What I'm confused about is where in the pipeline the Light ranker vs the heavy ranker are used? It seems that the light ranker is used to evaluate the potential candidate tweets in the candidate sourcing stage, whereas the heavy ranker is then utilized to filter the aggregated list?
the heavy ranker is the neural network looking thing in the middle, and the light ranker is right before the heavy ranker (but after the candidate sources). The light ranker is not shown in the diagram because they are trying to deprecate it.
Hey Pasha, really enjoyed this post!
What I'm confused about is where in the pipeline the Light ranker vs the heavy ranker are used? It seems that the light ranker is used to evaluate the potential candidate tweets in the candidate sourcing stage, whereas the heavy ranker is then utilized to filter the aggregated list?
Roughly yes. In the diagram: https://cdn.cms-twdigitalassets.com/content/dam/blog-twitter/engineering/en_us/open-source/2023/twitter-recommendation-algorithm/open-algorithm.png.img.fullhd.medium.png
the heavy ranker is the neural network looking thing in the middle, and the light ranker is right before the heavy ranker (but after the candidate sources). The light ranker is not shown in the diagram because they are trying to deprecate it.