acceptodds
Under review as a conference paper at ICLR 2027

GO-Flow: Positive-Only Graph Learning from Conditional Flow Fields

Abstract

In many graph prediction problems, the observed positives are reliable, whereas the unobserved graph objects are ambiguous: a missing edge may be an undiscovered relation, and an unlabeled node may still belong to the target class. We study positive-only graph learning, where trainable parameters are fit from observed positives without assigning negative labels to the remaining graph objects or using them as unlabeled training examples. Conditional flow matching provides a positive-density model, but density estimation alone does not determine how candidates should be ranked. We introduce GO-Flow (Graph learning from Observed positives with conditional Flow), which fits a conditional flow only on observed positive graph objects and uses it to score the unlabeled candidates; for relations, the score contrasts conditional and background densities so that it reflects what the candidate context adds. Relation prediction is the primary instantiation, evaluated across standard link prediction and drug-target interaction, while node positive-unlabeled ranking and a class-as-positive study apply the same conditional flow to nodes. With positive-only training and per-dataset selection, which uses sampled validation non-edges, GO-Flow improves over the learned one-class baselines on all eight link-prediction datasets and compares favorably with the positive-only baselines on seven; methods that additionally use unlabeled or negative examples remain stronger in some regimes, since the unlabeled pool and reliable negative labels carry information that positives alone do not, whereas hidden positives labeled negative degrade these methods.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
positive-only learning, graph learning, conditional flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XTCbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents GO-Flow, a positive-only graph learning framework for predicting unobserved relations and identifying candidate nodes. Motivated by the fact that missing edges or unlabeled nodes are not necessarily negative examples, GO-Flow trains a conditional flow using observed positives alone. For relation prediction, it scores candidate pairs by comparing the conditional likelihood of a target under its graph context with its likelihood under a background context. The method also supports a positive-only structural encoding to complement the flow-based score. A theoretical analysis connects the proposed density contrast to optimal ranking under specific assumptions. Experiments cover link prediction, drug–target interaction, and node ranking, with additional studies on false-negative contamination, model selection, alternative estimators, and computational efficiency.

Critical strengths
  1. I find the motivation both practical and well grounded. In applications such as drug–target interaction prediction, an unobserved relation may simply not have been discovered, so treating every missing edge as a negative example can introduce misleading supervision. GO-Flow addresses this issue by learning exclusively from confirmed positive relations and using graph-conditioned distributions to evaluate new candidates. What I particularly appreciate is that the authors clearly distinguish positive-only parameter training from model selection. This makes the scope of the proposed setting more transparent rather than presenting it as a completely negative-free pipeline.

  2. Another interesting contribution is the conditional-to-background density contrast. Instead of ranking a candidate solely by how likely its features are under the learned positive distribution, GO-Flow asks whether the corresponding graph context makes those features more plausible. This is a meaningful distinction, since common node features should not automatically imply a strong relation. The results in Table 6 provide direct evidence for this design, with the background contrast improving over conditional likelihood alone across the five evaluated datasets. I also appreciate the analysis in Section 4.4 showing that a better fit to the positive distribution does not necessarily produce better rankings. This observation gives the scoring mechanism a clearer justification beyond simply applying Flow Matching to graph data.

  3. I particularly appreciate the effort to examine the method's limitations rather than focusing only on favorable comparisons. GO-Flow performs strongly against positive-only baselines across the link-prediction benchmarks and shows encouraging results on drug–target interaction tasks. Beyond these results, the authors evaluate false-negative contamination, hard-negative ranking, direct density-ratio estimators, and positive-only model selection. Several of these experiments reveal situations where other methods perform better or where GO-Flow becomes less effective. Reporting these findings makes the empirical analysis more credible and helps readers understand the actual trade-offs of positive-only learning. The extensive supplementary experiments are also useful for identifying when the approach is most likely to work well.

Critical weaknesses
  1. I would like a clearer discussion of how much the reported performance depends on negative examples during model selection. Although GO-Flow trains its parameters exclusively on observed positives, its main evaluation protocol still uses sampled validation non-edges to select the encoder, conditioning strategy, scoring weights, and other configurations. I appreciate that the authors openly acknowledge this distinction and even investigate positive-only model selection in Appendix T. However, the reported average ROC-AUC drops by 0.049 when positive-only validation is used with structural encoding, with a particularly large decrease on PPI. This suggests that validation information plays a meaningful role in achieving the reported results. I do not consider this a contradiction of the positive-only training claim, but it does limit the method's applicability when reliable negative examples are unavailable throughout the entire pipeline. A more prominent discussion of this limitation, together with practical guidance on model selection in such settings, would strengthen the paper.

  2. The drug–target interaction evaluation protocol deserves further attention. Appendix J.2 explains that model configurations are selected using validation interactions drawn from the full interaction set, some of which later appear in the cross-validation test folds. Although the flow models themselves are not trained on these test interactions, their labels can still indirectly influence configuration selection. I appreciate the authors' transparency about this issue, but describing the resulting bias as small does not establish its actual magnitude. Since the DTI experiments provide some of the strongest results in the paper, I think an additional evaluation with strictly separated validation and test interactions would be valuable. For example, selecting configurations within the training portion of each fold would help determine whether the reported improvements remain consistent. This would also make the comparison with previously published PU methods more convincing.

  3. I am still somewhat uncertain whether the additional complexity of conditional flow matching is always justified. The conditional-to-background contrast is a sensible scoring principle, but estimating density ratios does not necessarily require learning a full conditional flow. I appreciate that Appendix S includes direct bilinear and MLP estimators trained using positive relations alone. Interestingly, these simpler alternatives outperform the flow-based score on UMLS and WN18RR, while requiring substantially less time to rank candidates. The reported scoring speed difference exceeds two orders of magnitude. GO-Flow performs better on the remaining six datasets in that comparison, so I do see value in the proposed approach. Still, a clearer discussion of the accuracy–efficiency trade-off would help explain when the flow model is worth its computational cost. I would also encourage the authors to distinguish the theoretical interpretation of the density contrast from a general ranking guarantee, since Proposition 1 relies on a specific independence assumption for non-relations.

