Information-Theoretic Bias Reduction via Causal View of Spurious Correlation

We propose an information-theoretic bias measurement technique through a causal interpretation of spurious correlation, which is effective to identify the feature-level algorithmic by taking advantage conditional mutual information. Although several methods have been proposed and widely investigated achieve fairness in various tasks such as face recognition, their accuracy- or logit-based metrics are susceptible leading trivial prediction score adjustment rather than fundamental reduction. Hence, we design novel debiasing framework against bias, incorporates regularization loss derived approach. In addition, present simple yet unsupervised based on stochastic label noise, does not require explicit supervision The approaches validated diverse realistic scenarios extensive experiments multiple standard benchmarks.