In this paper the problem of discovering association rules among items in extremely large databases has been considered. a novel mining algorithm named improved cluster based association rules (ICBAR) has been proposed which can explore efficiently the large itemsets. achieving this and intializing the cluster table (where transaction records with length k are placed in kth cluster table), database will be once scanned. simultaneously itemset array (IA)) will be created. here kth element in the array to each itemset indicates number of transaction records in kth cluster table which have that itemset. presented method not only prunes considerable amounts of data by comparing with the partial cluster tables but also reduces the number of large candidate itemset that must be checked in each cluster through itemset arrays. performance and efficiency of proposed method has been compared with CBAR and apriori algorithms. experiments illustrate that ICBAR will do better than both of them.