Empirical Framework for Building and Evaluating Bayesian Network Models for Defect Prediction
- When?
- Tuesday 27 April 2010, 11:15 to 12:15
- Where?
- 39BB02
- Open to:
- Staff, Students
- Speaker:
- Miss Ana Jakimovska
Software reliability is a crucial factor to consider when developing software systems or defining optimal release time. For many organisations ‘time to market’ is critical and avoiding unnecessary testing time whilst retaining reliable software is important.
The focus of this research work is on a defect prediction model (Phase model) incorporating genuine cause-effect relationship related to people, process, size and testing. The model was built using Bayesian network technology and has been used as a decision support tool in real projects to predict the number of residual defects of a software system during all the phases of the development lifecycle but prior to deployment. The Phase model was originally evaluated through validation trials in large industrial organisations (Philips, Israel Aircraft Industries and QinetiQ) with very encouraging defect prediction ability.
This research study defines a framework for the continued evaluation and improvement of large-scale Bayesian networks models such as the Phase model. The main concept of the framework is breaking down the models into independent parts (subnets), improving the accuracy of each part by using empirical software engineering methods and assembling the models with the improved subnets. In defining the framework we used the Phase model and its testing subnet as a case study. We identified areas for improvement of the existing testing subnet and conducted two experiments (small-scale and large-scale) that enabled the subnet to be refined with empirical data. The new testing subnet was validated using experimental data and demonstrated encouraging predictive power.