ASSESSING SEVERITY OF SOFTWARE DEFECT REPORTS USING MACHINE LEARNING TECHNIQUES

Abstract

As per the software development, software testing is one of the most important phases of software life cycle. And similarly, a defect report is a key document which is required for software testing. We need to maintain testing reports and defect reports to keep track of the behaviour of software, whether it is going on as desired or we need to make changes in the undergoing software development. But as the software complexity increases, the number of defects also increases. Our prime focus then relies on looking for the defects and classifying them on the basis of severity. Severity assessment is of prime focus for test engineers. Actually, most of the defect reports generated by almost any kind of software tool generate a log report. Such log reports contain description of the defects encountered. It is difficult to scan each and every line and find out the severity of the defects. So, there is a need for a system that scans various log reports and classifies it in various categories as low, medium, high on the basis of keywords encountered in the defect report. The main idea behind this paper can be broadly classified in two heads, text classification and machine learning techniques. As a subject, we have chosen the NASA’s Project and Issue Tracking System (PITS) dataset and TOMCAT dataset. Various text classification techniques have been applied to extract raw data from the log report. Then, we have applied machine learning techniques over it to get the severity report. To validate the result, k-fold cross validation method is applied over data in different machine learning techniques. The machine learning technique used here is Multilayer Perceptron and statistical method used is Multinominal Logistic Regression. It has been observed that MLP method has given better results in all of the cases as compared to MLR method.