Smart training and smart deployment of machine learning models

A computer-implemented method, comprising:

• accessing, by a system operatively coupled to a processor, a training dataset of training data candidates that are available for training of a machine learning model, wherein each training data candidate of the training dataset comprises a defined set of features;
• computing, by the system, a first group comprising first respective feature distributions for the features of the defined set for an entirety of the training dataset of training data candidates;
• iteratively constructing, by the system, from an empty set, a strict subset of training data candidates by inserting a different training data candidate from the training dataset to the strict subset at each iteration, until a second group comprising second respective feature distributions for the features of the defined set for an entirety of the strict subset of data training candidates matches the first group according to at least one statistical hypothesis test, wherein the strict subset is less that the entirety of the training dataset and mitigates at least of overtraining or overfitting of the machine learning model being trained using the training dataset;
• training, by the system, the machine learning model on the strict subset of training data candidates;
• and
• after the training of the machine learning model:
• accessing, by the system, a dataset of deployed data candidates that are available for inferencing using the machine learning model after being deployed for use on real-world data;
• and
• determining, by the system, whether a third group comprising third respective feature distributions for the features of the defined set for the dataset of deployed data candidates matches the first group according to the at least one statistical hypothesis test.

A computer program product for facilitating smart training and smart deployment of machine learning models, the computer program product comprising a computer-readable memory having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:

• access a training dataset of training data candidates that are available for training of a machine learning model, wherein each training data candidate of the training dataset comprises a defined set of features;
• compute a first group comprising first respective feature distributions for the features of the defined set for an entirety of the training dataset of training data candidates;
• iteratively construct, from an empty set, a strict subset of training data candidates by inserting a different training data candidate from the training dataset to the strict subset at each iteration, until a second group comprising second respective feature distributions for the features of the defined set for an entirety of the strict subset of data training candidates matches the first group according to at least one statistical hypothesis test, wherein the strict subset is less that the entirety of the training dataset and mitigates at least of overtraining or overfitting of the machine learning model being trained using the training dataset;
• train the machine learning model on the strict subset of training data candidates;
• and
• after the training of the machine learning model:
• access a data candidate on which the machine learning model is executable, and
• determine a level of confidence for execution of the machine learning model on the data candidate, wherein the level of confidence is based on a distance between at least one feature of the data candidate and at least one of a mean or median of at least one of the first respective feature distributions of the first group.