IEEE Access (Jan 2020)
A Bayesian Networks-Based Method to Analyze the Validity of the Data of Software Measurement Programs
Abstract
Measures are essential resources to improve quality and control costs during software development. One of the main factors for having successful software measurement programs is measure trustworthiness, defined as how much a user can trust a measure to use it with confidence. Such confidence enables the users to interpret them and use them for supporting decision-making. ISO/IEC 15939:2007 describes four stages that influence such interpretability: measure selection, measure validation, threshold definition, and data validation. The literature is scarce in supporting data validation, which directly impacts the measure's trustworthiness value. This article aims to detail a method that uses Bayesian networks for supporting data validation and shows its application in practice to four software development projects from one company. The proposed method uses Bayesian networks to calculate the degree to which a collected number or symbol represents the real value for the measures and is integrated with GQM for assessing the measurement program's goals. First, the measurement users must create GQM model hierarchical structures, use it as input for constructing the Bayesian network, validate the Bayesian network, and, finally, use it to support decision-making. A tool to support the proposed method was developed and is freely available. Further, herein, the results of the case study are presented. We identified four benefits in using the proposed method: Externalization, Diagnosis support, Measure interpretation improvement, and Decision-making support. Given this, even though the initial effort to use the proposed method lasted, on average, one hour and fourteen minutes, the benefits of using it outweighed the effort of applying it. Therefore, our findings suggest that there was a positive intention in adopting the proposed method in practice.
Keywords
