CN118245638B - Method, device, equipment and storage medium for predicting graph data based on generalization model - Google Patents

Abstract

The embodiments of the present application provide a method, device, equipment and storage medium for predicting graph data based on a generalization model, which belongs to the field of graph data processing technology. The method includes: obtaining target graph data, wherein the target graph data includes multiple target nodes and corresponding node connection structures; based on the node connection structure composed of multiple target nodes in the target graph data, finding the target subgraph of the target connection structure of each target node in the target association level; inputting the target subgraph into the generalization model to obtain the node category label of the target node in the target subgraph; wherein the generalization model is based on the training process of minimizing the target loss to perform causal prediction learning between the sample node connection structure and the sample node category label in the sample subgraph, and the target loss is composed of the first loss, the second loss and the third loss. The present application can improve the accuracy of graph data prediction.

Description

Graph data prediction method, device, equipment and storage medium based on generalization model

Technical Field

The present application relates to the technical field of graph data processing, and in particular to a graph data prediction method, device, equipment and storage medium based on a generalized model.

Background Art

Graph Neural Networks (GNNs) are deep learning models for processing graph-structured data. In a graph neural network, the features of nodes are updated and aggregated by considering the relationship between nodes (i.e., the existence of edges), so as to gradually build the relationship between nodes and learn the structural representation of the graph.

In the related art, during the training of graph neural networks, key features are usually extracted from the input graph data for prediction of the graph data under the premise of maximizing the mutual information between the graph data and the true label. However, this training method makes the graph neural network learn all the statistical correlations between the input features and labels in the graph data as much as possible, without distinguishing between the causal effect and the non-causal effect between the input features and the labels, which makes the graph neural network tend to access non-causal features as a shortcut to predict graph data, reducing the accuracy of graph data prediction after the graph neural network is trained.

Summary of the invention

The main purpose of the embodiments of the present application is to propose a graph data prediction method, device, equipment and storage medium based on a generalized model, which can improve the accuracy of graph data prediction.

To achieve the above-mentioned purpose, a first aspect of an embodiment of the present application proposes a graph data prediction method based on a generalization model, the method comprising:

Acquire target graph data, wherein the target graph data includes a plurality of target nodes and a node connection structure composed of the plurality of target nodes;

Based on the node connection structure formed between the plurality of target nodes in the target graph data, finding a target subgraph of the target connection structure of each target node in the target association level;

Inputting the target subgraph into a generalization model to obtain a node category label of a target node in the target subgraph;

The generalization model is obtained by performing causal prediction learning between the sample node connection structure and the sample node category label in the sample subgraph based on a training process of minimizing the target loss, and the target loss is composed of a first loss, a second loss and a third loss. The first loss is determined according to the first prediction probability of the core feature of the sample subgraph under the corresponding sample node category label; the second loss is determined according to the second prediction probability of the redundant feature of the sample subgraph under the corresponding sample node category label; the third loss is determined according to the sample distance between the core feature and the redundant feature, the core feature is a feature obtained by enhancing the sample node connection structure of the sample subgraph according to the edge mask matrix obtained by the attention mechanism; the redundant feature is a feature obtained by weakening the sample node connection structure of the sample subgraph according to the complement matrix of the edge mask matrix.

Accordingly, a second aspect of an embodiment of the present application proposes a graph data prediction device based on a generalization model, the device comprising:

An acquisition module, used to acquire target graph data, wherein the target graph data includes a plurality of target nodes and a node connection structure composed of the plurality of target nodes;

A search module, configured to search for a target subgraph of a target connection structure of each target node within a target association level based on a node connection structure formed between a plurality of the target nodes in the target graph data;

An input module is used to input the target subgraph into a generalization model to obtain a node category label of a target node in the target subgraph; wherein the generalization model is obtained by performing causal prediction learning between a sample node connection structure and a sample node category label in a sample subgraph based on a training process of minimizing a target loss, wherein the target loss is composed of a first loss, a second loss and a third loss, wherein the first loss is determined according to a first prediction probability of a core feature of the sample subgraph under the corresponding sample node category label; the second loss is determined according to a second prediction probability of a redundant feature of the sample subgraph under the corresponding sample node category label; the third loss is determined according to a sample distance between the core feature and the redundant feature, wherein the core feature is a feature obtained by enhancing the representation of the sample node connection structure of the sample subgraph according to an edge mask matrix obtained by an attention mechanism; and the redundant feature is a feature obtained by weakening the representation of the sample node connection structure of the sample subgraph according to a complement matrix of the edge mask matrix.

In some implementations, the sample nodes of the sample subgraph include a central sample node and a plurality of adjacent sample nodes; and the graph data prediction apparatus based on a generalization model further includes a training module for:

Acquire sample graph data for training a preset model, and generate a corresponding sample subgraph for each sample node in the sample graph data according to the sample node connection structure of the sample graph data;

Obtaining an initial feature matrix and an adjacency matrix of each sample subgraph, and sequentially inputting the initial feature matrix into a preset model for feature aggregation to obtain a fused feature matrix of each sample subgraph;

Determining the core features and redundant features of each of the sample subgraphs according to the adjacency matrix and the fusion feature matrix;

Determining a target loss of the preset model based on the core features and the redundant features;

The parameters of the preset model are updated based on the target loss, and when the preset model converges, a trained generalization model is obtained.

In some embodiments, the training module is further used to:

Based on the fusion feature matrix of the sample subgraph, any two sample nodes in the sample subgraph are input into a multi-layer perceptron to obtain an edge mask value between any two sample nodes;

Generating an edge mask matrix of the sample sub-graph according to a plurality of edge mask values in the sample sub-graph;

Generate core features corresponding to the sample subgraph based on the adjacency matrix, the fused feature matrix and the edge mask matrix;

Generate a corresponding complement matrix according to the edge mask matrix of the sample subgraph;

Based on the adjacency matrix, the fusion feature matrix and the complement matrix, redundant features of the sample subgraph are obtained.

In some embodiments, the training module is further used to:

Acquire basic graph data and a plurality of sub-patterns of different categories, and assign the plurality of sub-patterns to any sample node of the basic graph data to obtain sample graph data;

A preset number of edges to be added to the sample graph data is determined based on the number of nodes in the sample graph data, and edges are added to any two sample nodes in the sample graph data, wherein the number of edges added to the sample graph data is equal to the preset number.

In some embodiments, the training module is further used to:

Acquire the node density of the sample graph data, and determine the sample association level of each of the sample nodes based on the node density;

For each of the sample nodes as a central sample node, based on the sample node connection structure formed between the central sample node and adjacent sample nodes, a sample subgraph of the target sample node connection structure of each of the central sample nodes in the sample association level is found.

In some implementations, the graph data prediction apparatus based on the generalization model further includes a generation module, which is used to:

Get the preset graph theory library;

Obtaining preset characteristic indicators of the sample nodes, and generating node characteristic vectors of each sample node under the corresponding characteristic indicators through the graph theory library; wherein the characteristic indicators include at least one of node identification, node degree, clustering coefficient, betweenness centrality and closeness centrality;

A first feature matrix is generated according to a plurality of node feature vectors corresponding to a plurality of sample nodes.

In some embodiments, the training module is further used to:

In each of the sample subgraphs, determining a node identifier of each sample node;

Based on the node identifier, determining a node feature vector corresponding to each of the sample nodes from the first feature matrix;

Generating an initial feature matrix of the sample subgraph according to a plurality of node feature vectors corresponding to a plurality of sample nodes of the sample subgraph;

Based on the connection relationship between the sample nodes, an adjacency matrix corresponding to the feature matrix is generated.

In some embodiments, the training module is further used to:

Determining a node degree matrix of the sample subgraph from the first feature matrix according to the node identifier of each sample subgraph;

Obtaining an adjacency matrix of the sample subgraph, and normalizing the adjacency matrix using the node degree matrix to obtain a normalized adjacency matrix;

Taking the central sample node as the fusion center, based on the normalized adjacency matrix, performing layer-by-layer fusion of sample node features to obtain an updated central node feature vector corresponding to the central sample node;

The node feature vector of the central sample node of the initial feature matrix is updated according to the central node feature vector to obtain a fusion feature matrix of each of the sample subgraphs.

In some embodiments, the training module is further used to:

Determine a first loss based on a difference between a predicted node category label output for a core feature of the sample subgraph and a sample node category label of the core feature;

Determine the second loss based on the uniformity of probability distribution when predicting the category label of each sample node according to the redundant features of the sample subgraph based on the preset model;

Determining a third loss according to a sample distance between the core feature and the redundant feature;

The target loss of the preset model is constructed based on the sum of the first loss, the second loss and the third loss.

Correspondingly, the third aspect of the embodiments of the present application proposes a computer device, which includes a memory and a processor, the memory stores a computer program, and the processor implements the graph data prediction method based on the generalization model as described in any one of the embodiments of the first aspect of the present application when executing the computer program.

Correspondingly, the fourth aspect of the embodiments of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the graph data prediction method based on the generalization model as described in any one of the embodiments of the first aspect of the present application.

