CN116484911A - Distribution robust optimization method and system suitable for out-of-distribution generalization of graph neural networks - Google Patents

Abstract

The invention provides a distribution robust optimization method and a system suitable for map neural network distribution generalization, wherein the method comprises the following steps: selecting an arbitrary graph convolution neural network model, aggregating neighbor features of a target individual in graph data, and updating features of the target individual; obtaining a final result prediction of an individual through a model by using a loss function related to a task target, and calculating a corresponding loss function with an individual label to obtain a loss value corresponding to each individual; according to the calculated loss value, obtaining the weight of each individual through a distribution robust optimization algorithm based on KL divergence, wherein the weight distribution is the worst distribution of the current individual; carrying out smoothing operation on the weights of the individuals so that weights of training individuals with similar distances or similar local structures are also similar; calculating the overall loss of all training individuals through the adjusted weight values and the corresponding loss values; and optimizing the used graph convolution neural network model by utilizing the overall training loss.

Description

Distribution robust optimization method and system suitable for out-of-distribution generalization of graph neural networks

technical field

The invention relates to the field of graph neural networks, in particular to a distribution robust optimization method and system suitable for out-of-distribution generalization of graph neural networks.

Background technique

Nowadays, people pay more and more attention to graph-structured data, and various graph-related algorithms have emerged as the times require. Among them, graph neural networks have been extensively studied and applied to many fields. As a deep learning model, the graph neural network's credibility is very important, especially in the fields of finance or biological research. An important branch of trusted AI is the out-of-distribution generalization of the model. Since the data used by the model in the training process may be different from the data distribution faced in the prediction, the performance of the model during the test and the training phase will have a large gap, which makes the model lack the most basic credibility.

There are currently several works dedicated to addressing the out-of-distribution generalization problem of graph convolutional neural networks.

EERM (Explore-to-Extrapolate Risk Minimization) first uses the generator to divide all nodes into different environments, and then calculates the average loss of nodes in each environment, trains the classification model to minimize the mean and variance of all environmental losses, and trains the generator to maximize the loss variance, so that the classification model can have good classification effects on data in different environments and improve the generalization ability of the model. However, this method requires a large amount of calculation, so it is difficult to apply to a large-scale data set, and requires high computing resources. Shift-Robust GNNs (SRGNN) proposes to improve the generalization ability by reducing the distribution difference between training samples and test samples. This method first needs to sample a part of unbiased samples from the test set, input them into the graph convolutional neural network during training, obtain the corresponding feature representation, and then calculate the feature distribution difference between the training samples. The commonly used measurement method is the Maximum Mean Discrepancy (MMD), and the measurement loss will be used as a regular term to optimize the model together with the cross-entropy loss. This method needs to obtain the data of the test phase during the training phase, which is difficult to meet in actual scenarios. GraphOut-of-Distribution Benchmark (GOOD) proposes a test platform for graph convolutional neural networks on out-of-distribution generalization issues, and constructs multiple graph datasets with distribution drift, such as GOOD-Cora, GOOD-Arxiv, etc. Among them, some methods to solve the distribution drift problem in European data, such as invariant risk minimization (IRM), variance-risk extrapolation (Variance-Risk Extrapolation, V-REx), etc., are directly combined with the graph convolutional neural network. However, there is a common problem in these methods, that is, in the training phase, in addition to the classification label of the sample, it is also necessary to know the environmental label of each sample, and the environmental label is difficult to obtain in reality, so the cost of obtaining data labels in practical applications is too high. In solving the distribution drift problem of European data, there is another way, that is, the distribution robust optimization algorithm based on KL divergence, which improves the generalization ability of the model by making the model focus on the worse data distribution during the training process. This method has two advantages. One is that the environmental label of the sample does not need to be known during the training process, so the label cost is low; the other is that the computational complexity is low, and the training speed is basically the same as the traditional Empirical Risk Minimization (ERM) method, while the requirements for computing resources are not high. Distribution-robust optimization algorithms based on KL divergence have made great progress in solving the out-of-distribution generalization problem of deep learning models, but in graph neural networks, this method does not consider graph structure information. Therefore, how to improve the distribution robust optimization method to make it more suitable for graph data scenarios has become an urgent problem to be solved.

