Ordinal Regression Based on the Distributional Distance Between Labels
Ordinal Regression Based on the Distributional Distance Between Labels
复制标题
DOI：
10.1109/smc52423.2021.9658911
复制
发表时间：
2021-10
期刊：
影响因子：
--
通讯作者：
Yoshiyuki Tajima;T. Hamagami
中科院分区：
文献类型：
--
作者：
Yoshiyuki Tajima;T. Hamagami
序数回归是通过考虑标签之间的序数关系来对实例进行分类的问题。它用于各种应用。现有方法在坚持保留序数关系时往往会降低准确性。因此，我们提出了一种新的深度神经网络，可以在保持高精度的同时考虑序数关系。所提出的方法由两个子网络组成。第一个子网络通过考虑排序关系对实例进行分类。使用标签之间的分布距离将该关系嵌入到子网络中。第二子网络从交叉熵误差的角度修改第一子网络的输出。通过组合两个子网络的输出来调整准确性和关系。为了评估所提出的方法，我们在两个数据集上进行了实验。结果表明，该方法在精度和平均绝对误差方面与现有方法相当或优于现有方法。
Ordinal regression is a problem in classifying instances by considering ordinal relations between labels. It is used in various applications. The existing methods tend to decrease accuracy when they adhere to the preservation of the ordinal relation. Therefore, we propose a new deep neural network that can consider ordinal relations while maintaining high accuracy. The proposed method consists of two subnetworks. The first subnetwork classifies the instances by considering the ordering relation. The relation is embedded in the subnetwork using the distributional distance between labels. The second subnetwork modifies the output of the first subnetwork from the viewpoint of cross-entropy error. The accuracy and the relation are adjusted by combining the outputs of the two subnetworks. To evaluate the proposed method, we conducted experiments on two datasets. The results demonstrate that the proposed method is comparable to or superior to the existing methods in terms of accuracy and mean absolute error.