Pre-training method based on dynamic graph neural network

Abstract

A pre-training method based on a dynamic graph neural network learns node representation from three aspects of time, structure and semantics, and comprises the following steps: determining the size of a sampling sub-graph according to actual requirements and system performance, and performing sub-graph sampling on large-scale dynamic graph data by using a time sensitive sampling algorithm to obtain a sub-graph; for the sub-graph, performing covering processing on the sub-graph by using a time-sensitive edge covering algorithm and a node feature covering algorithm to obtain a processed new sub-graph; a dynamic graph generation algorithm is combined with a GNN model to predict covering edges and covering node features of the sub-graphs, optimal parameters are stored, and the pre-training process is ended; and loading the optimal parameters, and performing fine adjustment on the predicted graph data according to different downstream tasks to obtain a final result. According to the method, large-scale dynamic graph data can be processed, and compared with other pre-training methods, the method is higher in expression ability, learned nodes are more accurate in expression, and the method can be better applied to various downstream tasks.

Description

technical field [0001] The invention relates to the field of graph neural networks in deep learning, in particular to a pre-training method based on dynamic graph neural networks. Background technique [0002] A graph or network is an effective abstraction of real-world data structures, such as citation networks, social networks, protein networks, transportation networks, etc. Traditional network analysis methods require researchers to manually extract features, depending on network type, researcher's experience and specific tasks. In recent years, with the continuous development of deep learning, the GNN model has made breakthroughs in various graph mining tasks (such as node classification, link prediction, etc.) by virtue of the advantages of end-to-end learning and powerful reasoning methods. GNN takes node features and graph structural features as input, and through multi-layer information transfer (ie, graph convolution), obtains low-dimensional dense node representat...

AI Technical Summary

Problems solved by technology

[0004] However, the existing graph neural network pre-training models only focus on the features and structural information at the node or graph level, ignoring the time information carried by the edges in the graph.

Images