Towards automatically generating context-specific ML pipelines : a case study at adesso SE

Abstract

Machine learning pipelines are an essential component of modern data science. They play a crucial role in automating the development process of ML models. However, designing and configuring ML pipelines is a complex and time-consuming task. Furthermore, various challenges must be addressed to ensure a successful implementation. In this case study, a comprehensive approach for generating ML pipelines automatically is proposed, aiming to alleviate the burden of manual pipeline design and enable ML practitioners to focus more on model development and analysis. The ML practitioner should be able to provide a series of configurations that are used to generate an ML pipeline which could act as a base for the further ML workflow. For this, interviews with five experts from adesso SE were conducted to determine the components and quality requirements of ML pipelines in general as well as project specifications they might depend on. Using the results from these interviews and findings from related work, a prototypical approach was developed. In a second round of interviews with the same interviewees, the prototypical approach was evaluated using usefulness and ease of use as evaluation metrics. The results of the case study show that the interviewees deemed such an approach to automatically generate ML pipelines useful. The reduction in time to produce first ML models at the start of a project was highlighted. Based on these results, the prototypical approach could be developed further, becoming a useful tool in the ML workflow of every ML project.