融合Softmax回归和偏最小二乘的中药数据分析方法研究

中药呈现多成分、多靶点、多药效指标等特点，决定了中药数据具有多自变量、多因变量和非线性的特征，研究适合中药数据特点的多元分析方法显得尤为迫切。偏最小二乘方法在中药数据分析中具有优势，也存在局限：T平方椭圆图法在辅助分析高维中药数据特异点时存在不足；决定系数不能完全反映变量的重要性；模型内部的线性本质不能充分反映中药数据非线性特点。基于以上重要科学意义和偏最小二乘法的不足，本课题以近年来承担的中药领域课题数据为基础，收集整理实验数据，设计中药样本数据库；基于划分的对比散度受限波尔兹曼机快速学习算法，识别剔除样本数据中的特异点；基于随机森林算法分析变量的重要性，筛选重要的建模变量；模型内部嵌入Softmax非线性回归改进偏最小二乘建模。本研究聚焦设计有利于中药数据特点的多元非线性数据分析模式，为科学分析中药数据、揭示中医药内涵提供技术支撑。

Traditional Chinese Medicine (TCM) present features of more compositions, more targets and more efficacies.TCM data present multiple variables, multiple dependent variables and nonlinear characteristics. It is very urgent to research multivariate analysis method for TCM data. Partial least square method has advantages in the TCM data analysis, but there are some limits: T square ellipse diagram method has some deficiency in analyze the outlier data of high-dimensional TCM data; determination coefficients cannot fully reflect the importance of variables; the model cannot fully reflect nonlinear characteristics of TCM data by the linear nature of the internal model. For the above important scientific significance project and the shortcomings of the partial least squares, on the basis of project data of TCM in recent years, collect and collate the experiment data, design sample database of TCM; Using the method of divide Contrastive Divergence Restricted Boltzmann Machine to identifying and eliminate the outlier point from sample data; analyze the importance of variables and screening important modeling variable based on random forest; embed Softmax nonlinear regression internal model to improved partial least squares modeling. The research focus on design multivariate data analysis model propitious to the characteristics of TCM data, to provide technical support for scientific analysis of TCM data.