Handling implicit data

[Carl McGraw]

Hey guys.

My team has been using tensorflow-ranking for a couple years now at my company training and serving predictions at scale to millions of users a day.

Because of data privacy concerns we've been slowly changing to an anonymized world in which we track sessions and use implicit models (clicks vs impressions on converting sessions).

We currently use an MLPish BPR using a pairwise loss function from tfranking.

I am wondering what the state of the art is for tensorflow recommenders and the recommendation that you'd make about our use case. I am interested in examples, documentation or even thoughts you have.

[overflowtan(ZhenyuTan)]

Hi Carl:

 

Thanks for reaching out! Can you elaborate on what do you mean by data privacy concern and how it interacts with tfranking or tf recommenders? In particular, what are you expecting from tf recommenders SIG?

 

Cheers,

Zhenyu

--
You received this message because you are subscribed to the Google Groups "SIG Recommenders" group.
To unsubscribe from this group and stop receiving emails from it, send an email to recommenders...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/recommenders/90e5a093-6057-46c7-8d21-836185c212a4n%40tensorflow.org.
For more options, visit https://groups.google.com/a/tensorflow.org/d/optout.

[Carl McGraw]

For data privacy concerns we only collection session based metrics. So we present items to users in sessions, track session metrics (views/impressions), session metadata (country, etc), and have item data (metadata, purchase rates, etc).

Most traditional recommenders (like the one presented in the tutorial for tfrecommenders) use explicit feedback (i.e. spends, ratings, etc). I am wondering if you have recommendations about using tfrecommenders for implicit feedback only (clicks/impressions)