Research And Implementation Of Dynamic Social Network Alignment Based On Presentation Learning

With the rapid development of the Internet and mobile devices,a variety of online social network platforms continue to emerge,and more and more people participate in online social network platforms to share and exchange information.Because a user will be active in multiple social networks at the same time,social network alignment research is initiated.Its goal is to link users in different social networks through account information to find the accounts of the same user in different social networks.Linking the same user identity on multiple social networks can not only assist downstream tasks such as cross-domain recommendation and cross-domain link prediction,but also solve problems such as cold start and data sparseness.Network alignment task is a hot topic in the field of online social network analysis,which has great value in academic research and commercial application.Most of the existing social network alignment methods regard social networks as static,and use static social network information to complete the task of network alignment.Inspired by the dynamic nature of social networks,some scholars have begun to study dynamic social network alignment.However,the existing dynamic social network alignment methods only focus on the single factor dynamics of social networks,and fail to fully explore the dynamic factors of social networks.As a result,accurate user vector representation cannot be learned,thus affecting the accuracy of network alignment.Based on this,we propose a dynamic social network alignment model called DHomo-SNA based on multi-factor dynamics,which for the first time considers the dynamics of network structure,user location and user generated information in the social network alignment problem,and classifies the importance of user location and user generated information at different times.DHomo-SNA model can be divided into two parts:user dynamic feature extraction stage and network alignment stage.In the phase of user dynamic feature extraction,network structure,user location and user generated information are used to learn user dynamic feature.In the network alignment phase,the learned user vector representation is used to predict anchor connection.On the other hand,most existing social network alignment methods regard social networks as homogenous networks.With the continuous development of online social media,more and more heterogeneous information is contained within the social network.Building into a homogeneous social network causes weak data expression ability,and ignoring the social network of potential interactions between different types of objects,fails to learn accurate user vectors,which affects the alignment accuracy.Based on this,this paper for the first time comprehensively considers the heterogeneity and dynamics of social networks in social network alignment task,and proposes a dynamic heterogeneous social network alignment method DHete-SNA-DALPA based on multi-layer graph attentional neural network.The multi-layer graph attentional neural network of the model propagates and learns the dynamic heterogeneous social network according to the time slice order,which can not only learn the accurate user vector representation,but also capture the dynamic of the heterogeneous social network.Moreover,based on the prediction results of historical anchor connections,this paper proposes a dynamic anchor connection prediction algorithm DALPA with supervised learning.The algorithm selects trusted anchor connections from the prediction results of historical anchor connections to expand the known anchor connection set,so as to improve the accuracy of network alignment.In order to verify the effectiveness of DHomo-SNA and DHete-SNADALPA methods,a dynamic network alignment evaluation system is designed and implemented in this paper,and two different real datasets are used for experiments.Multiple groups of comparative experiments are conducted by changing the proportion of training sets,vector dimensions and anchor link prediction algorithm.Compared with other benchmark methods,it is proved that the proposed methods have obvious advantages in network alignment tasks.This thesis firstly introduces the existing social network alignment methods,and analyzes the problems and challenges of existing methods.Then,the dynamic social network alignment method DHomo-SNA based on multi-factor dynamics and the dynamic heterogeneous social network alignment method DHete-SNA-DALPA based on multi-layer graph attentional neural network are introduced in detail.Then it introduces the design and implementation of dynamic social network evaluation system,and verifies the effectiveness of the proposed scheme through experiments.Finally,this paper summarizes the achievements and future work.