一种基于RNN的社交消息爆发预测模型

发布时间：2018-04-08 14:12

本文选题：循环神经网络　切入点：点随机过程　出处：《软件学报》2017年11期

【摘要】：社交网络中,消息的爆发预测属于社交网络流行动态分析的范畴,是社会计算领域的研究热点之一.通过利用基于深度循环神经网络对社交消息的传播过程进行建模,提出了SMOP(social messages outbreak prediction model based on recurrent neural network)模型.与传统的基于机器学习的模型相比,SMOP直接对消息转发的到达过程进行建模,避免了传统方法中繁琐的特征工程;与基于点随机过程的模型相比,SMOP可以自动学习消息传播过程的速率函数,不需要手动定义消息传播速率的特征函数,具有较强的数据场景适应性.另外,SMOP采用了时间向量和用户向量的输入表示方法,将时间的周期性和用户的兴趣偏好建模到传播过程之中,提升了SMOP的预测效果.在Twitter和新浪微博数据集上的实验结果均表明,SMOP具有优良的数据适应能力,可以在消息传播的早期(0.5h),以较高的F1值预测某条社交消息是否爆发,验证了模型的有效性.
[Abstract]:In social networks, the burst prediction of messages belongs to the category of social network popular dynamic analysis, and is one of the research hotspots in the field of social computing.The SMOP(social messages outbreak prediction model based on recurrent neural network is proposed by modeling the propagation process of social message based on the deep loop neural network.Compared with the traditional model based on machine learning, SMOP models the arrival process of message forwarding directly and avoids the tedious feature engineering in the traditional method.Compared with the model based on point random process, SMOP can learn the rate function of message propagation process automatically, and does not need to define the characteristic function of message propagation rate manually, so it has strong adaptability to data scene.In addition, SMOP uses the input representation method of time vector and user vector to model the periodicity of time and user's interest preference into the propagation process, which improves the prediction effect of SMOP.The experimental results on the data sets of Twitter and Sina Weibo show that SMOP has good data adaptability and can predict whether a social message erupts with a higher F1 value at the early stage of message transmission, which proves the validity of the model.
【作者单位】：中国科学院网络数据科学与技术重点实验室(中国科学院计算技术研究所);中国科学院大学;中国人民解放军61755部队;
【分类号】：G206;TP18

【相似文献】