Architectural Components

Separate catalog retrieval from final ranking. Candidate generators collect promising titles from several sources; a common ranker then compares those titles for the viewer. User interactions supply training data for both stages.

It makes sense to consider generating the best recommendation from a large corpus of movies, as a multi-stage ranking problem. Let’s see why.

Scoring every title with an expensive model can consume too much of the request budget. A two-stage design spends less work collecting candidates, then applies more detailed scoring to the smaller set. Whether this split is needed depends on catalog size and measured serving cost.

Therefore, we split the recommendation task into two stages.

• Stage 1: Candidate generation

• Stage 2: Ranking of generated candidates

Stage 1 uses a simpler mechanism to sift through the entire corpus for possible recommendations. Stage 2 uses complex strategies only on the candidates given by stage 1 to come up with personalized recommendations.