基于自注意力移动平均线的时间序列预测  

Self-attentive moving average for time series prediction

作　　者：苏雅茜 崔超然[1] 曲浩 Su Yaxi;Cui Chaoran;Qu Hao(School of Computer Science and Technology,Shandong University of Finance and Economics,Ji'nan,250014,China;School of Software,ShandongUniversity,Ji'nan,250101,China)

机构地区：[1]山东财经大学计算机科学与技术学院,济南250014 [2]山东大学软件学院,济南250101

出　　处：《南京大学学报（自然科学版）》2022年第4期649-657,共9页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(62077033,61701281);山东省自然科学基金(ZR2020KF015)。

摘　　要：时间序列预测有广阔的应用前景,吸引了越来越多的研究人员对其进行深入的研究.移动平均线是最常用的技术指标之一,它可以对过去一段时间内时间序列的整体变化规律进行概括,常常用来进行时间序列趋势预测.然而,传统的移动平均线指标是通过赋予时间序列数据相等或预定义的权重计算得到的,忽略了不同时间点重要性的细微差别;另外,对于不同的时间序列数据采用相同的权重,忽视了不同时间序列内在特性的差异.为了解决上述问题,提出一个自适应的自注意力移动平均线(Self-Attentive Moving Average,SAMA)指标.利用循环神经网络对时间序列的输入信号进行编码后,引入自注意力机制来自适应地确定不同时间点上数据的权重以计算移动平均值.此外,还使用多个自注意头对不同尺度的SAMA指标进行建模,最后将它们组合起来进行时间序列预测.在两个真实数据集上的大量实验证明了该方法的有效性,数据集和代码已在https://github. com/YY-Susan/SAMA上发布.Time series prediction has a broad application prospect and attracts more and more researchers to conduct in-depth research on it. As one of the most popular technical indicators,moving average summarizes the overall changing patterns over a past period and is frequently used to predict the future trend of time series. However,traditional moving average indicators are calculated by averaging the time series data with equal or predefined weights,and ignore the subtle difference in the importance of different time steps. Moreover,unchanged data weights are applied across different time series,regardless of the differences in their inherent characteristics. In order to solve the above problems,this paper proposes a adaptive moving average indicator called SAMA(Self-Attentive Moving Average). After encoding the input signals of time series based on recurrent neural networks,we introduce the self-attention mechanism to adaptively determine the data weights at different time steps for calculating the moving average. Furthermore,we use multiple self-attention heads to model SAMA indicators of different scales,and finally combine them for time series prediction. Extensive experiments on two real-world datasets demonstrate the effectiveness of our approach. The data and codes of our work have been released at https://github.com/YY-Susan/SAMA.

关 键 词：时间序列预测 自注意力机制 移动平均线 多尺度指标融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]