BigInsight congratulates Lars Henry Berge Olsen from BigInsight and the Department of Mathematics, who successfully defended his thesis "What’s in the Black Box? Improving the Shapley Value Explanation Methodology" for the degree of Philosophiae Doctor on February 7, 2025.
Main research findings:
Artificial intelligence and machine learning models are increasingly used in predictive tasks across various sectors, e.g., finance, healthcare, and retail, due to their high-performance accuracy. However, their predictions are often impossible to understand due to their opaque decision-making processes. To address this issue and comply with regulations like the EU AI Act, explainable AI (XAI) has emerged to explain how and why the model reached specific predictions. A prominent XAI framework is Shapley values, which originated in cooperative game theory in the 1950s but is today extensively used as a tool that can explain predictions made by any model.
This thesis improves the Shapley value explanation methodology in four key areas: developing methods to incorporate feature dependencies into the explanations, making them more precise and trustworthy; categorizing both existing and novel methods into distinct classes for thorough comparison and evaluation, with accompanying user recommendations; introducing stabilization and correction strategies to improve the accuracy of Shapley value approximations consistently; and creating the open-source shapr package for R and Python to ensure that these advancements are accessible to researchers and practitioners.
Adjudication committee:
Professor Marvin Wright, University of Bremen
Associate professor Inga Strümke, Norwegian University of Science and Technology
Associate professor Johan Pensar, University of Oslo
Supervisors:
Professor Ingrid Kristine Glad, University of Oslo
Dr. Martin Jullum, Norwegian Computing Center
Dr. Philos Kjersti Aas, Norwegian Computing Center
Lars Henry Berge Olsen
