Question about the cutoff calculation.

[mb-Ma]

https://github.com/VEWOXIC/FITS/blob/0944309af2ca04f44553153102b3adaf431e15d9/run_longExp_F.py#L113

I am curious about the "+ 10" operation. Because I do not find the explanation in the paper.

Can you explain more?

[VEWOXIC]

This is a hyper-parameter and I almost randomly designit as this. the args.seq_len // args.base_T calculates its base frequency. The +1 and +10 is just adding some redundancy when cutting the spectrum. The *H_order to calculate the real cut_freq. so actually, when you use the H_order to assign a cut_freq, it is not cut at the exact frequency.

[mb-Ma]

Thank you for your quick reply. :-) It addresses my concerns. I tried some adaptive COF values, like top K amplitude-frequency. Badly, these all hurt the model. Do you have any insights into the COF design?

[VEWOXIC]

Yeah, we also tried the top-K selection and the result was not amazing. I think it might because the frequency components with larger amp are often some low freq components. Sometimes they are not a specific periodic pattern, but comes from the decomposition of some low frequency trend (e.g. try fft a sine wave segment with 1.5 period will result in some other frequency components). These do not contribute to our forecasting. However, the harmonic is the key to model the major periodicity pattern. Thus, we use harmonic to decide the COF.

Hope this can help! And looking forward to your insights!

[mb-Ma]

It is. From the visualization of spectrums, we can observe that most high amplitude components are located in low frequency. Another point is the top K amplitude selection strategy would break the continuity of the spectrum.

But as yet I cannot figure out one way that is better than fixed COF. It is hard to obtain an adaptive COF for each instance. There is a trade-off where a more optimal COF can be selected by power spectrum or information entropy but with more computation overhead.

[VEWOXIC]

So, you mean only model the top K frequency? Seems not a good idea hhh.

We have never considered the adaptive COF when designing the FITS and it may not be suitable for this. But is definitely a good idea to design a frequency-domain model that can adaptively attend to frequency components that is more important to the instance.

Further, we do not think further simplify the model is a meaningful topic since we design FITS is just to show there is not much information to learn with time series data alone.

[VEWOXIC]

if there is no further question, I will close this issue. Please feel free to reopen it.

[mb-Ma]

Sorry for my late reply. Thanks for your consecutive feedback.

Good work. 👍