Purpose: This study applied the Interaction of Person-Affect-Cognition-Execution (I-PACE) model and the Relational Development System Theory (RDS) to identify key individual and contextual correlates of adolescents? problematic Internet use (PIU) with machine learning approaches. Methods: Data from 68,425 adolescents were analyzed using five ensemble models (AdaBoost, Random Forest, LightGBM, Bagging, CatBoost) within a nested cross-validation framework. Key factors were identified through SHapley Additive exPlanations (SHAP), while bivariate partial dependence analyses were used to identify interactions. Results: The prevalence of PIU risk was 23.2%. Five algorithms achieved comparable performance. CatBoost achieved the best performance and was selected as the final predictive model. SHAP values showed that the top 17 features explained nearly 80% of the model. At the individual level, intolerance of uncertainty was the strongest risk factor, whereas mindfulness was the main protective factor. Additionally, weekend video game time was a major behavioral risk contributor. At the contextual level, home-leaving intentions and bullying perpetration were identified as key family- and peer-related risk factors, respectively. Bivariate partial dependence analyses found both within-individual (e.g., mindfulness intolerance of uncertainty) and individual-contextual (e.g., mindfulness home-leaving intentions) interaction effects. Conclusions: This study applied five machine learning algorithms to identify key individual and contextual factors associated with adolescent PIU risk and their interactions. The results suggest that risk factors accumulate across systems and impair adolescents? adaptive capacity, whereas mindfulness exerts crosssystem effects that buffer these risks, offering implications for targeted interventions.  Kulcsszavak: adolescent problematic Internet use, machine learning, I-PACE model, RDS theory, prediction model