To boost graph clustering based on power iteration by removing outliers

Abstract

Power Iteration Clustering (PIC) is an applicable and scalable graph clustering algorithm using Power Iteration (PI) for embedding the graph into the low-dimensional eigenspace what makes graph clustering tractable. This paper investigates the negative impacts of outliers on power iteration clustering and based on this understanding we present a novel approach to remove outlier nodes in a given graph what in turn brings significant benefits to graph clustering paradigms. In the original PIC algorithm, outliers have a high potential to be mis-clustered, and hence, they impress the output of PI embedding. As a result, the embedded space offered by PIC couldn't be deemed as a reasonable illustration of the graph in question. The statistical outlier detection method applied in this paper detects and removes outlier nodes in an iterative manner to enhance the embedded space for clustering. Experiments on several datasets across different domain show the advantages of the proposed method compared to the rival approaches. © Springer Science+Business Media Singapore 2014.