In practice, one difference is that datasets with large gaps get treated much different. kNN will pick far away points, but it is possible that relatively small Parzen Windows will actually enclose zero points. In this case, only the priors can be used to classify. Unfortunately, this dataset had many holes in it at the fringes, as was visible in the PDF graphs seen earlier.
I classified points using a variety of window sizes. I picked
20 equal-sized steps from the mininum to the maximum values of
distGrid, the matrix which stores the distances from
all points in the training set to all points in the testing set.
Feat. # | fwd | Bwd | rand | |
1 | 1 | 1 | 0 | |
2 | 0 | 1 | 0 | |
3 | 1 | 1 | 0 | |
4 | 0 | 0 | 0 | |
5 | 0 | 1 | 0 | |
6 | 0 | 0 | 1 | |
7 | 0 | 1 | 0 | |
8 | 0 | 0 | 0 | |
9 | 0 | 0 | 1 | |
10 | 0 | 0 | 1 | |
11 | 1 | 0 | 1 | |
12 | 0 | 0 | 0 | |
13 | 0 | 0 | 0 | |
14 | 0 | 0 | 0 | |
15 | 0 | 0 | 1 | |
16 | 0 | 0 | 0 | |
17 | 1 | 0 | 0 | |
18 | 0 | 0 | 1 | |
19 | 1 | 0 | 0 | |
20 | 1 | 0 | 0 | |
21 | 1 | 0 | 0 | |
22 | 1 | 0 | 0 | |
23 | 1 | 0 | 0 | |
24 | 0 | 0 | 0 | |
Validation Rate | 62.02% | 62.79% | 58.91% | |
Testing Rate | 58.8% | 59.6% | 58.1% |