mulang-cmu/school-dropout-detection

Explainable Early Warning System for School Dropout Prediction in African Education Systems # School Dropout Detection **Explainable Early Warning System for School Dropout Prediction in African Education Systems** Authors: Martin Mulang' & Collaborators Target: Deep Learning Indaba 2026 --- ## Project Overview This project builds an explainable machine learning system that: 1. **Predicts** which students are at risk of dropping out 2. **Explains** why using gradient-based saliency mapping 3. **Tracks** risk trajectories using Lyapunov stability analysis ### Key Innovation Most dropout models just say "High Risk." Our system says: > "Student #247 has 52.3% dropout risk. Primary driver: Peer Stability Index collapsed by 0.4 points (∇ = −0.41). Recommended intervention: peer mentoring program." --- ## Data Sources ### Young Lives Ethiopia (Primary Dataset) - **Source**: UK Data Service (Study 7483) - **Coverage**: 2,999 children tracked across 5 rounds (2002-2016) - **Observations**: 14,995 child-round records - **Variables**: 214 features including education, household, health - **Dropout Rate**: ~20.5% (614 children dropped out) ### Kaggle Education in Africa (Country-Level) - **Coverage**: 54 African countries, 2010-2023 - **Observations**: 756 country-year records - **Variables**: 60+ indicators on enrollment, attendance, teachers, expenditure --- ## Model Architecture ``` ┌─────────────────────────────────────────────────────────────┐ │ TWO-TIER ARCHITECTURE │ ├─────────────────────────────────────────────────────────────┤ │ │ │ TIER 1: Country-Level (Kaggle) TIER 2: Student-Level │ │ ┌─────────────────────────┐ ┌───────────────────┐ │ │ │ XGBoost Classifier │ │ Neural ODE │ │ │ │ - Country risk score │ │ - Lyapunov V(x) │ │ │ │ - Macro indicators │ │ - dV/dt tracking │ │ │ └─────────────────────────┘ └───────────────────┘ │ │ │ …