Relationship prediction method and device based on non-discrete masking variational graph auto-encoder

Abstract

The invention relates to the field of link prediction and node classification, in particular to a relation prediction method and device based on a non-discrete masking variational graph autoencoder. The method comprises the steps of obtaining a to-be-predicted population graph, obtaining a population masking graph and a feature matrix generated based on non-discrete bandwidth masking, inputting the population masking graph and the feature matrix into a trained graph self-supervised learning model, and predicting output of a dynamic variational graph self-encoder. A link prediction decoder or a node classification decoder is selected for processing according to different tasks, and the original attributes of the masking nodes are reconstructed through multiple decoders. According to the method, a potential feature matching loss function and an original attribute reconstruction loss function are calculated, model parameters are adjusted, adjustment is carried out in the iteration process according to a loss function result, and therefore efficiency is improved. The method can solve the problem that a current discrete masking strategy is not enough to learn a topological information representation class and a highly unstable local structure, and improves the certainty of masking data inference and the reliability of a self-supervision mechanism.

Description

Technical Field

[0001] The present invention relates to link prediction and node classification technology, and in particular to a relationship prediction method and device based on a non-discrete masked variational graph autoencoder. It can be specifically applied to the field of bioinformatics, such as protein subcellular localization prediction, genome sequence analysis, protein structure prediction, biological pathway analysis, drug targeted therapy, etc.; medical field, such as group disease prediction, protein-protein interaction relationship prediction, brain map functional area prediction, etc.; social network field, such as predicting friend relationships in link networks to improve user satisfaction; e-commerce field, such as product recommendations, etc.; aerospace field, such as identifying hidden link edges and abnormal link edges, improving network transmission capacity, and other application fields. Background Art

[0002] Currently, masked variational graph autoencoders have become a powerful tool in the field of self-supervised learning of graphs. However, from the perspective of message propagation in graph neural networks, existing discrete masking strategies cannot fully capture the topological structure information of the graph, which makes the reconstructed representation perform poorly when dealing with complex graph tasks. In addition, the non-Euclidean characteristics of graph data require more complex processing methods, and operating only in the original data space is often not enough to preserve the integrity and important information of the data. These shortcomings not only affect the training effect of the model, but also limit the performance of the model in practical applications.

[0003] Therefore, how to mask and reconstruct node samples in the high-order latent feature space has become a key issue in current research. Improving the certainty of masked data inference and the reliability of the self-supervision mechanism, while effectively retaining topological information during the masking process, is the core of solving this problem. By masking and reconstructing in the high-order latent feature space, the complex relationships of graph data can be better captured and retained, thereby improving the performance of the model in downstream tasks such as link prediction and node classification. This can not only significantly improve the accuracy and efficiency of the model, but also reduce the resource consumption of the model when processing large-scale complex graph data, and improve the feasibility and effectiveness of practical applications. Summary of the invention

[0004] In order to solve the technical problems that the current discrete masking strategy is not sufficient to learn the representation of topological information and that the highly unstable local structure increases the uncertainty of the inferred masked data and reduces the reliability of the utilized self-supervisory signal, the embodiment of the present invention provides a relationship prediction method and device based on a non-discrete masked variational graph autoencoder. The technical solution is as follows:

[0005] On the one hand, a relationship prediction method based on a non-discrete masked variational graph autoencoder is provided, the method is implemented by a relationship prediction device, and the method includes:

[0006] S1. Obtain the population map to be predicted;

[0007] S2. Obtain the masked reconstruction result based on the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model;

[0008] Among them, according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model, the masked reconstruction results are obtained, including:

[0009] Perform non-discrete bandwidth masking on the population map, calculate the bandwidth size, obtain the population mask map and the population visible map according to the bandwidth masking, obtain the mask map reorganization representation matrix according to the population mask map, input the mask map reorganization representation matrix into the application potential feature predictor to obtain the prediction representation matrix of the masked node, input the prediction representation matrix of the masked node into the multiple decoders to obtain the original attribute matrix of the reconstructed masked node, select link prediction decoding or node classification decoding according to the task, perform masked reconstruction operation on the original attribute matrix of the reconstructed masked node, and obtain the masked reconstruction result;

[0010] S3. Obtain link prediction results based on the masking reconstruction results.

[0011] Optionally, the training process of the graph neural network-graph convolutional network-graph self-supervised learning model in S2 includes:

[0012] S21. Construct a graph neural network-graph convolutional network-graph self-supervised learning network model.

[0013] S22. Obtain a sample population map, and apply non-discrete bandwidth masking to the sample population map to obtain a masked map and a visible map.

[0014] S23. Encode the masked graph through a variational graph autoencoder to obtain a reorganized representation matrix of the masked graph; encode the visible graph through a momentum variational encoder to obtain a representation matrix of the masked nodes.

[0015] S24, inputting the masked graph reconstructed representation matrix into the applied latent feature predictor to obtain the predicted representation matrix of the masked node, inputting the predicted representation matrix of the masked node into the multiple decoders to obtain the original attribute matrix of the reconstructed masked node.

[0016] S25. Construct a potential feature matching loss based on the representation matrix of the masked node and the predicted representation matrix of the masked node; construct an original attribute reconstruction loss based on the original attribute matrix of the reconstructed masked node and the original attribute matrix of the masked node.

[0017] S26. The potential feature matching loss and the original attribute reconstruction loss are combined by a weighted method to obtain the overall loss function.

[0018] S27. Adjust the parameters of the graph neural network-graph convolutional network-graph self-supervised learning network model through the overall loss function to obtain a trained graph neural network-graph convolutional network-graph self-supervised learning network.

