Scholarlink

<p>Sequential recommendation aims to address the information explosion problem in many online scenarios by modeling the sequence of historical interactions of a user and recommending the next potential interesting item to the user. The user’s long- and short-term preferences combine his overall tastes and latest needs, fully reflecting the persistence and variability of user preferences. Some state-of-the-art models enhance the expressiveness of item embedding by integrating various side information. In doing so, learning a user’s underlying interests from multiple perspectives is possible. To combine these two types of information, a recommendation model that incorporates Long-Short-Term preferences and Side Information(LST-SI) is proposed. Precisely, the LST-SI model first models the preference transition patterns of item IDs and side information using two independent self-attention networks. It then uses two attention modules to learn the long-short-term preferences of item IDs and side information. Lastly, we concatenate the learned feature representations of item IDs and side information and input them into a fully connected layer to obtain the final user interest representation. Experimental results on three real-world public datasets demonstrate that LST-SI outperforms other baseline models in terms of NDCG and recall, mainly when dealing with long historical user interaction sequences. Our source code is available at <a href="https://github.com/cyh512167618/LST-SI/tree/master.">https://github.com/cyh512167618/LST-SI/tree/master.</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

Sequential recommendation with long-short-term preference and side information

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Yongquan Fan,
  • Yuhang Cheng,
  • Yitong Wang,
  • Xianyong Li

摘要

Sequential recommendation aims to address the information explosion problem in many online scenarios by modeling the sequence of historical interactions of a user and recommending the next potential interesting item to the user. The user’s long- and short-term preferences combine his overall tastes and latest needs, fully reflecting the persistence and variability of user preferences. Some state-of-the-art models enhance the expressiveness of item embedding by integrating various side information. In doing so, learning a user’s underlying interests from multiple perspectives is possible. To combine these two types of information, a recommendation model that incorporates Long-Short-Term preferences and Side Information(LST-SI) is proposed. Precisely, the LST-SI model first models the preference transition patterns of item IDs and side information using two independent self-attention networks. It then uses two attention modules to learn the long-short-term preferences of item IDs and side information. Lastly, we concatenate the learned feature representations of item IDs and side information and input them into a fully connected layer to obtain the final user interest representation. Experimental results on three real-world public datasets demonstrate that LST-SI outperforms other baseline models in terms of NDCG and recall, mainly when dealing with long historical user interaction sequences. Our source code is available at https://github.com/cyh512167618/LST-SI/tree/master.