Prediction of Number of Software Defects based on SMOTE

Prediction of software defects is an effective way to improve system quality, and it is a key factor affecting the efficiency of defect detection and repair in software components. The purpose of this study is to improve the effectiveness of component defect prediction in the following two ways: for the imbalance of training data in defect prediction and the insufficient support of single regression in predicting the number of defects in components First, this study proposed to adopt SMOTE to construct a balanced sample dataset and oversample the defective components in the unbalanced sample dataset to take into account the proportion of different types of samples and improve the accuracy of prediction; second, this study proposes a method of multi-step prediction for the number of defects that supports regression after classification, and the method applies support vector machines to classify components and filter out non-defective components in the classification results, applies regression to establish a component defect number prediction model to effectively implement the multi-step prediction of component defect number, and further improves the accuracy of prediction. The evaluation of the experiment was completed on open-source datasets. The results show that the accuracy of multi-step prediction is better than the prediction by regression alone, and multi-step prediction has higher overall efficiency and applicability.