20.04.2014 Views

Combining Pattern Classifiers

Combining Pattern Classifiers

Combining Pattern Classifiers

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

20 FUNDAMENTALS OF PATTERN RECOGNITION<br />

and<br />

P (2) ¼ P (2)<br />

A P (2)<br />

B<br />

The estimated mean and variance of the differences, for this two-fold cross-validation<br />

run, are calculated as<br />

P ¼ P(1) þ P (2)<br />

; s 2 ¼ P (1) P 2<br />

þ P<br />

(2) P 2<br />

2<br />

(1:26)<br />

Let P (1)<br />

i denote the difference P (1) in the ith run, and s 2 i denote the estimated variance<br />

for run i, i ¼ 1, ..., 5. The proposed ~t statistic is<br />

P (1)<br />

1<br />

~t ¼ q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi<br />

(1=5) P (1:27)<br />

5<br />

i¼1 s2 i<br />

Note that only one of the ten differences that we will calculate throughout this experiment<br />

is used in the numerator of the formula. It is shown in Ref. [14] that under the<br />

null hypothesis, ~t has approximately a t distribution with five degrees of freedom.<br />

Example: Comparison of Two Classifier Models Through Cross-Validation<br />

Tests. The banana data set used in the previous examples is suitable for experimenting<br />

here because we can generate as many as necessary independent data sets from<br />

the same distribution. We chose the 9-nn and Parzen classifiers. The Matlab code for<br />

the three cross-validation methods discussed above is given in Appendices 1A to 1C<br />

at the end of this chapter. PRTOOLS toolbox for Matlab, version 2 [19], was used to<br />

train and test the two classifiers.<br />

K-Hold-Out Paired t-Test. The training and testing data sets used in the previous<br />

example were pooled and the K-hold-out paired t-test was run with K ¼ 30, as<br />

explained above. We chose to divide the data set into halves instead of a 2=3 to<br />

1=3 split. The test statistic (1.25) was found to be t ¼ 1:9796. At level of significance<br />

0.05, and degrees of freedom K 1 ¼ 29, the tabulated value is 2.045<br />

(two-tailed test). Since the calculated value is smaller than the tabulated value,<br />

we cannot reject the null hypothesis. This test suggests that 9-nn and Parzen classifiers<br />

do not differ in accuracy on the banana data. The averaged accuracies over the<br />

30 runs were 92.5 percent for 9-nn, and 91.83 percent for Parzen.<br />

K-Fold Cross-Validation Paired t-Test. We ran a 10-fold cross-validation for the<br />

set of 200 data points, so each testing set consisted of 20 objects. The ten testing<br />

accuracies for 9-nn and Parzen are shown in Table 1.5.<br />

From Eq. (1.25) we found t ¼ 1:0000. At level of significance 0.05, and degrees<br />

of freedom K 1 ¼ 9, the tabulated value is 2.262 (two-tailed test). Again, since the

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!