基于区间的时间序列分类算法的研究  

Research on interval time series classification algorithm

作　　者：李建平[1] 王兴伟 马连博 黄敏[1] Li Jianping;Wang Xingwei;Ma Lianbo;Huang Min(Northeastern University,LiaoningShenyang 110169)

机构地区：[1]东北大学

出　　处：《网络空间安全》2019年第8期84-92,101,共10页Cyberspace Security

基　　金：国家自然科学基金资助项目(项目编号：61872073,61572123);辽宁省高校创新团队支持计划资助项目(项目编号：LT2016007)

摘　　要：时间序列分类是时间序列数据挖掘的一个分支,针对传统时间序列分类模型存在的失真的问题,文章提出了基于区间权值的集成算法EAIW(Ensemble Algorithm of Interval Weights)。首先利用区间权值计算方法,为时间序列的不同区间赋予不同的权值,对计算做了并行化处理,以解决子序列特征不明显的问题。进而确定集成分类器的基分类器,以保证集成分类器的性能。然后,在训练集上训练集成分类器,并行化改进集成分类器训练、分类较为耗时的部分。文章将提出的算法在时间序列分类数据库上进行了实验,结果表明提出的算法比基准算法最优正确率数目高25%,并且算法在并行化之后具备可伸缩性。Time series classification is a branch of time series data mining.Targeting on the distortion of traditional time series classification model,the Ensemble Algorithm of Interval Weights is proposed.Firstly,the interval weight calculation,which assigns different weights to different intervals of the time series,solves the problem that the sub-sequence features are indistinguishable,and calculation process is parallelized.After that,according to the requirements of algorithm,to ensure the performance of the ensemble classifier,the base classifier is determined.Then,the ensemble classifier is trained by the training set,and the parallelization improvement is made for the time-consuming part of the ensemble classifier training and classification.Finally,the proposed algorithm is tested by the time series classification database and compared with the benchmark algorithm.The results show that the algorithm is 25%higher that of the benchmark algorithm in the number of optimal correctness,and also prove that the algorithm is scalable after parallelization.

关 键 词：时间序列 分类 数据挖掘 CUDA 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]