Treffer: Peaks and Valleys: A Journey Through Predictive Modelling for Software Engineering

Over the last decade automated predictive models have become very popular in a wide range of software engineering research areas, including software requirements, software design and development, testing and debugging and software maintenance. Despite the huge rise in these automated approaches and the investigation of their use in a wide range of areas, optimal results have not yet been reached, exper- imental and evaluation pitfalls still exist, and few studies have sought how they can be applied in industry. As a result, both researchers and practitioners still seek ways to achieve more accurate estimates, as well as increase the adoption of automated predictive models in practice. Therefore, enhancing the design, use and evaluation of predictive models is of great need. The work in this thesis seeks new ways to achieve machine-human coopera- tion to help ameliorate the performance and real-world applicability of automated prediction models. It also investigates current prediction evaluation measures and the use of different machine learning APIs as possible sources of conclusion insta- bility (i.e., inability to consistently and uniformly present the results of empirical software engineering models), in order to increase the robustness of the empirical studies. The approaches presented herein target two of the main areas for soft- ware engineering predictive models, specifically the areas of software effort estima- tion and software defect prediction, and advances them with both algorithmic and methodological contributions.