ZCA Whitening Restores Isotropy for More Reliable WEAT Bias Measurements
August 10, 2026
Zero-phase Component Analysis (ZCA) whitening is proposed as a geometric pre-processing step to correct anisotropy in embedding spaces. By transforming the covariance matrix into an identity matrix, ZCA ensures the isotropy assumption required for accurate Word Embedding Association Test (WEAT) bias measurements across various model architectures.
HOW THIS AFFECTS YOU
●
researcherYou can mitigate geometric artifacts that lead to unreliable bias assessments in embedding-based evaluations.
●
policyThis improves the technical rigor of fairness audits performed on large language model embeddings.