An Improved Cache-Based PTSVM Learning Algorithm

Support vector machine is gaining popularity due to many attractive features and promising empirical performance in the fields of nonlinear and high dimensional pattern recognition. TSVM (transductive support vector machine) takes into account a particular test set and tries to minimize misclassifications of just those particular examples. PTSVM (progressive transductive support vector machine) can automatically adapt to different data distributions and realize a transductive learning of support vectors in a more general sense. However, the process of pairwise labeling of PTSVM in the margin band is unnatural and products errors more easily. Although dynamical adjusting offers some sort of error recovery function, its ability is limited. In allusion to the shortcomings of PTSVM learning algorithm, ICPTSVM (an improved cache-based PTSVM) learning algorithm is presented. The algorithm uses pairwise labeling in the range and error-correcting on cache to replace pairwise labeling in the margin band and dynamical adjusting. Then it greatly reduces the number of mis-labeling and improves the speed and accuracy. Experiments data show the validity of this algorithm.