论文部分内容阅读
提出使用模拟切削算法的SVM增量学习机制.模拟切削算法在核函数映射的特征空间中计算每个样本的预期贡献率,仅选取预期贡献率较高的样本参与SVM增量学习,有效解决传统SVM增量学习代价高、目标样本选取准确性低、分类器缺乏鲁棒性的问题.一个样本的预期贡献率采用通过该样本的映射目标的合适分离面对两类样本的识别率来表示.对目标样本的选取酷似果蔬削皮的过程,所提算法由此得名.基准数据实验表明,文中算法在学习效率和分类器泛化性能上具有突出优势.在有限资源学习问题上的应用表明该算法在大规模学习任务上的良好性能.
This paper proposes an incremental learning mechanism of SVM using simulated cutting algorithm.The simulated cutting algorithm calculates the expected contribution rate of each sample in the feature space mapped by kernel function and selects only samples with higher expected contribution rate to participate in SVM incremental learning to effectively solve the traditional SVM incremental learning cost is high, the accuracy of the target sample selection is low and the classifier is not robust. The expected contribution rate of a sample is expressed by the recognition rate of the two types of samples with the appropriate separation through the mapping target of the sample. The selection of the target sample resembles the process of skin peeling of fruits and vegetables, hence the name of the proposed algorithm.Experimental results show that the proposed algorithm has outstanding advantages in learning efficiency and generalization performance of classifiers.The application of finite resource learning The algorithm performs well on large-scale learning tasks.