Architectural Components

Follow a feed request from candidate selection to ranking, then trace the resulting user actions back into training data. The architecture supports either one engagement model or separate predictions combined by the product’s value model.

Architecture

The diagram separates three responsibilities: choose candidate tweets, score them, and turn subsequent interactions into training examples.

Let’s briefly look at each component here. Further explanation will be provided in the following lessons.

Tweet selection

This component performs the first step in generating a user’s feed, i.e., it fetches a pool of Tweets from the user’s network (the followees), since their last login. This pool of Tweets is then forwarded to the ranker component.

Training data generation