Chunk #89 — The Methods — Variable Importance

Source

An introduction to recursive partitioning: rationale, application, and characteristics of classification and regression trees, bagging, and random forests.

Embedded

yes

Text

(1)VI(t)(Xj)=∑i∈B¯(t)I(yi=y^i(t))|B¯(t)|−∑i∈B¯(t)I(yi=y^i,ψj(t))|B¯(t)| where y^i(t)=f(t)(xi) is the predicted class for observation i before and y^i,ψj(t)=f(t)(xi,ψj) is the predicted class for observation i after permuting its value of variable Xj, i.e. with xi,ψj = (xi,1, …, xi,j−1, xψj(i),j, xi,j+1, …, xi,p). (Note that VI (t)(Xj) = 0 by definition, if variable Xj is not in tree t.) The raw importance score for each variable is then computed as the average importance over all trees