用强化学习设计防成瘾推荐系统，平衡用户健康与平台收益。
Balancing Benefits and Risks: RL Approaches for Addiction-Aware Social Media Recommenders
- 引入非平稳、非马尔可夫动态建模，适应用户行为多样性。
- 区分轻度与重度使用状态，提升对成瘾风险的识别能力。
- 适合关注心理健康与可持续平台设计的研究者与工程师。
社交媒体为用户提供信息获取、社交互动和娱乐机会，但其成瘾性带来过度使用及心理行为问题。本研究探索在保留平台价值与经济可持续性的前提下，缓解强迫性使用的方法，聚焦于促进均衡使用的推荐系统。通过分析用户内在差异与环境交互下的行为模式，提出基于强化学习的成瘾建模框架，构建能适应用户偏好的推荐系统，引入非平稳与非马尔可夫动态；采用用户与推荐器差异化状态表征，捕捉复杂交互；区分轻度与重度使用场景，克服传统强化学习在区分长期使用与健康使用上的局限。模拟实验揭示模型决策（MB）与无模型（MF）方法如何受环境动态影响用户行为与成瘾倾向。结果表明推荐系统在塑造用户行为中起关键作用，支持伦理化、自适应推荐设计，推动可持续社交媒体生态发展。
原文摘要 · Abstract (English)
Social media platforms provide valuable opportunities for users to gather information, interact with friends, and enjoy entertainment. However, their addictive potential poses significant challenges, including overuse and negative psycho-logical or behavioral impacts [4, 2, 8]. This study explores strategies to mitigate compulsive social media usage while preserving its benefits and ensuring economic sustainability, focusing on recommenders that promote balanced usage. We analyze user behaviors arising from intrinsic diversities and environmental interactions, offering insights for next-generation social media recommenders that prioritize well-being. Specifically, we examine the temporal predictability of overuse and addiction using measures available to recommenders, aiming to inform mechanisms that prevent addiction while avoiding user disengagement [7]. Building on RL-based computational frameworks for addiction modelling [6], our study introduces: - A recommender system adapting to user preferences, introducing non-stationary and non-Markovian dynamics. - Differentiated state representations for users and recommenders to capture nuanced interactions. - Distinct usage conditions-light and heavy use-addressing RL's limitations in distinguishing prolonged from healthy engagement. - Complexity in overuse impacts, highlighting their role in user adaptation [7]. Simulations demonstrate how model-based (MB) and model-free (MF) decision-making interact with environmental dynamics to influence user behavior and addiction. Results reveal the significant role of recommender systems in shaping addiction tendencies or fostering healthier engagement. These findings support ethical, adaptive recommender design, advancing sustainable social media ecosystems [9, 1]. Keywords: multi-agent systems, recommender systems, addiction, social media
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。