In an embodiment of the present application, target graph data is obtained, wherein the target graph data includes a plurality of target nodes and a node connection structure composed of a plurality of target nodes; based on the node connection structure composed of a plurality of target nodes in the target graph data, a target subgraph of the target connection structure of each target node in the target association level is found; the target subgraph is input into a generalization model to obtain a node category label of the target node in the target subgraph; wherein the generalization model is obtained by performing causal prediction learning between the sample node connection structure and the sample node category label in the sample subgraph based on a training process of minimizing the target loss, and the target loss is composed of a first loss, a second loss and a third loss, and the first loss is determined based on the difference between the predicted node category label output for the core feature of the sample subgraph and the sample node category label of the core feature; the second loss is determined based on the uniformity of the probability distribution when predicting each sample node category label for the redundant feature of the sample subgraph according to the preset model; the third loss is determined according to the sample distance between the core feature and the redundant feature, and the core feature is a feature obtained by enhancing the sample node connection structure of the sample subgraph according to the edge mask matrix obtained by the attention mechanism; the redundant feature is a feature obtained by weakening the sample node connection structure of the sample subgraph according to the complement matrix of the edge mask matrix. In this way, the preset model can be trained by constructing a target loss, and during the training process, the preset model tends to predict through core features, reduce the preference for redundant features, and reduce the sample distance between core features and redundant features, thereby enabling the trained generalized model to have stronger causal explanation ability and generalization ability. In addition, the edge mask matrix obtained by the attention mechanism enhances the sample connection structure representing the sample subgraph, and the complement matrix of the edge mask matrix weakens the sample node connection structure representing the sample subgraph, which helps the trained generalized model to better capture the important relationship between target nodes, reduce the interference of redundant information, and make the preset model more focused on the learning of core features. In the process of applying the generalized model, the target subgraph of the target node to be predicted can be obtained based on the target graph data, so that the generalized model does not need to predict the entire target graph data, but only needs to obtain the relevant target subgraph of the target node to be predicted, which helps the generalized model to analyze the causal relationship between target nodes more attentively, and improves the efficiency and accuracy of prediction. In summary, the present application can improve the accuracy of the trained generalized model in predicting graph data.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the architecture of a graph data prediction system based on a generalized model provided in an embodiment of the present application;

FIG2 is a flow chart of a graph data prediction method based on a generalization model provided in an embodiment of the present application;

FIG3 is a flowchart of the steps of training a preset model provided in an embodiment of the present application;

FIG4 is an overall flow chart of a graph data prediction method based on a generalized model provided in an embodiment of the present application;

FIG5 is a schematic diagram of functional modules of a graph data prediction device based on a generalized model provided in an embodiment of the present application;

FIG. 6 is a schematic diagram of the hardware structure of a computer device provided in an embodiment of the present application.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

Graph Neural Networks (GNNs) are deep learning models for processing graph-structured data. In a graph neural network, the features of nodes are updated and aggregated by considering the relationship between nodes (i.e., the existence of edges), so as to gradually build the relationship between nodes, learn the structural representation of the graph, and realize the prediction of the target node.

In the related art, during the training of graph neural networks, key features are usually extracted from the input graph data for prediction of the graph data under the premise of maximizing the mutual information between the graph data and the true label. However, this training method makes the graph neural network learn all the statistical correlations between the input features and labels in the graph data as much as possible, without distinguishing between the causal effect and the non-causal effect between the input features and the labels, which makes the graph neural network tend to access non-causal features as a shortcut to predict graph data, reducing the accuracy of graph data prediction after the graph neural network is trained.

Based on this, the embodiments of the present application provide a graph data prediction method, apparatus, device and storage medium based on a generalized model, which can improve the accuracy of graph data prediction.

The graph data prediction method, device, equipment and storage medium based on the generalized model provided in the embodiments of the present application are specifically illustrated through the following embodiments. First, the graph data prediction system based on the generalized model in the embodiments of the present application is described.

In some implementations, an embodiment of the present application provides a graph data prediction system based on a generalized model. Please refer to FIG. 1 . The graph data prediction system based on a generalized model includes a terminal 11 and a server 12 .

Specifically, the terminal 11 can be a mobile terminal device or a non-mobile terminal device. The mobile terminal device can be a mobile phone, a tablet computer, a laptop computer, a PDA, a wearable device, a super mobile personal computer, a netbook, a personal digital assistant, a wireless hotspot device, etc.; the non-mobile terminal device can be a personal computer, etc., which is not specifically limited in the implementation scheme of this application. The technician can directly interact with the server 12 through the terminal 11, send instructions for training a preset model to the server 12, set the number of preset model trainings, etc., and display the results after the generalized model is obtained through training.

Furthermore, the server 12 may include a cloud computing server, a data center server, etc., which is responsible for storing relevant data of the generalized model, and training the preset model to obtain the generalized model after receiving the model training instruction from the terminal 11. The server 12 has high computing, storage and network performance, and can support requests from multiple terminals 11.

The graph data prediction method based on the generalization model in the embodiments of the present application can be illustrated by the following embodiments.

It should be noted that in each specific implementation of the present application, when it comes to the need to perform relevant processing based on data related to user identity or characteristics such as user information, user behavior data, user historical data, and user location information, the user's permission or consent will be obtained first. Moreover, the collection, use, and processing of these data will comply with relevant laws, regulations, and standards. In addition, when the embodiment of the present application needs to obtain the user's sensitive personal information, the user's separate permission or separate consent will be obtained through a pop-up window or by jumping to a confirmation page. After clearly obtaining the user's separate permission or separate consent, the necessary user-related data for enabling the normal operation of the embodiment of the present application will be obtained.

In the embodiment of the present application, the description will be made from the dimension of a graph data prediction device based on a generalized model, which can be specifically integrated in a computer device. Referring to FIG. 2, FIG. 2 is a flowchart of the steps of a graph data prediction method based on a generalized model provided in the embodiment of the present application. The embodiment of the present application takes the graph data prediction device based on a generalized model as an example, in which the graph data prediction device based on a generalized model is specifically integrated in a terminal or a server. When the processor on the terminal or server executes the program instructions corresponding to the graph data prediction method based on a generalized model, the specific process is as follows:

Step 101, obtaining target graph data, wherein the target graph data includes a plurality of target nodes and a node connection structure composed of the plurality of target nodes.

It should be noted that in order to effectively perform classification prediction on the target graph data, the target graph data can be obtained so that the subsequent generalization model can understand the node connection structure composed of the target nodes of the target graph data, thereby performing classification more accurately.

Among them, the target graph data can be graph structure data composed of multiple target nodes and corresponding edges, such as social networks, user-item relationship graphs in recommendation systems, protein interaction networks in bioinformatics, etc.

Among them, the target node can be an entity in the target graph data, such as a user in a social network or an item in a recommendation system, etc. Each target node has its specific attributes or category labels.

The node connection structure can be the relationship or connection between target nodes. For example, in a social network, when two users are friends, there is an edge between the two users. The two users and the edge constitute a node connection structure. When a user has purchased an item, an edge can be used to connect the user node and the item node to indicate the user's purchase behavior of the item. The user, edge, and item can constitute a node connection structure, and so on.

Exemplarily, the target graph data to be identified can be obtained first. The target graph data can be preset or obtained in real time. For example, the data source of the target graph data to be obtained can be determined. For example, when the target graph data needs to be obtained from a social network, the web crawler technology can be used to collect user data and the relationship between users on the social media platform; when the target graph data related to bioinformatics needs to be obtained, the network data of protein interactions can be obtained from a public bioinformatics data platform, etc. Furthermore, after the raw data is collected, the raw data can be preprocessed, including data cleaning, data formatting, etc.

Obtaining the target graph data in the above manner can facilitate subsequent classification of the target graph data through a generalized model.

Step 102, based on the node connection structure formed by multiple target nodes in the target graph data, find out the target subgraph of the target connection structure of each target node in the target association level.

It is understandable that when classifying the target node to be predicted, if the entire target graph data is input into the generalization model, it will consume a lot of computing resources. By finding the target subgraph of the target connection structure of each target node within the target association hierarchy, the information related to the classification task of the target node to be predicted can be captured more accurately, thereby improving the accuracy of classification and saving computing resources.

Among them, the target association level can be a level determined by K-hop sampling through breadth-first search in the target graph data, and K-hop sampling can help the generalization model determine the local area information of the corresponding target node in the target graph data. For example, when it is necessary to predict the target node to be predicted, the number of steps from the target node to the center can be determined by the density of the target graph data, and the determined number of steps can be used as the target association level. Furthermore, when the target nodes in the target graph data are sparsely distributed, a larger target association level can be determined, and when the target nodes in the target graph data are densely distributed, a smaller target association level can be determined. Exemplarily, the target association level can be level 2, level 3, etc., and the embodiments of the present application do not impose specific restrictions on this.

The target connection structure is a node connection structure consisting of a target node to be predicted and other connected target nodes within a target association level.

The target subgraph may be a local graph structure obtained by a breadth-first search algorithm starting from the target node to be predicted in the target graph data. Each target subgraph includes the target node to be predicted as a central node and all other target nodes connected to the central node in the target association hierarchy.

Specifically, the target node to be predicted can be taken as the central starting point, and other target nodes associated with the target node to be predicted can be found in the target association hierarchy through breadth-first search, and the target subgraph corresponding to the target node to be predicted can be generated through the node connection structure between the target node to be predicted and other target nodes.

Exemplarily, when the target graph data is a social network graph, a new user node needs to be classified. Assuming that the target node to be predicted is user node A, starting from node A, the breadth-first search algorithm is used to determine other user nodes directly connected to node A in the target graph data. If the target association level is determined to be 3 according to the node density of the target graph data, then the node connection structure of the social network graph can be obtained, and other user nodes directly connected to node A can be found within the three target association levels starting from user node A. If user node A is directly connected to user nodes B, C, and D, then the node connection structure composed of user node A, user node B, user node C, and user node D can be used as the target connection structure, and the target connection structure can be used as the target subgraph.

In the above way, by finding the target subgraph corresponding to the target node to be predicted, the information related to the classification task of the target node can be captured more accurately, which is also beneficial to reduce the amount of calculation when the subsequent generalization model predicts the target subgraph.

