École Nationale Supérieure de
Techniques Avancées

Abstract

The increasing complexity of machine learning models, particularly deep neural networks for weather forecasting and large language models, has created a critical need for robust Explainable AI (XAI) techniques. These "black box" models operate on high-dimensional data, making it difficult to understand their predictions and trust their outputs. This work evaluates the effectiveness of XAI methods, specifically the Anchors technique, in providing actionable insights across two diverse data domains: tabular data for fairness auditing and high-dimensional meteorological data for weather prediction. We adopt a two-tiered methodology. First, we apply Anchors to the Folktables dataset to audit a binary income classifier, demonstrating its ability to identify precise, high-precision rules that reveal model vulnerabilities and biases, such as a reliance on sensitive attributes like gender. This provides a granular understanding of discriminatory mechanisms, a crucial step towards targeted bias mitigation. Second, we adapt and extend the Anchors framework for regression tasks to analyze a UNetR++ model trained on the Titan meteorological dataset. Our method successfully pinpoints the most influential spatial regions and variables (e.g., wind patterns, precipitation) for specific forecasts, transforming Anchors into a tool for regional influence analysis and proactive vulnerability mapping for extreme weather events. Our results show that the Anchors method provides focused, direct, and actionable insights regardless of data complexity. It proves to be a versatile tool not only for diagnosing societal biases in algorithms but also for auditing the physical reasoning of complex forecasting models. This work establishes a foundation for using example-based explainability as a guide for actionable intervention, enhancing trust and providing a pathway for improvement in both technical and societal systems.