acceptodds
Withdrawn from ICLR 2027

Topology-Conditioned Feasible-Set Generation for Continuous-Time Multigraphs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generating temporal subgraphs requires changing interaction structure while preserving observations and quantitative constraints. Structural edits alter relay equations and can make pinned observations incompatible with positive flow. We formulate this as topology conditioned feasible-set generation. Each structural proposal induces its own constraint system, and the minimum distortion required to complete it shapes which structure is selected. A typed edit grammar and a policy trained to reconstruct corrupted skeletons propose and score admissible changes while protecting pinned events. A conditional diffusion prior then generates quantities and event times for each candidate topology. Each candidate is completed by Dirichlet-Boundary-Anchored KL projection (DBA-KL), which finds the closest positive flow that preserves pinned quantities and satisfies relay conservation. Under strict feasibility this completion is unique, and its KL distortion separates into the minimum change the constraints require plus any extra distortion from another feasible completion. All completed outputs preserve positive quantities and pinned observations, and every output containing a relay satisfies the relay-balance tolerance. On three anti-money-laundering benchmarks from a single simulator family, our DBA-KL configurations attain higher four-classifier average precision than the ECI-DBA and PCFM-log adaptations on all three. Relative gains reach 23.5% over ECI-DBA and 12.7% over PCFM-log on LI-Medium, and the terminal variant exceeds real-data training on HI-Small. Downstream accuracy alone is nevertheless a poor guide to constrained generation: an ablation that removes flow completion scores higher on every dataset while violating relay balance on 610 of 610 audited relay-containing outputs. Predictive utility and quantitative validity therefore have to be measured separately.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Temporal graph generation, Constrained diffusion models, Boundary conditioning, Kullback–Leibler projection, Flow conservation

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

The paper studies how to generate variations of temporal interaction graphs while preserving observed events and satisfying flow and timing constraints. The central challenge is that changing the structure also changes which quantities and times are feasible. A learned policy proposes structures, and diffusion generates their attributes. DBA-KL finds the closest positive flow that respects fixed quantities and balance at intermediate nodes. Final selection considers both policy preference and the correction each candidate requires. Experiments on three IBM-AML subsets show a tradeoff between predictive performance and validity. Removing flow completion improves downstream AP, but every audited output containing a relay violates balance.

Critical strengths
  1. The idea is reasonable. Changing a graph’s structure also changes the constraints its attributes must satisfy. It makes sense to evaluate each candidate together with the corrections needed to make it feasible.
  2. Both the theory and experiments support the motivation. The theory establishes a unique optimal flow completion and convergence under the stated assumptions. The no-flow experiment provides complementary evidence: removing completion improves downstream AP, but all outputs containing relays violate balance.
Critical weaknesses
  1. The figures look unfinished, and the paper reads more like a draft. Figure 1 contains many small labels, equations, and tightly packed boxes, making the main workflow difficult to follow. The spacing and use of lines could also be cleaner. The results tables include implementation-style labels and paths, such as rq2_no_flow and results/rq3/audit/exports. Together, these details give the presentation a draft-like feel.
  2. The experimental results are hard to interpret. Understanding the comparisons requires frequent cross-referencing between the main tables and appendix. For example, Table 1 reports two selected detectors, while the four-detector averages appear in Table 5. Table 3 uses different denominators for edited outputs and valid new structures. Some outputs also pass the flow-balance checks but fail another validity check. Better results presentation would help.
  3. The experiments may not yet be sufficient to support the full conclusions. The results support the need for feasibility checks, but the benefit of the proposed structural selection mechanism remains less clear. The no-flow ablation removes correction, feasibility filtering, and cost weighting together. Separate cost-weighting and uniform-policy controls, with matched sources and positive counts, would help clarify each component’s contribution. Further analysis could also explain the increased proportion of relay-free outputs, which incur zero flow-completion cost, and the remaining vocabulary violations.
Questions for authors

I don't have any questions at this time, but I suggest a major revision of this draft to bring it up to ICLR standards.

Reviewer_E4WhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper is about generating temporal graphs to maintain the generated qualities consistent with the constraints. The proposed model uses structural edit policies to generate candidate graphs and then use a conditional diffusion models to generate the transactions for the temporal relationships. The final graph is selected from the structural preference and the completion cost. The paper includes experiments on three IBM-AML datasets that show the better performance compared to adapted constrained-generation baselines.

Critical strengths

The paper addresses an interesting problem of changing a temporal graph's structure also changes the constraints that its quantities must satisfy. The proposed framework considers structural changes, fixed observations, and quantitative feasibility, such that new graph structures to be generated while specified observations can be maintained.

The theoretical analysis shows the completion step from a unique positive solution and convergence under strict feasibility. The experiments evaluate predictive utility and quantitative validity separately.

Critical weaknesses

Figure 1 includes too many components and implementation details, and this makes the overall framework difficult to follow. The small text and dense layout cannot show the main steps and their connections. I would recommend simply simplifying the figure and moving secondary details to the caption or appendix would improve readability. If you use AI to generate such diagrams, which I believe so, you need to double check and polish the figure carefully. The figure need to show the main architecture and story but not all implementation details.

