Primary Submission Category: Machine Learning and Causal Inference
Causal Importance of Features in Machine Learning
Authors: Bo Liu, Fan Li,
Presenting Author: Bo Liu*
Measuring the importance of predictors or features on the outcome variable is crucial in supervised learning. Existing measures of feature importance rely solely on the black-box models and are often sensitive to model misspecification, because they require extrapolation to predict the outcome for values far outside of the training set. The degree of necessary extrapolation depends on the overlap between the remaining predictors at two levels of a particular predictor, which is also a key concept in causal inference. Common approaches in causal inference to reduce sensitivity to outcome model specification include alternative methods such as propensity score weighting and double robust estimators, and alternative estimands such as the average treatment effect on population with sufficient overlap. We borrow these ideas and point out the connections between several commonly used feature importance measures in machine learning literature and causal estimands. We then propose several new robust measures of feature importance for black-box supervised learning models.