How granular should waste composition predictions be? A compositional machine learning framework

Accurate waste composition data are essential for decision tools such as life cycle assessments and mathematical programming models. Yet existing models have key limitations: most predict only a few broad material categories and ignore the inherent interdependence in compositional data, potentially leading to biased estimates. This study introduces a machine learning framework for predicting county-level Municipal Solid Waste composition in the United States across 43 material categories. The framework combines a log-ratio transformation that renders compositional data suitable for machine learning, an Extra Trees regressor, and a compositional-data-consistent imputation method. We also extend a national dataset from 87 to 115 observations, increasing coverage from 21 to 33 states. The model achieves a variance-weighted R 2 of 0.29 in the transformed space, a twofold improvement over the corrected LASSO baseline. Repeated evaluation over 20 train-test splits reveals heterogeneous per-material performance: materials such as Newspaper, Magazine/Glossy, and Office paper are predicted reliably, while others such as Film Plastic and Electronics are not. Motivated by this heterogeneity, we propose a data-driven scheme that predicts only robustly predictable materials individually and groups the remainder, providing a principled way to balance compositional granularity against predictive accuracy. Improving accuracy further depends on the information provided by the predictors on the poorly predicted materials. Aggregating material subcategories into broader groups yields no improvement in predictions because subcategories grouped together can dilute or cancel each other's predictive signal when their relationships with predictors differ in strength and direction.