Understanding Douyin's Recommendation Algorithm: From Behavior Prediction to Value Modeling

Recommendation Only Estimates Behavioral Actions

When a user opens Douyin, the recommendation algorithm scores candidate videos and pushes the highest‑scoring video to the user.

User interactions such as likes, completions, and dislikes serve as feedback actions that indicate the user's interest level.

Each feedback carries positive or negative value, and the sorting model learns from these signals to deliver videos with the highest feedback value.

The core logic can be simplified to the "recommendation priority formula": predicted user‑action probability × behavior‑value weight = video recommendation priority.

Recommendation Priority Formula

Specific Steps of Douyin's Recommendation Action‑Rate Estimation

Algorithm Learning: User Feedback Input

The model predicts the probability of various user interactions and feedback. Before prediction, massive feature data—both real‑time and historical user behavior—is fed into the model. The diagram below shows the related model architecture.

Douyin's Deep Learning Algorithm for Learning User Feedback

Probability Model Prediction: Which User Actions Are Estimated

The recommendation algorithm predicts the probability (action rate) of user behaviors for each candidate video, combines this with the value weight of the action, and computes a score that determines which video is shown.

Typical actions include likes, follows, collections, shares, dislikes, clicking the author’s avatar, comment dwell time, long‑term consumption, and more.

Examples of User Actions Estimated by Douyin's Recommendation Algorithm

Value Model Evaluation: Defining the "Recommendation Value" of Actions

The value model reflects Douyin’s understanding of which actions are more important. Relying solely on interaction probability can disadvantage high‑quality mid‑length videos that have lower completion rates.

Therefore, the value model aims for a win‑win among content, users, creators, and the platform by calculating the value of each interaction and continuously adjusting parameters.

Following the "value maximization" principle, the algorithm computes a score for each candidate by weighting multiple factors, including content characteristics, creator earnings, and ecosystem health.

Through deep analysis and weighting, the algorithm assigns a potential‑value score to each candidate, providing a solid basis for the final recommendation.

Dynamic Adjustment: Real‑Time Feedback of Value Weights

With rapid algorithmic advances, the time from action occurrence to effective feedback is now very short, enabling minute‑level real‑time updates that help the system more accurately anticipate user behavior.

Human Establishes Content Order for the Algorithm

Regardless of its complexity, the algorithm’s core is to learn from user feedback data, compute probabilities, and push the video with the highest recommendation value.

The algorithm does not need to understand why a sunset video is appealing; it merely calculates watch time and share probability. However, without constraints, such blind optimization can lead to the spread of inappropriate content, so multiple safeguards are required to prevent the algorithm’s “cognitive blind spots.”

Source: https://95152.douyin.com/article/15381?enter_from=copy_link&channel=transparency