Information Technology Reference
In-Depth Information
Table 6.7.
Number of minimum training subsamples for outperforming the Accuracy
(%) of default class
Distribution I
J4.8
BPNN SVM CLR
OneR Bagged J4.8 Boosted J4.8 Stacking CAMLET
anneal
20
14
17
29
29
16
14
36
20
audiology
21
18
65
64
41
21
14
56
27
autos
38
28
76
77
70
28
28
77
31
balance-
scale
12
14
15
15
32
14
9
51
128
breast-
cancer
16
17
22
41
22
14
14
41
36
breast-w
7
10
10
18
14
10
6
19
11
8
8
9
22
14
8
8
24
8
colic
9
12
16
30
28
9
8
51
19
credit-a
Distribution II
J4.8
BPNN SVM CLR
OneR Bagged J4.8 Boosted J4.8 Stacking CAMLET
anneal
29
20
16
42
46
26
21
46
29
audiology
36
45
-
61
67
27
30
67
autos
49
39
49
123
88
44
34
74
44
balance-
scale
81
84
69
221
168
60
64
135
-
breast-
cancer
31
28
102
40
46
28
28
62
28
14
11
23
30
26
11
10
31
19
breast-w
colic
24
20
36
42
36
15
18
37
22
51
74
-
134
109
49
42
105
78
credit-a
Distribution III
J4.8
BPNN SVM CLR
OneR Bagged J4.8 Boosted J4.8 Stacking CAMLET
anneal
54
58
64
76
-
42
38
64
46
audiology
64
73
45
76
107
50
50
103
84
autos
66
102
84
121
98
45
39
76
76
balance-
scale
118
103
133
162
156
86
92
132
-
breast-
cancer
50
31
80
92
80
38
36
60
41
44
36
31
48
71
34
34
52
53
breast-w
colic
28
24
46
30
42
28
22
48
54
credit-a
118
159
-
-
173
76
76
120
109
class distribution (Distribution I), these learning algorithms were able to learn
valid models with less than 20% of the given training datasets. However, for
the datasets with imbalanced distributions (Distribution II & III), they needed
more training subsets to construct valid models, because their performances with
the entire training datasets fell to the percentages of the majority class of each
dataset, as shown in Table 6.6.
Comparison of results of the meta-learning algorithms.
Comparing
Stacking and CAMLET, CAMLET achieved a higher accuracy than Stacking,
as shown in Table 6.6. This shows that the approach of CAMLET, the de-
composition and re-construction of learning algorithms, is better than just com-
bining prepared learning algorithms. Although CAMLET can construct boosted
and bagged C4.5, which outperformed than the learning algorithms constructed
by CAMLET, CAMLET could not search for these algorithms as the appropriate
learning algorithms for these datasets. We need to improve the search method
