acceptodds
Under review as a conference paper at ICLR 2027

GraphBet: Prospective FDR-Controlled Discovery from Temporal Graph Predictions

Abstract

Temporal graph models rank likely future interactions, but a high score does not establish that a nominated group attracts more subsequent activity than its available opportunities explain. We introduce GRAPHBET, which freezes a nominated destination set and compounds predictable likelihood ratios from later interactions relative to pre-event opportunity probabilities. Under the directional null that the set attracts no more probability than those opportunities permit, the resulting wealth is an e-process. We prove graph-wide validity for adaptively selected, dependent candidates with history-dependent null labels; e-value Benjamini–Hochberg (e-BH) controls the false discovery rate (FDR) at a common reporting time, and the accumulated union of earlier e-BH selections also controls FDR. We additionally establish safe aggregation of betting experts, robustness to bounded reference error, and finite-sample links from predictive accuracy and betting regret to detection. We observe no final-snapshot false discoveries in 300 controlled global-null streams or 200 exact-law benchmark replays. In a crossed study of three prespecified training fits and 20 shared held-out semi-synthetic Wiki replays, learned candidates raise mean conditional testing power from .335 to .418 under a common fixed bet, with positive evidence-growth gains in every fit. A prespecified six-expert mixture raises power to .452 and achieves the highest mean end-to-end recall among the evaluated pipelines, .084 versus .069 for degree–recency candidates with the fixed bet, with no final-snapshot false discoveries. Thus the opportunity law governs validity, candidate learning governs what is nominated, and betting calibration governs how efficiently evidence accumulates.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
temporal graphs, anytime-valid inference, e-processes, e-values, false discovery rate, multiple testing, graph neural networks

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_fe7pOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GraphBet, a statistical framework for identifying temporal graph node sets that receive significantly more future interactions than expected under an opportunity model. It combines temporal graph predictions with sequential betting and multiple hypothesis testing, providing theoretical guarantees for false discovery rate control under adaptive candidate selection and statistical dependence. Experiments on synthetic and real temporal graph data, including simulated interaction streams, benchmark replays, and observational analyses, report improvements in testing power and overall recall. However, the practical significance of the problem remains insufficiently motivated, while the reliance on restrictive assumptions, synthetic ground truth, and limited empirical gains raises concerns about the validity and generalizability of the conclusions.

Critical strengths
  1. Clear problem formulation. The paper separates temporal graph prediction from statistical hypothesis testing, providing a structured framework for evaluating whether future interactions exceed expected opportunities.

  2. Theoretical guarantees. The authors establish formal FDR control under adaptive candidate selection, statistical dependence, and evolving graph histories, which constitutes a meaningful theoretical contribution.

  3. Detailed theoretical and empirical analysis. The paper provides extensive derivations, sensitivity analyses, and component evaluations to examine statistical validity and evidence accumulation.

Critical weaknesses
  1. Limited significance and novelty. The practical relevance of the proposed discovery problem to temporal graph learning is insufficiently justified. The framework largely integrates established statistical techniques, with limited methodological innovation.

  2. Unreliable key assumptions. The theoretical guarantees critically depend on the opportunity probability model, whose validity is not established for real world data. The substantial sensitivity of the Wiki results further undermines the reliability of the claimed discoveries.

  3. Questionable ground truth and experimental validity. The ground truth is defined through artificially injected probability shifts rather than independently verified interaction patterns. Recovering these synthetic labels does not necessarily demonstrate meaningful group discovery.

  4. Limited empirical improvements. The proposed method achieves modest overall recall, and the improvement from learned candidates under fixed betting is not statistically conclusive. The experiments provide insufficient evidence of practical advantages.

Questions for authors
  1. What concrete applications motivate the proposed discovery problem, and what fundamental methodological advances distinguish GraphBet from existing sequential testing and FDR control techniques?

  2. How can the opportunity probability model be reliably justified in real world temporal graphs, especially given the substantial sensitivity of the Wiki results to reference probability adjustments?

  3. Why should recovering artificially boosted destinations be considered meaningful group discovery? Can the authors provide evidence that the identified groups correspond to meaningful interaction patterns in real data?

  4. Given the limited overall recall and inconclusive improvement from learned candidates under fixed betting, can the authors provide a direct statistical comparison of the strongest complete pipelines and clarify the practical significance of the reported gains?

