For You ranking

How ForYouService turns candidates from four sources into a ranked, varied feed, and how sessions and cursors keep paging stable.

GET /userfeed/foryou?limit=20
GET /userfeed/foryou?limit=20&cursor=20
GET /userfeed/foryou?limit=20&profile=discovery

Every step below is set by a ranking profile; the numbers here are the defaults of the balanced preset, which serves any request that doesn’t name one (Ranking:DefaultProfile).

1. Candidates

ForYouService gathers post IDs from up to four sources, each switchable and capped:

Then, whatever the sources say:

• the member’s favourites’ live posts from the last 48 hours (up to 100) are added;
• a repost stands for the post it shares;
• posts the member pressed Not interested on in the last 30 days are removed;
• only live originals stay: not deleted, visible, not the member’s own, by an active account, and younger than the profile’s maximum age if it has one.

Topic and affinity scores come from Redis sorted sets the engagement worker keeps up to date from likes and comments, and that Not interested / Show more move directly.

2. Signals

FeatureBuilder works out the same 20 signals for every candidate, in a handful of batched queries. They are the model’s inputs, the formula’s inputs, the variables a custom formula can use, and what the feed log stores.

Rates are smoothed, so a post nobody has seen yet is treated as roughly average rather than zero.

3. Filters

The profile’s filters drop candidates outright: excluded tags, only these tags, photos only, a minimum number of likes. Then every remaining post is scored.

4. Scoring

How a score is made has every number in this step and the next, with worked examples. The profile’s scoring mode decides the score; the response’s ranker says what was used.

The model is five LightGBM classifiers, one per outcome, run in C# with ONNX Runtime. A post’s value is

value = 0.5·P(open) + 1·P(like) + 3·P(comment) + 2·P(long read ≥ 10 s) − 1.5·P(skip)

with the weights taken from the profile’s goals. RankerModel reads models/ranker/latest.json from blob storage, checks for a newer one at most once a minute, and refuses a model trained on a different list of signals.

The formula is recency × interest × quality, each part at least 1 so none can zero a post out: recency halves about every 2¾ hours; interest grows with topic match, meaning, this session, the author and following them; quality with engagement and open rates, time on screen and how fast it’s picking up likes now.

5. Adjustments

On top of the score, each multiplying it and each listed as a reason:

• Freshness: an extra half-life, if the profile sets one.
• Boosts: by tag, author, photos, following or not, or favourite (balanced lifts favourites ×1.5).
• Seen already: a post the member saw in the last 24 hours keeps 35% of its score. “Seen” means an impression was reported, or the member opened or engaged with it.

6. Variety and exploration

DiversityReranker turns the scored list into a feed, ten posts at a time:

• at most 2 posts by one author in any ten;
• no more than 2 posts on the same main topic in a row;
• at least 2 posts from accounts the member doesn’t follow in each ten, when there are any;
• 1 exploration slot in each ten (slot 6, counting from 0; two slots use 3 and 7, three use 2, 5 and 8), filled with a random post from below rank 30 that still fits the rules.

Each rule gives way when nothing left can meet it, so the feed never comes up short. Explored posts are marked explored: true and logged with the chance they had of being picked: they’re the only training data not shaped by the ranker’s own past choices. The random choice is seeded per member and minute, so refreshing within a minute explores the same posts.

7. Sessions and cursors

Ranking a few hundred candidates takes real work, and reshuffling while someone scrolls would show them the same posts twice. So:

• The first page (no cursor) always ranks afresh and saves the whole ordered list in Redis as the member’s session for that profile version, for 15 minutes.
• Later pages read the saved order. nextCursor is an offset ("20", "40"…), not a post ID; pass it back unchanged. It’s null at the end of the list.
• Each page is checked again as it’s served: a post deleted, held or whose author was suspended since the first page is left out, so a page can be shorter than limit even when nextCursor isn’t null.
• If the session has expired by the time a later page is asked for, the API ranks again and serves from the same offset of the new list. Drop posts you’ve already shown (by id) when appending a page.
• Sessions are per profile version (foryou:{member}:{profile}), so switching profiles never pages through another profile’s order.

{
  "posts": [
    {
      "id": "…", "authorUsername": "maya_runs", "content": "…", "tags": [ … ], "media": [ … ],
      "ranking": {
        "ranker": "model:20261105-041807", "position": 0, "score": 1.8412, "explored": false,
        "factors": [ { "label": "Your topics", "value": 2.31, "detail": "44% match with what you engage with" }, … ],
        "features": { "age_minutes": 37.2, "follows_author": 1, … }
      }
    }
  ],
  "nextCursor": "20",
  "ranker": "model:20261105-041807",
  "profile": "balanced",
  "candidates": 412
}

profile is <key> for a preset or <key>@<version> for a saved profile; candidates is how many posts were ranked for this session. ranking.factors are the reasons in words, for a Why this post? panel; ranking.features are the raw signals.

Not interested and Show more

POST /userfeed/posts/{postId}/not-interested
POST /userfeed/posts/{postId}/show-more

Both return 204. They’re a state, not a counter: pressing the same one again changes nothing, and switching from one to the other first undoes the earlier one.

• Not interested hides the post from For You for 30 days and counts its author (−3 affinity) and topics (−2 × the tag’s confidence) against similar posts for this member from now on. It works on any post the member was shown, even one held or removed since.
• Show more does the opposite with the same weight (more than a like), and lets a hidden post come back. It only works on posts everyone can see.

Both label the post’s feed log rows, so the ranker learns from them too. The current session isn’t re-ranked: remove a Not interested post from the screen yourself.

Previewing

Staff with the ranking-admin role can see what any settings would give any member, without saving or logging anything: POST /ranking/preview?viewer=<externalUserId>&limit=30 with a profile in the body. The Feed studio in the admin panel is built on it. See Ranking profiles.