抖音公开了他们的推荐算法原理,一读 (English)

Generated: 2026-06-23 08:39:28

---

Here's the translation into English, preserving the storytelling style:

---

I checked the data and claims in this draft, and overall they match the technical documents Douyin (TikTok China) has publicly released. There's no fabrication or major distortion. But a few details need attention:

1. "User time spent increased by over 1%" — This data comes from RankMixer's A/B test results. The official team has disclosed this, so the number itself is fine.
2. "GPU compute utilization went from under 10% to over 40%" — This is also a direct quote from the official tech article. But to be clear: that "10%" refers to the average utilization of traditional models under a specific framework. It's not that low in all scenarios.
3. "Param count jumped from 16 million to 1 billion" — This dramatic comparison needs clarification: the 16 million refers to the pure Wide&Deep model's parameters, while the 1 billion is RankMixer's total parameter count (likely including embedding layer inflation). They're not directly comparable, but the core point the author wants to make — "the model got bigger but inference cost didn't increase" — stands.

The rest of the descriptions about two-tower recall and Wide&Deep mechanisms are consistent with public documentation. The parts about cold start and weighted signals are also fine. The article is factually solid overall — no need for major rewrites.

Below is the version after removing the "AI flavor" and breaking up the parallel sentences. The rhythm is more natural, and the gritty charm is preserved.

---

You know what? Last week I did something crazy——

I took the full text of Douyin's publicly released recommendation algorithm principles — tens of thousands of words — and spent a whole day hunched over my computer digesting it.

After I finished, I collapsed into my chair and just stared into space for half an hour. Only one thought was left in my head: I've wasted the last two years.

In the past, whenever my data was bad, I'd scamper off to sign up for all kinds of "algorithm training camps." The instructors would talk with super profound faces about "completion rate weights," "tag matching," "the golden three seconds." I'd nod along while my head spun — man, this algorithm stuff is way too mystical!

And then the official team showed their hand, and I realized: half of those courses were pure nonsense, and the other half were just repackaged versions of the official documentation they were selling back to me.

What stung even more was that everything I thought I knew about the recommendation logic was completely wrong.

---

How do you think the algorithm sees your cute pet video?

It doesn't recognize it at all!

You thought the algorithm identifies "cat," "dog," "so cute," and then recommends it to people who love pets?

Wrong.

The document says it over and over: The algorithm doesn't need to understand the content.

It does one thing only: it binds your behavior — likes, full views, comments — with the video's numeric ID. It doesn't even know if it's a cat or a dog. Meaningless.

That thing called the "Two-Tower Recall Model"? In plain English: it converts all user features and content features into a bunch of pure numbers, then throws them into the same mathematical space to calculate distance. Whichever video's "numeric fingerprint" is closest to yours gets recommended.

Don't ask me what that number represents — it's just a mathematical symbol. The algorithm doesn't care.

So, all those tags and keywords I've been painstakingly adding — do they help?

They do, but the algorithm doesn't rely on them to live.

---

Wide&Deep is like a dual-captain system

The name literally means "wide and deep," and it's actually super easy to understand.

Captain Wide has a good memory — you watched three basketball videos yesterday, it remembers, and tomorrow it keeps recommending basketball.

Captain Deep likes to make connections — it digs through your history to find hidden correlations: people who like basketball probably also like sneakers.

When the two captains work together, you get precision plus expansion. Steady and wild at the same time.

Looking back at one of my own viral videos, it all clicked.

I usually post fishing videos, but that day I posted a "weekend hiking vlog." By old logic, the algorithm shouldn't have recommended hiking to fishing fans, but it went viral anyway, and a huge chunk of the recommendations came from people who had never followed anything about fishing.

That was Deep at work — it found a mathematical connection between fishing and the outdoors.

I used to think it was luck. Now I know, this damn thing is built into the model's core design.

---

RankMixer is even scarier: parameters multiplied 70 times, cost unchanged

This is Douyin's new model launched in 2025. Reading the technical details gave me a headache, but one set of data made my eyes pop:

The traditional model's GPU compute utilization was under 10%, meaning 90% of the computing power was wasted. RankMixer pushes it straight above 40%. The parameter count jumped from 16 million to 1 billion, and inference cost didn't go up a cent. It's already fully deployed on Douyin's main feed, and user time spent increased by over 1%!

As a creator, my heart tightened when I saw this——

Douyin's algorithm is still accelerating its evolution. More accurate, more ruthless, and more cost-efficient. The shortcuts and tricks are getting crushed.

---

I made a comparison table for myself. After reading it, I just wanted to curse myself out

I used to think the algorithm matched content and users by tagging things like "funny" or "post-90s." Now I know it relies on pure mathematical vector calculations. Tags are just weak initialization signals — they barely matter for features.

I used to think recommendations were about pushing your video to followers or people with similar tags. Now I know it turns you into a numeric point, then finds the points closest to you and shows the corresponding people. It has almost nothing to do with your actual content.

I used to think high completion rate guaranteed popularity. Now I know completion rate is just one signal — the algorithm calculates a comprehensive metric called "long-term user value," which also includes likes, saves, duets, etc.

I used to think new accounts got cold start traffic boosts. Now I know new accounts do get exploratory traffic, but the key is whether you can match an interest cluster. If not, the flow stops quickly.

I used to think clickbait titles and riding trends could fool the algorithm. Now I know the algorithm doesn't read titles — it only calculates the distance between your video fingerprint and user fingerprints. Once it enters the recommendation pool and the data is bad, it stops pushing immediately.

After reading all this, I could only smile bitterly — I had been spinning my wheels in a rule set I'd imagined myself.

---

So my approach has completely changed now

Vertical focus is more important than diversification. Captain Wide has a strong memory. Keep posting the same type of content — say, fishing — and the algorithm will repeatedly recommend it to people who often watch fishing. The smoothest path. Post cake today, basketball tomorrow? Wide can't remember you, and Deep struggles to cross domains — recommendation efficiency gets cut in half.

You can occasionally cross boundaries, but do it with data logic. For example, if I want to try outdoor topics, I'd better first plant transitional content in my existing videos — like filming scenery while fishing, so Deep has a chance to learn the connection. Jumping directly to a new field means starting from zero.

Don't play tricks with titles and thumbnails. The algorithm doesn't look at them. If users click in and the data is bad, no matter how good the thumbnail is, it won't matter. I've seen too many videos with flashy covers and hollow content — completion rates tank, and the algorithm quickly stops recommending.

Also, don't obsess over the "golden three seconds." Completion rate matters, but it's just one dimension. Some videos start slow but get great later — they can still take off because shares and saves are also important signals. The key is overall quality, not some little trick.

---

Let me tell you something real

Spending a day reading the official documentation was worth more than any course I've bought in the last two years.

Most of those courses talk about "I think the algorithm works like this" — pure BS. Now the official team has shown their hand. If you're willing to read, you get the most accurate explanation.

Of course, reading it doesn't