To develop multiple machine learning models to identify college students at high risk of social network addiction (SNA) and to select key predictors, so as to provide evidences for tiered screening and interventions.

A multi-stage cluster sampling was conducted among 12 541 undergraduates at five universities in Guizhou Province from September to October 2024. The Chinese version of Social Networking Addiction Scale (SNAS-C) was used to classify high-risk individuals. Candidate variables included demographic characteristics, the total scores of Adult ADHD Self-report Scale (ASRS) and the Chinese version of Diagnostic and Statistical Manual of Mental Disorders, 5th Edition Self-rated level 1 Cross-cutting Symptom Measure (DSM-5 CCSM), academic stress and academic satisfaction and 24-hour Movement Guidelines. Four machine learning models including multilayer perceptron neural network (MLP), XGBoost, K-nearest neighbors (KNN), and gradient boosting decision trees (GBDT) were trained with an 8∶2 train-test split. Model performance was evaluated using accuracy, recall, F1 score, and the area under the ROC curve (AUC), followed by feature importance analysis.

The detection rate of high-risk SNA was 8.60%, higher in girls (10.10%) than boys (6.74%), with a statistically significant difference (χ²=44.42, P < 0.01). Compared with the low-risk group for SNA, the high-risk group had higher ASRS scores, more severe of depression and anxiety, greater academic stress, and lower academic satisfaction (t=-21.23, -16.12, -16.73, -13.81, 11.58, all P < 0.01). Regarding 24-hour movement behaviors (physical activity, sleep, sedentary/screen time), the high-risk SNA group showed lower compliance rates (11.77%, 5.65%, 57.92%) than the low-risk group (15.57%, 13.16%, 75.32%) (χ²=11.04, 50.72, 154.40, all P < 0.01). All four models demonstrated good discrimination; XGBoost performed the best (accuracy 88.18%, F1=0.881), while KNN achieved the highest AUC (0.90). Feature importance analysis indicated ASRS as the most critical predictor of SNA, followed by academic stress, depression, age, and academic satisfaction.

Machine learning models can effectively identify college students at high risk for SNA. ADHD symptoms, internalizing problems, and academic stress are major risk factors, while 24-hour movement behaviors are key intervention targets, especially sleep and screen time.