混合范数趋势滤波时间序列分类算法研究

Research on Time Series Classification Algorithm with Hybrid-norm Trend Filtering

针对时间序列分类算法易受噪声影响的问题,传统的L1趋势滤波和L2趋势滤波都是有效的时间序列平滑方法,然而两者的趋势模型固定,仅适用于特定类型的数据.另外常用的欧氏距离度量在处理时间序列时没有考虑时序属性.因此,本文提出一种基于混合范数趋势滤波的时间序列分类算法.首先,在趋势滤波目标函数中引入混合范数模型作为正则项,获取更适于分类的时间序列趋势估计;然后,采用动态时间弯曲(DTW)距离度量消除时序干扰,训练基于DTW的k近邻分类器实现时间序列分类;最后,使用贝叶斯优化算法自适应地寻找最优的超参数组合,以降低问题求解的时间复杂度.在UCR时间序列数据库的40个数据集上进行实验,结果表明本文算法在时间序列分类任务中的性能更佳、分类错误率更低.

Aiming at the problem that time series classification algorithm is easily affected by noise,traditional L1 trend filtering and L2 trend filtering are effective methods for time series smoothing,whereas their trend models are fixed and only applies to certain kind of data.Furthermore,commonlyused Euclid distance metric ignores temporal attributes when dealing with time series data.This work proposes a time series classification algorithm based on hybrid norm trend filtering.First,introducing mixed norm model into object function of trend filtering as regular term,acquiring time series trend estimations that are more suitable for classification.Then,utilizing Dynamic Time Warping distance metric to eliminate temporal disturbance,the K-Nearest Neighbor classifier based on DTW were trained to implement time series classification.Finally,finding optimal hyper-parameters adaptively using Bayesian optimizing algorithm,reducing the time complexity of problem solving.Experiments were carried out on 40 data sets of UCR Time Series Classification Archive,the results show that proposed algorithm performs better in time series classification tasks,with lowererror rates.