Optimal Metrics Selection for Software Defect Prediction

Abstract

Software metrics are a very important component in software development area. The most important challenge of any software developer is that ‘the software should be 100% accurate or with minimal defect’ when it reaches to the end-user. The earlier the defect is detected, the earlier the development cost also gets reduced. This is a fact that more complex the data or software, there is a more probability of the defect. So it is always desirable to use only the relevant and important data for the software development which will definitely reduce time and cost of the developer. The main objective of this paper is to find the important and relevant metrics in the software modules which will reduce the attribute in a data set. More the number of variables, more complex will be the system and more the defects; so it is always preferable to select the small feature set using feature selection and focus only on the important variable. The data set for the software bug prediction Chidamber and Kemerer metrics and Object Oriented metrics were taken from the Promise repository which is publicly available. The experiment from the 10 software modules (Jureczko M. Software Engineering: An International Journal 2011;1(1):95), ant, ivy, tomcat, berek, camel, jedit, lucene, poi, synapse and velocity, using feature selection techniques, Boruta, regsubset and FSelector, has shown that response for a class, lines of code and weighted methods for class are the most optimal metrics, whereas number of children and Depth of Inheritance Tree are least significant.