Heat kernel analysis on graphs

In this thesis we aim to develop a framework for graph characterization by combining the methods from spectral graph theory and manifold learning theory. The algorithms are applied to graph clustering, graph matching and object recognition. Spectral graph theory has been widely applied in areas such as image recognition, image segmentation, motion tracking, image matching and etc. The heat kernel is an important component of spectral graph theory since it can be viewed as describing the flow of information across the edges of the graph with time. Our first contribution is to investigate how to extract useful and stable invariants from the graph heat kernel as a means of clustering graphs. The best set of invariants are the heat kernel trace, the zeta function and its derivative at the origin. We also study heat content invariants. The polynomial co-efficients can be computed from the Laplacian eigensystem. Graph clustering is performed by applying principal components analysis to vectors constructed from the invariants or simply based on the unitary features extracted from the graph heat kernel. We experiment with the algorithms on the COIL and Oxford-Caltech databases. We further investigate the heat kernel as a means of graph embedding. The second contribution of the thesis is the introduction of two graph embedding methods. The first of these uses the Euclidean distance between graph nodes. To do this we equate the spectral and parametric forms of the heat kernel to com-