融合时空行为与社交关系的用户轨迹识别模型  

A User Trajectory Identification Model with Fusion of Spatio-Temporal Behavior and Social Relation

作　　者：张伟[1] 李扬[1] 张吉 王建勇[3] ZHANG Wei;LI Yang;ZHANG Ji;WANG Jian-Yong(School of Computer Science and Technology,East China Normal University,Shanghai 200062;Zhijiang Lab,Hangzhou 311000;Department of Computer Science and Technology,Tsinghua University,Beijing 100084)

机构地区：[1]华东师范大学计算机科学与技术学院,上海200062 [2]之江实验室,杭州311000 [3]清华大学计算机科学与技术系,北京100084

出　　处：《计算机学报》2021年第11期2173-2188,共16页Chinese Journal of Computers

基　　金：国家自然科学基金(No.62072182,No.61702190,No.61532010);之江实验室(No.2019KB0AB04)资助.

摘　　要：随着位置社交媒体的流行,用户移动行为数据得到极大丰富,推动了用户轨迹识别问题相关研究.该问题着眼于判定目标轨迹所属用户,有助于理解用户移动模式,促进个性化推荐等下游应用.目前已有方法通常尝试采用多分类方法解决该问题.然而,这些研究仍然面临着两个亟待解决的挑战:用户轨迹稀疏和类别数量庞大.其中,前者产生的原因在于用户常常只在社交媒体中发布部分访问地点信息,并且用户兴趣地点分布具有局部性;后者则是因为多分类方法下每个类别即代表一个用户,而用户数量庞大.为应对上述挑战,本文致力于有效利用轨迹时间戳序列和用户社交关系这两类被相关研究忽略的信息.一方面时间戳信息能够缓解轨迹稀疏性,另一方面社交关系可以通过刻画用户(类别)之间相关性帮助用户表示学习.为此,本文提出了融合神经时间点过程与图神经网络的新模型NTPP-GNN(Neural Temporal Point Process with Graph Neural Network),包含空间、时间、社交关系三个模块.空间模块中,双向循环神经网络用来刻画地点间序列性;时间模块中,本文提出双向神经时间点过程从正反两个方向捕捉时间连续性,并以此促进轨迹的时间表征;社交关系模块中,图神经网络用于传播和学习用户表示.NTPP-GNN采用端到端方式学习上述三个模块,以确保模块之间彼此适配.为验证NTPP-GNN的有效性,本文在三组数据集(Foursquare、Gowalla和Brightkite)上进行实验.结果表明:(1)NTPP-GNN性能比最好的基准方法在ACC@1上平均提高7.0%;(2)NTPP-GNN的各个模块对于性能均有贡献;(3)所提出的双向神经时间点过程相比于只考虑先后顺序的传统神经点过程方法能够带来额外提升.With the flourish of location-aware online social platforms,user behavior of mobility has been greatly enriched,which promotes the relevant studies on the user trajectory identification problem.For a given target trajectory,this problem aims to identify a specific user that the trajectory belongs to,which is beneficial for understanding the mobility patterns behind user trajectories and could provide positive influence on a variety of downstream applications,such as personalized recommendation,to name a few.By far,a few existing relevant studies try to utilize multi-class classification methods to tackle the problem.However,these studies still face two main unresolved challenges which need to be addressed in the literature:the sparsity of user trajectories and the large number of categories to be used for classification.Among the two challenges,the reason of the first one is that users tend to choose only a limited number of visited locations to be published in online social medias and their preferred visited POIs are distributed in local regions;the second challenge is caused by the fact that in multi-class classification,each category denotes one user and the number of users is large.To address the above two challenges,this paper aims to effectively utilize the two types of information,i.e.,sequences of timestamps in user trajectories and social relations among users,both of which have not been investigated by previous studies for the considered problem.On the one hand,timestamps could be used as additional information to alleviate the sparsity issue.On the other hand,social relations could be leveraged to characterize the correlations between users,which in turn helps user representation learning.To effectively leverage the two types of information for the studied problem,we propose a novel model which couples Neural Temporal Point Process with Graph Neural Network,named NTPP-GNN.This model composes of three modules for the spatial aspect,the temporal aspect,and the social aspect,respectively.In the spatial mo

关 键 词：用户轨迹识别 循环神经网络 神经时间点过程 图神经网络 时空序列 社交关系 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]