acceptodds
Under review as a conference paper at ICLR 2027

Bidirectional Single-Class Learning for Partial Multi-Label Learning

Abstract

Partial multi-label learning (PML) is a special case of multi-label learning, in which label noise is confined to the observed positive labels. Although numerous PML methods have been developed to handle noisy label spaces, many of them rely on general noise disambiguation mechanisms without explicitly exploiting the asymmetric noise structure of PML. To better leverage this characteristic, we propose Bidirectional Single-Class Learning for Partial Multi-Label Learning (BSCPML), which treats positive and negative labels separately. Specifically, clean negative labels can be directly fitted through single-class learning, whereas positive labels may require purification before learning. To theoretically justify this asymmetric treatment, we develop a unified loss estimation framework that compares several representative PML losses in terms of bias, variance, and their combined estimation error. The analysis demonstrates that an appropriately weighted BSCPML loss can minimize the gap to the clean-label loss. Extensive experiments on 12 benchmark datasets demonstrate that BSCPML achieves competitive or superior performance compared with state-of-the-art PML methods, particularly under severe label noise.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-label learning, Single-class learning, Partial multi-label learning, Optimization, Bias-variance analysis

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_6PqfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Bidirectional Single-Class Learning for Partial Multi-Label Learning (BSCPML), a new method for partial multi-label learning (PML). BSCPML first leverages the reliability of negative labels in PML to encourage the classifier to focus more on negative-label information. It further incorporates label correlations, the low-rank structure of the label space, and the sparsity of noisy labels into a unified optimization objective. Experiment results demonstrate that BSCPML achieves the best performance on multiple benchmarks.

Critical strengths
  1. The paper provides extensive theoretical analysis to support the proposed method BSCPML.

  2. The proposed method is evaluated on 12 benchmark datasets and compared against 10 baseline methods.

Critical weaknesses
  1. The overall presentation of the proposed method needs improvement. Please refer to the Questions For Authors section.

  2. Additional experimental results (ablation studies) are needed to validate the effectiveness of each component.

Questions for authors
  1. The paper provides a relatively detailed explanation of Eq. (9). However, the motivation for replacing Tr(XWL(XW)T)Tr(XWL(XW)^T) in Eq. (8) with Tr(WLWT)Tr(WLW^T) in Eq. (9) remains unclear. What is the motivation for changing from the prediction space to the parameter space, and what is the intended benefit of this modification?

  2. The explanations of Eqs. (10) and (11) require further clarification. In Eq. (10), the authors introduce ZZ to represent the recovered ground-truth label matrix. Why is ZZ not used instead of YY in the first term of Eq. (10)? In Eq. (11), the authors replace Y=Z+NY = Z + N with Y=YZ+NY = YZ + N. Furthermore, what motivates the replacement of the l1l1 norm with the FF norm in Eq. (11)?

  3. In Section 5.1, the authors said that standard deviations are reported. However, the corresponding standard deviations are not presented in the tables.

  4. Figure 2 presents the ablation study results, but the results of Eq. (10) are missing, making it difficult to assess the performance difference between Eqs. (10) and (11). Moreover, in my opinion, the current box plots do not clearly illustrate the performance differences. The authors are encouraged to report the results as means and standard deviations to facilitate a clearer comparison of the contributions of individual components.

1 comment
AuthorsOpenReview ↗

Fully Addressed

Response 1: Laplacian Regularization in Prediction and Parameter Spaces

The basic principle of Laplacian regularization is to exploit label correlations by encouraging highly related labels to have similar representations. This constraint can be imposed either in the prediction space (XW) or in the parameter space (W). Although not mathematically equivalent, both approaches incorporate label correlations into model learning.

We choose parameter-space regularization mainly for computational efficiency. Regularizing W directly avoids the additional feature-matrix operations required by XW. Specifically, their gradients involve WL and XᵀXWL, respectively. Thus, given similar modeling objectives and comparable performance, parameter-space regularization reduces matrix operations and computational overhead during iterative optimization. This is a common trade-off between modeling and computational efficiency in graph regularization, which is why we did not elaborate on it in the original manuscript.

