随机森林中的袋外误差图

Question

I tried to fit Random Forest to my data set to do a classification between Control and Alzheimer group. In the first try I got the left OOB error plot and in the second try that I decreased the number of variables in my data set, I got the right side OOB error plot. My problem is comparing these two plots, what is the better OOB plot?should the class error for Alzhemier and Control be close to OOB curve of the Forest? if yes why?

Answer 1

The plot on the right has a better OOB error. I assume that the Alzheimer and control lines are also OOB errors but calculated for the particular classes. The random forest predictor is constructed by bootstrapping a fraction of the samples, the OOB error is calculated on the samples that were not selected (out of the bag) on each iteration of the algorithm. Therefore, OOB error is an estimation of the performance as you build the model as described by Breinman , and smaller errors are of course better.

"should the class error for Alzheimer and Control be closer to OOB curve of the Forest?." It depends on how good your model is at predicting each class. If the classification error is similar for both classes, then the OOB error will be close to both.