Shapley Values in Explainable Artificial Intelligence

Specialeforsvar ved Mads Herbert Kerrn

Titel: Shapley Values in Ecplainable Artificial Intelligence

Abstract: Explaining predictions made by complex machine learning models is an important prob- lem and increasing field of research. When high accuracy of model prediction is the goal, highly complex models are often preferred over classic statistical models. The price is lack of interpretability. Recent research has been focusing on methods to explain predictions from complex models by assigning a value to each feature that somehow corresponds to their contribution to the prediction. Some of the most popular methods rely on Shapley values, a concept from game theory, which is used to assign feature contribution for model predictions. We introduce the game theoretical foundation for Shapley values and define Shapley explanations, a novel generalization of methods relying on Shapley values. This generalization incorporates existing explanations and motivates new games for explana- tions based on Shapley values, including a method for explaining model predictions from structural causal models. The two most popular Shapley explanations rely on a marginal and conditional expectations respectively. We discuss the two and show how to fool expla- nations from both methods. We compare different explanation methods including LIME and methods based on Shapley values on simulated and real data and see that they agree to some extend when the features are independent and Gaussian distributed. Finally, we discuss our findings and the use of Shapley values in explainable artificial intelligence. 

Vejleder:  Niels Richard Hansen 
Censor:    Lars Nørvang Andersen, Aarhus Universitet