International audience; The Machine Learning (ML) world is in constant evolution, as the amount of different algorithms in this context is evolving quickly. Until now, it is the responsibility of data scientists to create ad-hoc ML pipelines for each situation they encounter, gaining knowledge about the adequacy between their context and the chosen pipeline. Considering that it is not possible at a human scale to analyze the exponential number of potential pipelines, picking the right pipeline that combines the proper preprocessing and algorithms is a hard task that requires knowledge and experience. In front of the complexity of building a right ML pipeline, algorithm portfolios aim to drive algorithm selection, learning from the past in a continuous process. However, building a portfolio requires that (i) data scientists develop and test pipelines and (ii) portfolio maintainers ensure the quality of the portfolio and enrich it. The firsts are the developers, while the seconds are the operators. In this paper, we present a set of criteria to be respected, and propose a pipeline-based meta-model, to support a DevOps approach in the context of Machine Learning Pipelines. The exploitation of this meta-model, both as a graph and as a logical expression, serves to ensure continuity between Dev and Ops. We depict our proposition through the simplified study of two primary use cases, one with developer's point-of-view, the other with ops'.
The different versions of the original document can be found in:
Are you one of the authors of this document?