Industry
Part of Recommendation systems: cold starts and feedback loops
Recommendation systems: six cases, including two audiences that cancel out
Six recommendation cases with names attached, from YouTube's cold start to TikTok's regional gate, and the flat average that hides two opposed audiences.
Six recommendation cases, each tied to a named system you can check. Reddit's vote-and-time sort and Instagram's feed signals come first. Then YouTube Shorts' watch-time weighting, Pinterest's related pins, TikTok's regional gate, and one flat average that hides two opposed audiences.
Ranking is not intent. It is an inference from the evidence a system could see. Each case below names the system, the evidence it had, and a check you can run in an hour without a platform contact.
What to take away
- A recommendation that looks wrong is usually an accurate prediction of the wrong thing. Ask what was predicted before asking what broke.
- Cold start, thresholds and aggregation cover most puzzles, and each leaves a different trace in your numbers.
- The distinguishing evidence is almost always something you can pull in an hour, without a platform contact.
- When an aggregate is flat but two segments are moving in opposite directions, the average is hiding the story.
- A gate is not a score. A regional block removes your item before ranking ever happens, so no amount of editing fixes it.
Case one: the old post that suddenly moved
Reddit and Pinterest both revive items that looked dead. Reddit's default best sort weighs votes against age, so a months-old post usually stays buried. Pinterest resurfaces older saves through the home feed and the "More like this" row under a pin.
What the system could see. The item's attributes never changed. The evidence around it did. A new post of yours may have pulled in readers whose behavior matches the old item's early viewers. Or the old item's attributes now match something trending, or a threshold tipped on a few new interactions.
What it inferred. That some cluster responds to this item. Late, because the evidence arrived late.
The check. Look at where the new viewers came from and what else they viewed. Arrival through a recent item points to the cluster explanation. Arrival through search or a topic surface points to attribute matching. What a recommender can and cannot infer separates the two families of evidence.
Case two: the follower who never sees you
Instagram's feed and the Facebook feed both produce this complaint. Meta publishes the ranking signals it uses: who posted, your past interactions with that account, and the format of the post.
What the system could see. A follow is one signal, recorded once. Since then, that person's behavior toward your posts has been a run of non-interactions, and non-interaction counts as evidence too. On a surface with more chosen sources than slots, a stale follow loses to an active one.
What it inferred. That the relationship is weak. Correct by its measure, wrong by the person's.
The check. Ask what they did with the last post of yours they saw. If they saw it and scrolled past, the inference is a fair reading. One follow, recorded once, is thin evidence against months of scrolling past.
Case three: the account with one hit and nothing since
YouTube Shorts and TikTok's For You feed both produce this pattern. YouTube's Shorts documentation lists "viewed vs swiped away" as a ranking signal. TikTok's For You description names watch time, likes, shares and comments.
What the system could see. One item produced strong early response and was pushed. Its audience assembled around that item, not the account. Later items do not resemble it, or do not produce the same early response from that audience, so the evidence that promoted the hit does not transfer.
What it inferred. Nothing about the account. It inferred something about one item, and the inference held.
The check. Compare later items' early response from the hit's audience against their early response from your previous audience. If the hit's audience is the cold one, the audience belonged to the item. The arithmetic of why a spike does not repeat applies.
Case four: the post that lands in the wrong place
Pinterest and YouTube classify items by surface features. A pin is sorted partly by its title and image. A new upload is sorted by its title, thumbnail and category, the cheap attributes that exist before any viewing data.
What the system could see. Whatever attributes are cheap to read: title words, a category picked from a list, a format, a thumbnail. If the cheap attributes resemble a popular category, the item is classified there before any behavior evidence exists. This is the cold start problem from the item's side. With no history, attributes are all there is.
What it inferred. That this belongs with the things it superficially resembles.
The check. Change nothing but the cheap attributes on the next comparable item and watch where it lands. If placement follows the attributes, classification is the cause, and it is one you control.
Case five: the two audiences that cancel out
Your overall numbers are flat. Yet two very different groups both say they see more of you than before. This is the case the title promises, and it is the one most often misread.
The pattern has a documented origin outside feeds. UC Berkeley's 1973 graduate admissions figures showed roughly 44% of men admitted against 35% of women. Department by department the gap narrowed or reversed, because women applied more often to departments with low admit rates.
Two audiences, opposite signals
Audience A
- Behavior
- Completes and saves
- Trend
- Growing
- Aggregate
- Flat
- Inference
- Middling
Audience B
- Behavior
- Opens and leaves
- Trend
- Growing
- Aggregate
- Flat
- Inference
- Middling
What the system could see. Two clusters, each growing, with opposite behavior toward you. One completes and saves. The other opens and leaves. Aggregated, the signals average to nothing, and the system's overall estimate of the item does not move.
What it inferred. That the item is middling, when it is strong for one group and wrong for the other.
The check. Split every metric by any segment you have: source, region, new against returning. A flat aggregate over strongly opposed segments is a known trap, Simpson's paradox, and it hides inside averages until someone splits them.
Case six: the plateau in one country
TikTok was blocked in India in June 2020, cutting off a market that had been among its largest. Canada's Online News Act became law in 2023, and Meta blocked news links on Facebook and Instagram that August. Google agreed in November 2023 to pay about CAD 100 million a year into Canadian news.
The EU Digital Services Act, which applied to the largest platforms from late August 2023, requires at least one recommendation option that is not based on profiling. TikTok, Meta and YouTube, all designated very large platforms, had to offer it.
What the system could see. Behavior evidence from that region is either absent or gated. Absent: your items are never shown there, so no evidence forms, so they are never shown. Gated: a constraint outside scoring, such as language, age classification, or a regional rule, removes the item before ranking.
What it inferred. In the first case, nothing, which is the problem. In the second, nothing again, because the constraint acted before inference.
The check. Constraints produce a step. Absence produces a flat line from the start. Look at the shape. Constraints are gates rather than scores, the distinction drawn in what a ranking input can and cannot be, and you check them by reading the rules, not by adjusting the item.
The six cases at a glance
| Case | Named system | What the system saw | Check |
|---|---|---|---|
| Old post revives | Reddit best sort, Pinterest home feed and "More like this" | New cluster evidence around old attributes | Where new viewers arrived from |
| Stale follower | Instagram feed, Facebook feed | Months of non-interaction against one follow | What they did with the last post |
| One hit, then nothing | YouTube Shorts, TikTok For You | Early watch time on one item, not the account | Hit audience against prior audience |
| Wrong-place post | Pinterest, YouTube | Cheap attributes only, no behavior history | Change attributes, watch placement |
| Two audiences cancel | UC Berkeley 1973 admissions, mixed feeds | Opposed clusters averaging to flat | Split every metric by segment |
| Regional plateau | TikTok in India, Canada Online News Act, EU DSA Article 38 | No evidence, or a gate before scoring | Step shape against flat line |
Common questions
Why does every case come back to evidence rather than platform intent?
Because intent is encoded as a countable objective, and the system only sees evidence about that objective. Explanations built on what a platform wants are stories. Explanations built on evidence are checkable.
Is there a case where the item really is the problem?
Cases four and sometimes three. In both, the fix sits on the item, and the evidence tells you which part of it.
Can I ask a platform for an explanation of any of these?
Some surfaces show a reason for a recommendation, such as TikTok's "Why this video". That reason is a simplification chosen for display, a hint about which family of evidence was heavy and nothing more.
What if none of the six fits?
Run the general reach checklist in order, starting with measurement. Most cases that fit none of the six are measurement errors, not ranking decisions.
