Who Makes Trends? Understanding Demographic Biases in Crowdsourced Recommendations

Abhijnan Chakraborty, Johnnatan Messias, Fabrício Benevenuto, Saptarshi Ghosh, Niloy Ganguly, Krishna P. Gummadi

Social media "Trending Topics" serve as powerful amplifiers, transforming niche conversations into national news. However, the democratic promise of these crowdsourced recommendations relies on the assumption that the "crowd" is representative. This research investigates the demographics of the "promoters"—the users who post about a topic before it goes viral—revealing significant systemic biases that effectively silence certain groups.

The Gatekeeping Problem

While previous research has focused on the mechanics of trending algorithms, this study addresses the demographic composition of the inputs. If the individuals who act as initial gatekeepers for trends do not reflect the broader population, the resulting recommendations may suffer from systemic biases, creating a "glass ceiling" for the interests and perspectives of marginalized groups.

Methodology: Quantifying Bias

The researchers developed a robust pipeline to analyze the demographics of approximately 1.7 million US-based Twitter users. Using the Twitter Streaming and REST APIs, they collected 340 million tweets and tracked nearly 12,000 trending topics over three months in 2016.

To infer demographics, the team utilized Face++, a deep-learning face recognition platform, to extract gender, race, and age from profile pictures. The accuracy of this inference was validated against a manually labeled ground truth set, achieving 88% accuracy for gender and 79% for race. Statistical significance was determined using Fisher’s Exact Test and the "80% rule" (a standard for identifying disparate impact) to compare promoter groups against the general Twitter population.

Widespread Demographic Divergence

The study found that for 88.25% of trending topics, the promoter group's demographics were significantly different from the overall Twitter population. This divergence is particularly visible when examining racial distributions across different topics.

Demographic distribution of trend promoters across racial groups. Each dot represents a trending topic, showing how its promoters deviate from the population mean (red dot).

The analysis revealed two distinct patterns in trend promotion:

  1. Niche Interests: Trends driven almost exclusively by specific groups (e.g., gaming by men, fashion by women).
  2. Event Facets: During major events, different demographic groups promote different hashtags, reflecting diverse perspectives on the same occurrence (e.g., #BlackLivesMatter vs. #PoliceLivesMatter).

The "Glass Ceiling" for Content

A critical finding is that "adopters"—those who tweet after a topic has already started trending—are much more demographically representative than the initial "promoters." This suggests that while the trending algorithm successfully exposes topics to a wide audience, the initial "gatekeeping" phase is where the bias is most acute.

The research highlights that women, Black, and Asian users are frequently under-represented in trend promotion. Intersectional analysis reveals even deeper disparities.

Percentage of trends where specific demographic groups are under-represented across (a) primary categories, (b) race-gender intersections, and (c) age-race-gender intersections.

As shown in the data, middle-aged Black females are the most marginalized group in the trend-making process. Despite being active users, their contributions are less likely to be part of the initial surge that triggers a "trending" status.

Conclusion and Future Directions

The findings underscore that Twitter's "crowdsourced" trends are not a neutral reflection of user interest but are shaped by a demographically skewed group of promoters. The authors suggest that future trending algorithms should explicitly account for demographic diversity to ensure a more equitable distribution of attention. While the study is limited by the broad racial categories of current inference models and its focus on Twitter in the US, it provides a crucial framework for auditing bias in recommendation systems.

Summary updated

BibTeX
@inproceedings{chakraborty2017who,
  author = {Abhijnan Chakraborty and Johnnatan Messias and Fabrício Benevenuto and Saptarshi Ghosh and Niloy Ganguly and Krishna P. Gummadi},
  title = {{Who Makes Trends? Understanding Demographic Biases in Crowdsourced Recommendations}},
  booktitle = {ICWSM},
  year = {2017},
  url = {https://openreview.net/forum?id=EvS6CwAH7uc},
}
Figure