The Research On Temporal Link Prediction Method Based On Node Behavior Characteristics

Networks are effective tools for modeling various complex systems,where nodes and links in the network are abstract representations of entities and relationships between entities.Using network analysis to explore the dynamics of various complex systems and guiding the production and construction process based on theoretical foundations has important practical significance for network science.Most real-world networks are in a continuous process of evolution,thus static networks with fixed links cannot reflect the dynamic characteristics of the network.Temporal networks that include time information can better express the dynamic nature of real-world networks.The task of link prediction on temporal networks is usually referred to as temporal link prediction,which aims to infer the likelihood of two nodes in the future network being linked based on their historical link information.Temporal link prediction is one of the fundamental problems in network science.In terms of theory,link prediction methods can provide effective ways for researchers to explore and understand the dynamic principles of complex evolving systems.In practical applications,prediction methods can be used for friend recommendations on social platforms,product recommendations on e-commerce websites,and predicting interactions in biological networks.Therefore,temporal link prediction has important theoretical research value and broad practical applications,and has received interdisciplinary attention in recent years.The key to temporal link prediction is to organize and utilize the complex entanglement relationship between time and structural information in the network reasonably.To achieve this task,this thesis designs relevant prediction methods based on node behavior characteristics and verifies the effectiveness of these characteristics.Node behavior characteristics refer to dynamic characteristics,link tendencies,and spatiotemporal patterns exhibited by nodes in their historical interactions in the network.This thesis will explore how to extract node behavior characteristics from node’s historical interaction sequences and apply them to the link prediction task.The main research contents of this thesis are as follows:(1)Proposing the Action-based Graph Neural Network(ActGNN)framework for capturing node behavior characteristics on a continuous-time form temporal network,which to some extent makes up for the lack of continuous-time modeling in existing research.ActGNN captures the behavior characteristics of nodes and their neighbors on a temporal network by first anonymizing the node interaction sequences and preserving the time information of interactions through time encoding to obtain the anonymous interaction sequences of nodes.Then,the gated recurrent unit(GRU)is used to encode the node’s anonymous interaction sequence and obtain the spatiotemporal features of the node interaction.The graph attention network(GAT)perceives the differences in behavior characteristics between nodes and their historical neighbors and conducts biased aggregation,finally obtaining the embedding expression of the node at a certain time.Through the above process of extraction,aggregation,derivation,and representation,the ActGNN model can obtain the temporal embedding expression of the node and apply it to link prediction.(2)Proposing a heuristic and interpretable prediction model: the Develop-Maintain Activity Backbone(DMAB)model.DMAB uses a heuristic modeling process to make the method highly interpretable as a whole,to a certain extent compensating for the lack of interpretability in existing methods.DMAB focuses on the dynamic process of node interactions,abstracts and defines two types of node dynamics in the link generation process: node activity and node loyalty.Activity refers to the level of node acquisition,and loyalty is the tendency of nodes to maintain the current link status.These two types of node dynamics are quantified through the solution of Hawkes processes and temporal entropy,respectively.Based on the quantified node dynamics,the two types of node behavior characteristics are organically combined using a heuristic modeling method to express the dynamic process of nodes when link generation occurs.This thesis the application of node behavioral characteristics in temporal link prediction from two perspectives: heuristic modeling methods and mainstream deep learning methods.The definition,quantification,and organization of node behavioral characteristics are investigated.The performance of the proposed method and the powerful predictive ability of node behavioral characteristics are validated through experiments on a large number of real networks.