Design-level metrics estimation based on code metrics

Abstract

Fault detection based on mining code and design metrics has been an active research area for many years. Basically module-based metrics for source code and design level are calculated or obtained and data mining is used to build predictor models. However, in many projects due to organizational or software process models, design level metrics are not available and/or accurate. It has been shown that performance of these classifiers or predictors decline if only source code features are used for training them. Based on best of our know knowledge no set of rule to estimate design level metrics based on code level metrics has been presented since it is believed that design level metrics have additional information and cannot be estimated without access to design artifacts. In this study we present a fuzzy modeling system to find and present these relationships for projects presented in NASA Metrics Data Repository (MDP) datasets. Interestingly, we could find a set of empirical rules that govern all the projects regardless of size, programming language and software development methodology. Comparison of fault detectors built based on estimated design metrics with actual design metrics on various projects showed a very small difference in accuracy of classifiers and validated our hypothesis that estimation of design metrics based on source code attributes can become a practical exercise.