Greenwashing Identification and Multidimensional Driving Mechanism of Heavily Polluting Enterprises Based on Interpretable Machine Learning

Against the global green transition and tightening ESG disclosure requirements, corporate greenwashing undermines capital market transparency and environmental governance. However, existing studies are limited in semantic quantification, model interpretability, and multidimensional feature interaction analysis. Using Chinese heavily polluting listed firms from 2020 to 2024, this study constructs a greenwashing indicator based on the divergence between environmental textual semantics in annual reports and substantive ESG performance and develops a comprehensive feature system including firm characteristics, corporate governance, green development, and external pressure. Seven machine learning models are employed for prediction, and SHAP is used to interpret nonlinear effects and key drivers. The random forest model achieves the best performance, with green development contributing 37.7% (RF) and 48.3% (SHAP). Key variables include environmental disclosure quality (Eidq), CSR disclosure (CSR), monitored pollution status (Monitored), and firm size (Size), with identified thresholds and interaction effects. EBM robustness tests confirm the reliability of results, providing evidence for greenwashing regulation and ESG disclosure standardization.