Predictors of loneliness at school: An Explainable Artificial Intelligence approach on a large-scale cross-cultural assessment

There is extensive evidence linking loneliness to adverse mental health outcomes in adolescents. During adolescence, it is possible to experience loneliness even when surrounded by peers if the quality of relationships and personal expectations do not align, as often happens in schools, where adolescents spend a significant portion of their time with classmates. Yet, research on school-based loneliness predictors is hindered by small sample sizes and a restricted number of variables tested in models. To fill this research gap, this study investigates predictive factors for loneliness at school using data from 60,498 students (30,909 aged 10 and 29,589 aged 15), derived from the SSES survey, which includes data from 10 large cities across 4 different continents. Adopting a data-driven approach, we employ 3 machine learning algorithms (Elastic Net, Random Forests, and XGBoost) alongside eXplainable Artificial Intelligence (XAI) techniques to analyze the relationship between a broad range of psychosocial variables and loneliness. Results showed that bullying is the strongest predictor for increases in loneliness in 10-year-olds, with its influence persisting but diminishing in 15-year-olds. In contrast, two personality traits, namely extraversion and emotional stability, emerge as key predictors for decreases in loneliness in both 10- and 15-year-olds. Additional important predictors of loneliness include low-quality relationships with parents or friends and high screen use. By integrating classical and interpretable machine learning techniques, this study provides a nuanced understanding of the relative importance of a large number of predictors of loneliness at school, also taking into account nonlinear relationships within the data.