应用于不平衡多分类问题的损失平衡函数

PDF下载HTML

Title:

Application of the loss balance function to the imbalanced multi-classification problems

作者:

黄庆康, 宋恺涛, 陆建峰

南京理工大学 计算机科学与工程学院, 江苏 南京 210094

Author(s):

HUANG Qingkang, SONG Kaitao, LU Jianfeng

School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China

关键词:

不平衡学习; 不平衡数据分类; 多分类不平衡; 损失平衡; 不平衡数据分类算法; 不平衡数据集; F₁调和平均; 卷积神经网络; 深度学习

Keywords:

imbalanced learning; imbalanced data classification; imbalanced multi-classification; loss balance; classification algorithm for imbalanced data; imbalanced dataset; F₁ measure; convolutional neural networks; deep learning

分类号:

TP391

DOI:

10.11992/tis.201808004

摘要:

传统分类算法一般要求数据集类别分布平衡，然而在实际情况中往往面临的是不平衡的类别分布。目前存在的数据层面和模型层面算法试图从不同角度解决该问题，但面临着参数选择以及重复采样产生的额外计算等问题。针对此问题，提出了一种在小批量内样本损失自适应均衡化的方法。该算法采用了一种动态学习损失函数的方式，根据小批量内样本标签信息调整各样本损失权重，从而实现在小批量内各类别样本总损失的平衡性。通过在caltech101和ILSVRC2014数据集上的实验表明，该算法能够有效地减少计算成本并提高分类精度，且一定程度上避免了过采样方法所带来的模型过拟合风险。

Abstract:

The traditional classification algorithms generally require a balanced distribution of various categories in datasets. However, the traditional classification algorithms often encounter an imbalanced class distribution in real life. The existing data- and classifier-level methods that attempt to solve this problem based on different perspectives exhibit some disadvantages, including the selection of parameters that have to be handled carefully and additional computing power because of repeated sampling. To solve these disadvantages, a method that can adaptively maintain the loss balance of examples in a mini-batch is proposed. This algorithm uses a dynamic loss-learnt function to adjust the loss ratio of each sample based on the information present in the label in every mini-batch, thereby achieving a balanced total loss for each class. The experiments conducted using the caltech101 and ILSVRC2014 datasets denote that this algorithm can effectively reduce the computational cost, improve the classification accuracy, and avoid the overfitting risk of the model that can be attributed to the oversampling method.