3.7 Figure 3.7(a) and 3.7(b) show the effect of shared nearest neighbor
(SNN) and our proposed LDAT, both built upon W(GLS)
and a positive
random walk normalization (RWC). It demonstrates LDAT’s
advantage of better recognizing density differences among clusters
than SNN and other algorithms shown in Figure 2.2 and 2.4. The
LDAT built upon AHK, shown in 3.7(c), has the best NMI result
through being aware of both density and manifold structures. . . . . 52