Causal Shapley Values: Exploiting Causal Knowledge to Explain Individual Predictions of Complex Models

Shapley values underlie one of the most popular model-agnostic methods within\nexplainable artificial intelligence. These values are designed to attribute the\ndifference between a model's prediction and an average baseline to the\ndifferent features used as input to the model. Being based on solid\ngame-theoretic principles, Shapley values uniquely satisfy several desirable\nproperties, which is why they are increasingly used to explain the predictions\nof possibly complex and highly non-linear machine learning models. Shapley\nvalues are well calibrated to a user's intuition when features are independent,\nbut may lead to undesirable, counterintuitive explanations when the\nindependence assumption is violated.\n In this paper, we propose a novel framework for computing Shapley values that\ngeneralizes recent work that aims to circumvent the independence assumption. By\nemploying Pearl's do-calculus, we show how these 'causal' Shapley values can be\nderived for general causal graphs without sacrificing any of their desirable\nproperties. Moreover, causal Shapley values enable us to separate the\ncontribution of direct and indirect effects. We provide a practical\nimplementation for computing causal Shapley values based on causal chain graphs\nwhen only partial information is available and illustrate their utility on a\nreal-world example.\n