A Robust Method to Measure the Global Feature Importance of Complex Prediction Models

Abstract

Because machine learning has been widely used in various domains, interpreting internal mechanisms and predictive results of models is crucial for further applications of complex machine learning models. However, the interpretability of complex machine learning models on biased data remains a difficult problem. When the important explanatory features of concerned data are highly influenced by contaminated distributions, particularly in risk-sensitive fields, such as self-driving vehicles and healthcare, it is crucial to provide a robust interpretation of complex models for users. The interpretation of complex models is often associated with analyzing model features by measuring feature importance. Therefore, this article proposes a novel method derived from high-dimensional model representation (HDMR) to measure feature importance. The proposed method can provide robust estimation when the input features follow contaminated distributions. Moreover, the method is model-agnostic, which can enhance its ability to compare different interpretations due to its generalizability. Experimental evaluations on artificial models and machine learning models show that the proposed method is more robust than the traditional method based on HDMR.