| 摘 要: 近年来,序列推荐系统在电商与内容平台中受到广泛关注,它的核心目标是建模用户行为的时序依赖以提升推荐效果。将Mamba与大语言模型(LLM)等新兴序列建模纳入体系,从数据类型视角提出覆盖ID、文本与混合特征的3层分类框架,总结序列推荐模型从基于ID的离散建模到文本语义增强再到多模态融合的技术演进。最后对代表性的基础模型进行系统复现与对比实验。在ML-1M 数据集上的实验结果表明,基于新型 Mamba架构的 Mamba4Rec相较于基于Transformer架构的BERT4Rec,HR@10提升约10%,每轮训练耗时减少约65%,显示它在高效捕获长短期依赖方面的优势,为后续模型融合设计与特征建模提供参考。 |
| 关键词: 序列推荐系统 文本增强 多模态融合 Mamba ID建模 |
|
中图分类号:
文献标识码: A
|
| 基金项目: :国家重点研发计划项目资助(2022YFB3105401) |
|
| Review of Research on Sequential Recommendation Systems |
|
JI Zhongqing, TIAN Shaocong, ZHOU Renjie, ZHANG We
|
(School of Computer Science, Hangzhou Dianzi University, Hangzhou 310018, China)
929779718@qq.com; chenyf@njupt.edu.cn; 1552817181@qq.com; zhangw@njupt.edu.cn
|
| Abstract: In recent years, sequential recommendation systems have received widespread attention in e-commerce and content platforms, aiming to model users’temporal behavior dependencies to improve recommendation performance. This paper incorporates emerging sequential modeling paradigms, including Mamba and Large Language Model (LLM), into a framework. From the perspective of data types, this paper proposes a three-level classification covering ID-based modeling, tex-t enhanced modeling, and multimodal fusion, and summarizes the evolution of sequential recommendation models from ID-based discrete modeling to tex-t enhanced modeling and then to multimodal fusion. Finally, systematic reproduction and comparative experiments are conducted on representative basic models. Experimental results on the ML-1M dataset show that, Mamba4Rec with the novel Mamba structure achieves about 10% improvement in HR@10 and reduces pe-r epoch training time by approximately 65% compared with the Transforme-r based BERT4Rec, demonstrating its advantage in efficiently capturing long-and shor-t term dependencies and providing reference for future model integration design and feature modeling. |
| Keywords: sequential recommendation systems text enhanced modeling multimodal fusion Mamba ID-based modeling |
