Local vs global importance
Let’s assume a simple model described as the interaction of X1 and X1 in values from 0 to 1 [0,1].
f(x1,x2)=x1∗x2
Globally both variables are equally important, because the model is symmetrical.
But if the instance we want to explain x_∗ has x1=0 and x2=1. Then the importance of X1 is larger than X2:
- h1x∗(z)=z as x2=1 for any value of z.
- h2x∗(z)=0 as x1=0 for any value of z.