Engineering and Selecting Machine Learning Features for Feature Models (Master’s Thesis Summary)

Software product lines are commonly used to handle the challenges introduced by growing variability needs in modern software. Feature models are often used to model this variability. Their formalized notation allows for automated analysis, but scalability issues can make the analysis of large feature models nearly impossible. One way to address this problem is through machine learning. In this thesis, we analyze the applicability of machine learning for common feature model analysis tasks. We especially compare different algorithms at most stages of a machine learning pipeline. We also create and analyze a new set of machine learning features for feature models. We find that no single pipeline configuration dominates all performance metrics. However, we identify advantages and disadvantages commonly displayed by different machine learning techniques. Our research paves the way for automated machine learning for feature model analysis and provides the means for easier machine learning application and ablation studies.