Response 2: Clarification of Eqs. (10) and (11)

(1) Why does the first term of Eq. (10) use Y rather than Z?

The first term preserves supervision from the original observed labels Y, while the third term introduces additional supervision from the purified labels Z. These terms serve different purposes and are not interchangeable.

In machine learning, an objective function often combines multiple loss and regularization terms to balance different learning goals. Our formulation follows this principle: some terms fit the observed labels Y, while others fit the purified labels Z. These complementary supervision sources constrain each other, reducing dependence on any single source and mitigating overfitting.

(2) Why replace Y = Z + N with Y = YZ + N?

PML label matrices typically contain many instances but relatively few labels, and may not satisfy a meaningful low-rank assumption. Direct low-rank decomposition of the entire label matrix therefore has limited value. In contrast, the dependency structure among labels may be low-rank. We consequently adopt Y = YZ + N, shifting the low-rank constraint from the label matrix to the label self-representation matrix Z to capture label correlations more directly. This motivation is already explained in Section 3.2.

(3) Why replace the L1 norm with the Frobenius norm?

Eq. (11) incorporates three complementary objectives: full-label fitting, reliable negative-label fitting, and purified positive-label fitting. Their mutual constraints reduce the risk of overfitting to any single supervision source, making a strong L1 sparsity constraint less necessary. Moreover, the L1 norm introduces nonsmooth optimization and additional computational complexity. We therefore replace it with the squared Frobenius norm, preserving the label reconstruction constraint while simplifying optimization and improving computational efficiency.

We will further clarify these design motivations in the manuscript.

Response 3: Missing Standard Deviations

Due to space limitations, we retained only the mean values in the main text but inadvertently omitted the corresponding standard deviations from the appendix. These results have now been added in Appendix D.3 (Tables 16–23) of the revised manuscript.

Response 4: Evaluation of Eq. (10) and Ablation Results

We have added a complete experimental comparison for Eq. (10) in Appendix D.3 (Tables 12–15). The results show that Eqs. (10) and (11) achieve comparable performance, consistent with our expectations. Eq. (11) retains the core idea of using purified labels as additional supervision while adopting a simpler reconstruction formulation and a smoother regularization term to reduce optimization complexity. We plan to move these results from the appendix to the main text in the final version to facilitate direct comparison.

Furthermore, following another reviewer's suggestion, we have added a more strictly controlled ablation study involving Laplacian regularization in Appendix D.1 (Fig. 10) to better isolate the contributions of individual components.

We retain the box plots because they clearly illustrate the performance distribution, medians, and variability across the 12 datasets.

Overall, these suggestions help clarify our methodological design and strengthen the presentation and rigor of the paper.

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

This paper proposes BSCPML, which exploits PML's asymmetric noise structure by fitting clean negative labels directly and down-weighting or purifying noisy positive labels. The authors provide a bias-variance analysis under a one-sided false-positive noise model, showing that the weighted BSCPML loss minimizes the gap to the clean-label loss.

Critical strengths

1.BSCPML is a simple and effective method, with thorough theoretical analysis provided. 2.The paper conducts extensive experiments, including evaluations on 12 benchmark datasets against 10 baselines, ablation study, robustness test, parameter sensitivity, and Nemenyi tests.

Critical weaknesses

1.The novelty is limited, the core idea of this paper is treating positive and negative labels differently. BSCPML essentially recombines existing components (single-class learning for negatives + purification for positives) without clearly articulating what's fundamentally new beyond a different weighting scheme. 2.Parameter sensitivity only varies one parameter at a time, ignoring interactions. 3.The proposed method doesn't consistently outperform other baselines — in many cases it's even worse , but the paper doesn't analyze or explain why.

Questions for authors

In partial multi-label learning, the label matrix contain noise, how to calculate the similarity matrix LL in equation 8?

