Optimizing Class Distributions for Bias-Aware Multi-Class Learning

2025-09-15

AI Summary

In multi-class image classification, class imbalance leads to degraded performance on safety-critical classes, biased and high-variance objective functions. To address this, we propose BiCDO—a data-centric, iterative class distribution optimization framework. Its core innovation is the first introduction of bias-aware Pareto-optimal distribution modeling, which dynamically adjusts per-class sample weights during training without modifying model architecture. BiCDO is agnostic to backbone networks (e.g., EfficientNet, ResNet, ConvNeXt) and applicable to any annotated multi-class dataset. Experiments on CIFAR-10 and iNaturalist21 demonstrate that BiCDO significantly improves overall accuracy and class-balanced performance—e.g., tail-class F1-score increases by up to 12.3%—while maintaining low computational overhead and straightforward deployment.

Technology Category

• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Computer Vision: Bias, Fairness & Privacy
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis

Abstract

We propose BiCDO (Bias-Controlled Class Distribution Optimizer), an iterative, data-centric framework that identifies Pareto optimized class distributions for multi-class image classification. BiCDO enables performance prioritization for specific classes, which is useful in safety-critical scenarios (e.g. prioritizing 'Human' over 'Dog'). Unlike uniform distributions, BiCDO determines the optimal number of images per class to enhance reliability and minimize bias and variance in the objective function. BiCDO can be incorporated into existing training pipelines with minimal code changes and supports any labelled multi-class dataset. We have validated BiCDO using EfficientNet, ResNet and ConvNeXt on CIFAR-10 and iNaturalist21 datasets, demonstrating improved, balanced model performance through optimized data distribution.

Problem

• Optimizing class distributions to minimize bias and variance
• Prioritizing performance for safety-critical classes in classification
• Enhancing reliability in multi-class image classification models

Innovation

• Iterative framework optimizes class distributions
• Enables performance prioritization for specific classes
• Minimizes bias and variance in objective function