Randomforests And Bootstrap Aggregating
As we know in simple averaging methods RFs perform better than Bootstrap Aggregating or bagging by considerably reducing the high variance (see bias-variance tradeoff) of a single classifier by continuous independent sampling with replacement from the same distribution(dataset). This performance gain of RFs is described very well in this Leo Breiman(2001) paper and also in The Elements of Statistical Learning : Hastie,Tibshirani,Friedman in the section RandomForests.
I will briefly describe the idea of performance gain in Rfs over bagging here.Since in both the ensembling methods a single classifier(decision trees) ,which is very prone to high variance, when makes use of all the available features,it tends to over-fit since a more complex tree dependence structure is generated. Averaging over different classifiers, with low collinearity among them, decreases the over-all complexity of the final ensembled model. But including all the feature is not a good idea,since all the features might not help much in overall information gain for individual classifier and many a time can prove to be a source of noise.To overcome this problem L. Breiman proposed the idea of selecting features randomly from some fixed number of features..