PaperlayerView exact revision on arXiv

Source-linked AI summary

ML-CLAPPS: A Probabilistic Framework for Imbalanced Multilabel Learning under Class Prioritization

Priel Hazan, Nir Almog, Noam Keidar, Yael Yaniv, Assaf Schuster

arXiv:2610.15816v1cs.LGcs.CVstat.ML

TL;DR

Severe class imbalance and label co-occurrence make multilabel learning challenging, while clinical ECG classification also calls for prioritizing significant conditions. ML-CLAPPS learns instance-sampling probabilities to align class exposure with a user-defined importance vector, and outperforms existing samplers on macro-averaged metrics and across most of the prevalence spectrum on a real-world imbalanced ECG image dataset.

  • Problem

    Severe class imbalance is harder to address in multilabel learning because labels co-occur, and clinical ECG classification also requires prioritizing conditions by significance.

    Source: §1 · p. 1

  • Method

    ML-CLAPPS learns a distribution over training instances to control class exposure and align it with a user-defined importance vector.

    Source: §1 · p. 2, §2.1 · p. 2, Notation and Expected Class Exposure · pp. 2–3

  • Results

    On a real-world imbalanced ECG image dataset, ML-CLAPPS outperforms existing samplers in macro-averaged metrics and across most of the prevalence spectrum.

    Source: §4 · p. 9

  • Takeaways & Limitations

    User-defined importance vectors shape class exposure and support controlled performance trade-offs across classes in clinical settings where condition prioritization is essential.

    Source: §4 · p. 9

Abstract

from arXiv · show

Real-world datasets often exhibit severe class imbalance, degrading the performance of deep learning models, particularly on rare classes. While imbalance is well studied in binary and multiclass settings, it is substantially more challenging in multilabel problems due to complex label co-occurrence. In clinical applications such as ECG classification, this challenge is coupled with a second critical challenge: the need to prioritize high-risk conditions, regardless of their prevalence, to mitigate the high cost of false negatives. We introduce the Multilabel Class-Prioritized Probabilistic Sampler (ML-CLAPPS), a learnable sampler that optimizes a distribution over training instances to control the model's exposure to different classes. This approach addresses multilabel imbalance while enabling prioritization of significant conditions via a user-defined importance vector (IV). Empirical evaluation on a real-world multilabel ECG image dataset demonstrates that ML-CLAPPS achieves state-of-the-art performance. We show that it aligns class exposure with the specified IV, and that increasing a class's importance leads to corresponding improvements in its classification performance, enabling explicit prioritization of clinically critical conditions.

1 Introduction

Multilabel ECG classification must handle both severe class imbalance and co-occurring conditions, while also allowing clinically significant conditions to receive priority. ML-CLAPPS addresses these challenges by learning instance-sampling probabilities that align class exposure with a user-defined importance vector.

1 Introduction

  • Multilabel ECG classification faces severe class imbalance complicated by label co-occurrence, alongside the need to prioritize conditions of varying clinical significance.Standard oversampling effective in binary or multiclass settings is considerably more challenging in multilabel learning.Source: §1 · p. 1
  • ML-CLAPPS learns a distribution over training instances to align class exposure with user-defined priorities, addressing imbalance and clinical prioritization in multilabel ECG classification.Existing resampling and class-weighting approaches offer limited exposure control or can destabilize optimization under extreme imbalance.Source: §1 · p. 2

2 Methods

ML-CLAPPS assigns sampling probabilities to training instances and optimizes them to align expected class exposure with user-defined importance while promoting dataset utilization. The resulting convex quadratic program is solved once before training, and its fixed solution is used to sample instances each epoch.

2.1 Defining the ML-CLAPPS Optimization Problem

  • ML-CLAPPS replaces discrete instance duplication with a continuous optimization that assigns each training instance a sampling probability for fine-grained exposure control.The approach contrasts with existing multilabel imbalance methods based on discrete resampling.Source: §2.1 · p. 2
  • Expected class exposure is d = Yp, where Y encodes instance labels and p gives each instance’s sampling probability; d_i is the probability that class i appears in a sampled instance.The exposure vector is not itself a probability distribution.Source: Notation and Expected Class Exposure · pp. 2–3
  • An importance vector specifies target relative class-exposure ratios, but label co-occurrence can make perfect alignment infeasible.The alignment loss compares expected exposure with an importance-proportional target at the same total exposure scale.Source: Exposure-Importance Alignment Term · p. 3
  • A dataset-utilization regularizer, balanced against alignment by γ, discourages probability mass from collapsing onto a small subset of instances.The objective is reformulated as a quadratic program over the probability simplex with a unique global minimizer.Source: Dataset Utilization Regularization · p. 3, Final ML-CLAPPS Objective · pp. 3–4

2.2 Efficient Solution of the ML-CLAPPS Problem

  • The QP is solved with a primal-dual interior-point method implemented in Clarabel, and the resulting distribution is fixed and used to sample instances at each epoch.The optimization is performed once before training.Source: §2.2 · p. 4, Final ML-CLAPPS Objective · p. 4

3 Experiments

Experiments on a multilabel ECG image dataset examine ML-CLAPPS’s sampling trade-off, alignment with user-defined class priorities, and classification performance. The results report improved macro-averaged metrics with a uniform importance vector and performance trends that follow prioritized importance.

3.1 Experimental Setup

  • The experiments use 5,197 multilabel ECG photographs from five hospitals, a 85/15 train-test split, and a fixed ResNet18 backbone across sampling methods.F1 and MCC are prioritized because accuracy can be misleading under severe imbalance.Source: §3.1 · pp. 4–5

3.2 Analysis of ML-CLAPPS Behavior

3.3 Performance under Class Imbalance and Prioritization

  • ML-CLAPPS with a uniform IV achieves superior macro-averaged performance across classification metrics, with substantial gains in F1 and MCC.Table 1 reports mean ± SD across four runs.Source: §3.3 · p. 7, Table 1 · p. 7
  • With LinR2F, performance is highest for the rarest conditions and gradually decreases as conditions become more frequent, following the assigned importance weights.LP ROS and UniformSampler yield near-zero F1 on the rarest conditions.Source: §3.3 · p. 8, Figure 5 · p. 8

4 Conclusion

The conclusion presents ML-CLAPPS as a probabilistic framework for imbalanced multilabel learning with explicit class prioritization. On a real-world imbalanced ECG image dataset, it outperforms existing samplers in macro-averaged metrics and across most of the prevalence spectrum.

  • ML-CLAPPS formulates multilabel sampling as a convex QP balancing exposure-importance alignment and dataset utilization, solved once in polynomial time with zero per-epoch overhead.Source: §4 · pp. 8–9
  • On a real-world imbalanced ECG image dataset, ML-CLAPPS outperforms existing samplers in macro-averaged metrics and across most of the prevalence spectrum.Source: §4 · p. 9
  • User-defined importance vectors shape class exposure and yield controlled performance trade-offs across classes, supporting prioritization of significant conditions in clinical settings.Source: §4 · p. 9

Related papers

Loading 2610.15816v1…