IDENTIFYING THE IMPACT OF THE COMPLEXITY OF DATASETS IN BAYESIAN OPTIMIZED XGBOOST ON THE PERFORMANCE OF CLASSIFICATIONS FOR IMBALANCED CLASS DISTRIBUTION DATASETS

Abstract

This study explores how dataset complexity affects the performance of XGBoost models optimized using Bayesian methods, focusing on datasets characterized by imbalanced class distributions. The main challenge is accurately identifying minority classes, which are often misdiagnosed due to the dominance of majority classes, impairing predictive power. Additionally, dataset complexity, as indicated by the coefficient of variation (14.64 % to 85.68 %), does not consistently correlate with improved model performance, highlighting the need for more targeted methods. High-dimensional datasets may not be as accurate as simpler ones and require the use of advanced approaches. By using Bayesian optimization, it is possible to fine-tune hyperparameters and improve classification performance on different types of datasets. This indicates that the selection of appropriate resampling techniques to match the characteristics of the dataset is critical, and that hyperparameter optimization plays an important role in building models with high accuracy. The applications extend to areas such as fraud detection and other fields where the categorization of minority groups is important. Through the use of efficient resampling techniques and advanced optimization methods, this study offers a comprehensive solution to the challenges of imbalanced datasets, enhancing the reliability of machine learning solutions. The variation in resampling techniques and optimizing model performance metrics can be attributed to the distribution of classes, the number of features, the complexity, and the characteristics of the datasets