How TikTok’s Secret Recommendation Engine Powers Its Global Addiction

In early Tuesday, U.S. President Donald J. Trump signed an executive order giving TikTok a 75‑day grace period after the Department of Justice’s shutdown, while demanding that the platform sell half of its equity to a U.S. entity to continue operating.

The move sparked discussion about TikTok’s addictive nature, which experts attribute to its powerful recommendation algorithm. Researchers from the University of Zurich, Maximilian Boeker and Aleksandra Urman, highlighted in their study “An Empirical Investigation of Personalization Factors on TikTok” that the recommendation system is the platform’s key success driver.

A 2022 ACM RecSys paper titled “Monolith: A Real‑Time Recommendation System with Collision‑Free Embedding Tables” provides insight into the engineering approaches used by ByteDance engineers. Although it does not describe TikTok’s exact algorithm, it reveals design choices such as a single unified model for all tasks, online training with real‑time feedback, and the use of Kafka and Apache Flink for feature ingestion.

The Monolith architecture addresses challenges like sparse features and concept drift by employing a “non‑collision” hash table (Cuckoo hashmap) and rapid incorporation of user feedback, enabling the system to adapt quickly to changing user interests.

According to former TikTok engineer Arman Khondker, the algorithm is “years ahead of competitors” and constitutes “the most valuable software”. The system evaluates signals such as likes, comments, and watch time to maximize user value, long‑term user value, creator value, and platform value.

People who use TikTok more often get a more accurate algorithm. — Zhao Zhengwei, Sun Yat‑sen University

Experts note that the more frequently users engage with TikTok, the more accurate the algorithm becomes, creating a feedback loop that fuels the platform’s growth.