We recently joined the Data Skeptic episode “Why Am I Seeing This?” to discuss our research on identifying and understanding hidden social-media recommender systems. The conversation walks through our approach, why exposure data is often missing, and what that means for researchers, platforms, and policy.
Why this matters
Recommenders quietly shape what we see—and what we believe. Because platforms rarely expose the underlying algorithms or full exposure logs, it’s hard to audit their real effects on behavior, polarization, or misinformation.
Our work tackles this gap by inferring the kind of recommender likely at play from observable network signals and user actions, without needing access to users’ impression data. That lets us reason about downstream impacts and compare “what if” scenarios responsibly.
What we contributed (and talked about on the show)
- Automatic Recommender Recognition with Graph Neural Networks (GNNs): we train a Recommender-Neutral User model to strip out platform-specific effects, then generate recommender-hypothesis synthetic datasets to see which algorithm best explains the observed world.
- Academic networks as an ethical testbed: we use scholarly collaboration networks as a proxy to study recommendation dynamics and run controlled “infosphere vs. behavior” simulations with GNNs.
- From recognition to mitigation: in parallel, our collaborators design collective-well-being-aware recommenders and a Virtual Learning Companion that coaches teens to handle toxic and misleading content—pairing automated detectors with educator input to improve community well-being.
A bit of motivation from the referenced work
- Opacity shouldn’t stall safety: if we can recognize recommender behavior from traces in the network, we can study systemic effects even when impressions are missing—turning a long-standing data-access problem into a modeling problem.
- Simulate before you legislate: synthetic, hypothesis-driven datasets let researchers and policymakers ask counterfactuals (“What if the platform used X instead of Y?”) without experimenting on real users.
- Well-being as an objective: when we optimize only for engagement, platforms can amplify popularity and outrage biases; framing recommendation around collective well-being offers a concrete alternative.
Papers & resources
- Why am I seeing this? Towards recognizing social media recommender systems with missing recommendations (2025).
- Modeling Social Media Recommendation Impacts Using Academic Networks: A Graph Neural Network Approach (2024).
- Challenging Social Media Threats using Collective Well-being-Aware Recommendation Algorithms and an Educational Virtual Companion (2023).
- Data Skeptic episode: “Why Am I Seeing This?”
- COURAGE project
People involved
This work and the podcast are the result of a collaborative effort. The episode features Dr. Sabrina Guidotti, Dr. Gregor Donabauer, and Prof. Dimitri Ognibene. We also build on contributions from Prof. Udo Kruschwitz, Dr. Davide Taibi, Prof. Giuseppe Vizzari, Dr. Simone Somazzi, and Dr. Sabrina Patania—with connected efforts through the COURAGE initiative involving Prof. Davinia Hernández Leo, Emily Theophilou, and Prof. Sabrina Eimler, among others.
Contact
- Dr. Sabrina Guidotti — s.guidotti2@campus.unimib.it
- Dr. Gregor Donabauer — gregor.donabauer@ur.de
- Prof. Dimitri Ognibene — dimitri.ognibene@unimib.it
- Dr. Davide Taibi — davide.taibi@cnr.it
- Dr. Sabrina Patania — sabrina.patania@unimib.it
- Prof. Udo Kruschwitz — Udo.Kruschwitz@ur.de
- Prof. Davinia Hernández Leo — davinia.hernandez-leo@upf.edu
- Prof. Giuseppe Vizzari — giuseppe.vizzari@unimib.it
- Prof. Sabrina Eimler — sabrinaeimler@gmail.com
- Dr. Emily Theophilou — emily.theophilou@upf.edu