Attribution via Distributional Paths for Information Revelation
Researchers have introduced Reveal-IG, a novel method for feature attribution in machine learning models. This technique shifts from input-space paths to a space of structured probe distributions, offering more control over how features are queried. Reveal-IG aims to provide stable, signed attributions and has shown promise in image classification and tabular regression tasks. AI
IMPACT Enhances interpretability of AI models by providing more stable and signed feature attributions.