Step 103, input the target subgraph into the generalization model to obtain the node category label of the target node in the target subgraph;

Among them, the generalization model is obtained by performing causal prediction learning between the sample node connection structure and the sample node category label in the sample subgraph based on the training process of minimizing the target loss. The target loss is composed of a first loss, a second loss and a third loss. The first loss is determined based on the difference between the predicted node category label output for the core feature of the sample subgraph and the sample node category label of the core feature; the second loss is determined based on the uniformity of the probability distribution when the preset model predicts the category label of each sample node for the redundant features of the sample subgraph; the third loss is determined according to the sample distance between the core feature and the redundant feature. The core feature is the feature obtained by enhancing the sample node connection structure of the sample subgraph according to the edge mask matrix obtained by the attention mechanism; the redundant feature is the feature obtained by weakening the sample node connection structure of the sample subgraph according to the complement matrix of the edge mask matrix.

In some embodiments, in order to improve the accuracy of prediction of target nodes in target graph data, a preset model can be pre-trained so that the preset model continuously learns the true causal relationship between sample nodes and true labels by focusing on causal features, while reducing attention to shortcut features (i.e., redundant features), so that the trained generalization model has higher generalization performance.

Among them, the generalization model can be a model with good generalization ability for predicting the node category label of the target node in the target subgraph, or it can be a model for predicting the category of the entire target graph data. The generalization model is pre-trained by a preset model, and when faced with new tasks and data, the generalization model can maintain good generalization performance. Exemplarily, the generalization model can be a graph neural network (Graph Neural Networks, GNNs), a graph attention network (Graph Attention Networks, GATs), a graph convolutional network (Graph Convolutional Networks, GCNs) and other neural networks suitable for processing target graph data and predicting target nodes.

The node category label may be a specific category to which each target node belongs. For example, in a social network, a target node may represent a user, and a node category label may represent the user's interest or group. In some embodiments, when the entire target graph data is to be predicted, the node category label represents the category of the entire target graph data. For example, in bioinformatics, the target graph data may represent a protein interaction network, and the node category label may represent the functional type of this network, and so on.

The sample subgraph can be a local sample graph structure obtained by a breadth-first search algorithm starting from the sample node to be predicted in the sample graph data. Each sample subgraph includes the sample node to be predicted as the sample center node and all other sample nodes connected to the sample center node in the sample association hierarchy.

The sample node connection structure may be the relationship or connection between sample nodes. For example, in a social network, when two users have a friend relationship, there is an edge between the two users. The two users and the friend relationship constitute a node connection structure. When a user has purchased an item, an edge may be used to connect the user node and the item node to indicate the user's purchase behavior of the item. The user, edge, and item may constitute a node connection structure, and so on.

Among them, the target loss can be a loss composed of a first loss, a second loss, and a third loss, which is used to indicate the adjustment of the parameters of the preset model to train the preset model to obtain a generalized model. Specifically, the first loss can be a cross entropy loss, the second loss can be an entropy loss, and the third loss can be a contrast loss. By forming the target loss with the first loss, the second loss, and the third loss, and using the target loss to update the model parameters, the generalized model can be made to learn the core features more attentively, avoid the generalized model from over-focusing on redundant features, and thus improve the quality of the overall feature representation.

The first loss may be a cross entropy loss, which is used to evaluate the difference between the probability distribution of each predicted node category label predicted by the preset model for the core feature and the sample node category label of the core feature. Specifically, an encoding may be set for each sample node category label, and each sample unique node category label may be represented by a binary vector, and only the real sample node category label corresponding to the core feature is set to 1, and the remaining sample node category labels are set to 0. Specifically, the first loss Can be:

in, is the one-hot encoding of the sample node category label, is the model's first predicted probability for the vth sample node category label. When the predicted node category label is consistent with the actual sample node category label corresponding to the core feature, the first loss is minimal. Therefore, by minimizing the first loss, the preset model can better learn the causal relationship of the data, thereby continuously making more accurate predictions.

The predicted node category labels are multiple node category labels predicted by the preset model for the core features. For example, for an animal image, the predicted node category labels may be cat, dog, and rabbit, and each predicted node category label corresponds to a predicted probability.

The second loss may be an entropy loss, which is used to measure the uniformity of the probability distribution when the preset model predicts the category labels of each sample node based on the redundant features of the sample subgraph. The more uniform the probability distribution, the smaller the classification preference of the preset model for redundant features. Thus, it can effectively prevent the preset model from learning shortcut features. Specifically, the second loss Can be:

in, Represents the uniform probability of each sample node category label when uniformly distributed, usually =1/number of categories, It is the second predicted probability of the preset model for the category label of the vth sample node.

Specifically, when redundant features are input into the preset model, the output second prediction probabilities are close to uniform distribution, indicating that redundant features are not helpful for the classification of the preset model. For example, if the sample category labels are 4, namely cat, rabbit, dog, and pig, then =1/4, if the redundant features are input into the preset model, and the predicted probabilities of cats, rabbits, dogs, and pigs are all 1/4, the second loss of the preset model is also the smallest. At this time, it shows that the preset model has no classification preference for redundant features, and the output probability distribution is highly uniform. Therefore, it is necessary to continuously train the preset model so that the predicted probability of the preset model for the redundant features in the category labels of each sample node is close to uniform distribution.

The uniformity of probability distribution may be the second preset probability of each sample node category label predicted and output by the preset model for the redundant features after the redundant features are input into the preset model. The more uniform the second preset probability distribution is, the higher the uniformity of probability distribution is. For example, the uniformity of probability distribution with the second preset probabilities of 1/3, 1/3, and 1/3 is greater than 2/3, 1/6, and 1/6.

Among them, the third loss can be a contrast loss. Specifically, the core features can be used as positive samples and the redundant features can be used as negative samples. The contrast loss can encourage the preset model to correctly distinguish between positive samples and negative samples, making the distance between positive samples closer and the distance between positive samples and negative samples farther. Furthermore, the third loss Can be:

in, is the sigmoid function, which is usually used to map the input to (0, 1), Q is a hyperparameter, C and are different instances of the core features, representing positive sample pairs, C and Both represent positive samples, i.e. core features; C and represents a positive and negative sample pair, Represents negative samples, i.e., redundant features; It represents the expected value, which can be obtained by averaging the first predicted probability of the core features of all positive sample pairs by the preset model and the number of second predicted probabilities of the redundant features of all negative sample pairs. By training the preset model with the third loss, the preset model can be encouraged to correctly distinguish positive samples from negative samples, so that the distance between positive samples is closer and the distance between positive samples is farther.

The sample distance may be the distance between two core features or the distance between a core feature and a redundant feature. The sample distance may be calculated by Euclidean distance or cosine similarity.

Among them, the core feature can be a feature with causal relationship or causal explanation ability in the sample subgraph, and the core feature has causal information of the sample node to be predicted and the sample node category label.

The edge mask matrix may be the edge mask values of node feature vectors of any two sample nodes in the sample subgraph after passing through a multi-layer perceptron, and may be a matrix composed of the edge mask values of all sample nodes in the sample subgraph.

Among them, redundant features can be shortcut features between core features and prediction results. Redundant features have no additional contribution to the prediction of sample nodes and sample node category labels, and do not have causal explanation characteristics. Moreover, redundant features will make the preset model tend to learn shortcut features to make decisions, thereby reducing the accuracy of the preset model's prediction of sample graph data, resulting in the performance of the trained generalized model in out-of-distribution test data, that is, the generalization performance is reduced. Exemplarily, redundant features can be noise features.

Among them, the complement matrix = 1-edge mask matrix, and the complement matrix can be used to represent redundant edges in the sample subgraph.

It can be understood that by inputting the target subgraph into the trained generalization model, the generalization model can quickly extract the core features in the target subgraph that are helpful for classifying the target node to be predicted, and predict the target node to be predicted based on the core features, thereby improving the efficiency and accuracy of the target node prediction.

For example, when the interest category of user A in the social network graph needs to be predicted, the target connection structure corresponding to user A can be determined as the target subgraph of user A through the node connection structure of the target graph data. For example, the target subgraph can be a target subgraph containing user A and his friends, and then the target subgraph is input into the generalization model to obtain the specific interest category of user A. In this process, the generalization model can extract useful information from the social network structure of user A and make accurate classification decisions based on it.

In an embodiment of the present application, target graph data is obtained, wherein the target graph data includes a plurality of target nodes and a node connection structure composed of a plurality of target nodes; based on the node connection structure composed of a plurality of target nodes in the target graph data, a target subgraph of the target connection structure of each target node in the target association level is found; the target subgraph is input into a generalization model to obtain a node category label of the target node in the target subgraph; wherein the generalization model is obtained by performing causal prediction learning between the sample node connection structure and the sample node category label in the sample subgraph based on a training process of minimizing the target loss, and the target loss is composed of a first loss, a second loss and a third loss, and the first loss is determined based on the difference between the predicted node category label output for the core feature of the sample subgraph and the sample node category label of the core feature; the second loss is determined based on the uniformity of the probability distribution when predicting each sample node category label for the redundant feature of the sample subgraph according to the preset model; the third loss is determined according to the sample distance between the core feature and the redundant feature, and the core feature is a feature obtained by enhancing the sample node connection structure of the sample subgraph according to the edge mask matrix obtained by the attention mechanism; the redundant feature is a feature obtained by weakening the sample node connection structure of the sample subgraph according to the complement matrix of the edge mask matrix. In this way, the preset model can be trained by constructing a target loss, and during the training process, the preset model tends to predict through core features, reduce the preference for redundant features, and reduce the sample distance between core features and redundant features, thereby enabling the trained generalized model to have stronger causal explanation ability and generalization ability. In addition, the edge mask matrix obtained by the attention mechanism enhances the sample connection structure representing the sample subgraph, and the complement matrix of the edge mask matrix weakens the sample node connection structure representing the sample subgraph, which helps the trained generalized model to better capture the important relationship between target nodes, reduce the interference of redundant information, and make the preset model more focused on the learning of core features. In the process of applying the generalized model, the target subgraph of the target node to be predicted can be obtained based on the target graph data, so that the generalized model does not need to predict the entire target graph data, but only needs to obtain the relevant target subgraph of the target node to be predicted, which helps the generalized model to analyze the causal relationship between target nodes more attentively, and improves the efficiency and accuracy of prediction. In summary, the present application can improve the accuracy of the trained generalized model in predicting graph data.

