Interpretable Surrogate Models for Explaining Black-Box Predictions in Data Science

Jaesik Jeong, Lusiana Efrizoni, Susi Erlinda — Sun, 04 Oct 2026 00:00:00 +0700

Background: Black-box machine-learning models often achieve strong predictive performance but provide limited transparency. Interpretable surrogate models approximate these models using simpler structures, although their explanations may not always be sufficiently faithful or stable.
Aims: This study evaluates global and local surrogate models for explaining black-box predictions and examines the trade-offs among fidelity, stability, complexity, sparsity, and computational efficiency.
Methods: Experiments were conducted on four public binary-classification datasets: Adult Income, Bank Marketing, Breast Cancer Wisconsin Diagnostic, and German Credit. Random Forest, XGBoost, and Multilayer Perceptron served as black-box classifiers. Their predictions were explained using global logistic regression, global decision tree, rule-based, local linear, and local decision tree surrogates. Each experiment was repeated using five random seeds. Explanation quality was assessed using fidelity, probability error, stability, directional consistency, structural complexity, sparsity, and computation time.
Result: The 40-rule global surrogate achieved the highest global fidelity of 0.918, followed by the depth-six decision tree at 0.915. The 15-feature local linear surrogate achieved the highest local fidelity of 0.927, while the depth-four local tree reached 0.916. Greater complexity improved fidelity but reduced the stability and simplicity of explanations. Global surrogates were generally more stable, whereas local surrogates provided stronger neighborhood-level approximation.
Conclusion: Interpretable surrogate models provide useful but approximate explanations of black-box predictions. Medium-complexity configurations offer a more practical balance among fidelity, stability, and interpretability. Surrogate selection should therefore consider the purpose of explanation, dataset characteristics, model architecture, and explanation reliability.

International Journal of Advances in Artificial Intelligence and Machine Learning

Interpretable Surrogate Models for Explaining Black-Box Predictions in Data Science