Compared with the other algorithms, DBSCAN fails miserably since it is
mainly defined in Euclidean space and suffers from the “curse of dimensionality”
and lack of manifold awareness. GDL, ST and SCDA, although based on the theory
that supports local density adaptation, are unable to maintain desirable performance
across all the datasets, which is mainly caused by their suboptimal local density approximations.
Originated from diffusion equations, kDM shows its stability on all
the three types of datasets/kernel functions. NJW has comparable performance on
Table 3.2 but not Table 3.3, partially due to that it does not have any correction for
local density bias.