Commute-Time Convolution Kernels for Graph Clustering
Commute time has proved to be a powerful attribute for clustering and characterising graph structure, and which is easily computed from the Laplacian spectrum. Moreover, commute time is robust to deletions of random edges and noisy edge weights. In this paper, we explore the idea of using convolution kernel to compare the distributions of commute time over pairs of graphs. We commence by computing the commute time distance in graphs. We then use a Gaussian convolution kernel to compare distributions. We use kernel kmeans for clustering and use kernel PCA for illustration using the COIL object recognition database.