基于增量核主成分分析的数据流在线分类框架

核主成分分析(Kernel principal component analysis, KPCA)是一种非线性降维工具, 在降低数据流分类处理量方面发挥着积极作用. 然而, 由于复杂性太高, 导致KPCA的降维能力有限. 为此, 本文给出了一种增量核主成分分析算法(Incremental KPCA for dimensionality-reduction, IKDR), 该算法在每步迭代估计中只需线性内存开销, 大大降低了复杂性. 在IKDR的基础上, 结合BP (Back propagation)神经网络提出了数据流在线分类框架: IKOCFrame (Online classification frame based on IKDR). 通过一系列真实和人工数据集上的实验, 检验了IKDR算法的收敛性, 并且验证了IKOCFrame相对于同类基于成分分析的分类算法的优越性.

Online Classification Framework for Data Stream Based on Incremental Kernel Principal Component Analysis

Kernel principal component analysis (KPCA) has been suggested for various data stream classification tasks requiring a nonlinear transformation scheme to reduce dimensions. However, the dimensionality reduction ability is restricted because of its high complexity. Therefore this paper proposes an incremental kernel principal component analysis algorithm: IKDR, which iteratively estimates the kernel principal components with only linear order storage complexity per iteration. On the basis of IKDR, this paper proposes an online classification framework for data stream: IKOCFrame. Extensive experiments on real and artificial datasets validate the convergence of IKDR and confirm the superiority of IKOCFrame over other recent classification schemes based on component analysis.