Information Technology Reference
In-Depth Information
algorithms. Although boosted J4.8 outperformed CAMLET on the entire train-
ing dataset, it had lower accuracy on LOO. This means boosted J4.8 had the
problem of overfitting. Thus, CAMLET showed higher adaptability than the
other selective meta-learning algorithms.
Learning curves of the learning algorithms.
Since the rule evaluation
model construction method required that the mined rules be evaluated by a
human expert, we investigated the learning curves of each learning algorithm
to estimate a minimum training subset to obtain a valid rule evaluation model.
The upper table in Fig. 6.4 shows the accuracies of the entire training dataset
for each subset of training dataset. The percentage for the achievements of each
learning algorithm compared with its accuracy over the entire dataset are shown
in the lower section of Fig. 6.4.
As observed in these results, SVM and CLR, which learn hype-planes, ob-
tained achievement ratios greater than 95% using less than 10% of the training
subset. Although decision tree learner, boosted J4.8, and BPNN could learn to
%training
sample
10
20
30
40
50
60
70
80
90
100
CAMLET
76.7
78.4
80.8
81.6
81.7
82.6
82.8
84.8
84.6
89.3
Stacking
69.6
77.8
75.3
77.9
72.2
82.2
75.4
83.4
86.5
81.1
Boosted J4.8
74.8
77.8
79.6
82.8
83.6
85.5
86.8
88.0
89.7
99.2
Bagged J4.8
77.5
79.5
80.5
81.4
81.8
82.1
83.2
83.2
84.1
87.3
J4.8
73.4
74.7
79.8
78.6
72.8
83.2
83.7
84.5
85.7
85.7
BPNN
74.8
78.1
80.6
81.1
82.7
83.7
85.3
86.1
87.2
86.9
SMO
78.1
78.6
79.8
79.8
79.8
80.0
79.9
80.2
80.4
81.6
CLR
76.6
78.5
80.3
80.2
80.3
80.7
80.9
81.4
81.0
82.8
OneR
75.2
73.4
77.5
78.0
77.7
77.5
79.0
77.8
78.9
82.4
100
95
90
85
80
Bagged J4.8
Boosted J4.8
Stacking
CAMLET
75
70
0
20
40
60
80
100
% training sample
Fig. 6.4.
Learning curves of Accuracies (%) on the learning algorithms over subsam-
pled training dataset
