Fairness as symmetry: New framework cuts bias by 90% with minimal accuracy loss
A novel approach treats bias as a symmetry violation, slashing discrimination without needing causal graphs.
A new preprint from Nishit Singh (arXiv, 2026) introduces a principled approach to detecting and mitigating bias in machine learning by treating fairness as a symmetry operation. The core idea: a classifier is fair if its predictions remain unchanged under a counterfactual swap of a sensitive attribute (e.g., flipping gender or race) while holding all merit-related features fixed. Bias, then, is a symmetry breaking — and the solution is to restore that symmetry through loss-based regularization.
On synthetic datasets with varying noise, correlation, and bias levels, the framework reduces fairness violations by over 90% while incurring only about 5% accuracy cost. Crucially, it does not require causal graph knowledge, is computationally lightweight, and generalizes to any sensitive attribute that can be represented as a bit-flip. This makes it a strong candidate for real-world high-stakes applications like hiring, lending, and criminal justice, where traditional debiasing methods are often too complex or assume unavailable causal structures.
- Formalizes bias as a symmetry-breaking operation: fairness means classifier outputs remain invariant under counterfactual sensitive attribute flips.
- Loss-based regularization restores symmetry, achieving over 90% violation reduction with only ~5% accuracy drop across four synthetic datasets.
- No causal graph knowledge required; method is lightweight and generalizes to any sensitive attribute representable as a bit-flip.
Why It Matters
A practical, low-cost bias mitigation technique that works without causal graphs — deployable now in high-stakes AI systems.