acceptodds
Under review as a conference paper at ICLR 2027

Joint Conformal Source Localization: Source Set Inference and Order-Adaptive Set Scoring

Abstract

Source localization models assign scores to candidate origins of an observed spreading process on a network, but these scores alone provide no guarantee of recovering the true sources. Existing conformal methods provide source recall guarantees, yet selecting candidates by individual confidence can require large sets because it does not directly account for how their combined selection meets the recall target. We propose Joint Conformal Source Localization (JCSL), a post hoc framework that uses propagation constraints to infer a distribution over source sets and prioritizes nodes by their expected contribution to normalized source recovery. Order-adaptive set scoring uses information from the original model to determine how far to extend the refined ordering to meet the recall target after calibration. Theoretically, we establish finite-sample marginal recall guarantees under exchangeability and prove that each prefix of the proposed ordering maximizes the expected recall-aware objective among all candidate sets of the same size. Across 15 spreading conditions on three real-world networks, JCSL reduces mean set size by 3.45–95.35% against the best baseline in the main comparison. The mean rate of meeting the recall target is 0.8999 at nominal confidence 0.90. Ablations show that propagation constraints and order-adaptive scoring each reduce mean prediction set size.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal prediction, Uncertainty quantification, Multi-source localization, Probabilistic inference

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_95tsOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes JCSL, a post-hoc method for returning compact source sets with a marginal recall guarantee under exchangeability. It combines joint source-set inference from detector scores and propagation constraints, recall-weighted node ordering, and order-adaptive conformal scoring. Experiments share a frozen detector across three contact networks and 15 SI/SIR conditions, comparing set size and the frequency of meeting the recall target.

Critical strengths

The method has a clear rationale. The individual node confidence can miss how possible sources jointly explain an observed spread. JCSL uses that structure to improve the ordering and then calibrates the resulting sets. The inference, ordering and calibration steps fit together well, and the results make their scope clear. I see a useful contribution beyond calibrating the detector's original node scores. The expedriments are also comprehensive and validated the claims.

Critical weaknesses

The practical scope still needs more evidence, like the experiments use complete observations taken very soon after seeding, though it is unclear how much of the reduction would remain when investigation starts later or reports are incomplete. Incomplete reports also make it harder to ensure that the candidate set contains every true source. Moreover, the cost of joint inference is another limit on practibility, the paper does not show how the model designs affect set efficiency as graph size and complexity increase.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 95ts (Part 1/4)

Thank you for your positive and thoughtful review, and for recognizing the rationale of joint source-set inference, how the inference, ordering, and calibration steps fit together, and the breadth of the experiments. Your comments on later investigations, incomplete reports, and computational cost concern the conditions under which JCSL would be used in practice, and addressing them has made the paper considerably stronger.

We have uploaded a revised version that addresses your comments. The new analyses are presented in the appendices listed below, each referenced from the corresponding section of the main text:

  • Appendix K: later investigations, with the first snapshot moved from step 1 to steps 2 through 5 (Table 39).
  • Appendix I.2: incomplete reports, where true sources may be hidden (Tables 29–31).
  • Appendix I.4: how source containment affects the recall guarantee (Proposition 8).
  • Appendix J: the width budget, larger and denser networks, and computational cost (Tables 35–38).

Appendix H (propagation model mismatch) provides further robustness results. For brevity, the tables below aggregate results over conditions; the corresponding appendices give further per-condition results.

AuthorsOpenReview ↗

Response to Reviewer 95ts (Part 2/4)

1. Later investigations

[Weakness] "The practical scope still needs more evidence, like the experiments use complete observations taken very soon after seeding, though it is unclear how much of the reduction would remain when investigation starts later ..."

1.1 Results (Appendix K, Table 39). We agree that this is an important practical case. Early snapshots are also where localization is most actionable (e.g., for containing an outbreak), yet existing conformal methods rank candidates by individual detector scores without jointly using the propagation constraints among candidate sources, which can leave unnecessarily large sets when early intervention matters most. To measure how much of the reduction remains when investigation starts later, we evaluate first snapshots taken one to five steps after seeding on highSchool, the largest of the three benchmark networks, under slow, medium, and fast spreading. JCSL fits its reference distribution P0P_0 on the fitting split; here we fit it either globally or conditionally on the observed candidate count (Appendix K); neither choice affects the recall guarantee. For each setting, the table reports the strongest of all evaluated baselines.

