摘要
目前学界普遍通过循环神经网络(RNN)建模强度函数来刻画时序点过程,然而此类模型不能捕捉到事件序列之间的长程依赖关系,并且强度函数具体的参数形式会限制模型的泛化能力。针对上述问题,提出一种无强度函数的注意力机制的时序点过程生成模型。该模型使用Wasserstein距离构建损失函数,便于衡量模型分布与真实分布之间的偏差,利用自注意力机制描述历史事件对当前事件的影响程度,使得模型具有可解释性且泛化能力更强。对比实验表明,在缺失强度函数先验信息的情况下,该方法比RNN类的生成模型和极大似然模型在QQ图斜率的偏差和经验强度偏差这两个指标总体上分别减少35.125%和24.200%,证实了所提模型的有效性。
At present,the academic circles generally describe the temporal point process by modeling the intensity function using recurrent neural network(RNN).However,this kind of model can’t capture the long-range dependence between event sequences,and the specific parameter form of the intensity function will limit the generalization ability of the model.In order to solve these problems,this paper proposed a temporal point process self-attention generation model without intensity function.The model used Wasserstein distance to construct the objective function,which was convenient to measure the deviation between the model distribution and the real distribution,and used the self-attention mechanism to describe the impact of historical events on current events,so that the model was interpretable and had stronger robustness.Comparative experiments show that,in the absence of prior knowledge of intensity function,the deviation of QQ graph slope and empirical intensity deviation of this method reduce 35.125%and 24.200%respectively compared with RNN generation model and maximum likelihood mo-del,which proves the effectiveness of the proposed model.
作者
芦佳明
李晨龙
魏毅强
Lu Jiaming;Li Chenlong;Wei Yiqiang(College of Mathematics,Taiyuan University of Technology,Jinzhong Shanxi 030600,China)
出处
《计算机应用研究》
CSCD
北大核心
2022年第2期456-460,共5页
Application Research of Computers
基金
国家自然科学基金资助项目(61901294)
山西省应用基础研究计划资助项目(201901D211105)。
