Scholarlink

<p>Side information-integrated sequential recommendation leverages supplementary information to alleviate the issue of data sparsity in pure item ID-based sequential recommendation. In this paper, we take online shopping as our modeling scenario and consider two important kinds of side information, i.e., the item categories and behavior types, because they can help to effectively represent users’ dynamic intentions in this practical scenario. Most state-of-the-art methods design an auxiliary task of category prediction during training, aiming to incorporate useful category information into the item representations. Despite their achieved success, we find that there are still two fundamental limitations need to be addressed: (1) the learned category sequence representations suffer from low quality; and (2) they overlook the learning of users’ fine-grained preferences. To address these limitations, we propose Dual-stage Scoring via Task Decoupling and Fine-grained Preference Learning (DSTD-FPL) to better leverage item category and behavioral information. Specifically, DSTD-FPL decouples the item and category prediction tasks and releases the potential of learning the improved category sequence representations. Moreover, it explicitly learns users’ intra-category, inter-category and target behavior-oriented fine-grained preferences for the items to refine preference learning. During model inference, DSTD-FPL computes the prediction scores for candidate items and their corresponding categories separately in two stages. The final ranking scores are obtained by summing the dual-stage scores. We conduct extensive experiments on two real-world datasets, and find that our DSTD-FPL achieves an average improvement of 5.61% and 22.45% on two datasets, respectively, compared with the most competitive baseline. Overall, this work advances the integration of side information for sequential recommendation by proposing a task-decoupled framework with fine-grained preference modeling, delivering both methodology advancement and performance improvement for real-world applications.</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

Dual-stage scoring via task decoupling and fine-grained preference learning for side-information integrated sequential recommendation

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Xiaolin Lin,
  • Jinwei Luo,
  • Mingkai He,
  • Weike Pan,
  • Zhong Ming

摘要

Side information-integrated sequential recommendation leverages supplementary information to alleviate the issue of data sparsity in pure item ID-based sequential recommendation. In this paper, we take online shopping as our modeling scenario and consider two important kinds of side information, i.e., the item categories and behavior types, because they can help to effectively represent users’ dynamic intentions in this practical scenario. Most state-of-the-art methods design an auxiliary task of category prediction during training, aiming to incorporate useful category information into the item representations. Despite their achieved success, we find that there are still two fundamental limitations need to be addressed: (1) the learned category sequence representations suffer from low quality; and (2) they overlook the learning of users’ fine-grained preferences. To address these limitations, we propose Dual-stage Scoring via Task Decoupling and Fine-grained Preference Learning (DSTD-FPL) to better leverage item category and behavioral information. Specifically, DSTD-FPL decouples the item and category prediction tasks and releases the potential of learning the improved category sequence representations. Moreover, it explicitly learns users’ intra-category, inter-category and target behavior-oriented fine-grained preferences for the items to refine preference learning. During model inference, DSTD-FPL computes the prediction scores for candidate items and their corresponding categories separately in two stages. The final ranking scores are obtained by summing the dual-stage scores. We conduct extensive experiments on two real-world datasets, and find that our DSTD-FPL achieves an average improvement of 5.61% and 22.45% on two datasets, respectively, compared with the most competitive baseline. Overall, this work advances the integration of side information for sequential recommendation by proposing a task-decoupled framework with fine-grained preference modeling, delivering both methodology advancement and performance improvement for real-world applications.