An Empirical Investigation of Personalization Factors on TikTok (2201.12271v1)

Abstract: TikTok currently is the fastest growing social media platform with over 1 billion active monthly users of which the majority is from generation Z. Arguably, its most important success driver is its recommendation system. Despite the importance of TikTok's algorithm to the platform's success and content distribution, little work has been done on the empirical analysis of the algorithm. Our work lays the foundation to fill this research gap. Using a sock-puppet audit methodology with a custom algorithm developed by us, we tested and analysed the effect of the language and location used to access TikTok, follow- and like-feature, as well as how the recommended content changes as a user watches certain posts longer than others. We provide evidence that all the tested factors influence the content recommended to TikTok users. Further, we identified that the follow-feature has the strongest influence, followed by the like-feature and video view rate. We also discuss the implications of our findings in the context of the formation of filter bubbles on TikTok and the proliferation of problematic content.

An Empirical Investigation of Personalization Factors on TikTok

The paper "An Empirical Investigation of Personalization Factors on TikTok" offers a rigorous analysis of user-side actions affecting TikTok’s recommendation system (RS), addressing the gap in empirical research regarding the platform's successful algorithmic content distribution. The authors employ sock-puppet audit methodology, creating a controlled environment to mimic realistic user behavior through virtual agents (bots), and analyze the impact of user interactions including likes, follows, and video view rates, alongside language and location preferences.

Methodology and Results

The authors run multiple experiments with bots that simulate user interactions, collecting data on post viewership and engagement attributes such as likes, comments, shares, and views. Across several scenarios, the experimenters manage variables for:

1. Likes: Users liking posts either randomly, based on predefined interests (personas defined by hashtags), or targeting specific creators/sounds.
2. Follows: Users following specific creators to observe variation in feed content.
3. Video View Rate (VVR): Users watching different percentages of video length, from 25% to multiple complete views.
4. Language/Location: Variations in users’ language settings and geographical locations.

Key findings reveal that all examined factors influence TikTok’s RS, with follow-feature exerting the strongest impact, succeeded by VVR and likes. Through their sophisticated control scenarios, the authors establish that inherent noise, introduced by randomization to combat filter bubbles, can be accounted for, becoming negligible compared to effects induced by active user interactions.

Implications for Recommendation Systems

This paper underscores the impact of user-side interactions on personalization algorithms, with particular insight into TikTok's RS design prioritizing user preferences inferred both from explicit actions like follows and implicit behaviors such as viewing patterns. By demonstrating that following specific content creators leads to the most significant divergence in personalized feeds, the authors highlight potential implications for content creators aiming to optimize their visibility.

The research also raises poignant concerns about algorithm-driven filter bubbles and their socio-political ramifications, particularly given TikTok's young user base. The influence of VVR suggests a substantial degree of inferred user interests based solely on video engagement durations, posing risks of inadvertently promoting problematic content among susceptible users.

Future Directions

Future research opportunities include expanding the scope of personalization factors to include other user interactions like comments and shares, and investigating long-term effects of repeated interactions on content diversity. The dynamic nature of TikTok’s RS can be further explored through temporal analyses, assessing algorithm adaptations over extended periods.

Providing comprehensive transparency tools for users, as recommended by the authors, could augment control over personal data utilization and alleviate biases within RS. Collaboration between platform providers and researchers may facilitate the development of groundbreaking solutions to enhance user autonomy and ethical content distribution practices in social media algorithms.

In summary, the paper serves as an insightful foray into the complexities of algorithmic personalization on TikTok, advancing understanding of RS mechanics in contemporary social media landscapes and fostering discourse on ethical digital content dissemination.