Scholarlink

<p>Sequential recommendation (SR) methods improve accuracy by considering the temporal sequence of user-item interactions rather than treating interaction histories as static sets. In this paper, we address the problem of implicit representation learning for multi-time-scale user interests by proposing a novel method based on user behavior sequence modeling—<b>m</b>ulti-<b>t</b>ime-<b>s</b>cale with <b>c</b>lockwork recurrent neural network modeling for sequential recommendation (MTSC). Specifically, firstly, we group the neurons of the hidden layer of the recurrent neural network (RNN) based on the clockwork RNN (CW-RNN) method according to the different degrees of dynamic changes of user interests. Secondly, we design different update frequencies to extract user interest features at multiple time scales. Finally, we model the dependency of user interest features at different time scales through scale-dimensional convolution to generate a unified representation of user interest features at multiple time scales, which can be used to predict the items of interest to users. To validate its effectiveness, we conducted extensive experiments on three public datasets, and the results demonstrate that the MTSC model achieves state-of-the-art performance across all baselines. More precisely, on the Steam dataset, the Precision@10 metric of the MTSC model improved by 4.73% compared to the best baseline model, robustly validating the effectiveness and superiority of the proposed method.</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

Multi-time-scale with clockwork recurrent neural network modeling for sequential recommendation

  • 期刊论文

  • 发表日期：

    2025年1月21日

查看全文
  • Nana Huang,
  • Hongwei Ding,
  • Ruimin Hu,
  • Pengfei Jiao,
  • Zhidong Zhao,
  • Bin Yang,
  • Qi Zheng

摘要

Sequential recommendation (SR) methods improve accuracy by considering the temporal sequence of user-item interactions rather than treating interaction histories as static sets. In this paper, we address the problem of implicit representation learning for multi-time-scale user interests by proposing a novel method based on user behavior sequence modeling—multi-time-scale with clockwork recurrent neural network modeling for sequential recommendation (MTSC). Specifically, firstly, we group the neurons of the hidden layer of the recurrent neural network (RNN) based on the clockwork RNN (CW-RNN) method according to the different degrees of dynamic changes of user interests. Secondly, we design different update frequencies to extract user interest features at multiple time scales. Finally, we model the dependency of user interest features at different time scales through scale-dimensional convolution to generate a unified representation of user interest features at multiple time scales, which can be used to predict the items of interest to users. To validate its effectiveness, we conducted extensive experiments on three public datasets, and the results demonstrate that the MTSC model achieves state-of-the-art performance across all baselines. More precisely, on the Steam dataset, the Precision@10 metric of the MTSC model improved by 4.73% compared to the best baseline model, robustly validating the effectiveness and superiority of the proposed method.