Constructing and Exploiting Software Metrics Networks for Software Quality Assessment

Abstract

Software quality is one of the important aspects of the reliable software. For this reason, software practitioners are interested in the methods that enable to predict defect with software metrics. Most studies focus on what kind of techniques would be fit for the prediction model; however, the research about what is the appropriate number of software metrics for building effective defect prediction model are relatively infrequent. In this paper, we propose a new approach for selecting a subset of software metrics by exploiting the correlation structure among them. We construct software metrics networks with two different relationship measures of MIC (Maximal Information Coefficient) and Pearson correlation coefficient. We then adjust the threshold for filtering out edges with weak correlation, and select representative features for each of the resulting connected components. We validate this subset selection approach with the Poisson regression model. Our results demonstrate that metrics selection using networks with relationship coefficient is efficient enough compared to model with the full set of features.