FWCE: A Feature-Weighted Calibrated Ensemble for Interpretable Tabular Risk Classification

Abstract

Tabular risk classification remains a central machine-learning problem because many high-value datasets are small, heterogeneous, and difficult to model with very large neural architectures. This paper presents FWCE, a feature-weighted calibrated ensemble for interpretable binary classification on structured clinical-style records. FWCE first treats zero-coded physiological fields as missing-like values, applies median imputation and z-score normalization, estimates feature relevance using standardized class separation, and trains three complementary learners: L2-regularized logistic regression, Gaussian naive Bayes, and weighted k-nearest neighbors. A validation set is then used to select nonnegative soft-voting weights and a probability threshold. We evaluate FWCE on the Pima Indians Diabetes dataset with 460 training examples, 154 validation examples, and 154 held-out test examples. Compared with logistic regression and kNN baselines, FWCE achieves 67.19% F1, 0.841 AUC, and a Brier score of 0.153. The results show that simple feature calibration and validation-based ensembling can improve recall-oriented performance while preserving transparent model components suitable for audit and revision.