16 and One More Reason to Know SHAP

The article, written by a researcher named Sabrina, provides an in-depth look at SHAP, a method for explaining machine learning model predictions based on cooperative game theory. It begins with the classic Shapley value introduced by Lloyd Shapley in 1953, which offers a unique fair allocation of a team's winnings based on each player's average marginal contribution. In 1988, Charnes, Golany, Keane, and Rousseau showed that the Shapley value can be computed as the solution of a weighted least squares problem, which later became the basis for KernelSHAP. The first adaptations to machine learning came in 2010 by Štrumbelj and Kononenko, who introduced a permutation-based estimation using Monte Carlo sampling, and in 2014 extended to subsets, revealing interactions and redundancy. The pivotal moment was the 2017 paper by Lundberg and Lee, 'A Unified Approach to Interpreting Model Predictions', which introduced the SHAP framework with properties like Local Accuracy, Missingness, and Consistency, and proved uniqueness. KernelSHAP estimates SHAP values by solving a weighted linear regression with a shapley kernel that emphasizes small and large coalitions. Despite its strengths, the article notes that the original implementation of background sampling may violate assumptions, leading to the development of 15 extensions to address various issues. The title '16 and one more reason' refers to the 15 extensions plus the axiomatic basis as the first reason, with a hidden final reason revealed at the end of the article.