A clustering framework for unbalanced partitioning and outlier filtering on high dimensional datasets

Abstract

In this study, we propose a better relationship based clustering framework for dealing with unbalanced clustering and outlier filtering on high dimensional datasets. Original relationship based clustering framework is based on a weighted graph partitioning system named METIS. However, it has two major drawbacks: no outlier filtering and forcing clusters to be balanced. Our proposed framework uses Graclus, an unbalanced kernel k-means based partitioning system. We have two major improvements over the original framework: First, we introduce a new space. It consists of tiny unbalanced partitions created using Graclus, hence we call it micro-partition space. We use a filtering approach to drop out singletons or micro-partitions that have fewer members than a threshold value. Second, we agglomerate the filtered micro-partition space and apply Graclus again for clustering. The visualization of the results has been carried out by CLUSION. Our experiments have shown that our proposed framework produces promising results on high dimensional datasets.