Optimized Machine Learning Input for Evolutionary Source Code to Architecture Mapping

Automatically mapping source code to architectural modules is an interesting and difficult problem. Mapping can be considered a classification problem, and machine learning approaches have been used to automatically generate mappings. Feature engineering is an essential element of machine learning. We study which source code features are important for an algorithm to function effectively. Additionally, we examine stemming and data cleaning. We systematically evaluate various combinations of features on five datasets created from JabRef, TeamMates, ProM, and two Hadoop subsystems. The systems are open-source with well-established mappings. We find that no single set of features consistently provides the highest performance, and even the subsystems of Hadoop have varied optimal feature combinations. Stemming provided minimal benefit, and cleaning the data is not worth the effort, as it also provided minimal benefit.