- 分享主题：Time-series forecasting
- 论文标题： Deep AR: Probabilistic Forecasting with Autoregressive Recurrent Networks
- 论文链接：https://arxiv.org/pdf/1704.04110.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？

DeepAR is a supervised algorithm that uses RNN to predict unit time series. Traditional time series prediction methods, such as ARIMA and ETS, will establish a separate model for each time series, while deep is to establish a model for a group of related time series for joint training. Using DeepAR, time series can be associated with multiple groups, nonlinear problems and scale problems that are difficult to deal with statistically can be handled, and cold start time series prediction can be performed with little historical data, which only needs to be supported by similar event sequences. Moreover, deep can reduce the requirements for data to a certain extent, and it can directly supplement the default values within the model. In deep, the output obtained by calculation is not a specific predicted value, but a probability model, such as the Gaussian probability model, is obtained first, and then a predicted value is sampled from this probability model. Compared with the traditional model, DeepAR can carry out joint training, and the results predicted by probability are more robust.

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
    • 预测结果为概率分布，从概率分布中采样获得最终结果，更具鲁棒性
    • 可以对多条时间序列进行联合训练，可以解决数据量的问题
  • 缺点
    • 要求多条时间序列之间是相关的，如何衡量这种相关性是一个新的挑战
    • 基于RNN进行的预测，无法解决长序列问题

      3. 其他

      【可选】

网络结构
image.png
两个网络，第一个用来训练；第二个用于预测，在输出后通过采样算法获取具体预测值。
z{i,t}表示第i条序列，时间t时的数据
x{i,t}表示协变量