TensorForest Estimator

TensorForest is a highly scalable implementation of random forests built by combining a variety of online HoeffdingTree algorithms with the extremely randomized approach.

TensorForest estimators are used to implementing the following algorithm:

Initialize the variables and sets
    Tree = [root]
    Fertile = {root}
    Stats(root) = 0
    Splits[root] = []

Divide training data into batches.
For each batch of training data:
    Compute leaf assignment for each feature vector
    Update the leaf stats in Stats
    For each  in Fertile set:
        if |Splits| < max_splits
            then add the split on a randomly selected feature to Splits
        else if is fertile and |Splits| = max_splits
            then update the split stats for 
    Calculate the fertile leaves that are finished. 
    For every non-stale finished leaf:
        turn the leaf into an internal node with its best scoring split 
        remove the leaf from Fertile
        add the leaf's two children to Tree as leaves
    If |Fertile| < max_fertile
        Then add the max_fertile ? |Fertile| leaves with 
        the highest weighted leaf scores to Fertile and 
        initialize their Splits and split statistics. 
Until |Tree| = max_nodes or |Tree| stays the same for max_batches_to_grow batches 

More details of this algorithm implementation can be found in the TensorForest paper.