[0019] Optionally, the construction of the graph neural network-graph convolutional network-graph self-supervised learning network model in S21 includes:

[0020] A variational encoder-decoder architecture model is constructed based on restricting the messages propagating through each edge to varying degrees.

[0021] The variational encoding-decoding architecture model is configured to obtain the graph neural network-graph convolutional network-graph self-supervised learning network model.

[0022] Optionally, the construction process of the variational graph autoencoder in S23 includes:

[0023] A variational graph autoencoder backbone is constructed, and a multi-layer perception layer is used as a projector to form the variational graph autoencoder; wherein the variational graph autoencoder backbone includes multiple graph attention layers or multiple graph isomorphism layers.

[0024] Optionally, the variational graph autoencoder in S23 is defined by an inference model, and the inference model is shown in the following equations (1) and (2):

[0025] (1)

[0026] (2)

[0027] In the formula, Represents the inference model formula of the variational graph autoencoder, represents the potential embedding representation matrix, represents the node feature matrix, represents the adjacency matrix, represents the standard Gaussian distribution, represents the posterior probability distribution of the vector representation, represents the mean of the node vector, Represents the variance of the node vector.

[0028] Optionally, the potential feature matching loss in S25 is as shown in the following formula (3):

[0029] (3)

[0030] In the formula, represents the potential feature matching loss, represents the set of masked nodes, represents a masked node, Indicated in The range of The representation vector of masked nodes, represents the predicted representation matrix of the mask node, Indicated in The range of The representation vector of masked nodes, Representation matrix representing the masked node.

[0031] Optionally, the original attribute reconstruction loss in S25 is as shown in the following formula (4):

[0032] (4)

[0033] In the formula, represents the original attribute reconstruction loss, represents the set of masked nodes, represents a masked node, Represents the reconstruction of the original attribute vector of the masked node, represents the original attribute vector of the mask node, Represents the scale factor.

[0034] On the other hand, a relationship prediction device based on a non-discrete masked variational graph autoencoder is provided, and the device is applied to a relationship prediction method based on a non-discrete masked variational graph autoencoder, and the device includes:

[0035] The acquisition module is used to obtain the population map to be predicted.

[0036] The masked reconstruction module is used to obtain the masked reconstruction result based on the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model.

[0037] Among them, according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model, the masked reconstruction results are obtained, including:

[0038] The population map is subjected to non-discrete bandwidth masking, and the bandwidth size is calculated. The population masked map and the population visible map are obtained according to the bandwidth masking. The masked map reorganization representation matrix is ​​obtained according to the population masked map. The masked map reorganization representation matrix is ​​input into the application potential feature predictor to obtain the predicted representation matrix of the masked nodes. The predicted representation matrix of the masked nodes is input into the multiple decoders to obtain the original attribute matrix of the reconstructed masked nodes. Link prediction decoding or node classification decoding is selected according to the task, and the original attribute matrix of the reconstructed masked nodes is subjected to masked reconstruction operation to obtain the masked reconstruction result.

[0039] The output module is used to obtain the link prediction result according to the masking reconstruction result.

[0040] Optionally, the masked reconstruction module is further used to:

[0041] S21. Construct a graph neural network-graph convolutional network-graph self-supervised learning network model.

[0042] S22. Obtain a sample population map, and apply non-discrete bandwidth masking to the sample population map to obtain a masked map and a visible map.

[0043] S23. Encode the masked graph through a variational graph autoencoder to obtain a reorganized representation matrix of the masked graph; encode the visible graph through a momentum variational encoder to obtain a representation matrix of the masked nodes.

[0044] S24, inputting the masked graph reconstructed representation matrix into the applied latent feature predictor to obtain the predicted representation matrix of the masked node, inputting the predicted representation matrix of the masked node into the multiple decoders to obtain the original attribute matrix of the reconstructed masked node.

[0045] S25. Construct a potential feature matching loss based on the representation matrix of the masked node and the predicted representation matrix of the masked node; construct an original attribute reconstruction loss based on the original attribute matrix of the reconstructed masked node and the original attribute matrix of the masked node.

[0046] S26. The potential feature matching loss and the original attribute reconstruction loss are combined by a weighted method to obtain the overall loss function.

[0047] S27. Adjust the parameters of the graph neural network-graph convolutional network-graph self-supervised learning network model through the overall loss function to obtain a trained graph neural network-graph convolutional network-graph self-supervised learning network.

[0048] Optionally, the masked reconstruction module is further used to:

[0049] A variational encoder-decoder architecture model is constructed based on restricting the messages propagating through each edge to varying degrees.

[0050] The variational encoding-decoding architecture model is configured to obtain the graph neural network-graph convolutional network-graph self-supervised learning network model.

[0051] Optionally, the masked reconstruction module is further used to:

[0052] A variational graph autoencoder backbone is constructed, and a multi-layer perception layer is used as a projector to form the variational graph autoencoder; wherein the variational graph autoencoder backbone includes multiple graph attention layers or multiple graph isomorphism layers.

[0053] Optionally, the variational graph autoencoder is defined by an inference model, and the inference model is shown in the following equations (1) and (2):

[0054] (1)

[0055] (2)

[0056] In the formula, Represents the inference model formula of the variational graph autoencoder, represents the potential embedding representation matrix, represents the node feature matrix, represents the adjacency matrix, represents the standard Gaussian distribution, represents the posterior probability distribution of the vector representation, represents the mean of the node vector, Represents the variance of the node vector.

