Enter your search text in the box above

Select a result to preview

Cold Start Problem

#tiktok #recommendation-systems #machine-learning

Cold Start Problem

The cold start problem is the challenge of providing good recommendations when you have little or no historical data about a user or content item. TikTok's solution to this problem is one of its most important architectural advantages over competitors.

Two flavors of cold start

New user cold start

A user opens TikTok for the first time. They have no watch history, no likes, no follows. How does the Two-Tower Neural Network generate a meaningful user embedding with no data?

New content cold start

A creator uploads a video with zero views, zero engagement. How does the algorithm decide who should see it?

TikTok solves both problems through fundamentally different mechanisms.

Solving new user cold start

The first 10 videos are a Bayesian experiment

When a new user opens TikTok, the system has almost no behavioral data. It uses three weak signals to generate an initial user embedding:

  1. Device type -- iOS vs Android correlates with demographic preferences
  2. Location -- Regional content preferences
  3. Time of day -- Morning users prefer different content than late-night scrollers

The system then serves a default mix of trending and popular content. But the real work happens in the first 5-10 interactions:

Each swipe is a data point

The first 10 videos TikTok shows a new user are a real-time Bayesian experiment. Each video is a probe testing a hypothesis about the user's interests. Each swipe narrows the interest profile. By video 10, TikTok knows more about your content preferences than most friends do.

The progression:

Video # What the algorithm knows How it updates
1 Almost nothing (device + location) Serves trending content
2-3 Weak preference signal (which trending content you watched) Adjusts user embedding slightly
4-6 Emerging interest clusters Begins ANN search with partial embedding
7-10 Strong preference signal Full personalized feed kicks in

The two-tower advantage

The Two-Tower Neural Network architecture is what makes this fast convergence possible. Even with a weak user embedding, the video tower still works perfectly. The system can find videos that are semantically similar to the few you've watched, using content features alone.

This is TikTok's architectural advantage over YouTube and Instagram:

Solving new content cold start

Every video is a new experiment

When a creator uploads a video, TikTok does not wait for organic discovery. The system immediately:

  1. Generates content embeddings at upload time using computer vision, NLP, and audio analysis
  2. Tests the video against a small micro-audience of 200-500 users
  3. Measures engagement velocity -- how quickly does the video accumulate completion, saves, and shares?
  4. Amplifies or suppresses based on performance

See Progressive Amplification Waves for the full distribution funnel.

Multimodal understanding

Before a single human sees the video, TikTok's AI "watches" and "listens" to it:

This multimodal understanding means the video can be categorized and matched to interested users purely through content features, without any engagement data.

The upload-to-FYP pipeline

TikTok's target: a video should be FYP-eligible within 60 seconds of upload. The moderation classifiers and basic embedding generation run before the full transcoding pipeline completes. The system publishes a "provisional" FYP flag after fast-path checks.

Why cold start matters for the platform

For users

Cold start determines whether a new user's first experience is engaging or frustrating. If the first 10 videos don't resonate, the user leaves. TikTok's fast convergence is a retention mechanism.

For creators

Cold start determines whether new creators can reach an audience. On YouTube or Instagram, a creator with zero followers has essentially zero reach. On TikTok, every video starts on equal footing. This is why TikTok has a stronger creator economy for newcomers.

For the platform

Cold start determines content supply diversity. If only established creators with large followings could get distribution, the content library would stagnate. By treating every video as a new experiment, TikTok ensures a constant influx of fresh content.

The follower-first 2026 shift

In 2026, TikTok introduced a modification to the cold start mechanism. New videos are now initially distributed primarily to the creator's existing followers for the first 24-72 hours before broader FYP distribution. This uses followers as algorithmic gatekeepers -- if the video performs well with followers, it earns the right to reach non-followers.

See Progressive Amplification Waves for the updated distribution model.

Cold start vs warm start

Aspect Cold Start Warm Start
User data None or minimal Extensive history
Embedding quality Weak, approximate Strong, precise
Recommendation source Content features + trending Full two-tower matching
Convergence time 5-10 interactions Already converged
Feed quality Good enough to retain Highly personalized

See also