Mediating The Marginal: A Quantitative Analysis of Curated LGBTQ+ Content on Instagram

Research Background and Issues

• What problems or challenges did the authors identify?

  1. Instagram's content recommendation algorithm exhibits significant bias within the LGBTQ+ community, particularly in terms of representation of race and gender expression.
  2. While content featuring darker skin tones receives higher user engagement, it is significantly underrepresented and excluded from "popular" #gay content. In contrast, content featuring lighter skin tones and conforming to "hyper-masculine" aesthetics is prioritized by the algorithm.
  3. The lack of clear metrics explaining algorithmic recommendation decisions reduces users' sense of participation and control in personalized recommendations.

• Why is this issue important?

  1. The visual culture of social media plays a crucial role in identity formation and community building, especially for the LGBTQ+ community. Algorithmic bias may further amplify marginalization and social inequality present in real life.
  2. The representation of visual culture has profound implications for identity, social status, and individual futures. Misrepresentation or exclusion of marginalized groups can lead to severe social and psychological harm.
  3. The LGBTQ+ community relies on these digital spaces not only for self-expression but also to find belonging and support. Therefore, algorithmic choices directly impact their safety and well-being both online and offline.

• Research Motivation and Related Work

  1. This study builds on discussions about algorithmic bias and representational harm, particularly in the intersectional domains of race, gender, and gender expression.
  2. Existing technical research lacks systematic exploration of racial and gender biases in social media platform design. This study extends the field through data analysis and GAN (Generative Adversarial Network) techniques.

Solutions

• What methods or solutions did the authors propose?

  1. The authors employed descriptive statistics and computer vision methods, such as skin tone quantification, visual clustering analysis, and Generative Adversarial Networks (GAN), to quantitatively and qualitatively analyze aesthetic trends in Instagram's recommended content.
  2. A virtual Instagram account labeled as a "gay user" was created to conduct a nominal audit of algorithmic behavior and explore how recommendations are made based on a single identity dimension (sexual orientation).

• What are the innovative aspects of the solution?

  1. The study introduced GANs for the first time to generate "normalized" visual samples, revealing which aesthetic or cultural labels are marginalized within the algorithm's learned content.
  2. It advanced academic research by moving beyond user report mechanisms to interdisciplinary methods combining computational and cultural analysis, uncovering biases in visual algorithms and their far-reaching impacts.

• What are the implementation steps and key technologies used?

  1. Data Collection and Processing: Public Instagram content was scraped using the "gay" tag and the Explore page, filtering and manually reviewing nearly 37,000 post images.
  2. Skin Tone Analysis: Facial region data was collected, using Multi-Task Convolutional Neural Networks (MTCNN) to extract skin tones, and HSV and LAb* color spaces for color quantification and matching.
  3. Visual Clustering and Generation: K-Means clustering and CLIP image encoders were used to extract 200-dimensional feature vectors for clustering. A Wasserstein GAN (WGAN) was trained to generate normalized content.
  4. User Behavior and Content Analysis: Engagement metrics such as likes and comments were analyzed to assess the distribution and engagement levels of content across tags and skin tones.

Research Findings

• What specific findings were achieved?

  1. Content featuring darker skin tones received higher user engagement (likes and comments) compared to lighter skin tones, but its proportion in "popular" content was significantly reduced, with darker skin representation decreasing by 7.27%.
  2. Instagram's Explore page predominantly recommended content to "gay" users that focused on white, masculine, fitness-oriented, and affluent aesthetics, while Black-related content was largely confined to cultural symbols such as sports, music, and fashion.
  3. GAN-generated visual content confirmed the normalization bias of the recommendation algorithm, with generated samples heavily emphasizing hyper-masculine white imagery, further erasing marginalized aesthetics.

• What advantages does it have compared to existing solutions?

  1. The study employed systematic, data-driven analysis combined with visual generation techniques, rather than relying solely on user surveys or subjective analysis, revealing more nuanced aesthetic biases in recommendation algorithms.
  2. It explored how recommendation algorithms become imbalanced across multidimensional intersectional identities (gender, race, sexual orientation), enhancing technical and social understanding of algorithmic accountability.

• What were the experimental or evaluation results?

  1. The average skin brightness value of "popular" #gay content was 7.7% higher than that of "recent" content.
  2. GAN-generated images clearly demonstrated the absence of darker skin tones and non-normative aesthetics in the primary data samples.
  3. Cluster analysis showed that Black representation was often confined to "hyper-sexualized" cultural spaces, severely disconnected from real-life contexts.

• Limitations and Future Directions

  1. The study was based solely on a virtual account with a single gender and sexual orientation dimension. Future research could expand to include more identity combinations, such as transgender or non-binary identities.
  2. GAN training and generation still have room for improvement. Testing different types of generators, such as StyleGAN, could explore other potential aesthetic themes.
  3. Combining user interviews or focus group studies could further reveal how algorithmic bias is perceived and resisted by real users.

Conclusion

This paper systematically examined biases in Instagram's recommendation algorithm for LGBTQ+ content, combining computer vision and cultural analysis to uncover the "invisibility" of marginalized identities. The research not only provides new methodological insights for algorithm design and social justice discussions but also suggests possibilities for more inclusive and transparent content curation directions.