Cold Start

Cold Start refers to the moment when a computer program is supposed to get going but doesn’t yet know what it would need for that. A new account with a music service is the typical example: the service is supposed to suggest songs, but has never heard what you like. The same thing happens in reverse with a song that has just been uploaded and that no one has clicked on yet. The term is also used for technology that has to start up again after a break and therefore responds slowly. What both cases have in common: the first run is clearly worse than all the following ones.

The most expensive moment is the first contact

Services on the internet earn money by getting people to stay. The decision about that is usually made within the first few minutes. It is precisely in these minutes that a recommendation system is at its weakest, because it still knows almost nothing about the new user. Anyone who first gets ten irrelevant suggestions often doesn’t log in again.

For providers of new content, the same problem applies in reverse. A video without clicks is rarely recommended, and without a recommendation it gets no clicks. This cycle can make good content permanently invisible. Experts call this a feedback loop: the system reinforces its own starting state.

For servers, on the other hand, a cold start mainly costs time. Half a second of delay sounds harmless. But with a payment page or a voice assistant, it’s enough to make the product feel sluggish. That’s why Cold Start regularly appears in reports about cloud providers, i.e. companies that rent out computing power.

How systems bridge the knowledge gap

The simplest way is to ask. Many services show new users a short selection of genres or artists. These few clicks serve as a makeshift substitute for the missing usage history. The downside: people check what they would like to like, not what they actually listen to.

The second way uses properties instead of behavior. Even without clicks, a new song has a genre, a length, a tempo, and an artist. This allows it to be compared with known songs and shown to matching listeners. Experts speak of content-based recommendations, as opposed to recommendations derived from the behavior of many users.

The third way is deliberate experimentation. The system deliberately mixes uncertain suggestions in among the safe ones and observes the reaction. Every reaction is new information that ends the cold start faster. For servers, the solution works differently: there, the provider keeps some machines warm, i.e. permanently ready for operation. This costs money, but saves waiting time.

Recognizing cold starts in everyday life

The effect is most noticeable with a freshly installed streaming or shopping service. The first suggestions seem arbitrary and rarely hit the mark. After one to two weeks of use, this changes noticeably. The same thing happens when you share an account with someone: the system sees two tastes as one and becomes imprecise.

In business news, Cold Start frequently appears as an argument for market power. Large platforms already have data on millions of users and therefore never really start from zero. A new competitor first has to catch up on this lead. This is precisely what makes market entry so difficult and is regularly discussed in antitrust proceedings.

A common misconception is that Cold Start is a bug in the program. That’s not true. It’s a data problem: the code works flawlessly, it just lacks the basis for good decisions. That’s why better technology alone helps little as long as the initial information is missing.