Empirical Study on Software Bug Prediction

Software defect prediction is a vital research direction in software engineering field. Software defect prediction predicts whether software errors are present in the software by using machine learning analysis on software metrics. It can help software developers to improve the quality of the software. Software defect prediction is usually a binary classification problem, which relies on software metrics and the use of classifiers. There have been many research efforts to improve accuracy in software defect prediction using a variety of classifiers and data preprocessing techniques. However, the "classic classifier validity" and "data preprocessing techniques can enhance the functionality of software defect prediction" has not yet been answered explicitly. Therefore, it is necessary to conduct an empirical analysis to compare these studies. In software defect prediction, the category of interest is a defective module, and the number of defective modules is much less than that of a non-defective module in data. This leads to a category of imbalance problem that reduces the accuracy of the prediction. Therefore, the problem of imbalance is a key problem that needs to be solved in software defect prediction. In this paper, we proposed an experimental model and used the NASA MDP data set to analyze the software defect prediction. Five research questions were defined and analyzed experimentally. In addition to experimental analysis, this paper focuses on the improvement of SMOTE. SMOTE ASMO algorithm has been proposed to overcome the shortcomings of SMOTE.