Online Modularity Clustering based on One-class SVMs

Abstract

Online clustering for unsupervised data requires fast and accurate analysis based on meaningful knowledge in endless stream of data. However as there are many problems to find an immediate solution of unlabeled data, most unsupervised clustering algorithms demand the prior knowledge or assumptions on the data. In this paper, we propose the online clustering method, an incremental learning to classify evolving data streams containing various types of observations. The proposed method updates the groups of streaming data set by exploiting the modularity clustering and one-class support vector machine, allowing it to identify data points that are considered distant from existing normal clusters. Our objective is to maximize the modularity quality for any given conditions. In the procedure of online learning, coherence criterion which determines the representatives of groups controls the computation complexity of algorithms. We evaluate the proposed method on both four two-dimensional synthetic data sets and three real data sets to report the accuracy and recall measure of our model to compare with other existing methods. The results show that the online adaptive clustering mechanism handles grouping problem of unbalanced data that is subject to change over time and produces acceptable outputs without knowing underlying distribution and input pattern of data.