Collaborative Optimization of Multiclass Imbalanced Learning: Density-Aware and Region-Guided Boosting

TL;DR

This work tackles multiclass imbalanced learning by proposing a Density-Aware and Region-Guided Boosting framework that tightly couples density-based sample weighting with a region-aware dynamic sampling strategy inside AdaBoost. It introduces a density factor $\rho$ derived from mutual nearest neighbors and a confidence factor $\delta$ based on classification hardness to guide noise-resistant weight updates, along with a region-partitioning scheme that divides minority samples into Dense, Boundary, and Noise regions. A region-guided generation strategy and a progressive sampling scheduler enable dynamic, region-aware synthetic sample creation that sharpens decision boundaries while mitigating class overlap. The authors provide a theoretical justification that their joint weighting minimizes a regularized exponential loss and validate the approach with extensive experiments on 20 KEEL datasets, showing state-of-the-art performance against eight baselines and emphasizing the benefits of the collaborative optimization. The method demonstrates robust minority-class performance and offers a path to extending collaborative optimization to other ensemble learners beyond AdaBoost.

Abstract

Numerous studies attempt to mitigate classification bias caused by class imbalance. However, existing studies have yet to explore the collaborative optimization of imbalanced learning and model training. This constraint hinders further performance improvements. To bridge this gap, this study proposes a collaborative optimization Boosting model of multiclass imbalanced learning. This model is simple but effective by integrating the density factor and the confidence factor, this study designs a noise-resistant weight update mechanism and a dynamic sampling strategy. Rather than functioning as independent components, these modules are tightly integrated to orchestrate weight updates, sample region partitioning, and region-guided sampling. Thus, this study achieves the collaborative optimization of imbalanced learning and model training. Extensive experiments on 20 public imbalanced datasets demonstrate that the proposed model significantly outperforms eight state-of-the-art baselines. The code for the proposed model is available at: https://github.com/ChuantaoLi/DARG.

Figures (4)

• Figure 1: Overall framework of the model proposed in this study. The density factor and confidence factor serve as bridge, they establish a deep coupling between the sample weight update mechanism and the dynamic sampling strategy, enabling the collaborative optimization of imbalanced learning and model training.
• Figure 2: Comparison of connectivity between $k$-nearest neighbors and mutual nearest neighbors with $k$=10. The $k$-nearest neighbors establishes 2,670 edges, the mutual nearest neighbors establishes only 1,141 edges.
• Figure 3: Evolution of sample weights during training process, darker shades indicate higher weights.
• Figure 4: Schematic diagram of dynamic sampling under collaborative optimization.