require('header.inc.php'); ?>
This section reports the experiments conducted to evaluate the effects of QuickRBF package.
In particular, we are interested in
how the mechanism proposed in this paper performs in comparison with the SVM
and other famous classifiers in data classification benchmark data sets.
The experiments in this section are conducted to evaluate the performance of
the QuickRBF classifier against other famous classifiers,
the RVKDE classifier [Oyang et al., 2005], APC-III [Hwang and Bang, 1997], LIBSVM [Chang and Lin, 2001]
and KNN. The discussions of the experiments will focus on the following two
issues: classification accuracy and execution efficiency. Also, regarding to
the parameter settings of other classifiers for comparison, we adopted the
parameter settings suggested by the authors in their original papers.
Table 1 lists the main characteristics of the
three data sets used in the experiments. All three data sets,
satimage, letter, and
shuttle, are from the Statlog collection [Michie et al., 1994]. The hardware platform used in the
experiments is a workstation with dual Pentium-III-1GHz CPUs, 2GB RAM, and the
FreeBSD UNIX-release 4.10.
Table 2 lists the number of support vectors of the
three data sets, and we use the support vectors which are selected by SVM as
our centers to compare with SVM.
Table 3 compares the accuracy delivered by
alternative classification algorithms with the three benchmark data sets. As
Table 3 shows, the proposed method basically
delivers the same level of accuracy with other famous classifiers, SVM and
RVKDE, while the KNN and APC-III based classifier do not produce comparable
generation results.
Table 4 compares the execution time of the RVKDE
classifier, the SVM, the APC-III based classifier and the proposed method with
the Statlog data sets. In Table 4, the
total time taken to construct classifiers based on the given training data
sets are listed in the rows marked by "Make classifier". The time listed in
"Make classifier" row are the time of cross validation for RVKDE classifier,
the time of model selection for SVM, and the time of clustering process,
calculating bandwidths and weights of APC-III based classifier. Also, for
proposed classifier, the reported time is the time of calculating weights. In
addition, the time taken by alternative classifiers to predict the classes of
the testing instances are listed in the rows marked by "Prediction".
The mechanism used in the QuickRBF package is more efficient than the SVM classifier
for constructing a data classifier. Also, the QuickRBF classifier is basically
at the same level or even more efficient than other classifiers. In addition,
the QuickRBF classifier delivers comparable execution efficiency
as the LIBSVM in the prediction phase and enjoys
a 10
times speedup over the RVKDE classifier in this regard.
Date set
number of
training datanumber of
testing datasatimage
4435
2000
letter
15000
5000
shuttle
43500
14500
Data set
number of
training datanumber of
support vectorssatimage
4435
1610
letter
15000
8945
shuttle
43500
286
Data set
RVKDE
LIBSVM
1NN
3NN
APC-III
QuickRBF
satimage
92.30
91.30
88.80
90.65
90.25
92.35
letter
97.12
97.98
95.68
95.16
91.16
97.68
shuttle
99.94
99.92
99.94
99.91
97.34
99.43
Average
96.45
96.40
94.84
95.24
92.92
96.49
Data set
RVKDE
LIBSVM
APC-III
QuickRBF
satimage
676
64644
136
314
Make Classifier
letter
2842
387096
712
59978
shuttle
98540
467955
2595
114
satimage
21.30
11.53
0.63
7.8
Prediction Time
letter
128.60
94.91
2.15
104.6
shuttle
996.10
2.13
0.48
2.95