Estimating Shapley Values for Explainable AI via Richer Model Approximations (Teal Witter, CMC)

Abstract: Modern machine learning is ultimately a simple process: We iteratively update the weights of machine learning models to minimize a problem-specific loss. When it works well, we deploy the model in human-facing domains like healthcare, finance, or the justice system. But even though we know how models are trained, we don’t understand why they make decisions the decision they do. A particularly compelling approach to explaining AI predictions is the Shapley value, a game-theoretic quantity that measures how each input to the model affects its output. Mathematically, the ith Shapley value is the average change in the ith dimension of a particular function defined on the d-dimensional hypercube. Because the hypercube has 2^d points, exactly computing Shapley values is infeasible. In this talk, we will instead leverage algorithmic insights to develop state-of-the-art approximation methods.