降低数据稀疏性的多维时序序列时间戳对齐方法

Timestamp alignment method of multi-dimensional time series sequence to reduce data sparsity

多维时序序列是指一组按照时间发生先后顺序进行排列的数据点序列,广泛存在于天文、医疗、交通等领域。囿于收集技术较差,或是序列的物理性质所致,时序序列记录中往往存在较多的缺失值和大量的不规则采样,使得时序序列的稀疏性大大增加。最终导致许多深度学习的时序序列分类算法等无法正常工作,出现算法效果差、算法训练时间过长等问题。面对这些问题,目前常用的做法是简单删减或是利用专家知识做重采样,前者会导致数据规模变小,后者使得算法成本增加。本文利用时序序列的时间戳数据构建了一种半自动化的预处理方法。在公共数据集MIMIC-Ⅲ、Physionet和肾移植数据集上的实验表明本文提出的方法在基本不损失算法效果的同时,能够有效降低数据稀疏规模,并且平均能够节约42.1%的算法训练时间。

Multi-dimensional time series refers to a series of data points arranged in the order of time,which is widely used in astronomy,medical treatment,transportation and other fields.Due to poor collection technology or the physical properties of the sequence,there are often more missing values and a large number of irregular sampling in the time series sequence record,which greatly increases the sparsity of the time series sequence.In the end,many deep learning time series sequence classification algorithms cannot work normally,and problems such as poor algorithm effect and long algorithm-training time occur.In the face of these problems,the current common method is to simply delete or use expert knowledge to do resampling.The former will result in a smaller data size,and the latter will increase the cost of the algorithm.In this paper,a semi-automatic preprocessing method is constructed using the timestamp data of the time series sequence.Experiments on the public data set MIMIC-Ⅲ,Physionet and kidney transplantation data set show that the method proposed in this paper can effectively reduce the sparse scale of the data while basically not losing the effect of the algorithm,and can save the algorithm training time on average by 42.1%.