search

jambo6 / generalised-signature-method

Code for: "A Generalised Signature Method for Time Series"
MIT License
58 stars 10 forks source link

CNN based results #2

Open frk1993 opened 3 years ago

frk1993 commented 3 years ago

Hi, thank you for the great work. I read the paper it is very informative. I would like to ask you if you can share also results based on CNN classifier instead of a random forest classifier. As I understand from the paper, you were running couple of classifier experiments on top of the signature features. So I would like to know if combining "signature features" + "a CNN classifier" improves the results.

And I would really appreciate if you can explaing your method for important feature selection.

Thank you!

jambo6 commented 3 years ago

Thank you for your kind words about the paper!

I don't believe I still have the CNN results I am afraid. However, I can say that it did not perform very well. The final model we used was a RandomForest which, given the sizes of the datasets, was a more sensible choice. All of the deep models performed relatively poorly. This could be due to poor tuning, but more likely the datasets were in general not large enough to warrant their use.

Do you have any specifics regarding important feature selection? I'm not sure what you mean.

frk1993 commented 3 years ago

Thanks for the answer.

Yes you say that you apply Lasso for the feature selection, can you explain your approach pls?