电子学报 ›› 2022, Vol. 50 ›› Issue (9): 2102-2109.DOI: 10.12263/DZXB.20210483

• 学术论文 • 上一篇    下一篇

基于方面和胶囊网络的跨域评分预测模型

梁顺攀(), 刘伟, 郑智中, 原福永   

  1. 燕山大学信息科学与工程学院, 河北 秦皇岛 066004
  • 收稿日期:2021-04-15 修回日期:2021-11-08 出版日期:2022-09-25
    • 通讯作者:
    • 梁顺攀
    • 作者简介:
    • 梁顺攀 男, 1976年6月出生于广东省江门市, 现为燕山大学副教授、 硕士生导师, 研究方向为推荐系统.在国内外发表学术论文20余篇.
    • 基金资助:
    • 河北省自然科学基金(G2021203010)

Cross Domain Rating Prediction Based on Aspect and Capsule Network

LIANG Shun-pan(), LIU Wei, ZHENG Zhi-zhong, YUAN Fu-yong   

  1. College of Information Science and Engineering, Yanshan University, Qinhuangdao, Hebei 066004, China
  • Received:2021-04-15 Revised:2021-11-08 Online:2022-09-25 Published:2022-10-26
    • Corresponding author:
    • LIANG Shun-pan
    • Supported by:
    • Program of Natural Science Foundation of Hebei Province(G2021203010)

摘要:

用户评论可以反映用户对项目的偏好信息,将用户在其他领域的偏好迁移到目标领域进行跨域推荐,可以缓解目标域数据稀疏引起的冷启动问题.本文针对传统的跨域推荐方法无法将完整的用户偏好进行迁移以及传统的方面提取方法预测精度不高两个问题,提出基于方面和胶囊网络的跨域评分预测模型ACN(Aspect and Capsule Network).ACN模型使用胶囊网络挖掘评论文档的多个方面,然后通过注意力机制筛选出对目标域最重要的特征,迁移到目标域进行评分预测.最后,通过实验证明ACN模型分别在单一源域和多源域的情况下,较基准模型最高有2.3%和20.8%的性能提升.

关键词: 跨域推荐, 评分预测, 胶囊网络, 方面提取, 注意力

Abstract:

Users'reviews on items can reflect users'preference information. Migrating users'preferences in other fields to the target domain for cross domain recommendation can alleviate the cold start problem caused by sparse data in the target domain. Aiming at the two problems that the traditional cross domain recommendation methods can not migrate the complete user preferences and the prediction accuracy of the traditional aspect extraction methods is incorrect, this paper proposes a cross domain score prediction model based on aspect and capsule network(ACN). ACN uses capsule network to mine multiple aspects of review documents, and then selects the most important features for the target domain through attention mechanism, and migrate to the target domain for score prediction. Compared with the benchmark model, ACN has the highest performance improvement of 2.3% and 20.8% when using single source domain and multi-source domain.

Key words: cross domain recommendation, rating prediction, capsule network, aspect extraction, attention

中图分类号: