Elias Ameli Bafandeh - Department of Engineering Islamic Azad University of Mashhad Mashhad, Iran
Hossein Deldari - Associated professor Islamic Azad University of Mashhad Mashhad, Iran

As web clicks, commercial, social networks, andscientific data sources growing with an extraordinary rate, it isvery necessary to analyze this data with powerful clusteringalgorithms. Current data mining Algorithms cannot deal withlarge datasets due to their large size and complexity. Forextracting useful information from these large datasets, newmining techniques are necessary. K-means is the most widely usedpartitional clustering algorithm and it is extremely sensitive to theinitial centroid selection. There is a growing development ofanalysis on large datasets using MapReduce jobs. MapReduceparallel processing is a framework for working with cloudcomputing, K-means is not suitable to be used in MapReducebecause of having repetitive calculation in working with largedata. For this reason, in recent years several research to optimizethe algorithm and reduce dependence on iterative computing isdone. In this article have been trying to work in the field toinvestigate the problem. The most important improvements havetaken place on the algorithm in order to reduce the number ofiterations and improvements done in the central parts of the initialselection.