Towards Artefact-based Requirements Engineering for Data-Centric Systems
This work tackles the complexity of RE for data-centric systems, which is an incremental step in improving software development practices for AI/ML-intensive applications.
The paper addresses the lack of guidance in Requirements Engineering (RE) for data-centric systems with AI/ML components, proposing an artefact-based approach with new requirements categories and data-centric artefacts, based on ongoing industry investigations.
Many modern software-intensive systems employ artificial intelligence / machine-learning (AI/ML) components and are, thus, inherently data-centric. The behaviour of such systems depends on typically large amounts of data processed at run-time rendering such non-deterministic systems as complex. This complexity growth affects our understanding on needs and practices in Requirements Engineering (RE). There is, however, still little guidance on how to handle requirements for such systems effectively: What are, for example, typical quality requirements classes? What modelling concepts do we rely on or which levels of abstraction do we need to consider? In fact, how to integrate such concepts into approaches for a more traditional RE still needs profound investigations. In this research preview paper, we report on ongoing efforts to establish an artefact-based RE approach for the development of datacentric systems (DCSs). To this end, we sketch a DCS development process with the newly proposed requirements categories and data-centric artefacts and briefly report on an ongoing investigation of current RE challenges in industry developing data-centric systems.