Interpretable graph neural network prediction method, system and device

Abstract

The invention provides an interpretable graph neural network prediction method, system and device, and the method comprises the steps: enabling a plurality of nodes in a graph neural network to be mapped into a plurality of N-dimensional one-hot codes through a trained decision tree model, and enabling each one-hot code to correspond to one leaf node of a decision tree; inputting the one-hot codeof the node and the one-hot code of the at least one neighbor node into a trained prediction model, the prediction model comprising a propagation layer, a fusion layer, a splicing layer and a regression layer, the propagation layer obtaining at least one N-dimensional propagation vector based on the trained propagation model; enabling the fusion layer to obtain an N-dimensional fusion vector basedon the at least one N-dimensional propagation vector; enabling the splicing layer to splice the N-dimensional one-hot codes of the nodes and the N-dimensional fusion vectors to obtain 2N-dimensionalrepresentation vectors of the nodes; enabling the regression layer to obtain a prediction result based on the representation vectors of the nodes; and obtaining explanation of a prediction result of the node based on the decision tree and the 2N-dimensional representation vector of the node.

Description

technical field [0001] The present application relates to the field of graph neural network interpretability, in particular to an interpretable graph neural network prediction method, system and device. Background technique [0002] Graph Neural Network (GNN) is a kind of data used to represent complex relationships between entity objects, usually including multiple nodes representing entity objects and edges connecting multiple nodes. The graph neural network model can predict entity objects (such as users) based on nodes and edges between nodes. In practical applications, it is necessary to explain the prediction results of the graph neural network model. However, the large number of parameters contained in the graph neural network model makes it difficult to explain the prediction results. [0003] Therefore, it is desirable to provide an interpretable prediction method for graph neural networks. Contents of the invention [0004] One aspect of this specification prov...

AI Technical Summary

Problems solved by technology

In practical applications, it is necessary to explain the prediction results of the graph neural network model. However, the large number of parameters contained in the graph neural network model makes it difficult to explain the prediction results.

Images