Comparing AI Frameworks for Predicting E-Waste Generation: A Practical Evaluation for Scalable E-Waste Management

Abstract

This paper presents a comprehensive, reproducible study that compares classical machine learning models for predicting electronic waste (e-waste) generation at regional scales. Accurate forecasting of e-waste volumes is a critical enabler for efficient collection planning, resource allocation, and policy design in circular-economy efforts. We assemble and harmonize official e-waste records with socio-economic and device-penetration covariates, design a standardized forecasting pipeline, and evaluate a set of classical models—ARIMA, Linear Regression, Random Forest, Gradient Boosting variants (XGBoost, LightGBM, CatBoost) and Support Vector Regression—using time-series cross-validation. Models are compared on predictive accuracy, robustness to missing data, computational cost, and interpretability. We report best-practice feature engineering, hyperparameter search spaces, and provide recommendations for practitioners and policymakers working in resource-constrained settings. The code, dataset processing scripts, and experiment logs are released to ensure reproducibility.