Robust Spectral Clustering by Maximizing Similarity Modularity
Download as PDF
Modularity is a metric of how well a given network is divided into communities. It has proven surprisingly effective in handling the ready-made real-world networks, including social networks, information networks, biological networks, and so forth. This paper, based on a notional instantiation of modularity to similarity modularity (s-mod), answers the question of how well the s-mod can measure the data-points clustering. This paper maximizes the s-mod by a new spectral algorithm. Compared to the existing typical spectral clustering techniques, the presented algorithm has two advantages: the higher robustness to the variation of input parameter and the capability of estimating the intrinsic clusters' number. This paper also analyzes theoretically why it performs robust to parameter variation and how to accurately estimate the intrinsic clusters' number. Experiments on typical UCI datasets show the good performance of the presented algorithm.
Graph, Spectral clustering, Similarity modularity