基于动态增强图注意力网络的突发事件预测  

Incident event prediction based on the dynamic enhanced graph attention network

作　　者：仲兆满 崔心如 张渝 吕慧慧 樊继冬 Zhong Zhaoman;Cui Xinru;Zhang Yu;LüHuihui;Fan Jidong(College of Computer Engineering,Jiangsu Ocean University,Lianyungang,222005,China)

机构地区：[1]江苏海洋大学计算机工程学院,连云港222005

出　　处：《南京大学学报(自然科学版)》2025年第1期94-104,共11页Journal of Nanjing University(Natural Science)

基　　金：国家自然科学基金(72174079);江苏省“青蓝工程”大数据优秀教学团队(2022-29);连云港市重点研发(产业前瞻与关键核心技术)项目(CG2323)

摘　　要：图神经网络在处理事件图的节点特征时会出现过平滑性的问题,导致较难获取全面的事件特征;同时,由于事件发展是动态的,网络在处理过程中会忽略短时间切片之间的数据变化,难以捕捉事件的全局时间特征.针对以上问题,提出基于动态增强图注意力网络的突发事件预测模型(Dynamic Enhanced Graph Attention Network,DEGAT),通过构建使用高斯扰动增强的图注意力网络(Enhanced Graph Attention Network,EGAT)来获取历史事件图全面的事件特征.将初始事件向量和EGAT输出的事件向量分别输入线性层进行融合,得到时间特征,再将多个不同历史时间特征序列输入多头注意力机制与LSTM相结合的时间编码层,获得全局时间特征.最后,将全局时间特征输入EGAT,经过非线性变换后输出预测结果.在四个社会突发事件数据集上的实验结果表明,提出的模型与DynamicGCN的方法相比,准确率和精确率分别提高了3.88%和4.12%.When processing the node features of the event graph,it is difficult to obtain comprehensive event features.Meanwhile,since the event development is dynamic,the network ignores the data changes between short time slices during the processing,making it difficult to capture the global temporal characteristics of the events.To address the above problems,an incident prediction model(Dynamic Enhanced Graph Attention Network,DEGAT)based on dynamic enhanced graph attention network is proposed.This model obtains the comprehensive event features of the historical event graph by building a Gaussian enhanced graph attention network(Enhanced Graph Attention Network,EGAT);we input the initial event vector and EGAT output into the linear layer to obtain the time features,and then input different historical time features into the time coding layer combined with multi⁃head attention mechanism and LSTM to obtain the global time features;finally,the global time feature is input into EGAT,and outputs the prediction results after nonlinear transformation.Experimental results on four social burst datasets show that the proposed model improves 3.88%in accuracy and 4.12%in precision compared with the DynamicGCN method.

关 键 词：事件预测 图注意力网络 特征增强 时间序列 多头注意力机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.1[自动化与计算机技术—控制科学与工程]