[0057] Optionally, the latent feature matching loss is as follows (3):

[0058] (3)

[0059] In the formula, represents the potential feature matching loss, represents the set of masked nodes, represents a masked node, Indicated in The range of The representation vector of masked nodes, represents the predicted representation matrix of the mask node, Indicated in The range of The representation vector of masked nodes, Representation matrix representing the masked node.

[0060] Optionally, the original attribute reconstruction loss is as shown in equation (4):

[0061] (4)

[0062] In the formula, represents the original attribute reconstruction loss, represents the set of masked nodes, represents a masked node, Represents the reconstruction of the original attribute vector of the masked node, represents the original attribute vector of the mask node, Represents the scale factor.

[0063] On the other hand, a relationship prediction device is provided, which includes: a processor; a memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, any one of the above-mentioned relationship prediction methods based on non-discrete masked variational graph autoencoders is implemented.

[0064] On the other hand, a computer-readable storage medium is provided, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement any one of the above-mentioned relationship prediction methods based on non-discrete masked variational graph autoencoders.

[0065] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0066] In the embodiments of the present invention, it is of great significance in link prediction and node classification. It uses a momentum variational graph autoencoder method based on a bandwidth masking strategy, without the need for time-consuming manual annotation and classification, and can effectively control errors and improve the accuracy and efficiency of link prediction and node classification. This method can not only better reveal the potential relationships in the network structure, but also accurately classify node attributes and improve the effect of overall network analysis. Therefore, in various application scenarios, such as social network analysis, recommendation systems, knowledge graph construction, etc., it significantly reduces the misclassification rate and missed detection rate, reduces the consumption of human resources, and improves work efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0068] Figure 1 It is a flow chart of a relationship prediction method based on a non-discrete masked variational graph autoencoder provided by an embodiment of the present invention;

[0069] Figure 2 It is a model building and configuration process framework diagram provided by an embodiment of the present invention;

[0070] Figure 3 is a diagram of a multiple decoder structure provided by an embodiment of the present invention;

[0071] Figure 4 is an overall flow chart of input image masking reconstruction provided by an embodiment of the present invention;

[0072] Figure 5 It is a framework diagram of a population map application device provided by an embodiment of the present invention;

[0073] Figure 6 It is a block diagram of a relationship prediction device based on a non-discrete masked variational graph autoencoder provided by an embodiment of the present invention;

[0074] Figure 7It is a structural schematic diagram of a relationship prediction device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0076] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0077] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same. "of", "corresponding, relevant" and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same.

[0078] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0079] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0080] The embodiment of the present invention provides a relationship prediction method based on a non-discrete masked variational graph autoencoder, which can be implemented by a relationship prediction device, which can be a terminal or a server. Figure 1 The flowchart of the relationship prediction method based on the non-discrete masked variational graph autoencoder shown in the figure may include the following steps:

[0081] S1. Obtain a population map to be predicted.

[0082] In a feasible implementation manner, the present invention has wide applications in group disease analysis, biological field, medical detection field, recommendation system, network security and other fields.

[0083] S2. According to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model, the masked reconstruction result is obtained.

[0084] Among them, according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model, the masked reconstruction results are obtained, including:

[0085] The population map is subjected to non-discrete bandwidth masking, and the bandwidth size is calculated. The population masked map and the population visible map are obtained according to the bandwidth masking. The masked map reorganization representation matrix is ​​obtained according to the population masked map. The masked map reorganization representation matrix is ​​input into the application potential feature predictor to obtain the predicted representation matrix of the masked nodes. The predicted representation matrix of the masked nodes is input into the multiple decoders to obtain the original attribute matrix of the reconstructed masked nodes. Link prediction decoding or node classification decoding is selected according to the task, and the original attribute matrix of the reconstructed masked nodes is subjected to masked reconstruction operation to obtain the masked reconstruction result.

[0086] Optionally, the training process of the graph neural network-graph convolutional network-graph self-supervised learning model in S2 may include the following steps S21-S27:

[0087] S21. Construct a graph neural network-graph convolutional network-graph self-supervised learning network model.

[0088] Optionally, the above step S21 can be: constructing a variational coding-decoding architecture model based on restricting the messages propagated through each edge to varying degrees; configuring the variational coding-decoding architecture model to obtain a graph neural network-graph convolutional network-graph self-supervised learning network model; wherein the graph neural network-graph convolutional network-graph self-supervised learning network model includes a population graph structure bandwidth masking module, a population graph analysis input module, an implicit and explicit self-supervised integration training method improvement module, a variational coding-decoding efficiency improvement module, a predicted attribute similarity detection module, an isolated node processing module, a disease prediction module, a node classification module, a graph visualization module, and a loss function module.

[0089] The specific model construction and configuration process is as follows: Figure 2As shown, the present invention is based on the CNN (Convolutional Neural Network) -GCN (Graph Convolutional Network) -Hybrid Self-Supervised Learning Network (self-supervised learning network) model. The model as a whole adopts a variational graph self-encoding-decoding structure, which is divided into masking, encoding-decoding, implicit-explicit modules, and loss function modules. The model is implemented on the PyTorch architecture. After the graph is input, it is masked. The masked bandwidth ratio is calculated according to the number of adjacent nodes. The size is randomly masked in the bandwidth ratio between [0%, 100%] to process this part of the nodes. The masked masking matrix is ​​input into the variational graph autoencoder. The variational graph autoencoder is responsible for converting the masked graph into a low-dimensional latent space. It adopts a graph neural network-based architecture, with a series of graph attention layers or graph isomorphism layers as the backbone of the variational graph autoencoder. The masked visible graph is then input into the momentum variational encoder, which is responsible for encoding the original data of the masked samples and using the complete representation of these samples to provide stable optimization guidance for the predicted samples. When calculating the loss function, a balance coefficient is added when obtaining the weighting of the latent feature matching loss function and the original attribute reconstruction loss function. ,The obtained graph embeddings can be saved and used for downstream evaluation such as node classification and link prediction tasks.

