source localization identifies the initial spreaders from node states observed after diffusion. It is an important task in epidemic tracing and information diffusion analysis. One way to assess a source estimate is to compare its predicted diffusion with the observation and refine the estimate by reducing their difference. Under noisy observations, however, we find that a closer fit can leave fewer true sources in the final predicted set. Better diffusion fit therefore does not necessarily improve localization, motivating refinement that directly targets source identification. We propose Generative–Discriminative Source Localization (GDSL). The method first produces an initial source estimate from the observation and network structure. It then predicts the resulting diffusion and computes node-level differences from the observation. A discriminative network combines these differences with the initial source scores and network structure, and learns to adjust node scores using source labels. Diffusion differences thus inform node decisions, while source supervision guides the correction. On SI and SIR simulations over four real-world networks, GDSL achieves the highest Top-K F1 among the evaluated implementations in all eight settings, with mean AUC, AP, and Top-K F1 of 0.963, 0.728, and 0.683. Stage analysis shows that most localization gains arise from discriminative correction, supporting source-label supervision alongside diffusion prediction.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Source Localization, Graph Diffusion, Graph Neural Networks, Inverse Problems, Generative–Discriminative Learning

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

The paper studies source localization from noisy diffusion observations. It proposes GDSL, which generates an initial source estimation, predicts spread, and uses prediction-observation diffs in a source-supervised correction network, followed by an refinement. Experiments use SI and SIR simulations on four networks and examine both localization performance and the relationship between diffusion fit and source recovery.

Critical strengths

The paper asks a good question: does fitting the observed diffusion better actually help recover its sources? The per-sample analysis makes this mismatch concrete and supports evaluating the source set directly. Moreover, the reported localization gains are consistent across several settings, the stage analysis is also helpful, including the observation that later refinement often adds little beyond discrimination correction.

Critical weaknesses

Though, I did not yet see a clear enough contribution beyond adding supervised source correction to a generative pipeline. The motivation is reasonable, but the gap between diffusion fit and localization does not by iteself explain why this design is needed/desired. Learned source correction is already part of IVGD (a famous WWW '22 work). The paper needs a clearer account of what its diffusion-discrepancy machanism adds, especially when the sources' labels are already available for training.

There are some concerns in experiments. Removing the discriminative stage -- also removes source-label supervision, so the large drop does not isolate the value of the proposed module.

Questions for authors

You may answer (i) what does the diffusion-discrepancy path add when the observation encoder and source-label supervision are held fixed? (ii) what distinguishes GDSL from existing learned source-correction methods?

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

This paper studies graph source localization under noisy diffusion observations. The authors identifies a mismatch between diffusion fit and source localization under noisy observations and proposes GDSL, a generative–discriminative framework that combines initial source estimation, propagation prediction, and discriminative source-score correction, followed by inference-time refinement. Experiments on four real-world network topologies under SI and SIR diffusion demonstrate improvements over the evaluated baseline implementations. Ablation further suggest that discriminative correction accounts for most of the performance gains.

Critical strengths
  1. The paper provides an informative empirical observation that improving the fit between predicted diffusion outcomes and noisy observations does not necessarily recover more true sources. This observation is meaningful for methods that refine source estimates through forward-diffusion consistency.
Critical weaknesses
  1. The comparison with existing methods is limited. In particular, PGSL† only evaluates the inverse-prediction branch of a local PGSL implementation, rather than the complete inference procedure. Moreover, several recent source localization methods discussed in the paper are not included in the experimental comparison.
  2. The evaluation of robustness to observation noise is insufficient. Although the method is developed for noisy diffusion observations, the main experiments are conducted under a fixed noise setting.
  3. In the reasoning stage, the true source count is used as a constraint. However, in most cases, the number of sources is variable or even unknown, but the author only conducted experiments using a fixed number of sources.
Questions for authors
  1. Appendix C.3 states that PGSL† uses only the inverse-prediction branch of a local PGSL implementation. Why was the complete PGSL inference procedure not evaluated?
  2. The discriminator is a central component of the proposed method, but its architecture is not described in sufficient detail. Could the authors clarify the network structure?
  3. Table 4 shows that fitting-based refinement provides only small gains in two settings and decreases F1 in five. What is the justification for retaining this stage in the final pipeline?
  4. Equation (12) introduces a source-count constraint based on the predefined source ratio, while the experimental setting assumes that the true source count (K) is known. How sensitive is the method to an unknown source-count estimate?
Reviewer_V3f4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

Graph source localization identifies the initial spreaders from node states observed after diffusion. They try to show how propagation dynamics and network structure can inform source localization. This paper examines how this information can refine an existing source estimate, and whether improved diffusion prediction translates into a better final source set. This paper introduces GDSL (generative–discriminative source localization), a novel method designed to resolve the disconnect between diffusion-fit error and accurate source localization in network diffusion processes. While conventional approaches rely purely on minimizing diffusion-fit error, which can often degrade true source recovery, GDSL uses a discriminative network guided by direct source-label supervision. By feeding prediction-observation differences back into node-level score corrections across successive inference stages, the method successfully bridges macroscopic diffusion prediction with microscopic source identification, as demonstrated through SI and SIR simulations on real-world networks.

Critical strengths
  1. The approach is interesting.
Critical weaknesses
  1. Presentation of the paper.
  2. The authors did not study the computational complexity of their algorithm.
  3. The authors did not consider the scalability of their algorithm.
Questions for authors
  1. The section defines Task and states, “At inference, the model uses the graph, specified diffusion parameters, and noisy observations,” but it does not define what the output of the model is. Specify it.
  2. Is SS the set of the actual nodes for which s=1s=1?
  3. What are did_i and di2d_i^2 in equation (1)? They are not defined.
  4. What is the dimensionality of the decoder output?
  5. What do the samples taken from the decoder output distribution, given in equation (3), show?
  6. How did you obtain the constant terms and form of equation (4)? what are the dimensionality of Pˉy~\bar{P}\tilde{y} and b?
  7. How do you train the network: end-to-end or part-by-part?
  8. Is sampling at the input of the generator differentiable?
  9. Why is the node correction bounded by 4 tanh?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.