Effect of adaptive interval configuration on parallel mining association rules

Abstract

All current parallel algorithms for mining the association rules follow the conventional level-wise approach. It imposes a synchronization in every iterative computation which degrades greatly their mining performance on a shared-memory multi-processor parallel machine. An asynchronous algorithm APM (asynchronous parallel mixing) has been proposed for mining the association rules on a shared-memory multi-processor machine. All participating processors in APM generate the candidate sets and count their supports independently without synchronization. Furthermore, it can finish the computation with fewer passes of database scanning than required in the level-wise approach. An optimization technique has been developed to enhance APM so that its performance would be insensitive to the data distribution. Two variants of APM and the synchronous algorithm Count Distribution, which is a parallel version of the popular serial mining algorithm Apriori, have been implemented on an SGI Power Challenge SMP parallel machine. The results show that the asynchronous algorithm APM is better than the synchronous algorithm in performance and extensibility.