[0090] S22. Obtain a graph structure that can represent the association and distinction between nodes with attribute similarity from a population graph, and use non-discrete bandwidth masking on the graph structure to obtain a masked graph and a visible graph.

[0091] In a feasible implementation, a masking graph masked by restricting information from a non-discrete distribution is obtained, and the bandwidth size is determined based on restricting the messages propagated through each edge to varying degrees, and then predicting the degree to which it is restricted. The graph neural network prediction module can be used to predict the potential relationship between diseases, such as predicting the relationship between diseases by inferring the similarity of their attributes, or predicting similar characteristic symptoms of different individual samples. Through this step, the model can be better trained, so that it can obtain better results in processing ordinary population graphs or complex multi-layer population graphs in downstream tasks, and can better distinguish between one-hop or multi-hop relationships between two nodes.

[0092] Specifically, the masking matrix is ​​constructed to deduce the masking graph, and an encoder is trained using the unmasked set, i.e., the visible set, and the entire node set. The specific steps are as follows:

[0093] S221, construct a masking matrix, determine whether a node is retained through an indicator, and calculate the masking bandwidth based on the surrounding nodes.

[0094] S221.1. Calculate the masked bandwidth. The bandwidth is between [0-100]. If there are R 1-hop adjacent nodes around the node, the bandwidth is 1 / R, that is, the maximum ratio of output messages to input messages through this edge in each message passing step is 1 / R.

[0095] S221.2. Construct a masking matrix M, where N is a continuous matrix probability simplex column vector, where each non-zero entry follows the Boltzmann-Gibbs distribution:

[0096] (1)

[0097] In the formula, represents the distribution calculation of a point in the masking matrix, represents a node in the neighborhood, Indicates the row where the point is located. Indicates the column where the point is located. represents a point in the feature matrix, Indicates temperature.

[0098] S221.3. Construct the visible matrix. After obtaining the masking matrix, the adjacency matrix of the visible matrix is , where A represents the adjacency matrix, represents the Hadamard product, and M represents the masking matrix.

[0099] S23. Encode the masked graph through a variational graph autoencoder to obtain a masked graph reorganization representation matrix; encode the visible graph through a momentum variational encoder to obtain a representation matrix of the masked nodes. The graph neural network prediction module can be used to predict the potential relationship between diseases, such as predicting the relationship between diseases by inferring the similarity of the attributes between diseases, or predicting similar characteristic symptoms of different individual samples. Through this step, the model can be better trained, so that it can get better results in processing ordinary population graphs or complex multi-layer population graphs in downstream tasks, and can better distinguish the one-hop or multi-hop relationship between two nodes.

[0100] In one feasible implementation, a graph neural network-based architecture is constructed, with a series of graph attention layers or graph isomorphism layers as the encoder backbone, and a multi-layer perception layer as a projector to form a variational graph autoencoder. The variational graph autoencoder is defined by obtaining the adjacency matrix and the node feature matrix. The specific steps are as follows:

[0101] S231.1. Build the backbone of the encoder, inputting the matrix of potential embedding representation matrix and mean vector.

[0102] S231.1.1. Input the adjacency matrix and node feature matrix and encode them. A variational graph encoder is defined by an inference model:

[0103] (2)

[0104] S231.1.2 For inference models For example:

[0105] (3)

[0106] In the formula, Represents the inference model formula of the variational graph autoencoder, represents the potential embedding representation matrix, represents the node feature matrix, represents the adjacency matrix, represents the standard Gaussian distribution, represents the posterior probability distribution of the vector representation, represents the mean of the node vector, Represents the variance of the node vector.

[0107] S24, inputting the masked graph reconstructed representation matrix into the applied latent feature predictor to obtain the predicted representation matrix of the masked node, inputting the predicted representation matrix of the masked node into the multiple decoders to obtain the original attribute matrix of the reconstructed masked node.

[0108] In a feasible implementation, the reassembled representation matrix obtained after the masking map is processed by the encoder is input into a potential feature predictor to obtain the predicted masked content.

[0109] The extracted masked node representation matrix is ​​trained based on the implicit and explicit self-supervision integration module to obtain the graph attention layer of the recovered masked content, which is responsible for restoring the original content of the masked nodes. Based on the restored content, the multi-layer perceptron (MLP) layer is used to further predict the potential features of the masked nodes.

[0110] Further, a training operation is performed on the prediction representation matrix of the masked node to generate a reconstructed original attribute matrix of the masked node. A decoding operation is performed on the prediction representation of the masked node based on a multiple decoder, and the multiple decoders are as follows: Figure 3 As shown in Figure 2. The multi-decoder is responsible for decoding the predicted representation matrix into the original attribute matrix. Through multiple iterations and adjustments, it ensures that the decoded matrix can accurately reflect the original attributes of the masked nodes. This process gradually reconstructs an attribute matrix that is as close to the original data as possible by restoring the features of the masked nodes layer by layer.

[0111] After the reconstructed original attribute matrix is ​​processed by the decoder, the explicit self-supervision signal and the reconstructed original data masked nodes are matched by minimizing the distance. Specifically, during the reconstruction process, the decoder predicts the original attributes of the masked nodes and optimizes the model by minimizing the distance between the predicted values ​​and the true values. This process ensures that the reconstructed original attributes can accurately reflect the characteristics of the masked nodes, thereby improving the overall performance and reliability of the model. In this way, the model can not only process graph data efficiently, but also significantly improve the quality of graph embedding and the performance of downstream tasks while maintaining data integrity. Whether in link prediction or node classification tasks, this method that integrates implicit and explicit self-supervision mechanisms can provide more accurate and reliable results, thereby showing higher practical value and application prospects in the application of graph neural networks.

