Interpretable Salary Prediction in the AI Sector Using Ensemble Learning with Uncertainty Quantification

Abstract

The rapid expansion of the global artificial intelligence and data science job market has intensified the need for accurate salary estimation to guide career decisions and recruitment strategies. Existing predictive models often face challenges in generalizing across diverse international markets due to high-cardinality categorical features such as job titles and the wide economic variance between geographic regions. This study introduces a comprehensive machine learning framework that employs domain-informed feature engineering, including eight-category job clustering and country tiering, combined with a Super Learner ensemble of gradient boosting machines. Evaluation on a dataset of 151,445 records reveals that the Super Learner ensemble achieves a cross-validated R^2 of 0.3047 ± 0.0076 and a held-out test R^2 of 0.2818, the highest among all 21 evaluated algorithms. Although the raw R^2 margin over standalone gradient boosting is not statistically significant at α=0.05, the Super Learner demonstrates superior bias calibration (GMR = 1.0021), better uncertainty quantification (82.3% empirical PI coverage), and a 7.8 percentage-point lift over a mean-prediction baseline at the Acc@20% threshold.