Predicting Software Defect Severity Level Using Deep-Learning Approach with Various Hidden Layers

Abstract

The severity value of the defect helps to decide how soon the defects needs to be fixed. The software used to handle defect reports gets more than ten thousands defects every week, and it is not feasible to assign severity value of the defect by manually reading defect reports. So, severity value prediction based on defect contents helps to find an appropriate severity value. These models have applications in appropriate scheduling of testing resources. The predictive power of severity value prediction models dependent on the input features computed from the description reports of defect. In this paper, we have applied eight different embedding techniques with an objective to compute n-dimensional numeric vector for defect report. Further, we have also applied feature selection techniques and data sampling technique to find relevant features and handle the unequal distribution of sample present in different classes. The predictive power of considered eight different embedding techniques have been assessed using eight different varieties of deep learning models. The experimental results on six projects highlight that the usage of feature selection techniques, embedding to extract feature from text, and SMOTE help in improving the predictive ability of defect severity level prediction models.