[0112] Different decoders are used in link prediction and node classification to obtain different downstream tasks and reconstruct the masked attribute matrix after training.

[0113] S25. Construct a potential feature matching loss based on the representation matrix of the masked node and the predicted representation matrix of the masked node; construct an original attribute reconstruction loss based on the original attribute matrix of the reconstructed masked node and the original attribute matrix of the masked node.

[0114] In a feasible implementation, the latent feature matching loss function and the original attribute reconstruction loss function are calculated for the reconstructed image obtained by the ordinary variational encoder-decoder and the predicted image obtained by the momentum variational encoder, respectively, to obtain the latent feature matching loss function and the original attribute reconstruction loss function. Through the calculation of the loss function and the reconstruction module of the model, the mask operation can better serve the disease similarity feature link prediction and the disease correlation node classification itself, thereby making the device more effective.

[0115] Specifically, the latent feature matching loss is obtained by calculating the similarity between the reconstructed nodes (or edges) and the original information of the original graph. This step ensures that the model can capture and preserve the topological structure of the graph and the relationship between nodes. The latent feature matching loss recovers the masked information in the latent space by constraining the predicted representation to match the representation of the underlying structural information of the graph that retains more information. Specifically, this loss function forces the predicted representation to be as close as possible to the implicit features of the original graph by comparing it with the representation of the underlying structural information. This approach not only captures the high-order structural characteristics of the graph data, but also ensures the integrity and consistency of the information. By using these implicit self-supervisory signals as a guide for model learning, the completeness and accuracy of the predicted representation can be ensured, resulting in higher quality graph embeddings.

[0116] Next, the original attribute matrix of the masked node and the original attribute matrix of the reconstructed masked node are calculated, and the improved scaled cosine function is used to measure the difference between the two to obtain the original attribute reconstruction loss.

[0117] Specifically, the high-order sample correlation representation matrix is ​​retained and the latent features of the masked nodes are predicted based on the recovered information to calculate the latent feature matching loss. The specific steps are as follows:

[0118] S251, maintains the similarity between the original information of the reconstructed nodes and the original graph, focusing on ensuring that the predicted representation accurately matches the underlying structural statistics calculated by the momentum graph encoder.

[0119] S251.1, and Respectively expressed in and The representation vector of the jth masked node in the range . The masked information in the latent space is recovered by constraining the predicted representation to match a representation that preserves the underlying structural information of the graph, which is more informative.

[0120] S252, approximated by minimizing the following formula arrive :

[0121] (4)

[0122] In the formula, represents the potential feature matching loss, represents the set of masked nodes, represents a masked node, Indicated in The range of The representation vector of masked nodes, represents the predicted representation matrix of the mask node, Indicated in The range of The representation vector of masked nodes, Representation matrix representing the masked node.

[0123] Furthermore, the graph attention layer that recovers the masked content from the observation is input into the multi-decoder for reconstruction, and downstream tasks such as link prediction and node classification are processed; specifically:

[0124] S253.1. The formulas for the node classification decoder and link prediction decoder involve the forward propagation process of the graph neural network (GNN) to calculate the new feature vector for each node. , where k is the number of propagation steps.

[0125] The formula of S253.1.1 decoder also involves a global pooling operation. After all the propagation steps are completed, the final feature vector of each propagation step can be obtained. The feature vectors of all nodes are aggregated into a global graph feature vector. The global pooling operation is assumed to have a graph G, such as Figure 4 As shown, the node set is V, the edge set is E, for each node V, is an initial eigenvector.

[0126] S253.1.2, in the first stage of pooling, it is obtained :

[0127] (5)

[0128] in, is the set of neighboring nodes of node V, and are the weights and offsets that need to be learned in the kth propagation, is a non-linear activation function.

[0129] S253.1.3, POOL(·) is a pooling function. is the global feature vector of graph G. Finally, this global feature vector can be converted into a graph label through a fully connected layer:

[0130] (6)

[0131] S253.1.4. The link prediction decoder has an additional edge scoring function, SCORE(·) is a scoring function such as dot product, cosine similarity, or a more complex and learnable scoring function. is the score of the edge between node u and node v:

[0132] (7)

[0133] Furthermore, the original attribute reconstruction loss is calculated by using the reconstructed original data masked nodes and the obtained explicit self-supervision signal through an improved scaled cosine loss function; specifically:

[0134] S254.1. In order to complete the original attributes of the mask node, the present invention uses a decoder ,Will Mapped into the original data space.

[0135] S254.2. Once the reconstructed original attribute matrix mask nodes are processed by the decoder, the present invention minimizes the explicit self-supervision signal and the reconstructed original data mask nodes. and The distance between.

[0136] S254.2.1. To this end, the present invention approximates by minimizing the following formula arrive :

[0137] (8)

[0138] In the formula, represents the original attribute reconstruction loss, represents the set of masked nodes, represents a masked node, Represents the reconstruction of the original attribute vector of the masked node, represents the original attribute vector of the mask node, Represents the scale factor.

[0139] S26. The potential feature matching loss and the original attribute reconstruction loss are combined by a weighted method to obtain the overall loss function.

[0140] In one feasible implementation, after the latent feature matching loss and the original attribute reconstruction loss are calculated separately, the two losses are combined by a weighted method to obtain the overall loss function:

[0141] (9)

[0142] in, is a balance coefficient. This overall loss function takes into account the graph structure information and node attribute information, so that the model can be better optimized and adjusted during the training process, improving the overall performance and accuracy.