Mean set sizes (global fit / conditional fit / strongest baseline), with the mean fraction of affected nodes in parentheses:

First snapshotSlowMediumFast
step 1(3%) 10.3 / 10.9 / 13.6(5%) 10.8 / 11.2 / 16.7(6%) 9.9 / 8.7 / 15.7
step 2(7%) 22.3 / 20.8 / 24.2(19%) 53.8 / 42.1 / 51.0(31%) 67.0 / 53.4 / 63.2
step 3(17%) 58.4 / 48.8 / 50.4(46%) 259.3 / 167.3 / 187.0(67%) 368.3 / 275.6 / 294.0
step 4(29%) 145.0 / 108.9 / 116.4(68%) 444.8 / 346.8 / 373.2(85%) 525.8 / 492.9 / 505.1
step 5(41%) 241.7 / 178.6 / 167.9(81%) 550.8 / 452.1 / 460.3(92%) 579.7 / 586.2 / 591.0

With the conditional fit, JCSL remains smaller than the strongest baseline in 11 of the 12 delayed settings, by 6.6% on average and by up to 17.5% (step 2, medium spreading). For example, at step 3 under medium spreading, when nearly half of the nodes are already affected, JCSL returns 167.3 nodes versus 187.0. Over the delayed settings, the mean inclusion rates are 0.898 (global fit), 0.892 (conditional fit), and 0.886 (strongest baselines), all close to the nominal 0.90. These results show that JCSL maintains its advantage over the baselines in most delayed settings, not only with early snapshots.

AuthorsOpenReview ↗

Response to Reviewer 95ts (Part 3/4)

2. Incomplete reports

[Weakness] "... though it is unclear how much of the reduction would remain when investigation starts later or reports are incomplete."

2.1 Missing reports (Appendix I.2, Tables 29–30). On all three networks under slow, medium, and fast spreading, we hide each node's reports over the whole observation window at rate 2–10%. Hidden nodes are encoded as unknown, and true sources can be hidden, so the candidate set may miss sources. Perturbed inputs pass through the same detector, and fitting and calibration use data with the same perturbation. Aggregated over the nine network–speed conditions and compared with the strongest baseline (the baseline with the smallest mean set in each condition):

Missing rateJCSL smaller than strongest baselineMean size reduction vs. strongest baselineInclusion (JCSL / strongest baseline)
0% (clean)9/922.2%0.895 / 0.898
2%9/920.7%0.894 / 0.902
4%9/918.7%0.905 / 0.906
6%9/917.0%0.904 / 0.908
8%9/914.1%0.905 / 0.902
10%8/911.0%0.900 / 0.901

The reductions decrease gradually but persist at every missing rate. At 10%, true sources are frequently excluded from the candidate set (containment 0.43–0.50), so this setting directly tests the case where the containment assumption fails. JCSL nevertheless remains smaller than the strongest baseline in 8 of 9 conditions while keeping its inclusion rates at 0.890–0.915.

2.2 Affected nodes reported as susceptible (Appendix I.2, Table 31). A related form of incomplete reporting is a false negative, which reports an infected or recovered node as susceptible. With 10% false negatives, JCSL is smaller than the strongest baseline in all nine conditions, by 11.1% on average, with an average inclusion rate of 0.893 for both JCSL and the strongest baseline.

3. Source containment under incomplete reports

[Weakness] "Incomplete reports also make it harder to ensure that the candidate set contains every true source."

3.1 When containment holds. Under complete SI/SIR observations, affected nodes never return to the susceptible state, so containment holds at any observation time, and Theorem 2 applies to the delayed snapshots in Section 1. We agree that incomplete reports can break containment; the missing-report setting in Section 2.1 tests this case directly.

3.2 What the guarantee needs (Appendix I.4, Proposition 8). Partial recall needs only enough sources in the candidate set, ∣Y∩VX∣≥rβ(∣Y∣)|Y\cap V_X|\ge r_\beta(|Y|) ("recall attainability"), not all of them. Any output restricted to the candidate set has recall at most ∣Y∩VX∣/∣Y∣|Y\cap V_X|/|Y|. If unattainable calibration examples receive score +∞+\infty and εD\varepsilon_D denotes the probability of unattainability given the fitting data DD, then max⁡{0,1−α−εD}≤Pr⁡{recall≥1−β∣D}≤1−εD.\max\lbrace 0,1-\alpha-\varepsilon_D\rbrace \le \Pr\lbrace \text{recall}\ge 1-\beta \mid D\rbrace \le 1-\varepsilon_D . Thus the guarantee of Theorem 2 holds whenever recall attainability holds almost surely, even if some sources are omitted, and degrades by at most εD\varepsilon_D otherwise. Inference, ordering, and calibration do not depend on whether the candidate set contains all sources; only this probability statement depends on recall attainability. Moreover, the lower bound 1−α−εD1-\alpha-\varepsilon_D is pessimistic. If calibration and test examples are i.i.d., unattainable test examples add to the failures only when the calibrated threshold becomes infinite, which requires about 10% or more of the calibration examples to be unattainable; otherwise, they count toward the α\alpha failures that calibration already allows. With our n=512n=512 calibration examples and α=0.1\alpha=0.1, the success probability is therefore at least 1−α−0.011-\alpha-0.01 whenever εD≤0.084\varepsilon_D\le 0.084. The unattainability rates observed with 10% missing reports or 10% false negatives (at most 0.083) are close to this level, consistent with the near-nominal inclusion observed there.

