CoSimRec: Measuring Coordinated-Content Penetration in Recommender Feedback Loops

Recommender systems are not static; they operate in feedback loops where user behavior influences future rankings, which in turn shape future user behavior. While previous research has examined how fake profiles can manipulate item rankings, it remains unclear whether such coordinated activity successfully penetrates the feeds of ordinary, non-coordinated users. This paper asks: to what extent does coordinated content gain exposure and engagement among non-bot users within a dynamic recommender feedback loop?

The authors introduce CoSimRec, an offline agent-based evaluation framework that simulates the interaction between coordinated accounts, recommender systems, and non-bot users. Unlike static robustness evaluations that only measure target rank, CoSimRec models the entire closed-loop process. It uses the Algorithmic Penetration Rate (APR) metric family to quantify the success of coordinated content by measuring its share of non-bot exposure and engagement, specifically calculating the 'lift' against matched no-attack baselines to isolate the effect of the coordination itself.

The study evaluates CoSimRec across three datasets (MIND, MovieLens, and LastFM) using various ranking strategies. The results show that random recommendation controls produce no statistically significant penetration. However, popularity-based and feedback-sensitive ranking algorithms consistently produce significant positive APR-Lift, meaning coordinated content successfully reaches ordinary users. The researchers also demonstrate that synchronization-aware ranking interventions can effectively reduce this penetration, providing a potential defense mechanism against coordinated manipulation.

This work shifts the focus of recommender robustness from simple rank manipulation to the broader problem of algorithmic amplification. By providing a framework to measure how coordinated content 'leaks' into ordinary user experiences, CoSimRec offers a way for researchers and platform designers to evaluate the vulnerability of their systems to coordinated influence campaigns in a controlled, reproducible environment.