Targeted Learning for Variable Importance

Machine Learning (ML) models offer high-quality predictions for complex data structures and have become indispensable across various fields, including civil engineering (Lu et al., 2023), sociology (Molina and Garip, 2019), and archaeology (Bickler, 2021), due to their versatility and predictive power. However, due to their complexity, ML models present an absence of interpretability for their internal mechanism (Hooker and Hooker, 2017; Hooker et al., 2021; Freiesleben et al., 2024). To address this issue, many researchers proposed interpretable machine learning tools (IML) to provide post hoc interpretability of ML models. Among these tools, variable importance, which measures the contribution of individual covariates to the response variable, is a widely adopted measure in IML (Molnar, 2020). Traditionally, it has been applied to assess the performance of fixed models, such as random forests (Breiman, 2001) and linear models (Grömping, 2007). Additionally, efforts have been made to create model-specific uncertainty quantification methods, as seen in Gan et al. (2022).