[0143] S27. Adjust the parameters of the graph neural network-graph convolutional network-graph self-supervised learning network model through the overall loss function to obtain a trained graph neural network-graph convolutional network-graph self-supervised learning network.

[0144] In a feasible implementation, the local model parameters are adjusted using the latent feature matching loss function and the original attribute reconstruction loss function, and adjustments are made in the iterative process based on the loss function results to improve efficiency.

[0145] Furthermore, when performing downstream tasks such as link prediction and node classification, the original attribute matrix of the reconstructed masked nodes and the predicted masked content are used to classify the nodes in the population network, so as to identify the similarity structure of disease characteristics in the network and further predict the disease status of the sample.

[0146] When performing downstream tasks such as link prediction and node classification, the original attribute matrix of the reconstructed masked node and the predicted masked content are used to classify the nodes in the network, so as to identify abnormal behaviors and network attacks in the network and improve network security.

[0147] S3. Obtain link prediction results based on the masking reconstruction results.

[0148] In a feasible implementation, downstream tasks such as link prediction and node classification are performed according to the masked reconstruction results. The framework diagram of the population graph application device is as follows: Figure 5 shown.

[0149] The present invention generates a masking graph and its masking feature matrix based on non-discrete bandwidth masking, and encodes two complementary masking graphs respectively; the masked content is predicted from the visible graph, and the predicted representation is similar to the output of the momentum variational graph autoencoder. The masking graph is input into a common encoder and processed by multiple decoders to reconstruct the original attributes of the masked nodes. The method calculates the potential feature matching loss function and the original attribute reconstruction loss function, and uses these loss functions to adjust the model parameters, improving the problem that discrete edge masking and binary link reconstruction strategies are insufficient to learn the representation of topological information. According to the processing of different downstream tasks by multiple decoders, the purpose of increasing the certainty of inferred masked data and improving the reliability of the utilized self-supervisory signal is achieved.

[0150] In the embodiments of the present invention, it is of great significance in link prediction and node classification. It uses a momentum variational graph autoencoder method based on a bandwidth masking strategy, without the need for time-consuming manual annotation and classification, and can effectively control errors and improve the accuracy and efficiency of link prediction and node classification. This method can not only better reveal the potential relationships in the network structure, but also accurately classify node attributes and improve the effect of overall network analysis. Therefore, in various application scenarios, such as social network analysis, recommendation systems, knowledge graph construction, etc., it significantly reduces the misclassification rate and missed detection rate, reduces the consumption of human resources, and improves work efficiency and accuracy.

[0151] Figure 6 1 is a block diagram of a relationship prediction device based on a non-discrete masked variational graph autoencoder according to an exemplary embodiment, wherein the device is used in a relationship prediction method based on a non-discrete masked variational graph autoencoder. Figure 6 The device includes an acquisition module 310, a masking reconstruction module 320 and an output module 330. Among them:

[0152] The acquisition module 310 is used to acquire the population map to be predicted.

[0153] The masked reconstruction module 320 is used to obtain a masked reconstruction result according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model;

[0154] Among them, according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model, the masked reconstruction results are obtained, including:

[0155] The population map is subjected to non-discrete bandwidth masking, and the bandwidth size is calculated. The population masked map and the population visible map are obtained according to the bandwidth masking. The masked map reorganization representation matrix is ​​obtained according to the population masked map. The masked map reorganization representation matrix is ​​input into the application potential feature predictor to obtain the predicted representation matrix of the masked nodes. The predicted representation matrix of the masked nodes is input into the multiple decoders to obtain the original attribute matrix of the reconstructed masked nodes. Link prediction decoding or node classification decoding is selected according to the task, and the original attribute matrix of the reconstructed masked nodes is subjected to masked reconstruction operation to obtain the masked reconstruction result.

[0156] The output module 330 is used to obtain a link prediction result according to the masking reconstruction result.

[0157] Optionally, the masked reconstruction module 320 is further configured to:

[0158] S21. Construct a graph neural network-graph convolutional network-graph self-supervised learning network model.

[0159] S22. Obtain a sample population map, and apply non-discrete bandwidth masking to the sample population map to obtain a masked map and a visible map.

[0160] S23. Encode the masked graph through a variational graph autoencoder to obtain a reorganized representation matrix of the masked graph; encode the visible graph through a momentum variational encoder to obtain a representation matrix of the masked nodes.

[0161] S24, inputting the masked graph reconstructed representation matrix into the applied latent feature predictor to obtain the predicted representation matrix of the masked node, inputting the predicted representation matrix of the masked node into the multiple decoders to obtain the original attribute matrix of the reconstructed masked node.

[0162] S25. Construct a potential feature matching loss based on the representation matrix of the masked node and the predicted representation matrix of the masked node; construct an original attribute reconstruction loss based on the original attribute matrix of the reconstructed masked node and the original attribute matrix of the masked node.

[0163] S26. The potential feature matching loss and the original attribute reconstruction loss are combined by a weighted method to obtain the overall loss function.

[0164] S27. Adjust the parameters of the graph neural network-graph convolutional network-graph self-supervised learning network model through the overall loss function to obtain a trained graph neural network-graph convolutional network-graph self-supervised learning network.

[0165] Optionally, the masked reconstruction module 320 is further configured to:

[0166] A variational encoder-decoder architecture model is constructed based on restricting the messages propagating through each edge to varying degrees.

[0167] The variational encoding-decoding architecture model is configured to obtain the graph neural network-graph convolutional network-graph self-supervised learning network model.

