acceptodds
Under review as a conference paper at ICLR 2027

Partial AUC Maximization from Positive-unlabeled Data

Abstract

The partial area under the receiver operating characteristic curve (pAUC) is an important performance metric for binary classification that summarizes true positive rates within a specific range of false positive rates (FPRs). Classifiers that achieve high pAUC need to be obtained in many real-world applications such as cybersecurity, medical care, and advertising. Although many methods for maximizing the pAUC have been proposed, they typically require both labeled positive and negative data for training. However, in practice, labeled negative data are often difficult to collect due to privacy concerns or the need for high expertise to annotate them. In this paper, we propose a method for maximizing the pAUC from positive and unlabeled (PU) data without negative data. Within an empirical risk minimization framework, we show that the pAUC, including its FPR-dependent thresholds, can be represented using only the positive and marginal densities, and derive an empirical estimator from PU data. A classifier is then trained by maximizing the derived smoothed empirical pAUC estimator. We experimentally demonstrate the effectiveness of the proposed method with ten real-world datasets.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Partial AUC maximization, PU learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gTiBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for maximizing partial AUC from positive and unlabeled data by reformulating the pAUC and its FPR dependent thresholds using positive and marginal distributions. It derives an empirical estimator, introduces a smoothed objective for classifier training, and establishes finite sample estimation bounds. Experiments on ten datasets show competitive performance against existing PU learning, AUC maximization, and pAUC maximization methods. However, the theoretical assumptions and the accuracy of threshold estimation in the low FPR regime require further examination.

Critical strengths
  1. The paper addresses the important problem of pAUC maximization without labeled negative data. Its reformulation avoids the heuristic pseudo negative assumption used by an existing PU based pAUC method.

  2. The paper provides a finite sample analysis of threshold estimation and pAUC estimation, including the effect of class prior estimation errors.

  3. The experiments cover multiple image and tabular datasets, different numbers of labeled positives, different class priors, and alternative surrogate functions.

Critical weaknesses
  1. The main methodological contribution is the reformulation of pAUC and its FPR-dependent thresholds using the positive and marginal distributions. However, the subsequent learning procedure relies on a smoothed pairwise ranking objective and score-based interval selection, which are also used in prior pAUC maximization methods. Although the proposed formulation avoids the heuristic pseudo-negative selection used by existing PU-based pAUC methods, the paper does not sufficiently clarify the methodological novelty beyond this difference. What are the substantive technical advances over existing PU risk rewriting and pAUC maximization approaches, and how do they contribute to the observed performance gains?

  2. The theoretical analysis relies on the assumption that the negative score distribution has no point masses, which ensures the continuity of the FPR and supports the threshold estimation arguments. However, the applicability of this assumption to the evaluated image datasets, where tied scores may occur, is not sufficiently discussed. Moreover, the finite-sample analysis primarily establishes estimation error bounds rather than guarantees on the excess pAUC of the classifier obtained through the proposed optimization procedure. How can the authors justify or relax the no-point-masses assumption, and what guarantees can be established for the actual learning procedure beyond estimator consistency?

  3. The threshold analysis in Appendix C.7 raises concerns about whether the proposed method accurately optimizes the intended low-FPR region. For example, when the target FPR is zero, the actual FPR induced by the estimated threshold reaches approximately 0.03–0.06 in several settings. Such deviations are particularly concerning when labeled positives are scarce, as they may cause the effective training interval to differ substantially from the intended interval. Can the authors quantify how these threshold estimation errors affect the final test pAUC and demonstrate that the proposed method remains effective when the target FPR is very low?

  4. Although the proposed method achieves competitive results across multiple datasets, its superiority over the strongest baselines is not consistent across all settings. The paper relies on paired t-tests to identify methods that are statistically indistinguishable from the best-performing method, but it does not discuss the potential impact of conducting multiple significance tests across numerous experimental settings. Can the authors provide further evidence of the robustness of the reported improvements, including an assessment of whether the conclusions remain unchanged after appropriate multiple-comparison correction?

Questions for authors

Please refer to the Weaknesses.

5 comments

We thank the reviewer for the careful reading and constructive feedback.

  1. What are the substantive technical advances over existing PU risk rewriting and pAUC maximization approaches, and how do they contribute to the observed performance gains?

Thank you for the question. We agree that the smoothed pairwise ranking loss and score-based interval selection are established techniques in existing pAUC maximization methods. However, our key technical contribution lies in the PU-based reformulation of both the pairwise risk and the FPR-dependent thresholds required for pAUC maximization.