The theoretical contribution is relatively incremental. The existence, uniqueness, and distortion decomposition mainly follow from classical convex KL projection properties. These results support the completion step, but do not explain whether selecting a topology based on its minimum completion cost produces more realistic or useful graphs. Meanwhile, the assumption Az=bAz=b requires equal incoming and outgoing quantities at every modeled relay, excluding losses and net accumulation. Since the paper reports that 88--97% of original records violate this condition, they for sure needstronger justification.

The empirical evaluation is limited to LI-Small, HI-Small, and LI-Medium, which are subsets of the same IBM-AML simulator family rather than independent application domains. Therefore, it is unclear whether the results generalize to other transaction mechanisms or datasets where the conservation assumption is less suitable.

Questions for authors
  1. How to justify the zero-net-accumulation assumption when 88--97% of original records violate it?

  2. How to distinguish useful topology selection from a preference for relay-free graphs with zero completion cost?

Reviewer_5kMTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies structural scenario generation for continuous-time transaction multigraphs. The task is to vary the connectivity, amounts and times of a source subgraph while keeping pinned events fixed and balancing flow at every relay account. Because each structural edit changes the balance equations, the paper poses generation as selection among topology-dependent feasible sets. The pipeline works as follows. It enumerates candidate topologies with a typed edit grammar and scores them with a learned policy. It proposes amounts and times with a conditional diffusion prior. It projects the amounts onto the balance constraints with DBA-KL, an iterative generalized-KL solver with closed-form row updates whose minimum distortion serves as a selection cost. It then completes the times and selects one graph. On three IBM AML subsets, with generated positives replacing real training positives, the paper reports a higher four-GNN average AP than adaptations of ECI and PCFM (up to +23.5% over ECI-DBA and +12.7% over PCFM-log on LI-Medium). It also reports relay balance in every relay-containing output. An ablation without completion obtains a higher AP but violates relay balance in all 610 audited relay-containing outputs. The paper takes this as evidence that utility and validity must be measured separately.

Critical strengths
  1. A well-posed problem with a principled and correct completion step. The paper observes that a structural edit changes the constraint system itself, so feasibility and the cost of reaching it must be computed for each candidate topology. For every candidate that admits a positive feasible completion, the proposed KL completion gives a unique, positive, relay-balanced solution and a cost that can be compared across candidates. This is the main merit I weigh in the paper's favour: the formulation is clear and the method delivers the guarantees it claims.

  2. Evidence that downstream utility does not certify validity. Removing flow completion raises the four-GNN average AP on every dataset, while independent audits show that all 610 audited relay-containing outputs violate relay balance. This is useful for how constrained graph generators are evaluated.

Critical weaknesses
  1. Weak statistical support for the main empirical claim. The paper claims that DBA-KL gives better downstream utility than alternative completions, but its main configuration (terminal Learned DBA-KL) exceeds the best adaptation by only 0.004, 0.010 and 0.005 in four-GNN average AP (computed from Table 5), no more than one detector-seed SD. The comparison uses a single generator fit, no paired test, synthetic-positive counts that are not equalized across arms, and a single simulator family. On the evidence provided, the gains cannot be separated from noise or from these differences in setup. This is the main limitation of the paper, and the answers to Questions 1-3 will determine whether my positive assessment holds.

  2. Limited novelty beyond the formulation. The completion solver applies Bregman's row-action method for the KL divergence, as used in classical matrix and network balancing (Lamond & Stewart, 1981; Censor & Zenios, 1997), with a recent ML precedent for entropic flow-balance constraints (Shi et al., ICLR 2025); none of these works is cited. The new element, selecting structures by completion cost, is not isolated by any control. The paper notes that this cost can favour relay-free candidates, about half of terminal Learned's emitted graphs, on which the completion has no effect. The fixed-topology arm, which makes no structural edits, performs comparably to the topology-editing variants. This matters because the paper does not show that its method adds value beyond the formulation.

Questions for authors
  1. In Sec. 4.2, the authors state that terminal Learned exceeds every ECI-DBA and PCFM-log configuration on every detector seed, and all arms share detector seeds 0-4 (0-3 on LI-Medium; App. D.1). Please report the seedwise differences in four-GNN average AP for terminal Learned vs. PCFM-log and vs. ECI-DBA (1 round). The mean differences are within one detector-seed SD (Table 5), so these paired values determine whether the main claim holds.

  2. App. D.2 states that the number of synthetic positives can differ across arms and is not equalized, and the BCE weight depends on these counts (Table 4). Please report the number of synthetic positives per arm and dataset. If the counts differ, the AP ordering may reflect training-set size rather than the completion method.

  3. The Real, Learned, Calibrated and fixed-topology series are a separately saved population, while the comparison series are filtered to generation seed 0 and one fitted-prior identifier (App. D.1). Do both use the same fitted prior, candidate bank and generation draw?

  4. Sec. 4.1 permits later-in-window context. Can context events that condition the generation of training positives fall inside the test period of the chronological split?

