Impact of Digitization and Artificial Intelligence on Carbon Emissions Considering Variable Interaction and Heterogeneity: An Interpretable Deep Learning Modeling Framework

Abstract

With growing attention on digitization and artificial intelligence (D&AI), interpreting their impact on urban carbon emissions has become critical. This study proposes an interpretable deep learning modeling framework considering variable feature interaction and heterogeneity to explore D&AI's carbon impacts. Random Forest (RF) is used to explore variable importance and optimize model parameters, followed by constructing a Decision Deep & Cross Feature-Transformation Network (DDCFTN) for high-fitting accuracy. SHapley Additive exPlanations (SHAP) and causal inference are employed to interpret variable effects and heterogeneity. Using data from 275 Chinese cities (2000–2021), DDCFTN outperforms traditional statistical or machine learning models (RMSE=579.88, MAE=440.91, $R_{adj}^2$=0.994). Key findings include: 1) D&AI's contribution to carbon emissions is underestimated—interactive effects increase the carbon impact by 665.569%. 2) Staged interaction patterns are observed: Digitization initially suppresses AI-related emissions (-34.017% to -96.361%) but later enhances them (13.191% to 45.353%). 3) Asymmetrical interactions exist, with AI's impact on digitization's emissions peaking at just 0.119%, following an inverted U-shaped trend due to retro-regulation effects. 4) City characteristics (e.g., Location Conditions and Urban Scale) reshape the heterogeneity of D&AI emissions through chain reactions, acting as key antecedents. This study introduces a novel analytical paradigm for interpreting D&AI's carbon impact and guiding urban decarbonization policies.