Abstract
Graduate tracer studies inform curriculum review and career services, but descriptive summaries may miss nonlinear interactions and provide limited guidance for targeted support. This study developed and explained a deep neural network for predicting current unemployment using 2,878 Build Bright University tracer-study responses collected in 2023 and 2024. The binary modeling sample contained 2,813 respondents with classifiable employment status. Predictors were restricted to information available during study—campus, sex, degree, faculty, program, employment category, position, and salary—to reduce target leakage. A class-weighted multilayer perceptron (MLP) was compared with XGBoost and logistic regression. Five-fold cross-validation on the 2023 cohort selected decision thresholds, and the unseen 2024 cohort was used for temporal validation. Integrated gradients and grouped permutation importance explained the MLP globally and locally, while error rates were audited by sex and campus. Current unemployment was 6.2% overall but increased from 4.1% in 2023 to 11.8% in 2024. The MLP achieved ROC-AUC=.939, PR-AUC=.759, F1=.737, and MCC=.703 in temporal validation. It was useful but did not outperform XGBoost on F1 and MCC or logistic regression on ROC-AUC, PR-AUC, and recall. Both explainability methods identified employment during study as the dominant predictor; shuffling it reduced MLP PR-AUC by .589. Secondary feature rankings were less stable across explanation methods. Sex-specific true-positive and false-positive gaps were small, although campus estimates were unstable where unemployed counts were low. The findings support explainable predictive analytics as a human-reviewed early-support tool, not an automated eligibility or employment decision system.
Keywords: graduate tracer study; deep learning; explainable AI; integrated gradients; unemployment prediction; temporal validation
About the presenter
Join us at the Forum, 5–6 November 2026, Institute of Technology of Cambodia.
Are you the author? Sign in to My NRF to share this abstract.