Databases Reference
In-Depth Information
Tabl e 2 .
Comparison of tree complexity (number of leafs) for decision tree (C4.5)
and proposed decision tree building methods (CMMC-1, CMMC-2 and CMMC-3)
Dataset
C4.5
CMMC1
CMMC2
CMMC3
amlall1
2.93
73
.
30
5
.
24
6
.
19
amlall2
2.93
75
.
57
5
.
38
5
.
45
amlall3
2.93
75
.
77
5
.
05
6
.
29
amlallAvg
2.93
74
.
88
5
.
22
5
.
98
breast1
6.26
46
.
35
18
.
78
11
.
22
breast2
7.12
11
.
37
15
.
65
14
.
53
breast3
6.76
11
.
92
14
.
68
23
.
05
breastAvg
6.71
23
.
21
16
.
37
16
.
27
lung1
3.99
75
.
32
5
.
05
4
.
35
lung2
4.00
64
.
05
9
.
21
4
.
31
lung3
4.00
72
.
71
5
.
52
4
.
21
lungAvg
4.00
70
.
69
6
.
59
4
.
29
mll1
3.00
115
.
94
6
.
39
4
.
88
mll2
3.00
118
.
31
8
.
68
5
.
24
mll3
3.92
121
.
76
7
.
60
5
.
00
mllAvg
3.31
118
.
67
7
.
56
5
.
04
Average
4.24
71
.
86
8
.
94
7
.
90
Tabl e 3 .
Comparison of accuracy by feature selection method
Feature selection
C4.5
CMMC1
CMMC2
CMMC3
RF
GainRatio
80.63
87.39
85.86
86.04
95.21
ReliefF
81.85
85.23
85.29
85.78
93.23
SVM-FS
83.65
85.54
85.55
84.56
96.12
generate only two times more rules than simple decision trees. Even better
results were obtained using CMMC-3 method. Low complexity at CMMC-3
based trees is a consequence of the decision tree building technique in which
trees are generated at the beginning using the initial training sets without
artificial data points. The artificial data points are added in later stages.
Table 3 presents the results based on average results on each dataset for
each gene selection method. The best results, with exception of CMMC-3
method, were achieved when SVM based feature selection method was used.
From this table it can also be seen that the majority of accuracy gain of the
CMMC-1 method compared to CMMC-2 was due to first method's better
accuracy when used with the GainRatio based feature selection method.
3.5 Subgrouping Leukemia Type
To demonstrate the practical advantage of our proposed adaptation to CMM
method, two sample trees were constructed from the Leukemia (AML-ALL)
