Disentangling the non-linear relationships and interaction effects of urban digital transformation on carbon emission intensity

Abstract

The inexorable rise of urban digital transformation (UDT) underscores the imperative of comprehending its complex relationships with carbon emissions intensity (CEI). Existing studies primarily focus on the linear relationships between individual UDT variables and CEI, overlooking non-linear dynamics and interactive effects, which may result in incomplete estimations. To address these gaps, this study develops an interpretable machine learning (IML) model that integrates machine learning (ML) techniques and SHAP (SHapley Additive exPlanations), to uncover the non-linear relationships and interaction effects of UDT on CEI. The results reveal the following: (1) The proposed IML model achieves high accuracy in modeling the relationships between multiple UDT variables and CEI (R2 = 0.932, RMSE = 0.899, MAE = 0.543, 2); (2) Non-linear relationships between all UDT variables and CEI are confirmed, and two types of threshold points are identified where variable impacts shift from negative to positive and vice versa; (3) Interactive effects among UDT variables are examined, with thresholds quantified and U-shaped and inverted U-shaped trends identified. These findings provide a foundation for policymakers and urban managers to implement strategies that simultaneously advance digital transformation and promote low-carbon development.