- 分享主题：Time Series Forecasting
- 论文标题：N-BEATS: Neural basis expansion analysis for interpretable time series forecasting.
- 论文链接：https://arxiv.org/pdf/1905.10437.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？

N-BEATS originates from research by Boris Oreshkin and its co-authors at unfortunately short-lived ElementAI. N-BEATS is an interesting step in applying deep learning to time series because it crafts an architecture dedicated to time series. The previous approach consists in translating sequences (sequences to sequence). Timepoints are given to the network one after the other, and the network updates some internal memory to update the internal representation of the state of the system. Then the output is computed using internal state representation and current output. N-BEATS is a deep neural architecture based on backward and forward residual links and a very deep stack of fully-connected layers. It consists of several blocks connected in a residual way: the first block tries to model past window (backcast) and future (forecast) the best it can, then the second block tries to model only the residual error of the past reconstruction done by the previous block (and also updates the forecast based only on this error) and so on. Such residual architecture allows to stack many blocks without risk of gradient vanishing and also has the advantages of boosting/ensembling technique: the forecast is the sum of predictions of several blocks, where the first block catches the main trends, the second specializes on smaller errors and so on. The architecture has several desirable properties, being interpretable, applicable without modification to a wide array of target domains, and fast to train.

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.训练快：所有的操作都可以在GPU上并行处理，训练比使用RNN快的多
2.好部署：N-BEATS的块可以删减增加，可以根据问题动态调整，更好部署
3.可解释性：提出了具有不同基底的模型，不同基底刻画不同模式，具有可解释性
缺点：
1.针对单变量时间序列提出的
2.其可解释性的特性是基于对基底的假设，而当融合专家经验时，其可解释性才更好
3.本质实际让神经网络做一个分解工作，而且该分解是黑盒的

3. 其他

【可选】

image.png

1.basic block

每个block分为两条路径，一个用于backcast，一个用于forecast。backcast的输出和下一个block的原始输入相减作为新的输入，forecast的输出用于组成最终stack的输出。输入进入block后，先经过一个四层的全链接，然后经过线性投影获得后向和前向系数。
image.png
之后对得到的系数再通过basis layer输出结果
image.png

2.block 之间

不同的block之间通过残差模块连接，多个block组成stack。stack的输出为每一个block输出的和。
image.png

3.可解释性

在可解释性层面，作者提出两种结构，一种generic architecture不依赖于特定经验知识，该变化可以看做在时间的波形重建，由于没有额外的限制在 Vlf上 ，导致生成的y**l解释性比较差。
image.png
另一种则通过引入专家经验来满足可解释性要求，即通过引入时间序列的趋势性(trend)和季节性(seasonality)来实现可解释性。具体方法为人为设定与θ相乘的V**矩阵使得该矩阵满足趋势性和季节性的要求。这方面带来的启示是，可以通过设计相乘矩阵的数值形式来实现专家经验的引入。
image.png
image.png
所以，根据时间序列的基本知识，我们可以知道，去除趋势和季节性后，只剩下短期相关性，所以在N-BEATS中通常堆叠两个或者三个stack就会有比较好的效果，因为其本质是借助神经网络对时间序列进行了分解。

提到了结合元学习