A Hybrid GDHS and GBDT Approach for Handling Multi-Class Imbalanced Data Classification

Abstract

Multiclass imbalanced classification remains a significant challenge in machine learning, particularly when datasets exhibit high Imbalance Ratios (IR) and overlapping feature distributions. Traditional classifiers often fail to accurately represent minority classes, leading to biased models and suboptimal performance. This study proposes a hybrid approach combining Generalization potential and learning Difficulty-based Hybrid Sampling (GDHS) as a preprocessing technique with Gradient Boosting Decision Tree (GBDT) as the classifier. GDHS enhances minority class representation through intelligent oversampling while cleaning majority classes to reduce noise and class overlap. GBDT is then applied to the resampled dataset, leveraging its adaptive learning capabilities. The performance of the proposed GDHS+GBDT model was evaluated across six benchmark datasets with varying IR levels, using metrics such as Matthews Correlation Coefficient (MCC), Precision, Recall, and F-Value. Results show that GDHS+GBDT consistently outperforms other methods, including SMOTE+XGBoost, CatBoost, and Select-SMOTE+LightGBM, particularly on high-IR datasets like Red Wine Quality (IR = 68.10) and Page-Blocks (IR = 188.72). The method improves classification performance, especially in detecting minority classes, while maintaining high accuracy.