[0168] Optionally, the masked reconstruction module 320 is further configured to:

[0169] A variational graph autoencoder backbone is constructed, and a multi-layer perception layer is used as a projector to form the variational graph autoencoder; wherein the variational graph autoencoder backbone includes multiple graph attention layers or multiple graph isomorphism layers.

[0170] Optionally, the variational graph autoencoder is defined by an inference model, and the inference model is shown in the following equations (1) and (2):

[0171] (1)

[0172] (2)

[0173] In the formula, Represents the inference model formula of the variational graph autoencoder, represents the potential embedding representation matrix, represents the node feature matrix, represents the adjacency matrix, represents the standard Gaussian distribution, represents the posterior probability distribution of the vector representation, represents the mean of the node vector, Represents the variance of the node vector.

[0174] Optionally, the latent feature matching loss is as follows (3):

[0175] (3)

[0176] In the formula, represents the potential feature matching loss, represents the set of masked nodes, represents a masked node, Indicated in The range of The representation vector of masked nodes, represents the predicted representation matrix of the mask node, Indicated in The range of The representation vector of masked nodes, Representation matrix representing the masked node.

[0177] Optionally, the original attribute reconstruction loss is as shown in equation (4):

[0178] (4)

[0179] In the formula, represents the original attribute reconstruction loss, represents the set of masked nodes, represents a masked node, Represents the reconstruction of the original attribute vector of the masked node, represents the original attribute vector of the mask node, Represents the scale factor.

[0180] In the embodiments of the present invention, it is of great significance in link prediction and node classification. It uses a momentum variational graph autoencoder method based on a bandwidth masking strategy, without the need for time-consuming manual annotation and classification, and can effectively control errors and improve the accuracy and efficiency of link prediction and node classification. This method can not only better reveal the potential relationships in the network structure, but also accurately classify node attributes and improve the effect of overall network analysis. Therefore, in various application scenarios, such as social network analysis, recommendation systems, knowledge graph construction, etc., it significantly reduces the misclassification rate and missed detection rate, reduces the consumption of human resources, and improves work efficiency and accuracy.

[0181] Figure 7 is a schematic diagram of the structure of a relationship prediction device provided by an embodiment of the present invention, such as Figure 7 As shown, the relationship prediction device may include the above Figure 6 The relationship prediction device based on non-discrete masked variational graph autoencoder is shown. Optionally, the relationship prediction device 410 may include a first processor 2001.

[0182] Optionally, the relationship prediction device 410 may further include a memory 2002 and a transceiver 2003 .

[0183] The first processor 2001, the memory 2002 and the transceiver 2003 may be connected via a communication bus.

[0184] Combine the following Figure 7 The components of the relationship prediction device 410 are specifically introduced as follows:

[0185] The first processor 2001 is the control center of the relationship prediction device 410, and may be a processor or a general term for multiple processing elements. For example, the first processor 2001 is one or more central processing units (CPUs), or may be application specific integrated circuits (ASICs), or may be one or more integrated circuits configured to implement the embodiments of the present invention, such as one or more microprocessors (digital signal processors, DSPs), or one or more field programmable gate arrays (field programmable gate arrays, FPGAs).

[0186] Optionally, the first processor 2001 may perform various functions of the relationship prediction device 410 by running or executing a software program stored in the memory 2002 , and calling data stored in the memory 2002 .

[0187] In a specific implementation, as an embodiment, the first processor 2001 may include one or more CPUs, such as Figure 7 CPU0 and CPU1 are shown in FIG.

[0188] In a specific implementation, as an embodiment, the relationship prediction device 410 may also include multiple processors, such as Figure 7 The first processor 2001 and the second processor 2004 are shown in FIG. Each of these processors may be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). The processor here may refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0189] The memory 2002 is used to store the software program for executing the solution of the present invention, and is controlled to be executed by the first processor 2001. The specific implementation method can refer to the above method embodiment, which will not be repeated here.

[0190] Optionally, the memory 2002 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 2002 may be integrated with the first processor 2001, or may exist independently, and may be accessed through the interface circuit ( Figure 7 (not shown) is coupled to the first processor 2001, which is not specifically limited in this embodiment of the present invention.

[0191] The transceiver 2003 is used to communicate with a network device or a terminal device.

[0192] Optionally, the transceiver 2003 may include a receiver and a transmitter ( Figure 7 The receiver is used to implement a receiving function, and the transmitter is used to implement a sending function.

[0193] Optionally, the transceiver 2003 may be integrated with the first processor 2001, or may exist independently and communicate with the first processor 2001 through the interface circuit ( Figure 7 (not shown) is coupled to the first processor 2001, which is not specifically limited in this embodiment of the present invention.

[0194] It should be noted that Figure 7 The structure of the relationship prediction device 410 shown in the figure does not constitute a limitation on the router, and the actual knowledge structure recognition device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0195] In addition, the technical effects of the relationship prediction device 410 can refer to the technical effects of the relationship prediction method based on the non-discrete masked variational graph autoencoder described in the above method embodiment, and will not be repeated here.

