Peningkatan Performa Rekognisi Code smell pada Bahasa Pemrograman Python Menggunakan Metode SMOTEENN dan Decision Tree Pruning

Abstract

This research addresses the enhancement of code smell recognition through an extensive investigation into two prominent types, Large Class and Long Method, utilizing advanced machine learning techniques. The study commences by establishing a comprehensive understanding of the code smell phenomenon, its detrimental impact on software quality, and the significance of efficient detection methods. Employing a novel approach, the research leverages two primary datasets, Large Class and Long Method, and introduces a meticulous data collection process, incorporating numerical features representing various program metrics. The preprocessing stage involves normalization and feature selection, refining the datasets for subsequent analysis. The utilization of the SMOTEENN technique proves instrumental in addressing class imbalance, resulting in a more balanced distribution and improved correlation between features and target variables. Model development encompasses hyperparameter optimization using RandomizedSearchCV and cross-validation, yielding refined decision trees, random forests, and boosting algorithms. The results showcase remarkable performance enhancements, outperforming previous studies, with accuracy and Matthews Correlation Coefficient reaching up to 99.69% and 99.38%, respectively, for Long Method detection using XGBoost and SMOTEENN. The overall findings underscore the efficacy of ensemble methods and data augmentation techniques in bolstering code smell recognition models. The research concludes by emphasizing the substantial advancements achieved in comparison to prior studies and suggests future investigations into diverse code smell types to further enrich dataset diversity and improve the robustness of the models.