Research On Dynamic Graph Representation Learning Based On Contrastive Learning

Graph structure data(such as social networks,citation networks,protein interaction networks,etc.)can naturally represent entities and relationships between entities.In recent years,graph data mining has become a hot research topic,in which graph representation learning is the core content of graph learning,which generally refers to embedding node attributes and structure into low-dimensional dense vector space,so as to better perform node classification,link prediction and other downstream tasks.Graph representation learning based on graph neural network(GNN)has received a lot of attention recently and has shown excellent performance.However,such methods still have the following challenges:(1)The mainstream graph neural network models based on supervised and semi-supervised learning paradigms over-rely on label information,while many graphs in the real world are unlabeled,and the cost of acquiring labels is high.(2)The structure of the realistic graph is constantly evolving,and the nodes and edges in the graph are dynamically changing,while most static graph representation learning methods fail to take advantage of the important time information of the original graph data.To solve the above problems,the following work is carried out in this paper:(1)This paper extends the contrast learning paradigm to continuous time graphs and proposes a self-supervised dynamic graph representation learning method(Dy Sub C)based on temporal subgraph contrast.Firstly,a new temporal subgraph sampling strategy is proposed to sample the corresponding temporal subgraph for each node.Then,the temporal subgraph is encoded by encoder,and the subgraph representation function is designed to get the temporal subgraph representation.Finally,the mutual information between the node representation and the temporal subgraph representation is maximized by comparing both structure and temporal.Experiments of link prediction and node classification on real dynamic graph datasets have verified the advance of the Dy Sub C method and the validity of the time enhancement strategy.(2)This paper extends the contrast learning paradigm to discrete time graphs,and proposes a self-supervised dynamic graph representation learning method(DGCP)based on contrast prediction.Firstly,the graph neural network is used to encode each snapshot to get the corresponding node representation.Then,the autoregressive model is used to predict the node representation in the next snapshot.Finally,the model is trained end to end by contrast loss and sliding window mechanism.The validity of DGCP method is verified by link prediction experiments on real dynamic graph datasets.In this paper,two dynamic graph representation learning methods based on contrast learning are studied from the perspective of two different definitions of dynamic graph data(continuous time graph and discrete time graph).In the future,we can continue to study self-supervised dynamic graph representation learning from the perspective of scalability and mobility.