Please refer to FIG. 3 , in some embodiments, the sample nodes of the sample subgraph may include a central sample node and adjacent sample nodes adjacent to the central sample node. In order to improve the accuracy of the generalization model for predicting the target graph data, the preset model may be trained so that the generalization model can learn causal features (i.e., core features) and avoid learning shortcut features (redundant features). Specifically, the generalization model is trained through steps 201 to 205 to obtain:

Step 201: obtain sample graph data for training a preset model, and generate a corresponding sample subgraph for each sample node in the sample graph data according to the sample node connection structure of the sample graph data.

The sample graph data may be a synthetic graph generated in a random manner, and sample nodes of different categories may be obtained by randomly adding graphs of different shapes to each node of the base graph, and edges may be added to any two sample nodes, and a sample node connection structure composed of sample nodes and edges may be used as sample graph data. In some implementations, existing graph data may also be obtained as sample graph data.

Exemplarily, the sample subgraph can be a local sample graph structure obtained by a breadth-first search algorithm from the sample node to be predicted in the sample graph data, so that the preset model can have a better ability to process local features, and although each computing node processes the sample subgraph independently, it can train the preset model, and improve the training efficiency of the preset model by synchronously updating and sharing the parameters of the preset model, and optimize the performance of the preset model. Each sample subgraph includes the sample node to be predicted as the sample center node, and all other sample nodes connected to the sample center node in the sample association hierarchy.

For example, suppose there is a social network graph, in which each sample node represents a user, and each edge represents a relationship between two users (such as a friend relationship). When it is necessary to generate a sample subgraph of user A based on the social network graph, other users connected to user A can be determined through breadth-first search. Assuming that user A is directly connected to users B and C, the generated sample subgraph will include users A, B, and C, with user A as the central node of the sample subgraph. Furthermore, the scope of the breadth-first search can also be set to avoid the sample subgraph being too large, so that the preset model cannot focus on the core features.

Through the above method, a sample subgraph of each sample node can be generated to train the preset model, improve the classification ability of the preset model for the sample nodes to be predicted, and improve the accuracy of the prediction.

It is understandable that in order to improve the generalization ability of the generalization model, sample graph data can be randomly generated, and synthetic graphs with diversified structures can be generated according to the randomly generated sample graph data, so as to help the preset model to more comprehensively understand and predict different types of graph structures and increase the robustness of the trained generalization model. For example, "obtaining sample graph data for training the preset model" in step 201 includes:

(201.a1) Obtaining basic graph data and multiple sub-patterns of different categories, and assigning the multiple sub-patterns to any sample node of the basic graph data to obtain sample graph data;

(201.a2) Determine a preset number of edges to be added to the sample graph data according to the number of nodes in the sample graph data, and add edges to any two sample nodes in the sample graph data, wherein the number of edges added to the sample graph data is equal to the preset number.

Among them, the basic graph data can be a scale-free network model graph, which is a mathematical model used to simulate the real-world network structure. The scale-free network model graph can be built from a single node, and nodes can be added to the network at a constant rate over time.

The sub-pattern may be a pattern of any shape, such as a house, a circle, a grid, or other patterns. By adding different sub-patterns to the nodes of the basic graph data, the corresponding nodes may have specific categories.

Among them, edges are elements that connect various nodes in the basic graph data.

Among them, the preset number can be determined according to the number of nodes in the sample graph data. For example, the preset number can be 10%, 15%, etc. of the number of nodes, or it can be a specific number, such as 20, 25, etc.

For example, a composite graph consisting of a basic graph and four shapes can be generated, and the network structure of the composite graph is ,in is a collection of nodes, is a set of edges. Specifically, the synthetic graph can be composed of a base graph scale-free network model graph and multiple sub-patterns of four shapes, where the sub-patterns of four shapes can be houses, circles, diamonds, grids or other shapes. Then, the sub-patterns of four shapes are randomly attached to a node of the base graph, and then 10% of random edges are added to further perturb the synthetic graph to generate sample graph data.

Furthermore, for all sample nodes in the sample graph data, a target number of sample nodes may be selected randomly for label masking, for example, 15% of all sample nodes may be selected for label masking.

Through the above methods, sample graph data with diverse structures can be generated, which will help the preset model to more comprehensively understand and predict different types of graph structures and improve the generalization performance of the generalization model in real scenarios. In addition, by adding random edges to the synthetic graph to stir it and randomly selecting sample nodes to mask the labels to obtain sample graph data, the generalization model can still maintain good performance in the face of a certain degree of noise and incomplete information.

In some implementations, in order to effectively control the computational complexity and make the training process of the preset model more efficient, for each sample node, a sample subgraph corresponding to the sample node can be generated, so that the preset model can be used to train the sample subgraph and improve the generalization ability of the preset model. For example, "generating a corresponding sample subgraph for each sample node in the sample graph data according to the sample node connection structure of the sample graph data" in 101.1 includes:

(201.b1) Obtain the node density of the sample graph data, and determine the sample association level of each sample node based on the node density;

(201.b2) For each sample node as a central sample node, based on the sample node connection structure formed between the central sample node and the adjacent sample nodes, find out the sample subgraph of the target sample node connection structure of each central sample node in the sample association level.

The node density may be the degree of closeness between sample nodes in the sample graph data. For example, the node density may be calculated by the ratio of the number of sample nodes to the area of the sample graph data, or by other methods, which are not limited here.

Among them, when the density of the sample association level is high, it means that the connection between the sample nodes is relatively close. The sample association level can be the number of levels searched by the breadth-first search algorithm starting from the sample node to be predicted. When the node density of the sample graph data is large, a smaller sample association level can be selected to limit the scope of the search, reduce the amount of calculation, and better capture the relationship between the sample nodes. On the contrary, when the node density of the sample graph data is small, it means that the connection between the sample nodes is relatively sparse. In order to fully consider the correlation between the sample nodes in the graph, it is necessary to select a larger sample association level. By expanding the scope of the search, the potential relationship between the sample nodes can be better captured. Therefore, the sample association level can also be the sample search step.

The adjacent sample nodes may be sample nodes that are directly connected to the sample node to be predicted or are connected to each other through intermediate nodes in the sample graph data through connectivity.

Through the above method, it is possible to efficiently select a suitable sample sub-graph from the sample graph data, and save computing resources as much as possible while ensuring that the selected sample sub-graph is representative.

It should be noted that the first feature matrix of the sample graph data can be generated so that when the sample subgraph needs to be trained later, the initial feature matrix of the sample subgraph can be directly searched through the first feature matrix to obtain the initial feature matrix of the sample subgraph. Therefore, before obtaining the initial feature matrix and adjacency matrix of each sample subgraph, that is, before step 202, the following is also included:

(A1) Obtain the preset graph theory library;

(A2) obtaining preset characteristic indicators of the sample nodes, and generating a node characteristic vector of each sample node under the corresponding characteristic indicator through a graph theory library; wherein the characteristic indicator includes at least one of a node identifier, a node degree, a clustering coefficient, a betweenness centrality, and a closeness centrality;

(A3) Generate a first feature matrix according to a plurality of node feature vectors corresponding to the plurality of sample nodes.

Among them, the graph theory library can be multiple open source graph theory libraries, such as NetworkX library, igraph library and SNAP library. Taking NetworkX library as an example, NetworkX library is an open source Python package for creating and manipulating the structure, dynamics and functions of complex networks. NetworkX library provides a comprehensive set of graph theory and network analysis tools, which is suitable for studying various types of networks, including social networks, bioinformatics networks, transportation networks, etc. NetworkX library provides a wealth of functions and methods to study the structural characteristics of these graphs, including node identification, searching and traversing graphs, calculating node degrees, clustering coefficients, betweenness centrality, closeness centrality and other graph theory features, etc.

Among them, the preset characteristic index can be at least one of node identification, node degree, clustering coefficient, betweenness centrality and closeness centrality. It is understandable that the node identification can also be generated when the sample graph data is generated. Specifically, the node degree can be the number of edges connecting the sample node with the adjacent sample nodes, the clustering coefficient can be an indicator to measure the closeness of the connection between the current sample node and the neighboring nodes, the betweenness centrality can be used to measure the importance of the current sample node in connecting other sample nodes in the sample data graph, and the closeness centrality measures the distance between the current sample node and other sample nodes.

Furthermore, the sample association level of each sample node may also be determined by the clustering coefficient. Sample nodes with a high clustering coefficient may select a smaller sample association level, and sample nodes with a low clustering coefficient may select a larger sample association level.

The node feature vector may be a node feature vector generated in the graph theory library for each preset feature index. The node feature vector may include feature values of feature indexes such as node identification, node degree, clustering coefficient, betweenness centrality, and closeness centrality.

The first feature matrix may be composed of node feature vectors of preset feature indicators.

Exemplarily, if the clustering coefficient, betweenness centrality, closeness centrality and a random feature are taken as the node feature vector of each sample node, then the node feature vector of sample node 1 can be [0.75, 0.33, 0.5, 0.42], and each number represents the value corresponding to a specific feature indicator. Further, when the feature indicator also includes node identification and node degree, the node feature vector of sample node 1 can be [1, 2, 0.75, 0.33, 0.5, 0.42], and the node feature vector of sample node 2 can be [2, 4, 0.85, 0.54, 0.6, 0.59]. Further, when the sample graph data includes sample node 1 and sample node 2, the first feature matrix is composed of the node feature vectors of sample node 1 and sample node 2, and each row in the first feature matrix represents a sample node, and each column represents the eigenvalue of a feature indicator.

