Information Technology Reference
In-Depth Information
In this chapter, 4 different types of asymmetry were tested. They have been chosen
in order to represent different degrees of asymmetry to provide one degree more of
freedom in the learning process. Fig. 2.6 (1) represents a high degree of asymmetry,
and Fig. 2.6 (2) is symmetric.
(1) (2)
Fig. 2.6.
Example of (1) non-symmetric and (2) symmetric 2-class output
The symmetric one is much better, because the fuzzy rules will be much easier
intelligible.
Regarding the stopping criteria, it is difficult to apply an early-stopping method
based on the test of a small quantity of data, since usually the minor-class is provided
with a reduce number of patterns. In this case, the genetic algorithm will stop when a
determinate threshold value in its fitness function is reached.
2.4 Experimental Results
The goal of the tests made, mainly, is to find a Fuzzy System that tries to improve the
results of the method used nowadays for the detection of the Down syndrome for the
second-trimester of pregnancy (called age/LR 18). In this section, we will apply
the FLAGID method to two types of problems. In the first subsection, the results of
the FLAGID method on the Down's syndrome dataset will be compared with those
obtained with the current method (age/LR). In the second subsection, the FLAGID
method will be compared with two of the best methods that work with imbalanced
datasets, SDC and KBA, in order to demonstrate that the new proposed method works
well with imbalanced datasets.
2.4.1 Experimental Results for Down's Syndrome Detection
The dataset has been obtained from the University Hospital Dr. Josep Trueta of Gi-
rona (Spain). The data contain many variables, and only a few are used: the race of
the mother, the number of fetuses, the age of the mother, its weight, the gestational
age of the fetus, the existence of diabetes, the consumption degree of tobacco and
alcohol, and the measure of the AFP and hCG hormones. Current methods also use a
reduction of variables used in medicine called MoM (Multiple of Median) 18. This
reduction is commonly used in this problem and consists in a regression based on the
median of the 2 hormones. The median values are previously modified by other vari-
ables, which produces 2 new variables that include the value of the rest of variables
