Non-Functional Requirements for Machine Learning: Challenges and New Directions

Machine Learning (ML) provides approaches which use big data to enable algorithms to "learn", producing outputs which would be difficult to obtain otherwise. Despite the advances allowed by ML, much recent attention has been paid to certain qualities of ML solutions, particularly fairness and transparency, but also qualities such as privacy, security, and testability. From a requirements engineering (RE) perspective, such qualities are also known as non-functional requirements (NFRs). In RE, the meaning of certain NFRs, how to refine those NFRs, and how to use NFRs for design and runtime decision making over traditional software is relatively well established and understood. However, in a context where the solution involves ML, much of our knowledge about NFRs no longer applies. First, the types of NFRs we are concerned with undergo a shift: NFRs like fairness and transparency become prominent, whereas other NFRs such as modularity may become less relevant. The meanings and interpretations of NFRs in an ML context (e.g., maintainability, interoperability, and usability) must be rethought, including how these qualities are decomposed into sub-qualities. Trade-offs between NFRs in an ML context must be re-examined. Beyond the changing landscape of NFRs, we can ask if our known approaches to understanding, formalizing, modeling, and reasoning over NFRs at design and runtime must also be adjusted, or can be applied as-is to this new area? Given these questions, this work outlines challenges and a proposed research agenda for the exploration of NFRs for ML-based solutions.