Page 253 - Introduction to Statistical Pattern Recognition

5 Parameter Estimation 235

TABLE 5-10

BIAS BETWEEN L AND R ERRORS FOR DATA 1-1 (%)

n

4 8 16 32 64

3 9.00 13.79 23.03 41.34 77.87
13.33 15.42 19.69 22.86 30.29
7.03 5.22 4.12 4.26 3.40

5 5.40 8.27 13.82 24.80 46.72
7.50 9.25 10.75 17.75 24.47
4.56 3.24 2.28 2.69 1.53

k 10 2.70 4.14 6.9 1 12.40 23.36
2.2s 4.63 6.34 9.58 16.01
1.84 2.02 1.59 1.61 1.24

20 1.35 2.07 3.45 6.20 1 1.68
1.38 2.09 3.14 5.05 9.56
1.05 1 .oo 0.64 0.53 0.45

40 0.67 1.03 1.73 3.10 5.84
0.44 1.08 1.55 2.96 5.21
0.30 0.39 0.30 0.30 0.36

Effect of Outliers

It is widely believed in the pattern recognition field that classifier perfor-
mance can be improved by removing outliers, points far from a class’s inferred
mean which seem to distort the distribution. The approach used in this section,
namely to analyze the difference between the R and L parameters, can be