iPDP: On Partial Dependence Plots in Dynamic Modeling Scenarios

Real-world AI systems change over time, but explanation tools haven't caught up: Traditional methods for understanding AI decisions assume models stay the same after training. This paper addresses a critical gap by developing tools that can track how model behavior shifts as new data arrives—essential for applications like fraud detection or medical diagnosis where patterns evolve.

The proposed iPDP method can detect when models are behaving differently due to changing data patterns: The framework successfully identifies "concept drift" (when underlying patterns shift) through a single adjustable parameter, offering practitioners a practical way to monitor whether their AI systems are still making decisions for the right reasons.

Explainability tools need to be as dynamic as the models they explain: By extending partial dependence plots—a standard technique for showing which features matter most—to work with continuously updating models, this work demonstrates that responsible AI requires explanation methods designed for real deployment conditions, not just laboratory settings.

The approach is model-agnostic and tested on real-world scenarios: iPDP works across different types of AI systems and was validated on actual data streams, suggesting it could be practically implemented across various industries where models must adapt to changing conditions while remaining interpretable.