文章摘要
欧阳纯萍,陈湘龙,刘永彬.基于网络表示学习的新闻用户影响力预测[J].情报工程,2021,7(5):115-125
基于网络表示学习的新闻用户影响力预测
Network News User Influence Prediction Based on Network Embedding
  
DOI:10.3772/j.issn.2095-915X.2021.05.010
中文关键词: 网络新闻评论;用户影响力预测;异构网络表示学习
英文关键词: Network news; user influence prediction; heterogeneous network embedding
基金项目:湖南省哲学社会科学基金 (16YBA323)。
作者单位
欧阳纯萍 1. 南华大学计算机学院 衡阳 421001
 
陈湘龙 1. 南华大学计算机学院 衡阳 421001
2. 湖南科技职业学院软件学院 长沙 410004 
刘永彬 1. 南华大学计算机学院 衡阳 421001
 
摘要点击次数: 65
全文下载次数: 53
中文摘要:
      [ 目的 / 意义 ] 现有新闻用户影响力预测研究大部分是利用微博社交网络中的全局特征进行分析,忽略了社交网络中异构节点的局部特征以及节点本身所包含的属性和文本信息。为融合更丰富的社交网络特征,提升新闻用户影响力预测性能,本文提出一种基于异构网络表示学习的新闻用户影响力预测模型。[ 方法 / 过程 ] 文章以新闻评论网络为研究对象,首先构建了评论信息网络、用户关注网络以及用户 - 评论网络三个基础网络,然后将三个异构网络进行联合学习,使得三个网络中的异构节点在向量空间中的分布近似于其在真实网络中的分布,并采用 KL 散度来刻画两种分布之间的关系。经过异构网络表示学习之后,用户、评论以及新闻文章被表征到一个低维的向量空间当中,通过保存网络中的局部结构使得具有相似潜在影响力的节点在低维的向量空间中的距离更近,从而通过计算节点间的相似度来构建节点影响力的概率转移矩阵,最后使用多变量随机游走算法进行迭代,预测用户的未来影响力。[ 结果 / 结论 ] 实验结果表明,在不同的 Top-K 下算法性能稳定,K 取值 20、50、100 和200 时,准确率分别达到 85%、82%、80% 和 77%。
英文摘要:
      [Objective/ Significance] The existing researches on network news user influence prediction only use global features of Weibo social network, which ignore the heterogeneity of nodes and the attributes and text information contained in nodes of social network. In order to make full use of richer social network features and improve the prediction performance of user influence, this paper proposes a network news user influence prediction model based on heterogeneous network embedding. [Methods/Process]. Firstly, we construct three basic networks including the comment information network, the user attention network and the user-comment network. Then, we combine the three networks to learn the heterogeneous node representation and ensure that the distribution of heterogeneous nodes in vector space is similar to that of the real network. KL dispersion is used to describe the relationship between the two distributions. After the heterogeneous network embedding, users, comments, and news articles are represented in the same low-dimensional vector space. By preserving the local structure of the network, nodes with similar potential influences are closer in the low-dimensional vector space, and the probability transfer matrix of node influence is constructed by calculating the similarity between nodes. Finally, the multivariable random walk algorithm is used to iterate to predict the user’s influence. [Results /Conclusions]. The experimental results show that the performance of the proposed algorithm is stable under different Top-K. When K is set to20, 50, 100, and 200, the accuracy rates reach 85%,82%, 80%, and 77%, respectively.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