Existing pAUC maximization methods usually require labeled negative data to determine the FPR-dependent thresholds and calculate the pairwise risk. On the other hand, existing PU risk rewriting techniques used for classification and AUC risks do not directly address the FPR-dependent thresholds. Thus, simply combining conventional pAUC maximization with PU-based pairwise risk rewriting is insufficient to recover the target pAUC objective. In Section 4, we address this issue by reformulating both components using only the positive and marginal distributions.

Furthermore, Iwata et al. (2020) heuristically treat selected unlabeled instances as negative without applying PU risk rewriting to the pairwise risk. In contrast, our method directly reformulates both the pairwise risk and the FPR-dependent thresholds, avoiding such heuristic assumptions.

To demonstrate the importance of our PU-based threshold reformulation, we conducted an additional comparison with a variant of our method without PU-based threshold estimation (Ours w/o PU thresholds). This variant uses our PU-rewritten pairwise risk in Eq. (15), but determines the thresholds from unlabeled data by simply treating them as negative data. The average test pAUCs are as follows:

Table A: Ablation results for the FPR interval [0, 0.1]

Dataπ\piOursOurs w/o PU thresholds
MNIST0.10.683 ± 0.0740.574 ± 0.068
MNIST0.20.633 ± 0.0710.421 ± 0.058
FMNIST0.10.924 ± 0.0250.655 ± 0.086
FMNIST0.20.914 ± 0.0280.470 ± 0.084
SVHN0.10.247 ± 0.0810.210 ± 0.056
SVHN0.20.191 ± 0.0720.155 ± 0.027
CIFAR100.10.478 ± 0.0780.337 ± 0.057
CIFAR100.20.433 ± 0.0750.236 ± 0.051
SUSY0.10.454 ± 0.0340.369 ± 0.067
SUSY0.20.424 ± 0.0450.222 ± 0.060
CODRNA0.10.741 ± 0.0470.555 ± 0.078
CODRNA0.20.716 ± 0.0560.451 ± 0.073
IJCNN0.10.620 ± 0.0950.468 ± 0.070
IJCNN0.20.565 ± 0.0960.321 ± 0.069
EPSILON0.10.193 ± 0.0410.171 ± 0.036
EPSILON0.20.163 ± 0.0450.141 ± 0.028
#Best—160

Table B: Ablation results for the FPR interval [0, 0.2]

Dataπ\piOursOurs w/o PU thresholds
MNIST0.10.755 ± 0.0560.731 ± 0.058
MNIST0.20.723 ± 0.0650.610 ± 0.062
FMNIST0.10.943 ± 0.0220.902 ± 0.052
FMNIST0.20.934 ± 0.0230.700 ± 0.069
SVHN0.10.342 ± 0.0920.329 ± 0.078
SVHN0.20.289 ± 0.0900.262 ± 0.043
CIFAR100.10.610 ± 0.0740.534 ± 0.093
CIFAR100.20.578 ± 0.0750.408 ± 0.079
SUSY0.10.553 ± 0.0320.529 ± 0.044
SUSY0.20.528 ± 0.0440.441 ± 0.073
CODRNA0.10.853 ± 0.0340.798 ± 0.074
CODRNA0.20.838 ± 0.0390.650 ± 0.079
IJCNN0.10.711 ± 0.0790.665 ± 0.078
IJCNN0.20.659 ± 0.1020.506 ± 0.078
EPSILON0.10.305 ± 0.0580.293 ± 0.045
EPSILON0.20.269 ± 0.0460.248 ± 0.035
#Best—161

Our method achieved higher mean test pAUC in all 32 settings, with average improvements of 0.164 and 0.080 for the FPR intervals [0, 0.1] and [0, 0.2], respectively. These results demonstrate that PU-based threshold estimation is essential for obtaining the observed performance gains and that our contribution goes beyond standard PU-based pairwise risk rewriting. We will include these results in the revised paper.

  1. The theoretical analysis relies on the assumption that the negative score distribution has no point masses, which ensures the continuity of the FPR and supports the threshold estimation arguments. However, the applicability of this assumption to the evaluated image datasets, where tied scores may occur, is not sufficiently discussed. ... How can the authors justify or relax the no-point-masses assumption?

Thank you for raising these important questions.

