发布时间:2023-04-19 文章分类:电脑基础 投稿人:樱花 字号: 默认 | | 超大 打印

ICLR 2023

1 intro

 2 TImesNet

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

2.1 从一维时间序列到二维张量的转化

可以用如下三个公式表示

  1. 论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS
  2. 论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS
  3. 论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS
  • 式1
    • FFT(.)表示将一维时间序列进行快速傅里叶变化(个人理解就是将N长的时域信号转换为N/2长度的谱域信号)
    • Amp就是算各个时间序列转换后,在各个频率上的振幅
    • Avg就是对不同时间序列的振幅求平均
    • —>A是不同频率的平均振幅
  • 式2
    • 取平均振幅最大的k个频率
  • 式3
    • 得到对应的k个周期
  • 为了简便考虑,论文将上述三个式子简化成:
  • 论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 基于论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS,可以将一维时间序列论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS转化成一组二维张量论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS,其中第i个二维张量论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS的计算方式为:论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

2.2 TimesBlock

2.2.1 捕获2D时间维度变化

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

2.2.2 自适应加和

3 实验

3.1 数据集

论文比较了长期时间序列预测、短期时间序列预测、时间序列补全、时间序列分类、异常检测五个问题

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS 

 3.2 主要结果

3.2.1 短期时间序列预测

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

3.2.2 长期时间序列预测 

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 3.2.3 时间序列补全

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

3.2.4 时间序列分类 

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

3.2.5 时间序列异常检测

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 3.3 2D时间序列张量可视化

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 3.4 超参数稳定性

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

3.5 消融实验

3.5.1 不用Inception,使用别的卷积架构

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 3.5.2 自适应加和部分

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 3.6 regression效果可视化

3.6.1 imputation

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 3.6.2  prediction

论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS

 论文笔记:TIMESNET: TEMPORAL 2D-VARIATION MODELINGFOR GENERAL TIME SERIES ANALYSIS