Database Reference
In-Depth Information
TABLE 6.5:
Comparative confusion matrices for 3 clusters of
Classic300.
fskmeans
spkmeans
hard-moVMF
soft-moVMF
med
cisi
cran
med
cisi
cran
med
cisi
cran
med
cisi
cran
29
38
22
29
38
22
3
72
1
0
98
0
31
27
38
31
27
38
62
28
17
99
2
0
40
35
40
40
35
40
35
0
82
1
0
100
TABLE 6.6:
Comparative confusion matrices for 3 clusters of
Classic400.
fskmeans
spkmeans
hard-moVMF
soft-moVMF
med
cisi
cran
med
cisi
cran
med
cisi
cran
med
cisi
cran
27
16
55
27
17
54
56
28
20
0
0
91
51
83
12
51
82
12
44
72
14
82
99
2
23
1
132
23
1
133
1
0
165
19
1
106
relatively pure segments, and merges the smaller two into one cluster. When
4 clusters are requested from soft-moVMF , it returns 4 very pure clusters (not
shown in the confusion matrices), two of which are almost equal sized segments
of the bigger cluster.
An insight into the working of the algorithms is provided by considering
their clustering performance when they are requested to produce greater than
the “natural” number of clusters. In Table 6.7 we show the confusion matrices
resulting from 5 clusters of the Classic3 corpus. The matrices suggest that
the moVMF algorithms have a tendency of trying to maintain larger clusters
intact as long as possible, and breaking them into reasonably pure and com-
parably sized parts when they absolutely must. This behavior of our moVMF
algorithms coupled with the observations in Table 6.6 suggest a clustering
method in which one could generate a slightly higher number of clusters than
required, and then agglomerate them appropriately.
TABLE 6.7:
Comparative confusion matrices for 5 clusters of Classic3.
fskmeans
spkmeans
hard-moVMF
soft-moVMF
med
cisi
cran
med
cisi
cran
med
cisi
cran
med
cisi
cran
312
323
292
1107
2
4
2
4
3
5
0
1
520
512
511
1455
8
10
8
9
1
0
5
14
936
944
514
526
5
6
5
6
1
0
2
1
1018
0
1
1018
0
1
0
2
1093 501
0
0
0
0
1069
0
0
1059
5
1451
13
1
2
276
The MI plots for the various Classic3 datasets are given in Figures 6.5(a) -(c).
For the full Classic3 dataset (Figure 6.5(a)), all the algorithms perform almost
 
Search WWH ::




Custom Search