खंड 3 No. issue3, May 2026, pg 2320-2336 (2026): Beyond Accuracy: Leakage-Controlled Evaluation of Interpretable and Ensemble Classifiers Using Average Precision and Probability Calibration

					View खंड 3 No. issue3, May 2026, pg 2320-2336 (2026): Beyond Accuracy: Leakage-Controlled Evaluation of Interpretable and Ensemble Classifiers Using Average Precision and Probability Calibration

Classifier evaluation becomes unreliable when predictive ranking, probability quality and the integrity of the evaluation process are treated as interchangeable properties. This article develops an auditable framework for comparing interpretable and ensemble classifiers under controlled information boundaries. Average precision is designated the primary ranking measure; probability assessment combines the Brier score, logarithmic loss, reliability diagrams and calibration intercept and slope. The framework separates base-model fitting, calibration, model selection and final testing, and records the information permitted at each stage. Its exploratory contribution is a two-stage decision rule: reject pipelines that violate information boundaries, then retain models that offer defensible trade-offs between ranking, probability quality and structural complexity. A fully specified four-observation example establishes that identical rankings and classification accuracy can coexist with materially different probability losses. In this example, average precision remains 0.8333 for two prediction vectors, whereas Brier scores are 0.2250 and 0.40505. These are analytical illustration values, not classifier benchmark results. An experimental protocol is provided for subsequent comparison of logistic regression, shallow decision trees, random forests and gradient boosting. The framework offers a testable route to more accountable model comparison without assuming that ensemble complexity necessarily improves deployment suitability. Empirical superiority, generalisation and practical utility remain to be established through preregistered multi-dataset experiments.

Keywords: average precision; probability calibration; data leakage; interpretable classification; ensemble learning; reproducible evaluation; model selection

प्रकाशित: 2026-11-14