What are Minimum Viable Machine Learning Models?

In an ideal world we would priortise all our machine learning projects according to their expected payoff. This would require a reasonable estimate of both the probability of success of a project and the return given success. Creating such estimates are difficult and instead we tend to be guided by heuristics, implicit and explicit, about the size of a problem and its difficulty.

Difficulty estimation is often limited to a discussion about what it would take to feasibly productionise a model. Consideration of the difficulty of building the model, to the extent that it is done at all, is usually an informal process relying on the experience of data scientist working on the project.

We have spent some time developing a structured approach to this problem. The MinViME application allows you to enter the fundamental business parameters of a problem you are looking to solve, and then returns an estimate of the minimally performant model that satisfies them.

The technique has been implemented into the open source application MinViME (Minimum Viable Model Estimator) which can be installed via the PyPI python package management system, or downloaded directly from the GitHub repository.