search

spotify-research / cosernn

Code for the paper "Contextual and Sequential User Embeddings for Large-Scale Music Recommendation".
Apache License 2.0
33 stars 8 forks source link

some question #2

Open arita37 opened 2 years ago

arita37 commented 2 years ago

Hello,

Thanks for the paper. You mentioned that : context vector at the start of session is more "efficient" than average vector preference.

Just wondering if you have tried to mix both type of preferences ? Long term preference context Short term preference context (ie at time of session start)

And, if you have done some evaluation on average vector preferences.

lucasmaystre commented 2 years ago

Hi @arita37 thanks for your interest and apologies for the delay.

It is likely that the sequential RNN model that we use encodes both long-term and short-term preferences. I believe this achieves what you mention, in a way.

Regarding evaluation of average vector preferences: if I understand correctly, the baseline we call "average, any context" is exactly what you mention.

arita37 commented 2 years ago

Thanks.

Just wondering if sequentil RNN really scales to 50 million of items and 50 million of users.

Isn’t just average user vector more scalable ? (But less precise)

On Jan 11, 2022, at 18:33, Lucas Maystre @.***> wrote:

 Hi @arita37 thanks for your interest and apologies for the delay.

It is likely that the sequential RNN model that we use encodes both long-term and short-term preferences. I believe this achieves what you mention, in a way.

Regarding evaluation of average vector preferences: if I understand correctly, the baseline we call "average, any context" is exactly what you mention.

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you were mentioned.