MULTIFRACTAL DISCRIMINATION MODEL (MDM) OF HIGH-FREQUENCY PUPIL DIAMETER MEASUREMENTS FOR HUMAN-COMPUTER INTERACTION - Quantitative Medical Data Analysis Using Mathematical Tools and Statistical Techniques - page 346

Control
#1
#2
#3

0.85

#3

0.8

Control

0.75

#1

0.7

#2

0.34 0.36 0.38 0.4 0.42 0.44

Spectral Mode

Fig. 7. Centroid points for bivariate measures: Spectral Mode and Broadness Measures.

To choose the nearest neighbor parameter k, the classier is built as
a learning process. The learning curve, which includes the test error and
training error corresponding to dierent parameters k, is given in Figure 4.
Although, relatively low training error could be achieved by choosing small
k, the test error is too big for a practically useful classier. To overcome
these drawbacks, we adapt the model by combining techniques. Model com-
bining is a technique of combining the predictions from dierent classiers.
The results have shown to be promising. For the details of this combining
technique, the reader is directed to Xu and colleagues 27 . The advantage
of using model combining is due to its ability of overcoming the instability
of the single classier. In fact, Shi and colleagues 28 provides a Bayesian
justication of the correctness of model combining. In our study, the single
k-nearest-neighbor classier is not very accurate and robust according to
Figure 8. By applying the model combining technique to these k-nearest-
neighbor classiers (3k10), the test errors get much smaller as we