In the above manner, a first characteristic matrix of the sample graph data can be generated, so that an initial characteristic matrix of each sample sub-graph can be subsequently generated according to the first characteristic matrix. It is not necessary to generate a characteristic matrix separately for each sample sub-graph, thereby improving the efficiency of data processing.

Step 202, obtaining the initial feature matrix and adjacency matrix of each sample subgraph, and sequentially inputting the initial feature matrix into a preset model for feature aggregation to obtain a fused feature matrix of each sample subgraph.

In some embodiments, the feature expression capability of the central node can be enhanced by fusing the information of the central sample node and the adjacent sample nodes of the central sample node vector in the sample subgraph, so that it can capture richer local structure and context information to facilitate the subsequent extraction of core features.

The initial feature matrix may be a feature matrix of each sample subgraph, and the initial feature matrix includes a vector for describing the node features of each sample node in the sample subgraph. Specifically, the initial feature matrix may search for the corresponding node feature vector from the first feature matrix corresponding to the sample graph data according to the node identifier of each sample node, and generate the initial feature matrix according to the node feature vectors of multiple sample nodes.

The adjacency matrix may be a matrix used to represent the connection relationship between each sample node in the sample subgraph, and the adjacency matrix is a two-dimensional matrix, in which the elements represent whether there is an edge between the sample nodes. Exemplarily, for a sample subgraph with n nodes, the adjacency matrix is an n×n matrix, in which the elements are used to record the connection status between the sample nodes: the elements in the adjacency matrix can be represented by 0 and 1, in which 1 represents the existence of an edge and 0 represents the absence of an edge.

Among them, the preset model can be a trained graph neural network model. By training the preset model and updating the model parameters, a generalized model can be obtained after the preset model is trained.

The fused feature matrix may be a feature matrix obtained by updating the feature vector of the central sample node in the initial feature matrix. The fused feature matrix may be obtained by a preset model by performing feature conversion and aggregation on the information of each sample node and its adjacent sample nodes.

Through the above method, for each sample subgraph, the important features of each sample node can be fused to obtain a fused feature matrix, so that the preset model has higher prediction accuracy for the sample subgraph, which is beneficial to subsequent classification tasks.

In some implementations, the node feature vector of the corresponding sample subgraph can be searched in the first feature matrix of the sample graph data according to the node identifier of each sample node in the sample subgraph, and an initial feature matrix of each sample subgraph can be generated, so as to improve the generation efficiency of the initial feature matrix. Exemplarily, "obtaining the initial feature matrix and adjacency matrix of each sample subgraph" in (101.2) includes:

(202.a1) In each sample subgraph, determine the node identity of each sample node;

(202.a2) Based on the node identifier, determine the node feature vector corresponding to each sample node from the first feature matrix;

(202.a3) Generate an initial feature matrix of the sample subgraph according to multiple node feature vectors corresponding to multiple sample nodes of the sample subgraph;

(202.a4) Generate the adjacency matrix corresponding to the feature matrix based on the connection relationship between sample nodes.

The node identifier may be a unique identifier or index value of each sample node in each sample graph data and sample subgraph. Through the node identifier, the node feature vector of the corresponding sample subgraph can be quickly located in the first feature matrix of the sample graph data.

For example, assume that the sample subgraph includes multiple sample nodes, each of which has a unique user ID, such as user123, user456, etc. Then, according to user123 and user456, the node feature vectors of the corresponding sample nodes are found from the sample graph data, such as [user123, 2, 0.75, 0.33, 0.5, 0.42], [user456, 4, 0.85, 0.54, 0.6, 0.59], [user789, 3, 0.66, 0.24, 0.6, 0.5], etc., that is, the node feature vectors of user123 can be found. The node feature vector is [user123, 2, 0.75, 0.33, 0.5, 0.42], the node feature vector of user456 is [user456, 4, 0.85, 0.54, 0.6, 0.59], and the initial feature matrix of the sample subgraph is composed according to [user123, 2, 0.75, 0.33, 0.5, 0.42] and [user456, 4, 0.85, 0.54, 0.6, 0.59].

Exemplarily, when there is a connection relationship between sample nodes, that is, when there is an edge, a corresponding adjacency matrix can be generated. For example, when the sample subgraph is a social network graph, if two users are friends, then there is a connection relationship between the two users, and the corresponding elements of the two users in the adjacency matrix are 1, otherwise they are 0.

It can be understood that directly using the existing first feature matrix to obtain the initial feature matrix can avoid recalculating the initial feature matrix in the sample subgraph, significantly improving the efficiency and accuracy of generating the initial feature matrix, and facilitating the subsequent acquisition of core features and redundant features.

It should be noted that the central sample node is the core of the sample subgraph, that is, the classification object of the sample subgraph. The quality of its features directly affects the analysis of the sample subgraph by the preset model. Therefore, the sample subgraph can be feature fused so that the central sample node can absorb and integrate the information of the connected adjacent sample nodes, thereby obtaining a more comprehensive and accurate feature representation of the central sample node. For example, the step 202 of "inputting the initial feature matrix into the preset model for feature aggregation to obtain the fused feature matrix of each sample subgraph" includes:

(202.b1) According to the node identifier of each sample subgraph, determine the node degree matrix of the sample subgraph from the first feature matrix;

(202.b2) Obtain the adjacency matrix of the sample subgraph and normalize the adjacency matrix using the node degree matrix to obtain the normalized adjacency matrix;

(202.b3) Take the central sample node as the fusion center, and perform layer-by-layer fusion of sample node features based on the normalized adjacency matrix to obtain the updated central node feature vector corresponding to the central sample node;

(202.b4) Update the node feature vector of the central sample node of the initial feature matrix according to the central node feature vector to obtain the fused feature matrix of each sample subgraph.

Among them, the node degree matrix can be a diagonal matrix of the sample subgraph, and each element on the diagonal of the node degree matrix represents the node degree of the corresponding node in the graph. The node degree of each sample node can be obtained from the first feature matrix through the node identification of each sample node in the sample subgraph. Exemplarily, when the node feature vector of each sample node is composed of node identification, node degree, clustering coefficient, betweenness centrality, closeness centrality and a random feature, such as [user123, 2, 0.75, 0.33, 0.5, 0.42], then when the node identification of the sample node is user123, the corresponding node feature vector can be found from the sample graph data as [user123, 2, 0.75, 0.33, 0.5, 0.42], then the node degree of the sample node can be found to be 2. The node degree matrix of the sample subgraph can be generated by multiple node degrees corresponding to multiple sample nodes in the sample subgraph.

The central node feature vector may be a node feature vector of a central sample node in each sample subgraph, the central sample node is a classification object of the sample subgraph, and the preset model may classify the central sample node based on the central node feature vector.

Specifically, the fusion process of the central node feature vector may include adjacency matrix normalization, feature aggregation, linear transformation and nonlinear activation, multi-layer iteration and readout. Exemplarily, the normalized Laplacian matrix can be calculated first to reduce the influence of sample nodes with large degrees in feature propagation, realize feature normalization, and maintain the stability of feature scale during propagation. After that, in each layer, the sample node feature vector and the normalized adjacency matrix are multiplied so that each sample node can receive the information of its adjacent sample nodes. After that, the aggregated features are linearly transformed through the weight matrix, and the activation function is applied to update the representation of the sample node. Further, through the iteration of multiple layers, the central sample node gradually integrates the information of the more distant adjacent sample nodes. In each layer, the representation of the central node is updated according to the node feature vector of its adjacent sample nodes. After completing the iteration of all layers, the node feature vector feature of the last layer of the central sample node can be directly used as the central node feature vector of the central sample node, or the node feature vector of all adjacent sample nodes can be pooled to obtain the representation of the entire sample subgraph and used as the central node feature vector of the central sample node.

For example, the subgraph information can be fused by means of graph convolutional neural networks (GCN) to obtain the central node feature vector of the central sample node. The specific formula is as follows:

in, represents the node feature vector of the sample node v in the l+1th layer; g(·) represents the activation function, such as ReLU or tanh, which is used to introduce nonlinearity; Represents the normalized node degree matrix The inverse square root of is the node degree matrix, Each element of represents the node degree of sample node i; A represents the adjacency matrix of the sample subgraph, which is used to represent the connection relationship between each sample node in the sample subgraph; represents the node feature vector of the sample node v of the lth layer; W represents the weight matrix of the lth layer, and W is the model parameter obtained after the preset model training and learning.

For each sample subgraph, the process of fusing subgraph information includes normalizing the adjacency matrix by the node degree matrix, that is, , in order to reduce the influence of sample nodes with large node degrees in feature propagation, realize feature normalization, and maintain the stability of feature scale during propagation.

Furthermore, the node feature vector of the sample node v at layer l can be Multiply it with the normalized adjacency matrix so that each sample node in layer l can receive information from its adjacent sample nodes. Furthermore, the node feature vector of the sample node v in layer l can be After multiplying it with the normalized adjacency matrix, it is multiplied by the weight matrix W and the activation function g(·) is applied to transform the node feature adjacency of each current sample node from the original space to the new feature space and introduce nonlinear activation. In this way, the updated feature vector of the sample subgraph can be obtained .

Furthermore, after completing the iteration of all layers, the node feature vector of the last layer of the central sample node can be directly used as the representation of the central sample node, or the node feature vectors of all sample nodes can be pooled to obtain the representation of the entire sample subgraph and use it as the global representation of the central sample node.

Furthermore, the central node feature vector of the central sample node can be updated, and the node feature vector of the central sample node in the initial feature matrix can be replaced by the central node feature vector, thereby obtaining a fusion feature matrix of each sample subgraph.

