How a score is made

Every number behind a For You post's place, from the built-in formula and the trained model to your own formula, freshness, boosts and the seen-already penalty, with worked examples.

A For You post’s place comes from one number, its score, worked out per member each time a first page is ranked. This section explains that number exactly: the parts FeedFabric computes for you, the parts a ranking profile sets, and the formula language you can use to write your own. The code is in src/FeedFabric.Application/Services/Ranking.

The pipeline

1. CandidatesUp to four sources plus favourites, minus hidden, held, own and too-old posts.
2. Signals20 numbers per post for this member (FeatureBuilder).
3. FiltersThe profile's tag, photo and minimum-likes filters drop posts outright.
4. Base scoreThe scoring mode: model, built-in formula, your formula, or newest.
5. AdjustmentsFreshness, then every matching boost, then the seen-already penalty, each a multiplier.
6. ArrangeSorted by score, then spread out for variety, with exploration slots mixed in.

Only steps 4 and 5 make the score. Steps 1 to 3 decide which posts get one, and step 6 decides where each one goes, using the scores’ order and nothing else: a post scoring 40 isn’t treated any differently from one scoring 4.01 if both are next in line. So you never need to keep scores in a particular range; you only need the right posts to score higher than the others.

The signals

Every scoring mode reads the same signals, worked out for the member reading the feed. They are also the variables a formula can use; Formula language lists them with their ranges.

The ones that do most of the work:

impressions, opens, ctr, engagement_rate, avg_dwell_s and the two viewer-author signals come from the reading signals your front end sends. Without them these stay at their starting values for every post, and the formula and model can only rank on topics, authors and age.

Base score: the four modes

scoring.mode picks how the base score is worked out. The ranker on the response, and on every feed log row, records which one actually ran.

The built-in formula

HeuristicRanker scores every post as nine parts multiplied together. Every part except recency is 1 or more, so a post that’s weak on one thing (no topic match, say) is never knocked to zero by it; it just doesn’t get that part’s lift.

formula = recency × topics × meaning × session × author × follows × popular × trending × attention

recency   = e^(−age_minutes / 240)                          1 when new, half after about 2¾ hours
topics    = 1 + 3 × topic_match                             up to ×4
meaning   = 1 + 1.5 × max(0, semantic_match − 0.2)          up to ×2.2; similarity under 0.2 counts for nothing
session   = 1 + 2 × session_topic_match                     up to ×3
author    = 1 + 0.6 × ln(1 + author_affinity)               keeps growing, more and more slowly
follows   = 1 + 0.4 × follows_author                        ×1.4 if you follow them
popular   = 1 + 8 × engagement_rate + 4 × ctr               ×1.6 for a post with no views yet
trending  = 1 + 0.15 × ln(1 + velocity_30m)
attention = 1 + 0.05 × min(avg_dwell_s, 20)                 up to ×2, at 20 seconds

Each part, with its value and a sentence about it, is one of the post’s ranking.factors, which is where a Why this post? panel gets its reasons:

A worked example. A post from two hours ago by someone the member follows and has liked four times, 30% on their topics, fairly close in meaning (0.5), with 8% engagement, 12% opens, 3 likes and comments in the last half hour, and six seconds’ average dwell:

The same post from someone they don’t follow and have never liked scores 15.31 ÷ 1.4 ÷ 1.966 ≈ 5.56, and a brand-new post nobody has seen yet, on none of their topics, from a stranger, scores about 1.6. That’s the formula’s character: interest and quality multiply, and age decays everything at the same rate.

The model’s value

The trained ranker is five LightGBM classifiers, each predicting one thing the member might do with the post. The profile’s goals say how much each one is worth, and the model’s value is the weighted sum:

model value = open × P(open) + like × P(like) + comment × P(comment) + longRead × P(long read) + skip × P(skip)

With the default goals, a post with P(open) 0.3, P(like) 0.12, P(comment) 0.03, P(long read) 0.25 and P(skip) 0.4 is worth 0.15 + 0.12 + 0.09 + 0.5 − 0.6 = 0.26. A negative weight makes the value go below zero for posts the model expects to be skipped; that’s fine for ordering, and the adjustments below are written to cope with it.

Changing goals changes what the feed is for without retraining: the community preset puts comments at 5 and opens at 0.3, so it favours posts that start conversations. Until a model has been trained and published (make train), model mode quietly uses the built-in formula and says heuristic.

Your own formula

With mode: "expression", scoring.expression is a formula in a small, safe language: arithmetic, comparisons, a dozen functions, and the signals as variables. Two extra variables bring in the scores above:

• formula is the built-in formula’s score.
• model is the model’s value shifted up to be 0 or more: FeedFabric adds the total of the negative goals back (1.5 with the defaults), so the example above is 0.26 + 1.5 = 1.76. The order is unchanged, and multiplying it by a decay or a boost can only lower or raise a post as you’d expect. Without the shift, halving a negative value would raise it. While no model is loaded, model is the built-in formula’s score instead.

The fresh preset is a formula:

formula * decay_exp(age_hours, 3) * (1 + 0.3 * log1p(velocity_30m))

The built-in score, halved every three more hours, and lifted by up to about ×2 for a post picking up speed. The cookbook has more, and Formula language everything a formula can contain. A post’s ranking.factors start with Profile formula, the formula’s value for it, followed by the built-in formula’s parts for context.

Adjustments

After the base score, up to three kinds of multiplier apply, in this order, each one added to the post’s ranking.factors:

Boosts compound: a photo post from a favourite under a profile with hasMedia ×1.2 and favourite ×1.5 gets ×1.8.

Because the model’s value can be negative, a multiplier isn’t applied as a plain multiplication:

So a boost above 1 always lifts a post and one below 1 always lowers it, whatever the sign, and a 0 multiplier is a reliable “put these last” without removing them. To remove posts, use filters instead.

freshness.maxAgeHours isn’t a multiplier: posts older than it are never candidates at all.

From scores to a feed

Posts are sorted by final score, highest first, and then DiversityReranker builds the feed ten at a time, taking the highest-scoring post that keeps to the profile’s variety rules (posts per author, topic runs, a share from outside the member’s follows) and filling exploration.slots with random picks from below exploration.fromRank. For You ranking has the details.

That’s why a post’s ranking.position doesn’t always follow its score: the post at position 3 may have a higher score than the one at position 2 that had to wait for a different author to come first, and an explored post (explored: true) may score far below the posts around it.

Seeing the numbers

• On a feed: every For You post has ranking.score (final), ranking.factors (each part and adjustment) and ranking.features (the raw signals).
• Before saving: the Feed studio, or POST /ranking/preview?viewer=<externalUserId>, ranks with any settings for any member without saving or logging anything. See A formula editor.
• After saving: the dashboard compares profiles and versions on what members actually did.