Representation Learning

Explanation

Representation learning fairness checks that the internal embeddings a model learns do not encode protected attributes — so that, for example, a customer-segmentation embedding does not implicitly cluster people by gender, caste, or religion and then leak that into downstream decisions. It matters because even when protected fields are excluded from inputs, biased embeddings act as hidden proxies that reintroduce discrimination. It is measured by Bias in Learned Embeddings, computed as Embedding Bias = Mutual Information(Embedding_Dimension, protected_Attribute); for customer segmentation the goal is to ensure embeddings don't cluster by protected attributes. To implement it, extract the learned embedding vectors for a labelled sample, estimate the mutual information between embedding dimensions and each protected attribute, and log per-dimension and aggregate MI scores as evidence, visualising clustering to confirm groups are not separable by protected status. Run it during model development and after any retraining of the representation. The threshold requires a mutual information score below 0.2; if MI reaches or exceeds 0.2 it triggers a breach — apply mitigation such as adversarial debiasing or removing leaky features, re-measure, and document sign-off, in line with RBI Fairness and Equity and NIST MAP 2.1/4.1 data-bias analysis.