This paper presents a method to speed up support vector classification, especially important when data is highdimensional. Unlike previous approaches which focus on less support vectors, we partition the data space into local regions, and perform approximation by linear functions. The experimental results on 31 datasets show that the performance degrades marginally, while the speedup is significant, up to three orders of magnitude.
Thang V. Pham, Arnold W. M. Smeulders