Uncovering nonlinear predictive patterns between artificial intelligence and green innovation efficiency in heavily polluting firms: an explainable machine learning approach

Abstract

Improving green innovation efficiency has become a central concern for firms pursuing sustainable development in the digital economy. Although artificial intelligence (AI) is increasingly incorporated into corporate strategies, existing studies have primarily relied on conventional linear econometric approaches to examine the association between AI-related characteristics and green innovation efficiency, which may be insufficient to capture the complex nonlinear predictive structures embedded in the data. To address this limitation, this study develops a prediction-oriented explainable machine learning framework to forecast firm-level green innovation efficiency. Using panel data from Chinese heavily polluting listed firms, we compare the predictive performance of Ordinary Least Squares (OLS) with six mainstream machine learning algorithms. Accumulated Local Effects (ALE) and SHapley Additive exPlanations (SHAP) are further employed to interpret nonlinear predictive relationships and feature contributions within the trained prediction model. The results show that ensemble learning models consistently outperform the linear benchmark in out-of-sample prediction, with CatBoost achieving the highest predictive accuracy and generalization capability. AI-related variables provide substantial incremental predictive information beyond conventional financial and governance characteristics. ALE analysis identifies pronounced nonlinear and stage-specific predictive patterns associated with AI-related features, with AI investment intensity exhibiting an inverted U-shaped predictive trajectory and AI attention demonstrating more stable predictive relevance across different ranges of model predictions. SHAP analysis further indicates that AI attention provides more consistent predictive information than AI investment and reveals heterogeneous predictive contribution patterns across governance and financial contexts. This study contributes to the literature by introducing an interpretable prediction framework for uncovering complex nonlinear predictive structures associated with AI-related characteristics and provides data-driven decision support for identifying firms with different expected levels of green innovation efficiency.