期刊
IEEE TRANSACTIONS ON FUZZY SYSTEMS
卷 30, 期 8, 页码 3176-3190出版社
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TFUZZ.2021.3106330
关键词
Training; Fuzzy systems; Bagging; Task analysis; Fuzzy sets; Firing; Complexity theory; Adaptive network-based fuzzy inference system (ANFIS); bagging; dropout; ensemble learning; interpretability
资金
- National Program on Key Basic Research Project [2019YFB1704900]
- National Natural Science Foundation of China [51635006, 51675199]
- Key Basic Research Project of Guangdong Province [2019B090918001]
In this paper, an ensemble classifier is proposed to enhance the accuracy and interpretability of fuzzy systems when dealing with high-dimensional data.
Improving the tradeoff between accuracy and interpretability is essential for the problem of handling high-dimensional data in Takagi-Sugeno-Kang (TSK) fuzzy systems and providing insights into real-world tasks. However, the TSK fuzzy system becomes complex and challenging to interpret as the data dimension increases. Here, we report an ensemble classifier, which is an enhanced adaptive network-based fuzzy inference system (ANFIS) integrating improved bagging and dropout to build concise fuzzy rule sets. First, the high-dimensional feature space is decomposed into a series of low-dimensional feature subsets using the bagging and random subspace method to train multiple ANFISs. An improved dropout strategy is then applied in training ANFISs by temporarily disabling rules in each training epoch and deleting rules after training to obtain sparse rulesets with high-quality rules. These sub-models are subsequently aggregated to perform the fuzzy inference. Results on high-dimensional benchmark datasets confirm that both the bagging and dropout strategies are effective, providing high interpretability by reducing the co-firing degrees and rules of sub-models while guaranteeing accuracy at the same time.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据