【摘要】 由于One-class支持向量机能用于无监督学习,被广泛用于信息安全、图像识别等领域中。而超球体One-class支持向量机能生成一个合适的球体,将训练样本包含其中,故更适合于呈球形分布的样本学习。但由于超球体One-class支持向量机没有一种快速训练算法,使其在应用中受到限制。SMO算法成功地训练了标准SVM,其训练思想也可用于超球体One-class支持向量机的训练。本文提出了超球体One-class支持向量机的SMO训练算法,并对其空间和时间复杂度进行了分析。实验表明,这种算法能迅速、有效地训练超球体One-class支持向量机。更多还原

【Abstract】 One-Class SVM, as an unsupervised learning algorithm, is used widely in the areas of information security and image recognition etc. Moreover, Hyper-Sphere One-Class SVM can product a right sphere including the training examples, so it is fit to learn the examples with sphere-shaped distribution. However, Hyper-Sphere One-Class SVM is limited in real applications because it lacks a fast training algorithm. Training standard SVM successfully, the idea of SMO algorithm can be used to train Hyper-Sphere One-Class SVM too. The SMO algorithm for Hyper-Sphere One-Class SVM is proposed, the space and time complexity degrees are also analyzed in this paper. As shown in our numeric experiments, the new algorithm can train Hyper-Sphere One-Class SVM precisely and efficiently.更多还原