PEBAM: A Profile-Based Evaluation Method for Bias Assessment on Mixed Datasets

Bias evaluation methods focus either on individual bias or on group bias, where groups are defined based on protected attributes such as gender or ethnicity. More generally, however, descriptively relevant combinations of feature values in the data space (profiles) may serve also as anchors for biased decisions. This paper introduces therefore a semi-hierarchical clustering method for profile extraction from mixed datasets. It elaborates on how profiles can be used to reveal historical, representational, aggregation and evaluation biases in algorithmic decision-making models, taking as example the German credit data set. Our experiments show that the proposed profile-based evaluation method for bias assessment on mixed datasets (PEBAM) can reveal forms of bias towards profiles expressed by the dataset that are undetected when using individual- or group-bias metrics alone.