AuthorsOpenReview ↗

Response to Reviewer 95ts (Part 4/4)

4. Set efficiency and computational cost on larger networks

[Weakness] "Moreover, the cost of joint inference is another limit on practibility, the paper does not show how the model designs affect set efficiency as graph size and complexity increase."

4.1 Set efficiency and cost on larger networks (Appendices J.2–J.3, Tables 36–37). We join two or four copies of highSchool with sparse or dense cross-module edges, giving connected networks with 1,548 and 3,096 nodes, draw sources over the whole graph, and observe from step 3. To show how JCSL's designs affect set efficiency at these scales, we also compare two ablations that keep its reference distribution, scoring, and calibration but order the nodes without the propagation constraints or by the raw detector scores. Mean set sizes at width 12 (three seeds), with JCSL's average computation time per input:

Network (nodes, cross edges)JCSLStrongest baselineNo propagation constraintsDetector orderJCSL time per input
774, original61.162.166.989.60.4 s
1,548, sparse48.458.755.084.90.9 s
1,548, dense55.281.461.5112.50.8 s
3,096, sparse40.062.448.579.01.6 s
3,096, dense45.784.950.2111.01.8 s

JCSL's advantage over the strongest baseline is larger on the larger networks, from 1.7% on the original graph to 17.6–32.2% at 1,548 nodes and 36.0–46.1% at 3,096 nodes, that is, 10–39 fewer candidates on the enlarged graphs at a cost of 0.8–1.8 s of computation per input. Since each additional candidate typically requires follow-up such as testing or contact tracing, this trade-off is favorable in investigation settings. Inclusion rates are 0.888–0.914 for JCSL and 0.893–0.911 for the strongest baseline. Both design components contribute on all five networks: the propagation constraints reduce the set by 8.8–17.6% relative to the ablation without them, and the refined order reduces it by 31.8–58.8% relative to the raw detector order.

4.2 Width budget (Appendices J.1–J.2, Tables 35–37). The width budget controls the cost of joint inference. Across the five networks of Section 4.1, width 4 saves 52–64% of the time relative to width 12, and JCSL with width 4 remains 16.0–45.0% smaller than the strongest baseline on the enlarged graphs. On the benchmark networks, width 4 lowers inference time by 4–23% and changes mean set sizes by at most 0.043 nodes.

4.3 Cost and practical use (Section 5.3 and Appendix J.3, Table 38). JCSL is slower than conformal baselines that only calibrate precomputed scores. Its cost is incurred once per input, is parallel across inputs, and does not depend on the calibration threshold; in the main experiments on highSchool, testing takes about 0.001–0.12 s per input (0.3–46 s for 400 inputs; Section 5.3), and our GPU implementation is 1.2–2.3 times faster than four CPU workers, with identical orders. JCSL is therefore best suited to settings in which a few seconds of computation per input are acceptable, as in investigations where each candidate requires follow-up; faster implementations are left for future work.

5. Summary

5.1 Taken together, these results extend the evidence on the practical scope of JCSL. The reductions persist with delayed first snapshots (Section 1) and under incomplete reports (Section 2), and the recall guarantee requires only that the candidate set contain enough sources. In the larger-network study, JCSL remains smaller than the strongest baseline on all five networks, and the cost of joint inference can be controlled through the width budget.

Thank you again for your positive assessment and constructive comments. We would be glad to address any further questions.

Reviewer_xLTbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Joint Conformal Source Localization (JCSL), a post-hoc framework for constructing compact prediction sets for network diffusion source localization while maintaining finite-sample partial-recall guarantees. The key idea is to move beyond independent node-level confidence scores by incorporating dependencies among candidate sources through propagation constraints.

Specifically, JCSL first constructs a reference distribution over source sets using the predictions of a frozen source detector and an estimated source-count distribution. It then conditions this distribution on propagation-distance constraints and combines multiple hypotheses to obtain a refined distribution over possible source sets. Based on this distribution, the method derives recall-weighted node priorities to construct a refined candidate ordering. An order-adaptive set scoring mechanism subsequently uses the reference distribution to determine the prediction set cutoff through conformal calibration.

The paper provides theoretical results on exact inference under bounded elimination width, optimal ordering for expected normalized source recovery, and finite-sample marginal recall validity under exchangeability. Experiments on three real-world contact networks across 15 simulated SI/SIR spreading conditions demonstrate that JCSL consistently produces smaller prediction sets than the evaluated baselines while achieving empirical inclusion rates close to the nominal confidence level. Additional ablations and controlled comparisons investigate the contributions of propagation conditioning, node ordering, and order-adaptive scoring.

