List of Figures
1.1 Our proposed physics-based unsupervised data modeling framework
and the derived techniques. . . . . . . . . . . . . . . . . . . . 6
2.1 The sensitivity example of NJW [109], one of the traditional spectral
clustering algorithms, with respect to different Gaussian scaling
parameter σ and noise appearance. The two output clusters are colored
with red or blue. A small variation to σ or data points (noise)
leads to radically-different results. Such an instability becomes an
issue to traditional spectral clustering algorithms. . . . . . . . . . . 14
2.2 Clustering results of different algorithms on a synthetic dataset with
heterogeneous density distributions. Figure 2.2(a) shows the original
dataset, where the green and blue clusters with Gaussian distributions
have higher density than the red cluster with a uniform
distribution. The clustering results of NJW (Figure 2.2(b)), RWC
(Figure 2.2(c)) and NN (Figure 2.2(d)) are shown respectively,
which are not capable of capturing the density variation. For the
localized method, ST (Figure 2.2(e)) has better result since it has
a locally adaptive scaling parameter (in Equation 2.9), while SCDA
(Figure 2.2(f)) reveals a similar density-awareness as NJW. In
short, none of the above methods provides a desirable separation
that is aware of both density change and manifold structures across
clusters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15