Incremental Support Vector Machine Learning: An Angle Approach

When new samples joining, classical Support Vector Machines must retrain the whole dataset which contains both historical samples and additional samples. Incremental Support Vector Machines can avoid retraining whole dataset through disposing of redundant samples. According to the angle, which is between the subtraction new sample from historical samples and the historical separation plane, MAISVM 1 (Minimum Angle Incremental Support Vector Machines 1) and MAISVM 2 (Minimum Angle Incremental Support Vector Machines 2) are proposed in this paper. The additional data, the support vectors and the samples, of which the angle between subtraction additional sample and the historical separation plane is minmum, are retained in MAISVM 1. Support vectors replace with generalized linear support vectors in MAISVM 2. Empirical results show that the MAISVM 1 has better accuracy than SVM-INC., and a faster speed than LISVM. The performance of MAISVM 2 is better than MAISVM 1. Its accuracy is no less than LISVM and its speed is faster than SVM-INC. MAISVM 1 can effectively discard the redundant samples in the neighborhoods of new sample. By selecting an appropriate subset of support vector set, MAISVM 2 is faster than SVM-INC.