Critical strengths

1.The paper addresses an important limitation of existing source localization methods: individual detector scores do not explicitly account for the dependencies among multiple possible sources. By incorporating propagation-distance constraints into a joint distribution over source sets, JCSL provides a principled mechanism for refining candidate priorities. The combination of structured inference and conformal calibration is technically coherent and well aligned with the goal of reducing prediction set sizes while maintaining statistical reliability. The accompanying theoretical analysis further strengthens the methodological contribution.

2.JCSL consistently achieves smaller mean prediction sets across the 15 evaluated spreading conditions, with particularly substantial improvements in fixed-source-count settings. The empirical inclusion rates remain close to the nominal confidence level, supporting the expected conformal behavior. Importantly, the authors go beyond the primary baseline comparison by providing matched-candidate-range experiments, comparisons with shared score corrections, alternative node-ordering controls, and component-wise ablations. These additional evaluations address several natural concerns about comparison fairness and provide meaningful evidence that the improvements are not solely attributable to a single implementation choice.

Critical weaknesses

1.Theorem 1 establishes that sorting candidates by recall-weighted marginal priorities maximizes expected normalized source recovery under the refined distribution for any fixed prediction set size. However, the primary objective of the paper is to minimize prediction set size subject to a prescribed probability of meeting a recall target. These objectives are not equivalent in general, particularly in the multi-source setting where dependencies among selected sources can substantially affect the probability of satisfying the recall requirement.

The paper provides a lower bound connecting expected normalized recovery to the recall-success probability, but maximizing this bound does not necessarily maximize the actual success probability or minimize the calibrated prediction set size. Moreover, the ordering optimality largely follows from the additive structure of the chosen objective. Therefore, the theoretical result provides a valid justification for the proposed heuristic ordering but is less conclusive regarding the claimed efficiency advantages under the actual constrained optimization problem.

2.The evaluation uses three contact networks with simulated SI/SIR spreading and assumes that the affected-node set extracted from the first observed snapshot contains every true source. This assumption is fundamental to the prediction procedure and its stated validity theorem, but it may not hold under partial observations, missing reports, delayed measurements, or more general diffusion processes.

Furthermore, the most dramatic reductions occur in fixed-source-count settings. In these cases, the reported mean prediction set sizes are close to the minimum number of candidates needed to meet the recall requirement, and the paper acknowledges that randomized calibration can yield empty prediction sets. While these findings are mathematically consistent and still demonstrate useful ranking performance, they raise questions about the practical interpretation of the reported percentage improvements.

The method also relies on structured inference with an elimination-width budget, introducing additional computational cost compared with lightweight conformal baselines. The current experiments do not sufficiently establish how the method behaves under propagation-model mismatch, incomplete observations, or substantially larger networks. These limitations constrain the generality of the empirical conclusions.

Questions for authors

Q1. Relationship between theoretical optimality and actual prediction-set efficiency.

Could the authors clarify the relationship between the expected normalized recovery objective in Theorem 1 and the probability of satisfying the recall requirement? In particular, can the proposed ordering be suboptimal for the actual recall-success probability under the refined source-set distribution? A small counterexample or a discussion of sufficient conditions under which these objectives align would help establish the precise scope of the optimality claim.

Q2. Contribution of propagation constraints versus candidate-set restriction.

The paper already provides matched-candidate-range comparisons in Appendix D.5, which are appreciated. Could the authors further clarify how much of the strong performance in fixed-count settings arises from the propagation feasibility constraints, as opposed to the source-count information and the relatively small affected-node candidate sets? The existing ablations suggest that conditioning is important, but a more explicit interpretation of these effects would strengthen the conclusions.

Q3. Practical interpretation of randomized empty prediction sets.

Appendix D.4 reports that a nontrivial fraction of outputs in fixed-count settings can be empty due to randomized calibration. Could the authors explain whether empty prediction sets are acceptable in the intended source investigation applications? It would also be helpful to emphasize the nonempty-output results alongside the main fixed-count comparisons, so that readers can more directly assess the operational significance of the reported reductions.

Q4. Robustness to propagation-model mismatch and incomplete observations.

How sensitive are the learned mixture weights and refined source rankings to misspecified propagation distances or partially observed infection states? Since the guarantee requires the true source set to lie within the selected candidate set, could the authors clarify which aspects of JCSL remain valid when this containment assumption is violated and whether a conservative candidate-set construction could address this limitation?

Q5. Scalability of structured source-set inference.

The complexity depends exponentially on the retained elimination width, and the reported multi-source runtime is noticeably higher than that of lightweight conformal approaches. Could the authors provide additional evidence or discussion of the tradeoff between elimination-width budget, prediction set size, and computational cost? In particular, how much accuracy or efficiency is lost when a smaller width budget is used?

