Fairness by Learning Orthogonal Disentangled Representations

Spyros Avlonitis, Alexander Papadatos, Danai Xezonaki

Keywords: fairness, learned representations, orthogonal, disentangled

Abstract: In order to assess the reproducibility of the original paper, we implemented the models proposed and ran the experiments with the listed, in the paper, datasets. We implemented three encoder-discriminator models for the Tabular, CIFAR, and YaleB datasets. Due to missing information regarding the models' architecture, we had to incorporate some assumptions. In addition, for some of the models, we had to make assumptions regarding datasets' versions, preprocessing, and targets' definition. Our experiments for the German and Adult dataset approached the reported accuracies by the authors. However, we were not able to reproduce the results for CIFAR-10, CIFAR-100, and YaleB datasets.

Paper Url: https://openreview.net/forum?id=uGujCu6SEH3