A Multivariate Analysis of Static Code Attributes for Defect Prediction

Defect prediction is important in order to reduce test times by allocating valuable test resources effectively. In this work, we propose a model using multivariate approaches in conjunction with Bayesian methods for defect predictions. The motivation behind using a multivariate approach is to overcome the independence assumption of univariate approaches about software attributes. Using Bayesian methods gives practitioners an idea about the defectiveness of software modules in a probabilistic framework rather than the hard classification methods such as decision trees. Furthermore the software attributes used in this work are chosen among the static code attributes that can easily be extracted from source code, which prevents human errors or subjectivity. These attributes are preprocessed with feature selection techniques to select the most relevant attributes for prediction. Finally we compared our proposed model with the best results reported so far on public datasets and we conclude that using multivariate approaches can perform better.