Stabilizing Estimates of Shapley Values with Control Variates

In layman's terms, Shapley values quantify how much information is gained from being told the value of each feature. Shapley values are among the most popular tools for explaining predictions of blackbox Shapley values are rarely computed exactly, as the machine learning models. However, their computational cost is exponential in the number of high computational cost motivates the use input features. Rather, they are typically estimated of sampling approximations, inducing a considerable using the Shapley Sampling or KernelSHAP algorithm degree of uncertainty. To stabilize (Lundberg and Lee, 2017; Strumbelj and Kononenko, these model explanations, we propose ControlSHAP, 2010, 2014). These algorithms, however, are subject an approach based on the Monte to sampling variability; as a result, running the same Carlo technique of control variates. Our procedure twice may yield different estimated Shapley methodology is applicable to any machine values, including different estimated orderings of learning model and requires virtually no extra features. This instability raises questions about the computation or modeling effort. On several trustworthiness of insights gleaned from Shapley values.