Contents of the invention

The purpose of the present invention is to overcome the deficiency of distribution robust optimization based on KL divergence in graph neural networks, propose a weight smoothing method, improve it, and make it applicable to graph convolutional neural networks. The present invention can further improve the performance of the graph neural network on the problem of out-of-distribution generalization,

At the same time, it has strong flexibility, plug and play.

The present invention is realized through at least one of the following technical solutions.

A distribution-robust optimization method suitable for out-of-distribution generalization of graph neural networks, comprising the following steps:

Step 1. Input the training individual into the graph convolutional neural network model, complete the aggregation, feature transformation and update of the target individual's own characteristics on the neighbor features of the target individual, and finally obtain the predicted output;

Step 2, calculate the loss value of each training individual;

Step 3, using the distribution robust optimization algorithm based on KL divergence to calculate the loss weight of each training individual; Step 4, smoothing the individual loss weight;

Step 5. Use weighted loss for training;

Step 6. Use the trained model to perform data prediction.

Further, the input of the graph convolutional neural network includes a feature matrix and graph structure information; wherein, the feature matrix includes a node feature matrix and an edge feature matrix, and the graph structure information is stored by an adjacency matrix, which represents the connection relationship between nodes;

The graph convolutional neural network mainly consists of three parts: feature transformation, information aggregation, and feature update.

Further, the graph convolutional neural network is a Graph Convolutional Networks (GCN) model:

Among them, W ^l represents the feature transformation matrix of layer l; H ^l represents the feature representation of the individual in layer l, and ^H0 is the original feature of the individual; D represents the degree matrix of nodes in the graph, representing the number of neighbors of each node; A represents the adjacency matrix of the graph It is to normalize the adjacency matrix; σ represents the activation function; by inputting the feature matrix of each node in the graph and the adjacency matrix representing the edge relationship between nodes into the graph convolutional neural network, the classification result of each node is finally obtained.

Furthermore, there are N individuals in the current graph data, and the model classifies each individual, and there are M risk categories in total. The cross-entropy loss calculated by the label is used to optimize the model parameters. The calculation formula is as follows:

Among them, N is the number of training individuals, L _i is the loss corresponding to individual i, p _ic represents the predicted probability of the model for individual i in category c, y _ic is the corresponding label, and the final total loss L is obtained through this formula.

Further, according to the obtained loss value, the distribution robust optimization algorithm based on KL divergence is used to calculate the current loss weight of all individuals, and the loss weight distribution is the worst distribution that the current algorithm can find.

Further, the optimization objective is as follows:

Among them, θ is the parameter of the graphical model, Θ represents the possible parameter set of the model, p represents the characteristic distribution of possible individual data in the future, and P represents the uncertain set composed of all possible distributions, and its uncertain range is determined by Sure. D(P,Q) represents the distance function between distributions P and Q. Taking KL divergence as an example, when D(P,Q) represents the KL divergence between P and Q, D(P,Q) can be expressed as /> f(X, θ) represents the prediction result of the model for each individual, and L(f(X, θ), y) represents the loss value corresponding to each individual, where L uses the cross entropy loss function. (X, y) ~ p means that the sample (X, y) is randomly sampled from the data of distribution p. X represents the information of the individual, including the characteristic information of the individual and the structure information of the subgraph, and the structure information of the subgraph contains the association information between the current individual and other individuals. /> is to calculate the expectation of the individual loss L under the distribution p. /> Represents the data distribution of the training set, through the constraint of the distance threshold ρ, the worst distribution is limited to As the center, within the range of a sphere with radius ρ.

Further, the distance function is Wasserstein distance or KL divergence.

Further, the individual loss weights obtained in step 3 are smoothed, so that individuals adjacent or with similar local neighbor structures can have closer weight values, thereby introducing graph structure information.

Further, the smoothing operation is interpolation smoothing. The smoothing operation is as follows,

Among them, W _s ⁿ represents the weight after smoothing n times, and W _s ⁰ is the initial weight calculated in step 3, Represents the smoothing matrix, 1 represents the importance of the initial weight, using λ can make the smoothed weight retain part of the weight information obtained in step 3.

A system for implementing the described distribution robust optimization method suitable for out-of-distribution generalization of graph neural networks, comprising:

Operation module: Input the graph data into the graph convolutional neural network to obtain the prediction of the psychological risk level of each individual, and use the label to calculate the corresponding loss value.

