TikTok's AI Algorithm
Published by Sid Chadha on July 19, 2020
In recent years, TikTok has grown to be one of the most popular video streaming apps in the industry. Over time, viewers keep getting more and more personalized content, which is why they keep scrolling and watching more videos on this app. The reason TikTok is able to keep millions of high schoolers latched on to their platform is because of the artificial intelligence algorithms they use. There are numerous algorithms TikTok uses but there are 3 major ones that make up the TIkTok interface that has become so popular today.
The Submission Process
The submission process can be described by the below image. The machine review tries to review the content based on factors such as inappropriate keywords, hashtags, and others.
How TikTok's AI Algorithm works, adapted from towardsdatascience.com
The User and Content Creator
TikTok uses multiple factors to determine the best possible content. For both the user and creator, TikTok utilizes their personal information, including their location, internet searches, and characteristics, creating much controversy over their misuse of data.
In terms of the user, the algorithm reads video information, including hashtags, sounds, captions, coupled with user interactions like the videos you like/share, followed accounts, posted comments, and the type of content that you create to build a personalized reccomendation engine. Surprisingly TikTok also uses your language preference, country setting, and device type to also predict the best content for you. But this would also account for a lower weight since preferences wouldn't account for the type of content you would want to watch.
As you may have noticed, TikTok doesn't force you to make an account when downloading the app and the app instead tries to keep you on the app for as long as it can. This helps TikTok build a dataset of videos that you may spend longer or shorter amounts of time watching. This creates a score or weight for each genre on your profile and anytime a video is rewatched, that genre is given a higher weightage in the reccomendation system.
Analysis of how a new user is drawn in, adapted from veed.io
On the content creator side, TikTok reads the music, hashtags, and filters on a video takes into account. Interesting enough, TikTok also uses natural language processing, a form of machine learning, to create a transcript of the video and understand it better. Lastly, it uses computer vision to understand what is happening in the video. It compiles all of this to put it in different categories (example: informational?). If a separate user has liked, shared, and rewatched these informational videos, this one has a high probability of showing up on this users "For You" page.
How TikTok analyzes a video, Adapted from veed.io
Due to ByteDance's (the company behind TikTok) secrecy, it is unknown how these inputs work in a reccomendation engine. There are three theories, though, of how this works:
The first theory is called batch theory and essentially involves a pool of batches of users with different preferences being shown a video and how each group reacts (Like, share, commenting, rewatching) pushes your content to a larger batch to watch.
The second theory is known as authority ranking and gaining popularity in your videos depends on the first few videos. If the first few videos are able to gain momentum, then the following content you create will get more popular on the for you page. All you have to do now is engage the audience with the content you create.
The last theory is delay momentum and this theory involves that if you have not posted in a while, the platform will use its algorithm to gain more viewers on the older content you have produce. This will thus help to encourage a given user to start making content again.
Analysis of how a video posted to TikTok could work, Adapted from veed.io
After the data is processed and videos are recommended, the longer the user watches a video, there is a greater success and weight given to that type of content for future recommendations. Each reaction, such as a like or comment, will be given different weightage, with some being more important than others. This builds a sizable amount of data to train a model predicting optimal videos for a user on. After training the model on videos that the user would be most likely to watch, they are ranked on each for you feed. This could be done by a simple logistic regression algorithm or a deep feed forward neural network. The hard part, though, would be scaling this algorithm to millions of users and combining algorithms, such as natural language processing with convolutional neural networks to recognize the song in this video. Additionally, a video that receives more views will lead to a higher likelihood of it being on your for you page but at the same time, it will not be determined by if the creator is popular. The best fit reccomendation will continue to be adjusted as you react positively to different videos and switch up your internet search history all the time. The algorithm behind the "For You" page also tries to recommend content that is unique, so that two videos in a row with the same sound/creator will not be shown.
These 3 algorithms are so effective that the average TikTok user spends 52 minutes per day on the app, equating to almost 200 videos. This is the exact reason why TikTok is the world's most valuable start up!