A SIMPLE AND ACCURATE CLASSIFICATION METHOD BASED ON CLASS ASSOCIATION RULES: ADAPTIVE SUPERVISED DISCRETIZATION WITH RANDOM FOREST ENSEMBLE

Abstract

Class Association Rule (CAR)-based classifiers combine the transparency of rule-based models with the predictive power of supervised learning, yet classical methods such as CBA, CMAR, and CPAR suffer from rule explosion, sensitivity to threshold parameters, and degraded performance on imbalanced datasets. This paper proposes an improved two-phase classification pipeline — Adaptive Supervised Discretization combined with a Random Forest ensemble (ASD-RF) — that addresses these limitations. In the first phase, continuous features are discretized using shallow decision trees trained per feature on the class label, deriving cut points that maximise class homogeneity within each bin. In the second phase, a Random Forest of 100 trees is trained on the resulting one-hot-encoded binary feature matrix. Empirical evaluation on two public benchmark datasets (Pima Indians Diabetes and Banknote Authentication) demonstrates that ASD-RF achieves mean accuracy of 90.14%, outperforming CBA (87.28%), CPAR (88.38%), CMAR (79.03%), and an unprocessed Decision Tree baseline (86.14%). The method is fully reproducible, requires no manual threshold tuning, and produces a compact, interpretable feature representation suitable for high-stakes domains such as healthcare and finance.