Bagging applies bootstrap sampling to the train set, creating a number of N bootstrap samples. It then creates the same number N of base learners, using the same machine learning algorithm. Each base learner is trained on the corresponding train set and all base learners are combined by voting (hard voting for classification, and averaging for regression). The procedure is depicted as follows:
By using bootstrap samples with the same size as the original train set, each instance has a probability of 0.632 of appearing in any given bootstrap sample. Thus, in many cases, this type of bootstrap estimate is referred to as the 0.632 bootstrap estimate. In our case, this means that we can use the remaining 36.8% of the original train set in order to estimate the individual base learner's performance. This is called the out-of-bag score, and the 36.8% of instances are called out-of-bag instances.