|
In their paper, Zhang and King suggest that there is an interesting relationship between the separation of classes via SVMs and similarly using the Gabriel graph of the total dataset.
The Gabriel graph of a set of points involves forming an edge between any two unique points for which the diametrical sphere (or hypersphere in higher dimensions) does not contain any other points in the set.
The following image looks at the similarity of classification using both the SVM (the background) and the Gabriel graph's boundary between nodes from different classes. (The circled nodes represent Gabriel-edit points, which will be defined in the following section.)
[image of results for previously shown NL points]
Another sample result for a spiral dataset is shown below.
From these examples, it seems that both SVMs and Gabriel graphs are capable of defining similar boundaries in nonlinearly separable cases.
|