acceptodds
Under review as a conference paper at ICLR 2027

IHC-LDL: Imbalanced Hierarchical Classification with Label Distribution Learning

Abstract

Utilizing multi-label learning(MLL) to solve hierarchical classification problems is a challenging task. Recently a novel paradigm named label distribution learning (LDL) handing label polysemy well can mine the relative importance of different labels in label hierarchy. Although numerous hierarchical MLL and LDL algorithms have been proposed respectively, the combined application of the two lacks in-depth exploration. The approaches constructing the hierarchical label distribution often lack addressing feature space correlation and sample imbalance. To address this problem, we propose a LDL-based imbalanced hierarchical classification approach whose main idea is to construct and learn implicit label distribution with hierarchical information, and decompose it into three sub-tasks: label enhancement(LE), imbalanced resampling and adaptive multi-label learning. In this approach, we design a joint strategy of feature-based label distribution and hierarchical label distribution to enhance label and an adaptive weighting for multi-label learning to tackle the task with unobvious label polysemy. Experimental results on widely used open datasets show the effectiveness of our proposal.

Primary area:
learning theory
Keywords:
Hierarchical Classification, Imbalanced Learning, Label Distribution Learning, Label Enhancement

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1MMAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes IHC-LDL for imbalanced hierarchical classification. It combines label enhancement, random oversampling, and a hierarchical consistency module. The method builds soft labels from sample features, label hierarchy, and original labels. It also adjusts the weights of classification and label distribution losses based on the difference between soft and original labels.

Experiments on seven datasets compare the method with several baselines and test the effects of label distribution learning and oversampling. The method improves some metrics on several datasets, but the gains are not consistent across datasets.

Critical strengths
  1. The method considers both label hierarchy and class imbalance in one framework. This is useful for tasks where rare, fine-grained classes must be predicted consistently with their parent classes.

  2. Experiments on seven datasets include images, animal attributes, and proteins. The baseline comparisons and ablations provide useful evidence about where the approach helps, although the gains are mixed.

Critical weaknesses
  1. Eq.2 does not ensure a valid label distribution, and Eq.5 can produce values above 1. Eqs.6–8 also mix prediction outputs with scalar losses. These issues make the training objective difficult to verify and reproduce.

  2. Removing LDL improves all four metrics on F194 and DD. The paper also lacks separate ablations for the adaptive weighting and the modified hierarchy constraint. Since the method largely combines existing components, this missing evidence weakens its contribution.

Questions for authors
  1. How do you ensure that the label distributions in Eqs.2-4 are nonnegative and sum to one, especially when the original labels contain multiple positive entries?

  2. How do you prevent Eq. 5 from producing values above 1? Does restricting the maximum operation to k descendant levels still guarantee hierarchical consistency?

  3. Are HCCM_d and HCCM_l prediction vectors or scalar losses? Please clarify the exact training objective implemented in Eqs. 6-8.

  4. Why should a larger difference between soft and hard labels lead to greater weight on the soft labels? Do you have ablations comparing this rule with fixed weighting?

5.Removing LDL improves all four metrics on F194 and DD. Are these results consistent across random seeds, and what evidence supports the claimed benefit of LDL despite these failures?

Reviewer_uNkLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes the IHL-LDL method for imbalanced hierarchical classification tasks. It applies label distribution learning to hierarchical classification, constructs a joint label enhancement module that integrates features, hierarchical structures and logical labels, and combines it with an imbalanced resampling strategy. It further designs an adaptively weighted loss with hierarchical consistency constraints, which dynamically balances the classification loss and label distribution learning loss based on the MSE between logical labels and label distributions, to adapt to scenarios with different degrees of label polysemy. Experiments are conducted on seven public datasets, comparing flat classification methods, existing label distribution learning approaches and state-of-the-art hierarchical classification baselines, using metrics including micro-F1, macro-F1, hF1 and TE. Module ablation studies are also carried out to verify the function of each component. However, the overall performance improvement is limited, and the method does not outperform the compared baselines on several datasets.

Critical strengths
  • Existing hierarchical label‑distribution‑learning work seldom explicitly copes with sample imbalance across label hierarchies. This paper combines feature‑aware label enhancement, hierarchical prior information and original logical labels to construct comprehensive label distributions. Further, it dynamically adjusts the trade‑off between classification loss and LDL loss according to the MSE gap between logical labels and learned label distributions, which enables adaptive handling of scenarios with varying degrees of label polysemy.