Weight calculation module: use KL divergence-based distribution robust optimization algorithm to calculate the loss weight of each individual.

Weight smoothing module: smooth the obtained loss weights, so that individuals with adjacent or similar local neighbor structures can have closer weight values.

Training module: use the smoothed weights and the loss of each individual to calculate the overall weighted loss, and train the graph convolutional neural network.

Execution module: When the psychological questionnaire survey results of students in a certain area are obtained, it is constructed into graph data and input into the trained model to predict the risk of mental illness.

Compared with the existing technology, the beneficial effect of the present invention is: the performance of the present invention on data with distribution drift is higher than that of the currently commonly used out-of-distribution generalization method; at the same time, after the traditional distribution robust optimization method is combined with the weight smoothing method proposed by the present invention, the model performance has been significantly improved, which also proves the effectiveness of the present invention.

Description of drawings

Fig. 1 is an individual similar psychological relationship diagram of an embodiment of the present invention;

Fig. 2 is the training block diagram of the embodiment of the present invention;

FIG. 3 is a flowchart of a distribution robust optimization method suitable for out-of-distribution generalization of a graph neural network according to an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. Apparently, the described embodiments are only some of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In this embodiment, a distribution-robust-optimized mental illness detection method suitable for the out-of-distribution generalization of graph neural networks takes all college students of a certain university as detection objects, each college student is an individual, and the psychological questionnaire survey done by the students is used as their characteristics. In order to mine more characteristic information, a student similar psychological relationship graph is constructed by using the student questionnaire results, so that there is an edge between two students with similar psychological conditions. In general, the graph model is used to assess the individual risk of an individual, and the characteristic information of the individual and the corresponding relationship subgraph are input into the graph model to predict its mental illness, or the entire graph data is fully input. At this time, the graph model only uses a simple cross-entropy loss for optimization training. When using a school's student data to train a classifier and want to extend this classifier to other regions or schools, the results will often be poor. This is because in different regions or different schools, the psychological state of students and the distribution of characteristics of the questionnaire are not the same. The above-mentioned differences in distribution are mainly caused by environmental factors. For example, different schools have different curriculum and learning pressures. Students in some schools have higher academic pressures, while others have lower pressures; different regions have different pressures on living standards. The above factors will lead to different distributions of mental health status of students in different regions and schools. This also leads to the fact that the data distribution faced by the model during inference is very different from the distribution during actual training, so simply using the cross-entropy loss does not make the model have a good generalization ability. The distribution robust optimization algorithm based on KL divergence is a commonly used algorithm to improve the generalization ability of the model. It mainly assigns higher weights to individuals with large losses during training, thereby increasing their proportion in the total loss, so that the model can pay more attention to these individuals. However, it is unreasonable to simply introduce this method into the training process of the graph model. For example, in the similar mental relationship graph data of students, an individual will have relationships with other individuals besides itself, so the weight of the individual training loss should also be similar to the weight of adjacent individuals. After the weight of each individual is calculated by the distribution robust optimization algorithm, the present invention further designs a weight smoothing operation, so that the smoothed weight not only considers the size of the individual loss, but also considers the connection relationship between individuals.

As shown in Figure 2, the distribution robust optimization method suitable for the generalization of the distribution of the graph neural network is provided to provide an optimization method for alleviating the low accuracy and untrustworthiness of the graph neural network in the face of out-of-distribution data, including the following steps:

Step 1. In the detection of mental illness, the student’s similar psychological relationship graph is used as input, and the graph neural network is used to obtain the psychological risk prediction output of each student in the graph. It mainly completes the aggregation of each student’s neighbor information, feature transformation, and updates the target student’s own characteristics, so as to fully consider the structural information of the graph. Among them, each node in the student similarity psychological relationship graph represents a student, and each node has some characteristic information. These characteristics are information mined from the psychological questionnaires made by the students; at the same time, each edge in the graph represents the similarity relationship between nodes.

