Seminar: "Regional effect plots for the interpretation of black box machine learning models"

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/

Presentations pdf can be found here