First, we clarify that Assumption A.1 is introduced as a sufficient condition for our finite-sample theoretical analysis, rather than for the PU-based rewriting itself. Specifically, the PU-based rewriting of the pairwise expression and FPR in Eqs. (8)–(10) does not require this assumption. Assumption A.1 is used to ensure FPR continuity when bounding the errors at the estimated thresholds.

To examine the practical relevance of this assumption, we additionally analyzed negative-score ties across 480 trained models on four image datasets, using 2,500 negative test instances per model. No ties were observed in 404 of the 480 models (84.2%). For each dataset, the mean pairwise tie rate, defined as the average proportion of unordered pairs of negative test instances with identical scores across its 120 models, was below 10−710^{-7}. The largest observed tied-score group contained only two instances (0.08% of the negative test set). These results indicate that score ties were extremely rare in our experiments.

Furthermore, the threshold estimation bound in Theorem A.3 can be extended to allow point masses. Let ms=sup⁡tP(s(xn)=t)m_s=\sup_t P(s(\mathbf{x}^{\rm n})=t) denote the maximum point mass of the negative-score distribution. Using this quantity, the bound in Theorem A.3 can be extended to

\leq\epsilon_{\mathrm{FPR}}+m_s.$$ Specifically, in the proof of Theorem A.3, Eq. (50) remains valid without Assumption A.1. Taking $t\uparrow\hat t_\gamma$ in Eq. (50), we obtain $$\mathrm{FPR}(\hat t_\gamma^-) \geq\gamma-\epsilon_{\mathrm{FPR}},$$ where $\mathrm{FPR}(\hat t_\gamma^-)$ denotes the left limit of $\mathrm{FPR}(t)$ at $\hat t_\gamma$. Since $$\mathrm{FPR}(\hat t_\gamma^-) -\mathrm{FPR}(\hat t_\gamma) =P(s(\mathbf{x}^{\rm n})=\hat t_\gamma) \leq m_s,$$ Eq. (51) can be relaxed to $$\mathrm{FPR}(\hat t_\gamma) \geq\gamma-\epsilon_{\mathrm{FPR}}-m_s.$$ Combining this with Eq. (46) yields the extended bound above, corresponding to Eq. (52). Thus, allowing small point masses introduces only an additional error term of $m_s$ into the threshold estimation bound. > 2. The finite-sample analysis primarily establishes estimation error bounds rather than guarantees on the excess pAUC of the classifier obtained through the proposed optimization procedure. What guarantees can be established for the actual learning procedure beyond estimator consistency? Theorem A.6 establishes a uniform pAUC estimation error bound over the score-function class. Consequently, an exact maximizer of the empirical pAUC in Eq. (14) satisfies an excess pAUC bound of $2B_{N,N^p}$ relative to the best score function in the class. However, our actual learning procedure (Algorithm 1) uses the smoothed objective in Eq. (15) and mini-batch SGD. Extending the guarantee to this procedure requires controlling the surrogate approximation and optimization errors, which are not bounded by our current analysis. We will clarify this distinction and the corresponding limitation in the revised paper.
  1. The threshold analysis in Appendix C.7 raises concerns about whether the proposed method accurately optimizes the intended low-FPR region. ... Can the authors quantify how these threshold estimation errors affect the final test pAUC and demonstrate that the proposed method remains effective when the target FPR is very low?

Thank you for the insightful comment. The FPR boundary at α=0\alpha=0 corresponds to the extreme tail of the negative-score distribution, where accurate threshold estimation is particularly challenging with finite data. This explains why the deviations at α=0\alpha=0 tend to be larger than those at β=0.1\beta = 0.1 or 0.20.2. As shown in Table 9, increasing the number of labeled positive instances from 30 to 100 consistently reduces these deviations.

Regarding the effect of threshold estimation errors, Eq. (91) bounds the error in the pAUC objective caused by replacing the population thresholds with their estimated values, for a fixed score function, by 2ϵFPRβ−α.\frac{2\epsilon_{\mathrm{FPR}}}{\beta-\alpha}. Theorem A.6 further combines this threshold estimation error with the empirical pAUC estimation error.

Importantly, the test pAUCs in Tables 1 and 2 are evaluated using the true positive and negative test labeled data over the specified FPR ranges, rather than the estimated FPR thresholds. Despite the threshold estimation errors, our method achieved the highest mean test pAUC in 24 out of 32 settings and was best or not significantly different from the best in 31 out of 32 settings (Tables 1 and 2). These results demonstrate that our method remains empirically effective for maximizing the target pAUC even with finite-sample threshold estimation errors. Furthermore, the ablation study in Tables A and B (our response to Question 1) provides additional evidence of the importance of PU-based threshold estimation. With the same PU-rewritten pairwise risk, our method achieved higher mean test pAUC than the variant treating unlabeled data as negative for threshold estimation in all 32 settings.

