近红外光谱（NIR）分析具有测试方便、 不破坏样本、 响应快速等优势, 但是, 由于在谱带分布和结构分析中存在着许多复杂因素, 使得在提取特征光谱信息时存在许多困难。 现阶段, 虽然已经有多种光谱数据降维方式被广泛使用, 但是这些传统的数据降维方式都有一个局限性, 就是数据的降维仅仅针对于一个数据集, 当数据集中有多个关键因素形成干扰时, 数据降维和分类的结果往往不是很理想, 得不到想要分析的信息。 这一问题造成了在分析近红外光谱时建立的数据降维模型极差, 无法正确的对样品进行预测分类。 对比主成分分析（contrastive principle component analysis, cPCA）是一种基于主成分分析（PCA）的改进算法, 起源于对比学习, 并应用于基因组信息解析。 cPCA算法的优势就是能够将一个数据集中的降维推广到两个相关联数据集之间的降维, 从而能够得到数据集中的关键信息。 将cPCA算法应用于近红外光谱处理中, 建立了准确的近红外光谱数据降维模型。 在实验验证中, 使用cPCA算法对不同类型水果（苹果和梨）表面农药残留进行分析。 结果表明, 在对不同类型的水果进行农药残留分析时, 使用PCA算法进行数据降维只能区分出不同的水果类型, 而水果表面是否喷洒农药这一关键的特征信息并不能分析出来; 而使用cPCA算法进行数据降维分析时, 由于对背景光谱的约束作用, 能够清晰的将有无喷洒农药的样本分类。 这说明了, cPCA在近红外光谱数据降维中有着明显的优势, 解决了近红外光谱数据降维模型中数据集受限和特征信息的提取问题, 进而建立准确的近红外光谱数据降维模型。

Near-infrared spectroscopy (NIR) analysis is considered as a promising chemical analysis technique because its advantages of convenient-testing, no damaging and fast response. However, due to the many unknown factors in the band distribution and structural analysis of the near-infrared spectrum, there are many difficulties in extracting the characteristic spectral information. Nowadays, although a variety of spectral data dimensionality reduction methods have been widely used, the traditional data dimensionality reduction methods have a limitation that the dimensionality reduction is restricted in one dataset. The results of data dimensionality reduction are often not ideal when there are many factors in dataset . This problem makes the data establish dimensionality reduction model extremely hard in near-infrared spectrum. Comparative Principal Component Analysis (cPCA) is an improved algorithm based on principal component analysis (PCA), which originated from Comparative Learning and applied to genomic information analysis. The advantage of the cPCA algorithm is that it can realize the dimensionality reduction between two related data sets. In this paper, the cPCA algorithm is applied to near-infrared spectroscopy for the first time and establish an accurate spectral dimensionality reduction model. In the experimental, we used the cPCA algorithm to analyze the surface of different types of fruits (apples and pears) with pesticide residues and without pesticide residues . The result showed that the PCA algorithm just distinguishes different fruit types, while the cPCA algorithm classifies the fruits with or without pesticides due to the constraint of the background dataset. This showed that cPCA outperforms in data dimensionality reduction of near-infrared spectra. It solves the problem of dataset limitation and feature information extraction in the near-infrared spectral data dimensionality, and cPCA could establish an accurate spectral data dimensionality reduction model.