Preliminary Investigation on the Stability of Bias-aware Classifiers

Abstract

We here discuss a bias-aware classifier, which is designed to predict a class while removing the bias caused by the influence of specific information. Such type of classifiers are used for, say, taking fairness into account by removing socially sensitive information. We define the stability of such bias-aware classifiers as how similar predictions are made from the same information other than the bias-source information to be removed. We report the stability on the datasets that is influenced by different kinds of cognitive biases, collected through a crowdsourcing service.