Abstract: Discovering association rules is a basic problem in data mining. Finding frequent item sets is the most expensive step in assnciation rule discovery. Analysing a frequent pattern growth (FP-growth) method is effieient for mining both long and short frequent patterns without candidate generation，but FP_growth would generate a huge nnmber of conditional FP-trees and then occupied memory space ,so proposing a new efficient algorithm not only heirs all the advantages in FP-growth method, but also avoids its bottleneck. By Establishing several conditional FP_trees（the number is equal the number of database’s items）to mine long and short frequent item sets,the improved algorithm could save memory space significantly.Performance study also shows that the improved method is effieient .