Relational Diversification-based Anti-Mapping for Concealing User Preference

Personalized recommendation engines often rely on preference mapping, which infers user interests from historical behaviors and social graph structures. Such mapping can trigger a content homogenization loop, gradually limiting users’ exposure to diverse content and resulting in information cocoons. While existing approaches focus on diversifying content recommendations, they often overlook the role of social topology in shaping information exposure. This paper seeks to examine how the structure of social connections shapes users’ exposure to diverse information and to develop measurable strategies for reducing information homogenization. We find that users with heterogeneous, strategically chosen links have diverse information exposure, which can break the information cocoon. To quantify the homogeneity of the environment, we introduce a Diversity-Scale score that quantifies both the diversity and the scope of content exposure by each user. We then establish a correlation between preference features and social connection structure, independent of explicit personal preference data, and demonstrate that strategies with diverse connections significantly weaken this correlation. We validate our approach on two real-world platforms with distinct interaction patterns (Yelp and Steam), showing that constructing diverse and dynamically heterogeneous graph structures as an anti-mapping approach is a practical and lightweight mechanism to resist preference fingerprinting and its resulting content homogeneity.