  4. Some aspects of the evaluation could better reflect the challenges of real-world link discovery. In particular, the positive-only structural encoding includes endpoint degrees and path-based statistics, while conventional link-prediction benchmarks with randomly sampled non-edges can favor methods that exploit degree information. This issue is discussed in Implicit Degree Bias in the Link Prediction Task (Aiyappa et al., ICML 2025, https://proceedings.mlr.press/v267/aiyappa25a.html) and Bias-aware Training and Evaluation of Link Prediction Algorithms in Network Biology (Yılmaz et al., PNAS 2025, https://doi.org/10.1073/pnas.2416646122). I think these studies are particularly relevant given the biomedical applications considered here. The authors already provide HeaRT hard-negative results, which I appreciate, but GO-Flow is noticeably weaker than several negative-trained methods in that setting. Another issue is the difference between training and evaluation graph contexts. As discussed in Appendix C, training target edges remain available during message passing, whereas evaluated edges are excluded. The authors investigate target-edge masking, but its effect varies across datasets and configurations. A clearer interpretation of these findings, together with a degree-aware evaluation or additional analysis using existing results, would help establish how reliably GO-Flow identifies genuinely difficult and previously unseen relations.

  5. I think the presentation could better reflect the amount of experimental work that has been done. Figure 1 provides a useful conceptual overview, but it is quite simplified and does not fully explain how the graph encoder, conditional flow, background density estimation, and positive-only structural encoding fit together. A more complete architecture diagram would make the proposed framework easier to understand without requiring readers to reconstruct the workflow from the equations. I also noticed that almost all experimental findings are presented in tables, with very little visual analysis. This makes the paper feel somewhat dense despite the considerable amount of evaluation. Several existing results would be particularly suitable for visualization, such as performance changes under false-negative contamination, the relationship between positive-density quality and ranking performance, and the accuracy–efficiency trade-off against direct density-ratio estimators. A few carefully selected plots could make these findings much more intuitive. I do not think additional experiments are necessarily required here. Better organization and visualization of the existing evidence would already improve the paper substantially.

Questions for authors

Please see Weaknesses 1–4 above. I would particularly appreciate clarification on how much the DTI results are affected by the overlap between validation and cross-validation test interactions, and whether the main improvements remain consistent under a strictly separated evaluation protocol. I would also be interested in a clearer explanation of when conditional flow matching offers sufficient benefits over simpler density-ratio estimators to justify its additional computational cost. Finally, the limitations of positive-only model selection and the generalization to more challenging candidate relations deserve further discussion.

4 comments
AuthorsOpenReview ↗

Authors Responses (Part 1)

Thank you for the thorough and constructive review, and for the references on degree bias [1, 2]. We are glad that you found the motivation "both practical and well grounded", appreciated that we "clearly distinguish positive-only parameter training from model selection", found that the conditional-to-background contrast draws "a meaningful distinction", and valued "the effort to examine the method's limitations". We hope that you find our responses satisfactory, and that you will consider revising your score.


In short: your review asks whether the reported results hold under stricter protocols and harder candidates, and how the existing evidence can be organized and visualized. Following it, we:

  1. measured a configuration fixed in advance, a positive-only selection criterion, and false-negative contamination in training and validation;
  2. re-selected the drug–target configuration inside every cross-validation fold;
  3. added an accuracy–cost discussion, and stated Proposition 1 as an interpretation of the score;
  4. added a degree-aware evaluation on all eight link graphs and the four drug–target families, with SEAL, BUDDY, NCNC and LPFormer on the same negatives of the six graphs of Table 4;
  5. added an architecture diagram and three figures of existing results.

Regarding model selection and practical guidance (W1, Q). We appreciate the question on how much the reported performance depends on validation non-edges. We evaluated one configuration fixed in advance, the configuration of Appendix C (GCN, FiLM, cross-endpoint conditioning, the full contrast, no reflow, fusion weight α = 0.5 with the positive-only structural encoding, PSE), on all eight link datasets with no validation non-edge anywhere (Table 25):

ROC-AUCCoraCiteSeerUMLSWN18RRDISEASESDrug–DiseasePPIOhmnetMean
PSE-only (Table 1)0.6950.5920.8560.7350.7080.8000.8840.8700.768
GO-Flow, per-dataset selection (Table 1)0.9130.9350.9370.7510.9200.7970.8890.8940.880
GO-Flow, fixed configuration (Table 25)0.8550.7980.8730.6910.8790.7470.8120.8940.819

It averages 0.819 ROC-AUC, compared with 0.880 with per-dataset selection and 0.768 for PSE-only, the strongest positive-only baseline of Table 1 on average. We also introduce a criterion built only from observed positives, the ROC-AUC with which validation positives rank above recombined validation positives. It selects every setting of the flow without the PSE, and its choice is 0.011 ROC-AUC below the choice made with validation non-edges on average (0.095 lower on PPI, where the positive-only criteria of Appendix T lose 0.264). With the validation non-edges contaminated at the training rate for GO-Flow, BUDDY and NCNC, GO-Flow retains at least 97% of its ROC-AUC at 40% contamination on all five graphs:

Retention at 40% contamination, validation contaminated (Table 26)CoraCiteSeerPPIWN18RROhmnet
BUDDY (trained with negatives)0.9630.9750.9340.8790.974
NCNC (trained with negatives)0.9530.9740.9980.9450.994
GO-Flow (observed positives only)0.9971.0000.9971.0010.972

Following your comment, Sections 3.2 and 4.4 now state this dependence with practical guidance: without reliable validation negatives, use the fixed configuration, which needs no selection criterion and is 0.061 lower on average and 0.137 lower on CiteSeer; with a small set of trusted non-edges, select on them.

AuthorsOpenReview ↗

Authors Responses (Part 2)

Regarding a strictly separated protocol for drug–target interaction (W2, Q). Thank you for proposing a strictly separated evaluation. As you suggested, we selected configurations within the training portion of each fold: inside every one of the 50 cross-validation folds (150 per family over three seeds), on 15% of the training interactions of that fold, with ten non-interactions each drawn from pairs that are not test candidates of the fold. No test pair, positive or negative, informs any choice (Tables 2 and 13):

ROC-AUCEnzymeIon channelGPCRNuclear receptor
Strongest positive-only baseline (Table 2)0.7980.8250.7900.739
PUDTI [3] (as published)0.8980.8750.8940.907
PUDT-Lan [4] (as published)0.8840.8310.8780.885
GO-Flow, submitted protocol (Table 2)0.9680.9690.9560.894
GO-Flow, nested selection0.9630.9660.9540.854

Under nested selection, GO-Flow reaches 0.963, 0.966 and 0.954 on Enzyme, Ion channel and GPCR, within 0.005 of the submitted protocol and 0.060–0.091 above every published positive–unlabeled (PU) method. On Nuclear receptor, the smallest family (90 interactions, about 11 validation positives per inner split), it reaches 0.854, below PUDTI and PUDT-Lan. On all four families it remains 0.115–0.165 above the strongest positive-only baseline. In the revision, Appendix J reports the nested protocol beside the submitted one, Tables 2 and 13 add the nested rows, and Appendix K adds nested ranking metrics.

