Minimum number of data points estimation for supervised ML problems

“How many data points are needed to fit a machine learning model?”. This question is ubiquitous in the planning of a data science project. An estimation of the appropriate sample size is crucial in terms of time, resource allocation, and model quality. Literature review shows a lack of effective methods to address the problem. With this thesis, we propose an original work that does not suffer from the main limitations of the other available approaches. It is based on a metamodel that predicts the minimum number of points required depending on the characteristics of the dataset under consideration. Moreover, we specify what conditions have to be met in the planning of a project before the above question can be formulated.