Information-Theoretic Bias Reduction via Causal View of Spurious Correlation

Spurious correlations are a root cause of algorithmic bias: The paper identifies that AI systems often rely on irrelevant features (like background details in face recognition) rather than the actual target feature, and proposes measuring this through information theory to catch bias that traditional accuracy metrics miss.

Current fairness metrics may create a false sense of progress: The researchers found that existing accuracy- or score-based bias measurements can be "gamed" through superficial adjustments to prediction scores without actually fixing the underlying bias, highlighting the need for deeper measurement approaches.

A new measurement technique using conditional mutual information can identify feature-level bias: The proposed method specifically measures which features the AI model is inappropriately relying on, offering a more precise diagnostic tool than previous approaches for understanding where bias originates.

Debiasing can work without explicit bias labels: The paper demonstrates an unsupervised debiasing technique using stochastic label noise that doesn't require manually labeled bias information, potentially making bias reduction more practical to implement across different applications.

The approach was validated across multiple real-world scenarios: Testing on standard benchmarks in diverse realistic settings suggests this information-theoretic framework could be broadly applicable beyond specific domains like face recognition to other AI fairness challenges.