Bursting the Filter Bubble with Knowledge Graph Inversion

As recommender systems increasingly mediate our online experiences, ethical concerns arise as they often reinforce filter bubblesnarrow content loops that isolate users from diverse perspectives.In this paper, we propose a novel approach to mitigating filter bubbles by combining personalized knowledge graph (KG) completion with federated learning (FL) and KG edge inversion techniques.User-specific KGs are constructed from private interaction data and remain entirely on-device, preserving privacy while enabling the system to learn fine-grained preferences.A central recommendation model is trained collaboratively via FL, allowing KG completion without exposing sensitive user data.To disrupt filter bubbles, we introduce KG edge inversion, a method that strategically inverts selected relations in the user's KG to simulate alternative viewpoints.The model then produces plausible yet diverse recommendations, effectively guiding users beyond their typical content landscape.Our approach offers a transparent and controllable framework for promoting opinion diversity and combating the isolating effects of algorithmic personalization.