Ranking

Rank candidates from several retrieval sources with one watch-prediction model. Start with a baseline, then compare a neural model that combines dense features with pooled watch-history and search embeddings. Finish by applying list-level diversity rules.

The candidate-generation lesson describes several retrieval sources. Combine their candidates, then rank the resulting set by the model’s prediction of whether the user will watch each item.

Here, your goal is to rank the content based on the probability of a user watching a media given a user and a candidate media, i.e., P(watch|(User, Media)).