Exhaustive learning techniques for machine learning algorithms

A non-transitory, computer-readable medium having instructions stored thereon that are executable by a computer system to perform operations comprising:

• performing an iterative machine learning operation that includes:
• training a first version of a machine learning model based on a current version of a training dataset that includes a plurality of data samples, wherein the first version of the machine learning model includes a plurality of decision branches;
• identifying, within the first version of the machine learning model, a particular node of a particular decision branch of the plurality of decision branches that does not satisfy a set of user-specified criteria, wherein the set of user-specified criteria includes one or both of a minimum net-to-gross loss ratio and a minimum number of users impacted;
• pruning, based on the identifying, one or more nodes in the particular decision branch that are at lower levels of the particular decision branch than the particular node;
• determining whether a first subset of the plurality of data samples included in the pruned particular decision branch satisfies one or more criteria in the set of user-specified criteria;
• based on the first subset satisfying one or more criteria in the set of user-specified criteria (a) removing the pruned particular decision branch from the plurality of decision branches included in the first version of the machine learning model to generate an updated version of the machine learning model and (b) updating the training dataset having a second plurality of data samples;
• outputting the particular decision branch from the first version of the machine learning model as a decision rule;
• and
• repeating the iterative machine learning operation using the updated version of the training dataset and the updated version of the machine learning model to produce a final trained version of the machine learning model.

A method, comprising:

• training, by a computer system, a first version of a machine learning model, using a white box machine learning algorithm, based on a first version of a training dataset that includes a plurality of data samples, wherein the first version of the machine learning model specifies a plurality of decision branches;
• selecting a first decision rule that is based on a first decision branch of the plurality of decision branches;
• identifying, within the first version the machine learning model, a particular node of the first decision branch that does not satisfy a set of user-specified criteria, wherein the user-specified criteria includes one or both of a false-positive rate (FPR) and a number of users impacted;
• pruning, based on the identifying, the first version of the machine learning model, including removing one or more nodes in the first decision branch that are at lower levels of the first decision branch than the particular node;
• determining, by the computer system, whether a first subset of the plurality of data samples included in the pruned first decision branch satisfies one or more criteria in the set of user-specified criteria;
• based on the first subset satisfying one or more criteria in the set of user-specified criteria (a) removing, by the computer system, the pruned first decision branch from the plurality of decision branches included in the first version of the machine learning model to generate a second version of the machine learning model and (b) updating the training dataset by removing the first subset from the plurality of data samples;
• outputting the first decision branch from the first version of the machine learning model as a decision rule;
• and
• training, by the computer system, a second machine learning model based on the second version of the machine learning model and the updated training dataset.