TikTok Sheds Some Light on Why Videos Appear in Users’ For You Feeds

TikTok provided some insights into how it recommends videos for users’ For You feeds. The video creation and sharing platform said in a blog post that the For You feed is unique for each user, and its machine learning system recommended content based on factors including: User interactions, meaning videos people like or share, accounts they follow, comments they post and content they create. Video information, including details such as captions, sounds and hashtags Device and account settings, including language preference, country setting and device type. TikTok said these factors are included to ensure that the user’s system is optimized for performance, but they receive lower weight in its recommendation system compared with the factors listed above. TikTok’s recommendation system processes those factors and assigns them a weight based on value to the user. The company said as an example that a user watching a longer video from beginning to end would have much more impact that a viewer and creator being in the same country. Videos are then ranked to determine the likelihood of interest to the user, and served to the For You feed. Follower count is not a factor in TikTok’s recommendation system, nor is whether the account that created the video had high-performing videos in the past. TikTok invites new users to select categories of interest to help build the initial For You feed, tweaking recommendations based on early interactions. Those who do not choose categories are shown a generalized feed of popular videos, and TikTok said their first set of likes, comments and replays will trigger the process and help its system learn more. The company wrote, “Every new interaction helps the system learn about your interests and suggest content, so the best way to curate your For You feed is to simply use and enjoy the app. Over time, your For You feed should increasingly be able to surface recommendations that are relevant to your interests.” Other actions that help TikTok refine recommendations include following new accounts and exploring hashtags, sounds, effects and trending topics on the Discover tab. When a user sees a video that doesn’t interest them, long-pressing on it and tapping “not interested” will influence future recommendations, as will choosing to hide videos from specific creators or containing certain sounds, or videos that appear to violate TikTok’s guidelines. TikTok wrote, “One of the inherent challenges with recommendation engines is that they can inadvertently limit your experience—what is sometimes referred to as a ‘filter bubble.’ By optimizing for personalization and relevance, there is a risk of presenting an increasingly homogenous stream of videos. This is a concern we take seriously as we maintain our recommendation system.” In order to help prevent this, TikTok’s recommendation system works to intersperse diverse types of content with fare the user already likes, and the For You feed will generally not show consecutive videos with the same sound or from the same creator. TikTok does not recommend duplicated content, content the user has seen before or spam, but it may recommend videos that have been well-received by other users with similar interests. The company wrote, “Diversity is essential to maintaining a thriving global community, and it brings the many corners of TikTok closer together. To that end, sometimes you may come across a video in your feed that doesn't appear to be relevant to your expressed interests or have amassed a huge number of likes. This is an important and intentional component of our approach to recommendation: Bringing a diversity of videos into your For You feed gives you additional opportunities to stumble upon new content categories, discover new creators and experience new perspectives and ideas as you scroll through your feed.” On the subject of user safety, videos depicting things such as graphic medical procedures or legal consumption of regulated goods may not be eligible for recommendation, and the same applies to videos that have just been uploaded or are under review, as well as to spam content aimed at artificially increasing traffic. TikTok concluded, “Developing and maintaining TikTok's recommendation system is a continuous process as we work to refine accuracy, adjust models and reassess the factors and weights that contribute to recommendations based on feedback from users, research and data. We are committed to further research and investment as we work to build in even more protections against the engagement bias that can affect any recommendation system. This work spans many teams—including product, safety and security—whose work helps improve the relevance of the recommendation system and its accuracy in suggesting content and categories you're more likely to enjoy.”