BPRMF optimization (hyperparameter tuning)

[Djoels]

What would be the preferred way of optimizing (hyperparameter-tuning) the BPRMF method?

I would think num_factors, and then all of the 4 regularization parameters (which are in a sense confusing to me, maybe I need to re-read the paper), and then the learning rate.

Are there any default values that are supposed to work well for some of these parameters (and closely resemble the original BPR setup)?

[Zeno Gantner]

Hi, 

I would first set a num_factors (because maybe you have performance considerations), then determine a good learn_rate (by watching convergence on the training set), and then look at the regularization parameters, first using the same values, then decoupling users and items, then distinguishing positive and negative items, etc. 

If I remember correctly there recently was a post about this on the list...

Cheers, 

  Z. 

--
You received this message because you are subscribed to the Google Groups "MyMediaLite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mymedialite+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Djoels]

I'm currently wondering whether parameters like learning rate and regularization parameters are meant to be between 0 and 1? 

Or is this choice rather arbitrary?

Regards,

Julien

Op vrijdag 24 februari 2017 18:56:45 UTC+1 schreef Zeno Gantner:

  Z. 

To unsubscribe from this group and stop receiving emails from it, send an email to mymedialite...@googlegroups.com.

[Zeno Gantner]

Below 0 does not make sense for neither.

And a learning rate of 0 means no updates.

I am not aware of a hard upper limit. I seldom have seen values above 1, though.

If you search exponentially  (..., 0.001, 0.01, 0.1, 1, 10, ...) you will quickly find the rough area of good values.

Z.

To unsubscribe from this group and stop receiving emails from it, send an email to mymedialite+unsubscribe@googlegroups.com.

[Djoels]

Thank you for the help. I've already submitted my master's thesis, for which I modified MyMediaLite a little. I want to thank you for your guidance wrt some of the matters discussed in this group.

The document can be found here, if you're interested: https://www.dropbox.com/s/havll6rn8pzvpzv/20170623_recsys_krook_thesis-final.pdf?dl=0

Op donderdag 3 augustus 2017 08:55:50 UTC+2 schreef Zeno Gantner: