A Random PRIM Based Algorithm for Interpretable Classification and Advanced Subgroup Discovery

Abstract

Machine-learning algorithms have made significant strides, achieving high accuracy in many applications. However, traditional models often need large datasets, as they typically peel substantial portions of the data in each iteration, complicating the development of a classifier without sufficient data. In critical fields like healthcare, there is a growing need to identify and analyze small yet significant subgroups within data. To address these challenges, we introduce a novel classifier based on the patient rule-induction method (PRIM), a subgroup-discovery algorithm. PRIM finds rules by peeling minimal data at each iteration, enabling the discovery of highly relevant regions. Unlike traditional classifiers, PRIM requires experts to select input spaces manually. Our innovation transforms PRIM into an interpretable classifier by starting with random input space selections for each class, then pruning rules using metarules, and finally selecting definitive rules for the classifier. Tested against popular algorithms such as random forest, logistic regression, and XG-Boost, our random PRIM-based classifier (R-PRIM-Cl) demonstrates comparable robustness, superior interpretability, and the ability to handle categorical and numeric variables. It discovers more rules in certain datasets, making it especially valuable in fields where understanding the model’s decision-making process is as important as its predictive accuracy.