Automated algorithmic decisions are widely used by organizations on a daily basis and thus developing efficient methods is of top priority. Organizational efficiency is a sophisticated target since there are multiple aspects that need to be optimized simultaneously, which include but are not limited to algorithmic fairness, i.e. minimize systemic bias against certain disadvantageous social groups, and economic efficiency, minimize the cost induced by wrong decisions.

With these targets in mind, we utilize a dual-focused Neyman-Pearson (NP) classification paradigm, which seeks minimal type II error under simultaneous control over both of the type I error and bias against fairness. By leveraging a LDA model, we develop for the first time an oracle framework for dual-focused NP classificat...[ Read more ]

Automated algorithmic decisions are widely used by organizations on a daily basis and thus developing efficient methods is of top priority. Organizational efficiency is a sophisticated target since there are multiple aspects that need to be optimized simultaneously, which include but are not limited to algorithmic fairness, i.e. minimize systemic bias against certain disadvantageous social groups, and economic efficiency, minimize the cost induced by wrong decisions.

With these targets in mind, we utilize a dual-focused Neyman-Pearson (NP) classification paradigm, which seeks minimal type II error under simultaneous control over both of the type I error and bias against fairness. By leveraging a LDA model, we develop for the first time an oracle framework for dual-focused NP classification. In particular, we propose finite-sample based classifiers that satisfy at population-level both the fairness constraint and type I error constraint with high probability, and derive oracle bounds on the excess type II error. It is worth noting that the new classifier does not require sample splitting, which was necessary for most of the existing NP methods, and thus further increase data efficiency. Numerical and real data analyses demonstrate its superiority.