基于共享矢量链的多任务概念漂移分类方法

Multi-task concept drift classification method based on shared vector chain

增强型时间自适应支持向量机在针对单一概念漂移问题时展现出了良好效果,但是无法协同求解多个概念漂移问题.然而,在很多应用场景中,有时会包含数个具有内在相关性的非静态数据集,它们各自的分类模型应充分考虑这种关联.为了反映出各概念漂移分类模型之间的相关部分,提出共享矢量链的概念,并开发面向多任务概念漂移问题的共享矢量链支持向量机(SVC-SVM).在模拟数据集及气体传感器阵列漂移数据集上的实验结果显示,协同求解多个具有相关性的概念漂移问题能够有效提升各自的泛化能力.

The improved time adaptive supported vector machine performs well for a single concept drift problem in theoretical analysis and experimental studies, but cannot solve multiple concept drift problem. However, several parallel non-stationary datasets with inherent correlation are sometimes involved in many practical scenarios, whose respective classification models should take such cohesions into account. In this paper, the concept of shared supported vector chain is proposed to reflect the common vectors in each classification model for each nonstationary dataset and accordingly,a multi-task concept drift classification approach, called shared vector chain supported vector machines（SVC-SVM）, is developed. Experimental results on synthetic datasets and gas sensor array drift dataset show the effectiveness of the proposed algorithm when solving several correlative concept drift problems together.