Recommendation systems: cold starts and feedback loops

A search box gets told what you want. A recommendation system has to guess, drawing on two broad families of evidence: what an item is, and what people have done. Context, user settings, and engineering also shape the result.

Knowing which family a system is drawing on at a given moment explains most of its behavior, including the parts that feel arbitrary. The failures are not bugs. They are the characteristic weaknesses of each family, showing up on schedule.

What to take away

Two families of evidence

The behavior side usually decomposes records of who responded to what into a smaller set of latent factors, which is what matrix factorization does here. Real systems combine both approaches, because each covers the other's blind spot.

That combination is where the interesting behavior lives. A single explanation for a recommendation is usually unavailable even to the people who built the system.

Examples from YouTube, TikTok, and Instagram

These are product-level judgments about visible recommendation surfaces, not claims about each platform’s undisclosed ranking code.

YouTube recommendations on Home and Up next hold up when viewing history helps surface related videos. The weakness is the feedback loop described above: past viewing can keep recommendations close to familiar topics, while new videos start with little behavior evidence.

TikTok’s For You feed holds up as a discovery surface for videos from accounts a viewer may not follow. It does not escape the loop: interactions inform later recommendations, and a new post has little engagement history.

Instagram’s Explore and Reels recommendations hold up for discovery beyond followed accounts. They remain vulnerable to feedback loops and cold starts, and the rationale for an individual recommendation is not fully visible to the user.

The cold start problem

A new item has no behavior attached. A new reader has no history. A system that runs mainly on behavior evidence therefore has nothing to say about either, and this is not a temporary state of affairs, it recurs every time anything is new.

Common cold-start moves include an onboarding survey or ask-the-user prompt, which adds friction; a content-based fallback, which is less personalized; and exploration traffic, which spends some exposure on uncertain recommendations instead of established bets.

If you publish anything, the cold start is the part that affects you directly. Early distribution of a new item is a measurement problem for the system, not a verdict on the item, and the two are easy to confuse from outside.

Why popularity concentrates

Attention flows to what is shown, and what is shown is chosen partly on the evidence of previous attention. That is a positive feedback loop, and it runs whether or not anyone wants it.

The mechanism has no villain in it. Each step is reasonable: an item with more evidence behind it is a safer prediction, so it gets more exposure, which produces more evidence. The aggregate outcome is a heavy concentration at the top and a long stretch of items that never accumulate enough evidence to be judged properly.

The resulting distribution is the familiar power law shape, where a small number of items hold most of the attention and the count falls away sharply behind them.

Systems push back on this deliberately, with exploration budgets, diversity constraints and dampening of raw popularity. Those are corrections applied against the grain of the objective, which is why they get weakened whenever the measured numbers are under pressure.

The loop nobody escapes

Here is the part that is genuinely awkward. A recommender learns from logged behavior. The logged behavior was produced by people responding to what the recommender showed them. So the training data is not a record of what people like. It is a record of what people did with the options they were given.

Two consequences follow.

An item that was never shown generates no evidence of being unwanted. It generates no evidence at all, and a system that treats absence of evidence as evidence of absence will keep it hidden indefinitely.

And measured gains can confirm themselves. A change that narrows what gets shown tends to improve the immediate numbers, since the system is scored on the same narrow distribution it created.

Catching that takes holding some traffic outside the loop. A randomized holdout or exploration traffic can test recommendations beyond the system’s usual choices, but it diverts impressions from the current ranking objective.

What "relevant" means inside the system

It does not mean useful, or true, or good. It means: predicted to produce the logged outcome the system was trained to predict.

That definition is worth holding onto, because it explains disagreements that otherwise look like incompetence. A recommendation that felt insulting may have been an accurate prediction of what you would in fact click. A recommendation that felt useful may have scored badly. The system is not measuring the thing you mean by relevance, and it has no way to.

It also sets the limit on what any advice about recommenders can be worth. You can make an item easier to classify, easier to complete, and easier for someone to want. You cannot make it relevant in the system's sense, because that word belongs to the log rather than to the item.

Common questions

Why do I get recommended something I already have?

Because "already consumed" is a separate piece of state from "predicted to be liked", and it has to be tracked and applied deliberately. Where the two systems do not share information, the prediction wins.

Does interacting with something I dislike teach the system to show me more of it?

Interaction is evidence, and most systems cannot tell approval from irritation, because both produce the same log entry. Where an explicit negative control exists, it carries information that a click cannot.

Why does a small change to an item sometimes change its distribution a lot?

Because ranking is relative and thresholds are involved. Near a cut off, a small movement in a predicted score changes whether the item survives a stage at all, which is a much larger difference than the score change suggests.

Can a recommender be made to show a wider range?

Yes, and it is a deliberate constraint rather than an emergent property. Diversity has to be imposed against the objective, and the cost is measured immediately while the benefit shows up slowly. That asymmetry, not indifference, is why it is a recurring argument.