9 comments
AuthorsOpenReview ↗

Response to Reviewer xLTb (Part 1/8)

Thank you for your detailed review, and for recognizing that modeling dependencies among candidate sources addresses an important limitation of individual detector scores and that the additional controls support the reported gains. Your questions have helped us sharpen both the theory and the experiments.

We have uploaded a revised version that addresses your comments. The new analyses are presented in the appendices listed below, each referenced from the corresponding section of the main text:

  • Appendix F: the scope of the ordering optimality, with an exact counterexample (Table 21), a sufficient condition for agreement (Proposition 7), and a discussion of nesting and calibrated set size.
  • Appendix G: the sources of the fixed-count gains (Tables 22–23) and nonempty outputs (Tables 24–25).
  • Appendix H: propagation-model mismatch (Tables 26–28).
  • Appendix I: missing reports and state errors (Tables 29–32), recall when the candidate set omits sources (Proposition 8), and conservative candidate sets (Tables 33–34).
  • Appendix J: the width budget, larger networks, and computational cost (Tables 35–38).
  • Appendix K: delayed observations (Table 39).

For brevity, the tables below aggregate results over conditions; the corresponding appendices give further per-condition results.

AuthorsOpenReview ↗

Response to Reviewer xLTb (Part 2/8)

1. Theorem 1 and the recall-success objective

[Weakness 1] "... However, the primary objective of the paper is to minimize prediction set size subject to a prescribed probability of meeting a recall target. These objectives are not equivalent in general ..."

1.1 Scope of Theorem 1 (Appendix F). Theorem 1 is designed to justify the ordering rather than to solve the constrained problem directly, and the revision makes this scope explicit. For a candidate set AA of size kk and a distribution QQ over source sets (for JCSL, Q=PR(⋅∣x)Q=P_{\mathrm R}(\cdot\mid x)), Appendix F distinguishes three goals:

  • (O1) Expected normalized recovery, EQ[∣A∩Y∣/rβ(∣Y∣)]\mathbb E_Q[|A\cap Y|/r_\beta(|Y|)]: the expected number of sources in AA relative to the number rβ(∣Y∣)r_\beta(|Y|) required, which gives partial credit for each source found.
  • (O2) Recall-success probability, Q{∣A∩Y∣≥rβ(∣Y∣)}Q\lbrace |A\cap Y|\ge r_\beta(|Y|)\rbrace: the probability that AA contains enough sources to meet the recall target.
  • (O3) Calibrated set size: the mean size of the final conformal sets.

Theorem 1 shows that a single order of the nodes solves (O1) for every size kk at once, and Corollary 6 shows that the same sets maximize a lower bound on (O2). Neither result solves (O2) or (O3) in general, so JCSL's efficiency is supported empirically. The recall guarantee of Theorem 2 holds for any order that depends only on the input and the fitting data, so the scope of Theorem 1 affects efficiency, not validity.

[Q1] "... can the proposed ordering be suboptimal for the actual recall-success probability under the refined source-set distribution? A small counterexample ..."

1.2 An exact counterexample (Appendix F.1, Table 21). The following exact example shows that (O1) and (O2) can differ at some sizes. Take four candidates {a,b,c,d} and exactly two sources, so meeting the target (rβ(2)=2r_\beta(2)=2) requires finding both. Let QQ assign probability 0.26 to {a,b}, 0.22 to {a,c}, 0.17 to {a,d}, and 0.35 to {c,d}. Node aa appears in three of the four source sets and receives the highest priority, and the order is (a,c,d,b)(a,c,d,b). At size 2:

SetExpected normalized recoveryRecall-success probability
{a,c} (JCSL prefix)0.6100.220
{c,d}0.5450.350

The JCSL prefix finds the most sources on average, but {c,d} is the most likely complete source set and therefore meets the target more often. This QQ arises from JCSL's constraints: it is exactly the constrained distribution on a star graph centered at aa, with a,ba,b infected, c,dc,d recovered, and distance hypothesis h=2h=2, where the constraints rule out {b,c} and {b,d}; our implementation retains all four constraints and returns the same feasible sets. A refined distribution of the form 0.9Q+0.1P00.9Q+0.1P_0, which mixes it with P0P_0 as JCSL does and gives all six pairs positive probability, keeps the gap (0.209 versus 0.333). At size 3, the JCSL prefix is optimal for both goals.

[Q1] "... or a discussion of sufficient conditions under which these objectives align ..."

1.3 A sufficient condition for agreement (Appendix F.2, Proposition 7). The two goals agree when, given the number of sources, the probability of a source set is proportional to the product of node weights, Q(y)=η(∣y∣)∏v∈ywv/e∣y∣(Vx)Q(y)=\eta(|y|)\prod_{v\in y}w_v/e_{|y|}(V_x), as in the reference distribution P0P_0. Proposition 7 shows that every prefix of the JCSL order then maximizes both goals at every size, for any count law and any β\beta: the priorities follow the weights, and replacing a node with one of larger weight never lowers the success probability. A gap can therefore arise only when propagation constraints add joint structure that this product form cannot express, such as ruling out {b,c} and {b,d} in Section 1.2, as you anticipated. In that case, (O1) serves as a tractable surrogate for (O2).

