More or Less features? about patchtst

[yuqinie98]

Thanks @lesego94 for the question and @ikvision for the great discussion! There are 2 more additional points that we may want to address here: 1. One key point here we address is that Transformer is really easy to be overfitted, so we should treat the feature combination very carefully. Here we use channel-independence and see it clearly outperforms channel-mixing on those selected datasets. In future we are aiming at finding better ways to combine channels. 2. Financial dataset is different. We have discussed it in Appendix A.1.1. Hope this helps.

[ikvision]

@lesego94 I think your question relates to this paper https://openreview.net/forum?id=Fp7__phQszn.
It is more focused on tabular data, but I think it is informative for time series too.
As can be seen in Figure 4 b. deep learning system are suffering in performance when adding non informative features.
Feature selection for deep learning system is a very wide topic I don't think there is an off the shelf solution for.