Towards a Moderate-Granularity Incremental Clustering Algorithm for GPU

Abstract

The incremental clustering algorithm plays a vital role in big data processing. The massive data problems generally raise high computation demand on the hardware platform. GPU-based parallel computing is a promising method to satisfy this demand. However, the existing incremental clustering algorithms face an accuracy-parallelism dilemma when accelerated by GPU. The block-wise algorithms evolve the clusters in coarse granularity and sacrifice accuracy for parallelism, while the point-wise algorithms proceed in fine granularity and barter parallelism for accuracy. We propose a moderate-granularity algorithm. This algorithm constantly generates micro-clusters from the incoming data blocks, and then evolves the clusters in the granularity of a micro-cluster. The proposed algorithm takes the following advantages: first, it avoids predefining a cluster number searching range like block-wise algorithms, second, it alleviates the accuracy problem caused by coarse granularity, third, it adopts the parallel-friendly algorithm to generate micro-clusters and decreases the amount of serial operations, so that it is parallelism-scalable compared to point-wise algorithms. Experiments on a CPU-GPU hybrid platform show that our algorithm can achieve comparable accuracy to its batch counterpart and is scalable in terms of parallelism.