AUDA: An Analytical Workflow for Sparse and Heterogeneous Plastic-Related Data

Abstract

Sparse and heterogeneous plastic-related data complicate the construction of analytical datasets and the evaluation of predictive models. This study presents AUDA, a workflow that organizes human-reviewed records extracted by PRISM into country–year analytical tables and supports descriptive analysis, interpolation, and forecasting. The analytical database contains over 28,000 records from 195 countries and regions. The case studies use smaller datasets constructed from available observations, along with an external global plastic-production series. Correlation and random-forest feature-importance analyses describe associations with plastic waste generation. Predictive evaluations compare interpolation, statistical, and machine-learning methods using repeated masking and rolling-origin forecasting. Cubic spline interpolation and Gaussian process regression achieved the lowest mean weighted absolute percentage error in the Japan and United States resin-consumption interpolation tasks, respectively. Theil–Sen regression achieved the lowest error using this metric in global plastic-production forecasting, while persistence and ARIMA tied for the lowest error in forecasting Japanese plastic-waste generation. These rankings are specific to the evaluated series and data partitions. An exploratory sensitivity analysis also examines how a standard-error tolerance coefficient and the chosen complexity ordering affect hyperparameter selection; it does not establish improved performance on unseen data. AUDA provides workflows for analyzing available records but does not reconcile differences in source measurement methods or estimate policy effects.