Cold start, interest graphs, and why feeds differ

A new item has been seen by nobody. It therefore has no engagement history. But engagement history is exactly what most recommenders use to decide what to show. So it is never shown, so it never gets history. That circle is the cold-start problem, and it appears in three forms:

• New item. Uploaded a minute ago, zero signal.
• New user. Just joined, no history to match against.
• New system. A launch with no data at all.

Cold start is not a corner case for a social platform. On a large video service, the overwhelming majority of the corpus was uploaded recently and is competing for a first impression. A system that could only rank things with history would be a system that could only ever recommend old content, which is useless on a platform whose value is freshness.

How a platform solves this determines almost everything about how it feels to post there.

The classic approach is collaborative filtering: recommend based on patterns of co-consumption, with no understanding of the items themselves.

The logic is "people like you liked this". It never asks what the item is. It only knows that users who consumed A also consumed B, so if you consumed A, here is B.

That indifference to content is its strength. It needs no understanding of video, audio, or language, and it discovers non-obvious pairings a topic model would never guess, because taste does not respect categories.

It is also its wall. An item nobody has consumed yet has no co-consumption pattern. It is not scored low, it is invisible: there is no row for it. Collaborative filtering cannot cold start by construction, not by weakness of implementation. The same applies to a new user, who has no consumption to match against anyone.

So any platform whose corpus is mostly new content needs something else.

The escape is to score an item by what it is rather than by who has touched it.

Run the raw content through a model: transcribe the speech, embed the visual frames, identify the audio track, read the caption. You now have a vector describing the item that required zero engagement data to produce. Drop it into the same space as user vectors and it can be retrieved on its first second of existence.

# available immediately at upload, before anyone has seen it
item_vec = item_tower([
    embed_video_frames(clip),      # what is depicted
    embed_transcript(asr(clip)),   # what is said
    embed_audio(track),            # the sound used
    embed_text(caption),
])
# no likes, no views, no history required

This is why content is embedded rather than tagged. Modern systems infer the topic from the artifact itself, which means hashtags matter far less than creators assume: the model already knows what the video is about without being told.

Content embeddings get an item its first impression. They do not prove it is any good. So the standard mechanism is to treat every new item as a hypothesis to be tested cheaply.

Show it to a small audience the content vector suggests is a good match. Measure how they respond, normalised against what similar items achieve with similar viewers. If it clears the bar, show it to a larger pool. Repeat. If it stalls at any tier, stop spending impressions on it.

This is why a first post from an account with no followers can travel a long way, and equally why a large account's weak post goes nowhere. Each item is re-tested rather than inheriting its creator's status.

Be careful with the specific numbers circulating online, exact pool sizes and precise completion thresholds. Platforms do not publish them, they differ by surface and country, and they change constantly. The mechanism is stable and worth knowing. The folklore around it is not, and treating it as fact is how creators end up optimising for a rule that was never real.

The deepest split between platforms is not their models. It is what they consider the primary structure of the world.

A social graph organises around who you know. Distribution follows connections: you see things because you follow someone, or because someone you follow interacted. Following is the unit of distribution.

An interest graph organises around what you respond to. Distribution follows predicted affinity, computed from behavior. Following is a weak hint, not a channel.

The consequence for a creator is enormous:

Neither is better. They are different bets about where relevance comes from, and each produces its own pathologies.

Every large platform runs retrieval, then ranking, on blended objectives. The differences are in emphasis, and they follow from what each product is for.

• YouTube. Long-form, so watch time dominates and the cost of a bad recommendation is high: a wasted twenty minutes, not a wasted eight seconds. Subscriptions matter, but the recommender drives a large share of consumption.
• TikTok. Short-form and interest-graph-first. Cheap items make aggressive exploration affordable, since a bad guess costs the viewer seconds. That is what makes cold start reach plausible.
• Instagram. Mixed, and unusually well documented, because its head Adam Mosseri publicly describes the system. He has said there is no single algorithm: Feed, Reels, Stories and Explore each rank differently, so it is one product running several. He has also named the signals that matter most across surfaces: watch time, sends per reach, and likes per reach. Note that two of those three are ratios, not totals, which is why a small post with a high send rate can outperform a large one.
• Pinterest. Closer to search than to social. Much intent is explicit and forward-looking, people are planning something, so the objective leans toward usefulness for a task, and items stay relevant for months rather than hours.

The pattern: format length, item lifespan, and whether intent is stated or inferred explain more about a feed's behavior than any architectural detail.

The mechanisms imply a few conclusions, and they are duller than most advice.

Consistency of topic helps, because the item vector must land near an identifiable audience. Content that is hard to categorise is hard to route to anyone.

Early relative response matters, because the tiered test decides whether to keep spending impressions. Not raw counts, but performance against comparable items.

Hashtags are weak where content is embedded. The model watched the video.

Followers are not reach on an interest graph. They are a prior, and each item is re-tested.

What does not follow is any specific tactic. Post times, hashtag counts, and magic thresholds are not mechanisms, they are folklore about parameters that change without announcement. The stable knowledge is structural: a system embeds your item, tests it on a small matched audience, and expands it if the response beats comparable items. Anything phrased as a precise number is a claim about a parameter nobody outside the platform can see.

Content embeddings give an item a score before anyone has seen it, which is what collaborative filtering cannot do. From there it is a tiered experiment: match, test, expand or stop.