From a Data Science Driven Process to a Continuous Delivery Process for Machine Learning Systems

Paper i proceeding, 2020

Development of machine learning (ML) enabled applications in real-world settings is challenging and requires the consideration of sound software engineering (SE) principles and practices. A large body of knowledge exists on the use of modern approaches to developing traditional software components, but not ML components. Using exploratory case study approach, this study investigates the adoption and use of existing software development approaches, specifically continuous delivery (CD), to development of ML components. Research data was collected using a multivocal literature review (MLR) and focus group technique with ten practitioners involved in developing ML-enabled systems at a large telecommunication company. The results of our MLR show that companies do not outright apply CD to the development of ML components rather as a result of improving their development practices and infrastructure over time. A process improvement conceptual model, that includes the description of CD application to ML components is developed and initially validated in the study.