Information Technology Reference
In-Depth Information
(1) Classifier C
0
is generated by using C4.5 algorithm and testifying 10 times in
learning set LS, which has the predict accuracy
P
0
after tested in TS;
Table 7.4 Predict accuracy comparison
Dataset
P0
P1
P2
Anneal
breast-cancer
credit
genetics
glass
heart
hypo
letter
sonar
soybean
voting
diabetes
96.9±10.4
95.7±2.1
84.8±2.5
98.7±4.4
68.9±9.2
77.8±4.3
93.2±4.2
88.4±9.8
65.4±7.1
78.9±5.9
95.8±1.3
74.3±3.0
97.1±9.7
96.2±1.7
87.4±1.8
98.7±4.4
75.2±6.5
79.6±3.9
92.7±3.9
93.7±8.7
74.7±12.1
84.8±6.5
95.9±1.6
76.1±3.1
97.8±10.8
97.2±2.5
89.5±2.1
98.7±4.4
75.2±6.5
79.8±3.7
92.7±3.9
93.7±8.7
74.7±12.1
84.8±6.5
95.9±1.6
77.2±2.3
(2) Classifier C
1
is generated by using algorithm BSDT and testifying 10 times in
learning set LS, which has the predict accuracy
1
after tested in TS;
(3) Classifier C
2
is generated by using algorithm BSDT with attribute selecting
function ASF and testifying 10 times in learning set LS, which has the predict
accuracy
P
2
after tested in TS;
(4) For each algorithm we compute average value of 15 experiments, so that
getting stable predict model;
(5) Compare results of
P
P
0
,
P
1
and
P
2
and predict accuracy are listed in Table 7.4.
It can be seen from Table 7.4 that after classified by decision tree
classification predict model generated by algorithm BSDT, precisions of all
dataset are improved. Predict precision of ASF function defined by cost
coefficient and bias coefficient is obviously improved. Experiment results show
that, ASF function and algorithm BSDT is effective.
We implement a method integrated various decision tree learning algorithms
in algorithm BSDT. It is an effective approach to integrate various machine
learning algorithms, and it assures optimum selection of representation bias and
procedure bias, so it supports multiple strategy learning algorithm.
