The increasing popularity of Machine Learning (ML) is generating challenges also for developers. The multitude of programming languages, libraries and available resources allow them to easily build their own models or algorithms. However, ML models are tightly connected to their data implying a different development process from other types of software. Software projects often rely on version control platforms, such as GitHub, but these platforms have not yet been extended to support ML projects. There is poor support for data versioning and no link between ML and software artifacts. Thus, traceability and model evolution can become challenging for developers. While some specific ML platforms exist, they still require considerable manual specification of ML artifacts and links between them. In this work, we propose a framework for automatic identification and traceability of links between data, code and ML model through Mining Software Repositories (MSR) techniques. Our tool combines static code analysis and mining commit data to identify ML, code and data artifacts, reconstruct links between them and retrieve commits that affect each end of the link. The objective is to increase productivity and the developers’ awareness of their project through the recovered traceability.
Paper
Full text
MSR4ML: Reconstructing Artifact Traceability in Machine Learning Repositories
Semantic Scholar · Computer Science · 2021
Abstract
The increasing popularity of Machine Learning (ML) is generating challenges also for developers. The multitude of programming languages, libraries and available resources allow them to easily build their own models or algorithms. However, ML models are tightly connected to their data implying a different development process from other types of software. Software projects often rely on version control platforms, such as GitHub, but these platforms have not yet been extended to support ML projects. There is poor support for data versioning and no link between ML and software artifacts. Thus, traceability and model evolution can become challenging for developers. While some specific ML platforms exist, they still require considerable manual specification of ML artifacts and links between them. In this work, we propose a framework for automatic identification and traceability of links between data, code and ML model through Mining Software Repositories (MSR) techniques. Our tool combines static code analysis and mining commit data to identify ML, code and data artifacts, reconstruct links between them and retrieve commits that affect each end of the link. The objective is to increase productivity and the developers’ awareness of their project through the recovered traceability.