稀疏贝叶斯混合专家模型及其在光谱数据标定中的应用

在光谱数据的多元校正中, 光谱数据通常是在多种不同的环境条件下收集的. 为了建模来源于不同环境中的高维光谱数据, 本文提出了一种新的稀疏贝叶斯混合专家模型, 并将其用来选择多元校正模型的稀疏特征. 混合专家模型能够把训练数据划分到不同的子类, 之后使用不同的预测模型来分别对划分后的数据进行预测, 因此这种方法适合于建模来自于多种环境下的光谱数据. 本文提出的稀疏的混合专家模型利用稀疏贝叶斯的方法来进行特征选择, 不依赖于事先指定的参数; 同时利用probit模型作为门函数以得到解析的后验分布, 避免了在门函数分类模型中进行特征提取时需要的近似. 本文提出的模型与其他几种常用的回归模型在人工数据集和几个公开的光谱数据集上进行了比较, 比较结果显示本文提出的模型对多个来源的光谱数据进行浓度预测时精度比传统的回归方法有一定的提高.

Sparse Bayesian Mixture of Experts and Its Application to Spectral Multivariate Calibration

In spectral multivariate calibration, high dimensional spectral data are often measured on different conditions. To predict the property value of a spectrum without knowing its source, a new sparse Bayesian mixture experts (ME) model is proposed and applied to the multivariate calibration model for selecting the sparse features. The technique of mixture of experts can divide the training data into some different classes and estimate the different predictive functions for each class. Therefore, ME is suitable for prediction of multiple-source spectral data. To analyze high dimensional spectral data, the new ME model employs the sparse Bayesian method to select certain features without tuning parameters. Moreover, the multinomial probit model is used as the gate function to obtain the analytic posterior distribution in this model. This new method is compared with some classical multivariate calibration methods on artificial and some real-world datasets. Experimental results show the advantage of proposed model for high dimensional spectral data.