Analyzing Human Dimensions of Software Engineering Processes

We propose to build CodeFault, a system for learning models of faulty software engineering based on patterns of human behavior. Flawed sofwtare is notably costly in terms of maintenance, but the costs are even higher when vulnerabilities introduced by such code are exploited by data breaches and ransomware. Our Phase II proposal builds on the design and feasibility investigation work we completed in Phase I, which showed that we can aggregate and integrate public data about software engineering from muliple heterogeneous sources to learn models related to how human behavior affects software development. Our approach is distinguished by the range of sources we consider, our methods for linking users across sources, the way we identify and classify faults, the range of machine learning techniques we employ, and our dual focus on general patterns as well as those associated with certain types of vulnerabilities. The end result will be a system that can look at a new codebase and generate predictions about which types of software changes may introduce significant risk.