The present invention discloses a fast algorithm for mining high utility itemsets, wherein some transaction data and item utilities are recorded in a tree structure. The method to construct a tree structure is recording on a node the item utilities appearing from the root node to the node. Some techniques are used to reduce the mining space, whereby the fast algorithm can directly generate high utility itemsets from the tree structure without generating any candidates. The fast algorithm of the present invention is more efficient than the existing highest-efficiency algorithm. The present invention further proposes a compression method to effectively save memory space.