计算机科学与探索 ›› 2021, Vol. 15 ›› Issue (11): 2222-2232.DOI: 10.3778/j.issn.1673-9418.2009035

• 人工智能 • 上一篇    下一篇

融合短文本层级注意力和时间信息的推荐方法

邢长征,郭亚兰,张全贵,赵宏宝   

  1. 辽宁工程技术大学 电子与信息工程学院,辽宁 葫芦岛 125105
  • 出版日期:2021-11-01 发布日期:2021-11-09

Recommendation Method Integrating Review Text Hierarchical Attention with Time Information

XING Changzheng, GUO Yalan, ZHANG Quangui, ZHAO Hongbao   

  1. School of Electronics and Information Engineering, Liaoning Technical University, Huludao, Liaoning 125105, China
  • Online:2021-11-01 Published:2021-11-09

摘要:

信息过载造成的数据稀疏性问题制约着基于评分数据的矩阵分解模型的推荐性能,融合评论文本的推荐模型能够有效缓解评分数据稀疏性。当前的推荐系统利用评论文本为用户和项目建模时,大多仅将用户对项目的评论作为数据来源,而忽视了时间信息对用户和项目属性的影响。针对此问题,提出了一种融合短文本层级注意力和时间信息的推荐方法(RHATR),该方法能够充分地挖掘评论文本潜在的语义信息,并为用户偏好和项目特征的动态变化进行建模。通过对单条评论文本应用单词级注意力,挖掘单条评论文本中情感词和关键词等有效信息,学习用户和项目表示;对含有时间因素的用户评论集和项目评论集分别应用评论级注意力,提取有效的评论,进一步学习用户偏好和项目特征动态表示。将从评论文本中学到的用户和项目表示以及基于ID的项目和用户嵌入作为最终特征,来捕获各用户和项目的潜在因素。实验结果表明,提出的方法相对于当前基线方法在Amazon和Yelp数据集上的均方根误差(RMSE)取得了较好的效果。

关键词: 推荐系统, 评论文本, 层级注意力, 时间信息

Abstract:

The data sparsity problem caused by information overload restricts the recommendation performance of the matrix factorization model based on scoring data. The recommendation model integrated with reviews text can effectively alleviate the sparsity of scoring data. The current recommendation system uses reviews text to model users and items, most of them use users' reviews text on items as the data sources, and ignore the impacts of time information on user and item attributes. In response to this problem, a recommendation method which integrates review text hierarchical attention with time information (RHATR) is proposed. This method can fully mine the potential semantic information of review text and model the dynamic changes of user perferences and item features. This method mines effective information such as sentiment words and keywords in a single review text, learns user and item representation by applying word level attention to a single review text, applies review level attention to extracting effective reviews respectively for the user review set and item review set with temporal factor and further learns the dynamic representation of user preferences and item features. Finally, the user and item representation learned from the review text, and ID based item and user embedding as the final features, it captures the potential factors of each user and item. The experimental results show that the proposed method has a better effect in root mean square error (RMSE) on Amazon and Yelp datasets than the current baseline method.

Key words: recommendation system, review text, hierarchical attention, time information