Class measuring biggest absolute difference in prospective child counts if split. More...
Class measuring biggest absolute difference in prospective child counts if split.
This is proportional to the L1 error between histogram estimate spn and histogram estimate child nodes of spn.
L1 error is 1/N * { | n- 2n_l | + |n - 2n_r | } where N is total data points in the entire histogram, n is data points in the node, n_l is data points going to left child and n_r is data points going to right child.
n_r = n - n_l so n = 2n_r = n - 2(n - n_l) = 2n_l - n So 1/N * { | n- 2n_l | + |n - 2n_r | } = 2/N * | n- 2n_l | and n - 2n_l = n- n_l - n_l = n_r - n_l is the difference between prospective child counts if split, so this measure can be treated as proportional to the L1 error between the histogram estimate on spn and the histogram estimate on the children of spn.