Leveraging Rights of Data Subjects for Social Media Analysis: Studying TikTok via Data Donations

Savvas Zannettou, Olivia-Nemes Nemeth, Oshrat Ayalon, Angelica Goetzen, Krishna P. Gummadi, Elissa M. Redmiles, Franziska RoesnerarXiv 2023 · OpenReview

Introduction

Short-form video (SFV) platforms, led by TikTok, have fundamentally altered digital consumption patterns. Despite their ubiquity, these platforms remain "black boxes" to the research community. Traditional auditing methods often rely on "sock-puppet" bots, which fail to capture the evolving complexity of real human behavior. This study addresses this gap by leveraging GDPR-mandated "data donations" from 347 real users to perform a large-scale audit of TikTok’s recommendation engine, specifically examining how engagement and attention change over a user's tenure on the platform.

Methodology

The researchers developed the Social Media Donator (SMD), a system that allows users to contribute their personal data packages—provided by TikTok under "Right of Access" laws—in a privacy-preserving manner.

Data and Participants

The dataset consists of 9.2 million video views and 1.1 million "likes" from a diverse cohort. Notably, 52% of participants were from Africa and 32% from North/Central America, with 91% of the group aged 34 or younger. The data spans from July 2020 to February 2022.

Inferring Viewing Behavior

Because TikTok’s data exports do not explicitly state "watch time," the researchers employed a two-component Gaussian mixture model to analyze video start timestamps. This allowed them to infer actual viewing durations versus instances where a user simply closed the app or took a break.

Comparison of actual video durations versus inferred viewing durations, illustrating the prevalence of content skipping.

Key Insights and Results

The analysis reveals a significant evolution in user behavior over time, characterized by a doubling of usage volume alongside stable attention metrics.

The Usage Growth and Attention Paradox

As users spend more time on the platform, their daily consumption increases dramatically. Within the first 120 days, the average daily time spent and the number of videos watched doubles, growing from approximately 29 to 50 minutes. However, "attention"—the percentage of a video watched to completion—remains flat at roughly 45%. This suggests that TikTok's algorithm excels at maintaining a high-volume stream of content that keeps users scrolling, even if they skip more than half of the videos recommended to them.

Growth in the average and median daily time spent watching videos over a 120-day period.

The Dominance of the Interest Graph

The study confirms that TikTok functions primarily as a discovery engine rather than a social network. Content from followed accounts represents only a median of 10% of views, a figure that remains stable even as users follow more accounts. Furthermore, users watch non-followed videos to completion more often than those from followed accounts. This is likely because the algorithm prioritizes "global hits"—videos with significantly higher play counts (median 755K vs. 85K for followed accounts)—which possess higher viral potential and "stickiness."

Behavioral Reinforcement

While attention remains stable, interaction metrics like "liking" increase over time. Users were found to be 1.5x to 2x more likely to like a video after 120 days of use compared to their first day. The researchers hypothesize that TikTok utilizes an intermittent reward schedule, mixing high-engagement "positive reinforcement" with content that is quickly skipped, a psychological driver known to foster addictive behavior.

Discussion and Limitations

While this study provides unprecedented insight into real-world TikTok usage, it is limited by its sample size and potential survivorship bias, as 17% of videos in the logs were deleted or made private before metadata could be scraped. Future work should utilize the SMD tool to conduct cross-platform comparisons with YouTube Shorts and Instagram Reels to determine if these engagement patterns are inherent to the short-form video format itself.

Summary updated

BibTeX
@misc{zannettou2023leveraging,
  author = {Savvas Zannettou and Olivia-Nemes Nemeth and Oshrat Ayalon and Angelica Goetzen and Krishna P. Gummadi and Elissa M. Redmiles and Franziska Roesner},
  title = {{Leveraging Rights of Data Subjects for Social Media Analysis: Studying TikTok via Data Donations}},
  year = {2023},
  url = {https://openreview.net/forum?id=ryp2yvwvqw},
}
Figure