Capuchin: Causal Database Repair for Algorithmic Fairness — mentioned by hansvm

Thu, 09 Jul 2020 05:34:37 GMT

There are plenty of options if you know the kinds of bias you want to protect against. I'm having trouble finding the project to reference, but one solution I saw recently used a data preprocessing (reweighting) step that would remove bias in any of a broad class of models built on the processed data. The crux of the idea was that you can pick any properties against which your model should be biased (e.g. gender), additionally _allow_ such a bias if it came through other specified properties (e.g. gender bias in college admissions is arguably fine if it exists solely because one group only…

Paper and discussion history. A mention is not an endorsement.

Capuchin: Causal Database Repair for Algorithmic Fairness · HackerPapers discussions

Capuchin: Causal Database Repair for Algorithmic Fairness — mentioned by hansvm