Using model ensembles is a popular technique that allows to improve pred=
iction accuracy and obtain additional statistical measures for every predic=
tion. Suppose we have an ensemble of M predictors {P1i,P2i,...,PMi} for a molecule <=
em>i. The ensemble average: 1/M=E2=88=91k Pk
CORREL measure uses ensemble predictions for molecules = i,j to determine similarity of their ensemble responses. It is calculated a= s Pearson linear or Spearman rank correlation coefficients.
The main assumption of CORREL is that if the two molecules are under-pre= dicted or over-predicted by the same ensemble models, then the ensemble res= ponse for these two molecules is "similar".
CORREL is used in the Associate Neural Network (AS= NN) method [Tetko 2008] to provide bias correction of ensemble predicti= ons.