Unbalanced node classification method based on minority class weighted graph neural network

Abstract

The invention discloses an unbalanced node classification method based on a minority class weighted graph neural network, and the method comprises the steps: 1, calculating a node membership value based on adjacency information for input graph structure data, so as to obtain the weighted feature information of a node in an embedded space; 2, executing a data oversampling operation in the embedding space to generate a new node; 3, cost sensitive learning is introduced in the training process of the edge predictor, and the trained edge predictor is used for obtaining adjacency information of a new node; and 4, constructing balanced graph structure data in combination with the features of the new nodes and the existing nodes and the adjacency information, performing node classification by using a graph neural network, and using a Gumbel distribution function as an activation function. According to the method, the neighbor aggregation of minority class nodes is enhanced, the bias to majority class nodes in the edge generation process is reduced, the convergence speed of the graph neural network model is improved, and the unbalanced node classification effect based on the graph neural network is remarkably improved.

Description

technical field [0001] The invention relates to the field of machine learning, in particular to an unbalanced node classification method based on minority class weighted graph neural network. Background technique [0002] Graph neural network plays a very important role in the analysis and mining of graph structure data, and has been effectively used in node classification tasks such as social, text, image and bioengineering. The node classification of graphs usually adopts the form of semi-supervised classification. In the graph structure data, only the label information of some nodes is known, and the label information of the remaining nodes is unknown. At present, most graph neural network algorithms are aimed at balanced data. However, in practice, due to the influence of data sampling bias and other objective reasons, the distribution of labeled data between categories is often unbalanced. For example, in the classification of webpage topics, some topics are easy to co...

AI Technical Summary

Problems solved by technology

However, the current method tends to ignore the influence of minority data on classification results in the stage of aggregation feature information and edge prediction. At the same time, there is a problem of convergence of node characteristics of different types in the process of information aggregation, which makes the classification effect decline.

Images