Regarding the cost of the flow and the role of the theory (W3, Q). Thank you for asking when the flow is worth its cost. Table 24 (Table 22 in the original submission) compares the flow score with the bilinear and MLP estimators of Appendix S, which estimate the full contrast (w = 1) by classification:

ROC-AUCCoraCiteSeerUMLSWN18RRDISEASESDrug–DiseasePPIOhmnet
Bilinear0.8170.8050.9850.6710.6780.6320.6760.617
MLP0.8110.9180.9740.7150.7820.6280.7110.635
GO-Flow without the PSE (α = 0)0.9130.9350.9110.6560.9200.6390.8290.746

The flow score is higher on six of the eight graphs, by 0.11–0.14 on DISEASES, PPI and Ohmnet, and the direct estimators are higher on UMLS and WN18RR; they score 100,000 candidates in 0.03–0.07 s, compared with 9.68 s. The flow is worth its cost where the direct estimators are behind on validation; where a direct estimator matches the flow on validation, it is the better choice. Five integration steps instead of twenty run 3.3–4.0× faster for a median ROC-AUC decrease of 0.007 (Appendix Y). Following your suggestion, Section 4.4 now discusses this trade-off with Figure 5. On the theory, Proposition 1 states optimality under a model of non-relations, and w is selected because that model is unknown. In the revision, Section 3.2 presents Proposition 1 as an explanation of the score under this model, and Appendix B distinguishes this theoretical interpretation from a general ranking guarantee.

AuthorsOpenReview ↗

Authors Responses (Part 3)

Regarding degree-aware evaluation and difficult relations (W4, Q). We appreciate the references and the question on how reliably GO-Flow identifies difficult relations. As you suggested, we added a degree-aware evaluation on all eight link datasets with the splits, configurations and seeds of the paper: degree-corrected non-edges, with endpoints drawn in proportion to their training degree [1, 2], degree-matched non-edges in the same log2-degree bins as each positive, and ROC-AUC stratified by degree. SEAL, BUDDY, NCNC and LPFormer are scored on the same negatives of the six graphs of Table 4 (Appendix Z, Tables 31 to 33). The evaluation complements the HeaRT evaluation of Section 4.3 [5], where negatives are chosen by heuristics to resemble true edges, which negative-trained methods can learn to reject:

ROC-AUC, degree-corrected non-edgesCoraCiteSeerUMLSWN18RRPPIOhmnet
SEAL0.8360.8350.9720.8700.6340.731
BUDDY0.8850.9220.9750.6860.5750.706
NCNC0.9360.9490.9380.7620.6010.715
LPFormer0.8880.9170.9460.7520.6120.730
GO-Flow without the PSE (α = 0)0.8670.9030.8400.6300.5050.535
GO-Flow with the PSE0.8490.8490.8770.5710.5460.621

On Cora, CiteSeer and UMLS, the four negative-trained methods decrease by 0.01–0.06 from Table 4 under degree correction, about as much as the flow (0.03–0.07), while PSE-only reaches 0.540 on Cora and 0.459 on CiteSeer; on the lowest-degree tertile of Cora positives and the standard non-edges, the flow reaches 0.831, compared with 0.408 for PSE-only. On WN18RR, PPI and Ohmnet, the negative-trained methods decrease by 0.06–0.29 and reach 0.57–0.87, and the flow reaches 0.505–0.630. On the four drug–target families, we scored the configurations selected in every fold of the nested protocol on degree-corrected non-interactions (Table 33):

ROC-AUC, GO-Flow under nested selectionEnzymeIon channelGPCRNuclear receptor
Standard non-interactions0.9630.9660.9540.854
Degree-corrected non-interactions0.9450.9430.8990.811

The drug–target results therefore decrease by 0.018–0.055 under degree correction; with the flow without the PSE, Nuclear receptor reaches 0.839, while PSE-only reaches 0.296–0.562 and the weighted-profile baseline reaches 0.578–0.650 (Table 33). Appendix Z of the revision reports this evaluation.

Regarding training and evaluation contexts (W4). Thank you also for asking how to read the masking study of Appendix C. Removing the target edges of each minibatch from message passing during training raises ROC-AUC on Cora and CiteSeer by 0.013 and 0.004 beyond random edge dropout, and on these graphs the flow does not depend on seeing the target edge; on PPI it lowers ROC-AUC by 0.082 beyond random dropout, and with the PSE the effect is within 0.021 of random dropout on all three graphs:

Test ROC-AUC, configuration of Appendix C (Table 8)CoraCiteSeerPPI
Without the PSE, unmasked0.9170.9250.642
Without the PSE, random edge drop0.9290.9550.613
Without the PSE, target edges masked0.9420.9590.530
With the PSE (α = 0.5), unmasked0.8550.7980.812
With the PSE (α = 0.5), random edge drop0.8790.8470.804
With the PSE (α = 0.5), target edges masked0.8840.8330.783

Evaluated edges are never in the message-passing graph, and the reported numbers therefore assess unseen relations. Following your comment, Appendix C now gives this interpretation.

AuthorsOpenReview ↗

Authors Responses (Part 4)

Regarding figures of the existing evidence (W5). We are grateful for the suggestions on presentation. As you proposed, we added three figures from existing results: (i) ROC-AUC under contamination on Cora and retention on all five graphs (Figure 3); (ii) held-out log-likelihood and ROC-AUC for independent, optimal-transport and reflow couplings on the five graphs of Section 4.4, showing the two objectives moving in opposite directions on four of the five (Figure 4); (iii) ROC-AUC and scoring time of the direct estimators and of the flow at each integration budget (Figure 5). We also added an architecture diagram (Figure 2, Appendix C, referenced from Section 3) with a training panel and a scoring panel. It shows the encoder, the target–context construction under cross-endpoint and joint conditioning, the conditional and null contexts of the one velocity field, the change-of-variables log-density, the contrast, the node-pair score, and the rank fusion with the PSE; its caption names the equation of each stage.


We sincerely appreciate the constructive feedback, which strengthened both the evaluation and the presentation of our work. We hope that the nested protocol, the degree-aware evaluation, the model-selection measurements and the new figures address your concerns, and that you will consider revising your score. We are happy to discuss any remaining questions.


References:

[1] Implicit degree bias in the link prediction task

[2] Bias-aware training and evaluation of link prediction algorithms in network biology

[3] Screening drug-target interactions with positive-unlabeled learning

[4] Predicting drug–target interaction using positive-unlabeled learning

[5] Evaluating Graph Neural Networks for Link Prediction: Current Pitfalls and New Benchmarking

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

This paper introduces GO-Flow for ranking graph relations and nodes using positive-only training. It jointly trains a graph encoder and a conditional flow model on observed positives. For relations, the main scoring principle contrasts conditional and background densities to measure the evidence supplied by graph context, optionally combined with structural scores. The paper establishes a ranking justification under specified distributional assumptions and evaluates the method on link prediction, drug–target interaction, and node-ranking tasks, with additional ablations and supervision controls.

