Using Machine Learning Techniques to Predict Bugs in Classes: An Empirical Study

Abstract

Software bug prediction is an important step in the software development life cycle that aims to identify bug-prone software modules. Identification of such modules can reduce the overall cost and effort of the software testing phase. Many approaches have been introduced in the literature that have investigated the performance of machine learning techniques when used in software bug prediction activities. However, in most of these approaches, the empirical investigations were conducted using bug datasets that are small or have erroneous data leading to results with limited generality. Therefore, this study empirically investigates the performance of 8 commonly used machine learning techniques based on the Unified Bug Dataset which is a large and clean bug dataset that was published recently. A set of experiments are conducted to construct bug prediction models using the considered machine learning techniques. Each constructed model is evaluated using three performance metrics: accuracy, area under the curve, and F-measure. The results of the experiments show that logistic regression has better performance for bug prediction compared to other considered techniques.