Mining software defect data to support software testing management

Achieving high quality software would be easier if effective software development practices were known and deployed in appropriate contexts. Because our theoretical knowledge of the underlying principles of software development is far from complete, empirical analysis of past experience in software projects is essential for acquiring useful software practices. As advances in software technology continue to facilitate automated tracking and data collection, more software data become available. Our research aims to develop methods to exploit such data for improving software development practices. This paper proposes an empirical approach, based on the analysis of defect data, that provides support for software testing management in two ways: (1) construction of a predictive model for defect repair times, and (2) a method for assessing testing quality across multiple releases. The approach employs data mining techniques including statistical methods and machine learning. To illustrate the proposed approach, we present a case study using the defect reports created during the development of three releases of a large medical software system, produced by a large well-established software company. We validate our proposed testing quality assessment using a statistical test at a significance level of 0.1. Despite the limitations of the available data, our predictive models give accuracies as high as 93%.