Through the above method, the node degree matrix and the adjacency matrix can be used to normalize the node feature vectors of the sample subgraph, which can reduce the influence of sample nodes with larger node degrees in the sample subgraph on feature propagation and maintain the scale stability in the feature propagation process, so that each sample node can receive information from adjacent sample nodes more fairly. In addition, by fusing the node feature vectors of the adjacent sample nodes of the central sample node layer by layer, the information in the sample subgraph can be effectively aggregated to enhance the representation ability of the central sample node.

Step 203: Determine the core features and redundant features of each sample subgraph according to the adjacency matrix and the fusion feature matrix.

In some embodiments, the core features and redundant features of the sample subgraph can be calculated through the adjacency matrix and the fusion feature matrix, so that the preset model can be trained through the core features and redundant features later, so that the preset model focuses more on causal features and reduces the attention to redundant features.

In some embodiments, after the central node feature vector of each sample subgraph is updated, the central node feature vector of the central sample node can be updated in the first feature matrix of the sample graph data according to the node identifier of the central sample node. And after each central node feature vector in the first feature matrix is updated, the corresponding node feature vector is updated synchronously in all sample subgraphs. In this way, the representation of the node feature vector of the sample node in each sample subgraph can be enriched as much as possible. Alternatively, only the central node feature vector of the central sample node in each sample subgraph can be updated, so that the preset model can pay more attention to the core features during the training process.

Exemplarily, the node feature vectors of any two sample nodes in the sample subgraph can be input into a multilayer perceptron to obtain the corresponding edge mask values, and an edge mask matrix can be generated based on all edge mask values corresponding to the sample subgraph. Furthermore, the edge mask matrix can be multiplied with the adjacency matrix of the corresponding sample subgraph to adjust the weights of each edge, and then combined with the fusion feature matrix to obtain the core structure diagram, and the core network diagram is processed by a preset model to obtain the core features. It can be understood that through the edge mask matrix, the weights of each existing edge in the adjacency matrix can be adjusted to make the important edges between sample nodes more prominent.

Exemplarily, the complement matrix of the edge mask matrix can be multiplied with the adjacency matrix of the corresponding sample subgraph to adjust the weight of each edge, and then combined with the fusion feature matrix to obtain a redundant network graph, and the redundant structure graph is processed by a preset model to obtain the core features. It can be understood that since the edge mask matrix contains the weight values of the edges between each sample node, the redundant matrix of the edge mask matrix contains the redundant parts of the edges between each sample node.

The core features and redundant features calculated in the above manner can facilitate the subsequent training of the preset model, so that the preset model can focus more on causal features and reduce the focus on shortcut features.

In some implementations, in order to accurately obtain core features and redundant features, the weight of the adjacency matrix may be adjusted by an edge mask matrix to quantify the importance of the edge between two sample nodes. Exemplarily, step 203 may include:

(203.1) Based on the fusion feature matrix of the sample subgraph, any two sample nodes in the sample subgraph are input into the multi-layer perceptron to obtain the edge mask value between any two sample nodes;

(203.2) generating an edge mask matrix of the sample subgraph according to a plurality of edge mask values in the sample subgraph;

(203.3) Generate the core features corresponding to the sample subgraph based on the adjacency matrix, fusion feature matrix and edge mask matrix;

(203.4) Generate the corresponding complement matrix based on the edge mask matrix of the sample subgraph;

(203.5) Based on the adjacency matrix, fusion feature matrix and complement matrix, the redundant features of the sample subgraph are obtained.

The edge mask value can be a scalar value obtained by calculating the multi-layer perceptron for each edge in the sample subgraph. The edge mask value indicates the importance of the edge between two sample nodes. Specifically, the edge mask value can be calculated by inputting any two sample nodes in the sample subgraph into the multi-layer perceptron. For example, the edge mask value of the edge can be 0, 0.1, 0.6, etc.

The edge mask matrix may be a matrix generated according to all edge mask values by sequentially inputting node feature vectors of any two sample nodes of the sample subgraph into the multilayer perceptron until all sample nodes are input.

Specifically, you can find it in the sample subgraph In the example, after pairing any two sample nodes, the node feature vectors of the two sample nodes are input into the multilayer perceptron (MLP) to obtain an edge mask value between 0 and 1. After obtaining the edge mask values between all sample nodes in the sample, the sample subgraph can be constructed. The edge mask matrix .

Furthermore, the edge mask matrix can be and sample subgraph The adjacency matrix A is multiplied element by element and combined with the fusion feature matrix X to obtain the core structure diagram .

Furthermore, the complement matrix is , the complement matrix and sample subgraph The adjacency matrix A is multiplied element by element and combined with the fusion feature matrix X to obtain the redundant structure diagram .

Specifically, the core structure graph and the redundant structure graph can be processed by the preset model to obtain the core features and redundant features .

Through the above methods, the structural information of the sample subgraph can be better captured, and the prediction accuracy and interpretability of the preset model for the sample subgraph can be improved as much as possible. By generating redundant features, redundant data, such as the negative impact of noise data on the performance of the preset model, can be reduced, so as to subsequently improve the preset model's attention to core features and reduce dependence on redundant features.

Step 204, determining the target loss of the preset model based on the core features and the redundant features.

The generalized model is a model obtained by training a preset model and stopping training after the preset model converges or reaches a preset number of training times.

Among them, the target loss can be a loss composed of a first loss, a second loss, and a third loss, which is used to indicate the adjustment of the parameters of the preset model to train the preset model to obtain a generalized model. Specifically, the first loss can be a cross entropy loss, the second loss can be an entropy loss, and the third loss can be a contrast loss. By forming the target loss with the first loss, the second loss, and the third loss, and using the target loss to update the model parameters, the generalized model can be made to learn the core features more attentively, avoid the generalized model from over-focusing on redundant features, and thus improve the quality of the overall feature representation.

It can be understood that by inputting the core features and redundant features into the preset model, the first loss can be determined by the difference between the predicted node category label output by the core features of the sample subgraph and the sample node category label of the core features, the second loss can be determined based on the uniformity of the probability distribution when predicting each sample node category label for the redundant features of the sample subgraph based on the preset model, and the third loss can be determined according to the sample distance between the core features and the redundant features. In this way, the generalization model can focus on the causal features, i.e., the core features, reduce the reliance on unstable redundant features, reduce the risk of overfitting, and make more reliable decisions.

In some implementations, in order to enable the generalization model to better capture the essential structure of the data, thereby having better generalization ability when facing test data with different distributions, the preset model can be trained based on the core features and redundant features so that the preset model has better classification performance. For example, step 204 may include:

(204.1) determining a first loss based on a difference between a predicted node category label output for a core feature of a sample subgraph and a sample node category label of the core feature;

(204.2) Determine the second loss based on the uniformity of probability distribution when predicting the category label of each sample node according to the redundant features of the sample subgraph based on the preset model;

(204.3) Determine the third loss based on the sample distance between the core features and the redundant features;

(204.4) Based on the sum of the first loss, the second loss and the third loss, the target loss of the preset model is constructed.

The first loss may be a cross entropy loss, which is used to evaluate the difference between the probability distribution of each predicted node category label predicted by the preset model for the core feature and the sample node category label of the core feature. Specifically, a unique hot encoding may be set for each sample node category label, each sample node category label is represented by a binary vector, and only the real sample node category label corresponding to the core feature is set to 1, and the remaining sample node category labels are set to 0. Specifically, the first loss may be:

in, is the one-hot encoding of the sample node category label, is the model's first predicted probability for the vth sample node category label. When the predicted node category label is consistent with the actual sample node category label corresponding to the core feature, the first loss is minimal. Therefore, by minimizing the first loss, the preset model can better learn the causal relationship of the data, thereby continuously making more accurate predictions.

The predicted node category labels are multiple node category labels predicted by the preset model for the core features. For example, for an animal image, the predicted node category labels may be cat, dog, and rabbit, and each predicted node category label corresponds to a predicted probability.

The second loss may be an entropy loss, which is used to measure the uniformity of the probability distribution of the preset model when predicting the category labels of each sample node based on the redundant features of the sample subgraph. The more uniform the probability distribution is, the smaller the classification preference of the preset model for the redundant features is, thereby effectively preventing the preset model from learning shortcut features. Specifically, the second loss may be:

in, Represents the uniform probability of each sample node category label when uniformly distributed, usually =1/number of categories, It is the second predicted probability of the preset model for the category label of the vth sample node.

Specifically, when redundant features are input into the preset model, the output second prediction probabilities are close to uniform distribution, indicating that redundant features are not helpful for the classification of the preset model. For example, if the sample category labels are 4, namely cat, rabbit, dog, and pig, then =1/4, if redundant feature 1 is input into the preset model, and the predicted probabilities of cat, rabbit, dog, and pig are all 1/4, the second loss of the preset model is also the smallest. At this time, it shows that the preset model has no classification preference for redundant features, and the output probability distribution is highly uniform. Therefore, it is necessary to continuously train the preset model so that the predicted probability of the preset model for the redundant features in the category labels of each sample node is close to uniform distribution.

The uniformity of probability distribution may be the second preset probability of each sample node category label predicted and output by the preset model for the redundant features after the redundant features are input into the preset model. The more uniform the second preset probability distribution is, the higher the uniformity of probability distribution is. For example, the uniformity of probability distribution with the second preset probabilities of 1/3, 1/3, and 1/3 is greater than 2/3, 1/6, and 1/6.

