面向单调分类的简洁单调TSK模糊系统

Concise monotonic TSK fuzzy system for monotonic classification

Takagi-Sugeno-Kang(TSK)模糊系统的一致逼近能力和可解释性使其可以直观高效地描述复杂的非线性不确定系统,可以有效地应用于模式分类.然而,对于单调分类任务,现有的模糊分类算法并没有考虑单调数据存在的有序关系,因此这些算法对于单调分类任务在模型的复杂度和分类性能方面有待改进.针对此问题,提出了面向单调分类的简洁单调TSK模糊系统建模方法(Concise Monotonic TSK Fuzzy System for Monotonic Classification,CM-TSK-FS),引入有序互信息进行单调特征选择,然后利用抽取的特征来训练TSK模糊系统进行分类识别.该方法有如下优点:(1)由于对单调数据进行了特征选择,新方法降低了TSK模糊系统规则的复杂性,因而得到的模糊系统更加简洁;(2)由于在特征抽取时考虑了单调数据的特征值和决策值之间的单调性,使得训练的模型的分类性能也有了一定程度的提高.在多个单调数据集上进行了实验验证,实验结果表明:面向单调分类的简洁单调TSK模糊系统在处理单调数据集时,通过选取重要的单调数据特征,不仅可以降低其模型的复杂性,还可以提高分类精度.

Universal approximation and interpretability of Takagi-Sugeno-Kang（TSK）fuzzy systems make it possible to describe complex nonlinear uncertain systems intuitively and efficiently.TSK fuzzy systems have been effectively applied to pattern classification.However,for the monotonic classification task,the existing fuzzy classification algorithms still do not consider the ordinal relationship of the monotonic data adequately,so that the classification performance of these algorithms for monotonic classification tasks in the model needs to be improved.To solve this problem,Concise Monotonic TSK Fuzzy System for Monotonic Classification （CM-TSK-FS） is proposed.It uses rank mutual information for monotonic feature selection,and then uses the extracted features to train TSK fuzzy system for classification and recognition.The proposed method has the following advantages：1）The new method reduces the complexity of TSK fuzzy system due to the feature selection of monotonic data which makes the fuzzy system more concise.2）Since the monotonicity between the features and the decision values of the monotonic data is taken into account when extracting the features,the classification performance of the training model has been improved to a certain extent.The experimental results show that the concise monotonic TSK fuzzy system for monotonic classification,and the simple monotonic TSK fuzzy system for monotonic classification can not only reduce the complexity of the model by selecting important monotonic data features when dealing with monotonic data sets,but also improve the classification accuracy.