Under-representation and error

Measure how error on the smaller group grows as its share shrinks, repeated over several seeds.

IntermediateAlgorithmic Bias Lab
العربية

The idea

Sweeps several shares and reports mean and standard deviation of error per group.

Data description

Synthetic data generated in your browser from stated distributions; no real personal data. Feature–outcome relations differ between groups by the heterogeneity you set.

Suggested next experiment

Concepts in this experiment

Representation bias

Occurs when some groups are under-represented in training data, so the model learns the majority pattern and errs more on the minority. Gender Shades showed commercial systems erred far more on darker-skinned women.

Scientific sources for this experiment

  1. Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification (PMLR 81:77–91)J. Buolamwini, T. Gebru (2018) — Paper
  2. A Framework for Understanding Sources of Harm throughout the ML Life Cycle (arXiv:1901.10002)H. Suresh, J. Guttag (2019) — Paper

More in this module