Software Defect Discovery and Resolution Modeling incorporating Severity

Traditional software reliability growth models only consider defect discovery data, yet the primary concern of software engineers is defect removal. Past attempts to model defect resolution emphasize approaches based on differential equations and queueing theory. However, these models do not explicitly identify the activities performed to remove defects and resources allocated to these activities according to their severity. Models should consider these practical factors to enable more detailed resource allocation and planning. This paper presents a model to predict the number of defects resolved according to the discrete Cox proportional hazard model with covariates, demonstrating the approach with covariates on the number of low, medium, and high severity defects that were discovered but not resolved in successive intervals. A comparison with differential equation-based and distributional approaches reveals that the covariate model performs better on each goodness of fit measure considered and requires less time to apply. The covariate model also better tracks unresolved defects and exhibits low predictive error, even when as little as 10-20% of testing has been completed. These results suggest that collecting information on defect resolution activities and the corresponding effort dedicated could substantially improve defect resolution modeling to guide process improvement.