Quantifying the Potential to Escape Filter Bubbles: A Behavior-Aware Measure via Contrastive Simulation

TL;DR

This work tackles the persistence of filter bubbles in recommender systems by introducing Bubble Escape Potential (BEP), a behavior-aware metric that decouples algorithmic confinement from user preference. BEP is computed via a contrastive simulation framework using LLM-driven positive and negative user behaviors to quantify how easily users can escape narrow exposure. The study demonstrates a trade-off between predictive accuracy and bubble severity across models, with sequential models achieving higher accuracy but stronger confinement, and shows that mild randomization is not an effective long-term solution. Overall, BEP provides a principled, actionable diagnostic tool for designing more inclusive recommender systems and informing mitigation strategies beyond traditional diversity metrics.

Abstract

Nowadays, recommendation systems have become crucial to online platforms, shaping user exposure by accurate preference modeling. However, such an exposure strategy can also reinforce users' existing preferences, leading to a notorious phenomenon named filter bubbles. Given its negative effects, such as group polarization, increasing attention has been paid to exploring reasonable measures to filter bubbles. However, most existing evaluation metrics simply measure the diversity of user exposure, failing to distinguish between algorithmic preference modeling and actual information confinement. In view of this, we introduce Bubble Escape Potential (BEP), a behavior-aware measure that quantifies how easily users can escape from filter bubbles. Specifically, BEP leverages a contrastive simulation framework that assigns different behavioral tendencies (e.g., positive vs. negative) to synthetic users and compares the induced exposure patterns. This design enables decoupling the effect of filter bubbles and preference modeling, allowing for more precise diagnosis of bubble severity. We conduct extensive experiments across multiple recommendation models to examine the relationship between predictive accuracy and bubble escape potential across different groups. To the best of our knowledge, our empirical results are the first to quantitatively validate the dilemma between preference modeling and filter bubbles. What's more, we observe a counter-intuitive phenomenon that mild random recommendations are ineffective in alleviating filter bubbles, which can offer a principled foundation for further work in this direction.

Figures (6)

• Figure 1: Overview of the Bubble Escape Potential ($\mathsf{BEP}$) evaluation framework. We simulate two groups of users—positive and negative—who share user profiles (conformity and tastes). By modifying the user's activity, diversity and different attitudes in prompts, we generate positive users and negative users. Both groups interact with a selected recommendation system across multiple rounds, receiving page-by-page recommendations and making choices. By comparing their exposure, we quantify the system's Bubble Escape Potential ($\mathsf{BEP}$).
• Figure 2: Change in diversity under different settings: A-LightGCN, B-TiCoSeRec, and C-LightGCN(positive) in ml-1m. Users' positive behaviors in LightGCN surpasses the gap between different recommendation systems.
• Figure 3: The variation in the diversity of recommendations received by users under different behavioral patterns changes over time as the simulation rounds increase in different recommendations in ml-1m. Blue lines represent positive users, while red lines represent negative users. Each of these pictures represents a different recommendation system: (a) MF. (b) LightGCN. (c) DiffRec. (d) TiCoSeRec.
• Figure 4: The accuracy and the Bubble Escape Potential ($\mathsf{BEP}$) corresponding to different recommendation systems under ml-1m and Amazon-Books: (a) HR@20 vs. $\mathsf{BEP}$ in ml-1m. (b) NDCG@20 vs. $\mathsf{BEP}$ in ml-1m. (a) HR@20 vs. $\mathsf{BEP}$ in Amazon-Books. (b) NDCG@20 vs. $\mathsf{BEP}$ in Amazon-Books. This set of results reveals the Accuracy-Bubble Dilemma.
• Figure 5: The accuracy and the Bubble Escape Potential ($\mathsf{BEP}$) corresponding to different recommendation systems after adding randomization under ml-1m: (a) LightGCN. (b) TiCoSeRec.