基于会话的双层注意力机制新闻推荐方法
News recommendation method for session based neural neural network with two layer attention mechanism
  
DOI:
中文关键词:  新闻推荐;基于会话的推荐;注意力机制;神经网络
英文关键词:news recommendations; session based recommendation; attention mechanism; neural network
基金项目:国家自然科学基金(61772285,61373185)资助项目
作者单位
王海艳 南京邮电大学 计算机学院,江苏南京210023
南京邮电大学 江苏省无线传感网高技术研究重点实验室,江苏南京210023
南京邮电大学 江苏省大数据安全与智能处理重点实验室,江苏南京210023 
胡阳 南京邮电大学 计算机学院,江苏南京210023 
摘要点击次数: 1890
全文下载次数: 1440
中文摘要:
      随着新闻信息的爆炸式增长,个性化新闻推荐对用户快速找到感兴趣的新闻变得非常重要。基于会话的推荐旨在根据用户的行为序列向用户推荐下一个可能感兴趣的项目。但是,现有方法大都忽略了会话内用户阅读行为的随机性和偶然性,难以捕捉用户的主要兴趣,或者将用户会话视为单个序列,忽视了用户兴趣在不同会话之间的演变和关联。文中提出了一种基于会话的双层注意力机制新闻推荐方法,该方法将用户行为序列划分成多个会话:首先使用卷积神经网络对用户点击的新闻特征进行提取;其次在兴趣感知注意力层结合门控循环单元和自注意力机制获取用户在每个会话内的主要兴趣;最后,在会话感知注意力层使用注意力机制建模当前会话和历史会话之间用户兴趣的关联程度以形成最终的用户兴趣表示。通过在真实世界的数据集上与基准方法进行对比实验,结果表明了所提方法在新闻推荐上的有效性。
英文摘要:
      With the explosive growth of news information, personalized news recommendations have become very important for users to specially push news they are really interested in. The task of session based recommendations is to recommend the next item that may be of interest to the user based on the user history behavior sequence. However, most of the existing methods ignore the randomness and the contingency of the user reading in a session, thus it is difficult to capture the user main interest. Some other methods treat the user sessions as a single sequence, ignoring the evolution and the association of user interests between different sessions. A news recommendation method for session based neural network with two layer attention mechanism is proposed, by dividing the user behavior sequence into multiple sessions. Firstly, it uses a convolutional neural network to extract the news features clicked by the user. Then, the gate recurrent unit and the self attention mechanism are combined to learn the user main interest in the interest aware attention layer. Finally, the attention mechanism is used in the session aware attention layer to model the correlation of user interests between tcurrent sessions and historical sessions to form the final representation of user interests. The simulation experiment is conducted on real world news datasets. Experimental results demonstrate the effectiveness of the proposed method for news recommendations.
查看全文  查看/发表评论  下载PDF阅读器

你是第3766431访问者
版权所有《南京邮电大学学报(自然科学版)》编辑部
Tel:86-25-85866913 E-mail:xb@njupt.edu.cn
技术支持:本系统由北京勤云科技发展有限公司设计