Stenobird

Episode

Why Am I Seeing This?

Podcast
Data Skeptic
Published
Sep 8, 2025
Duration seconds
2976
Processing state
processed
Canonical source
https://dataskeptic.com/blog/episodes/2025/why-am-i-seeing-this
Audio
https://pscrb.fm/rss/p/mgln.ai/e/35/traffic.libsyn.com/secure/dataskeptic/Dimitri_Gregor_Sabrina_Podcast_with_Ads.mp3?dest-id=201630
JSON
/v1/public/podcasts/data-skeptic/episodes/why-am-i-seeing-this
Markdown
/podcast/data-skeptic/why-am-i-seeing-this.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/data-skeptic/episodes/why-am-i-seeing-this/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/data-skeptic/why-am-i-seeing-this.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Researchers present a 'recommender neutral user model' designed to simulate social media dynamics without needing proprietary platform logs. The discussion explores how optimizing for engagement-based reward functions can inadvertently drive polarization and echo chambers.

Topics

  • Recommender Systems
  • Social Media Algorithms
  • Graph Neural Networks
  • Algorithmic Bias
  • Echo Chambers
  • User Behavior Modeling
  • Reinforcement Learning
  • Data Privacy

Highlights

  • Main idea: The recommender neutral user model uses a graph-neutral approach to learn user behavior without being biased by specific platform objectives
  • Failure mode: Optimizing for engagement-based reward functions can inadvertently promote polarizing content because high-conflict interactions drive time-on-platform
  • Practical takeaway: Researchers can use synthetic data and computational models to simulate how different recommendation strategies might steer users even without private company logs
  • Challenge: Studying social media impact is difficult because most high-quality datasets are proprietary and lack essential social connection data
  • Core tension: There is a significant gap between the well-intentioned goal of surfacing relevant content and the unintended consequence of creating information silos

Chapters

  1. 1:00 The Recommender Neutral User Model: Introduction to a graph-neutral approach that simulates user behavior without needing privileged access to platform logs.
  2. 4:25 The Scarcity of Social Datasets: A look at why existing datasets like Amazon reviews are insufficient for studying social connections and network dynamics.
  3. 8:35 Modeling Social Dynamics: Discussing the necessity of using computational models to study social media when platform data is kept private by companies.
  4. 12:00 The Difficulty of Predicting Neural Models: Why having code or pseudocode is insufficient for predicting the behavior of large-scale neural recommendation engines.
  5. 16:10 Learning Baseline Behavior: The goal of minimizing the influence of specific recommenders to establish a baseline of user behavior.
  6. 20:05 Learning Directly from Graph Data: How the model learns to behave like a user under different recommendation environments using graph-based learning.
  7. 23:45 Experimental Results on Bias: Comparing user decision-making when exposed to fair versus biased recommendation distributions.