Prediction of defect severity by mining software project reports

With ever increasing demands from the software organizations, the rate of the defects being introduced in the software cannot be ignored. This has now become a serious cause of concern and must be dealt with seriously. Defects which creep into the software come with varying severity levels ranging from mild to catastrophic. The severity associated with each defect is the most critical aspect of the defect. In this paper, we intend to predict the models which will be used to assign an appropriate severity level (high, medium, low and very low) to the defects present in the defect reports. We have considered the defect reports from the public domain PITS dataset (PITS A, PITS C, PITS D and PITS E) which are being popularly used by NASA’s engineers. Extraction of the relevant data from the defect reports is accomplished by using text mining techniques and thereafter model prediction is carried out by using one statistical method i.e. Multi-nominal Multivariate Logistic Regression (MMLR) and two machine learning methods viz. Multi-layer Perceptron (MLP) and Decision Tree (DT). The performance of the models has been evaluated using receiver operating characteristics analysis and it was observed that the performance of DT model is the best as compared to the performance of MMLR and MLP models.