Node classification method, model training method and device

Abstract

The invention discloses a model training method, which comprises the following steps: acquiring a node subset to be trained from a node set to be trained; the number of nodes in the node subset to betrained is smaller than the number of nodes in the node set to be trained; Determining a node characteristic vector quantum set according to a node subset to be trained; According to the subset of thetarget node and the characteristic of the node to the quantum set, the subset of the prediction class probability is determined, and the prediction class probability and the target node have the corresponding relation; Determining a second model parameter according to the predicted class probability subset and the first model parameter; A node classification model is obtained by training according to the parameters of the second model. The invention also discloses a node classification method and a server. As for that large-scale map, the invention can only carry out train based on a part ofnode, and part of nodes in the map are calculated in each iteration, and each node in the map need not be traversed, thereby greatly reducing computational overhead and save computational resources.

Description

technical field [0001] The invention relates to the technical field of the Internet, in particular to a node classification method, a model training method and a device. Background technique [0002] With the development of machine learning technology, image classification tasks have made remarkable progress, and related technologies have been applied to many scenarios such as autonomous driving, security, and games. Similar to pictures, graphs are also a common type of data, such as social networks, knowledge graphs, and drug molecular structures. Different from pictures, there are still relatively few studies on the use of machine learning methods to classify graph nodes. [0003] At present, for large-scale graphs, the integration of graph node information requires special iterative calculations. see figure 1 , figure 1 A schematic diagram of the graph convolutional neural network (graph convolutional networks, GCN) classification model for graph node classification i...

AI Technical Summary

Problems solved by technology

However, in order to update the feature information of each node, GCN needs to traverse all the neighbor nodes of each node in one feature information calculation, resulting in high calculation cost and excessive calculation resource consumption

Images