[Weakness 1] "... maximizing this bound does not necessarily maximize the actual success probability ... Moreover, the ordering optimality largely follows from the additive structure of the chosen objective."

1.4 Role of the additive objective (Appendix F.3). The additive structure is a deliberate choice. Conformal calibration selects a prefix of one fixed order, so it needs nested candidate sets, where the set of size k+1k+1 contains the set of size kk. (O1) yields such a family that is optimal at every size. The best sets for (O2) need not be nested: if {a,b} has probability 0.4 and {c,d} and {c,e} each have 0.3, the best 2-node set is {a,b}, but the best 3-node set is {c,d,e}, which does not contain it. The calibrated size (O3) also depends on the calibration data and the scoring rule. Theorem 1 thus justifies the ordering, and efficiency is supported by the matched comparisons, ordering controls (Appendix D.8), and ablations.

AuthorsOpenReview ↗

Response to Reviewer xLTb (Part 3/8)

2. Sources of the fixed-count gains

[Weakness 2] "Furthermore, the most dramatic reductions occur in fixed-source-count settings."

[Q2] "Could the authors further clarify how much of the strong performance in fixed-count settings arises from the propagation feasibility constraints, as opposed to the source-count information and the relatively small affected-node candidate sets?"

2.1 Controlled factorial study (Appendix G.1, Tables 22–23). On highSchool SI and SIR with 7 or 10 sources, we fix the candidate set, detector scores, score correction, and distance mixture, and cross three count laws (exact: point mass at the true count; coarse: uniform on {6,…,10}; broad: uniform on {1,…,15}) with two node orders: full JCSL, which uses the propagation constraints, and no conditioning, which uses the reference distribution P0P_0 alone. Both share the same scoring and are recalibrated on shared data. To give the baselines the same count information, we also feed setCP-rec/pre/min and FavMac the node-level source probabilities of P0P_0 under each count law in place of the detector scores.

2.2 Propagation constraints. With the candidate set fixed, conditioning reduces the mean set under every count law, by 35.2–65.7%, in all 12 condition–law combinations and all 36 repeat-level comparisons. For example (SI, 7 sources; set size / inclusion):

Count lawNo conditioningFull JCSL
exact13.48 / 0.9114.62 / 0.887
broad11.66 / 0.9146.89 / 0.904

2.3 Source-count information. Count information helps, but mainly through conditioning. Within full JCSL, the exact count gives the smallest sets (coarsening adds 0.03–0.04 nodes for 7 sources and 3.8–4.1 nodes for 10). Within the unconditioned control, however, the broad law gives smaller sets than the exact count in all four conditions. Baselines given the same count information improve, yet full JCSL remains 28.8–61.1% smaller than the strongest count-informed baseline in all 12 combinations (e.g., SIR, 7 sources, exact: 4.66 vs. 10.08 for count-informed setCP-min). Its inclusion rate is lower in 6, equal in 2, and higher in 4 of the 12 combinations.

2.4 Candidate restriction. In the existing matched-candidate study (Appendix D.5), restricting the baselines to VxV_x leaves setCP-rec and setCP-min unchanged in the fixed multi-source conditions and shrinks ArbiTree from about 770 to 42–59 nodes, still far above JCSL's 4.6–6.9.

Together, these controls indicate that the fixed-count gains arise mainly from propagation conditioning, amplified by count information, rather than from count information or the small candidate set alone. Because the factors interact, we do not split the gain into additive shares.

[Weakness 2] "In these cases, the reported mean prediction set sizes are close to the minimum number of candidates needed to meet the recall requirement ..."

2.5 Why mean sizes fall below the minimum (Appendix G.2). In the main fixed-count experiments, every JCSL output is either the first rβ(s)r_\beta(s) nodes of the order, the smallest set that can meet the target, or, through randomization, empty. The mean size therefore equals rβ(s)r_\beta(s) times the fraction of nonempty outputs. A mean below rβ(s)r_\beta(s) thus reflects randomized empty outputs, not successful investigations with fewer than rβ(s)r_\beta(s) nodes. With empty outputs removed, the sets have exactly 1, 5, and 7 nodes, and the reductions remain large (Section 3).

Reviewer_6g6hOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Joint Conformal Source Localization (JCSL), a post-hoc framework that constructs compact candidate-source sets with finite-sample marginal partial-recall guarantees. It builds a reference source-set distribution from frozen detector scores and source-count probabilities, then refines the distribution using propagation-distance constraints. Recall-weighted node priorities determine a new ordering, while an order-adaptive scoring function and randomized conformal calibration determine the final prediction set. The paper establishes optimality for an expected normalized-recovery objective and tractable inference under bounded elimination width. Experiments across three real-world network topologies and 15 simulated spreading conditions demonstrate reductions in mean prediction set size against the evaluated baselines.