Reviewer_uwQ8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

GraphBet tests whether destination sets nominated by temporal graph models attract more subsequent interactions than specified opportunity probabilities allow. It accumulates evidence using e-processes and applies e-BH to control FDR over registered candidates, including retained discoveries under the stated reporting protocol. Controlled streams, benchmark replays, and component studies examine validity, reference sensitivity, and the roles of candidate generation and betting.

Critical strengths
  1. The paper clearly formulates prospective enrichment and provides a coherent treatment of adaptive candidate selection, shared graph histories, and FDR control.
  2. The crossed candidate and betting experiments provide useful insight into why candidate quality and efficient evidence accumulation need to be evaluated separately.
Critical weaknesses
  1. Establishing a scientifically meaningful opportunity reference remains the main obstacle to practical use. Wiki discoveries nearly disappear after the diagnostic reference adjustment, while Review yields no discoveries. The paper therefore does not yet demonstrate useful discoveries in an application where exposure data, an assignment mechanism, or other independent evidence supports the reference.
  2. Evidence for the benefits of learned nomination and safe betting mixtures comes mainly from a single Wiki replay with strong injected enrichment. Learned candidates improve evidence growth under a common betting rule, but their lower nomination recall makes the overall benefit to discovery less clear. It remains unclear whether these gains persist under weaker enrichment or changes in the opportunity structure.
  3. Sequential monitoring and retaining discoveries over time are central to the framework, yet the experiments evaluate only a prespecified final snapshot. They therefore do not show whether these features lead to earlier detection or a more useful accumulated discovery set, leaving their practical value insufficiently demonstrated.
Questions for authors
  1. What useful discoveries emerge in an application where independent exposure or assignment information supports the opportunity reference, and how should those discoveries be interpreted?
  2. How do learned nomination and safe betting mixtures perform under weaker enrichment, changing opportunity patterns, and different graph structures?
  3. How do detection delay and discovery power compare between sequential monitoring with retained discoveries and reporting only at the final snapshot?
Reviewer_LvdeOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GraphBet, a framework for prospective discovery in temporal graphs with false discovery rate (FDR) control. The method first uses temporal graph predictors to identify candidate destination sets and then evaluates whether these sets receive more future interactions than expected under a reference opportunity distribution. It constructs sequential evidence using betting-based e-processes and applies e-value Benjamini–Hochberg (e-BH) to control FDR across adaptively selected candidates. The authors provide theoretical guarantees for sequential validity and FDR control, along with extensions for reference-model uncertainty and betting-strategy aggregation. Experiments on synthetic temporal graphs, real-world temporal interaction datasets, and semi-synthetic replays evaluate the proposed framework performance.

Critical strengths
  1. The paper provides a comprehensive theoretical analysis, establishing statistical guarantees for the proposed framework.
Critical weaknesses
  1. Insufficient motivation and unclear significance of the research problem. The paper studies whether a prospectively selected set of destination nodes receives more future interactions than expected under a reference opportunity distribution. However, the introduction does not sufficiently motivate why this particular problem is important or worth investigating. Although the authors distinguish their objective from conventional temporal graph prediction, it remains unclear what practical or scientific needs motivate this formulation, in which real-world scenarios such discoveries would be useful, and what meaningful insights they would provide beyond standard prediction tasks. The paper would benefit from a more compelling motivation and concrete application examples that establish the significance of the proposed research problem.

  2. Limited methodological novelty. The methodological novelty of this work appears limited. The proposed GraphBet framework mainly combines several well-established techniques, including temporal graph prediction, betting-based e-processes, e-BH for FDR control, and mixtures of betting strategies. Most of these components have been extensively studied in prior literature, and their combination does not appear to introduce a fundamentally new learning algorithm or statistical methodology. The paper appears to be primarily an integration and adaptation of existing tools to temporal graph data, rather than a substantial methodological.

  3. Limited baseline comparisons. The experimental evaluation lacks sufficiently comprehensive comparisons with existing methods. The main baselines are largely restricted to degree–recency heuristics and a TGN-style architectural ablation, while most other comparisons focus on different components or betting strategies within the proposed framework. In particular, the paper does not provide direct end-to-end comparisons with relevant existing approaches to temporal graph discovery or sequential statistical testing under a comparable problem formulation.

Questions for authors

See the weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.