Bipartite Dynamic Longitudinal Sparse Network Model And Its Application

Since Euler put forward the "Seven Bridges Problem of Gonesburg" in 1973,social network analysis and complex network theory have emerged successively in the 1970 s and the end of the 20 th century,adding a significant touch to the process of network research.Network data provides an intuitive visual way to express data structures in real life,such as transportation networks,social networks and community networks.Different from traditional network data,bimodal data is composed of two types of participants,and the observable network relationship only exists between different types of participants.Therefore,how to obtain the potential network relationship of the same type of participants from the observable network relationship between different participants has become a hot issue for modern statisticians.With the advent of the big data era,researchers began to extract effective information from network data.It has also become an important topic to apply big data analysis methods to mining the evolution model of international relations.In political research,the behavior of one country often affects another country,and such ability is called "national influence".Through the identification and monitoring of global events and national behaviors,the "national influence" among different countries can be quantified,thus forming a "national influence" network.As a big data oriented information mining method,network mining is widely used in national influence network analysis because of its concrete network structure,which is an important method for international influence network research.However,at present,most of the research on potential networks is static,and there are still many gaps in the research on dynamic potential networks.However,in practice,due to the complexity of the real world,the edges,nodes and node relationships in the real network world will change over time.Dynamics is also a crucial part of traditional network research.Network evolution and network dynamics are very important topics.On the other hand,in the bipartite longitudinal relational data,because the relationship between nodes only exists between different types of nodes and the relationship is relatively simple,the network data often has zero inflated characteristics.Therefore,this paper attempts to promote and research the existing Bipartite Longitudinal Network Model(BLIN)in both theoretical and empirical aspects,further expanding the scope of application of the model and improving the prediction accuracy of the model.This paper mainly studies the following three points:First,in the general BLIN model,we characterize the dynamics of potential networks in bipartite longitudinal relational data.Considering the dynamic evolution of the network,this paper improves the existing BLIN model,expands the original potential network that does not change with time to have time-varying characteristics,and proposes a Bipartite Dynamic Longitudinal Network Model(BDLN),which can better capture the dynamic changes of the potential network,and improve the flexibility and interpretability of the model.At the same time,social network characteristics such as input and output are introduced to explore the changes of network structure,further enriching the practical application of network research.Second,the zero inflated characteristics of observable networks and the sparsity of potential networks are considered simultaneously in the bipartite longitudinal relational data.In the observable network,we consider zero inflated characteristics and non-truncation,accurately depict the zero,positive and negative characteristics of network data,consider the sparsity of the network,and implement LASSO penalty on the model.A Bipartite Dynamic Longitudinal Sparse Network Model(BDLSN)is proposed and extended to a Multipartite Dynamic Longitudinal Sparse Network Model(MDLSN),which is estimated by the penalty likelihood method.This model can not only reduce the estimation dimension and avoid the occurrence of "curse of dimensionality",but also study the influence degree of both sides of the relationship,determine whether the two sides of the relationship are related,and predict the future relationship more accurately.Third,we apply the model proposed in this paper to the research of national influence network.Based on the Integrated Crisis Early Warning System(ICEWS)data,we identify the potential networks of source countries,target countries and interaction types,and analyze the potential networks in combination with the international background.At the same time,ICEWS data is used to compare the MDLSN model proposed in this paper with the Multipartite Longitudinal Network Model(MLIN).The study found that the smaller the lag order of the potential network of the source country and target country,the closer the network is,and the country’s response to an event is "timely" and the country’s series of words and deeds are "continuous".Further,the target country has "competition" for resources in the type of positive action interaction(MP).In addition,the model comparison results show that the estimation error of the proposed model is significantly reduced.The MDLSN model enhances the ability to interpret actual data,helps to explore the changing trend of international relations,and also provides a new direction for the evolution of international relations in the future.