A Ball Vector Machine based on improved enclosing ball iterative solution strategies

摘要

Aimed at the problem of Ball Vector Machine's long training time for large scale data, an improved enclosing ball vector machine (IEBVM) based on new iterative solution strategy was proposed. When solving enclosing ball (EB) problem, IEBVM caches the dot product of training points and ball center for the distance solution next time, making the solution independent of support vectors weights. The training points which are unable to become the furthest point are ruled out. In addition, the support vectors weights can be updated once in a certain number of iterations to reduce the calculation amount. Moreover, the number of search times in the support vectors set is increased. Compared with BVM and LIBSVM in large scale datasets, IEBVM significantly reduces the training time and the number of support vectors, simultaneously keeping high testing accuracies.