An Empirical Analysis of Similarity Matrix for Spectral Clustering

Constructing the similarity matrix is the key step for spectral clustering, and its' goal is to model the local neighborhood relationships between the data points. In order to evaluate the influence of similarity matrix on performance of the different spectral clustering algorithms and find the rules on how to construct an appropriate similarity matrix, a system empirical study was carried out. In the study, six recently proposed spectral clustering algorithms were selected as evaluation object, and normalized mutual information, F-measures and Rand Index were used as evaluation metrics. Then experiments were carried out on eight synthetic datasets and eleven real word datasets respectively. The experimental results show that with multiple metrics the results are more comprehensive and confident, and the comprehensive performance of locality spectral clustering algorithm is better than other five algorithms on synthetic datasets and real word datasets.