Critical strengths

The paper provides empirical evidence that positive-only conditional modeling can support useful graph rankings. The approach avoids assigning negative training labels to unobserved candidates. Experiments across link prediction and drug–target interaction, together with component ablations, demonstrate competitive performance among positive-only methods and provide evidence for the usefulness of the conditional-to-background scoring principle.

Critical weaknesses

W1. The demonstrated scope is limited to ranking, and broader applicability is not established.

The paper adopts the broad framing of “positive-only graph learning,” while its demonstrated capabilities concern relation and node ranking. Ranking is a valid objective for link discovery, but the contribution remains narrower than a general framework for learning predictive decisions from positive-only graph supervision. I find this limitation particularly relevant to the node setting, where the experiments establish ranking performance rather than a complete classification procedure.

The density-based scores are not directly calibrated positive-class probabilities. The paper does not provide or evaluate a procedure for converting these scores into binary decisions through threshold selection or probability calibration. Consequently, the reported ranking results do not establish performance in applications requiring explicit positive/negative decisions or controlled false-positive and false-negative rates.

The practical positive-only scope is further restricted by model selection. The main protocol uses validation positives and sampled non-edges to select configurations. Although this dependence is explicitly disclosed, Appendix T shows that selection using positive-only statistics reduces average performance. This leaves the applicability of the full pipeline less established in settings where reliable negative validation information is unavailable.

W2. The methodological contribution beyond existing components could be more clearly articulated.

GO-Flow builds on established conditional flow matching and incorporates a correction from Contrastive Flow Matching, with graph representations providing the conditioning information. This is a reasonable adaptation, although the paper could more clearly identify the graph-specific methodological insights introduced by this integration.

The conditional-to-background contrast offers a useful ranking principle related to existing likelihood-ratio approaches. Proposition 1 supports its interpretation under stated assumptions. A clearer distinction between the inherited principles and the new insights for positive-only graph ranking would help readers assess the contribution.

The direct-estimator comparisons in Appendix S support the flow model’s effectiveness on six of eight datasets. Since the simpler alternatives remain stronger on two datasets and offer substantially faster scoring, a more explicit discussion of the accuracy–cost trade-off would strengthen the motivation for choosing flow matching.

Questions for authors

Could you clarify the intended boundaries of “positive-only graph learning” in this work? Are binary node or relation decisions within the intended scope, and what additional information or assumptions would be required to obtain such decisions from the proposed scores?

What is the principal graph-specific methodological contribution beyond conditional flow matching and likelihood-ratio scoring? In light of Appendix S, what justifies using the flow model over substantially faster direct estimators, and in which regimes is that trade-off most compelling?

8 comments
AuthorsOpenReview ↗

Authors Responses (Part 1)

Thank you for your detailed and actionable review. We are glad that you found that the paper "provides empirical evidence that positive-only conditional modeling can support useful graph rankings", that the approach "avoids assigning negative training labels to unobserved candidates", and that the experiments "provide evidence for the usefulness of the conditional-to-background scoring principle". We hope that you find our responses satisfactory, and that you will consider revising your score.


In short: your review asks what exactly the paper contributes: the scope of positive-only graph learning beyond ranking, and the graph-specific contribution of GO-Flow beyond conditional flow matching and likelihood ratios. Following it, we:

  1. stated the scope as ranking, and added a thresholding procedure that targets a chosen recall from observed positives and a calibration map fitted on them;
  2. measured a configuration fixed in advance, a positive-only selection criterion, and contaminated validation;
  3. revised the contribution paragraph to separate inherited and new components;
  4. added an accuracy–cost discussion of the flow and the direct estimators.

Regarding the scope of positive-only graph learning (W1, Q1). Thank you for asking us to state the boundaries of the setting. The intended scope is ranking graph objects, relations and nodes, from observed positives; following your comment, the revised conclusion states that the demonstrated capability is ranking, and the title keeps "positive-only graph learning" as the name of the supervision regime: the trainable parameters are fit from observed positives only. Binary decisions build on these rankings: the next block gives what they need and what we added, for relations and for the node tasks of Appendix J.

Regarding decisions and calibration from observed positives (W1, Q1). We appreciate this question on thresholds and calibration. Under its assumptions, Proposition 1 gives the posterior log-odds as the score plus the constant log{πζ_w/(1 − π)}, which is log(π/(1 − π)) at w = 1. A probability therefore needs the class prior π, which observed positives do not determine; positive–unlabeled learning estimates the corresponding constant from unlabeled data [1]. To address your question directly, we fitted a calibration map from observed positives, a temperature and a bias on held-out training positives and their recombination. Recombining positives draws pairs from the model of non-relations in Proposition 1 at w = 1. Therefore, given the class prior π, the fitted map estimates the posterior probability of a relation under that model, and we measured its expected calibration error (ECE; Table 35):

GraphECE with the true prior, before calibrationECE with the true prior, after calibrationFitted temperature
Cora0.3320.0406.3
CiteSeer0.3280.0498.9
UMLS0.1210.0272.1
WN18RR0.4820.05685
DISEASES0.0620.0284.7
Drug–Disease0.7490.000-780
PPI0.3720.001-4,000
Ohmnet0.2530.0772,900

On Cora, CiteSeer, UMLS and DISEASES, the map lowers ECE to 0.03–0.05. On WN18RR, Drug–Disease, PPI and Ohmnet the fitted map is flat or nearly flat: it returns almost the same probability for every pair, and its ECE is that of a nearly constant prediction. What observed positives do support is a threshold for a target recall, which needs no class prior. Where held-out validation positives and test positives score alike, the threshold at the q-quantile of validation-positive scores gives a test recall close to 1 − q. At targets of 0.98, 0.95, 0.90, 0.80 and 0.50, the test recall is within 0.03 of the target on seven of the eight link datasets (Table 34; on DISEASES the test recall is 0.07 to 0.21 below targets of 0.95 to 0.50) and, with the same procedure on the two node tasks of Appendix J (reported here), in nine of the ten task–target cells (the tenth, CiteSeer at 0.80, gives 0.756). At a target of 0.90:

Target recall 0.90 (Table 34; nodes reported here)CoraCiteSeerUMLSWN18RRDISEASESDrug–DiseasePPIOhmnetCora (nodes)CiteSeer (nodes)
Test recall0.9000.8960.9050.8770.8160.9050.9120.9010.8760.883
False-positive rate0.3270.2560.3320.7020.1190.8040.4610.6090.2350.209

The false-positive rate depends on the negatives, and controlling it needs unlabeled or negative pairs at evaluation time. In the revision, Appendix Z describes the thresholding procedure and this boundary, and Section 5 states what observed positives determine.

