Bridging the Drift-Performance Paradox: A Causal Framework for Proactive Anomaly Attribution in Deployed Machine Learning Systems

This research addresses the 'drift-performance paradox' observed in deployed machine learning systems, where statistically significant drifts in features or predictions do not always lead to measurable performance degradation. The proposed work aims to develop a proactive, causal inference-driven framework that (1) investigates under what real-world conditions feature or prediction drift translates into actual performance degradation, and (2) attributes root causes of anomalies before they become operationally significant. The framework will integrate causal discovery algorithms to map relationships among input features, model predictions, and performance metrics over time; counterfactual analysis to simulate 'what-if' scenarios estimating potential performance impacts; and explainable anomaly attribution modules that provide interpretable rationales for why certain drifts matter. This approach contrasts with existing monitoring systems by bridging drift and performance through causality, prioritizing actionable, context-aware insights to reduce false positives, prevent performance collapses, and enhance trustworthiness. The framework is designed to generalize across domains such as supply chain, predictive maintenance, and environmental modeling, enabling proactive model monitoring with improved transparency and interpretability.

References:

1. A monitoring framework for deployed machine learning models with supply chain examples. Bradley Eck, Duygu Kabakci-Zorlu, Yan Chen, France Savard, Xiao-Hua Bao (2022). 2022 IEEE International Conference on Big Data (Big Data).