pAUC: Partial Area Under ROC Curve

• pAUC is defined as the partial area under the ROC curve that focuses on performance within a specified false-positive or true-positive interval relevant to the application.
• It enables practitioners to optimize models in domains like object detection, speaker verification, and medical triage where only a narrow operating range is critical.
• Various methodological variants—such as one-way, two-way, and time-dependent pAUC—support direct optimization strategies to enhance targeted ranking performance.

PAUC, more commonly written pAUC in the ROC literature, denotes the partial area under the receiver operating characteristic curve: the area under only the application-relevant segment of the ROC curve rather than the full curve. It is used when performance outside a prescribed operating regime is irrelevant, so that optimizing or reporting full AUC would be misaligned with deployment. In that sense, pAUC generalizes AUC, with full AUC recovered when the retained range is the entire ROC curve (Paisitkriangkrai et al., 2013, Bai et al., 2019).

1. Conceptual role in ROC analysis

The central motivation for pAUC is operational selectivity. In object detection, speaker verification, screening, anomaly detection, and related decision problems, systems are typically deployed only within a restricted false-positive regime. A detector can therefore be globally strong on the full ROC curve yet still be suboptimal in the narrow region that matters in practice. This observation is explicit in pedestrian detection, where many applications operate within a prescribed false-positive range, and in speaker verification, where real systems work on only part of the ROC or DET curve rather than the entire curve (Paisitkriangkrai et al., 2013, Bai et al., 2019).

This restricted-view interpretation gives pAUC a different semantics from ordinary AUC. Full AUC averages ranking quality across all operating points; pAUC suppresses performance outside the chosen interval and concentrates the objective on the relevant frontier. In the speaker-verification literature, this is described as a scalar summary of the application-relevant portion of ROC that remains threshold-free and calibration-insensitive (Bai et al., 2019). In high-specificity clinical settings, the same logic appears as emphasis on the left-most part of the ROC curve, where false positives are costly and specificity above $80\%$ may be required (Gerke et al., 2016).

The choice of interval is domain dependent. In cascade-based detection, node classifiers are described as needing maximal detection rate at a moderate false positive rate, for example around $40\%$ to $50\%$ (Paisitkriangkrai et al., 2013). In security-sensitive speaker verification, the motivating region is very low FPR, with illustrative cases such as a bank security system requiring FPR below $0.01\%$ (Bai et al., 2019). In medical triage, the emphasis can instead be framed as maintaining high sensitivity while controlling false alarms (Hasan et al., 3 Jun 2025).

2. Formal definitions and principal variants

A standard one-way definition restricts the false-positive axis to an interval $(t_0,t_1)$: $$\text{pAUC}(t_0,t_1)=\int_{t_0}^{t_1}\text{ROC}(t)\,dt.$$ When $(t_0,t_1)=(0,1)$, pAUC reduces to full AUC (Gerke et al., 2016). In left-tail usage, the focus is often $\text{pAUC}_t$, meaning the segment $(0,t)$ of the FPR axis (Gerke et al., 2016).

Several papers make the ranking structure explicit. In speaker verification, pAUC over an FPR interval $[\alpha,\beta]$ is implemented by selecting the subset $40\%$0 of negative trials whose ranks correspond to that FPR band, and then computing a normalized AUC over all positive trials $40\%$1 and only those selected negatives. This makes pAUC a restricted positive-negative ranking problem rather than an unrestricted pairwise ranking problem (Bai et al., 2019). The same idea appears in deep speaker embeddings, where negatives are sorted, the subset corresponding to $40\%$2 is retained, and a ranking loss is applied only to positive trials against those selected negatives (Bai et al., 2019).

The literature also contains important variants. A two-way pAUC constrains both FPR and TPR rather than only FPR; the corresponding work argues that FPR-only pAUC can be misleading conceptually and practically, and proposes a nonparametric estimator, asymptotic normality properties, bootstrap-based comparison, and a regression framework for covariate effects (Yang et al., 2015). A time-dependent pAUC extends the notion to censored survival data by redefining cases and controls at time $40\%$3 and estimating the restricted ROC area as a function of time, together with variance-covariance estimation and confidence intervals (Hung et al., 2011).

Not all parameterizations are FPR-first. In the ISIC 2024 skin-cancer setting, pAUC is defined in the high-sensitivity region as

$40\%$4

with the score normalized to $40\%$5, matching the width of the retained TPR interval from $40\%$6 to $40\%$7 (Hasan et al., 3 Jun 2025). That formulation is mathematically different from the usual $40\%$8 FPR truncation, but conceptually analogous: only the clinically relevant segment of the operating curve counts. Likewise, in anomalous sound detection for DCASE 2020 Task 2, pAUC is defined operationally over the low-FPR interval $40\%$9 (Huang et al., 17 Sep 2025).

