Comparing fault-proneness estimation models

Over the last, years, software quality has become one of the most important requirements in the development of systems. Fault-proneness estimation could play a key role in quality control of software products. In this area, much effort has been spent in defining metrics and identifying models for system assessment. Using this metrics to assess which parts of the system are more fault-proneness is of primary importance. This paper reports a research study begun with the analysis of more than 100 metrics and aimed at producing suitable models for fault-proneness estimation and prediction of software modules/files. The objective has been to find a compromise between the fault-proneness estimation rate and the size of the estimation model in terms of number of metrics used in the model itself. To this end, two different methodologies have been used, compared, and some synergies exploited. The methodologies were the logistic regression and the discriminant analyses. The corresponding models produced for fault-proneness estimation and prediction have been based on metrics addressing different aspects of computer programming. The comparison has produced satisfactory results in terms of fault-proneness prediction. The produced models have been cross validated by using data sets derived from source codes provided by two application scenarios.