To further evaluate the effectiveness of our method in low-FPR regions, we additionally conducted experiments on a narrower FPR interval of [0, 0.05]. The mean test pAUCs are reported in the following table.

Dataπ\piOursNNPUPUAUCPAUCPUpAUCh2step
MNIST0.10.5940 ± 0.08460.4392 ± 0.12720.5265 ± 0.06030.4376 ± 0.06030.5955 ± 0.07720.5334 ± 0.0621
MNIST0.20.5424 ± 0.08450.4713 ± 0.07550.4887 ± 0.05370.3244 ± 0.06930.5320 ± 0.07510.4863 ± 0.0562
FMNIST0.10.8979 ± 0.03150.7403 ± 0.02860.8608 ± 0.04750.6355 ± 0.08850.9040 ± 0.02930.8768 ± 0.0464
FMNIST0.20.8819 ± 0.04450.7521 ± 0.03800.8239 ± 0.06060.4972 ± 0.11320.8894 ± 0.03340.8319 ± 0.0597
SVHN0.10.1689 ± 0.05850.0287 ± 0.00380.1378 ± 0.05860.1161 ± 0.02590.1698 ± 0.06800.1398 ± 0.0599
SVHN0.20.1213 ± 0.05580.0287 ± 0.00380.1092 ± 0.04210.0847 ± 0.02580.0993 ± 0.06040.0757 ± 0.0422
CIFAR100.10.3390 ± 0.08170.0298 ± 0.02730.3279 ± 0.07640.2096 ± 0.03920.3280 ± 0.07170.3345 ± 0.0837
CIFAR100.20.3028 ± 0.08010.0436 ± 0.04830.2931 ± 0.06890.1378 ± 0.03820.2886 ± 0.06690.2945 ± 0.0697
SUSY0.10.3675 ± 0.04010.2780 ± 0.07830.3217 ± 0.06170.2042 ± 0.04110.3458 ± 0.04130.3183 ± 0.0624
SUSY0.20.3356 ± 0.04580.2861 ± 0.05970.2949 ± 0.06180.1297 ± 0.03080.2868 ± 0.04950.2900 ± 0.0629
CODRNA0.10.5837 ± 0.06030.2491 ± 0.22350.6045 ± 0.07240.4147 ± 0.07280.5877 ± 0.05970.6054 ± 0.0697
CODRNA0.20.5627 ± 0.05300.5569 ± 0.06450.5710 ± 0.07040.3226 ± 0.08170.5620 ± 0.05260.5759 ± 0.0692
IJCNN0.10.5326 ± 0.09890.1907 ± 0.10760.4437 ± 0.08680.3569 ± 0.06850.4838 ± 0.13420.4489 ± 0.0910
IJCNN0.20.4195 ± 0.10170.3508 ± 0.07910.4052 ± 0.07470.2619 ± 0.05470.3013 ± 0.17720.4003 ± 0.0759
EPSILON0.10.1183 ± 0.03270.0510 ± 0.00540.1193 ± 0.03210.0981 ± 0.02430.1173 ± 0.03040.1196 ± 0.0342
EPSILON0.20.0979 ± 0.03030.0503 ± 0.00750.1046 ± 0.02900.0805 ± 0.01670.0942 ± 0.02740.1083 ± 0.0318
Avg0.4291 ± 0.24900.2842 ± 0.25570.4021 ± 0.23760.2695 ± 0.17190.4116 ± 0.25840.4025 ± 0.2435
#Best1406097

Our method achieved the highest average test pAUC across all experimental settings (0.4291) and performed best or comparably to the best method in 14 out of 16 settings. These results further support the effectiveness of our method in low-FPR regions. We will include these results in the revised paper.

Reviewer_G7vaOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method for learning a classifier that maximizes the partial area under the ROC curve (pAUC) using only positive and unlabeled data. Specifically, the authors express the false positive rate (FPR) and pAUC risk in terms of expectations over the positive and unlabeled data distributions, thereby eliminating the need for labeled negative examples. A classifier is then trained by minimizing the corresponding empirical risk estimator. Experiments on benchmark datasets show that the proposed method significantly outperforms all five competing methods in 14 of the 32 experimental settings.