3. Direct optimization and surrogate objectives

A large part of the pAUC literature is concerned not merely with evaluating models by pAUC, but with optimizing it directly. In pedestrian detection, structured ensemble learning was proposed to maximize detection rate in a user-defined false-positive range by directly optimizing pAUC. Because the interval is user-specified, the same framework can train either a single strong classifier or a node classifier within a cascade (Paisitkriangkrai et al., 2013). A later extension incorporated spatially pooled low-level visual features and a structured ensemble method, again targeting pAUC rather than generic classification loss (Paisitkriangkrai et al., 2014).

Speaker verification produced several of the clearest optimization frameworks. One line introduced pAUCMetric, a Mahalanobis distance metric learning back-end that optimizes pAUC over a chosen low-FPR interval, typically $50\%$0, $50\%$1, using a hinge-relaxed ranking objective. Because the score is linear in the Mahalanobis matrix, the resulting objective is convex and admits a global optimum (Bai et al., 2019). A second line moved pAUC optimization into deep embeddings: a squared-hinge verification loss was defined over positive trials and selected hard negatives, and a class-center trial-construction strategy was introduced to make pAUC-based training competitive with identification losses such as softmax variants (Bai et al., 2019).

A different strategy is to optimize a simpler surrogate that induces good left-tail ranking behavior without exactly matching pAUC. Logistic push is the canonical example: it turns pAUC-oriented ranking into weighted logistic regression, with heavier penalties on high-scoring controls, and can therefore use standard penalized regression machinery such as lasso in high-dimensional biomarker settings. Its argument is explicitly heuristic rather than an exact pAUC derivation, but it is designed to improve the low-FPR part of the ROC curve (Gerke et al., 2016).

Scalable stochastic optimization required another line of work. Online and mini-batch methods treated pAUC as a non-decomposable loss depending on the top $50\%$2-fraction of negatives, established sublinear regret, and derived stochastic subgradient procedures with per-epoch complexity $50\%$3 due to top-$50\%$4 negative selection (Kar et al., 2014). More recent deep-learning work reformulated one-way and two-way pAUC via distributionally robust optimization. In that framework, a CVaR-based DRO yields an exact but non-smooth estimator for pAUC, whereas a KL-regularized DRO yields a smooth but inexact soft estimator; both come with gradient-based algorithms and non-convex convergence guarantees (Zhu et al., 2022).

The most recent formulations in the supplied corpus pursue instance-wise minimax reformulations. One line develops asymptotically unbiased instance-wise regularized objectives with linear per-iteration computational complexity in sample size (Shao et al., 2022). Another explicitly frames the remaining issue as an approximation gap and proposes two minimax formulations: one with an asymptotically vanishing gap, the other unbiased at the cost of more variables, together with linear per-iteration complexity and a sharp generalization order $50\%$5 for typical one-way and two-way PAUCs (Jiang et al., 1 Dec 2025).

4. Application domains and operating regimes

The historical driver of direct pAUC optimization in computer vision was pedestrian and object detection. There the concern is not generic rank discrimination but detector quality within a prescribed low-false-positive regime, often expressed in false positives per image or its ROC analogue. Structured pAUC optimization was shown to be effective on both synthetic and real-world data, and later combined with spatially pooled covariance and pooled LBP features to obtain state-of-the-art pedestrian detection on Caltech-USA at the time (Paisitkriangkrai et al., 2013, Paisitkriangkrai et al., 2014).

In speaker verification, pAUC addresses the mismatch between full-curve ranking metrics and deployment at stringent operating points. The literature uses low-FPR regions such as $50\%$6, interprets DET curves as reparameterized ROC curves, and shows that pAUC-focused back-ends or losses outperform or rival strong baselines when the verification regime is the real target (Bai et al., 2019, Bai et al., 2019).

In medical AI, pAUC is often tied to screening or triage utility rather than generic discrimination. In the ISIC 2024 skin-cancer setting, performance is measured by pAUC above $50\%$7 TPR, explicitly to ensure that models prioritize high sensitivity. The reported best system is a hybrid ensemble combining image models, engineered metadata, and GBDT fusion, with pAUC $50\%$8 on the normalized $50\%$9 scale (Hasan et al., 3 Jun 2025). In prostate cancer detection on bpMRI, lesion-level performance is summarized by pAUC over the FROC interval $0.01\%$0 to $0.01\%$1 false positives per patient, and encoding anatomical priors yields an average increase of $0.01\%$2 pAUC across FROC setups (Saha et al., 2020).