1 comment
AuthorsOpenReview ↗

Concerns on Novelty Fully Addressed

1. Limited Novelty ?

Existing PML methods employ various label disambiguation and noise purification techniques, but mainly focus on identifying or correcting noisy candidate positive labels, rather than explicitly exploiting the one-sided noise structure of PML at the loss-function level. Recognizing the unequal reliability of positive and negative labels and translating this asymmetry into a learning objective are two different contributions. Our core innovation lies in the latter: a learning strategy specifically designed for the asymmetric noise structure of PML, rather than a simple combination of existing purification techniques.

More importantly, we provide theoretical justification for this strategy. By analyzing bias, variance, and expected squared error, we reveal a fundamental limitation of conventional unbiased loss correction: eliminating bias does not necessarily produce an estimate closer to the clean loss, and noise correction may substantially amplify variance. In contrast, we derive element-wise optimal weights for the bidirectional loss and show that, under the specified theoretical conditions, the estimated loss of our proposed method can exactly match the clean-label loss.

2. Parameter Sensitivity

We have added joint parameter sensitivity experiments in Appendix D.2 (Fig. 11) to examine interactions between hyperparameters. Specifically, we jointly vary parameter pairs for BSCPMLn and BSCPMLr and evaluate Hamming Loss and Ranking Loss across 12 datasets.

3. Inconsistent Performance Across Datasets

No machine learning algorithm can be expected to rank first across every dataset, metric, and experimental setting. The purpose of methodological research is to develop theoretically motivated and empirically effective approaches to specific learning problems, rather than guarantee universal superiority. Our work focuses on the one-sided noise structure of PML and investigates whether class-separated fitting provides a more effective learning objective than conventional joint fitting.

To validate this central idea, we develop two variants: BSCPMLn, which requires no additional label purification, and BSCPMLr, which incorporates matrix factorization and purification. Even without purification, BSCPMLn outperforms existing PML methods on multiple datasets and evaluation metrics. This directly demonstrates the independent value of class-separated fitting and shows that our main performance gains do not rely on existing label purification techniques.

Reviewer_Ci27OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Bidirectional Single-Class Learning for Partial Multi-Label Learning (BSCPML), which explicitly exploits asymmetric label noise by treating reliable negative labels and potentially corrupted positive labels differently. The authors develop two variants: BSCPMLn, which strengthens negative-label fitting, and BSCPMLr, which further introduces positive-label purification through low-rank label dependency modeling. The paper also presents a bias-variance analysis comparing different loss estimators under uniform one-sided false-positive noise. Experiments on 12 benchmark datasets against 10 existing methods demonstrate competitive performance across multiple evaluation metrics, particularly under severe label noise.

Critical strengths
  1. The paper identifies a relevant characteristic of partial multi-label learning: observed negative labels are reliable, whereas observed positive labels may contain false positives. Explicitly incorporating this asymmetry into the loss function is a reasonable design choice and provides an interpretable alternative to general-purpose label disambiguation methods.
  2. The experiments cover 12 datasets from different domains and compare BSCPML against 10 representative methods using four evaluation metrics. The inclusion of different noise levels, ablation studies, and runtime comparisons provides useful evidence of the method's effectiveness. The reported improvements, particularly under severe label noise, support the practical value of asymmetric label treatment.
Critical weaknesses
  1. The bias-variance analysis in Section 4 focuses on a simplified entry-wise bidirectional squared loss under uniform one-sided false-positive noise. However, the actual BSCPMLn and BSCPMLr objectives additionally incorporate label correlation regularization or low-rank label purification, whose effects are not covered by the theoretical analysis. Furthermore, the optimal weighting coefficients derived (36) depend on the unobserved clean labels and current predictions, whereas the implemented algorithms use fixed hyperparameters. Therefore, the analysis does not rigorously establish the claimed benefits of the complete learning framework.
  2. In Section 5.4, B4 is described as extending B3 with positive-label purification. However, the robust BSCPMLr objective in (11) removes the Laplacian regularizer used by BSCPMLn while introducing low-rank reconstruction and purification. Therefore, multiple components change simultaneously, making it difficult to attribute the performance differences specifically to positive-label purification. A more controlled ablation should retain the other components while adding or removing only the purification mechanism.
  3. BSCPMLr improvement over the simpler BSCPMLn is small in the main results. For example, under 30% noise, the average Accuracy increases only from 0.5265 to 0.5290, while Ranking Loss decreases from 0.1658 to 0.1653. Moreover, the robust variant does not consistently outperform the naive variant on every dataset and metric. The paper should more clearly demonstrate when positive-label purification is necessary and whether its additional computational cost is justified by meaningful performance improvements.
Questions for authors

