Longitudinal and Graph-Augmented Prediction of Adolescent Substance Use Onset in the ABCD Study

Temporal XGBoost outperforms cross-sectional and graph-based models for predicting adolescent alcohol and marijuana use onset in the ABCD Study dataset of 11,860 participants. While Temporal Graph Convolutional Networks using family and school graphs do not exceed standalone temporal XGBoost performance, their risk scores offer complementary predictive signal when combined.