A Noise Clustering-induced Robust Adaptive Network-based Fuzzy Inference System for Classification
A Noise Clustering-induced Robust Adaptive Network-based Fuzzy Inference System for Classification
复制标题
DOI：
10.1109/ijcnn55064.2022.9892766
复制
发表时间：
2022-07
期刊：
影响因子：
--
通讯作者：
Katsuhiro Honda;Koki Kitamori;S. Ubukata;A. Notsu
中科院分区：
文献类型：
--
作者：
Katsuhiro Honda;Koki Kitamori;S. Ubukata;A. Notsu
自适应网络模糊推理系统（ANFIS）是一种基于神经网络的模糊推理系统（FIS）模型，是一种很有前途的可解释神经网络模型。本文从分类应用的角度考虑ANFIS的鲁棒性，引入了噪声模糊聚类的噪声抑制机制。在真实的世界分类任务中，我们经常遭受不可靠的类标签，其中一些对象具有不正确的类标签，并且应该从分类器构造中被拒绝。Davé提出的噪声模糊聚类是模糊C均值（FCM）算法在鲁棒聚类中的一种扩展，其中增加了一个噪声聚类来吸收噪声对象。由于噪声聚类方案与单一的集群情况下，可以识别与鲁棒最小二乘估计，鲁棒机制已被应用到几个最小二乘类型的数据分析。在本文中，噪声聚类方案被引入到ANFIS分类模型中，其中利用非噪声模糊隶属度，使得额外的噪声聚类吸收噪声对象，并且通过拒绝噪声对象来鲁棒地构造ANFIS分类器。最小化的成员加权最小二乘目标函数和非噪声模糊成员的估计迭代实现，直到收敛。通过使用真实的世界基准数据集的数值实验证明了所提出的方法的特点。
Adaptive Network-based Fuzzy Inference System (ANFIS) is a neural network-based model for fuzzy inference system (FIS) and can be a promising approach for explainable neural networks. In this paper, robustification of ANFIS is considered from the classification application viewpoint, where the noise rejection mechanism is introduced induced by noise fuzzy clustering. In real world classification tasks, we often suffer from unreliable class labels, where some objects have incorrect class labels and should be rejected from classifier construction. Noise fuzzy clustering proposed by Davé is an extension of fuzzy c-means (FCM) to robust clustering, where an additional noise cluster works for absorbing noise objects. Because the noise clustering scheme with single cluster cases can be identified with robust least square estimation, the robustifying mechanism has been applied to several least square-type data analyses. In this paper, the noise clustering scheme is introduced into the ANFIS classification model, where non-noise fuzzy memberships are utilized such that the additional noise cluster absorbs noise objects and the ANFIS classifier is robustly constructed by rejecting noise objects. The minimization of the membership-weighted least square objective function and the estimation of non-noise fuzzy memberships are iteratively implemented until convergence. The characteristics of the proposed method are demonstrated through numerical experiments using real world benchmark datasets.