Home | Publications | Kol25

Approximating the Shapley Value and Shapley Interactions

MCML Authors

Patrick Kolpaczki

Dr.

→ Group Eyke Hüllermeier
Artificial Intelligence and Machine Learning

Abstract

This dissertation develops efficient approximation algorithms for the Shapley value and Shapley interactions, enabling scalable fair attribution in cooperative games and machine learning. By introducing novel representations based on mean estimation and weighted regression with advanced variance reduction, the methods achieve high accuracy with fewer samples. The proposed domain-independent algorithms come with theoretical guarantees and are empirically shown to outperform existing approaches in explainable AI and model attribution tasks. (Shortened.)

phdthesis Kol25