Multiobjective Fuzzy Genetics-Based Machine Learning for Multi-Label Classification
Multiobjective Fuzzy Genetics-Based Machine Learning for Multi-Label Classification
复制标题
DOI：
10.1109/fuzz48607.2020.9177804
复制
发表时间：
2020-07
期刊：
影响因子：
--
通讯作者：
Yuichi Omozaki;Naoki Masuyama;Y. Nojima;H. Ishibuchi
中科院分区：
文献类型：
--
作者：
Yuichi Omozaki;Naoki Masuyama;Y. Nojima;H. Ishibuchi
在多标签分类问题中，多个类别标签被分配给每个实例。文献中研究了两种方法。一种是数据转换方法，它将多标签数据集转换为多个单标签数据集。然而，这种方法往往失去了多类分配的类之间的相关性信息。另一种是方法自适应方法，其中传统的分类方法扩展到多标签分类。近年来，人们提出了一些可解释的多标签分类模型。在分类过程的透明度方面，也讨论了其高度可解释性。虽然可解释性是模糊系统的一个众所周知的优点，但它们在多标签分类中的应用还没有得到很好的研究。由于多标签分类问题往往具有模糊的类边界，模糊系统似乎是一个很有前途的多标签分类方法。在本文中，我们提出了一个新的多目标进化模糊系统，它可以被归类为一种方法自适应方法。该算法产生的非支配分类器的准确性和复杂性之间的不同权衡。我们使用合成的多标签数据集检查所提出的算法的行为。我们还比较了所提出的算法与五个代表性的算法。我们在真实世界数据集上的实验结果表明，所获得的模糊分类器具有少量的模糊规则具有较高的透明度和可比的泛化能力，其他检查多标签分类算法。
In multi-label classification problems, multiple class labels are assigned to each instance. Two approaches have been studied in the literature. One is a data transformation approach, which transforms a multi-label dataset into a number of singlelabel datasets. However, this approach often loses the correlation information among classes in the multi-class assignment. The other is a method adaptation approach where a conventional classification method is extended to multi-label classification. Recently, some explainable classification models for multi-label classification have been proposed. Their high interpretability has also been discussed with respect to the transparency of the classification process. Although the explainability is a well-known advantage of fuzzy systems, their applications to multi-label classification have not been well studied. Since multi-label classification problems often have vague class boundaries, fuzzy systems seem to be a promising approach to multi-label classification. In this paper, we propose a new multiobjective evolutionary fuzzy system, which can be categorized as a method adaptation approach. The proposed algorithm produces nondominated classifiers with different tradeoffs between accuracy and complexity. We examine the behavior of the proposed algorithm using synthetic multi-label datasets. We also compare the proposed algorithm with five representative algorithms. Our experimental results on real-world datasets show that the obtained fuzzy classifiers with a small number of fuzzy rules have high transparency and comparable generalization ability to the other examined multi-label classification algorithms.