Balancing of Samples in Class Hierarchy
Balancing of Samples in Class Hierarchy
复制标题
DOI：
10.1007/978-3-030-89691-1_22
复制
发表时间：
2021
期刊：
影响因子：
--
通讯作者：
Shuhei Aoki;Mineichi Kudo
中科院分区：
文献类型：
--
作者：
Shuhei Aoki;Mineichi Kudo
在现实世界的分类问题中，经常会出现一些类(头类)具有大量样本而另一些类(尾类)具有少量样本的情况。这种不平衡问题已经被广泛研究了很长一段时间，并提出了各种方法，如从尾类进行过采样或对尾类进行重加权。然而，当类的数量很大且不平衡很明显时，这些方法就失去了有效性。这种问题被称为长尾问题，其中有几个头部类和许多个尾部类。在本文中，我们构造了一个类层次(二叉树)，其中每个节点的左子节点和右子节点的样本数量几乎平衡。实验证明了该方法的有效性。
In real-world classification problems, it is often the case that some classes (head classes) have large numbers of samples and the other classes (tail classes) have small numbers of samples. Such imbalance problems have been widely studied for a long time, and various methods have been proposed, such as oversampling from tail classes or heavy weighting to tail classes. However, these approaches lose the effectiveness when the number of classes is very large and imbalance is remarkable. Such a problem is called a long-tailed problem where there are a few head classes and many tail classes. In this paper, we construct a class hierarchy (a binary tree) where the numbers of samples are almost balanced in left and right children of each node. Some experiments demonstrated the effectiveness of the proposed approach.