Unbiased machine learning-assisted approach for conditional discretization of human performances

Performance discretization maps numerical performance values to ordinal categories or performance ranking labels. Norm-referenced performance discretization is extensively applied in human performance evaluation such as grading academic achievements and determining salary increases for employees. These tasks stipulate a common condition that certain performance ranking labels might have no associated performance values and are referred to as conditional discretization. Currently, the only statistical method available for norm-referenced performance discretization is Z score, which merely addresses partial conditions. To achieve a fully conditionally norm-referenced performance discretization, this article proposes four novel approaches enlisting a multi-modal technique that incorporates unsupervised machine-learning algorithms and a heuristic method as well as a novel decision function ensuring conditional unbiasedness. The machine-learning-based methods demonstrate superiority over the heuristic one across most testing data sets, achieving a conditional unbiasedness degree ranging from 0.11 to 0.82. On the other hand, the heuristic method notably outperforms for a specific data set, exhibiting a conditional unbiasedness degree up to 0.76. Leveraging the strengths of these constituent methods enable the effectiveness of the proposed multi-modal approach for conditionally norm-referenced performance discretization.