高维数据分类方法研究

Study on Classification Methods for High-dimensional Data

在对高维度数据进行模式分类时,能否有效进行降维是一个关键问题。提出了一种结合高斯过程潜变量模型(GPLVM)和支持向量机(SVM)的阶梯跳跃降维分类框架方法,能有效的降低样本数据维数,同时提高分类器性能。利用GPLVM实现数据的平滑映射,对输入样本进行非线性降维后,根据SVM的分类校验结果进行下一步降维迭代操作;计算新的阶梯维数,根据反馈动态调整降维输入数据。利用该方法对UCI上的数据集进行分类,仿真结果验证了方法的有效性。

Effective dimensionality reduction is a key issue in high-dimensional data classification. A new ladder jumping dimensional reduction classification framework was proposed which combined the Gaussian process latent variable model （GPLVM） and the Support Vector Machine （SFM）. The data dimensions were reduced remarkably, while at the same time improving the performance of SFM classifiers. For the purpose of nonlinear low dimensional embedding of sample datasets, GPLVM provides a smooth probabilistic mapping from latent to data space. According to the feedback results of SVM, the renewed ladder dimension was calculated and the input data was adjusted dynamically. The proposed approach was applied to four benchmark problems, and the simulation results show its validity.