基于全局和局部目标点相关性的多标签学习

Multi-Label Learning Based on Global and Local Target Point Correlation

传统单标签学习一般都是假设一个示例仅仅与一个标签相关。然而,随着科技的发展和数据形式的日益复杂,现实中的学习问题,往往一个示例同时与多个标签相关,面对这种情况,多标签学习框架被提出。现有的多标签学习大多都是从标签相关性的角度考虑,而目标点的相关性也是一个值得研究的问题。目前目标点相关性算法都是只从全局角度,或者只是从局部角度研究多标签学习问题,这对实际情况并不完全适用。本文提出一种新的同时利用全局目标点相关性和局部目标点相关性的多标签学习算法。首先运用深度神经网络学习得到全局目标点的相关性,再使用k近邻算法将全局目标点划分为局部目标点,进而运用欧式距离度量局部目标点的相关性。最后,将全局目标点相关性和局部目标点相关性结合在一起得到一个最优解。8个多标签数据集上的实验结果验证了本文所提出算法的有效性和可行性。

Traditional single-label learning generally assumes that an example is related to only one label. However, with the development of technology and the increasingly complex form of data, an example is often related to multiple labels at the same time in real life scenarios. In the face of this situation, the multi-label learning framework is proposed. Most of the existing multi-label learning is considered from the perspective of label correlation, however, the correlation of target points is also a problem worth studying. At present, the target point correlation algorithms are all from the global perspective, or only from the local perspective to study the multi-label learning problem, which is not completely applicable to the actual situation. In this paper, we propose a new multi-label learning algorithm that utilizes both global object point correlation and local object point correlation. Firstly, the correlation of global target points is obtained by deep neural network, then the global target points are divided into local target points by K-nearest neighbor algorithm, and the correlation of local target points is measured by Euclidean distance. Finally, the global target point correlation and local target point correlation are combined to get an optimal solution. Experimental results on 8 multi-label data sets validate the effectiveness and feasibility of the proposed algorithm.