Abstract: Engineering machine learning (ML)-enabled systems poses various challenges from both a theoretical and a practical perspective. Among those challenges are how to effectively align the specification of stakeholder needs, requirements, and constraints with the engineering activities composed by interdisciplinary teams and activities (including data science and software engineering). In this paper, we tackle this challenge with PerSpecML, our perspective-based approach for specifying ML-enabled systems that involves analyzing 51 concerns grouped into five perspectives: ML objectives, user experience, infrastructure, model, and data. To conceive, evaluate, and evolve PerSpecML, we followed the model for technology transfer to industry, conducting three evaluations: (i) in academia, (ii) with industry representatives, and (iii) in two real industrial case studies. We report on both the approach and the results from our evaluations which corroborate how the approach was considered useful for specifying ML-enabled systems in practice, particularly helping to reveal important requirements that would have been missed without using the approach. The evaluations allowed us to evolve PerSpecML towards its industry-readiness and strengthen our confidence its suitability for specifying ML-enabled systems.

Paper submitted to the International Conference on Software Engineering (ICSE) 2023.