An information theoretic approach to reducing algorithmic bias for machine learning

Algorithmic bias indicates the discrimination caused by algorithms, which occurs with protected features such as gender and race. Many researchers have tried to define the fairness and devise methods to mitigate bias, but it is still premature to reach the unanimous definition and evaluation metrics of fairness according to society, times and cultures. In this paper, we introduce three evaluation metrics such as parity difference, equalized opportunity and equalized odds that can deal with various definitions of algorithmic bias, and concretely divide the three general approaches further into seven methods with some challenges, resulting in relabeling, generation, fair representation (for pre-processing), constraint optimization, regularization (for in-processing), calibration and thresholding (for post-processing). Among them, the pre-processing method is widely used due to its versatility, but it has limitation to deal with the information on data and features related with bias appropriately. In order to preserve the characteristics of the original data while excluding the information about the features causing bias, we propose a preprocessing approach based on information theory that avoids collision in the dual optimization, where the latent space is divided into two subspaces. Experiments are conducted with the well-known benchmark datasets of Census and COMPAS, and two real-world tasks: facial emotion recognition and text sentiment analysis. The information theoretic approach is promising to achieve fair machine learning by reducing the bias caused by several features such as age, race and gender.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Algorithmic bias refers to discrimination caused by algorithms, often involving protected features such as gender and race. There is currently no unanimous definition and evaluation metrics for fairness, but this paper introduces three evaluation metrics and seven methods to address algorithmic bias, with pre-processing being widely used but having limitations.