The input of the graph convolutional neural network model mainly includes feature matrix and graph structure information. The feature matrix includes node feature matrix and edge feature matrix. The graph structure information is mainly stored by the adjacency matrix, which represents the connection relationship between nodes. The working process of the graph convolutional neural network model is mainly composed of three parts: feature transformation, information aggregation, and feature update. The core is to update its own feature information by aggregating the neighbor features of the target node.

Taking GCN as a preferred embodiment, its calculation formula is shown in the following formula:

Among them, W ^l represents the feature transformation matrix of layer l; H ^l represents the feature matrix of individuals in layer l; D represents the degree matrix of nodes in the graph, representing the number of neighbors of each node; A represents the adjacency matrix of the graph;

It is to normalize the adjacency matrix; σ represents the activation function.

Taking the mental illness detection mentioned in this embodiment as an example, it is assumed that there are currently N individuals in total, and the number of individual features is C. The prediction process of graph convolutional neural network can be divided into the following three steps:

Feature transformation: (H ^l )W ^l , when l=0, H ⁰ represents the original feature of the individual, the size of the matrix dimension is (N,C), and the feature can be the information mined from the psychological questionnaire made by the students. The features are linearly transformed through a linear layer to obtain an updated feature matrix.

Information aggregation: A represents the adjacency matrix of the graph, and the matrix dimension size is (N, N). The element in the matrix is 1, which means that there is a high similarity relationship between the corresponding two individuals, and 0 means that there is no dissimilarity. D represents the degree matrix, and the dimension size of the matrix is (N, 1), and each value indicates how many individuals the corresponding individual has a high similarity with in the entire relationship graph. Taking individual a as an example, when individuals b, c, d, and e are very similar to them, there will be an edge between them. Through the current step, individual a can aggregate and obtain the feature information of individuals b, c, d, and e. After weighted summation, as a new feature of individual a, the obtained feature of individual a not only contains the original feature information, but also contains neighbor information, which improves the information content of the feature.

Feature update: Finally, an activation function is used to further update the features.

The first is to use a graph convolutional neural network model to perform category prediction on individuals in the training set. Through the graph convolutional neural network model, students are similar to each individual in the psychological relationship graph, that is, each node in the graph, and can aggregate the feature information of multi-level neighbor nodes to obtain the final prediction result.

Step 2. Calculate the loss of each individual.

Through step 1, the prediction output of the graph convolutional neural network model for each individual training set can be obtained. There are N individuals in the data in the current graph, and the model classifies the mental disease risk category of each individual. The label of the individual indicates the risk of psychological problems. There are M risk categories in total. Use the cross-entropy loss calculated by the label to optimize the model parameters. The calculation formula is as follows:

Among them, N is the number of training individuals, l _i is the corresponding loss of each individual, p _ic represents the predicted probability of the model for individual i in category c, y _ic is the corresponding label, and the final total loss L is obtained through this formula.

Step 3. Calculate the loss weight of each individual using the distribution robust optimization algorithm based on KL divergence.

For general machine learning classification tasks, by training a classification model under limited training data, only limited data can be used to estimate the real data distribution. However, due to the limitation of the amount of data, there is a certain difference between the observable training set distribution and the actual distribution, so that the model cannot learn the real data distribution, so the performance of the model will be affected by the training set distribution. How to reduce this effect, or further optimize the learned classifier, has become a research topic.

Distributionally Robust Optimization (Distributionally Robust Optimization, DRO) was proposed to alleviate the impact of model performance degradation caused by distribution drift. Distribution robust optimization focuses on the data distribution of the training set, and finds the worst data distribution for the current model near it, so that the corresponding overall loss is the largest, and then optimizes the model so that it can have better results under the worst distribution, thus improving the generalization ability and robustness of the model. Its optimization goal is as follows:

Among them, θ is the parameter of the graphical model, Θ represents the possible parameter set of the model, p represents the characteristic distribution of possible individual data in the future, and P represents the uncertain set composed of all possible distributions, and its uncertain range is determined by OK, D(P,Q) represents the distance function between distribution P and Q, taking KL divergence as an example, when D(P,Q) represents the KL divergence between P and Q, D(P,Q) can be expressed as /> f(X, θ) represents the prediction result of the model for each individual, and L(f(X, θ), y) represents the loss value corresponding to each individual, where L uses the cross entropy loss function. (X, y) ~ p means that the sample (X, y) is randomly sampled from the data of distribution p. X represents individual information, including individual feature information and subgraph structure information, and the subgraph structure information includes the similarity relationship between the current individual and other individuals. /> is to calculate the expectation of the individual loss L under the distribution p. /> Represents the data distribution of the training set, through the constraint of the distance threshold ρ, the worst distribution is limited to As the center, within the range of a sphere with radius ρ.

