- 分享主题：TS —MF
- 论文标题：Think Globally, Act Locally: A Deep Neural Network Approach to High-Dimensional Time Series Forecasting
- 论文链接：https://arxiv.org/pdf/1905.03806.pdf
- 分享人：唐共勇

1. Summary

【必写】，推荐使用 grammarly 检查语法问题，尽量参考论文 introduction 的写作方式。需要写出

  1. 这篇文章解决了什么问题？
  2. 作者使用了什么方法（不用太细节）来解决了这个问题？
  3. 你觉得你需要继续去研究哪些概念才会加深你对这篇文章的理解？

This paper points out that in the multi-dimensional time series prediction, considering the historical data of relevant time series can improve the accuracy, and there are problems of different scales between different time series, so it is difficult to initialize the network. In this paper, Leveledinit is used so that the training process does not need Apriori normalization to solve the problem of initialization. The existing models only consider the global feature (MF) or local feature (TCN) of time series, and can not better combine the two features of multidimensional time series to obtain higher prediction accuracy. In this paper, TCN-MF is used to predict the multi-dimensional time series to obtain the global characteristics of the time series. Then, the obtained prediction results considering the global characteristics and the combination of the original sequence covariables are input into the TCN network as new covariables for prediction again, that is, hybrid model DeepGLO, which considers the global and local characteristics at the same time.

2. 你对于论文的思考

需要写出你自己对于论文的思考，例如优缺点，你的takeaways

优点：
1.多维多重时间序列的全局信息与局部信息结合使用
2.解决了不同时间序列具有不同尺度的问题
缺点：
1.多重时间序列之间需要相似，度量相似性是一个复杂的工程
2.不同尺度的时间序列虽然糅合到一起，但效果有待考证

3. 其他

【可选】

LeveledInit

image.png

TCN-MF

image.png
整体逻辑
image.png