Resilience of Supervised Learning Algorithms to Discriminatory Poisoning of Training Data

EasyChair Preprint 6274

Abstract

We address the concern that training datasets for supervised learning could have been poisoned through discrimination by i) defining and modeling discrimination as a dataset poisoning, ii) proposing novel interventional mixtures to inhibit discrimination, iii) and evaluating these and other methods addressing discrimination on synthetic and real-world datasets.

Keyphrases: Data Poisoning, decision systems, discrimination, fairness, machine learning

@booklet{EasyChair:6274,
  author    = {Przemyslaw Grabowicz and Nicholas Perello},
  title     = {Resilience of Supervised Learning Algorithms to Discriminatory Poisoning of Training Data},
  howpublished = {EasyChair Preprint 6274},
  year      = {EasyChair, 2021}}