Computes fast (relative to other implementations) approximate
Shapley values for any supervised learning model. Shapley values help to
explain the predictions from any black box model using ideas from game
theory; see Strumbel and Kononenko (2014)
for details.