Evolution of Graphs for Unsupervised Learning

Abstract

In this paper, we propose a novel method which adopts evolutionary techniques so as to optimize a graph structure. The method that was developed has been applied in clustering problems, where spectral graph clustering technique has been used. In order to use evolutionary algorithms initial population has been created consisting of nearest neighbor graphs and variations of these graphs, which have been properly altered in order to form chromosomes. Since it was observed that initial population is crucial for the performance of the algorithm, several techniques have been considered for the creation. A fitness function was used in order to decide about the appropriateness of the chromosomes. The major advantage of our approach is that the algorithm is generic and can be used to all problems that are, or can be, modeled as graphs, such as dimensionality reduction and classification. Experiments have been conducted on a traditional dance dataset and other multidimensional datasets, providing encouraging results.