ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2020, Vol. 57 ›› Issue (6): 1312-1322.doi: 10.7544/issn1000-1239.2020.20190584

• 网络技术 • 上一篇    下一篇



  1. (黑龙江大学计算机科学技术学院 哈尔滨 150080) (
  • 出版日期: 2020-06-01
  • 基金资助: 

NT-EP: A Non-Topology Method for Predicting the Scope of Social Message Propogation

Liu Zitu, Quan Ziwei, Mao Rubai, Liu Yong, Zhu Jinghua   

  1. (College of Computer Science and Technology, Heilongjiang University, Harbin 150080)
  • Online: 2020-06-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61972135, 61602159), the Natural Science Foundation of Heilongjiang Province of China (F201430), the Innovation Talents Project of Science and Technology Bureau of Harbin (2017RAQXJ094, 2017RAQXJ131), and the Fundamental Research Funds of Universities in Heilongjiang Province (HDJCCX-201608, KJCX201815, KJCX201816).

摘要: 准确预测社交网络中消息的传播范围是舆情分析的重要内容,该问题受到了数据挖掘领域的广泛关注.目前的大部分研究主要利用社交网络拓扑结构和用户的动作日志来预测社交消息的传播范围.在实际应用中用户的动作日志中通常容易获得,但是社交网络的拓扑结构(例如用户之间的朋友关系)并不容易获得,因此无拓扑结构的社交消息预测具有更广泛的应用前景.提出了一种新的社交消息传播范围预测方法NT-EP,该方法由4部分构成:1)利用消息传播随时间衰减的特性为消息构造加权传播图,使用随机游走策略获取多条传播路径;2)把目标消息的传播路径输入到Bi-GRU(bidirectional gated recurrent unite),结合注意力机制计算出目标消息的传播特征向量;3)使用梯度下降方法计算出其他消息的影响向量;4)将目标消息的传播特征向量和其他消息的影响向量结合在一起,预测目标消息的传播范围.在Sina微博和Flixster数据集上的实验结果表明:NT-EP方法在均方误差(mean squared error, MSE),F1-score等多个指标上都优于现有的社交消息预测方法.

关键词: 社交网络, 传播范围, 拓扑结构, 随机游走, 梯度下降

Abstract: Predicting the scope of a message accurately in social networks is an important part of public opinion analysis, which has received extensive attention in the field of data mining. Most of the current research mainly uses social network topology and user action logs to predict the spread of social messages. It is usually easy to obtain action log about users in real applications, but the topology of the social network (for example, the friend relationship between users) is not easy to obtain. Therefore, non-topology social message prediction has good prospects for broader applications. In this paper, we propose a new method called NT-EP for predicting the propagation scope of social messages. NT-EP consists of four parts: 1)We construct a weighted propagation graph for each message based on the characteristics of message propagation decay over time, and then use a random walk strategy to obtain multiple propagation paths on the propagation graph; 2)We put multiple propagation paths of the target message into Bi-GRU, and combine the attention mechanism to obtain the propagation feature representation for the target message; 3)We use the gradient descent method to calculate the influence representation about other messages; 4)Combining the propagation feature representation for the target message with the influence representation about other events, we predict the propagation scope of the target message. The experimental results on Sina microblog and Flixster dataset show that our method is superior to existing social event prediction methods in terms of many indicators such as MSE and F1-score.

Key words: social network, scope of propagation, topology structure, random walk, gradient descent