AuthorsOpenReview ↗

Authors Responses (Part 2)

Regarding model selection without reliable validation negatives (W1). We value the point on the practical scope of the full pipeline. We added three measurements of how much it depends on negative validation information. First, one configuration fixed in advance, the configuration of Appendix C fused with the positive-only structural encoding (PSE), uses no validation non-edge anywhere (Table 25):

DatasetPSE-only (Table 1)GO-Flow, per-dataset selection (Table 1)GO-Flow, fixed configuration (Table 25)Difference
Cora0.6950.9130.855-0.058
CiteSeer0.5920.9350.798-0.137
UMLS0.8560.9370.873-0.064
WN18RR0.7350.7510.691-0.060
DISEASES0.7080.9200.879-0.041
Drug–Disease0.8000.7970.747-0.050
PPI0.8840.8890.812-0.077
Ohmnet0.8700.8940.8940.000
Mean0.7680.8800.819-0.061

It averages 0.819 ROC-AUC, compared with 0.880 with per-dataset selection and 0.768 for PSE-only, the strongest positive-only baseline of Table 1 on average. Second, a criterion built only from observed positives, the ROC-AUC with which validation positives rank above recombined validation positives, selects every setting of the flow without the PSE. Its choice is 0.011 ROC-AUC below the choice made with validation non-edges on average (higher on Ohmnet and UMLS, 0.055 lower on WN18RR and 0.095 lower on PPI), compared with 0.065 for the positive-only criteria of Appendix T without the PSE (0.049 with it). Third, with the validation non-edges contaminated at the training rate, GO-Flow retains at least 97% of its ROC-AUC at 40% contamination on all five graphs; the table shows GO-Flow, BUDDY and NCNC:

Retention at 40% contamination, validation contaminated (Table 26)CoraCiteSeerPPIWN18RROhmnet
BUDDY (trained with negatives)0.9630.9750.9340.8790.974
NCNC (trained with negatives)0.9530.9740.9980.9450.994
GO-Flow (observed positives only)0.9971.0000.9971.0010.972

Following your comment, Sections 3.2 and 4.4 now state this dependence with practical guidance, and Appendix T reports Tables 25 and 26.

Regarding the graph-specific contribution (W2, Q2). We appreciate the request to separate the inherited components from the new ones. In the revision, the contributions in Section 1 make this separation. Inherited: conditional flow matching as the density model [2], the contrastive field correction [3], conditioning dropout from classifier-free guidance [4], and likelihood ratios to a background model [5]. New for graphs: (i) the cross-endpoint construction, in which the target is the attributes of one endpoint and the context is the graph representation of the other, averaged over both directions, which makes a conditional density model usable on candidate pairs; (ii) the null-context density of the same flow as the background of the ratio, which needs no second model; (iii) the finding that density quality and ranking are different objectives on graphs: minibatch optimal transport raises held-out log-likelihood on all five tested graphs but lowers ROC-AUC on four (Section 4.4; Figure 4 in Appendix R); (iv) computing and rank-fusing standard structural features without any negative label (Appendix D). The principal graph-specific contribution is the cross-endpoint construction with a null context on the same flow.

AuthorsOpenReview ↗

Authors Responses (Part 3)

Regarding inherited principles and new insights (W2). We are grateful for this question on the relation to likelihood-ratio approaches. The likelihood ratio to a background model comes from out-of-distribution detection [5], where the background is a separate model trained on perturbed inputs. For positive-only graph ranking, the background is the positive target marginal under the same flow, obtained from the null context, and the contrast asks whether the graph context makes a target more plausible than it is on its own. Proposition 1 then states the model of non-relations under which ranking candidate relations by this contrast, computed with exact densities, is Bayes-optimal, with w interpolating between two models of a non-relation. As we report in Table 6, subtracting this background improves ROC-AUC on all five datasets that we tested:

ROC-AUC (Table 6)CoraUMLSDISEASESPPIDrug–Disease
Conditional log-density, w = 00.8030.8760.9080.7360.561
Full contrast, w = 10.9110.9150.9150.8080.622
Contrast, selected w0.9110.9080.9120.8230.640

Following your comment, the contributions in Section 1 state this distinction, and Section 3.2 presents Proposition 1 as the interpretation of the score under its assumptions.

Regarding the accuracy and cost of the flow (W2, Q2). Thank you for this question on what justifies the flow over faster direct estimators. Table 24 (Table 22 in the original submission) compares the flow score with the bilinear and MLP estimators of Appendix S, which estimate the full contrast (w = 1) by classification:

ROC-AUCCoraCiteSeerUMLSWN18RRDISEASESDrug–DiseasePPIOhmnet
Bilinear0.8170.8050.9850.6710.6780.6320.6760.617
MLP0.8110.9180.9740.7150.7820.6280.7110.635
GO-Flow without the PSE (α = 0)0.9130.9350.9110.6560.9200.6390.8290.746

The flow score is higher on six of the eight graphs, by 0.11–0.14 on DISEASES, PPI and Ohmnet, while the direct estimators are higher on the two knowledge graphs, UMLS and WN18RR, and score 100,000 candidates in 0.03–0.07 s, compared with 9.68 s. The preferable estimator is a property of the graph that validation can decide: a direct estimator where it matches the flow on validation, and the flow elsewhere, at a reduced budget. Five integration steps instead of twenty run 3.3–4.0× faster for a median ROC-AUC decrease of 0.007 (Appendix Y). Following your suggestion, Section 4.4 now discusses this trade-off with Figure 5 and states this guidance.


Thank you again for the precise and constructive review, which helped us state the scope and the contribution of the paper. We hope that the decision procedures, the model-selection measurements and the revised contributions answer your questions, and that you will consider revising your score. We are happy to discuss any remaining questions.


References:

[1] Learning classifiers from only positive and unlabeled data

[2] Flow Matching for Generative Modeling

[3] Contrastive Flow Matching

[4] Classifier-Free Diffusion Guidance

[5] Likelihood Ratios for Out-of-Distribution Detection

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

This paper proposes GO-Flow, a positive-only graph learning framework based on conditional flow matching. The method learns from observed positive examples without using unlabeled or negative samples for training. It ranks candidate relations by comparing their conditional and background densities, optionally combined with structural features. Experiments on link prediction, drug–target interaction, and node ranking demonstrate promising performance against positive-only baselines.

Critical strengths

The paper addresses a meaningful problem in graph learning, particularly in applications where negative labels are difficult to obtain or unreliable.