The third loss may be a contrast loss. Specifically, the core features may be used as positive samples and the redundant features may be used as negative samples. The contrast loss may encourage the preset model to correctly distinguish between positive samples and negative samples, so that the distance between positive samples is closer and the distance between positive samples and negative samples is farther. Furthermore, the third loss may be:

in, is the sigmoid function, which is usually used to map the input to (0, 1), Q is a hyperparameter, C and are different instances of the core features, representing positive sample pairs, C and Both represent positive samples, i.e. core features; C and represents a positive and negative sample pair, Represents negative samples, i.e., redundant features; It represents the expected value, which can be obtained by averaging the first predicted probability of the core features of all positive sample pairs by the preset model and the number of second predicted probabilities of the redundant features of all negative sample pairs. By training the preset model with the third loss, the preset model can be encouraged to correctly distinguish positive samples from negative samples, so that the distance between positive samples is closer and the distance between positive samples is farther.

The sample distance may be the distance between two core features or the distance between a core feature and a redundant feature. The sample distance may be calculated by Euclidean distance, cosine similarity, and the like.

Among them, the core feature can be a feature with causal relationship or causal explanation ability in the sample subgraph, and the core feature has causal information of the sample node to be predicted and the sample node category label.

Among them, redundant features can be shortcut features between core features and prediction results, which have no additional contribution to the prediction of sample nodes and sample node category labels, and do not have causal explanation characteristics. Moreover, redundant features will make the preset model tend to learn shortcut features to make decisions, reduce the accuracy of the preset model's prediction of sample graph data, and thus cause the performance of the generalized model obtained by training to degrade in out-of-distribution test data. Exemplarily, redundant features can be noise features.

Exemplarily, if the sample graph data is a social network graph, the preset model needs to classify users in the social network, and the labels include "student", "worker", "freelancer", etc. The sample subgraph is the activity record of the user (i.e., the sample node to be predicted) on the social network, such as posts, photos, interactions, etc. The core features may include the user's occupational information, educational background, etc. After the core features are input into the preset model, the preset model outputs the first predicted probability that the user is a student, a worker, and a freelancer. Further, the cross entropy loss can be used as the first loss to calculate the difference between the probability distribution under each prediction node category label output by the preset model and the actual category label (represented by one-hot encoding). For example, if the actual sample category label is a student, the one-hot encoding is [1, 0, 0]; if the predicted probability is [0.7, 0.2, 0.1], the first loss will calculate the cross entropy loss between the two, and adjust the parameters of the preset model to encourage the preset model to test the sample nodes to be predicted in the sample subgraph, thereby improving the prediction accuracy of the student category.

Exemplarily, redundant features may be weather information at the user's geographic location. Although this appears to be relevant to the classification task, redundant features may actually introduce bias and are of no practical help in classifying the user's occupational information. Therefore, entropy loss may be calculated as the second loss. When the output of the preset model for redundant features tends to be evenly distributed, it indicates that the preset model has not learned any biased classification decisions from these redundant features, thereby encouraging the preset model to focus mainly on core features that are of substantial help in classification.

Furthermore, in the process of training the preset model, the contrast loss can be used to enable the preset model to distinguish relevant features (i.e., core features) from irrelevant features (i.e., redundant features), and learn to bring core features (positive samples) closer to each other as much as possible, while moving redundant features (negative samples) away from core features. Furthermore, the contrast loss is used as the third loss to calculate the similarity between positive sample pairs, and to penalize the similarity between positive and negative sample pairs, so that the learned embedding space can better distinguish relevant and irrelevant information.

It can be understood that by combining the first, second and third losses to form the target loss, it can ensure that the preset model has a balanced performance improvement in different aspects. The optimization of the target loss will help the preset model to learn the causal characteristics of the samples while ignoring redundant features that are not helpful for classification, and better distinguish the core features related to the prediction task.

Step 205, updating the parameters of the preset model based on the target loss, and when the preset model converges, a trained generalized model is obtained.

Furthermore, during the training of the preset model, the parameters of the preset model can be continuously updated according to the target loss, for example, by a gradient descent algorithm. When the loss of the preset model is no longer significantly reduced, or a certain termination condition is reached (for example, a preset number of training times is reached or the preset model converges), the training of the preset model can be stopped to obtain a generalized model.

In summary, by combining multiple loss functions to train the preset model, the preset model is more likely to learn the causal features that are decisive for classification, rather than simply relying on shortcut features that are easy to learn but may lead to overfitting. In this way, a generalized model with better generalization performance and higher prediction accuracy can be obtained.

Please refer to Figure 4. In some embodiments, the overall process of the present application is introduced in conjunction with Figure 4. Exemplarily, the sample graph data can be generated by random generation, and a sample subgraph is generated in the sample association level for each sample node in the sample graph data. Furthermore, feature fusion is performed for the central sample node of each sample subgraph as a fusion center to generate a central node feature vector of the central sample node, and the initial feature matrix of the corresponding sample subgraph is updated according to the central node feature vector to obtain a fused feature matrix, so as to increase the attention to the central sample node, that is, the sample node to be predicted.

Furthermore, any two sample nodes in the sample subgraph are passed through a multi-layer perceptron to generate edge mask values, and then an edge mask matrix is generated based on the direct edge mask values of all sample nodes. The edge mask matrix is multiplied element by element with the adjacency matrix, and then combined with the fusion feature matrix to obtain the core features. Furthermore, the complementary matrix of the edge mask matrix is multiplied element by element with the adjacency matrix, and then combined with the fusion feature matrix to obtain redundant features, thereby quantifying the importance of the edge.

Furthermore, the first loss can be determined based on the difference between the predicted node category label output for the core feature of the sample subgraph and the sample node category label of the core feature, the second loss can be determined based on the uniformity of the probability distribution when predicting each sample node category label for the redundant features of the sample subgraph based on the preset model, the third loss can be determined based on the sample distance between the core feature and the redundant feature, and the target loss of the preset model can be constructed based on the sum of the first loss, the second loss and the third loss. In this way, the parameters of the preset model can be updated based on the target loss, and when the preset model converges, a well-trained generalization model with good generalization ability is obtained, and then the generalization model can be used to accurately predict and evaluate data outside the distribution range.

In an embodiment of the present application, target graph data is obtained, wherein the target graph data includes a plurality of target nodes and a node connection structure composed of a plurality of target nodes; based on the node connection structure composed of a plurality of target nodes in the target graph data, a target subgraph of the target connection structure of each target node in the target association level is found; the target subgraph is input into a generalization model to obtain a node category label of the target node in the target subgraph; wherein the generalization model is obtained by performing causal prediction learning between the sample node connection structure and the sample node category label in the sample subgraph based on a training process of minimizing the target loss, and the target loss is composed of a first loss, a second loss and a third loss, and the first loss is determined based on the difference between the predicted node category label output for the core feature of the sample subgraph and the sample node category label of the core feature; the second loss is determined based on the uniformity of the probability distribution when predicting each sample node category label for the redundant feature of the sample subgraph according to the preset model; the third loss is determined according to the sample distance between the core feature and the redundant feature, and the core feature is a feature obtained by enhancing the sample node connection structure of the sample subgraph according to the edge mask matrix obtained by the attention mechanism; the redundant feature is a feature obtained by weakening the sample node connection structure of the sample subgraph according to the complement matrix of the edge mask matrix. In this way, the preset model can be trained by constructing a target loss, and during the training process, the preset model tends to predict through core features, reduce the preference for redundant features, and reduce the sample distance between core features and redundant features, thereby enabling the trained generalized model to have stronger causal explanation ability and generalization ability. In addition, the edge mask matrix obtained by the attention mechanism enhances the sample connection structure representing the sample subgraph, and the complement matrix of the edge mask matrix weakens the sample node connection structure representing the sample subgraph, which helps the trained generalized model to better capture the important relationship between target nodes, reduce the interference of redundant information, and make the preset model more focused on the learning of core features. In the process of applying the generalized model, the target subgraph of the target node to be predicted can be obtained based on the target graph data, so that the generalized model does not need to predict the entire target graph data, but only needs to obtain the relevant target subgraph of the target node to be predicted, which helps the generalized model to analyze the causal relationship between target nodes more attentively, and improves the efficiency and accuracy of prediction. In summary, the present application can improve the accuracy of the trained generalized model in predicting graph data.

Referring to FIG. 5 , an embodiment of the present application further provides a graph data prediction device based on a generalization model, which can implement the above-mentioned graph data prediction method based on a generalization model. The graph data prediction device based on a generalization model includes:

An acquisition module 51 is used to acquire target graph data, wherein the target graph data includes a plurality of target nodes and a node connection structure composed of the plurality of target nodes;

A search module 52, configured to search for a target subgraph of a target connection structure of each target node in a target association level based on a node connection structure formed between multiple target nodes in the target graph data;

The input module 53 is used to input the target subgraph into the generalization model to obtain the node category label of the target node in the target subgraph; wherein the generalization model is obtained by performing causal prediction learning between the sample node connection structure and the sample node category label in the sample subgraph based on a training process of minimizing the target loss, and the target loss is composed of a first loss, a second loss and a third loss. The first loss is determined based on the difference between the predicted node category label output for the core feature of the sample subgraph and the sample node category label of the core feature; the second loss is determined based on the uniformity of the probability distribution when predicting each sample node category label for the redundant features of the sample subgraph based on the preset model; the third loss is determined according to the sample distance between the core feature and the redundant feature, the core feature is the feature obtained by enhancing the sample node connection structure of the sample subgraph according to the edge mask matrix obtained by the attention mechanism; the redundant feature is the feature obtained by weakening the sample node connection structure of the sample subgraph according to the complement matrix of the edge mask matrix.

The specific implementation of the graph data prediction device based on the generalization model is basically the same as the specific implementation of the graph data prediction method based on the generalization model mentioned above, and will not be repeated here. On the premise of meeting the requirements of the embodiments of this application, the graph data prediction device based on the generalization model can also be provided with other functional modules to implement the graph data prediction method based on the generalization model in the above embodiments.

The embodiment of the present application also provides a computer device, the computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the above-mentioned graph data prediction method based on the generalization model when executing the computer program. The computer device can be any intelligent terminal including a tablet computer, a car computer, etc.

