基于用户行为的情感影响力和易感性学习
发布时间:2018-03-29 20:33
本文选题:在线社交网络 切入点:观点传播 出处:《计算机学报》2017年04期
【摘要】:在不同情感极性上建模用户间的影响力是观点形成和病毒式营销的一个关键问题.已有工作将用户间影响力直接定义在用户对上,无法刻画未观测到用户对之间的关联关系,造成用户影响力学习的过拟合问题.此外,目前尚无针对不同情感极性的用户间影响力建模的有效方法.因此,该文提出一种融合情感因素的用户分布式表达模型.该模型首先构建两个低维参数矩阵度量在不同情感极性上传播者的影响力和接受者的易感性,然后通过生存分析模型刻画级联的传播行为,最后利用负采样方法解决模型中存在正负例严重不平衡的问题.基于带有情感观点的微博转发所形成级联数据集的实验结果表明,与基准方法对比,该文方法在"预测动态级联"和"谁将会被转发"任务上MRR指标分别提高了273%和32.4%,在"级联大小预测"任务上MAPE指标下降了10.46%,很好地验证了该文模型的有效性.此外,该文分析用户的情感影响力和易感性分布并发现了一些重要的现象.
[Abstract]:Modeling the influence between users on different emotional polarities is a key issue in view formation and viral marketing. In addition, there is no effective method for modeling the influence between users with different affective polarities. In this paper, a user distributed representation model combining emotional factors is proposed. Firstly, two low-dimensional parameter matrices are constructed to measure the influence of the communicator and the susceptibility of the receiver in different affective polarities. Then the propagation behavior of cascades is described by survival analysis model. Finally, negative sampling method is used to solve the problem of serious imbalance of positive and negative cases in the model. Compared with the baseline method, In this paper, the MRR index of "predicting dynamic cascade" and "who will be forwarded" is increased by 27.3% and 32.4% respectively, and the MAPE index of "cascaded size prediction" task is decreased by 10.46%, which verifies the validity of the model. This paper analyzes the affective influence and susceptibility distribution of users and finds some important phenomena.
【作者单位】: 福州大学数学与计算机科学学院;福建省网络计算与智能信息处理重点实验室(福州大学);中国科学院网络数据科学与技术重点实验室;
【基金】:国家“九七三”重点基础研究发展规划项目基金(2013CB329606,2013CB329602) 国家自然科学基金项目(61572467,61300105) 中国科学院网络数据科学与技术重点实验室开放基金课题(CASNDST20140X)资助~~
【分类号】:TP391.1
,
本文编号:1682737
本文链接:https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/1682737.html