A Large-Scale Study of ML-Related Python Projects

Paper i proceeding, 2024

The rise of machine learning (ML) for solving current and future problems increased the production of ML-enabled software systems. Unfortunately, standardized tool chains for developing, employing, and maintaining such projects are not yet mature, which can mainly be attributed to a lack of understanding of the properties of ML-enabled software. For instance, it is still unclear how to manage and evolve ML-specific assets together with other software-engineering assets. In particular, ML-specific tools and processes, such as those for managing ML experiments, are often perceived as incompatible with practitioners' software engineering tools and processes. To design new tools for developing ML-enabled software, it is crucial to understand the properties and current problems of developing these projects by eliciting empirical data from real projects, including the evolution of the different assets involved. Moreover, while studies in this direction have recently been conducted, identifying certain types of ML-enabled projects (e.g., experiments, libraries and software systems) remains a challenge for researchers. We present a large-scale study of over 31,066 ML projects found on GitHub, with an emphasis on their development stages and evolution. Our contributions include a dataset, together with empirical data providing an overview of the existing project types and analysis of the projects' properties and characteristics, especially regarding the implementation of different ML development stages and their evolution. We believe that our results support researchers, practitioners, and tool builders conduct follow-up studies and especially build novel tools for managing ML projects, ideally unified with traditional software-engineering tools.