并行集成具有高可解释的TSK模糊分类器  

Ensemble TSK fuzzy classifiers with parallel learning and high interpretability

作　　者：张雄涛 蒋云良[2] 胡文军[2] 王士同[1] ZHANG Xiong tao;JIANG Yun-liang;HU Wen-jun;WANG Shi-tong(School of Digital Media,Jiangnan University,Wuxi 214122,China;School of Information Engineer,Huzhou University,Huzhou 313000,China)

机构地区：[1]江南大学数字媒体学院,江苏无锡214122 [2]湖州师范学院信息工程学院,浙江湖州313000

出　　处：《控制与决策》2020年第10期2535-2542,共8页Control and Decision

基　　金：国家自然科学基金项目(61572236,61300151,61772198,61771193)。

摘　　要：针对分层Takagi-Sugeno-Kang (TSK)模糊分类器可解释性差,以及当增加或删除一个TSK模糊子分类器时Boosting模糊分类器需要重新训练所有TSK模糊子分类器等问题,提出一种并行集成具有高可解释的TSK模糊分类器EP-Q-TSK.该集成模糊分类器每个TSK模糊子分类器可以使用最小学习机(LLM)被并行地快速构建.作为一种新的集成学习方式,该分类器利用每个TSK模糊子分类器的增量输出来扩展原始验证数据空间,然后采用经典的模糊聚类算法FCM获取一系列代表性中心点,最后利用KNN对测试数据进行分类.在标准UCI数据集上,分别从分类性能和可解释性两方面验证了EP-Q-TSK的有效性.Traditional ensemble Takagi-Sugeno-Kang(TSK) fuzzy subclassifiers face such challenges, hierarchical learning have no interpretability, because of the presence of intermediate variables, and when a new TSK fuzzy subclassifier is added to or removed from the structure of the current fuzzy classifier, boosting learning must retrain each TSK fuzzy subclassifier by appropriately assigning new weights. Therefore, An ensemble framework EP-Q-TSK of TSK fuzzy subclassifiers with parallel learning way is proposed. The proposed framework has the following distinctive characteristics: 1) Each TSK fuzzy subclassifier can be built quickly with least learning machine(LLM) in parallel;2)As a novel ensemble learning, the proposed framework augments the original validation data space with the outputs of each TSK subclassifier in an incremental and inexpensive way, and then speed up the final classification on the validation data by using the FCM and the KNN method;3) Enhanced classification performance by FCM & KNN is experimentally revealed, and the experimental results on benchmark datasets indicate the effectiveness of EP-Q-TSK and its parallel learning method in the sense of both enhanced classification performance and interpretability.

关 键 词：集成TSK模糊分类器 并行学习 最小学习机 代表性中心点 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]