Information Technology Reference
In-Depth Information
RIL
RIL
1
1
0.8
0.8
0.6
0.6
AVE
TMSE
EBC
0.4
0.4
0.2
0.2
AVE
TMSE
EBC
0
0
0.5
0.2
0.1
0.001
1.0
0.999
0.99
0.9
(a)
K
=2
RIL
RIL
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
AVE
TMSE
EBC
AVE
TMSE
EBC
0
0
0.5
0.2
0.1
0.001
1.0
0.999
0.99
0.9
(b)
K
=5
RIL
RIL
1
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
AVE
TMSE
EBC
AVE
TMSE
EBC
0
0
0.5
0.2
0.1
0.001
1.0
0.999
0.99
0.9
(c)
K
=10
Fig. 15.3.
Experimental results on artificial data of incomplete orders
NOTE: See note in Figure 15.2.
We begin with the variation of estimation performance according to the decrease of
intra-cluster cohesion. If the cohesion is 1, sample orders are exactly concordant with
their corresponding true central orders. In this trivial case, the
AVE
method succeeds
almost perfectly in recovering the embedded cluster structure. Because the dissimilari-
ties between sample orders are 0 if and only if they are in the same cluster, this method
could lead to perfect clusters. Though both the
EBC
and
TMSE
methods found almost
perfect clusters in the
K
=2 case, the performance gradually worsened when
K
in-
creased. In a
k
-o'means clustering, a central order is chosen from the finite set,
(
L
∗
).
This is contrasted to the fact that a domain of centers is an infinite set in the clustering
S
