Test Smell Prediction Via Automated Machine Learning

Abstract

Test smells are poor design or implementation practices in test code that can degrade test quality and hinder maintenance, making their early detection crucial. While traditional heuristic-based methods have been used for test smell detection, they often suffer from limited generalizability and manual tuning needs; recently, machine learning (ml) approaches have emerged as a promising alternative, though ml applications in test smell prediction remain underexplored. In this study, we conducted an empirical evaluation using TPOT, an AutoML tool, on two benchmark datasets covering Eager Test and Mystery Guest smells. We systematically compared models across three data imbalance handling strategies and selected the top performers for cross-comparison. The results demonstrate that TPOT-based models achieve robust and competitive predictive performance, with an F1-Score of 0.6779 on the Mystery Guest dataset. These findings highlight the potential of TPOT as an effective and efficient approach for improving test smell detection beyond traditional ML models.