电子学报 ›› 2023, Vol. 51 ›› Issue (1): 93-104.DOI: 10.12263/DZXB.20210570

• 学术论文 • 上一篇    下一篇

基于注意力增强的热点感知新闻推荐模型

丁琪1,2, 田萱1,2(), 孙国栋1,2   

  1. 1.北京林业大学信息学院,北京 100083
    2.国家林业草原林业智能信息处理工程技术研究中心,北京 100083
  • 收稿日期:2021-05-06 修回日期:2021-12-21 出版日期:2023-01-25
    • 通讯作者:
    • 田萱
    • 作者简介:
    • 丁 琪 女,1996年生,山东邹城人.北京林业大学信息学院硕士研究生.主要研究方向为智能信息处理、个性化推荐等.E-mail: 947515733@qq.com
      田萱(通讯作者) 女,1976年生,山东济宁人.2008年于中国人民大学获得博士学位,现为北京林业大学副教授,CCF高级会员.主要研究方向为智能信息处理、文本挖掘等.
      孙国栋 男,1981年生,黑龙江哈尔滨人.2009年与哈尔滨工业大学获得博士学位,现为北京林业大学副教授.主要研究方向为无线传感器网络、移动计算、数据挖掘等.E-mail: sungd@bjfu.edu.cn
    • 基金资助:
    • 国家重点研发计划 (2018YFC1603305)

HAN: Hotspot-Aware Attention Enhanced News Recommendation

DING Qi1,2, TIAN Xuan1,2(), SUN Guo-dong1,2   

  1. 1.School of Information Science and Technology, Beijing Forestry University, Beijing 100083, China
    2.Engineering Research Center for Forestry-Oriented Intelligent Information Processing of National Forestry and Grassland Administration, Beijing 100083, China
  • Received:2021-05-06 Revised:2021-12-21 Online:2023-01-25 Published:2023-02-23
    • Corresponding author:
    • TIAN Xuan
    • Supported by:
    • National Key R&D Program of China (2018YFC1603305)

摘要:

完全个性化的新闻推荐工作通常只基于用户兴趣,可能会导致推荐结果与点击过的内容过于相似甚至重复.事实上即使一些热点新闻并不完全符合用户兴趣,用户也可能希望点击类似的新闻.目前基于热点的新闻推荐方法不能很好挖掘潜在新闻的热点特征、灵活平衡用户兴趣和热点特征.本文提出一种新颖的注意力增强的热点感知新闻推荐模型(Hotspot-aware Attention enhaNced model,HAN),充分利用注意力网络和自注意力网络等深度神经网络的优势,在个性化推荐中将个性化兴趣与新闻热点性进行更好平衡与利用.该模型包括新闻编码器、热点特征提取器、用户兴趣提取器和点击预测器四个组件.提出一个热点特征提取器,使用注意力网络动态聚合热点新闻学习热点表示以更好挖掘热点特征;提出一个新颖的点击预测器来灵活融合热点特征、用户兴趣和候选新闻,以提升候选新闻的点击预测准确率.真实数据集上的实验表明HAN模型在AUC(Area Under the Curve of ROC)和F1两项指标上分别提升了7.51%和8.63%,且能够有效缓解用户冷启动问题.

关键词: 新闻推荐, 热点感知, 自注意力网络, 注意力网络, 卷积神经网络

Abstract:

Personalized news recommendation is usually based on users' interests only, which may cause the recommendation results to be too similar with or even repeat the content that has been clicked. In fact, even if some hot news may not meet the user's interests, users may also want to click on similar news. At present, hotspot-based approaches usually can not well mine the potential news hotspot features and flexibly balance the user interest and hotspot features. In this paper, a hotspot-aware attention enhanced model (HAN) for news recommendation is proposed, which makes full use of the advantages of deep neural networks such as attention networks and self-attention networks to better balance and utilize personalized interests and news hotspot in recommendation algorithm. HAN includes four components: news encoder, hotspot feature extractor, user interests extractor and click predictor. In order to effectively mine hotspot features, a hotspot feature extractor is proposed, which uses an attention network to dynamically aggregate hot news and learn hotspot feature representation; in order to improve the accuracy of predicting the click probability of candidate news, a click predictor is proposed to flexibly fuse hotspot features, user interests feature and candidate news representation. Experiments on a real-world dataset show that the area under the curve of ROC (AUC) and F1 increase by 7.51% and 8.63% respectively. At the same time, the model also helps to alleviate the cold-start problem of users.

Key words: news recommendation, hotspot-aware, self-attention networks, attention networks, convolutional neural networks

中图分类号: