GeCo: Quality Counterfactual Explanations in Real Time

In this talk, I will provide an overview of our recent work on explanations of classification outcomes. In particular, I will present a system called GeCo, which computes high-quality counterfactual explanations in real time. To compute high-quality explanations, GeCo combines a constraint language with a customized genetic algorithm. To achieve real-time performance, GeCo employs two novel optimizations: a compressed data representation for candidate counterfactuals, as well as specialization of the classifier using partial evaluation techniques. In a comparison with four competitive systems, GeCo is the only system that can compute high quality explanations in real time. GeCo is presented in an upcoming VLDB paper, the preprint can be found here: https://arxiv.org/abs/2101.01292