ALOP-Kolloquium

Im Rahmen des Kolloquiums des Graduiertenkollegs Algorithmic Optimization findet am

Montag, dem 23. Januar 2023 16:00 Uhr c.t. Hörsaal 9

folgender Vortrag statt:

Deriving Robust Counterfactual Explanations for Machine Learning Models

via Two-Stage Robust Optimization

Jannis Kurtz, Universiteit van Amsterdam

Nowadays machine learning models are often used to make decisions based on personal data which significantly affect individuals, e.g. the decision of granting a loan to a person or not. However many machine learning models behave like black-boxes and can incorporate biases contained in the data into their decision which is why in 2016 the European Union enacted the "right to explanation".

Counterfactual explanations (CE) play an important role in providing such explanations and hence improving explainability. The idea is to find the smallest changes which have to be made to a data point such that the model would have changed its decision. Unfortunately most of the common methods can only provide one CE which can be unrealistic to reach for the user.

In this work we present a model to calculate robust CEs, i.e. CEs where after changing each attribute in a certain range the perturbed data point still remains a CE. Hence our method provides a whole set of CEs and the user can choose the one which is most reasonable. We use algorithmic ideas from two-stage robust optimization to calculate robust CEs for the most popular ML methods including logistic regression, decision trees and neural networks. We show that our method is able to calculate CEs efficiently for common data sets.