王凯丽, 李海明. 基于多通道CNN-BiGRU与多特征融合方法[J]. 微电子学与计算机, 2022, 39(3): 41-49. DOI: 10.19304/J.ISSN1000-7180.2021.0987
引用本文: 王凯丽, 李海明. 基于多通道CNN-BiGRU与多特征融合方法[J]. 微电子学与计算机, 2022, 39(3): 41-49. DOI: 10.19304/J.ISSN1000-7180.2021.0987
WANG Kaili, LI Haiming. A method with multi-channel CNN-BiGRU and Multi-feature fusion[J]. Microelectronics & Computer, 2022, 39(3): 41-49. DOI: 10.19304/J.ISSN1000-7180.2021.0987
Citation: WANG Kaili, LI Haiming. A method with multi-channel CNN-BiGRU and Multi-feature fusion[J]. Microelectronics & Computer, 2022, 39(3): 41-49. DOI: 10.19304/J.ISSN1000-7180.2021.0987

基于多通道CNN-BiGRU与多特征融合方法

A method with multi-channel CNN-BiGRU and Multi-feature fusion

  • 摘要: 新闻推荐是根据用户的阅读习惯,为其推送更符合需求的内容,然而现有的方法仍存在特征学习不足的问题.针对此问题,提出了一种基于多通道CNN-BiGRU与多特征融合方法,主要由以下四部分组成:(1)词嵌入层.在词向量中融入实体嵌入向量,弥补单独仅使用词向量的不足,完成多通道词向量的构建;(2)多通道CNN-BiGRU模型.此部分模型使用卷积神经网络(CNN)提取语句的局部特征,使用双向门控循环单元(BiGRU)提取文本的长序列依赖关系,同时,借助最大池化操作以减少参数数量,避免冗余,借助注意力网络以关注重要词汇的特征,获取精确的特征表示;(3)多特征融合.借助注意力机制融合多个新闻特征,并关注重要特征的内容,完成新闻表征的构建;(4)用户表征提取.通过多头注意力机制提取用户历史浏览记录中新闻间的交互关系,以构建准确的用户表征,完成推荐.实验结果表明,所提出的CBMR模型相对现有的CNN、DKN、TANR和NRMS模型,在AUC、MRR、nDCG@5和nDCG@10指标上表现更优异.

     

    Abstract: According to the reading habits of users, news recommendations can push more suitable content for them, but the existing methods still have the problem of insufficient feature learning. To solve this problem, a method based on multi-channel CNN-BiGRU and multi-feature fusion is proposed, which is mainly composed of the following four parts: (1) Word Embedding Layer. Entity embedding vectors are integrated into word vectors to make up for the deficiency of using word embedding vectors and construct multi-channel word vectors; (2) Multi-channel CNN-BiGRU Model. In the model, a convolutional neural network (CNN) is used to extract the local features of sentences, and a bidirectional gated recurrent unit (BiGRU) is used to extract the long sequence dependency of text. At the same time, the maximum pool algorithm is used to reduce the number of parameters and avoid redundancy, and the attention network is used to pay attention to the features of key words to extract accurate features; (3) Multi-feature Fusion. With the help of attention mechanism, integrate multiple news features and pay attention to important features to complete the construction of news representation; (4) User Encoder. The interaction of news in the browsing histories of a user is extracted by multi-headed self-attention mechanism to construct user representation and complete recommendation.The experimental results show that compared with CNN, DKN, TANR, and NRMS models, CBMR performs better in AUC, MRR, nDCG@5, and nDCG@10 indicators.

     

/

返回文章
返回