  • The method unifies label enhancement, imbalanced resampling, and hierarchical consistency constraint modules into one framework. Evaluations are conducted over seven public datasets covering image, protein and attribute‑related tasks, with standard hierarchical classification metrics including hF1 and tree‑induced error.

Critical weaknesses
  • Even with the full proposed pipeline, IHL‑LDL only achieves best‑on‑dataset results on a minority of the seven tested benchmarks. On several datasets, competitive or better performance is obtained by simple flat baselines and existing hierarchical methods. The observed performance gains are numerically small.

  • Most building blocks are adapted from existing prior art. label enhancement is built upon the existing LEVI‑MLP, imbalanced resampling directly adopts off‑the‑shelf resampling algorithms, and the hierarchical consistency constraint draws from published hierarchical classification literature. The adaptive weighting βi\beta_i is a simple MSE‑based heuristic without theoretical justification.

  • The figures are too blurry.

Questions for authors
  • The adaptive weight βi\beta_i is heuristically computed from the MSE between logical labels and constructed label distribution. Is there any empirical analysis demonstrating that this MSE‑driven weight reliably reflects the real‑world label polysemy degree of individual samples? Could you provide statistics showing how βi\beta_i distributes across different samples in your datasets?

  • Your full method only yields limited and inconsistent performance gains over simple flat baselines on multiple datasets. Could you identify what dataset characteristics determine whether IHL‑LDL brings observable improvements? For example, label hierarchy depth, imbalance severity, or degree of label polysemy.

  • In the ablation study, combining label enhancement and resampling only brings improvements for three out of seven datasets. Could you clarify why joint usage of the two modules fails to deliver consistent gains across all benchmarks? Are there cases where one component may counteract the effect of the other?

  • The hierarchical consistency constraint HCCM adopts a fixed hyper‑parameter k for descendant label range selection. How sensitive is final performance to the choice of k? Have you conducted controlled experiments to rule out the possibility that good results are heavily dependent on this manually set hyper‑parameter?

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

This paper proposes IHL-LDL, a label distribution learning-based framework for imbalanced hierarchical classification. The method aims to address label dependencies and class imbalance in hierarchical multi-label classification. Specifically, it integrates feature-based and hierarchy-based label enhancement, imbalanced resampling, and adaptive loss weighting with hierarchical consistency constraints. Experiments are conducted on seven public datasets to evaluate the effectiveness of the proposed framework.

Although the problem is practically relevant, the paper suffers from significant weaknesses in its theoretical justification, experimental validation, and presentation quality. In particular, several key mathematical formulations are insufficiently justified, and the experimental results do not convincingly support the claimed effectiveness of the proposed components.

Critical strengths
  1. Practically relevant problem. The paper addresses class imbalance and label dependencies in hierarchical multi-label classification, which are important challenges in real-world applications. Incorporating label distribution learning into this setting is a potentially meaningful research direction.
  2. A unified framework with evaluation across multiple datasets. The paper integrates label enhancement, imbalanced resampling, and hierarchical consistency constraints into a unified framework.
Critical weaknesses
  1. Poor figure quality. Several key figures, particularly Figures 2 and 3, are blurry and have insufficient resolution. Important text and architectural details are difficult to distinguish, significantly affecting the readability and clarity of the proposed method.
  2. Insufficient theoretical justification. The framework mainly combines existing techniques without sufficient theoretical analysis. Moreover, Eqs. (2)–(4) do not explicitly guarantee valid probability distributions, while Eq. (5) may produce predictions exceeding 1, raising concerns about the mathematical soundness of the method.
  3. Insufficient experimental validation. The performance improvements are generally marginal, and statistical significance tests and hyperparameter sensitivity analyses are missing. Furthermore, some ablation results in Table 3 outperform the full model, such as removing LDL on the DD dataset. The lack of tail-class evaluations also weakens the claims regarding imbalanced classification.
Questions for authors
  1. In Eqs. (2)–(4), how do the authors ensure that the weighted combination of feature-based label distributions and multi-hot logical labels satisfies the non-negativity and normalization constraints?
  2. In Eq. (5), adding δ may result in probabilities exceeding 1. How is this handled in practice? Could the authors also clarify the computation and normalization procedures of the loss functions in Eqs. (6)–(8)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.