0
$\begingroup$

I've been working on a RecSys model recently (using HRNNs), and when thinking about the features used for users and itens, I thought that many of them ended up being biased by the old system recommendation model, biasing the propensity of an item to appear based on it, and not what the user really wants. This feedback loop in training new models could be mitigated by weighting the features on the current propensity of being recommended to users? With Inverse Propensity Scoring or something similar?

Does someone know anything on the literature about it?

$\endgroup$
0
$\begingroup$

First note that you can get very good results even without solving (the very complex problem) of feedback loops. Also, weighting your loss with inverse propensity is usually not enough and it can cause stability problems. When using IPS you will also need to deal with IPS overfitting and controlling the variance (i.e. regulaizer that make your new policy don't get very far away from the original policy).

  Regarding literature: This is a good explanation about IPS: 

https://arxiv.org/pdf/1602.05352.pdf

This paper from criteo explain about IPS and how to address the two problem i mentioned (variance and overfitting)

https://arxiv.org/pdf/1808.00232.pdf

And a spotify paper that uses IPS in their recommendations algorithm

https://static1.squarespace.com/static/5ae0d0b48ab7227d232c2bea/t/5ba849e3c83025fa56814f45/1537755637453/BartRecSys.pdf

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.