acceptodds
Under review as a conference paper at ICLR 2027

Granularity Gating-Guided Label-Specific Metric Learning for Multi-Label Classification

Abstract

Label-specific local metric learning improves nearest-neighbor multi-label classification by adding label- and region-specific private transformations to a shared one. The regions assigned to a training anchor, its support, select the anchor's transformation and also determine which private transformations its training constraints update. Existing methods partition the data, forcing instances with overlapping semantics into one region. Soft cluster memberships avoid this choice but send every anchor's supervision to every region, and with uniform memberships all private transformations of a label become identical. We therefore restrict each local metric to the anchors that provide enough evidence for it. Granularity-Gated Fuzzy Metric Learning (GFML) sets this restriction with a gate derived from the principle of justifiable granularity. Each region receives a granule whose radius balances rival-adjusted membership coverage against specificity. An anchor inside a granule uses and trains only the private transformations whose granules contain it, so excluded transformations receive none of its supervision. On eight benchmarks with two KNN classifiers, GFML matches or improves on the strongest metric-learning and specialized baselines; ablations show that the gate improves on dense support.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Granular Computing, Fuzzy Set, Multi-Label Classification, Metric Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4sFjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Granularity-Gated Fuzzy Metric Learning (GFML) for multi-label classification. It combines fuzzy clustering, granularity-based gating, and label-specific local metric learning to selectively assign training samples to local metric transformations. However, the methodological novelty is limited, and the experimental evaluation relies on a relatively small number of benchmark datasets. Furthermore, the paper lacks sufficient discussion of and comparisons with recent state-of-the-art methods.

I suggest the author switch to journals such as KBS and increase the number of experiments, especially the dataset and the latest comparative algorithms.

Critical strengths

The paper identifies a meaningful limitation of existing local metric learning methods: hard partitioning restricts samples to a single region, while dense fuzzy assignments may cause different local transformations to become similar.

Critical weaknesses
  1. Limited methodological novelty. The proposed method largely integrates existing techniques, including fuzzy C-means clustering, the principle of justifiable granularity, and label-specific local metric learning. The overall contribution appears incremental rather than a substantial methodological advance. 
  2. Limited theoretical depth. The theoretical analysis mainly establishes that uniform support weights can lead to identical private transformations under positive regularization. However, this result does not establish the necessity or superiority of the proposed gating mechanism, nor does it provide guarantees on generalization or classification performance. 
  3. Insufficient experimental justification. The experiments are conducted on only eight relatively small-scale benchmark datasets, with limited diversity and representativeness. The experimental evaluation is therefore insufficient to convincingly demonstrate the generalizability and robustness of the proposed method.
Questions for authors

see Critical Weaknesses

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

This paper proposes GFML, a multi-label method that learns label-specific local metrics with a granularity-gated fuzzy support mechanism. It uses FCM to obtain prototypes and memberships, then applies PJG to select an optimal radius for each prototype and gate which anchors contribute to each label-specific transformation. The method is evaluated on eight benchmarks against metric-learning and multi-label baselines, with ablation and theoretical analysis.

Critical strengths

1.The paper applies Justifiable Granularity to construct an admission gate for local metric learning, addressing the near-collapse of dense support where private transformations coincide at stationary points. 2.The paper shows uniform memberships cause private transformations to coincide at stationary points, yielding nearly identical private factors. The gate establishes distinct admission sets to break this equality, and the integration of granularity-based regions with label-specific learning is reasonably motivated.

Critical weaknesses

1.The novelty is limited, the method is a straightforward combination of FCM, PJG, and the shared-private metric framework. It lacks a deep theoretical analysis of why this specific combination works and offers limited conceptual contribution. 2.The method is evaluated only with KNN classifiers and compared mainly with older baselines, lacking recent deep multi-label methods. The improvements are marginal and not statistically compelling. 3. Learning a label-specific metric matrix leads to high parameter and computational costs that scale poorly with the number of labels and samples. The paper lacks complexity analysis, training time/memory reporting, or experiments on large-scale datasets to validate scalability.

Questions for authors
  1. How does GFML compare with recent deep metric learning methods for multi-label classification (e.g., deep embedding learning, contrastive learning-based approaches)?
  2. The support matrix W is computed once from training features and held fixed during optimization. Has the authors considered joint optimization of WW and the transformation factors?
  3. Since GFML learns a label-specific metric matrix per label, how do parameters and computation scale with the number of labels and samples?
Reviewer_oqDZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Granularity-Gated Fuzzy Metric Learning (GFML) for multi-label classification, integrating fuzzy clustering, justifiable granularity, and shared–private metric learning to selectively allocate supervision to label-specific local transformations. The method is motivated by the observation that dense fuzzy memberships may propagate supervision to irrelevant regions, potentially leading to the collapse of private metric components. Experiments on eight multi-label benchmarks demonstrate competitive performance against existing metric-learning methods and specialized multi-label classifiers. Although the problem is relevant and the framework is reasonably formulated, I have substantial concerns regarding the underlying motivation, methodological justification, and strength of the experiments.

Critical strengths
  1. The paper points out that hard partitioning may fail to adequately characterize overlapping local structures, while dense fuzzy membership values ​​might disperse supervisory information across irrelevant local components.
  2. GFML integrates fuzzy membership estimation, granularity-based gating, and label-specific metric learning into a coherent framework, and its effectiveness has been demonstrated through experiments.
Critical weaknesses
  1. The motivation for this work is unclear, methodological innovation is limited, and there is insufficient differentiation from existing methods.
  2. The presentation of the paper, as well as the formal definition and articulation of the methodology, require further improvement. At present, the readability of paper is poor.
Questions for authors
  1. The gating mechanism is constructed based on unsupervised feature-space geometry and fuzzy membership margins, without explicitly considering label-specific discriminative information or gradient compatibility. Moreover, its advantages over simpler sparse-support strategies remain insufficiently established. What motivates the adoption of this particular gating mechanism, and what distinctive benefits does it offer over simpler alternatives?
  2. The paper suggests that non-zero gradient contributions under dense fuzzy support may introduce harmful supervision and lead to local metric collapse. To what extent does this argument hold in general, and how does it relate to existing research on fuzzy metric learning?
  3. Does the proposed method originate from PJG itself, or from a general support sparsification operation?
  4. Does maintaining diversity among private transformation factors necessarily lead to more effective local metrics with greater discriminative power? Furthermore, how does this relate to the diversity of the resulting Mahalanobis metrics, or to downstream predictive performance?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.