Launch a Ranking Model for the Home Feed

The home feed ranks posts by recency and like count. Session length has been flat for four quarters, and only 12% of sessions scroll past the first ten posts.

A gradient boosted model predicts the probability that a user clicks each candidate post, and the feed sorts by that score.

Training features are computed in the nightly Spark pipeline. At serving time, the ranking service recomputes the same features in Java from the written definitions.

Training data covers every active user, including users who turned off personalized recommendations in their privacy settings.

The model retrains every night on clicks from the feed it ranked the day before.

The ranking service scores up to 500 candidates per request. If it doesn't respond, the feed waits for it and retries once before rendering.

Offline AUC improved from 0.71 to 0.78, so the model launches to all users in one step, and we compare session length for the month before and the month after.

Success is a 5% increase in average session length across all users.