Metode Cost-Sensitive Ensemble dalam Menyelesaikan Multi-Class Imbalanced Classification Problem

Abstract

This research aims to develop an effective Cost-sensitive learning method for handling imbalanced multi-class Datasets. The method takes into consideration the cost of classification errors and adapts its cost function to focus more on minority classes, which are often underrepresented in the data. The core concept of this research involves assigning different cost weights to each class within the cost function in the context of Logistic Regression modeling. The classification cost assigned to each class may vary depending on the importance of correctly classifying that particular class (true positive rate). This study utilizes optimization techniques to determine the optimal cost for each class, thereby improving the overall classification performance. Additionally, in an effort to enhance classification performance, this research applies data preprocessing using Principal Component Analysis (PCA) before implementing the classification method on multi-class Datasets. The test results indicate that the use of PCA has a positive impact on improving the classification method's performance in various scenarios. The proposed method is tested on seven different Datasets, including lymphography, wine, glass identification, new-thyroid, e-coli, ispu and dermatology. Average performance results, including accuracy, precision, recall, f1-score, and Area Under the Curve (AUC), demonstrate that Cost-sensitive learning with cost optimization outperforms conventional classification methods. This research holds significant potential in the development of Cost-sensitive learning methods for predicting data in imbalanced multi-class Datasets. The findings underscore the importance of considering classification error costs and modifying cost functions to achieve more accurate classification of minority classes in the context of imbalanced Datasets.