search

xinzzzhou / ScalableTransformer4HighDimensionMTSF

10 stars 2 forks source link

Why the correlation is learned from GNN rather than using pearson correlation coefficient matrix directly? #1

Open erikalien5595 opened 3 months ago

erikalien5595 commented 3 months ago

Hi, I have two questions that are troubling me, could you please help me answer them? Thank you very much. The questions are:

1)In your papers and code, I see that you use the complete sequence to learn the correlation matrix, rather than learning the correlation matrix within each look-back window. Why is that? 2) Why do you use DeepGraph to learn the correlation instead of directly using the Pearson correlation coefficient matrix?

xinzzzhou commented 2 months ago

Hi, thanks for your interest and thoughtful questions.

Regarding question 1), we prioritized computational efficiency for handling high-dimensional data. We selected the simplest method to learn correlations. However, you're right, modeling the relationships among windows could yield more accurate results. We look forward to seeing your improved design! As for question 2), DeepGraph supports parallelization, which can significantly reduce computation time. Please refer to Figure 6 for more details.

wtt6668888 commented 1 month ago

Hi, thanks for your interest and thoughtful questions.

Regarding question 1), we prioritized computational efficiency for handling high-dimensional data. We selected the simplest method to learn correlations. However, you're right, modeling the relationships among windows could yield more accurate results. We look forward to seeing your improved design! As for question 2), DeepGraph supports parallelization, which can significantly reduce computation time. Please refer to Figure 6 for more details.

Hello, I have two questions that are bothering me. I’m still not quite clear on how you handle high-dimensional data for multivariate time series forecasting. Additionally, how is the correlation of high-dimensional variables used for forecasting in high-dimensional data? I look forward to your response!