The fairness-accuracy landscape of neural classifiers

That machine learning algorithms can demonstrate bias is well-documented by now. This work confronts the challenge of bias mitigation in feedforward fully-connected neural nets from the lens of causal inference and multiobjective optimisation. Regarding the former, a new causal notion of fairness is introduced that is particularly suited to giving a nuanced treatment of datasets collected under unfair practices. In particular, special attention is paid to subjects whose covariates could appear with substantial probability in either value of the sensitive attribute. Next, recognising that fairness and accuracy are competing objectives, the proposed methodology uses techniques from multiobjective optimisation to ascertain the fairness-accuracy landscape of a neural net classifier. Experimental results suggest that the proposed method produces neural net classifiers that distribute evenly across the Pareto front of the fairness-accuracy space and is more efficient at finding non-dominated points than an adversarial approach.