Scalable Parallel Algorithms for Shared Nearest Neighbor Clustering

Clustering is a popular data mining technique which discovers structure in unlabeled data by grouping objects together on the basis of a similarity criterion. Traditional similarity measures lose their meaning as the number of dimensions increases and as a consequence, distance or density based clustering algorithms become less meaningful. Shared Nearest Neighbor (SNN) is a solution to clustering high-dimensional data with the ability to find clusters of varying density. SNN assigns objects to a cluster, which share a large number of their nearest neighbors. However, SNN is compute and memory intensive for data of large size and/or dimensionality. Nearest neighbor queries are responsible for a major proportion of computations in SNN, resulting in lower efficiency for higher value of number of nearest neighbors (k). The main motivation of this work is to improve the efficiency of SNN and to parallelize it so that it can be used for clustering large high-dimensional datasets and for large values of k. Existing SNN algorithms become inefficient in these situations. In this paper, we present a new sequential SNN algorithm, R-SNN, which uses R-tree for executing neighborhood queries efficiently and exploiting spatial locality to minimize memory usage. R-SNN is benchmarked against the best available implementation of SNN and is found up to 77 times faster when tested on various real datasets. R-SNN is parallelized for distributed memory, shared memory, and hybrid systems. Significant speedup and scalability achieved can be attributed to parallelization and good load balancing strategies and also to exploitation of spatial locality. Experimental results demonstrate the same for datasets of varying dimensionality and size. The maximum speedup achieved for shared, distributed, and hybrid models are 427.19 using 48 threads, 394.24 using 32 processes, and 1380.69 on 32 nodes (with each node spawning 4 threads), respectively. Super-linear speedup for some datasets is attributed to optimized neighborhood queries. All the proposed algorithms produce identical clustering results as that of the classical SNN.