Toward recognizing Social Media Recommenders under Absent Recommendations: a Graph Neural Network-based Approach

Sabrina Guidotti, Gregor Donabauer, Davide Taibi, Giuseppe Vizzari, Udo Kruschwitz, Dimitri Ognibene

Keywords: social media, recommender systems, graph neural networks, algorithm auditing, transparency, user behavior modeling

TL;DR: Social media platforms shape discourse but keep recommender logs opaque, making impact assessment hard. We formalize the problem and propose SM-ARR-G that tests simulated infospheres to infer the hidden recommender.

Abstract:

Recommender algorithms shape public discourse on social media, often intensifying polarization and accelerating misinformation. Their opacity and integration within social networks make assessing their true impact exceptionally difficult.

We introduce the problem of Social Media Recommenders Recognition under Absent Recommendations, which accounts for the complexity induced by platforms not publicly releasing interaction logs and algorithmic details, which hinders detecting specific misbehaviors, e.g., dark patterns. We also present a proof-of-concept implementation, SM-ARR-G (Social Media Automatic Recommender Recognition through Graph Neural Networks), a graph neural network (GNN) framework designed to identify which recommendation algorithm is shaping interactions during a given period, without requiring access to the platform’s internal recommendation logs. SM-ARR-G learns to forecast user actions by combining their past behavior with candidate "infospheres", simulated content exposure patterns generated by alternative recommender models. The candidate that produces the most accurate generalization performance is taken as the best explanation of the observed dynamics.

Our evaluation draws on the DBLP-Citation-Network V14 dataset, chosen for its scale and structural richness as a proxy for social media data. While not a typical social media graph, it offers many of the same relational patterns, dense connectivity, influence dynamics, and recommendation-like exposure. By embedding multiple recommender algorithms into this environment, we create controlled yet varied scenarios. This enables us to assess how consistently SM-ARR-G can identify the hidden recommender across different conditions, without relying on access to proprietary platform data.

Our experiments demonstrate that the SM-ARR-G approach is promising and can reliably detect the hidden recommender while highlighting how alternative algorithms alter user interaction patterns. While further improvements are necessary (e.g., handling unknown or evolving recommenders), we expect the framework to complement existing audit strategies, broaden the possibilities for recommender assessment, and aid bias detection.

Area: Modelling and Simluation of Societies (SIM)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1745