基于分层注意力的信息级联预测模型

信息级联预测(Information Cascade Prediction)是社交网络分析领域的一个研究热点,其通过信息级联的扩散序列与拓扑图来学习在线社交媒体中信息的传播模式。当前的信息级联预测模型大多以循环神经网络为基础,仅考虑信息级联的时序结构信息或者序列内部的空间结构信息,无法学习序列之间的拓扑关系。而现有的级联图结构学习方法无法为节点的邻居分配不同的权重,导致节点之间的关联性学习较差。针对上述问题,文中提出了基于节点表示的信息级联采样方法,将信息级联建模为节点表示而非序列表示。随后提出一种基于分层注意力的信息级联预测(Information Cascade Prediction with Hierarchical Attention,ICPHA)模型。该模型首先通过结合了自注意力机制的循环神经网络来学习节点序列的时序结构信息;然后通过多头注意力机制学习节点表示之间的空间结构信息;最后通过分层的注意力网络对信息级联的结构信息进行联合建模。所提模型在Twitter,Memes,Digg这3种数据集上达到了领先的预测效果,并且具有良好的泛化能力。

Information Cascade Prediction Model Based on Hierarchical Attention

Information cascade prediction is a research hotspot in the field of social network analysis.It learns the propagation mode of information in online social media through the diffusion sequence and topology map of the information cascade.Most current models for solving this task are based on recurrent neural networks and only consider information cascading time series structure information or spatial structure information inside sequences,and cannot learn topological relationships between sequences.And the existing cascade graph structure learning methods cannot assign different weights to the neighbors of the nodes,resulting in poor association learning between the nodes.In response to the above problems,this paper proposes an information cascade sampling method based on node representation,which models the information cascade as a node representation rather than a sequence representation.This paper also proposes an information cascade prediction model based on hierarchical attention network(ICPHA),which learns the time series structure information of the node sequence through a recurrent neural network layer with self-attention mechanism,and learns the spatial structure information between node representations through a multi-head attention mechanism.By this way,ICPHA jointly models the structural information of the information cascade through a hierarchical attention network.ICPHA has achieved leading prediction results on Twitter,Memes,and Digg,and has good generalization ability.