Why Shapley Value Regressionο
When building a regression model, one of the most common questions is:
Which predictors are most important in explaining or predicting the target variable?
A common approach is to examine:
regression coefficients (beta values)
standardized coefficients
correlations
sequential changes in \(R^2\)
These methods are computationally inexpensive and easy to interpret, but they can become misleading when predictors are correlated or share explanatory information.
Shapley Value Regression provides an alternative perspective by allocating model explanatory power among predictors using principles from cooperative game theory.
Shapley value regressionο
Shapley Value Regression is a different strategy for assessing the contribution of regressor variables to the target variable. It originates from cooperative game theory (Shapley, 1953).
In this framing, the \(R^2\) from the linear model \(y = X\beta + u\) is treated as the total value of a cooperative game. The predictors in \(X\) are the players, and their individual contributions are imputed from the observed joint explanatory power.
An algorithm to impute individual contributionsο
Let there be m regressors in the model \(y = X\beta + u\).
Let \(X(p, r)\) be an
r-member subset of \(X\) that contains regressorp.Let \(X(q, r)\) be an
r-member subset of \(X\) that does not contain regressorp.Let \(R^2(p, r)\) be the \(R^2\) from regressing \(y\) on \(X(p, r)\).
Let \(R^2(q, r)\) be the \(R^2\) from regressing \(y\) on \(X(q, r)\).
Then, the share of regressor p (that is, \(x_p \in X\)) is:
with \(R^2(q, 0) = 0\), and k equal to the number of evaluated cases.
The sum of all shares equals total model explanatory power: