基于输入扰动的分层混合模糊-神经网络敏感性分析  

Sensitivity Analysis of Hierarchical Hybrid Fuzzy - Neural Network to Input Perturbation

作　　者：邢海花[1,2] 陈焕东[1] 林红燕[1] 

机构地区：[1]海南师范大学信息科学与技术学院,海口571158 [2]北京师范大学信息科学与技术学院,北京100875

出　　处：《软件》2013年第2期52-55,共4页Software

基　　金：国家自然科学基金(41072245;41272359);海南省自然科学基金(612123)

摘　　要：分层混合模糊-神经网络(HHFNN)能较好地处理离散和连续变量并存的高维数据,为了提高其解释能力和在数据挖掘应用中的说服力。本文针对输入中离散和连续数据并存的特点,提出一种基于输入扰动的分层混合模糊-神经网络敏感性分析方法,首先优化HHFNN模型,然后依次给测试数据集的各个输入变量添加高斯白噪声,计算模型加入噪声前后的均方误差(MSE)。MSE差值越大说明相应的输入对模型输出影响越大,其重要性程度越高。为了验证本文提出方法的有效性,选用人工模拟数据集和UCI-Pyrimidines数据集进行实验,结果表明本文所提方法能够筛选出模型重要性程度高的属性,有效地约简了模型。The hierarchical hybrid fuzzy - neural network（HHFNN） can deal with high-dimensional inputs containing both discrete and continuous variables. In order to provide explanatory insight into the contributions of the input variables in the model and to improve its credibility in data mining applications. In this study, we proposed a sensitivity analysis method of HHFNN model based on the input perturbations. Firstly, we optimize HHFNN model, and then add white Gaussian noise to the test dataset for each input variable, and calculate the mean square error of model output. The greater the difference of MSE the greater impact of the input to the model output, and it means the input variable more important. To verify the effectiveness of the proposed method, this study employed a man-made examples and a UCI-Pyrimidines dataset to test the performance of the method. The results showed that the sensitivity analysis method can really identify important variables of model and simplify the model effectively.

关 键 词：敏感性分析 输入扰动 分层混合模糊-神经网络(HHFNN) 变量重要性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]