Parallel clustering of high dimensional data by integrating multi-objective genetic algorithm with divide and conquer

Abstract

This paper applies divide and conquer approach in an iterative way to handle the clustering process. The target is a parallelized effective and efficient approach that produces the intended clustering result. We achieve scalability by first partitioning a large dataset into subsets of manageable sizes based on the specifications of the machine to be used in the clustering process; then cluster the partitions separately in parallel. The centroid of each obtained cluster is treated like the root of a tree with instances in its cluster as leaves. The partitioning and clustering process is iteratively applied on the centroids with the trees growing up until we get the final clustering; the outcome is a forest with one tree per cluster. Finally, a conquer process is performed to get the actual intended clustering, where each instance (leaf node) belongs to the final cluster represented by the root of its tree. We use multi-objective genetic algorithm combined with validity indices to decide on the number of classes. This approach fits well for interactive online clustering. It facilitates for incremental clustering because chunks of instances are clustered as stand alone sets, and then the results are merged with existing clusters. This is attractive and feasible because we consider the clustering of only centroids after the first clustering stage. The reported test results demonstrate the applicability and effectiveness of the proposed approach.