It is a kernel operating on ExamplePairs applying the following formula:
\(K( < x_1, x_2 >, < y_1,y_2 > ) = BK(x_1, y_1) + BK(x_2, y_2) - BK(x_1, y_2) - BK(x_2, y_1)\)
where BK is another kernel the preference kernel relies on.
The preference kernel was firstly introduced in:
[Shen and Joshi, 2003] L. Shen and A. K. Joshi. An SVM based voting algorithm
with application to parse reranking. In Proc. of CoNLL. 2003.