Shedding Light on the Techniques for Building Bayesian Networks in Software Engineering

Abstract

Context Bayesian networks (BNs) have been used to tackle several software engineering (SE) problems, such as risk management and effort estimation. They enable reasoning under uncertainty and have the advantage of incorporating expert knowledge to build more accurate models when sufficient historical data are not available. Software practitioners often encounter a lack of substantial evidence concerning the usability, limitations, risks, and benefits of BNs, as is the case with many other topics in the SE literature. Therefore, there is a need to organize and systematize the existing knowledge in this area. Objective This paper aims to provide researchers and practitioners with an overview of the techniques for building BNs in SE. Method We conducted a tertiary study following the guidelines available in the SE literature. Results We examined six secondary studies. Our findings revealed that expert knowledge emerges as the predominant technique for structure learning and, in conjunction with learning from data using automated tools, is widely employed for parameter learning in BNs. Conclusion Despite the attention given to data-driven approaches in SE, it is worth acknowledging the significant value that expert knowledge continues to hold in constructing more accurate and robust models. This observation underscores potential opportunities for developing expert-driven solutions to enhance model building and foster the adoption of BNs in the software industry.