See weakness above.

1 comment
AuthorsOpenReview ↗

Theoretical Concerns Fully Addressed

Response 1: Limited Theoretical Scope and Misalignment Between Theory and Algorithm Design

Our theoretical analysis directly addresses the central research question: Can class-separated fitting theoretically outperform conventional joint fitting of positive and negative labels? We establish a unified loss-estimation framework and analyze different losses in terms of bias, variance, and expected squared error. This analysis targets our core methodological contribution, rather than the convergence or generalization properties of the entire optimization framework. Treating the absence of theoretical analysis for auxiliary regularization and label purification as a fundamental flaw conflates the central research question with auxiliary algorithmic components.

The optimal weights in Eq. (36) demonstrate the theoretical advantage of class-separated fitting, rather than prescribing instance-wise hyperparameters for implementation. We use fixed hyperparameters to ensure fair comparisons: introducing instance-adaptive weighting would add another optimization mechanism and confound the comparison.

More importantly, fixed hyperparameters do not invalidate the theoretical advantage of class-separated fitting. Conventional joint losses impose a fixed relative weighting between positive and negative labels, whereas our additional negative-label fitting term introduces an independent degree of freedom. Even without instance-wise optimal weights, asymmetric weighting allows the model to exploit reliable negative supervision. Our theoretical results therefore provide a direct justification for the proposed loss design.

Response 2: Ablation Study

We have added a more strictly controlled ablation study in Appendix D.1 (Fig. 10), which isolates the contributions of individual components.

The results further demonstrate that negative-label fitting achieves competitive performance even without sophisticated label purification, directly supporting our central research claim.

Response 3: Performance Gains of BSCPMLr

This concern requires distinguishing our core contribution from the incremental gains of an auxiliary purification module. The central question of this paper is whether class-separated fitting improves upon conventional joint multi-label losses, not whether a particular matrix purification technique consistently improves every dataset.

We develop two variants to investigate this question. BSCPMLn directly exploits reliable negative labels without purification, while BSCPMLr integrates existing matrix factorization and label purification techniques. Since many established PML methods employ such techniques, including a purification-based variant also enables a more comprehensive comparison.

Crucially, BSCPMLn achieves competitive or superior results on multiple datasets and metrics without any label purification. This demonstrates that the improvements do not depend on existing matrix purification techniques but can arise directly from our proposed class-separated fitting strategy. This is one of the paper's key experimental findings.

For fair comparisons, we deliberately use fixed hyperparameters rather than instance-adaptive weights that could further exploit the theoretical optimum. Moreover, real-world label uncertainty may extend beyond false-positive noise, making matrix-based purification potentially valuable. BSCPMLr accordingly provides modest additional gains under the main 30% noise setting.

Therefore, the contribution of this paper cannot be judged solely by the incremental improvement of BSCPMLr over BSCPMLn. BSCPMLn offers a simple and efficient realization of our central idea, while BSCPMLr demonstrates its compatibility with existing purification techniques.

The conditions under which different purification and label-correlation methods offer additional benefits constitute a separate research question. They do not undermine our theoretical or experimental conclusions regarding class-separated fitting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.