Fair Graduate Employment Prediction for Chinese Universities: DML Assisted Stratified Panel XGBoost with Time Varying Loss

Abstract

The existing employment-prediction models overlook key challenges including panel-data endogeneity, cross-tier prediction unfairness, and year-on-year temporal distribution drift. To address these issues, this study proposes a novel deconfounded and fair employment-prediction framework, built upon panel stratification and a time-varying loss function within a DML-enhanced panel XGBoost architecture. From the Input-Process-Output (IPO) theoretical perspective, the raw dataset forms a ten-year balanced panel covering 30 universities over 2015-2024. After introducing a two-year lag for key predictors, the analytical sample spans 2017-2024 and yields 240 valid observations. Double Machine Learning (DML) mitigates confounding-driven bias to produce conditionally de-confounded associative estimates under explicitly stated identification assumptions for this observational panel dataset. Meanwhile, the time-varying loss function alleviates cross-tier prediction disparities amid temporal-distribution shifts, and panel stratification further reduces prediction biases across university subgroups. The evaluation suite includes baseline-model comparisons, ablation studies, and robustness analyses. Experimental results under the rolling-origin evaluation protocol demonstrate that the proposed method achieves lower prediction error and improved subgroup fairness relative to benchmark models. SHAP-based interpretability analysis reveals prominent nonlinear threshold and interaction effects between teaching quality and per-capita funding.