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

How to handle session embeddings without any song in it? #4

Open n0obcoder opened 1 year ago

n0obcoder commented 1 year ago

Great paper! I really enjoyed reading it : D I have one doubt though. It was mentioned in the paper, that for every session, you make 2 separate groups of songs, first session+ contains all consumed songs, second session- contains all the songs those were skipped. Then you take the average of all the songs in these 2 groups to create embeddings for session+ and session-, and then these embeddings are passed into the 2 LSTMs for various time-steps.

How do you handle the cases when, let's say, there are no skipped songs in a session? How do we find the embedding for such a session- which is empty?

animesh-wynk commented 5 months ago

@lucasmaystre i would really appreciate if you could please help me with this : )

lucasmaystre commented 5 months ago

This is handled around line 131 in this file. It looks like, in the case of no skipped songs, the embedding for "session-" as you call it is the all-zeros vector.

animesh-wynk commented 5 months ago

i really appreciate your help, thanks @lucasmaystre : )