[0196] It should be understood that the first processor 2001 in the embodiment of the present invention may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0197] It should also be understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0198] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0199] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0200] In the present invention, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0201] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0202] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0203] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0204] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0205] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0206] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0207] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0208] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A relationship prediction method based on non-discrete masked variational graph autoencoder, characterized in that: The method comprises: S1. Obtain the population map to be predicted; S2. Obtain a masked reconstruction result according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model; Wherein, obtaining the masked reconstruction result according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model includes: The population map is subjected to non-discrete bandwidth masking, the bandwidth size is calculated, a population mask map and a population visible map are obtained according to the bandwidth masking, a mask map reorganization representation matrix is ​​obtained according to the population mask map, the mask map reorganization representation matrix is ​​input into an application potential feature predictor to obtain a prediction representation matrix of a masked node, the prediction representation matrix of the masked node is input into a multiple decoder to obtain an original attribute matrix of a reconstructed masked node, link prediction decoding or node classification decoding is selected according to the task, a masked reconstruction operation is performed on the original attribute matrix of the reconstructed masked node, and a masked reconstruction result is obtained; S3. Obtain a link prediction result according to the masked reconstruction result.

2. The relationship prediction method based on non-discrete masked variational graph autoencoder according to claim 1, characterized in that: The training process of the graph neural network-graph convolutional network-graph self-supervised learning model in S2 includes: S21. Construct graph neural network-graph convolutional network-graph self-supervised learning network model; S22, obtaining a sample population map, and applying non-discrete bandwidth masking to the sample population map to obtain a masked map and a visible map; S23, encoding the masked graph through a variational graph autoencoder to obtain a masked graph reorganization representation matrix; encoding the visible graph through a momentum variational encoder to obtain a representation matrix of masked nodes; S24, inputting the masking graph reconstructed representation matrix into an application potential feature predictor to obtain a prediction representation matrix of the masked node, and inputting the prediction representation matrix of the masked node into a multiple decoder to obtain an original attribute matrix of the reconstructed masked node; S25, constructing a potential feature matching loss according to the representation matrix of the masked node and the predicted representation matrix of the masked node; constructing an original attribute reconstruction loss according to the original attribute matrix of the reconstructed masked node and the original attribute matrix of the masked node; S26, combining the potential feature matching loss and the original attribute reconstruction loss by a weighted method to obtain an overall loss function; S27. Adjust the parameters of the graph neural network-graph convolutional network-graph self-supervised learning network model through the overall loss function to obtain a trained graph neural network-graph convolutional network-graph self-supervised learning network.

3. The relationship prediction method based on non-discrete masked variational graph autoencoder according to claim 2, characterized in that: The construction of the graph neural network-graph convolutional network-graph self-supervised learning network model in S21 includes: Model a variational encoder-decoder architecture by restricting the messages propagating through each edge to varying degrees. The variational encoding-decoding architecture model is configured to obtain a graph neural network-graph convolutional network-graph self-supervised learning network model.

4. The relationship prediction method based on non-discrete masked variational graph autoencoder according to claim 2, characterized in that: The construction process of the variational graph autoencoder in S23 includes: A variational graph autoencoder backbone is constructed, and a multi-layer perception layer is used as a projector to form the variational graph autoencoder; wherein the variational graph autoencoder backbone includes multiple graph attention layers or multiple graph isomorphism layers.

5. The relationship prediction method based on non-discrete masked variational graph autoencoder according to claim 2, characterized in that: The variational graph autoencoder in S23 is defined by an inference model, and the inference model is shown in the following equations (1) and (2): （1） （2） In the formula, Represents the inference model formula of the variational graph autoencoder, represents the potential embedding representation matrix, represents the node feature matrix, represents the adjacency matrix, represents the standard Gaussian distribution, represents the posterior probability distribution of the vector representation, represents the mean of the node vector, Represents the variance of the node vector.

6. The relationship prediction method based on non-discrete masked variational graph autoencoder according to claim 2, characterized in that: The potential feature matching loss in S25 is shown in the following formula (3): （3） In the formula, represents the potential feature matching loss, represents the set of masked nodes, represents a masked node, Indicated in The range of The representation vector of masked nodes, represents the predicted representation matrix of the mask node, Indicated in The range of The representation vector of masked nodes, Representation matrix representing the masked node.

7. The relationship prediction method based on non-discrete masked variational graph autoencoder according to claim 2, characterized in that: The original attribute reconstruction loss in S25 is shown in the following formula (4): （4） In the formula, represents the original attribute reconstruction loss, represents the set of masked nodes, represents a masked node, Represents the reconstruction of the original attribute vector of the masked node, represents the original attribute vector of the mask node, Represents the scale factor.

8. A relationship prediction device based on a non-discrete masked variational graph autoencoder, wherein the relationship prediction device based on a non-discrete masked variational graph autoencoder is used to implement the relationship prediction method based on a non-discrete masked variational graph autoencoder as claimed in any one of claims 1 to 7, characterized in that: The device comprises: An acquisition module, used for acquiring a population map to be predicted; A masked reconstruction module, used to obtain a masked reconstruction result according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model; Wherein, obtaining the masked reconstruction result according to the population map and the trained graph neural network-graph convolutional network-graph self-supervised learning model includes: The population map is subjected to non-discrete bandwidth masking, the bandwidth size is calculated, a population mask map and a population visible map are obtained according to the bandwidth masking, a mask map reorganization representation matrix is ​​obtained according to the population mask map, the mask map reorganization representation matrix is ​​input into an application potential feature predictor to obtain a prediction representation matrix of a masked node, the prediction representation matrix of the masked node is input into a multiple decoder to obtain an original attribute matrix of a reconstructed masked node, link prediction decoding or node classification decoding is selected according to the task, a masked reconstruction operation is performed on the original attribute matrix of the reconstructed masked node, and a masked reconstruction result is obtained; The output module is used to obtain a link prediction result according to the masking reconstruction result.

9. A relationship prediction device, characterized in that: The relationship prediction device comprises: processor; A memory having computer-readable instructions stored thereon, wherein when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes, which can be called by a processor to execute the method according to any one of claims 1 to 7.

Images