实现兼类样本增量学习的一种算法

AN ALGORITHM REALIZING INCREMENTAL LEARNING FOR MULTI-CLASS SAMPLE

针对兼类样本,提出一种增量学习算法。利用超球支持向量机,在特征空间对属于同一类别的样本求得一个能包围该类尽可能多样本的最小超球,使各类样本之间通过超球隔开。增量学习过程中,只对新增样本以及与新增样本具有相同兼类的旧样本集中的支持向量进行训练,且每次训练只针对一类样本,使得算法在很小的样本集、很小的空间代价下实现兼类样本增量学习,同时保留了与新增样本类别无关的历史训练结果。分类过程中,通过计算待分类样本到各超球球心的距离判定其所属类别,分类准确快捷。实验结果证明了该算法的有效性。

In light of multi-class sample, an incremental learning algorithm is proposed. Hyper-sphere support vector machine is used in feature space for samples belong to the same class to find a smallest hyper-sphere that contains most samples of that class, to have different class samples be separated from each other. In the process of incremental learning, only the new samples and the support vectors in historical sample set that have the same multi-class as the new samples are trained, and just one class of samples are trained each time, therefore, the multi-class incremental learning can be realized in a small set of samples and a small memory space, and the historical training results of the classes that are irrelative to the new samples are reserved at the same time. For the sample to be classified, the distances from it to the centre of each hyper-sphere are used to determine what class the sample belongs to. The classification method is precise and the fast. The experimental result shows the validity of the proposed algorithm.