The conditional-to-background scoring strategy is intuitive and supported by theoretical analysis.

The experiments cover diverse datasets and include useful ablation studies to examine the contributions of different components.

The paper is clearly written and easy to follow.

Critical weaknesses

While GO-Flow performs well among positive-only methods, it generally underperforms approaches using negative or unlabeled samples on the same prediction tasks. It would be helpful to further clarify when positive-only learning is preferable to these alternatives.

Although training uses only positive examples, model selection still relies on sampled validation non-edges. This somewhat limits the practical advantages of the positive-only setting.

GO-Flow shows weaker performance under hard-negative evaluation, like Section 4.4. Further analysis of its ability to distinguish challenging candidate relations would strengthen the empirical evaluation.

Flow-based scoring is substantially slower than direct estimators. A more detailed discussion of the performance–efficiency trade-off would be valuable.

Theoretical assumptions (Section 3.2 and Appendix B). The optimality result relies on assumptions about the distribution of non-relations. It would be helpful to discuss how well these assumptions reflect real-world graph prediction scenarios.

Questions for authors

Please refer to the weaknesses.

4 comments
AuthorsOpenReview ↗

Authors Responses (Part 1)

Thank you for the careful reading. We are glad that you found that the paper addresses "a meaningful problem in graph learning", that the conditional-to-background scoring strategy is "intuitive and supported by theoretical analysis", that the experiments "cover diverse datasets and include useful ablation studies", and that the paper is "clearly written and easy to follow". We hope that you find our responses satisfactory and that you will consider revising your score.


In short: your review asks when positive-only learning with GO-Flow is preferable, and what supports its assumptions and its cost. Following it, we:

  1. measured robustness to false negatives in validation and on structurally plausible pairs;
  2. measured a configuration fixed in advance and added a positive-only selection criterion;
  3. added a degree-aware evaluation on all eight link datasets;
  4. added an accuracy–cost discussion of the flow and the direct estimators;
  5. stated Proposition 1 as an interpretation of the score under its assumptions.

Regarding the settings that favor positive-only learning (W1). Thank you for raising when positive-only learning is preferable. GO-Flow is designed for settings where the missing pairs cannot be trusted as negatives, as in discovery: its training data do not change when hidden positives are labeled negative. Under the standard protocol of Table 4, GO-Flow is within 0.024–0.057 ROC-AUC of the strongest method trained with sampled negatives or unlabeled pairs on five of six datasets and reaches 0.751 on WN18RR, and it is above all three positive–unlabeled (PU) methods on Cora and CiteSeer, above SEAL on Cora and CiteSeer, and above BUDDY on PPI and Ohmnet:

ROC-AUC (Table 4)CoraCiteSeerUMLSWN18RRPPIOhmnet
Naive PU0.7640.7790.9850.7440.8880.904
nnPU0.7370.7430.9640.6660.8880.897
Bagging-PU0.8450.8580.9940.7550.9080.915
SEAL0.8890.8610.9870.9300.9130.933
BUDDY0.9450.9640.9860.7770.8360.884
LPFormer0.9170.9460.9640.8270.8820.903
NCNC0.9650.9660.9700.8370.8960.928
GO-Flow (observed positives only)0.9130.9350.9370.7510.8890.894

The contamination study of Section 4.3 measures the other regime: when 40% of the sampled negatives are hidden true edges, PULL with a frozen negative pool loses 14% to 18% on Cora and CiteSeer, BUDDY 6.3% on PPI and LPFormer 5.4% to 7.5% on Cora, CiteSeer and PPI, NCNC and SEAL are less sensitive, and the training data of GO-Flow are unchanged (Table 15). We also contaminated the validation non-edges at the training rate (GO-Flow, BUDDY, NCNC):

Retention at 40%, validation contaminated (Table 26)CoraCiteSeerPPIWN18RROhmnet
BUDDY (trained with negatives)0.9630.9750.9340.8790.974
NCNC (trained with negatives)0.9530.9740.9980.9450.994
GO-Flow (observed positives only)0.9971.0000.9971.0010.972

GO-Flow retains at least 99.7% of its ROC-AUC on Cora, CiteSeer, PPI and WN18RR and 97.2% on Ohmnet. In a real graph, missing relations can concentrate on structurally plausible pairs. We therefore placed the same number of false negatives on the hidden edges with the highest Adamic–Adar index, with every other input identical (Table 16):

Average precision at 20%, localized minus uniformCoraCiteSeerPPIWN18RROhmnet
SEAL-0.038-0.029-0.150-0.050-0.098
BUDDY0.000+0.002-0.019-0.008-0.017
NCNC-0.006-0.002-0.041-0.012-0.063
LPFormer-0.006+0.003-0.084-0.017-0.082
GO-Flow (reads no negatives)00000

On PPI, Ohmnet and WN18RR, the sampled-negative methods lose more average precision than under uniform placement in all three seeds in 33 of 36 method-level cells up to 20% contamination (in two seeds in the other three, on WN18RR), by up to 0.150 (SEAL on PPI); on Cora and CiteSeer the changes are within 0.05, and GO-Flow is unchanged under both placements. The test positives that resemble the false negatives lose rank: for the third with the highest Adamic–Adar index, ROC-AUC is lower under localized placement in all 36 cells up to 20% contamination on PPI, Ohmnet and WN18RR, by up to 0.087; for the lowest third it changes by +0.004 on average (Appendix M, Table 17):

ROC-AUC of the highest third at 20%, localized minus uniformCoraCiteSeerPPIWN18RROhmnet
SEAL-0.062-0.029-0.051-0.016-0.047
BUDDY-0.007-0.002-0.024-0.019-0.016
NCNC-0.024-0.005-0.013-0.010-0.021
LPFormer-0.010-0.003-0.087-0.010-0.045

Following your comment, Section 4.3 now states when each supervision regime is preferable.

