Graph neural network data sampling method and device, equipment and storage medium

Abstract

The invention discloses a graph neural network data sampling method and device, equipment and a storage medium. According to the scheme, vertexes of an original image data set are clustered, and the training vertexes are sorted according to the clustering categories of the training vertexes, so that a batch of training vertexes can be sampled in the same cluster at the same time during sampling in the sampling process, and the data locality of sampling is improved; moreover, as the training vertexes in the same cluster generally have more similar attributes and are closely connected parts, and connection among different clusters is little, the neighborhood vertexes expanded in the same cluster are concentrated in the same cluster, and the vertexes in the same cluster are close in storage, so that the sampling data locality can be improved, the range of neighborhood expansion is limited, and the efficiency of sub-graph sampling is improved.

Description

technical field [0001] The present invention relates to the technical field of graph data sampling, and more specifically, to a graph neural network data sampling method, device, equipment and storage medium. Background technique [0002] At present, graph data, as a kind of unstructured data, has been widely used in recommender systems, social networks, knowledge graphs and other fields. Graph neural networks have emerged as powerful tools for processing graph data. Different from image recognition and sentence processing, the existing graph neural network model training process needs to sample each batch of training samples on the CPU (Central processing unit, central processing unit) before data loading, and process them on the GPU (graphics processing unit). , graphics processor) for model calculations. However, graph data in the real world is a complex and irregular structure. The sampling and traversal process of graph data involves irregular memory access, which lea...

AI Technical Summary

Problems solved by technology

However, graph data in the real world is a complex and irregular structure. The sampling and traversal process of graph data involves irregular memory access, which leads to randomization of graph data access and poor data locality, and poor locality will lead to frequent global data access, and significantly increase the memory access time, which in turn causes the sample batch sampling time on the CPU to be longer than the model calculation time on the GPU, which leads to an unbalanced workload of the pipeline unit based on the CPU-GPU architecture, which damages the Pipeline Training Performance

Moreover, another problem brought by sample batch sampling is that neighborhood expansion introduces significant computational overhead

For example, when training an L-layer graph neural network model, it is necessary to sequentially sample the 1st to L-order neighborhoods of the target vertices, which leads to an exponential increase in time complexity with the depth of the graph neural network

This neighborhood explosion problem involves a lot of random memory access, resulting in poor data locality, and the sampled data may even exceed the GPU memory

Images