TikTok currently has close to a billion monthly active users and is the most addictive of all social media platforms. Algorithms determine your preferences within a very short period of time due to the intensity of the signals they pick up in concentrated form.
A leaked paper from the NY Times explains that in an effort to grow its daily active users, TikTok has decided to optimize two closely related metrics for video views. The first is whether the user is returning to the same type of content. And second, how long they watch a particular video versus videos from other categories.
The document says watch time isn’t the only factor that TikTok takes into account. The document provides a rough equation for how videos are scored, which summarizes machine learning predictions and actual user behavior. It also shows in detail how the company tunes its system to identify and use “bait for likes” (videos designed to play with the algorithm by explicitly asking people to like them).
TikTok is different from other social media in that it doesn’t rely on knowing who your friends are when deciding what content to put in front of you. Instead, TikTok’s algorithms figure out what exactly attracts you.
AI is very good at understanding what you really want to see and then feeds it to you. But beyond that, the algorithms also understand that the user may soon get bored with constantly watching the same material. So the algorithms will start testing related but different content to see if you like it.