Empirical Analysis of Open Source projects using Feature Selection and Filtering Techniques

Abstract: To achieve software quality for the large systems results very costly. Developers and testers put a lot of effort to investigate a large number of modules in order to ensure the software quality. This tends to be very time consuming process. The machine learning techniques are used for fault prediction of the modules. However, there are many modules which are very low in priority for quality investigation. In this research, Wrapper subset evaluation method has been used to identify that subset of attributes which are most prominent. 10, 20 and 30% of less complex faulty software modules are filtered out from each attribute. The remaining modules are used to build models against four classifiers: Naïve Bayes, Support Vector Machine, k nearest neighbors and C4.5 decision trees. The results of the classifiers were analyzed and compared against the filtering of less complex instances from LOC and NPM metrics. The classifiers based on wrapper subset evaluation method gave better results than the filtering of LOC and NPM metrics. Keywords: Software fault, complexity, CBO, NOA, NMI, DIT, NOC, NAI, NPRIM, NPM, FAN-IN, FANOUT