DLPR: a distributed locality preserving dimension reduction algorithm

Document indexing using dimension reduction has been widely studied in recent years. Application of these methods in large distributed systems may be inefficient due to the required computational, storage, and communication costs. In this paper, we propose DLPR, a distributed locality preserving dimension reduction algorithm, to project a large distributed data set into a lower dimensional space. Partitioning methods are applied to divide the data set into several clusters. The system nodes communicate through virtual groups to project the clusters to the target space, independently or in conjunction with each other. The actual computation of reduction transforms is performed using Locality Preserving Indexing, which is a less studied method in distributed environments. Experimental results demonstrate the efficiency of DLPR in terms of preserving the local structure of the data set, and reducing the computing and storage costs.