In addition, when choosing the question at each node, only a small, random subset of the features is considered. With these two modifications, each run may result in a slightly different tree.