Random forest: how to derive the characteristics of the predicted classes? Decision Trees stratify the feature space into different regions and fit a basic model to each region (average for regression, most frequent class for classification). With this method it's not difficult to derive the characteristics of the individuals falling in each predicted class or even (I'm not sure) to derive the characteristics of the group of individuals that are missed.
The problem for the decision trees is its accuracy. Random Forest can resolve this problem efficiently but is there any way to derive the 2 points above as in Decision Trees?
 A: If you train a RF model in R on some training set (let's call it data.train) using model=randomForest(Y~., data=data.train), where Y is the column of data.train corresponding to the target variable, you can use the command predict(model) to get the "out-of-bag" predictions for your model. Then, it is easy to compare whether or not the predicted class differs from the true class for each observation in the training set: something along the lines of data.train[which(predict(model)!=Y),] will get you all the observations (the rows of data.train) that were missed. Once you have this subset of your training set, you can investigate it to try to see if these observations match a particular profile (with respect to the predictor variables).
Now, if you're using cross-validation, or a separate test set, let's call it data.test, you can also get the predictions using predict(model,newdata=data.test), and then investigate the observations for which wrong predictions were generated in the exact same way.
The "out-of-bag" (OOB) predictions are obtained as follows: any observation in data.train can be fed to all the trees in the Forest that were trained on boostrap samples devoid of this observation. Approximately one third (~37%) of the total number of trees will meet this condition. Further, by letting these tree vote and taking the most popular class, a prediction (the OOB prediction) can be obtained for the observation. 
