Different results when using oobPredict for TreeBagger

I've noticed that averaging the predicted OOB responses for individual trees using 'oobPredict(B, 'trees', i)' produces different results than when using oobPredict(B, 'trees', 'all'). Ex:

b = TreeBagger(no_trees,xtrain,ytrain,'method','r','OOBVarImp','off',... 'OOBPred','on','minleaf',1); yhat_oob_tree = zeros(no_trees,numel(ytrain)); for i = 1:no_trees % one row per tree yhat_oob_tree(i,:) = oobPredict(b,'trees',i); end avg_yhat_oob_tree = mean(yhat_oob_tree); %avg. by observations(columns) yhat_oob_all = oobPredict(b);

That is, 'avg_yhat_oob_tree' and 'yhat_oob_all' are different, doen anyone why this happens?

Thank you!

Answer 1

[Ilya]

See http://www.mathworks.com/matlabcentral/answers/9638-difference-between-individual-and-cumulative-oobmargin-of-treebagger It's the same problem here. When you ask for obbPredict from one tree, you get predictions filled with default values for in-bag observations.