Detection of code smells using machine learning techniques combined with data-balancing methods

Abstract

Code smells are prevalent issues in software design that arise when implementation or design principles are violated. These issues manifest as symptoms or anomalies in the source code. Timely identification of code smells plays a crucial role in enhancing software quality and facilitating software maintenance. Previous studies have shown that code smell detection can be accomplished through the utilization of machine learning (ML) methods. However, despite their increasing popularity, research suggests that the suitability of these methods are not always appropriate due to the problem of imbalanced data. Consequently, the effectiveness of ML models may be negatively affected. This study aims to propose a novel method for detecting code smells by employing five ML algorithms, namely decision tree (DT), k-nearest neighbors (K-NN), support vector machine (SVM), XGboost (XGB), and multi-layer perceptron (MLP). Additionally, to tackle the challenge of imbalanced data, the proposed method incorporates the random oversampling technique. Experiments were conducted in this study using four datasets that encompassed code smells, specifically god-class, data-class, long-method, and feature-envy. The experimental outcomes were evaluated and compared using various performance metrics. Upon comparing the outcomes of our models on both the balanced and original datasets, we found that the XGB model achieved the highest accuracy of 100% for detecting the data class and long method on the original datasets. In contrast, the highest accuracy of 100% was obtained for the data class and long method using DT, SVM, and XGB models on the balanced datasets. According to the empirical findings, there is significant promise in using ML techniques for the accurate prediction of code smells.