Data Skeptic Why Am I Seeing This?
13 snips
Sep 8, 2025 Dimitri Ognibene, Director of the Biconnaire Club at the University of Milano Bicocca, dives into the intricate world of social media recommender systems. The conversation highlights the lack of accessible exposure data and the resulting challenges in understanding algorithmic influence. They discuss innovative solutions like the 'recommender neutral user model' to address biases and promote data privacy. Insightful reflections on user perceptions, especially among teenagers, reveal how opaque algorithms can lead to polarization and dissatisfaction, underscoring the need for transparency.
AI Snips
Chapters
Transcript
Episode notes
Hidden Exposure Distorts User Preferences
- Missing exposure data skews our interpretation of user preferences and behavior.
- Recommender influence can make observed clicks reflect algorithm choices, not true user intent.
Datasets Lack Recommendation Labels
- Public datasets rarely include both interaction histories and the recommendations that produced them.
- That forces researchers to use proxies like citation graphs to approximate social media structure.
Collect Social And Exposure-Rich Data
- Collect richer datasets including social connections and exposure when possible.
- Use platforms like the Fediverse and crowdsourced uploads to build usable anonymized corpora.