An Explainable AI Approach to Employee Performance Evaluation Using Predictive Modeling

Abstract

While employee performance analytics can aid in more consistent workforce evaluations, prediction models may exhibit unfair correlations and opaque predictions. The paper focuses on an analysis of the explainable machine learning approach to classification of employee performance grades based on a publicly available synthetic dataset with 100,000 records and 20 features. Four classifiers, namely multinomial logistic regression, decision tree, random forest, and histogram-based gradient boosting were compared by means of a stratified 80:20 train-test split. In a feature ablation study, both with and without inclusion of the monthly salary feature, logistic regression, random forest, and histogram-based gradient boosting showed 100% accuracy, and decision tree – 94.32%. With the removal of salary, all the classifiers showed almost similar performance around the 20% random guess baseline across five near-balanced classes. Additionally, salary proved to be most important in decision trees, contributing an estimated 76.36% towards the features used for classification. This is a clear indication that the predictive ability observed in this case can be a function of the salary feature as a proxy or possibly a leaking feature and not a reflection of the broader characteristics of employees. This is a demonstration of the reason why high accuracy is not enough proof of validity in employee-performance prediction. The data is entirely artificial in nature; hence the findings are diagnostic in nature only.