Social norm bias: residual harms of fairness-aware algorithms

Many modern machine learning algorithms mitigate bias by enforcing fairness constraints across coarsely-defined groups related to a sensitive attribute like gender or race. However, these seldom account for within-group heterogeneity and biases that may disproportionately affect some members of group. In this work, we characterize Social Norm Bias (SNoB), subtle but consequential type algorithmic discrimination be exhibited models, even when systems achieve group objectives. We study issue through the lens in occupation classification. quantify SNoB measuring how an algorithm’s predictions are associated with conformity inferred norms. When predicting if individual belongs male-dominated occupation, framework reveals “fair” classifiers still favor biographies written ways align masculine compare techniques show it is frequently residual bias, post-processing approaches do not at all.