New CoSimRec framework reveals how coordinated accounts exploit recommender feedback loops

A new paper from researchers (Li et al.) introduces CoSimRec, an offline agent-based evaluation framework that models coordinated accounts, dynamic ranking, non-bot responses, and ranking interventions in a shared closed-loop recommender feedback process. The framework introduces the Algorithmic Penetration Rate (APR) metric family, which measures target content's share of non-bot exposure and engagement, lift against matched no-attack baselines, and exposure gained per coordinated interaction. The authors evaluated CoSimRec on MIND, MovieLens, and LastFM datasets using random, popularity-based, feedback-sensitive, MF, and BPR-MF recommenders, with population-scale experiments up to 1,000 users.

The results are stark: random controls show no statistically supported positive penetration, while popularity-based and feedback-sensitive ranking produce significant positive APR-Lift across all six master-worker dataset-recommender settings, reaching as high as 0.4505 on LastFM. Critically, synchronization-aware ranking reduces APR in every corresponding defense setting. This work provides the first systematic measurement of how coordinated activity can be amplified beyond the accounts that initiate it, revealing a fundamental vulnerability in widely-used recommendation algorithms. The framework offers a practical tool for evaluating and designing more robust recommenders against coordinated content manipulation campaigns.