On Counterfactual Explanations (Lecture 25-05-2020 at 13:00 CET)

As per our schedule, the online monthly meeting will take place next Monday the 25th of May at 1pm CET. Mr Martin Pawelczyk from the Tuebingen group will deliver a talk entitled “On Counterfactual Explanations“. The abstract of the talk and a brief CV are given below.


Counterfactual explanations (CE) can be obtained by identifying the smallest change made to an input vector to influence a prediction in a positive way from a user’s viewpoint; for example, from ’loan rejected’ to ’awarded’ or from ’high risk of cardiovascular disease’ to ’low risk’. The literature distinguishes two types of counterfactual explanations: those with data support and those with lowest costs (i.e. smallest change). In the first part of this talk, we will present a new method to generate counterfactual explanations with data support that draws ideas from the manifold learning literature. Our method relies on data density approximators in the form of variational autoencoders. In the second part of the talk, we introduce the notion of greedy counterfactual explanations. Greedy CEs are obtained by identifying the smallest change made to an input to change a prediction made by a fixed model. We then compare data supported counterfactual explanations to those greedy explanations along several dimensions (both theoretically and empirically).

Short Bio

Martin Pawelczyk is a PhD candidate at the Data Science and Analytics Lab of the University of Tübingen. Prior to this, he obtained his M.Sc. in Statistics (Research) at the London School of Economics (LSE), where he focused on statistical learning. Before graduating from LSE, he studied Econometrics at the University of Edinburgh (M.Sc.) and Economics at the University of Cologne (B.Sc.). During his graduate studies, he was supported by the German National Merit Foundation. As a first year PhD student, his research interests include machine learning, fairness and explainability.