The paper you link does not describe YouTube's algorithm. YouTube's selection algorithm is proprietary and not revealed to the public. You just linked a paper from researchers at Google studying the topic of video recommendation. The extent to which it describes youtube's actual algorithm is not at all obvious.

Some researchers at tiktok's parent company released a paper on a recommender system called Monolith here: https://arxiv.org/abs/2209.07663.

I'm not sure its actually what tiktok is using, but they do say that "Monolith has successfully landed in the BytePlus Recommend product".

Also interested been looking into this myself. Some materials of potential interest:

​

Official TikTok blog post describing with very limited detail how it works: How TikTok recommends videos #ForYou

​

Papers:

• An Empirical Investigation of Personalization Factors on TikTok (2022) - sock puppet methodology to identify the parameters and their strength in influencing the algo
• Analysis on the “Douyin (Tiktok) Mania” Phenomenon Based on Recommendation Algorithms (2021)
• Trick and Please. A Mixed-Method Study On User Assumptions About the TikTok Algorithm (2021)
• Leveraging Rights of Data Subjects for Social Media Analysis: Studying TikTok via Data Donations (2023) - may be of interest

​

WSJ did a video on them trying to reverse engineer the algorithm, not too technical though Investigation: How TikTok's Algorithm Figures Out Your Deepest Desires

Some blogs I came across that may or may not be reliable:

• Why TikTok made its user so obsessive? The AI Algorithm that got you hooked.
• The App That Knows You Better than You Know Yourself: An Analysis of the TikTok Algorithm

Typically, large companies don't use a single model, but rather a large number of different models, all performing different tasks (recommending, filtering, etc). It would very difficult to describe the complete recommendations pipeline (i.e from user request to final candidate) in a single academic paper.

These are their features:

https://www.cs.princeton.edu/courses/archive/spring21/cos598D/icde_2021_camera_ready.pdf

The paper also references older neural network architectures used in late stages of the recsys stack.

Whatever you can read are outdated. They don't reveal what they actually use. They are rumored to have the best recommendation system.