Agenda

Speaker: Agus Sudjianto

Automated Machine Learning (AutoML) has simplified model development by efficiently optimizing predictive models based on aggregate performance metrics. However, despite their widespread use, these metrics often obscure underlying model weaknesses, particularly in high-stakes decision-making applications where conceptual soundness and reliability are paramount. This paper highlights critical pitfalls in evaluating machine learning models purely based on aggregate metrics and introduces a comprehensive framework to diagnose, interpret, and improve model reliability.

We illustrate these challenges using a credit default prediction case study, demonstrating how high-performance models on leaderboards can exhibit severe flaws, such as non-monotonic feature effects, regional performance inconsistencies, and excessive sensitivity to input perturbations. Through a systematic validation approach—including interpretable modeling via depth-2 Gradient Boosted Decision Trees (XGBoost), supervised residual error analysis, distribution drift impact assessment, reliability quantification using conformal prediction, robustness evaluation under input perturbations, and a Mixture of Experts (MoE) modeling approach—we reveal how to detect and correct these vulnerabilities.

Our findings emphasize the necessity of rigorous validation frameworks that go beyond conventional metrics, ensuring conceptual alignment with domain knowledge, improving predictive stability, and enhancing resilience against distribution shifts. This study provides a structured methodology for practitioners to develop and deploy machine learning models that are not only high performance but also reliable, interpretable, and trustworthy in dynamic and high-stakes environments.