Scored Rule Sets: A Unified Framework for Interpretable Classification, Regression, and Cluster Explanation

Abstract

Interpretable machine learning on tabular data has traditionally relied on fragmented rule-induction ecosystems, where distinct paradigms are used for discrete and continuous targets, and transparent post hoc explanation of unsupervised structures is rarely integrated under a single model class. This article presents a unified, task-aware formalization of the scored-rule-set framework that spans multiclass classification, continuous regression, and transparent cluster explanation while preserving logical comprehensibility and modularity as primary design goals. By establishing an explicit task-aware formalization R=(S,τ,Gτ) that decouples rule-condition discovery from task-specific score-vector semantics and commutative aggregation operators, we consolidate foundations established in prior classification research, introduce a native regression runtime with scikit-learn compatibility, and formulate a two-stage surrogate pipeline for explaining arbitrary clustering assignments via transparent rule sets. Across extensive empirical evaluations on standard tabular benchmarks, we compare predictive performance, explanation fidelity, and model simplicity across heuristic, genetic programming, and multi-objective evolutionary backends. The results demonstrate that task-aware scored rule sets achieve competitive accuracy across both supervised regimes and high explanation fidelity in unsupervised settings, establishing scored rule sets as a versatile, holistic foundation for transparent tabular machine learning.