Thursday, March 5, 2020

Multi Task Learning in Homefeed Ranking

Recently I have been asked a few times how we do ranking at Pinterest homefeed, what type of models we used and what we optimize for.

Probably I shall prepare a new blog post with all the recent updates we have had in the ranking model in the past year and we have had quite a few with the introduction of MTL to an AutoML-based NN, complemented with a transfer layer on top for calibration of predictions as well as combining calibrated scores into a final utility value for ranking.

But until such a blogpost, I thought it would make sense to share a video recording from the latest Mad Science Fair at Pinterest. It is a bit dated, but if we assume most of what is discussed to be in the works in the below video as currently in production, it should provide a good overview of our current system. Hope you enjoy...