关键词: 核主成分分析, 数据流, 降维

Abstract: Theory and implementation of two data stream dimension reduction algorithms, PCA and KPCA, are analyzed. Due to linear PCA and KPCA can not effectively reduce data stream dimension when applied over large scale stream data, a new dimension reduction technique called GKPCA is proposed. With GKPCA, data sets are first partitioned into groups, and then KPCA is applied over each group. Data sets are filtered and regrouped into a new dataset. KPCA is again evaluated over the new data sets. This process is preceding recursively when some reduction threshold is reached which simplifies data stream sampling space and reduces time and space complexity of KPCA. Experimental analysis over different datasets illustrates that GKPCA can reduce data stream dimension excellently with less time consumption.

Key words: Kernel Principal Component Analysis（KPCA）, data stream, dimension reduction