|
Abstract:
|
There is a body of recent empirical work that attributes many instances of algorithmic bias to distributional shifts between training and real-world data. Broadly speaking, research has identified two types of algorithmic bias caused by distributional shifts: (i) the model is trained to predict the wrong target, and (ii) the model is trained to predict the correct target, but its predictions are inaccurate for demographic groups that are underrepresented in the training data. Statistically speaking, the first type of algorithmic bias is caused by posterior-drift between the training and real-world data. while the second type of algorithmic biases is caused by covariate shift between the training and real-world data. Unfortunately, current algorithmic fairness practices were not developed with this statistical perspective of algorithmic bias in mind. In this roundtable, we bring together stakeholders and statisticians to align their goals and promote mutual understanding.
|
