Spectral Clustering Using Evolving Similarity Graphs

Abstract

In this paper, we propose a novel spectral graph clustering method that uses evolutionary algorithms in order to optimise the structure of a graph, by using a fitness function, applied in clustering problems. Nearest neighbour graphs and variants of these graphs are used in order to form the initial population. These graphs are transformed in such a way so as to play the role of chromosomes in the evolutionary algorithm. Multiple techniques have been examined for the creation of the initial population, since it was observed that it plays an important role in the algorithm's performance. The advantage of our approach is that, although we emphasise in clustering applications, the algorithm may be applied to several other problems that can be modeled as graphs, including dimensionality reduction and classification. Experiments on traditional dance dataset and on other various multidimensional datasets were conducted using both internal and external clustering criteria as evaluation methods, which provided encouraging results.