Critical strengths
  1. The paper explicitly states the assumptions required for its finite-sample marginal recall guarantee. The experimental section includes extensive comparisons and statistical analyses across different propagation conditions. Additional controls involving matched candidate ranges, shared score correction, alternative node orderings, and mechanism ablations strengthen the evidence for the reported improvements.
  2. Unlike setCP (Jian et al., AAAI 2026), which directly uses the original source detector's ranking, JCSL incorporates propagation constraints into joint source-set inference and uses the resulting distribution to refine candidate priorities. The bounded-width variable-elimination procedure provides a computationally tractable way to implement this idea without enumerating all source subsets.
Critical weaknesses
  1. JCSL constructs its candidate set from affected nodes in the first observed snapshot and assumes that all true sources belong to this set. In the main experiments, this snapshot is obtained only one or two simulator steps after source initialization. Consequently, the propagation constraints may be particularly informative under the evaluated conditions. However, practical source localization often involves delayed observations, incomplete infection reports, and noisy node states. The paper does not establish whether the substantial reductions in prediction set size persist under these more challenging conditions. This limits the practical applicability of the framework.
Questions for authors
  1. How robust is JCSL to delayed snapshots, incomplete observation of affected nodes, and noisy node states? Can the authors evaluate such settings and clarify how the source-containment assumption affects the applicability of the recall guarantee?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 6g6h (Part 1/4)

Thank you for your thoughtful review and for recognizing the explicit statement of our assumptions, the breadth of the experimental controls, and the difference between JCSL and setCP. Your question about delayed, incomplete, and noisy observations is highly relevant to real investigations, and addressing it has made the paper considerably stronger.

We have uploaded a revised version that addresses your comments. The new analyses are presented in the appendices listed below, each referenced from the corresponding section of the main text:

  • Appendix K: delayed observations, with the first snapshot moved from step 1 to steps 2 through 5 (Table 39).
  • Appendix I.2: missing reports and misreported node states, where true sources may be hidden (Tables 29–31).
  • Appendix I.3: how the fitted model responds to these errors (Table 32).
  • Appendix I.4: how source containment affects the recall guarantee (Proposition 8).

Appendix H (propagation model mismatch) and Appendix J (larger networks and runtime) provide further robustness and cost results. For brevity, the tables below aggregate results over conditions; the corresponding appendices give further per-condition results.

AuthorsOpenReview ↗

Response to Reviewer 6g6h (Part 2/4)

1. Delayed snapshots

[Weakness 1] "In the main experiments, this snapshot is obtained only one or two simulator steps after source initialization. Consequently, the propagation constraints may be particularly informative under the evaluated conditions."

[Question 1] "How robust is JCSL to delayed snapshots ...?"

1.1 Results (Appendix K, Table 39). We agree that this is an important practical case. Early snapshots are also where localization is most actionable (e.g., for containing an outbreak), yet existing conformal methods rank candidates by individual detector scores without jointly using the propagation constraints among candidate sources, which can leave unnecessarily large sets when early intervention matters most. To assess robustness beyond this stage, we evaluate first snapshots taken one to five steps after seeding on highSchool, the largest of the three benchmark networks, under slow, medium, and fast spreading. JCSL fits its reference distribution P0P_0 on the fitting split; here we fit it either globally or conditionally on the observed candidate count (Appendix K); neither choice affects the recall guarantee. For each setting, the table reports the strongest of all evaluated baselines.

Mean set sizes (global fit / conditional fit / strongest baseline), with the mean fraction of affected nodes in parentheses:

First snapshotSlowMediumFast
step 1(3%) 10.3 / 10.9 / 13.6(5%) 10.8 / 11.2 / 16.7(6%) 9.9 / 8.7 / 15.7
step 2(7%) 22.3 / 20.8 / 24.2(19%) 53.8 / 42.1 / 51.0(31%) 67.0 / 53.4 / 63.2
step 3(17%) 58.4 / 48.8 / 50.4(46%) 259.3 / 167.3 / 187.0(67%) 368.3 / 275.6 / 294.0
step 4(29%) 145.0 / 108.9 / 116.4(68%) 444.8 / 346.8 / 373.2(85%) 525.8 / 492.9 / 505.1
step 5(41%) 241.7 / 178.6 / 167.9(81%) 550.8 / 452.1 / 460.3(92%) 579.7 / 586.2 / 591.0

With the conditional fit, JCSL remains smaller than the strongest baseline in 11 of the 12 delayed settings, by 6.6% on average and by up to 17.5% (step 2, medium spreading). For example, at step 3 under medium spreading, when nearly half of the nodes are already affected, JCSL returns 167.3 nodes versus 187.0. Over the delayed settings, the mean inclusion rates are 0.898 (global fit), 0.892 (conditional fit), and 0.886 (strongest baselines), all close to the nominal 0.90. These results show that JCSL maintains its advantage over the baselines in most delayed settings, not only with early snapshots.

