Emile-Lucky-Muhigira/Modeling-School-Dropout-as-a-Dynamical-System-Neural-ODEs-with-Lyapunov-Stability-for-Early-Warning

I developed an early warning system for school dropout prediction using data from five East African countries. Our approach models dropout as a dynamical system using Neural Ordinary Differential Equations (Neural ODEs) with Lyapunov stability constraints. PROJECT TITLE: Modeling School Dropout as a Dynamical System: Neural ODEs with Lyapunov Stability for Early Warning in East Africa AUTHOR: Emile Lucky Muhigira AFFILIATION: Carnegie Mellon University Africa DATE: May 2026 I. PROJECT OVERVIEW ==================== This research presents a novel early warning system for school dropout prediction in Sub-Saharan Africa. Unlike conventional models that treat dropout as a static event, this system models educational disengagement as a continuous dynamical process. The project leverages a hybrid approach: 1. A Two-Tier Ensemble Model for broad screening across multi-country contexts. 2. A Neural Ordinary Differential Equation (Neural ODE) component for high-fidelity trajectory tracking and prioritization. The system was validated on a combined dataset of 158,684 students across five East African countries (Ethiopia, Kenya, Rwanda, Tanzania, and Uganda), achieving state-of-the-art performance in identifying at-risk populations within imbalanced datasets. II. KEY CONTRIBUTIONS & INNOVATIONS =================================== * DYNAMICAL MODELING: Introduced Neural ODEs to handle irregularly sampled longitudinal survey data, enabling the modeling of student disengagement as a continuous trajectory. * LYAPUNOV STABILITY ANALYSIS: Integrated control theory principles to produce a "Trajectory Signal". This allows practitioners to distinguish between students who are stabilizing and those whose risk is actively worsening. * TWO-TIER ARCHITECTURE: Developed a system that encodes country-level systemic awareness (national enrollment and dropout trends) combined with individual-level feature prediction to improve regional generalization. * SYSTEMATIC ERROR ANALYSIS: Conducted a rigorous post-hoc analysis identifying Rural Bias (78.8% of missed cases) and age-specific failure modes, providing a concrete roadmap for future data-driven educational policy. III. TECHNICAL SPECIFICATIONS ============================= * ALGORITH …