Clustering of transactions can find potential useful patterns to improve the product profit. In this paper, a two-step clustering algorithm——CATD is proposed, applicable in large transaction databases. First, the database is divided into partitions in which transactions are partially clustered into a number of subclusters. A hierarchical clustering algorithm is used to control the distance between these subclusters. In the global clustering, a k-medoids clustering algorithm is performed on the subclusters to get a set of k global clusters and identify noise. The algorithm is feasible for large databases because it only scans the original databases once and the clustering process can be performed in main memory due to the partitioning scheme and the support vector representative of subclusters.