General Linear Discriminant

I wanted to try using a GLD next, because it is a fairly simple algorithm. It executes very rapidly, so I could do plenty of experimentation with different combinations of features. Based on what I had seen of the distributions of data earlier, though, I suspected it might not work well. This is because GLD builds boundaries from polynomials. I suspected the boundaries of my data were very complex. To test this I graphed some features against each other. As you can see, the two datasets "good" and "bad" are almost hopelessly intertwined for most features.

1 vs. 2
5 vs. 8
17 vs. 10
14 vs. 17 vs. 19

The Good Features

Similar to what I had done with Bayes, I used random feature-seeking to find the best features. They seemed to be 5, 6, 10, 12 and 17.
Feat. # Rand1 Rand2 Rand3
1 0 0 0
2 0 0 0
3 1 0 0
4 1 0 0
5 1 0 1
6 1 0 1
7 0 0 0
8 1 0 0
9 0 0 0
10 0 1 1
11 0 0 0
12 1 1 0
13 0 0 1
14 1 0 0
15 1 0 1
16 0 0 0
17 1 1 1
18 0 0 0
19 0 1 0
20 0 0 0
21 1 0 0
22 1 0 0
23 1 0 0
24 0 0 1
rate:65.13% 65.79% 66.48%

Then, I graphed these features against each other. It is pretty clear that there is no low order polynomial boundary to separate them. But scientists have to be very, very sure. For example, some graphs have the same color at both fringes that might be useful in classifcation:

17 vs. 10
6 vs. 5
12 vs. 17

Results

Therefore, I experimented with values 2 and 4 for K. K=2 improved classification somewhat, probably picking up those fringes with its parabola shape. K=4 performed miserably at 50/50!

Polynomial Order (K) Validation Classification Rate Correct Classification Rate
1 57.2% 55.9%
2 58.1% 56.6%
4 52.3% 50.0%

--->Fischer Linear Discriminant

Index