Abstract:

Hyper-sphere Support Vector Machine (SVM) is very effective for solving multi-class classification problems. Considering data distribution is very important for convergence of solving support vectors, a weight factor is imported into the original hyper-sphere SVM. After computing data for each training class, this weight factor is decided by its center-distance ratio. In the training process, data with bigger weight is put into the data processing thread first and is then followed by smaller ones. To save computation cost, a parallel genetic algorithm based SMO multi-threading is adopted. For a test sample, its class decision is based on its position with each classification of hyper-sphere. If all class-specific hyper-spheres are independent of each other, a new test sample can be classified correctly. But, if some hyper-spheres have common spaces, that is, one hyper-sphere intersects with one or more hyper-spheres, it is hard to decide the class of the test sample. Based on detailed analysis of three decision rules for the intersection data classification, one decision rule that combines the kNN method is put forward in this paper. For other simple inclusion cases, the simple decision rule is defined. Through two real experimental results of navigation tracking and ship meeting situations classification, our new proposed algorithm has a higher classification accuracy and boasts a lower computation cost than other algorithms.