BioJ48: adaptando o método J48 para mlassificação de dados biológicos desbalanceados

Abstract

Decision tree-based classification models are widely used in Machine Learning (ML) due to their high interpretability and accuracy. However, traditional decision tree algorithms, such as J48 (an implementation of C4.5), face challenges when applied to imbalanced biological datasets, where the number of records per class varies significantly. This imbalance can lead to misleading performance metrics, as models tend to favor the majority class while neglecting the minority class, which is often crucial in medical and biological applications. This study proposes modifications to the J48 algorithm to improve its sensitivity and specificity when classifying imbalanced biological data. The research explores adjustments in information gain calculation, aiming to enhance the algorithm’s performance without compromising interpretability. Various evaluation metrics are analyzed to ensure a more balanced classification approach. The research methodology involves the implementation and testing of the modified J48, named bioJ48, on real biological datasets with different degrees of class imbalance. Comparative experiments between traditional J48 and bioJ48 are conducted, evaluating their predictive effectiveness based on key performance metrics such as accuracy, precision, recall, and F1-score. The expected results include an improved classification model that better identifies minority classes in biological datasets, providing a more reliable evaluation of performance through diversified metrics. The proposed modifications aim to contribute to the field of ML by offering a practical solution to class imbalance issues in biological data analysis.