Explainable AI: A Unified Approach Based on Cooperative Game Theory

Explainable Artificial Intelligence (XAI) aims to enhance the interpretability of machine learning (ML) models, yet existing feature-based explanation methods remain fragmented across local and global approaches. This thesis presents a unified framework based on cooperative game theory to systematically characterize and distinguish feature-based explanations. By integrating functional ANOVA decompositions with the Möbius transform (MT) and Shapley Interactions (SIs) from cooperative game theory, we bridge perturbation- and gradient-based local explanations, as well as sensitivity-based and performance-based global explanations, providing a comprehensive perspective on feature attributions, joint feature effects and feature interactions. To address computational challenges of SIs, we introduce SHAP-IQ, an efficient model-agnostic approximation method for any-order SIs, and extend KernelSHAP to higher-order interactions. Additionally, TreeSHAP-IQ and GraphSHAP-IQ improve model-specific computations for tree-based models and graph neural networks (GNNs), respectively. Recognizing the dynamic nature of real-world AI systems, we further develop incremental Permutation Feature Importance (iPFI) for online learning environments, ensuring efficient tracking of global feature importance during evolving model behaviors. Our contributions offer a principled and scalable approach to XAI, advancing the robustness and applicability of feature-based explanations across diverse ML paradigms.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Bielefeld

Year

2025

Author dc:creator

• Fumagalli, Fabian

Repository record source_url

https://pub.uni-bielefeld.de/record/3006445

OAI identifier oai:identifier

oai:pub.uni-bielefeld.de:3006445