AuthorsOpenReview ↗

Response to Reviewer 6g6h (Part 3/4)

2. Incomplete observation of affected nodes

[Weakness 1] "However, practical source localization often involves delayed observations, incomplete infection reports, and noisy node states. The paper does not establish whether the substantial reductions in prediction set size persist under these more challenging conditions."

[Question 1] "How robust is JCSL to ... incomplete observation of affected nodes ...? Can the authors evaluate such settings ...?"

2.1 Missing reports (Appendix I.2, Tables 29–30). On all three networks under slow, medium, and fast spreading, we hide each node's reports over the whole observation window at rate 2–10%. Hidden nodes are encoded as unknown, and true sources can be hidden, so the candidate set may miss sources. Perturbed inputs pass through the same detector, and fitting and calibration use data with the same perturbation. Aggregated over the nine network–speed conditions and compared with the strongest baseline (the baseline with the smallest mean set in each condition):

Missing rateJCSL smaller than strongest baselineMean size reduction vs. strongest baselineInclusion (JCSL / strongest baseline)
0% (clean)9/922.2%0.895 / 0.898
2%9/920.7%0.894 / 0.902
4%9/918.7%0.905 / 0.906
6%9/917.0%0.904 / 0.908
8%9/914.1%0.905 / 0.902
10%8/911.0%0.900 / 0.901

The reductions decrease gradually but persist at every missing rate. At 10%, true sources are frequently excluded from the candidate set (containment 0.43–0.50), so this setting directly tests the case where the containment assumption fails. JCSL nevertheless remains smaller than the strongest baseline in 8 of 9 conditions while keeping its inclusion rates at 0.890–0.915.

3. Noisy node states

[Question 1] "How robust is JCSL to ... noisy node states?"

3.1 State misclassification (Appendix I.2, Table 31). We misreport node states: false negatives (FN) report an infected or recovered node as susceptible, and false positives (FP) report a susceptible node as infected. Since the FP rate applies to the much larger set of susceptible nodes, we use 10% FN and 1% FP, alone and combined, and 5% FP as a stress test; 1% FP already enlarges the candidate set by more than a third on highSchool (slow condition). Aggregated over the nine network–speed conditions and compared with the strongest baseline in each condition:

Error typeJCSL smaller than strongest baselineMean size reduction vs. strongest baselineInclusion (JCSL / strongest baseline)
FP 1%9/97.6%0.898 / 0.900
FN 10%9/911.1%0.893 / 0.893
FN 10% + FP 1%9/94.1%0.892 / 0.891
FP 5% (stress test)8/93.7%0.892 / 0.904

JCSL is smaller than the strongest baseline in all 27 conditions with false negatives, false positives, or both, with essentially the same average inclusion rates as these baselines. Even under 5% FP, which nearly triples the candidate set on highSchool, JCSL remains smaller in 8 of 9 conditions.

3.2 How JCSL responds to frequent false positives (Appendix I.3, Table 32). A false positive creates an "infected" node that is often beyond the reach of the true sources within the assumed propagation distances, so fitting examples can become inconsistent with every distance hypothesis. JCSL's fitted mixture responds to this automatically: the weight on the unconstrained reference component grows with the fraction of inconsistent fitting examples (Table 32). When false positives are frequent, the ranking moves toward the reference distribution and can reduce to it entirely. JCSL therefore relies less on constraints that the observations contradict. This is one reason the efficiency gain shrinks under heavy false positives, but it also limits the influence of contradicted constraints on the ranking, and the recall guarantee is unaffected because false positives keep every source in the candidate set and calibration uses data with the same errors.

3.3 Practical guidance. JCSL tolerates moderate rates of missed reports and false negatives, as well as occasional false positives; frequent false alarms are the main case to avoid, so positive reports should be confirmed where possible. In all cases, the calibration data should reflect the same reporting process as the deployment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their constructive feedback and have uploaded a revised version. The new analyses are presented in Appendices F–K, and the corresponding parts of the main text (Sections 4.2, 4.3, 5.1, 5.2, and 5.3) now refer to them:

  • Appendix F: scope of the ordering optimality, with a counterexample and a sufficient condition (xLTb).
  • Appendix G: sources of the fixed-count gains and nonempty outputs (xLTb).
  • Appendix H: propagation-model mismatch (xLTb, 6g6h).
  • Appendix I: missing reports, state errors, and recall when the candidate set omits sources (all reviewers).
  • Appendix J: width budget, larger networks, and computational cost (xLTb, 95ts).
  • Appendix K: delayed observations (all reviewers).

Our responses to each reviewer summarize the corresponding results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.