AuthorsOpenReview ↗

Authors Response (Part 2)

Regarding model selection without validation non-edges (W2). We appreciate this point on the practical scope of positive-only training. To measure how much the results depend on selection, we evaluated one configuration fixed in advance, the configuration of Appendix C (GCN, FiLM, cross-endpoint conditioning, the full contrast, no reflow, fusion weight 0.5 with the positive-only structural encoding, PSE), on all eight link datasets with no validation non-edge anywhere (Table 25):

ROC-AUCCoraCiteSeerUMLSWN18RRDISEASESDrug–DiseasePPIOhmnetMean
PSE-only (Table 1)0.6950.5920.8560.7350.7080.8000.8840.8700.768
GO-Flow, per-dataset selection (Table 1)0.9130.9350.9370.7510.9200.7970.8890.8940.880
GO-Flow, fixed configuration (Table 25)0.8550.7980.8730.6910.8790.7470.8120.8940.819

It averages 0.819 ROC-AUC, compared with 0.880 for per-dataset selection and 0.768 for PSE-only, the strongest positive-only baseline of Table 1 on average. In addition, we introduce a criterion built only from observed positives: the ROC-AUC with which validation positives rank above recombined validation positives (the target of one pair with the context of another). It selects every setting of the flow without the PSE, and its choice is 0.011 ROC-AUC below the choice made with validation non-edges on average: higher on Ohmnet and UMLS, and lower by 0.055 on WN18RR and by 0.095 on PPI. In the revision, Sections 3.2 and 4.4 state this dependence with practical guidance: without reliable validation negatives, use the fixed configuration, which needs no selection criterion and is 0.061 lower on average; with a small set of trusted non-edges, select on them.

Regarding ranking challenging candidates (W3). Thank you for asking how GO-Flow distinguishes challenging candidate relations. Beyond the HeaRT evaluation of Section 4.3 [1], where negatives are chosen to resemble true edges, which methods trained with sampled negatives can learn to reject, we added a degree-aware evaluation on all eight link datasets with the splits, configurations and seeds of the paper: degree-corrected non-edges, whose endpoints are drawn in proportion to their training degree [2, 3], degree-matched non-edges in the same log2-degree bins as each positive, and ROC-AUC stratified by degree (Appendix Z, Table 31):

ROC-AUCCoraCiteSeerUMLSDISEASES
PSE-only, degree-corrected non-edges0.5400.4590.7590.530
GO-Flow without the PSE, standard non-edges (Table 21, Table 19 in the original submission)0.9130.9350.9110.920
GO-Flow without the PSE, degree-corrected non-edges0.8670.9030.8400.915

On these four graphs, the flow decreases by 0.005 to 0.071 under degree correction, while PSE-only decreases to 0.459–0.759; on the lowest-degree tertile of Cora positives and the standard non-edges, the flow reaches 0.831, compared with 0.408 for PSE-only. On WN18RR, Drug–Disease, PPI and Ohmnet, GO-Flow and PSE-only reach 0.505–0.630 and the parameter-free baselines 0.500–0.666. Appendix Z of the revision reports this evaluation on all eight graphs.

AuthorsOpenReview ↗

Authors Responses (Part 3)

Regarding the accuracy and cost of the flow (W4). We appreciate this question on the trade-off between performance and efficiency. Table 24 (Table 22 in the original submission) compares the flow score with the bilinear and MLP estimators of Appendix S, which estimate the full contrast (w = 1) by classification:

ROC-AUCCoraCiteSeerUMLSWN18RRDISEASESDrug–DiseasePPIOhmnet
Bilinear0.8170.8050.9850.6710.6780.6320.6760.617
MLP0.8110.9180.9740.7150.7820.6280.7110.635
GO-Flow without the PSE (α = 0)0.9130.9350.9110.6560.9200.6390.8290.746

The flow score is higher on six of the eight graphs, by 0.11–0.14 on DISEASES, PPI and Ohmnet, and the direct estimators are higher on UMLS and WN18RR; they score 100,000 candidates in 0.03–0.07 s, compared with 9.68 s for the flow. The cost of the flow can be reduced at evaluation time without retraining: five integration steps instead of twenty run 3.3–4.0× faster for a median ROC-AUC decrease of 0.007 (Appendix Y). Following your suggestion, Section 4.4 now discusses this trade-off with Figure 5 and gives the guidance: where a direct estimator matches the flow on validation, use it; otherwise use the flow with the reduced budget.

Regarding the assumptions of Proposition 1 on real graphs (W5). We are grateful for the question on how well these assumptions reflect real graphs. Proposition 1 assumes that the observed positives are selected completely at random, and that a non-relation pairs an independent context with a target drawn from the positive target marginal raised to the power w. The two ends of the range are two models of a non-relation: at w = 1, its target is distributed like a positive target, and the dependence on the context separates the classes; at w = 0, its target is flat, and typical targets are also rewarded. On real graphs the target distribution of non-relations is unknown, for example when candidate endpoints are less typical than observed ones, which is why w is selected on validation rather than derived (Appendix B). Table 6 shows the effect of w:

ROC-AUC (Table 6)CoraUMLSDISEASESPPIDrug–Disease
Conditional log-density, w = 00.8030.8760.9080.7360.561
Full contrast, w = 10.9110.9150.9150.8080.622
Contrast, selected w0.9110.9080.9120.8230.640

The full contrast improves on the conditional log-density on all five datasets, and the selected w improves further on PPI and Drug–Disease. The model of a non-relation in Proposition 1 is also approximate on real graphs, because the non-relations a ranker meets share endpoints, and therefore contexts, with true relations (Appendix B). Following your comment, Section 3.2 now presents Proposition 1 as an explanation of the score under these assumptions, and Appendix B discusses how well they reflect real graphs.


Thank you again for the detailed review. We hope that the added measurements of robustness, model selection, degree-aware ranking and cost address your concerns, and that you will consider revising your score. We are happy to discuss any remaining questions.


References:

[1] Evaluating Graph Neural Networks for Link Prediction: Current Pitfalls and New Benchmarking

[2] Implicit degree bias in the link prediction task

[3] Bias-aware training and evaluation of link prediction algorithms in network biology

Dear Reviewer XgKi,

Thank you for your careful and constructive review. Your comments led directly to several substantial additions to the paper. With the discussion closing shortly, we wanted to briefly summarize our responses that address each of your comment:

  • Positive-only model selection: one configuration fixed in advance across all 8 link datasets, requiring no validation non-edge, reaches 0.819 mean ROC-AUC vs. 0.768 for the strongest positive-only baseline. Our new positive-only selection criterion is only 0.011 ROC-AUC below selection with validation non-edges on average.
  • False negatives: we added contaminated-validation experiments on 5 graphs and structurally localized false negatives. At 40% contamination, GO-Flow retains at least 97.2% of its ROC-AUC on all 5 graphs and at least 99.7% on 4/5.
  • Difficult candidates: we added degree-corrected, degree-matched, and degree-stratified evaluation across all 8 link datasets, and re-evaluated 4 negative-trained baselines on the same degree-aware negatives.
  • Accuracy–cost trade-off: the flow outperforms the better of 2 direct estimators on 6/8 datasets, with gains of 0.11–0.14 ROC-AUC on DISEASES, PPI and Ohmnet. Using 5 rather than 20 integration steps gives a 3.3–4.0× speedup for only 0.007 median ROC-AUC decrease.
  • We also revised the discussion of when positive-only learning is preferable and clarified the assumptions and scope of Proposition 1.

