生成式推荐系统综述OA

With the rapid growth of social media content scale,traditional collaborative filtering recommender systems increasingly exhibit limitations in data sparsity and cold start problems.In recent years,the powerful data feature analys-is and content generation capabilities of generative models have brought new development opportunities for recom-mender systems.This paper systematically reviews the technical frameworks and research progress in generative recom-mender systems,focusing on five key aspects:feature tokenization methods,core model architectural designs,main-stream evaluation protocols and typical application scenarios.Through comparative analysis and literature review,we demonstrate that generative recommender systems significantly outperform conventional approaches in recommenda-tion accuracy,personality,and scenario adaptability.The study further identifies critical challenges including computa-tional overhead,privacy risks,and standardization of evaluation metrics.Practical solutions and future research direc-tions are proposed to address these challenges,breaking the cognitive bottleneck of generative recommender systems.