Research On Characterization Methods Of Time Series Dynamic Network

With the era of information explosion,the scale of the network is also facing explosive development,so whether the network can make effective and reasonable data analysis is a hot topic in today's academic research.There are two forms of network: static and dynamic.Traditional representation learning is based on static network,but in the real world,the network is dynamic,so the research on dynamic network has more practical significance.There are many forms of dynamic network: dynamic network with topology change,dynamic network with interaction change,dynamic network with both topology and interaction change.Temporal dynamic network and non temporal dynamic network.Topological changes refer to changes in the network structure,including the increase of nodes,the decrease of nodes,the increase of edges,and the decrease of edges.Interaction refers to the communication between nodes through the edge,such as the information transfer between nodes.A network with changes in both interaction and topology is a combination of the first two types of networks,and has the characteristics of topology change dynamic network and interaction change dynamic network.Temporal dynamic network is based on temporal variation,such as temporal network with topological variation,while non temporal dynamic network is based on sequential variation,such as topological variation dynamic network and interactive variation dynamic network.According to the size of network,the network can be divided into small-scale dynamic network and large-scale dynamic network.In the past,only a single kind of dynamic network is considered,and other dynamic networks are ignored.This paper studies three kinds of dynamic networks,which are topology change time series dynamic network,large-scale dynamic network of topology change time series,dynamic network of interactive change time series,and proposes three dynamic network representation methods.Firstly,this paper proposes a new method RGAT(recurrent graph attention networks)based on recurrent neural network(RNN)and graph attention network(GAT)to learn topological change time series network representation.There are many kinds of dynamic network with topology change.In this paper,we mainly focus on the situation of edge adding and deleting.For topology changing time-series network,it is necessary to learn not only the topology structure but also the change rule based on time-series topology.Therefore,this paper uses GAT to learn the topology structure and node information of the network,and then uses RNN to learn the time-series change rule and update the node representation.Secondly,this paper proposes a new method RSage(recurrent sampling to aggregate neighbor nodes)based on recurrent sampling,which is used to learn the representation of large-scale topology changing sequential networks.When the edges appear and disappear frequently in the network,the topological structure of the graph will be changed directly,so the representation also needs to be adjusted.But if all the node representations are adjusted according to the traditional method,the calculation cost of the whole model will be very large.The research finds that the node representation changes far from the distance(the minimum steps taken from the source node to the destination node)is not obvious.Therefore,it is not necessary to update all the node representations,only the node representation of the hop number comparison needs to be updated.This model is more reasonable for the learning of a dynamic graph.Therefore,this paper proposes RSage model to learn large-scale dynamic network representation.In this paper,we first preprocess the large-scale dynamic network,then obtain the initial representation of the current node by sampling and aggregating the neighbor node information,and then use RNN to capture the rule of topology change at any time and update the node representation to get the final node representation.Finally,this paper proposes a new model RUP(recurrent update propagation)to learn the representation of interactive temporal networks.There are many types of interaction changes.For example,the practical significance of node interaction in social networks is that there are message exchanges between users,and that there is mail delivery between users in e-mail networks.The occurrence of an interaction action represents the exchange of information.The interaction of nodes in the graph may change the characteristics of nodes,but the existing research often ignores the impact of interaction changes on the network,so interaction changes are also a research focus of dynamic graph.Therefore,this paper proposes RUP model to learn dynamic network representation of interactive change,which is mainly divided into two modules: update module and propagation module.The update module is used to update the representation of the node with interaction,and the propagation module is used to update the representation of the node affected by interaction.