M2Rec
收藏资源简介:
本文介绍了一种名为M2Rec的序列推荐系统框架,旨在解决现有推荐系统在效率和多尺度模式识别方面的局限性。M2Rec整合了多尺度Mamba、傅里叶分析和大型语言模型(LLM),并引入了自适应门控机制。该框架通过傅里叶变换增强Mamba,明确地在频率域中建模周期性模式,并利用LLM提取丰富的语义信息。最后,通过学习门控机制动态平衡不同模态的特征,实现多模态融合。实验表明,M2Rec在保持Transformer基线模型20%的推理速度的同时,比现有的Mamba模型提高了3.2%的命中率@10。该框架适用于电子商务、流媒体平台等领域,旨在解决用户交互序列中多尺度周期性模式识别问题。
This paper introduces a sequential recommendation framework named M2Rec, which aims to address the limitations of existing recommendation systems in terms of efficiency and multi-scale pattern recognition. M2Rec integrates multi-scale Mamba, Fourier analysis, and Large Language Models (LLMs), and introduces an adaptive gating mechanism. This framework enhances Mamba via Fourier transform, explicitly models periodic patterns in the frequency domain, and leverages LLMs to extract rich semantic information. Finally, it achieves multimodal fusion by learning the gating mechanism to dynamically balance features from different modalities. Experiments demonstrate that M2Rec maintains 20% of the inference speed of the Transformer baseline model, while improving the Hit Rate@10 by 3.2% compared to existing Mamba-based models. This framework is applicable to domains such as e-commerce and streaming media platforms, and aims to solve the problem of multi-scale periodic pattern recognition in user interaction sequences.




