Unsupervised relation prediction method based on depth map network auto-encoder

Abstract

The invention discloses an unsupervised relationship prediction method based on a depth map network auto-encoder. The method comprises the following steps: collecting an entity pair data set X with mpairs of target relationships and n entities; preprocessing the entity pair data, and dividing the entity pair data into K combinations of a test set and a training set; converting the test set into atest matrix Atest, and converting the training set into a training matrix Atrain; normalizing the training set matrix Atest to define an encoder and a decoder for an adjacent matrix, and building a prediction model; traversing k test sets and training sets by the prediction model for training and relationship reconstruction to obtain k trained prediction models; and traversing the k trained prediction models to obtain prediction of the target relationship. According to the method, target relationship prediction is completed by adopting a graph convolutional neural network, and a transductionproblem of a node embedding vector is changed into a conclusion problem during edge prediction of a graph.

Description

technical field [0001] The invention relates to the field of big data analysis, in particular to an unsupervised relationship prediction method based on a deep graph network autoencoder. Background technique [0002] The use of computation-based relationship prediction algorithms is an important application of machine learning. The current main method is to migrate the matrix factorization algorithm of the recommendation problem to this problem. However, when this method is migrated to the node representation on the graph, it will become a direct encoding method, independently training a unique embedding vector for each node, which has the following shortcomings: [0003] (1) There is no parameter sharing between nodes inside the encoder (i.e., the encoder is only queried about the embedding vectors of arbitrary node identification numbers). Since parameter sharing can be used as a powerful regularization form, this will cause the problem of low statistical validity, and b...

AI Technical Summary

Problems solved by technology

Since parameter sharing can be used as a powerful regularization form, this will cause the problem of low statistical validity, and because the calculation of the embedding vector of each node is independent, it will lead to low computational efficiency. This means that the number of parameters necessarily increases with increased by

[0004] (2) The direct encoding method cannot take advantage of the attributes of the nodes during encoding

[0005] (3) The direct encoding method has an inherent transduction problem, that is, it cannot generate embedding vectors for nodes not seen in the training set

However, the general graph neural network edge prediction model can only solve the first two of the above problems, and cannot complete the node embedding vector from the transduction problem to the induction problem when edge prediction is performed.

Images