基于SHAP值的类重叠识别方法对FSVM的改进

Improvement of FSVM by the method of class overlap recognition based on SHAP value

在分类问题中,类重叠现象会大大影响分类模型的效果,针对类重叠样本的识别问题,提出了一种基于SHAP值的类重叠识别新方法,基于SHAP值构造出样本在所属类中的分类作用能力隶属属性,对类间重叠样本进行有效识别,然后利用仿真实验验证了基于SHAP的类重叠识别方法的适用性;将样本的分类作用能力归一化后构造出样本隶属度度量,并将该隶属度应用于模糊支持向量机(Fuzzy Support Vector Machine,FSVM)算法后得到FSVM_SHAP模型,通过在多个经典二分类数据集上实验得到了较好的效果,体现了该模型的有效性.

In the classification problem,the phenomenon of class overlap will greatly affect the effectiveness of the classification model.A new method of class overlap recognition based on SHAP values is proposed for the identification of class overlap samples.Based on the SHAP value,the membership attribute of the sample's classification ability in the class to which it belongs is constructed to effectively identify the overlapping samples between classes.The applicability of the class overlap recognition method based on SHAP is verified by using simulation experiments;After normalizing the classification capability of samples,the membership measure of samples is constructed and applied to the fuzzy support vector machine(FSVM)algorithm to obtain FSVM_SHAP,The model has been tested on several classic binary data sets and achieved good results,which reflects the effectiveness of the model.