Attraction and retention dynamics in recommendation graphs: a cross-dataset analysis using uniform and degree-biased random walks

Recommendation systems can unintentionally confine users to narrow regions of the recommendation graph, creating structural ’traps’ that influence user discovery and persistence. While previous research often treats these traps as single phenomena, this study decouples their dynamics into two distinct behaviors: Attraction (the ease of discovery from varying topological distances) and Retention (the persistence of traversal once inside). We study these dynamics in YouTube’s Watch-Next recommendation network using hop-aware random-walk simulations under both neutral (uniform) and popularity-skewed (degree-biased) traversal, combined with multiple structural extraction methods. Across three socio-political datasets (China–Uyghur discourse, Cheng Ho propaganda conflict, and discourse surrounding the 2024 Trump assassination attempt), we observe a consistent structural specialization: Focal Structure Analysis (FSA) tends to surface highly accessible gateway structures that act as entry interfaces, whereas community partitions tend to capture dense regions that sustain prolonged within-region traversal. By distinguishing these roles, our framework provides a diagnostic lens for auditing recommendation graphs and identifying topological conditions under which algorithmic confinement is more likely, with patterns that replicate across distinct topical ecosystems.