Quantifying Algorithmic Entrapment in YouTube Recommendation Network: A Composite Measure of Structure and Persuasion

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Algorithmic recommendation systems shape information exposure and engagement, yet scalable measures of “content traps” rarely integrate both network structure and content-level influence. This study proposes TrapIntensity, a sociotechnical construct for quantifying content traps in three YouTube recommendation networks case studies by combining structural entrapment and persuasive intensity. Structural entrapment is estimated through hop-aware attraction and retention dynamics using random-walk simulations over multi-hop recommendation graphs, with candidate trap regions extracted using Collective Influence–Focal Structure Analysis (CI-FSA) and assessed via resiliency-based fragmentation tests. Persuasive intensity is measured using a structured large language model pipeline grounded in four persuasion theories applied to video titles, descriptions, and transcripts. Using engagement outcomes as external behavioral indicators, we compare an equal-weight composite against a context-sensitive weighting scheme derived from Cliff’s delta. Results show that dataset-specific weighting improves engagement alignment and that the dominant driver of trap intensity varies across contexts, alternating between structural lock-in and persuasive reinforcement. The work contributes an auditable measurement framework for recommendation system governance.

Links and sources

Comments