GEMEX: Model-Agnostic XAI via Geodesic Entropic Manifold Analysis

KÖSE U.

8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2026, Ankara, Türkiye, 21 - 23 Mayıs 2026, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/ichora69329.2026.11537244
  • Basıldığı Şehir: Ankara
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: fisher information matrix, GEMEX, machine learning, model-agnostic explainability, riemann information geometry, statistical manifold, XAI
  • Süleyman Demirel Üniversitesi Adresli: Evet

Özet

This study introduces GEMEX (Geodesic Entropic Manifold Explainability), a novel model-agnostic Explainable Artificial Intelligence (XAI) framework. GEMEX transforms a trained machine learning model into a structure that defines a statistical manifold using the Fisher Information Matrix (FIM) as a Riemann metric tensor and derives all explanations from the intrinsic geometry of this manifold. It uses Geodesic Sensitivity Field (GSF) for re-parametrizationinvariant feature attribution through a 4th-order Runge-Kutta geodesic integrator, the Parallel Transport Interaction (PTI) matrix for holonomy-based pairwise feature interactions, the Riemannian Curvature Triplet (RCT) to get three-way feature interactions via the Riemann curvature tensor, the Riemannian Saliency Tensor (RST) for principal explanation directions, the Feature Attention Sequence (FAS) for analysis of geodesic attention trajectory, and the Bias Trap Detector (BTD) for geometric bias and confounder detection. GEMEX supports tabular, time series, and image data; offers a fully modelagnostic approach that can wrap any model with probabilistic estimations / predict_proba() output. GEMEX can also produce spatially consistent image descriptions without requiring access to model-specific structure or gradients. GEMEX has been evaluated through multiple tests including comparative ones with SHAP, LIME, and ELI5. The study introduces the GEMEX and presents the promising findings from the tests.