How Predicted Links Influence Network Evolution: Disentangling Choice and Algorithmic Feedback in Dynamic Graphs

A team of researchers has published a significant paper titled 'How Predicted Links Influence Network Evolution: Disentangling Choice and Algorithmic Feedback in Dynamic Graphs' on arXiv. The work addresses a critical blind spot in how we understand social networks and recommendation engines. Currently, platforms use link prediction models to suggest connections (like 'People You May Know'), but their long-term impact on the network's structure is typically assessed from static snapshots. This conflates natural user behavior with the amplifying feedback loops created by the algorithms themselves. The authors argue that observed trends like homophily—the tendency for similar users to connect—are not purely organic but are significantly shaped by these AI systems.

The researchers propose a novel temporal framework based on multivariate Hawkes processes, a type of statistical model for analyzing event sequences over time. This allows them to separate two key forces: intrinsic user choice and algorithmic feedback. A central contribution is their 'instantaneous bias measure,' derived from interaction intensities, which captures reinforcement dynamics as they happen, moving beyond cumulative metrics. The paper provides a theoretical analysis of the stability of these AI-influenced dynamics and validates the measure through experiments across different prediction strategies. This work provides a formal, quantifiable method to audit how recommendation algorithms actively steer the evolution of digital communities, revealing their hidden role in shaping online ecosystems.