AUEB Stats seminar (Christos Diou)
Papastamoulis Panagiotis
<papapast@yahoo.gr>AUEB STATISTICS SEMINAR SERIES 2023-24
Christos Diou (Assistant Professor | Department of Informatics and Telematics, Harokopio University of Athens)
Regional effect plots for the interpretation of black box machine learning models
FRIDAY 19/4/2024, 11:15
Room: Troias Amphitheater
ABSTRACT
During the past decade, multivariate and highly nonlinear machine
learning models such as deep neural networks have become popular in
research and development of solutions for multiple practical
applications. An inherent drawback of such methods, which hinders their
further adoption, is the inability to describe their properties based on
the value of their parameters, in contrast to simpler linear or
polynomial models. To this end, global effect methods such as Partial
Dependence Plots (PDP) and Accumulated Local Effects (ALE) have been
proposed in the bibliography. These methods can be used to succinctly
describe the global model behaviour, however they become less useful
when there is high heterogeneity in feature effects.
This talk will have three parts: In part 1, we will introduce global
feature effect methods, such as PDP and ALE. In part 2, we will present
Robust and Heterogeneity-aware Accumulated Local Effects (RHALE,
ECAI-2023), a method designed to quantify and minimize the uncertainty
of the global explanation, also taking advantage of
auto-differentiation, in the case of differentiable models. Finally, in
part 3, we will discuss regional effect plots, where explanations are
produced for selected subregions of the feature space to minimize the
uncertainty of the explanation. In the practical examples of this talk
we will also introduce "effector", a newly developed python package for
global and regional effect plots, which can be found at https://xai-effector.github.io/