基于Fisher判别的半监督天体光谱数据特征降维

降维是天体光谱数据预处理常用的手段之一,如何利用标号天体光谱数据,克服降维过程中的过分拟合,是提高降维效果的有效途径之一。采用半监督学习,给出了一种天体光谱数据特征降维方法。该方法首先针对具有标号天体光谱数据,建立Fisher判别分析和PCA可变动选择的不确定关系;其次构建其半监督降维的全局最优化形式,通过特征值分解计算降维结果,从而有效地克服了天体光谱降维过程中的过分拟合问题;最后采用高红移类星体和晚型星SDSS天体光谱特征线数据集,实验验证了该方法的有效性。

Semi-supervised Dimension Reduction of Spectral Characteristic Based on Fisher Discriminant Analysis

Dimensionality reduction is one of the common methods to preprocess astronomical spectral data.How to use label astronomical spectral data for conquering over-fitting of the dimensionality reduction process is an effective way to improve the effect of dimensionality reduction.In this paper,a dimension reduction method of astronomical spectrum data feature is presented by utilizing semi-supervised learning.Firstly,for the label Celestial spectral data,an uncertainty relation is established in which Fisher discriminant analysis and PCA can be selected variably.Secondly,the global optimization of semi-supervised dimensionality reduction is built.Dimensionality reduction results are calculated through the eigenvalue decomposition,so that the problem of over-fitting is solved in astronomical spectral data dimensionality reduction.In the end,the experimental results proved the validity of the method by using the hzqso and mstar astronomical spectral features line data sets.