Speaker: Prof Jiji Zhang, The Chinese University of Hong Kong
Abstract:
It is now common to call biased algorithmic predictions “unfair”, and many criteria of fairness have been proposed to serve as side constraints in the training of predictive models for optimal predictive accuracy or as parts of the objective function to be optimized along with some metric of predictive accuracy. This practice is usually interpreted as reflecting a trade-off between the accuracy and the fairness of algorithmic predictions. In this talk, I argue that the alleged trade-off is due either to a misconception of fairness (as applied to predictions) or to a misconception of accuracy (due to a misidentification of the target of prediction). I contend instead that when it comes to predictions, a “fairness constraint” is justifiable only if it contributes to (the properly defined) accuracy. From this perspective, I suggest that the causal approach to achieving fairness in algorithmic predictions is more reasonable than the purely statistical approach.
