A New Spectral Clustering Algorithm
This work addresses clustering challenges in network analysis, particularly for overlapping weighted networks, but is incremental as it builds on existing spectral methods with specific enhancements.
The authors tackled the problem of clustering in networks by proposing a new spectral algorithm that searches for gaps in eigenvectors of the Laplacian, and found it outperforms other spectral methods in certain parameter regimes on the LFR benchmark, with concrete improvements in performance metrics.
We present a new clustering algorithm that is based on searching for natural gaps in the components of the lowest energy eigenvectors of the Laplacian of a graph. In comparing the performance of the proposed method with a set of other popular methods (KMEANS, spectral-KMEANS, and an agglomerative method) in the context of the Lancichinetti-Fortunato-Radicchi (LFR) Benchmark for undirected weighted overlapping networks, we find that the new method outperforms the other spectral methods considered in certain parameter regimes. Finally, in an application to climate data involving one of the most important modes of interannual climate variability, the El Nino Southern Oscillation phenomenon, we demonstrate the ability of the new algorithm to readily identify different flavors of the phenomenon.