In anomalous sound detection, pAUC is central because the task is dominated by false alarms in the low-FPR regime. For DCASE 2020 Task 2, pAUC is evaluated on $0.01\%$3 FPR. The OS-SCL framework reports $0.01\%$4 pAUC using Log-Mel features and $0.01\%$5 pAUC with TFSTgram, with especially notable gains on the ToyConveyor machine type, which is described as the hardest case for false alarms among similar machine IDs (Huang et al., 17 Sep 2025).

In ERP-based brain-computer interfaces, pAUC is not the sole preferred metric but emerges as one of the strongest correlates of spelling rate under class imbalance. Across a private LARESI ERP dataset and the public OpenBMI ERP dataset, low-FPR pAUC—especially $0.01\%$6—belongs to the metrics most strongly supported by the experiments, and the paper recommends reporting pAUC together with PRAUC, AP, ROCAUC, Brier score, and MCC in ERP-based BCI experiments (Bekhelifi et al., 1 Jul 2026).

5. Inference, protocol dependence, and common pitfalls

pAUC is often discussed as an optimization target, but it is also an object of formal statistical inference. For censored time-to-event data, nonparametric estimators with closed-form expressions have been derived for time-dependent pAUC, together with asymptotic Gaussian processes, estimated variance-covariance functions, pointwise confidence intervals, simultaneous confidence bands, and between-group comparisons (Hung et al., 2011). Two-way pAUC work similarly emphasizes asymptotic normality, bootstrap comparison, and regression analysis for covariate effects (Yang et al., 2015).

A recurring practical issue is that pAUC values are definition dependent. Raw magnitudes are not directly interpretable without the retained interval and normalization convention. In the ISIC 2024 setting, for example, the score is normalized to $0.01\%$7, so numerically small values such as $0.01\%$8 are in fact close to the top of the permitted range. The same paper also notes missing implementation details: it does not specify a trapezoidal interpolation rule, a library routine, or whether the final values come from aggregated out-of-fold predictions or averaging per-fold scores (Hasan et al., 3 Jun 2025). This makes code-level equivalence difficult even when the conceptual metric is clear.

Another common misconception is that pAUC comparisons are portable across evaluation pipelines. They are not. The same symbol may refer to low-FPR one-way pAUC, high-sensitivity TPR-restricted pAUC, two-way pAUC, time-dependent pAUC, FROC pAUC, or even pAUC computed on non-ROC curves. In biometric quality assessment, for example, pAUC is computed over Error versus Discard Characteristic curves rather than ROC curves, lower pAUC is better because the y-axis is error, and rankings are sensitive to discard-fraction limits, starting error, interpolation method, and quality-score normalization (Schlett et al., 2023).

The literature also distinguishes sharply between evaluation use and training use. Some systems optimize direct surrogates for pAUC, whereas others use pAUC strictly for model comparison and selection. The ISIC 2024 skin-cancer system is explicit that pAUC is the main reporting metric but not the direct training objective; image models are trained with binary cross-entropy or multitask losses, and GBDTs are trained as ordinary boosted classifiers (Hasan et al., 3 Jun 2025). By contrast, pAUCMetric, structured ensemble learning, and DRO-based deep objectives all target pAUC or its surrogates directly (Bai et al., 2019, Paisitkriangkrai et al., 2013, Zhu et al., 2022).

6. Acronym reuse and terminological ambiguity

Although PAUC overwhelmingly denotes partial area under the ROC curve in detection, verification, and biostatistics, the acronym is not unique. In self-supervised learning, PAUC was introduced as “Prototypical representation through Alignment, Uniformity and Correlation”, a modification of ProtoNCE designed to mitigate prototype “coagulation.” That method adds prototype-level alignment, uniformity, and correlation losses and reports $0.01\%$9 top-1 on ImageNet-100 and $(t_0,t_1)$0 top-1 on ImageNet-1K under the reported settings (Mo et al., 2022).

A second unrelated reuse appears in proactive multimodal evaluation. There PAUC stands for “Proactive Area Under Curve”, a time-aware metric for proactive video question answering. Instead of ROC area, it measures the normalized area under a time-score polyline built from accumulated responses during a ground-truth reply span, and the paper reports that timing-aware PAUC agrees better with human preferences than a time-agnostic $(t_0,t_1)$1 baseline (Wang et al., 12 Jul 2025).

In most of the detection, verification, medical, and statistical literature represented here, however, PAUC refers to the restricted ROC-area functional now standardly written as pAUC. Within that dominant usage, the term covers a family of related objects: one-way FPR-restricted pAUC, two-way pAUC with joint FPR/TPR constraints, time-dependent pAUC for censored outcomes, and application-specific high-sensitivity parameterizations.