融合拓扑和属性的动态网络链路预测方法  

Dynamic Network Link Prediction Method for Fusion Topology and Attributes

作　　者：罗世杰 吕文韬 李凡 崔家熙 相洁 LUO Shijie;LYU Wentao;LI Fan;CUI Jiaxi;XIANG Jie(School of Information and Computer Science,Taiyuan University of Technology,Jinzhong,Shanxi 030600,China)

机构地区：[1]太原理工大学信息与计算机学院,山西晋中030600

出　　处：《计算机工程与应用》2023年第5期122-130,共9页Computer Engineering and Applications

基　　金：国家自然科学基金(61873178,61876124,61906130);山西省国际科技合作项目(201803D421047)。

摘　　要：网络数据中出现的大量节点属性和随时间变化的特征,给链路预测提出了新挑战。基于注意力机制和循环神经网络对随时间演化网络进行建模,提出了DTA-LP模型。与传统的静态链路预测算法相比,DTA-LP使用LSTM捕获时序信息,动态预测可以更好应用于现实网络;与基于网络拓扑的动态链路预测算法相比,DTA-LP可以聚集高阶拓扑特征,有效挖掘网络邻域信息;与基于属性网络的动态链路预测算法相比,DTA-LP可以加权融合网络拓扑属性,提高预测精度。在4种真实数据上的实验结果表明,该方法能结合网络已有先验知识,以较高的MAP值来预测未来网络中的边,验证了模型的有效性。A large number of node attributes and temporal features in network data provide new challenges for the link prediction. On the basis of the attention mechanism and recurrent neural network, this paper analyses the evolution network over time and proposes a DTA-LP model for the computer simulation of this network. Compared with the traditional static link prediction algorithm, DTA-LP uses LSTM to capture temporal information, the dynamic prediction can be better applied to real networks. Unlike the dynamic link prediction algorithm based on the network topology, DTA-LP can gather high-order topology features and effectively obtain network domain information. Different from the dynamic link prediction algorithm based on attribute networks, DTA-LP can weight and fuse network topology and attributes to improve the prediction accuracy. The experimental results on four kinds of real data show that DTA-LP can combine the prior knowledge of the network to predict the middle edge of the future network with a high MAP value, which verifies the effectiveness of the model.

关 键 词：链路预测 动态图 属性融合 注意力机制 节点嵌入 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]