We propose a simple and efficient way for pattern recognition and signal classification within the Diffusion Framework. Our proposed Node Connectivity Matching (NCM) method is derived from the diffusion distance. However, instead of computing the eigenvalues / eigenvectors of the normalized diffusion matrix on the graph constructed from the data, as required when approximating the diffusion distance, we treat each row of the normalized diffusion matrix as a training histogram of node connectivities. To classify an unlabeled data point, we compare its node connectivities to the training histograms using the L2 norm as a bin-by-bin histogram discriminant measure. Through numerical examples we show that our NCM method is more accurate than using the diffusion distance.
Keywords: Diffusion distance, normalized diffusion matrix, Markov transition probabilities, directed diffusion, histogram matching.