Software defect prediction using neural networks

Abstract

Defect severity assessment is the most crucial step in large industries and organizations where the complexity of the software is increasing at an exponential rate. Assigning the correct severity level to the defects encountered in large and complex software, would help the software practitioners to allocate their resources and plan for subsequent defect fixing activities. In order to accomplish this, we have developed a model based on text mining techniques that will be used to assign the severity level to each defect report based on the classification of existing reports done using the machine learning method namely, Radial Basis Function of neural network. The proposed model is validated using an open source NASA dataset available in the PITS database. Receiver Operating Characteristics (ROC) analysis is done to interpret the results obtained from model prediction by using the value of Area Under the Curve (AUC), sensitivity and a suitable threshold criterion known as the cut-off point. It is evident from the results that the model has performed very well in predicting high severity defects than in predicting the defects of the other severity levels. This observation is irrespective of the number of words taken into consideration as independent variables.