Introduction to the TikTok recommendation system

Your For You Feed

Building the best TikTok experience for our community

More than one billion people around the world come to TikTok every month to create, discover and share. This includes creators looking for new audiences, small businesses that want to grow, and the many millions who come to TikTok to find entertainment. We continually focus on providing them with inspiring, enriching and joyful TikTok experiences.

Central to delivering this goal is our recommendation system. It powers the For You feed, a core part of the TikTok experience. As one of the most important building blocks of our app, we aim to build the recommendation system in line with these fundamental criteria:

• Providing a safe experience for a broad audience, and in particular teens
• Respecting local contexts and cultural norms
• Maintaining content neutrality, or in other words, the recommendation system is designed to be inclusive of all communities and impartial to the content it recommends on the For You feed
• Championing opportunities for original and creative expression
• Enabling new content exploration by promoting a variety of content and topics

How TikTok’s For You feed works

When you interact with content on TikTok, you provide us with a variety of signals. These signals assist our recommendation system in understanding the type of content you might enjoy. We use these signals to create a prediction score for different videos. This score is an estimate of how likely you are to interact with each of those videos. Videos with the highest score are selected to be served to your For You feed. Below is a snapshot of the factors that shape what you see in your feed.

Signals from you

The signals you provide when you engage with content on TikTok encompass a range of interactions, such as the videos you have viewed, liked, or shared. By interacting with content on TikTok, you inform the system about the content you might find relevant and useful, helping shape your unique For You feed experience. We also take into account how others interact with content to help inform your experience – for example, if another user has engaged with two of the same videos that you have interacted with, the system may predict that you are likely to also engage with a third video that this user has interacted with.

See the appendix for examples of key signals utilized in the recommendation system.

Recommending your video feed

The recommendation system uses the signals you provide to predict which content you’re likely to enjoy.

To deliver your For You feed, the recommendation system, which is made up of different machine learning models, goes through a number of different steps, which can be broadly summarized as:

1. Selecting videos for recommendation
2. Predicting
3. Ranking
4. Similarity check
5. Recommendation rules

Selecting videos for recommendation

At this stage, the system retrieves a large number of videos that have passed our initial content moderation processes and is eligible for recommendation to the For You feed.

Prediction

Predictions estimate the likelihood you will interact with a given video; for example, the probability of liking, sharing, commenting, or skipping a given video. By combining prediction scores for each of these interactions, we can arrive at an overall score for each video, with higher scores signifying a closer match to your preferences.See the appendix for examples of the recommendation system’s key predictions.

Ranking

The ranking process sorts videos by their prediction scores, from high to low of the likelihood you’ll enjoy the video. Your For You feed is created from these top-ranked videos, provided they pass a final similarity check before being recommended.

Similarity check

Content similarity measures how different consecutive videos in your For You feed are. If some of the videos that the system has chosen as top-ranked for you are too similar to each other – for example, they use the same sound – we replace them with others to ensure a greater variety in the content you see. These efforts aim to reduce the chance of you viewing videos with repetitive themes and help you instead explore and discover different types of content.

Recommendation rules

As the system goes through the actions outlined above, we have additional recommendation rules that are applied at different points. Rules provide an extra layer of assessment, which helps ensure that video recommendations create the best TikTok experience for our community. For example, we have rules in place that ensure videos in your For You feed include content from creators in your region, and that content created by accounts aged under 16 is ineligible for recommendation.

Building for safety in the recommendation system

Moderation system: Prioritizing safety when recommending content

Our Community Guidelines set the rules that determine which types of content are not allowed on the platform. They also outline the categories of content that are not eligible for recommendation in the For You feed, which is an experience intended for a range of audiences, from teenagers to great grandparents. We deploy a mix of machine learning models and human reviews to enforce our Community Guidelines. If you come across a video you think goes against our rules, you can long-press on the video and select ‘Report’ from the pop-up menu to send send the video to our moderation team for review.

When a video is first uploaded to TikTok, it goes through our content moderation system, which is a combination of both automated systems and moderation teams. Content that violates our Community Guidelines is removed, and content that is not suitable for a broad audience is made ineligible for recommendation. Only videos that pass moderation are included in the pool of videos that can be recommended to our community in the For You feed.

Using the prediction scores, the recommendation system picks from the approved videos those we believe you are likely to enjoy.

Youth safety

We prioritize safety so that we can provide teens on our platform with an age-appropriate TikTok experience. We recognize that while the vast breadth of content that is shared on TikTok adds to the rich diversity of the experience, not all of it may be suitable for younger audiences. We limit overtly mature content, so it is not available anywhere on TikTok for accounts aged 13-17, and only viewable by adults 18 years and older. In addition, any content created by an account aged under 16 is not eligible for recommendation to the FYF. Our Guardian’s Guide provides more information about our setting and tools that create a safer experience for teens on TikTok.

How you can influence your For You feed

Aside from the signals you provide by how you interact with content on TikTok, there are additional tools we have built to help you better control what kind of content is recommended to you.

• Not interested: You can long-press on the video in your For You feed and select ‘Not interested’ from the pop-up menu. This will let us know you are not interested in this type of content and we will limit how much of that content we recommend in your feed.
• Video keyword filters: You can add keywords – both words or hashtags – you’d like to filter from your For You feed.
• For You refresh: To help you discover new content, you can refresh your For You feed, enabling you to explore entirely new sides of TikTok.

Appendix

This non-exhaustive list offers insight about the variety of signals and interactions the system utilizes to deliver personalized content to your FYF. The signals that serve as input to our system are dynamic and constantly updated to encourage evolving interests and a thriving ecosystem. If you are interested in learning more, head to our Help Center to view up-to-date information on how TikTok recommends content.TikTok personalizes your For You feed based on various factors, including:

• Your region, aligning recommendations to local standards and cultural norms.
• The language your device is set to.
• Your device’s operating system.
• Videos you’ve recently interacted with, including those you watched, finished, liked, or skipped.
• Your follow relationships with other people on TikTok.

The system predicts a variety of interactions, including your likelihood to:

• Like, share, comment on, or mark a video as ‘Not Interested.’
• Follow the video’s author or interact with their profile.
• Finish, skip, or favorite a video.
• Spend a certain amount of time viewing a video.
• Tap the video’s soundtrack.

In addition to signals you create by interacting with content, we also take into account various signals about the video, including:

• The time the video was posted.
• The region the video was posted from.
• The language setting of the video’s author.
• The video’s soundtrack.
• The length of the video.
• The hashtags associated with the video.