Scholarlink

<p>Sequential recommendation systems’ (SRSs) main goal is to predict what users could be interested in in the future by capturing their dynamic preferences based on their past interaction records. Although existing Transformer-based models perform well in this domain, the quadratic time complexity of their attention mechanism limits their efficiency in handling long interaction sequences. Recently, Mamba has been introduced into sequential recommendation systems, demonstrating impressive capabilities. However, as Mamba is a unidirectional model that only records previous user actions, it cannot leverage future contextual information. Our proposed GruMambaRec retains this unidirectional design, combining selective GRU and gating mechanisms to capture long-range dependencies, without incorporating future information, unlike bidirectional models such as BERT4Rec. Furthermore, user behavior data within the sequence inevitably contains noise, causing deep recommendation models to overfit on this noisy data, adversely affecting their recommendation performance. To address these issues, we have designed the GruMambaRec model. This model incorporates a filtering layer to mitigate the influence of noise on subsequent processing effectively. Additionally, it integrates the design principles of the GLINT-RU model by combining Dense Selective GRU with Mamba. Mamba is responsible for capturing long-term dependencies within the sequence. At the same time, Dense Selective GRU further extracts local temporal features through convolution, gating mechanisms, and nonlinear activation functions, improving the model’s capacity to choose important data. These two components work in concert to provide the model a more thorough grasp of user behavior sequences, which increases the efficacy and accuracy of recommendations. The detailed code is available at <a href="https://github.com/WEIWEIWEI369/GruMambaRec.">https://github.com/WEIWEIWEI369/GruMambaRec.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GruMambaRec: a sequential recommendation system combining GRU and Mamba

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Yu Huang,
  • Xiaojia Mo,
  • Xinyue Zhang,
  • Chunzhi Xie,
  • Jia Liu,
  • Zhisheng Gao

摘要

Sequential recommendation systems’ (SRSs) main goal is to predict what users could be interested in in the future by capturing their dynamic preferences based on their past interaction records. Although existing Transformer-based models perform well in this domain, the quadratic time complexity of their attention mechanism limits their efficiency in handling long interaction sequences. Recently, Mamba has been introduced into sequential recommendation systems, demonstrating impressive capabilities. However, as Mamba is a unidirectional model that only records previous user actions, it cannot leverage future contextual information. Our proposed GruMambaRec retains this unidirectional design, combining selective GRU and gating mechanisms to capture long-range dependencies, without incorporating future information, unlike bidirectional models such as BERT4Rec. Furthermore, user behavior data within the sequence inevitably contains noise, causing deep recommendation models to overfit on this noisy data, adversely affecting their recommendation performance. To address these issues, we have designed the GruMambaRec model. This model incorporates a filtering layer to mitigate the influence of noise on subsequent processing effectively. Additionally, it integrates the design principles of the GLINT-RU model by combining Dense Selective GRU with Mamba. Mamba is responsible for capturing long-term dependencies within the sequence. At the same time, Dense Selective GRU further extracts local temporal features through convolution, gating mechanisms, and nonlinear activation functions, improving the model’s capacity to choose important data. These two components work in concert to provide the model a more thorough grasp of user behavior sequences, which increases the efficacy and accuracy of recommendations. The detailed code is available at https://github.com/WEIWEIWEI369/GruMambaRec.