Research Focus

My research goal is to develop approximation techniques for the Shapley value and Shapley interactions. Both notions stem from game theory and convince with a theoretical foundation inspired by economics. The Shapley value fairly distributes collective benefit achieved by a group of players. It is commonly used in Explainable AI to attribute feature importance providing understanding of black box models to the human user or developer. Further, it is used in machine learning on a more broader scale for data valuation and the selection of features, base learners in ensembles, or neurons in neural networks. Shapley interactions capture synergies between players and how their cooperation results in additional benefit beyond the addition of the individual player's contribution. Recently, theses are increasingly used to reveal interactions between features which do not independently influence a model's prediction, such as one would expect from today's highly complex models. I develop approximation algorithms on a sound analytical basis such that these can be applied across the wide spectrum of domains in machine learning to reliably generate explanantions containing Shapley values and Shapley interactions.

Publications