Research On Dynamic Network Characterization Method Based On Node Interest

Recently,computer network technology has grown rapidly.a wide range of software has arised,often generating large amounts of data.Studying the patterns behind existing data can help people improve efficiency and accomplish some specific tasks.The data in life are connected with each other.Through some mathematical modeling methods,complex data relationships can be abstracted into graph structures.The study of graph structures is an important means to analyze the laws behind data.Previous research is limited to the study of static graphs,while things in reality often change over time,and the corresponding abstract graph structure also changes over time.This kind of graph that changes over time is for dynamic graphs.The research on dynamic graphs is more in line with real life and is one of the most popular research topics.Changes in dynamic graphs mainly include changes in nodes and edges.Changes in nodes often represent changes in entity members,while changes in edges represent changes in relationships and interactions.In order to study the meaning behind dynamic graphs,it is most basic to study the representation of dynamic graphs.This paper proposes three different approaches for representation learning on dynamic graphs.First,we use a novel representation model called ETIV(Embedding Temporal Network via Interest Vector)based on multi-head attention mechanism and interest vector.Inspired by real-life entity-to-entity interactions,entity interactions often depend on their own interests.In the process of learning the representation vector,the multi head attention mechanism can get the information of the historical interaction sequence of nodes,so as to generate the interest vector of nodes.In the process of processing historical interactive node information,the attenuation effect of time on information is integrated to simulate the loss and forgetting of information.Finally,the possibility of interaction between nodes is optimized,and the representation of nodes is updated synchronously,so as to learn node representation of dynamic network.Secondly,we use a novel representation model called Dy ITrans(Learning dynamic representations via interest transfer functions)based on transfer functions and interest vectors.The existence form of node representation vector is in European space.In order to describe the changes of interest vector,this paper constructs a spatial transfer function to represent the changes of interest vector in European space.Through the transfer function in Euclidean space,the interest vector of nodes can be easily obtained.Then,the optimization algorithm can trim the function and correct the vector of nodes.By combining the interest vector and transfer function of nodes to describe the change of node interest,we can synchronously learn the node representation in dynamic networks.Finally,we use a representation model called Evo IV(Learning Dynamic Representations via Interest Evolution)that evolves node interest vectors.The interaction of nodes in the time-series network exists in the form of sequence,and the recurrent neural network is conducive to learning the ordered data.Therefore,this paper uses recurrent neural network to evolve the interest vector of nodes.In addition,the temporal attention mechanism is introduced to capture the interest of nodes at different times.The evolution of node interest vectors is described by a recurrent neural network.And we composed it with a attention mechanism to generate interest vectors for different moments.The final optimisation of the objective function is achieved by updating the node representations simultaneously,thus enabling the learning of node representations.