Kernel-Based Clustering of Big Data

There has been a rapid increase in the volume of digital data over the recent years. Analysis of this data, popularly known as big data, necessitates highly scalable data analysis techniques. Clustering is an exploratory data analysis tool used to discover the underlying groups and structures in the data. Stateof-the-art scalable clustering algorithms assume “linear separability” of the clusters, and do not perform well on real-world data sets. Kernel-based clustering algorithms, which use non-linear similarity measures, are more accurate, but are not scalable to data sets containing billions of highdimensional points from thousands of clusters. We propose scalable approximate kernel-based clustering algorithms, based on random sampling and efficient optimization. The proposed algorithms are as efficient as linear clustering algorithms, and achieve cluster quality comparable to that of classical kernel-based clustering algorithms. We demonstrate the efficiency and effectiveness of the proposed algorithms on several diverse large-scale data sets.