No really an issue but more like a chat? Or message board.

[trbKnl]

Great Rcpp stuff Davide, I am going to use it 👍 :).
and I think while I am at it, I may also, program the CV in cpp, because it is too slow as is in case of big data.

[davidevdt]

Yoo Niek! Thx for the feedback. Sure let me know how your CV works! I might do the same for my functions.

[trbKnl]

Hi Davide, I coppied your Rcpp function!! changed it a little bit here and there, but I kept almost all your major changes, they were great :).
The things I have different that are worth checking out:

• I calculate the sign differently, I found a way to do that in one line, saves a little bit of code. Maybe interesting to look at.
• and the update of P, you index the columns of U at 1, a mistake maybe?

[davidevdt]

hey niek! mm do you mean the indices of U, that i put that goes from 0 to Q-1? I think it's right like that..why do you think it should go from 1?
Also, do you think it's a good idea to rescale the columns of W so that they have euclidean norm 1 at the end of the algorithm (or of each iteration)?

[trbKnl]

Huh, I think a 0 changed to a 1 when looking at your code in a txt file on my machine. It should be 0 and Q-1 as you also have it! My bad.

Mmm in end maybe, but I don't see a benefit? But not after each iteration, that would mess up the algorithm I feel. After-all you are supposed to converge to a stable W state, I think rescaling could prevent that.

[davidevdt]

Ah ok no worries!
Yes in the paper by Zou , Hastie and Tibshirani (2006) they say to normalize the columns in the end of the iterations. I don't think it has a benefit in terms of computations, but it could be a benefit when comparing matrices (e.g. by tucker congruence). This would imply also re-scale the columns of W when generating the data. Once again, I don't know whether it makes sense, but it seems the proper way to do it to me