We are very grateful for your comments, which we feel strengthened our paper.

We hope you find our responses satisfactory, and that you will consider revising your score. We are also happy to address any remaining questions.

Thank you again for your time and careful review.

Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response (Part 1)

We thank all reviewers for their careful reading and constructive feedback.

The reviewers rated the presentation "Clear and easy to follow" (XgKi, Erxd, XTCb) and found that the paper addresses "a meaningful problem in graph learning" (XgKi), with a motivation that is "both practical and well grounded" (XTCb). They noted that the approach "avoids assigning negative training labels to unobserved candidates" (Erxd), and described the conditional-to-background scoring as "intuitive and supported by theoretical analysis" (XgKi) and "a useful ranking principle" (Erxd), and the distinction it draws as "a meaningful distinction" (XTCb). They found that the experiments "cover diverse datasets and include useful ablation studies" (XgKi), and appreciated "the effort to examine the method's limitations", which makes the empirical analysis "more credible" (XTCb).

The reviews share one question: how well GO-Flow ranks when no reliable negatives are available at any stage, from model selection to hard candidates, and when its flow is worth its cost. Following the reviews, we added the following results and analyses.

  • Model selection without validation non-edges (XgKi, Erxd, XTCb). One configuration fixed in advance, the configuration of Appendix C, averages 0.819 ROC-AUC on the eight link datasets with no validation non-edge, compared with 0.880 with per-dataset selection and 0.768 for the positive-only structural encoding (PSE-only), the strongest positive-only baseline of Table 1 on average (Table 25). A criterion built only from observed positives, the ROC-AUC with which validation positives rank above recombined validation positives, selects every setting of the flow without the PSE. Its choice is 0.011 ROC-AUC below the choice made with validation non-edges on average.
  • False negatives in training and validation (XgKi, Erxd, XTCb). With the validation non-edges contaminated at the training rate, GO-Flow retains at least 99.7% of its ROC-AUC at 40% contamination on Cora, CiteSeer, PPI and WN18RR and 97.2% on Ohmnet, while BUDDY retains 87.9% to 97.5% and NCNC 94.5% to 99.8% (Table 26). With the false negatives placed on the hidden edges with the highest Adamic–Adar index, the sampled-negative methods lose up to 0.150 more average precision than under uniform placement, while GO-Flow is unchanged (Table 16). The test positives with the highest Adamic–Adar index lose rank: their ROC-AUC is lower under localized placement in all 36 method-level cells up to 20% contamination on PPI, Ohmnet and WN18RR (Appendix M, Table 17).
  • Nested selection for drug–target interaction (XTCb). With the configuration re-selected inside every cross-validation fold, where no test pair informs any choice, GO-Flow reaches 0.963, 0.966 and 0.954 on Enzyme, Ion channel and GPCR, within 0.005 of the submitted protocol and above every published positive–unlabeled (PU) method, and 0.854 on Nuclear receptor, the smallest family; it remains 0.115–0.165 above the strongest positive-only baseline on all four (Tables 2 and 13).
  • Degree-aware evaluation (XgKi, XTCb). Following the studies that XTCb pointed to [1, 2], we scored GO-Flow on degree-corrected and degree-matched non-edges on all eight link datasets, and SEAL, BUDDY, NCNC and LPFormer on the same non-edges of the six graphs of Table 4. On degree-corrected non-edges, the flow reaches 0.867, 0.903, 0.840 and 0.915 on Cora, CiteSeer, UMLS and DISEASES, where PSE-only reaches 0.459–0.759, and GO-Flow reaches 0.811–0.945 on the four drug–target families (Appendix Z, Tables 31 to 33). On WN18RR, Drug–Disease, PPI and Ohmnet, GO-Flow and PSE-only reach 0.505–0.630 and the parameter-free baselines 0.500–0.666.
  • Decisions from observed positives (Erxd). A threshold at a quantile of validation-positive scores meets the target recall on test within 0.03 on seven of the eight link datasets (Table 34) and in nine of ten cells on the two node tasks, and, with the known class prior, a temperature and bias fitted on held-out positives and their recombination lower the expected calibration error (ECE) from 0.06–0.33 to 0.03–0.05 on Cora, CiteSeer, UMLS and DISEASES (Table 35).
  • Accuracy and cost of the flow (XgKi, Erxd, XTCb). Section 4.4 now discusses the trade-off with Figure 5: the flow score is higher than the better direct estimator on six of eight graphs, by 0.11–0.14 on DISEASES, PPI and Ohmnet, while the direct estimators score 100,000 candidates in 0.03–0.07 s, compared with 9.68 s, and are higher on UMLS and WN18RR (Table 24, Table 22 in the original submission). Five integration steps instead of twenty run 3.3–4.0× faster for a median ROC-AUC decrease of 0.007 (Appendix Y).
AuthorsOpenReview ↗

General Response (Part 2)

Added material. All additions are marked in blue in the revised PDF; a result listed without a location is reported in the responses.

  • Section 1: the contribution paragraph separates inherited and new components (Erxd).
  • Sections 3.2, 4.4 and 5: the dependence of model selection on validation non-edges, with practical guidance (XgKi, Erxd, XTCb).
  • Section 3.2 and Appendix B: Proposition 1 as the interpretation of the score under its assumptions (XgKi, Erxd, XTCb).
  • Section 4.3: when each supervision regime is preferable (XgKi).
  • Section 4.4 and Figure 5: the accuracy–cost trade-off, with guidance (XgKi, Erxd, XTCb).
  • Section 5: the scope of ranking, and decisions at a target recall (Erxd).
  • Section 3, Appendix C and Figure 2: an architecture diagram, and the interpretation of target-edge masking (XTCb).
  • Sections 4.2 and 4.4, Appendices E, J and K, Tables 2 and 13: the nested protocol for drug–target interaction (XTCb).
  • Appendix M and Tables 16 and 17: false negatives on structurally plausible pairs, and the test positives that lose rank (XgKi).
  • Appendix T and Tables 25 and 26: the fixed configuration and contaminated validation (XgKi, Erxd, XTCb).
  • Section 4.3, Appendix Z and Tables 31 to 35: the degree-aware evaluation, the decision thresholds and the calibration map (XgKi, Erxd, XTCb).
  • Figures 3 to 5: contamination, density quality and ranking, accuracy and cost (XTCb).

We appreciate the constructive reviews, which strengthened the paper. We hope that the added results and revisions address your questions, and that you will consider revising your scores. We are happy to discuss any remaining questions.


References:

[1] Implicit degree bias in the link prediction task

[2] Bias-aware training and evaluation of link prediction algorithms in network biology

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.