Social feed ranking

How this question breaks down

1. 1SourcesCore18 minA feed ranker can only order what it is given. Pull a few hundred posts per request from several sources (friends, groups, followed pages, what friends engaged with, interest embeddings, fresh local posts), trim each with a cheap model, and size each source's share by how many future engagements its extra slots actually bring in.
2. 2EngagementCore19 minFor each of 600 candidates, predict what this viewer will do: like, comment, share, skip or hide. One multi-head model trained on a day-fresh impression log, with sampled quiet rows weighted back, a learned position term that is switched off at serving, and streaming counters for posts that are minutes old.
3. 3Value modelCore22 minA feed's value model decides what the engagement predictions are worth: which actions count, how much a comment is worth against a like, what the author gains when someone replies, how survey answers enter, and how each viewer's own habits scale the weights. The weights are estimates of long-term return, checked by experiments that run for months.
4. 4IntegrityCore19 minSome posts must not be shown and some should travel less, however engaging they are. Wire policy classifiers and quality signals into ranking as three actions (remove, reduce, inform), handle posts that haven't been scored yet, and measure the result by prevalence: how many of every 10,000 views were of violating content.

Work through the parts in order and you’ve covered the full design. Each part is one topic.