A Clustering-Based Oversampling Approach for Class Imbalance in Mixed Data

Class imbalance is a recurring problem in machine learning, especially in domains like SME's, healthcare, finance, and fraud detection, where rare but important events are often underrepresented in data. Traditional oversampling techniques such as random oversampling and SMOTE frequently suffer from overfitting or the generation of unrealistic synthetic samples, particularly when handling mixed data types that include both numerical and categorical features. This study proposes a clustering-based oversampling method that combines Mini-Batch K-Means with genetic-inspired inheritance mechanisms to address this challenge. The approach first clusters minority instances and then generates synthetic samples using dominanceweighted averaging for numerical features and probabilistic random inheritance for categorical attributes. A dynamic multigenerational augmentation loop is applied to incrementally balance the dataset. The proposed method is evaluated on a mobile price classification dataset, using a Random Forest classifier as the predictive model. Compared to the baseline (no oversampling) and random oversampling, the proposed method achieves higher classification accuracy (0.9000) and F1-score (0.8970), demonstrating improved generalization and minority class recognition without distorting data distributions.