search

lss-1138 / SegRNN

The official repository of the SegRNN paper: "Segment Recurrent Neural Network for Long-Term Time Series Forecasting." This work is developed by the Lab of Professor Weiwei Lin (linww@scut.edu.cn), South China University of Technology; Peng Cheng Laboratory.
Apache License 2.0
87 stars 15 forks source link

question about model's channel_id #2

Closed YoonKiwooong closed 4 months ago

YoonKiwooong commented 6 months ago

I am studying your great model, but I am having hard time understanding the architecture of the model. In the model's code, we can give model's channel_id to True or False, but when should we give True to channel_id?

From the paper Table4, False was given to channel when the data set was 'Traffic'. Is it related to the fact whether data is univariate or multivariate time series forcasting? If it is right, can I also give True to channel_id when doing univariate timeseries forcasting?

lss-1138 commented 6 months ago

Thank you for your interest. When performing univariate time series forecasting, the channel_id should be set to False since there is only one channel and there is no need to distinguish between different channels. However, when performing multivariate time series forecasting, you can choose to set it to either False or True.

For multivariate time series forecasting, if the channel_id is set to False, it represents an independent channel scheme. If it is set to True, it introduces an additional channel identifier to distinguish between different channels on the basis of channel independence, a technique first proposed in SITD(https://github.com/zezhishao/STID).

We set the channel_id to False for the Traffic dataset based on the ablation results in Table 6. This suggests that different channels in Traffic may have relatively consistent patterns, and there is no need for further channel encoding to distinguish between them.

Finally, whether to model the relationships between multiple variables and how to model them during multivariate time series forecasting is still a challenging issue that perplexes the community. Currently, most papers indicate that the independent channel scheme is a more powerful choice, which may contradict the intuition of performing multivariate forecasting. If you are interested, you can further explore and research this issue.

If you have any further questions, please feel free to ask. We're always happy to help.

YoonKiwooong commented 6 months ago

Oh Thank you for your kind response. I will study further and come back to ask if there is anything else I don't know.