Stability analysis of game-theoretic feature attributions for machine learning models

Event Start
Event End
Location
Building 1, Level 3, Room 3119

Abstract

In this work, we study the feature attributions of machine learning (ML) models originating from linear game values and coalitional values, defined as operators on appropriate functional spaces. The main focus is on random games based on conditional and marginal expectations. It is well-known from the Rashomon effect that, under predictor dependencies, distinct models that approximate the same data well can have different representations. To understand the impact of the Rashomon effect on the explanation maps, we formulate a stability theory for these explanation operators by establishing certain bounds for both marginal and conditional explanations. We elucidate the differences between the two games, showing that marginal explanations can become discontinuous in some naturally designed domains, while conditional explanations remain stable. In the second part of our work, group explanation methodologies are devised using game values with coalition structures, where features are grouped based on dependencies. We analytically show that computing group attributions this way has a stabilizing effect on the marginal operator and allows for the unification of marginal and conditional explanations. 

Brief Biography

Alexey Miroshnikov is a Senior Principal Research Scientist at Discover Financial Services. Previously, he held various academic positions: Assistant Adjunct Professor at UCLA Mathematics Department (2016-2019), Postdoctoral Research Associate in the Department of Biostatistics and Epidemiology at UMass Amherst (2015-2016), and Visiting Assistant Professor in the Department of Mathematics and Statistics at UMass Amherst (2012-2015). During his Ph.D., he was fortunate to work as a researcher at the Institute of Applied and Computational Mathematics at FORTH in Crete, Greece.

Contact Person