Comparative Assessment of Software Quality Classification Techniques: An Empirical Case Study

• Authors: Taghi M. Khoshgoftaar;Naeem Seliya
• Venue: Empirical Software Engineering
• Year: 2004

Software metrics-based quality classification models predict a software module as either fault-prone (fp) or not fault-prone (nfp). Timely application of such models can assist in directing quality improvement efforts to modules that are likely to be fp during operations, thereby cost-effectively utilizing the software quality testing and enhancement resources. Since several classification techniques are available, a relative comparative study of some commonly used classification techniques can be useful to practitioners. We present a comprehensive evaluation of the relative performances of seven classification techniques and/or tools. These include logistic regression, case-based reasoning, classification and regression trees (CART), tree-based classification with S-PLUS, and the Sprint-Sliq, C4.5, and Treedisc algorithms. The use of expected cost of misclassification (ECM), is introduced as a singular unified measure to compare the performances of different software quality classification models. A function of the costs of the Type I (a nfp module misclassified as fp) and Type II (a fp module misclassified as nfp) misclassifications, ECM is computed for different cost ratios. Evaluating software quality classification models in the presence of varying cost ratios is important, because the usefulness of a model is dependent on the system-specific costs of misclassifications. Moreover, models should be compared and preferred for cost ratios that fall within the range of interest for the given system and project domain. Software metrics were collected from four successive releases of a large legacy telecommunications system. A two-way ANOVA randomized-complete block design modeling approach is used, in which the system release is treated as a block, while the modeling method is treated as a factor. It is observed that predictive performances of the models is significantly different across the system releases, implying that in the software engineering domain prediction models are influenced by the characteristics of the data and the system being modeled. Multiple-pairwise comparisons are performed to evaluate the relative performances of the seven models for the cost ratios of interest to the case study. In addition, the performance of the seven classification techniques is also compared with a classification based on lines of code. The comparative approach presented in this paper can also be applied to other software systems.