Figure 2.4: Clustering results of Diffusion Maps (DM) and Multiscale Diffusion
Maps (MDM) on the synthetic dataset in Figure 2.2(a). The global Gaussian kernel
is used here with σ(G) = 2. Figure 2.4(a) to 2.4(d) show the results of DM from t =
1 to t = 100. Although DM with t = 50 obtains better separation in the boundary
area among the three clusters, it is hard to guess the best range of t unsupervisedly.
MDM, in spite of the elimination of parameter t, easily gets over-diffusion without
perception of density change (see Figure 2.4(e)).