Database Reference
In-Depth Information
Table 5.8. Results using the S 2 and s 2 sets obtained from the EAs.
GGA
SGA
Database
1-NN
C4.5
%
1-NN
C4.5
%
Pen-based
95.60
99.10
96.13
96.12
99.10
94.18
Satimage
85.41
96.90
97.46
86.65
97.10
95.82
Thyroid
90.52
99.90
97.58
88.46
99.90
96.15
Average
90.51
98.63
97.06
90.41
98.70
95.38
CHC
PBIL
Database
1-NN
C4.5
%
1-NN
C4.5
%
Pen-based
98.32
99.10
85.70
79.37
99.10
99.75
Satimage
86.48
97.30
91.96
84.88
97.10
96.87
Thyroid
91.36
100.00 88.58
87.64
99.90
96.64
Average
92.05
98.80
88.75
83.96
98.70
97.75
Average
Database
1-NN
C4.5
%
Pen-based
92.35
99.10
93.94
Satimage
85.86
97.10
95.53
Thyroid
89.50
99.93
94.74
Average
89.23
98.71
94.74
We want to offer the following conclusions:
z The C4.5 algorithm obtains a very good test accuracy (approximately 98%),
for all the IS algorithms. This indicates that the stratified mechanism used in
this chapter for the TSS (Fig. 5.3) is a suitable and robust method for finding
training sets for this learning algorithm.
z The EAs return ( S 1 s 1 ) and ( S 2 s 2 ) pairs with a better mix of 1-NN and C4.5
test accuracy and reduction percentage than the ones reached with the
classical algorithms.
z The best EA with regard to the 1-NN and C4.5 test accuracy is CHC.
5.8 Concluding Remarks
This chapter presented the analysis of the evolutionary IS algorithms and its use
for data reduction in KD. An experimental study has been carried out for
comparing the results of four EA models against classical IS algorithms on two
particular applications, the PS and the TSS. The principal conclusions reached are
the following:
 
Search WWH ::




Custom Search