首页|基于局部数据增强动态图的事件预测

基于局部数据增强动态图的事件预测

扫码查看
原文链接
  • 万方数据
  • 维普
事件指在真实世界中特定的时间和地点发生的与特定主题相关的活动,例如,社会动乱、暴恐袭击、自然灾害和传染病流行等事件会对国家安全和人民群众的生活产生重大威胁.如果能对此类事件的发生进行有效预测,将最大程度地减少负面事件带来的影响或最大化正面事件带来的利益.关于事件的研究中,准确预测事件仍然是一个非常具有挑战性的任务.文中提出了一种基于图注意力网络的事件预测方法LAT-GAT(Local Augmented Temporal-GAT),该方法使用条件变分编码器,在所构建的事件图中对目标节点的邻居节点生成新的特征样本,与节点原有特征进行拼合,形成新的节点特征,实现了对事件的传播结构的利用;另外,LAT-GAT还考虑了历史事件发生的时间先后顺序,将网络在上一时间点的输出结果集成到当前时间的特征中,从而实现了对事件传播时间特性的利用.最后,在泰国、印度、埃及和俄罗斯这4个国家真实事件数据集上,与多种代表性基线方法进行了对比实验.实验结果表明,LAT-GAT在4个国家数据上的F1评分都优于基线方法;在泰国、俄罗斯和印度数据集上召回率优于基线方法;在泰国、埃及和印度数据集上也获得了最高的准确率.还通过消融实验考察了模型参数对最终结果的影响.
Event Prediction Based on Dynamic Graph with Local Data Augmentation
Event refers to activities that occur in real world at specific time and places.For instance,unrest,violent terrorist at-tacks,natural disasters and the spread of infectious diseases,will bring great threats and losses to national security and human life.If the occurrence of such events could be predicted more precisely and effectively,the impact of negative events will be mini-mized,and it is possible to maximize the benefits of the positive events.It is still a very challenging task to predict events accu-rately.An event prediction method named local augmented temporal-GAT(LAT-GAT)based on graph attention network is pro-posed in this paper.It uses conditional variational encoders to generate new features,which will be concatenated with the original features to new one,based on neighbors of the current node.With this approach,our model can utilize the propagation structure of events.In addition,the chronological order of events occurrence is considered by our model.The feature of events in last time point is integrated into the output of the neural network in current time.The temporal property of event propagation is exploited through temporal data integration.And finally,the proposed method is compared with a number of representative baseline me-thods on the real-world datasets,including Thailand,India,Egypt and Russia.The results show that LAT-GAT has the best F1 scores in all datasets.The recall of our model exceeds that of any other baseline methods in the datasets of Thailand,Russia and India.In Thailand,Egypt and India,our model achieves the best precision.Ablation experiments are also conducted to investigate the influence of the model parameters on the final results.

Event predictionGraph attention networkDynamic graphConditional variational auto-encoderData augmentation

潘磊、刘欣、陈君益、程章桃、刘乐源、周帆

展开 >

中国电子科技集团公司第十研究所 成都 610036

电子科技大学信息与软件工程学院 成都 610054

喀什地区电子信息产业技术研究院 新疆喀什 844099

事件预测 图注意力网络 动态图 条件变分编码器 数据增强

国家自然科学基金国家自然科学基金四川省自然科学基金四川省科技计划厅市共建智能终端四川省重点实验室开放课题

62176043620720772022NSFC05052022YFSY0006SCITLAB-20006

2024

10.11896/jsjkx.221200054

计算机科学
重庆西南信息有限公司（原科技部西南信息中心）

计算机科学

CSTPCD北大核心
影响因子：0.944
ISSN：1002-137X
年,卷(期)：2024.51(3)
  • 48