Please refer to FIG. 6 , which schematically shows the hardware structure of a computer device according to another embodiment. The computer device includes:

The processor 61 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

The memory 62 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 62 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 62, and the processor 61 calls and executes the graph data prediction method based on the generalized model in the embodiment of the present application;

Input/output interface 63, used to implement information input and output;

Communication interface 64, used to realize communication interaction between the device and other devices, which can be realized by wired mode (such as USB, network cable, etc.) or wireless mode (such as mobile network, WIFI, Bluetooth, etc.);

A bus 66 that transmits information between the various components of the device (e.g., the processor 61, the memory 62, the input/output interface 63, and the communication interface 64);

The processor 61 , the memory 62 , the input/output interface 63 and the communication interface 64 are connected to each other in communication within the device via a bus 66 .

An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned graph data prediction method based on the generalization model.

The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules/units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

It should be understood that in the present application, "at least one (item)" and "several" refer to one or more, and "plurality" refers to two or more. "And/or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and/or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character "/" generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

In the several embodiments provided in the present application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of the above 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 system, 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 an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

The units described above 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.

In addition, each functional unit in each embodiment of the present application 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. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

The preferred embodiments of the present application are described above with reference to the accompanying drawings, but the scope of the rights of the present application is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present application should be within the scope of the rights of the present application.

Claims (11)

1. A graph data prediction method based on a generalization model, the method comprising:

obtaining target graph data, wherein the target graph data comprises a plurality of target nodes and a node connection structure formed by the plurality of target nodes;

Searching a target subgraph of a target connection structure of each target node in a target association level based on a node connection structure formed by a plurality of target nodes in the target graph data;

Inputting the target subgraph into a generalization model to obtain a node class label of a target node in the target subgraph;

The generalization model carries out causal prediction learning between a sample node connection structure and a sample node class label in a sample sub-graph based on a training process of minimizing target loss, wherein the target loss is formed by a first loss, a second loss and a third loss, and the first loss is determined based on the difference between a predicted node class label output for core characteristics of the sample sub-graph and a sample node class label of the core characteristics; the second loss is determined based on probability distribution uniformity when the redundancy features of the sample subgraph are used for predicting the category labels of each sample node according to a preset model; the third loss is determined according to the sample distance between the core feature and the redundant feature, wherein the core feature is a feature obtained by enhancing and representing a sample node connection structure of a sample sub-graph according to an edge mask matrix obtained according to an attention mechanism; the redundant features are features obtained by weakening and representing a sample node connection structure of a sample sub-graph according to a complement matrix of the edge mask matrix;

The sample nodes of the sample subgraph comprise a central sample node and a plurality of adjacent sample nodes; the generalization model is obtained by training in the following way:

Acquiring sample graph data for training a preset model, and generating a corresponding sample subgraph for each sample node in the sample graph data according to a sample node connection structure of the sample graph data;

acquiring an initial feature matrix and an adjacent matrix of each sample sub-graph, and sequentially inputting the initial feature matrix into a preset model to perform feature aggregation to obtain a fusion feature matrix of each sample sub-graph;

according to the adjacency matrix and the fusion feature matrix, determining core features and redundant features of each sample subgraph;

determining a target loss of the preset model based on the core feature and the redundancy feature;

And updating parameters of the preset model based on the target loss, and obtaining a trained generalization model when the preset model converges.

2. The generalized model based graph data prediction method according to claim 1, wherein the determining the core feature and the redundancy feature of each of the sample subgraphs according to the adjacency matrix and the fusion feature matrix includes:

Inputting any two sample nodes in the sample subgraph to a multi-layer perceptron based on a fusion feature matrix of the sample subgraph to obtain an edge mask value between any two sample nodes;

generating an edge mask matrix of the sample subgraph according to a plurality of edge mask values in the sample subgraph;

Generating core features corresponding to the sample subgraph based on the adjacency matrix, the fusion feature matrix and the edge mask matrix;

generating a corresponding complement matrix according to the edge mask matrix of the sample subgraph;

and obtaining the redundant features of the sample subgraph based on the adjacent matrix, the fusion feature matrix and the complement matrix.

3. The method for predicting map data based on a generalization model according to claim 1, wherein the obtaining sample map data for training a preset model comprises:

Acquiring basic graph data and a plurality of sub-patterns of different categories, and distributing the plurality of sub-patterns to any sample node of the basic graph data to obtain sample graph data;

determining a preset number of edges added to the sample graph data according to the number of nodes of the sample graph data, and adding edges to any two sample nodes in the sample graph data, wherein the number of edges added to the sample graph data is equal to the preset number.

4. The method for predicting graph data based on a generalization model according to claim 1, wherein the generating a corresponding sample subgraph for each sample node in the sample graph data according to a sample node connection structure of the sample graph data comprises:

Acquiring node density of the sample graph data, and determining a sample association level of each sample node based on the node density;

And aiming at each sample node serving as a central sample node, based on a sample node connection structure formed between the central sample node and the adjacent sample nodes, searching a sample subgraph of a target sample node connection structure of each central sample node in a sample association hierarchy.

5. The method for predicting map data based on a generalization model according to claim 1, further comprising, before obtaining the initial feature matrix and the adjacency matrix of each sample subgraph:

Acquiring a preset graph theory library;

acquiring preset characteristic indexes of the sample nodes, and generating node characteristic vectors of each sample node under the corresponding characteristic indexes through the graph theory library; the characteristic index comprises at least one of node identification, node degree, clustering coefficient, medium centrality and near centrality;

And generating a first feature matrix according to a plurality of node feature vectors corresponding to the plurality of sample nodes.

6. The method for predicting map data based on a generalization model according to claim 5, wherein said obtaining an initial feature matrix and an adjacency matrix of each sample subgraph comprises:

determining node identification of each sample node in each sample subgraph;

determining a node feature vector corresponding to each sample node from the first feature matrix based on the node identification;

Generating an initial feature matrix of the sample subgraph according to a plurality of node feature vectors corresponding to a plurality of sample nodes of the sample subgraph;

and generating an adjacent matrix corresponding to the feature matrix based on the connection relation between the sample nodes.

7. The method for predicting graph data based on a generalization model according to claim 5, wherein the sequentially inputting the initial feature matrix into a preset model for feature aggregation to obtain a fused feature matrix of each sample subgraph comprises:

determining a node degree matrix of each sample subgraph from the first feature matrix according to the node identification of each sample subgraph;

Acquiring an adjacent matrix of the sample subgraph, and normalizing the adjacent matrix through the node degree matrix to obtain a normalized adjacent matrix;

taking the central sample node as a fusion center, and carrying out layer-by-layer fusion of sample node characteristics based on the normalized adjacency matrix to obtain a central node characteristic vector after updating corresponding to the central sample node;

And updating the node characteristic vector of the central sample node of the initial characteristic matrix according to the central node characteristic vector to obtain a fusion characteristic matrix of each sample subgraph.

8. The generalized model based graph data prediction method according to claim 1, wherein the determining the target loss of the preset model based on the core feature and the redundancy feature includes:

determining a first penalty based on a difference between a predicted node class label output for a core feature of the sample subgraph relative to a sample node class label of the core feature;

Determining a second loss based on a preset model for predicting probability distribution uniformity when each sample node class label is predicted for the redundant features of the sample subgraph;

Determining a third loss based on a sample distance between the core feature and the redundant feature;

And constructing a target loss of the preset model based on the sum of the first loss, the second loss and the third loss.

9. A graph data prediction apparatus based on a generalization model, the apparatus comprising:

The system comprises an acquisition module, a storage module and a storage module, wherein the acquisition module is used for acquiring target graph data, and the target graph data comprises a plurality of target nodes and a node connection structure formed by the plurality of target nodes;

the searching module is used for searching a target subgraph of a target connection structure of each target node in a target association level based on a node connection structure formed by a plurality of target nodes in the target graph data;

The input module is used for inputting the target subgraph into the generalization model to obtain a node class label of a target node in the target subgraph; the generalization model carries out causal prediction learning between a sample node connection structure and a sample node class label in a sample sub-graph based on a training process of minimizing target loss, wherein the target loss is formed by a first loss, a second loss and a third loss, and the first loss is determined based on the difference between a predicted node class label output for core characteristics of the sample sub-graph and a sample node class label of the core characteristics; the second loss is determined based on probability distribution uniformity when the redundancy features of the sample subgraph are used for predicting the category labels of each sample node according to a preset model; the third loss is determined according to the sample distance between the core feature and the redundant feature, wherein the core feature is a feature obtained by enhancing and representing a sample node connection structure of a sample sub-graph according to an edge mask matrix obtained according to an attention mechanism; the redundant features are features obtained by weakening and representing a sample node connection structure of a sample sub-graph according to a complement matrix of the edge mask matrix; the sample nodes of the sample subgraph comprise a central sample node and a plurality of adjacent sample nodes; the generalization model is obtained by training in the following way: acquiring sample graph data for training a preset model, and generating a corresponding sample subgraph for each sample node in the sample graph data according to a sample node connection structure of the sample graph data; acquiring an initial feature matrix and an adjacent matrix of each sample sub-graph, and sequentially inputting the initial feature matrix into a preset model to perform feature aggregation to obtain a fusion feature matrix of each sample sub-graph; according to the adjacency matrix and the fusion feature matrix, determining core features and redundant features of each sample subgraph; determining a target loss of the preset model based on the core feature and the redundancy feature; and updating parameters of the preset model based on the target loss, and obtaining a trained generalization model when the preset model converges.

10. A computer device, characterized in that it comprises a memory storing a computer program and a processor implementing the generalization model based graph data prediction method according to any of claims 1 to 8 when said computer program is executed.

11. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the generalization model-based graph data prediction method of any one of claims 1 to 8.