On Minimizing the Impact of Dataset Shifts on Actionable Explanations

Explore a 26-minute conference talk from the Uncertainty in Artificial Intelligence (UAI) 2023 Oral Session 1 that delves into minimizing the impact of dataset shifts on actionable explanations. Discover the challenges associated with providing actionable explanations for algorithmic decisions in light of the Right to Explanation regulatory principle. Learn about the factors influencing explanation stability when models are retrained to handle dataset shifts. Gain insights from rigorous theoretical analysis and extensive experimentation on real-world datasets, which highlight the key factors determining explanation (in)stability, including model curvature, weight decay parameters, and the magnitude of dataset shift. Understand how these factors significantly impact the stability of explanations produced by state-of-the-art methods, and explore the implications for maintaining valid and actionable explanations in practice.

UAI 2023 Oral Session 1: On Minimizing the Impact of Dataset Shifts on Actionable Explanations