what is calculation time

[Sandy4321]

in the paper you did not write about calculation time,
by the way, did you tried ~1000 features data?
where is the code for other methods mentioned in paper/
thanks

[aflapan]

Hi Sandy,

Sorry about the late response time. We did not try any data sets with ~1000 features. I think that would be a great simulation to try out. Also, we have not provided the code for the other methods in our simulation study, but I will talk with my advisor about what to do with that code. However, we used only publicly available R packages as cited in the paper.

You are correct in that we did not write about computation time. That will depend on multiple things:

1. Convergence threshold and maximum number of iterations allowed.
2. Any parallelizing you may do
3. The maximum number of iterations allowed in the coordinate descent weight update step.

Running one (out of the 100) of the simulations from Simulation Model 1 with all default threshold and maximum iteration parameters took 17.43379 mins on my personal computer. This is with using the automatic parameter selection methods. With the parameters pre-specified, the same simulation took 7.03811 secs. To run the entirety of simulation model 1, we parallelized over 50 nodes, and the computation took around half an hour.

Unfortunately, the algorithm is computationally expensive, but I fear that may just be the nature of the problem itself. We have implemented C++ code to speed up the worst parts in development.

If you have any more questions, please let me know!

-Alex

[Sandy4321]

sure
may you recommend some easy idea level about gaussian kernel for LDA
"Sorry about the late response time. We did not try any data sets with ~1000 features. I think that would be a great simulation to try out."
great idea , do you want I help you find this kind of data: many features and sparse data?

[aflapan]

I think that would be a great simulation. Have you tried the package on such a data set? If not, please let me know what a good data set might be.

[Sandy4321]

super lets do it
data can be taken from
http://srome.github.io/Leveraging-Factorization-Machines-for-Sparse-Data-and-Supervised-Visualization/
this blog also gives you base line based on FactorizationMachines
https://github.com/srome/PyFactorizationMachines
I would be happy to help you with python code
lets communicate by email
my email
sndr.stpnv@gmail.com
what is your email?

[aflapan]

My email is aflapan@gmail.com. I’m currently at a conference, and do I haven’t been able to look much into the simulation yet. However, I’ll have more time in a few days!

[aflapan]

Hi Sandy, my email is aflapan@gmail.com. I’m currently at a conference right now, but I can look more into simulation when I get back.

…

On Friday, April 12, 2019, Sandy4321 ***@***.***> wrote: super lets do it data can be taken from http://srome.github.io/Leveraging-Factorization-Machines- for-Sparse-Data-and-Supervised-Visualization/ this blog also gives you base line based on FactorizationMachines https://github.com/srome/PyFactorizationMachines I would be happy to help you with python code lets communicate by email my email ***@***.*** what is your email? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AB4MYWIEBQMDP3LAMVCSRILPQCF6DANCNFSM4GXPJ25Q> .