文献类型：期刊文献

中文题名：基于Attention机制的链接预测算法

英文题名：Link prediction algorithm based on attention mechanism

作者：程华[1];张林[1];房一泉[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2019

卷号：47

期号：2

起止页码：109

中文期刊名：华中科技大学学报（自然科学版）

外文期刊名：Journal of Huazhong University of Science and Technology(Natural Science Edition)

收录：CSTPCD;;EI(收录号：20192407026505);Scopus;北大核心:【北大核心2017】；CSCD:【CSCD2019_2020】；

基金：国家自然科学基金资助项目(61501187)

语种：中文

中文关键词：注意力机制;双向循环神经网络;拓扑序列化;链接预测;局部网络

外文关键词：link prediction;attention mechanism;bi-directional recurrent neural network (Bi-RNN);topological serialization;local network

摘要：针对社会网络中链接预测问题,提出了基于注意力(Attention)机制的链接表示及其预测算法.基于待预测节点的共邻关系构建其链接局部网络,设计了基于紧密游走的网络拓扑序列化方法.采用双向循环神经网络(Bi-RNN)对链接序列进行向量编码,以充分挖掘序列相关节点间的上下文依赖信息.通过Attention机制对链接中的节点进行关注和加权,强化重要节点对链接预测任务的贡献,实现链接拓扑特征的自动提取与准确分类预测.实验结果表明,在4种不同类型的社会网络数据集中,该算法的准确率和运算效率都有较大提高且普适性较强.
Aiming at the link prediction task in social networks,a link representation method and a link prediction algorithm based on attention mechanism were proposed.A link local network was designed which contained a node pair to be link predicted and their co-neighbors,and was serialized by multiple closely walks.Links sequences were encoded by bi-directional recurrent neural network (Bi-RNN) for its ability on mining the context information from the sequential nodes.The attention mechanism,used to focus and weight the nodes in the link,strengthens the contribution of important nodes,which can promote the accuracy of link prediction.Experiments on four types of social network datasets show that the algorithm has significant improvement in accuracy and computational efficiency,and is suitable for multiple social networks.