基于特征子集相关度和偏最小二乘法的特征选择策略

Feature Selection Method based on Correlation based Feature Selection and Partial Least Square

在中药方剂的量-效关系分析中,需要寻找药理指标(因变量)受哪些血液指标(自变量)影响。本文提出一种基于特征子集相关度和偏最小二乘法的特征选择策略,利用特征子集相关度对药理指标进行评估预选出特征子集,然后将其放入偏最小二乘法中进行训练,利用训练后得出的残差平方和评估该特征子集是否可取,并结合顺序前向浮动混合搜索策略与顺序后向浮动混合搜索策略,综合评估以分析药理指标受哪些血液指标的影响。分别采用麻杏石甘汤君药止咳数据集及UCI数据集进行分析处理,实验结果表明该特征选择策略能较好寻找一个较优的特征子集。

In the analysis of the dose-effect relationship TCM prescription,it need to find the effect of pharmacological indexes(dependent variables)by which blood indexes(independent variables).In this paper,we put forward a kind of analysis of TCM data method based on correlation based feature selection(CFS)and partial least square(PLS),using CFS to evaluate the characteristics of the selected feature subset,then put it into the PLS for training.Using sum of Squares for Error of Train to evaluate to the desirability of this feature subset combined with(SFFS)and sequential backward floating search strategy(SBFS)to select the better feature subset.Comprehensive assessment to analyze the effect of pharmacological indicators by which blood parameters.Using TCM prescription data Ma Xing Shi Gan Tang and UCI data sets were analyzed.The experimental results show that,the feature selection strategy can find an optimal feature subset.