Research And Implementation Of Distributed Graph Inference Computing Technology Oriented To Graph Neural Network Model

A graph is a structure that can effectively express the relationship between objects and objects,and the emergence of graph neural networks enables deep learning techniques to be applied in the field of graphs.However,the acquisition cycle of the graph neural network model is relatively long,and it needs to go through processes such as data cleaning,model training,and parameter tuning.For scenarios with very high real-time requirements,a long period of time will bring many potential risks,resulting in offline training of graph neural network models can only be used as one of the auxiliary means.In addition,due to the huge amount of graph data,a single computer is often unable to carry an excessive amount of data and computation.Distributed graph computing has become a mainstream trend.Therefore,how to use the graph neural network model for real-time graph inference computation in a distributed cluster environment is an urgent problem to be solved.In addition to real-time performance,the accuracy of the graph inference calculation process is also very important,which also means that the effect of the graph neural network model in the continuous iterative inference process needs to be maintained.Aiming at these problems,this paper proposes the corresponding framework and algorithm,implements the corresponding system and conducts a large number of experimental tests to verify its effectiveness.The main contribution of this paper consists of the following three parts:(1)A distributed graph inference framework based on graph neural network model is proposed.The framework consists of three modules,namely incremental composition,GNN encoder,and GNN decoder.The incremental composition module will firstly store the graph data in partitions,then construct the full graph based on the partitioned vertices and edges,and then perform the basic graph computation on the full graph.The GNN encoder module and the GNN decoder module disassemble the graph neural network module,and the disassembled"encoding-decoding" form reduces the coupling degree of the inference process.Among them,the encoder is responsible for encoding the graph features to obtain the embedding to update the vertex features,and the decoder is responsible for decoding the full-graph features to obtain the execution effect of the task.These two models follow the principle of "distributed storage and dynamic invocation" in a distributed environment.The experimental results show that the framework can apply the GNN model to the distributed environment for real-time inference.It greatly improve the upper limit of the system at a small cost of time consumption,and maintain a certain model effect.(2)A distributed graph inference calculation method based on neighborhood information is proposed.On the basis of the above framework,the method additionally introduces the neighborhood information of graph events,that is, the second-degree subgraph module and the mailbox mechanism are added.The computation of second-degree subgraphs will expand the scope of influence of a single graph event.The mailbox mechanism sends the relevant information of the graph events to the neighbors in the form of characteristic messages,so that the neighbors can directly obtain historical activity data when inferring.In addition,the method adopts the idea of "single-point inference,message passing, distributed computing" to improve the calling inference process of the graph model.The experimental results of this method show that the neighborhood information can provide richer graph information,so that the inference computing task achieves better results and still maintains considerable timeliness.(3)Realize a complete distributed graph inference computing system based on the above framework and algorithm.Reasonable design ideas and implementation methods make the system have good flexibility and robustness.