The algorithm behind the For You Page of Tik Tok
Tik Tok has quickly become one of the most valuable private companies in the world after becoming vastly popular amongst younger demographics. A lot of the success of this application can be attributed to the For You Page, where users scroll through videos of 15-60 seconds that are generated by a powerful recommendation system. This system utilizes a user-centric design approach to provide for a deeply personalized experience. The algorithm behind this system works with 3 different types of data: content data, profile data, and scenario data. Content data is analyzed using NLP and computer vision to identify and distinguish specific traits of each video. Profile data looks at different traits of the user, such as their age, indicated areas of interest, gender, career, and other observations regarding their identity. Lastly, scenario data takes into account the preferences and behaviors of the user during different scenarios such as time of the day, national events, and activities they may be involved in at the time. Between these three sets of data, correlations and connections are made through modeling to produce a constant stream of videos for the user.
In a macro sense, the app uses this initial data to send new users broadly liked videos and will begin to tweak their models as it compiles the user’s interaction with each video such as sharing, commenting, rewatching, or skipping. Once the app gets a stronger understanding of the user, it begins to send more specific and targeted videos and continues to adjust as the user spends more time on the app. In creating a cycle that relies on the feedback of the user, the app aims to create an addicting never-ending stimulus for the user to continue watching videos.
In relation to some of the concepts of search covered in this course, such as Hubs and Authorities and Page Rank, the search algorithm of Tik Tok operates to provide a more personalized result, and works on a different network of content. Google for example, while it likely also considers these different types of data, has search that acts on a query inputted by the user, while Tik Tok essentially considers the user to be its own search query. Considering the specifics of Page Rank, it works on a network that consists of web pages as nodes, and links as the edges connecting these nodes. This network is constructed using web crawlers that can navigate through the connected components of the web. Tik Tok’s search algorithm, therefore, would have to be fundamentally different as there are no direct structural links between videos, with the exception of duets, which allows new videos to be generated overlaying other videos, and a weak tagging system that is inconsistent, subjective, and broad. This would mean that the network of videos Tik Tok must navigate through, requires edges to be generated through Tik Tok’s own algorithms through its analysis of the content data. Overall, Tik Tok’s innovation of video, query-less search, has resulted in a multi-billion dollar valuation and demonstrates the continued importance of these algorithms in the tech industry.
Sources:
https://newsroom.tiktok.com/en-us/how-tiktok-recommends-videos-for-you
https://towardsdatascience.com/the-inescapable-ai-algorithm-tiktok-ad4c6fd981b8
https://towardsdatascience.com/why-tiktok-made-its-user-so-obsessive-the-ai-algorithm-that-got-you-hooked-7895bb1ab423