Task-Parallel FP-Growth on Cluster Computers

Frequent itemset mining (FIM) is one of the most deeply studied data mining task. A number of algorithms, employing different approaches and advanced data structures, have already been proposed to solve the task efficiently. Even the fastest serial FIM algorithms fail to scale up with the rapid growth of database sizes. Hence, parallel FIM algorithms are the only viable solutions in many domains as serial so- lutions have almost reached the physical barriers. To this end, parallel versions of a few serial FIM algorithms, including FP-Growth, have al- ready been developed. In this study, we develop three different parallel FP-Growth implementations for cluster computers. They, all MPI based, are (i) Static Parallel FP-Growth, (ii) Dynamic Parallel FP-Growth, and (iii) (Tree-Sharing) Dynamic Parallel FP-Growth. All the three variants are task-parallel, i.e., not based on horizontal or vertical partitioning of database. The algorithms are experimentally evaluated on a 16-node cluster computer. Our results demonstrate the utility of the algorithms.