  5. The values of T and lambda_tau used in the AML experiments are not reported. Please give them and the split on which they were chosen. Please also report the share of relay-free candidates in the menus and in the selected outputs, which shows whether the completion cost shifts selection toward relay-free graphs.

  6. The completion solver is an instance of Bregman's row-action method for the KL divergence (Bregman, 1967, which the paper cites): each relay update is the KL projection onto one balance equation, the same scheme that underlies classical matrix and network balancing (Lamond & Stewart, 1981; Censor & Zenios, 1997) and recent entropic flow-balance solvers (Shi et al., ICLR 2025). None of these three works is cited.

Suggestions for future work (not expected in the rebuttal):

  • Repeat the comparison over several generator fits and draws, with a count-matched control and the positive-QP completion as an additional baseline.
  • Evaluate in a domain where conservation holds in the data, or with a downstream task whose outcome depends on conservation.
Reviewer_bruxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper generates structurally varied continuous-time transaction multigraphs while keeping pinned observations fixed and enforcing flow balance at relay nodes. A learned structural policy scores a bounded menu of topologies, a diffusion prior proposes amounts and timestamps, and DBA-KL projects each proposal onto the positive, relay-balanced set. Evaluation uses three IBM-AML simulator subsets. DBA-KL variants beat the authors’ own ECI-DBA and PCFM-log adaptations on four-GNN average AP, but a no-flow ablation scores higher while violating relay balance in every audited relay-containing output.

Critical strengths
  1. Relay validity is enforced by construction and audited, and the paper shows utility does not measure validity. All DBA-KL outputs satisfy relay balance within tolerance (Table 3). The no-flow ablation has higher AP on every dataset yet fails the relay audit on all 610 relay-containing graphs (Appendix D.5). This argues against judging constrained generated scenarios by downstream accuracy alone, which is the most transferable finding in the paper.

  2. The completion step is formally specified and checked. The fixed-reference KL projection, the exact required/excess split (Eq. 3), the analytic one-row updates (Propositions 2–4), and the toy-problem verification in Appendix E make clear what the completion step guarantees, which aids reproducibility.

Critical weaknesses
  1. The utility advantage over the adaptations is within noise, and the mechanism behind AP is unidentified. On LI-Small the best DBA-KL variant scores 0.193 ± 0.016 against 0.189 ± 0.014 for PCFM-log, and on LI-Medium 0.050 ± 0.028 against 0.045 ± 0.024. Neither gap is tested with paired, seed-matched comparisons. The DBA-KL and adaptation rows also appear to come from different saved populations (Appendix D.1). The no-flow ablation removes three components at once, so it cannot show which one drives AP.

  2. The selection law and several reported numbers are inconsistent. Eq. (5) uses exp(+C_e/T), while Appendix E uses exp(−m_e/(s0T)), so the sign is unclear. Table 11 reports medians (2.45×10³, 7.35×10³) larger than the corresponding maxima (0.82, 91.2). The omission counts in Table 3 do not match the breakdown in Section 4.4 and Appendix D.6. The “17–78×” cost-study speedup cannot be reproduced from Table 10. Because the selection law is central to the claim that completion cost shapes the emitted structure, these errors matter.

Questions for authors

Material questions:

  1. Which sign is implemented in Eq. (5): exp(+C_e/T) or exp(−C_e/T)? Please correct the equation and state whether the relay-free result in Section 4.4 depends on it.

  2. Table 11: how can the median excess (2.45×10³) exceed the maximum (0.82)? Please correct the values or the units.

  3. Please report paired, seed-matched differences (with confidence intervals or tests) between Learned DBA-KL and PCFM-log/ECI-DBA on all three datasets.

  4. Please rerun all completion arms on the same generation population (same seed, same fitted prior) so that the DBA-KL and adaptation rows are directly comparable.

  5. Please provide an ablation that separates quantity correction, feasibility filtering, and cost weighting, for example with rejection sampling on the no-flow arm.

  6. Please reconcile the omission counts in Table 3 with the breakdown in Section 4.4 and Appendix D.6.

  7. Please explain the saturation of the normalized relay residuals at about 10⁸ in Table 9.

  8. Please release the ECI and PCFM adaptation code, since the main comparison depends on it.

Clarification requests:

  1. Proposition 4 says "Σ_k D(z_{k+1}‖z_k) < 1" by telescoping. Telescoping gives ≤ D(z⁺‖q), a finite bound that does not equal 1. Please correct the statement.

  2. ECI-naive and PCFM-naive have identical numbers in Tables 1, 5, 6, 7, and 8. The “eleven configurations” therefore include duplicates. Please clarify.

  3. FraudGT and smart-grid appear in Appendix D and H without context or results. Please remove them or explain.

  4. The 60–113 sweep range in Section 4.5 does not match the medians in Table 10 (60 and 106). Please clarify.

  5. Please state how the practitioner-reported motivation in the introduction relates to the evaluation, or remove it. Appendix F says it is not evidence and provides no data.

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

Related papers

Discussion (0)

Sign in to comment.