Evaluating Empirical Models for the Detection of High-Risk Components: Some Lessons Learned

Introduction Software complexity metrics are often used as indirect metrics of reliability since they can be obtained relatively early in the software development life cycle. Using complexity metrics to identify high-risk components, i.e. components which likely contain faults, allows software engineers to focus the verification effort on them, thus achieving a reliable product at a lower cost. Since in this study the direct metric of reliability is the class to which the software component belongs (high-risk or low-risk), the prediction problem is reduced to a classification model. Classification problems have traditionally been solved by various methods, which originate from different problem-solving paradigms such as statistical analysis, machine learning, and neural networks. This study compares different modeling techniques which cover all the three classification paradigms: principal component analysis, discriminant analysis, logistic regression, logical classification