Critical strengths
  • Unlike existing heuristic approaches that treat a subset of the unlabeled examples as negative, the proposed method derives a principled objective for pAUC optimization using only positive and unlabeled data.
  • The paper establishes an estimation error bound of the pAUC risk. The bound explicitly characterizes its dependence on the width of the target FPR range and the numbers of positive and unlabeled samples.
  • The proposed method is evaluated on multiple benchmark datasets and performs favorably against existing methods. Across the 32 main experimental settings, the proposed method achieves the highest pAUC or does not differ significantly from the best-performing method in 31 settings. In 14 settings, it significantly outperforms all five competing methods.
Critical weaknesses
  • The proposed approach extends a standard idea in positive-unlabeled (PU) learning—expressing a supervised risk in terms of expectations over positive and unlabeled data—to pAUC optimization. Thus, although the extension is technically meaningful, its methodological novelty appears somewhat limited.
  • The proposed method significantly outperforms all five competing methods in 14 of the 32 settings. In the remaining settings, however, its advantage over existing methods is less clear, suggesting that the magnitude of the improvement depends on the data or evaluation setting.
Questions for authors
  • Do the authors have any insight into the dataset characteristics or experimental factors associated with significant improvements over the strongest competing method? A comparison with the settings in which no significant improvement is observed would help clarify when the proposed method is likely to provide a meaningful practical advantage.
3 comments

We thank the reviewer for the careful reading and positive feedback.

The proposed approach extends a standard idea in positive-unlabeled (PU) learning—expressing a supervised risk in terms of expectations over positive and unlabeled data—to pAUC optimization. Thus, although the extension is technically meaningful, its methodological novelty appears somewhat limited.

Thank you for recognizing the technical significance of our extension. We agree that the underlying PU risk rewriting principle is well established. However, extending it to pAUC maximization is not straightforward because, unlike standard classification and AUC risks, pAUC depends on FPR-dependent thresholds determined by the unknown negative-score distribution. Thus, rewriting the pairwise risk alone is insufficient to obtain a PU-based pAUC objective.

As detailed in Section 4, our main technical contribution is to reformulate not only the pairwise risk but also the FPR-dependent thresholds using positive and marginal distributions, enabling direct pAUC estimation from PU data without heuristic pseudo-negative selection. We further derive an empirical estimator and establish finite-sample error guarantees for its estimation (Theorem A.6). These contributions go beyond a direct application of standard PU risk rewriting.

To further demonstrate the importance of our PU-based threshold reformulation, we conducted an additional ablation that retains our PU-rewritten pairwise risk but estimates thresholds by treating unlabeled data as negative. Detailed results are provided in Tables A and B in our response to Reviewer gTiB (Question 1). Our method achieved higher mean test pAUC in all 32 experimental settings, supporting the practical importance of this contribution.

Do the authors have any insight into the dataset characteristics or experimental factors associated with significant improvements over the strongest competing method? A comparison with the settings in which no significant improvement is observed would help clarify when the proposed method is likely to provide a meaningful practical advantage.

Thank you for the question. We examined the settings with and without statistically significant improvements over the strongest baseline.

From Tables 1 and 2 and Appendix C.1, our method significantly outperformed the strongest baseline in 8 out of 16 settings for [0, 0.1] and 6 out of 16 for [0, 0.2], but in none of the 16 settings for [0.1, 0.2]. These results suggest that our method tends to be more advantageous for FPR intervals starting at 0. Since many real-world applications prioritize performance at low FPRs (Narasimhan & Agarwal, 2013a; Iwata et al., 2020), this tendency is particularly relevant in practice.

Regarding datasets, for the two FPR intervals starting at 0, significant improvements were observed in all four settings on SUSY and IJCNN, and in three of four settings on MNIST and SVHN. In contrast, no significant improvements were observed on FMNIST, CIFAR10, CODRNA, or EPSILON.

Although the average improvement tended to be larger for higher class priors (0.0108 for π=0.2\pi=0.2 vs. 0.0068 for π=0.1\pi=0.1), the numbers of significant improvements were identical (7 out of 16 settings each).

These observations suggest that our method tends to be particularly advantageous for FPR intervals including the highest-score region, although the gains vary across datasets. We will clarify these findings in the revised paper.

Thank you for the response and the additional experiments. Although I still think the methodological novelty is limited, the ablation clarifies the practical value of the proposed FPR threshold estimator.

