Application of algebraic graph descriptors for clustering of real-world structures

摘要

We propose several vector graph descriptors created on the basis of vertex rank measures such as PageRank, Hubs and Authorities or Betweenness Centrality. The descriptors are used for clustering artificial and real-world data. We present the comparison of descriptors with the use of criteria such as computational complexity, size and quality of clustering. The experiments were performed mainly on the sets of aerial photos transformed to graphs with the use of Harris corner detection and Delaunay triangulation. The results show that the introduced pattern vectors can be a lower dimensional, less computationally expensive and graph size independent alternative for spectral descriptors, such as defined by Wilson, Hancock and Luo in [1].