Empirically validating software metrics for risk prediction based on intelligent methods

Abstract

The software systems which are related to national science and technology projects are very crucial. This kind of systems always involves high technical factors and has to spend a large amount of money, so the quality and reliability of the software deserve to be further studied. Hence, we propose to apply four intelligent classification techniques most used in data mining fields, including Bayesian belief networks (BBN), nearest neighbor (NN), rough set (RS) and decision tree (DT), to validate the usefulness of software metrics for risk prediction. Results show that comparing with metrics such as Lines of code (LOC) and Cyclomatic complexity (V(G)) which are traditionally used for risk prediction, Halstead program difficulty (D), Number of executable statements (EXEC) and Halstead program volume (V) are the more effective metrics as risk predictors. By analyzing obtained results we also found that BBN was more effective than the other three methods in risk prediction.