Thank you for your follow-up and for acknowledging the practical value of our proposed FPR threshold estimator.

While our approach builds on established PU learning principles, we believe that providing a principled and effective solution to the practically important problem of pAUC maximization without labeled negative data constitutes a meaningful contribution. In addition to reformulating both the pairwise risk and FPR-dependent thresholds, our work establishes finite-sample error bounds for the resulting estimators. We hope that these theoretical results, together with the additional ablation, help clarify the technical and practical contributions of our work.

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

The paper presents a method for classification that uses the AUC as objective function. The goal is to maximize the AUC between a defined interval, so it's called partial AUC (pAUC). It derives an algorithm towards this end for an arbitrary model.

Critical strengths

The paper is easy to follow and does a good job following through on what it proposes. The results are better than what is compared with.

Critical weaknesses

Th problem addressed is not of great significance in my subjective opinion. Rather than maximize the full AUC, a method is presented to maximize the AUC over a subinterval of [0,1]. I don't have much direct experience with the problem, but the resulting approach seems fairly unsurprising and straightforward.

Questions for authors

I have no questions.

2 comments

We thank the reviewer for the careful reading and constructive feedback.

The problem addressed is not of great significance in my subjective opinion. Rather than maximize the full AUC, a method is presented to maximize the AUC over a subinterval of [0,1]. I don't have much direct experience with the problem, but the resulting approach seems fairly unsurprising and straightforward.

Thank you for your comment. We would like to clarify the practical importance and technical contributions of our work.

Maximizing pAUC is important in applications such as cybersecurity threat detection, medical screening, and online advertising, where false positives can lead to unnecessary investigations, additional medical examinations, or wasted advertising costs. In such applications, the goal is to achieve a high true positive rate (TPR) while maintaining a low false positive rate (FPR). As discussed in previous pAUC maximization studies, maximizing full AUC does not necessarily lead to high TPR in the low-FPR region, since full AUC evaluates performance over the entire FPR range. In contrast, pAUC directly focuses on performance over the FPR region of practical interest.

Furthermore, restricting the FPR to a subinterval introduces an additional technical challenge: the pAUC objective depends on score-dependent thresholds determined by the negative-class distribution, unlike full AUC. In the PU setting, this distribution is unknown, so simply rewriting the pairwise risk using PU data is insufficient. In Section 4, we address this challenge by reformulating both the pairwise risk and the FPR-dependent thresholds using positive and marginal distributions. We also establish finite-sample error bounds for the estimated FPR thresholds (Theorem A.3) and the empirical pAUC estimator (Theorem A.6).

As shown in Tables 1 and 2, our method outperformed PU-based full AUC maximization (PUAUC) in many settings, supporting the importance of directly optimizing the target FPR region. Furthermore, to demonstrate the importance of our PU-based threshold reformulation, we conducted an additional ablation that retains our PU-rewritten pairwise risk but estimates the thresholds by treating unlabeled data as negative. Detailed results are provided in Tables A and B in our response to Reviewer gTiB (Question 1). Our method achieved higher mean test pAUC in all 32 settings, supporting the practical importance of the proposed threshold reformulation.

Overall, these theoretical and experimental results support the significance of maximizing pAUC directly from PU data and show that our contribution goes beyond a straightforward application of existing PU risk rewriting.

Thank you for taking the time to review our paper.

As the author–reviewer discussion period is nearing its end, we would appreciate it if you could review our responses and the revised paper.

Please let us know if you have any further questions or concerns.

Comments on the paper

2 comments

We have uploaded a revised version of our paper incorporating the additional theoretical clarifications and experimental results discussed in our responses. The added material is highlighted in blue.

The main revisions are as follows:

  • Appendix A.2–A.3: Added remarks on relaxing the no-point-mass assumption for threshold estimation and clarifying the scope of the excess pAUC guarantee.
  • Appendix C.9–C.10: Added an ablation study on PU-based threshold estimation and experiments on a narrower low-FPR interval, [0, 0.05].
  • Appendix C.11–C.12: Added an analysis of negative-score ties and an analysis of the experimental settings in which significant improvements are observed.

We thank the reviewers for their constructive feedback and welcome any further questions or comments.

Area_Chair_XpewOpenReview ↗

Reminder: Please Review the Author Responses

Dear Reviewers,

As the discussion period is coming to an end, please take a moment to review the authors' responses and update your ratings if your assessment has changed. If you have any remaining concerns, please feel free to share them in the discussion.

Thank you for your efforts!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.