Geoscience Reference
In-Depth Information
1
0.8
0.6
0.4
0.2
0
2
4
6
8
10
12
14
16
Number of features
Fig. 12.7
Feature selection using cross validation
pkfssvm -t training_features.sqlite -cv 5 -n 16 -sm sffs -v 1
...
65 25 (cost: 0.720811)
65 25 0 (cost: 0.844122)
65 25 0 50 (cost: 0.909954)
65 25 0 50 83 (cost: 0.928493)
65 25 0 50 83 9 (cost: 0.937573)
9 65 25 0 50 83 69 (cost: 0.950058)
69 9 65 25 0 50 83 85 (cost: 0.960273)
25 0 50 83 85 69 9 65 116 (cost: 0.967083)
85 69 9 65 116 25 0 50 83 67 (cost: 0.970489)
116 25 0 50 83 67 85 69 9 65 16 (cost: 0.971623)
0 50 83 67 85 69 9 65 16 116 25 78 (cost: 0.972002)
85 69 9 65 16 116 25 78 0 50 83 67 133 (cost: 0.972759)
78 0 50 83 67 133 85 69 9 65 16 116 25 102 (cost: 0.975785)
0 50 83 67 133 85 69 9 65 16 116 25 102 78 44 (cost: 0.976542)
83 67 133 85 69 9 65 16 116 25 102 78 44 0 50 98
(cost:
→
0.977299)
...
The output shows how the floating routine adds features after each iteration.
The cost represents the Kappa value based on a cross validation using the selected
features. In a floating search, previously selected features can also be rejected and
replaced by new ones. More simple approaches such as the sequential forward search
(SFS) fix all previously selected features. They have a higher risk to be trapped in
local maximum, which can result in sub-optimal solutions. In Fig.
12.7
, the accuracy
measure, based on the Kappa value, is shown in function of the number of selected
features.
Search WWH ::
Custom Search