How can I set the parameter "low_rank" in different application scenarios?

[DingSizhe]

First of all, your open source work is very beautiful! Let me have a good understanding of the main content of your paper.

Here I would like to discuss an initial parameter setting pointed out in the code or the paper. In the code, the value of initial parameter low_rank needs to be specified when executing BGCP notebook. How can the value be defined according to the given different time series? Or is the definition of this value completely random? I'm troubled by that.

If convenient, please reply. I will be very grateful! Thank you again for your open source work.

[xinychen]

Thanks for this great feedback! You can consider to set a series of low ranks and validate the imputation performance in different missing data scenarios. I do not have any better solution to the setting of low rank. If I have a new idea for this in the future, I will update the Jupyter notebooks accordingly.

[DingSizhe]

Thank you for your reply. At present, I am thinking about similar problems.

As you know, a good model needs strict theoretical support, which you have done very well. At the same time, a simple and easy-to-use solution will make the complex model can be used without being completely mastered by non-professional users (for example, some practice researchers), which will greatly enhance the practical significance of the model itself. So is the parameter control of the model. Perheps, there is much more work to be done; however, the tool is really noteworthy.

In fact, my confusion is not limited to the replied question. For example, in the paper "Bayesian Temporal Factorization for Multidimensional Time Series Prediction", I noticed that the time lags parameter always appears suddenly. In single-step prediction, the parameter is set to \mathcal{L}={1, 2, T_0}, while in multi-step prediction, the parameter is set to \mathcal{L}={1, 2, 3, T_0, T_0+1, T_0+2, 7T_0, 7T_0+1, 7T_0+2}. This may be a priori knowledge, but there seems to be no further explanation in the paper. I know that this will not cause any problems for the model itself, but perhaps similar writing will cause difficulties in application.

Thank you again for your work!

[xinychen]

Thanks for this great feedback on our work! It could help our current and future work a lot. In the transdim project, I also want to find some simple but effective solutions to the real-world domain problems. I am still trying to do it. If you have any other questions, please let me know and I would appreciate your feedback. Thank you!