Improving Machine Learning-based Test Smell Prediction using Data Processing Techniques

Abstract

Test smells are indicative symptoms of poor design choices or bad implementations in test code, potentially reducing maintainability and compromising test effectiveness. Although machine learning-based approaches have been proposed to automate test smell detection, their predictive performance remains limited due to severe class imbalance and the presence of redundant or irrelevant features. This study presents an integrated framework that combines data balancing techniques with feature selection methods to enhance model accuracy. Specifically, four data balancing techniques, namely Random Undersampling, NearMiss-3, Random Oversampling, and Borderline SMOTE, are applied alongside two feature selection methods, Chi-Square and Lasso Regression. Predictive models are built using four widely used machine learning algorithms, including Random Forest, K-Nearest Neighbors, Support Vector Machine, and Extreme Gradient Boosting, and evaluated on a benchmark dataset comprising two test smells: Mystery Guest and Duplicate Assert. Experimental results demonstrate substantial improvements in key metrics, especially F1-score, and AUC-PR, in which Random Forest consistently outperforms other classifiers, achieving F1-scores of 0.868 and 0.869 for the two test smells, respectively. The findings also confirm that oversampling methods are generally more effective than undersampling in addressing data imbalance.