面向复杂时间序列的k近邻分类器

发布时间：2018-03-14 21:23

本文选题：复杂时间序列　切入点：k近邻　出处：《软件学报》2017年11期 　论文类型：期刊论文

【摘要】：基于时序对齐的k近邻分类器是时间序列分类的基准算法.在实际应用中,同类复杂时间序列经常展现出不同的全局特性.由于传统时序对齐方法平等对待实例特征并忽略其局部辨别特性,因此难以准确、高效地处理此类具有挑战性的时间序列.为了有效对齐并分类复杂时间序列,提出了一种具有辨别性的局部加权动态时间扭曲方法,用于发现同类复杂时间序列的共同点以及异类序列间的不同点.同时,通过迭代学习时间序列对齐点的正例集与负例集,获取每条复杂时间序列中每个特征的辨别性权重.在多个人工和真实数据集上的实验结果表明了基于局部加权对齐策略的k近邻分类器所具有的可解释性与有效性,并将所提出方法扩展至多变量时间序列分类问题中.
[Abstract]:The k-nearest neighbor classifier based on timing alignment is a benchmark algorithm for time series classification. The same complex time series often show different global characteristics. Because the traditional time series alignment methods treat the case features equally and ignore their local discriminations, it is difficult to be accurate. In order to effectively align and classify complex time series, a discriminative locally weighted dynamic time distortion method is proposed. It is used to discover the common points of the same complex time series and the differences between different kinds of time series. At the same time, the positive and negative sample sets of the alignment points of the time series are learned by iterative learning. The discriminative weights of each feature in each complex time series are obtained. Experimental results on multiple artificial and real data sets show the interpretability and effectiveness of the k-nearest neighbor classifier based on locally weighted alignment strategy. The proposed method is extended to multivariable time series classification.
【作者单位】：北京交通大学计算机与信息技术学院;北京交通大学交通数据分析与挖掘北京市重点实验室;
【分类号】：TP18

【相似文献】