item_alpha and user_alpha meaning

[bcc008]

Are Item_alpha and user_alpha L2 penalties just for item and user_features or do they also apply to the weights of the user/item latent factor as well? From the documentation it seems like they're applied to just the optional metadata/descriptive features. However, from cross-validation, I noticed that the best performing hyperparameters usually have small item and user_alphas so I am assuming they are also applied to the latent factors. In the original BPR paper, L2 regularization is similarly applied to the weights of the latent factors.

[maciejkula]

Regularization applies to parameters you learn: in this case, the user and item latent factors. Weights in the item and user features matrices are fixed inputs into the model. Feature latent factors and user/item latent factors are the same thing as far as LightFM is concerned.

…

On Wed, 29 May 2019 at 15:31, bcc008 ***@***.***> wrote: Are Item_alpha and user_alpha L2 penalties just for item and user_features or do they also apply to the weights of the user/item latent factor as well? From the documentation it seems like they're applied to just the optional descriptive features. However, from cross-validation, I noticed that the best performing hyperparameters usually have small item and user_alphas so I am assuming they are also applied to the latent factors. In the original BPR paper, L2 regularization is similarly applied to the weights of the latent factors. [image: image] <https://user-images.githubusercontent.com/33501320/58595859-df67ca00-8226-11e9-9912-c2adfb8099bc.png> — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#461?email_source=notifications&email_token=AASIEA5W3BJT5CB74PXTLSLPX375TA5CNFSM4HQW5HP2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4GWTGNVQ>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AASIEA42YHILTEYPAA5KUHLPX375TANCNFSM4HQW5HPQ> .

[SimonCW]

I’m closing this issue because it has been inactive for a long time. If you still encounter the problem, please open a new issue.

Thank you!