Machine Learning-Based Classification and Feature Analysis of Heterogeneous Environmental Sustainability Disclosure

Environmental sustainability disclosure has become increasingly critical as climate risks intensify and regulatory and investor demands for transparent, decision-useful information continue to grow. It plays a key role in reducing information asymmetry and supporting capital allocation, risk assessment, and regulatory oversight. However, prior studies predominantly rely on aggregated ESG indicators and linear models, which often fail to capture the structural heterogeneity and nonlinear relationships inherent in environmental data. This study develops a machine learning-based analytical framework to examine environmental disclosure using corporate data from the Taiwan Economic Journal (TEJ) from 2022 to 2024. A polarized sampling design is employed by selecting firms in the top and bottom 20% of ESG performance to identify and compare the distinctive disclosure characteristics of companies with high versus low environmental performance. Five models are evaluated using Accuracy, Precision, Recall, F1-score, and AUROC. The results show that ensemble models outperform traditional approaches, with CatBoost achieving the most robust performance. Feature importance analysis reveals a concentrated structure dominated by carbon emissions, energy efficiency, and waste management, while the importance of renewable energy variables increases over time. These findings highlight the nonlinear and multidimensional nature of environmental disclosure and demonstrate the value of machine learning in enhancing environmental sustainability analysis, investment decision-making, and regulatory effectiveness. As this study is based on a single-country dataset (Taiwan), future research may incorporate cross-country datasets to improve external validity.