Interpretable machine learning for predicting school absenteeism at-scale in Brazil: evidence from 22,000 schools

School attendance is a fundamental prerequisite for children's and adolescents' academic, social, and personal development. Yet, despite decades of research and policy efforts, rates of problematic school absenteeism remain persistently high. This study investigates the predictive potential of school-level factors to identify schools at secondary level in Brazil at risk of elevated absenteeism. Using data from 22,476 schools, we consolidate three main categories of predictors: school infrastructure, human resources, and regional characteristics, combined with attendance records and demographic information. Five machine learning models were compared, and Random Forest was selected due to its balance of predictive performance, interpretability, and statistical confirmation through the DeLong test. Building on this, we apply interpretable machine learning techniques, particularly SHapley Additive exPlanations (SHAP), to reveal how the importance of predictors varies across different school complexity levels, defined as the structural and organizational conditions under which schools operate. Furthermore, we conduct counterfactual "what-if" scenarios, systematically perturbing key features to explore how policy-relevant adjustments could shift schools from high to low absenteeism risk. Our results highlight how structural, staff-related, and contextual features interact differently depending on school complexity, offering actionable insights to inform more equitable educational interventions and policy decisions.