FP-growth based algorithms (used in CPSTree [34], CanTree [21]) use a bottom-up tree traversal method to build a projection-tree (or conditional-tree in FP-growth algorithm) for each data item. If transactions have many items, the con-
struction overhead of a projection-tree by bottom-up traversing is high, because all nodes need to be visited, regardless
of their frequency in the base-tree. On the other hand, we can improve the performance of a mining algorithm with the top-down traversing method in GTree construction.