For a proper neighborhood size k, we set its value as half of the average cluster
size k = n/(2c), where n is the number of instances in a dataset and c is the number
of clusters. We assume that this is a safe choice for each instance to assemble its
true local density. In Section 3.6.5 we will further test the algorithm sensitivity to
k in the range of [10%, 100%] of n
c with 10% as step size to verify the rationality of
k = 50%.
In Section 3.4, we already discussed the effect of α on LDAT. In our general
experiments, we use α = 1 by default but we will also test RWC+LDAT and
AHK+LDAT with different value of α ∈ [0, 2] in Section 3.6.6.
For DBSCAN experiments, we set Eps, the neighborhood radius, in the same
way as we set σq. We assign minP ts, minimal number of instances considered as
a cluster, in the range of [10, min(n/c, 300)], and only record the best result among
them.