The Research And Implementation Of Dynamic Graph Representation Learning Based On Ordinary Differential Equation

The network is a ubiquitous and complicated structure,which characterizes the real-world relations ranging from social media networks,academic networks to e-commerce user-item networks.Representation learning on networks has been the focal study topic from researchers in recent years.The core idea behind is to learn low-dimensional vectors preserving the structure and attribute properties at the node level.which can be used for various downstream tasks,such as node classification,link prediction,community detection and graph classification,etc.More recently,with considerable efforts from researchers,graph neural networks,have been widely used pipelines for network representation learning.These aforementioned methods are generally designed for a given static graph.However,most real-world networks are naturally dynamic and evolve over time.The inherent dynamics of networks urge the dynamic graph methods should encode the temporal evolution of relational data.Dynamic network representation has aroused increasing attention from researchers since it reflects the characteristics of real-world networks and has been widely used in various applications.Enormous efforts have been made to this problem.Existing studies often use fixed-time steps.However,in fact,the dynamics in real-world networks are usually irregular over time.To address this problem,the thesis employs ordinary differential equations and proposes a novel graph representation approach for dynamic networks,namely Graph-ODE.The core idea lies in that it leverages ordinary differential equations to model the continuous latent dynamics of the network and therefore can naturally capture the evolution pattern.Specifically,to deal with the irregularity of node interactions,the thesis first proposes a differential form of GRU aggregator to aggregate the chronological neighbors.In addition,the thesis employs neighbor sampling strategies for efficiency.Besides,the designed inductive algorithm owns the ability to deal with new emerging nodes.Finally,this thesis optimizes the deficiencies of GraphODE,and proposes GraphCDE based neural controlled differential equations,which solved the influence of node interaction on the dynamic evolution of the network system and can be applied to downstream tasks perfectly.Experimental results on various real-world data sets demonstrate that the dynamic graph network representation learning algorithm based on differential equations and its improved algorithm proposed in this thesis can effectively model the dynamic representation of nodes in the graph network to perform more accurate prediction tasks and improve the performance of node classification and link prediction.