As another preferred embodiment, the distance function is, for example, Wasserstein distance, K1 distance and the like.

Through the above-mentioned minmax confrontation training framework, using the optimization algorithm, the loss weight coefficient of each individual under the current loss can be calculated.

Step 4. Smooth the loss weight of the individual obtained in step 3. By smoothing the loss weight of each individual obtained, the adjacent or similar local neighbor structure individuals can have closer weight values, thereby further introducing graph structure information.

After step 3, the loss weight corresponding to each individual can be obtained. However, the weight at this time only considers the individual loss value of each individual, and does not consider the structural information of the relationship graph. Due to the assortment of graph data in structure, the weights of training nodes with similar distances on the graph should also be similar. Therefore, the present invention proposes a weight smoothing module. After calculating the weight of each individual in step 3, a smoothing operation is performed on the weight value, so that the weights of similar training nodes on the graph are similar, thereby considering the unique structural information of graph data. The smoothing operation is as follows,

Among them, W _s ⁿ represents the weight after smoothing n times, and W _s ⁰ is the initial weight calculated in step 3, Represents the smoothing matrix, 1 represents the importance of the initial weight, using λ can make the smoothed weight retain part of the weight information obtained in step 3.

Through the improvement of this step, the distribution robust optimization algorithm can be more suitable for graph data and incorporate more information, thus improving the adaptability of the method on graph data.

Step 5: Train with weighted loss, use the smoothed loss weight and corresponding loss value to calculate the weighted sum of the losses of all individuals as the final total loss; after completing the above steps, use the gradient descent optimization algorithm to minimize the total prediction loss, and update and optimize the parameters of the graph convolutional neural network model.

Step 6. Use the trained model to predict the individual mental disease risk category.

A system for implementing the described distribution robust optimization method suitable for out-of-distribution generalization of graph neural networks, comprising:

Operation module: input the similar psychological relationship graph data of students into the graph convolutional neural network to obtain the prediction of each individual's psychological risk level, and use the labels to calculate the corresponding loss value.

Weight calculation module: use KL divergence-based distribution robust optimization algorithm to calculate the loss weight of each individual.

Weight smoothing module: smooth the obtained loss weights, so that individuals with adjacent or similar local neighbor structures can have closer weight values.

Training module: use the smoothed weights and the loss of each individual to calculate the overall weighted loss, and train the graph convolutional neural network.

Execution module: When the psychological questionnaire survey results of students in a certain area are obtained, it is constructed into graph data and input into the trained graph convolutional neural network for risk prediction of mental diseases.

The preferred embodiments of the invention disclosed above are only to help illustrate the invention. The preferred embodiments are not exhaustive in all detail, nor are the inventions limited to specific embodiments described. Obviously, many modifications and variations can be made based on the contents of this specification. This description selects and specifically describes these embodiments in order to better explain the principle and practical application of the present invention, so that those skilled in the art can well understand and utilize the present invention. The invention is to be limited only by the claims, along with their full scope and equivalents.

Claims (10)

1. The distribution robust optimization method suitable for the map neural network distribution generalization is characterized by comprising the following steps of:

step 1, inputting a training individual into a graph convolution neural network model, completing aggregation and feature transformation of neighbor features of a target individual, updating the features of the target individual, and finally obtaining prediction output;

step 2, calculating a loss value of each training individual;

step 3, calculating the loss weight of each training individual by using a distribution robust optimization algorithm based on KL divergence;

step 4, smoothing the loss weight of the individual;

step 5, training by using the weight loss;

and 6, carrying out data prediction by using the trained model.

2. The distributed robust optimization method suitable for the outward generalization of the graph neural network distribution according to claim 1, wherein the input of the graph convolution neural network comprises a feature matrix and graph structure information, wherein the feature matrix comprises a node feature matrix and an edge feature matrix, and the graph structure information is stored by an adjacent matrix and represents the connection relation between nodes;

the graph convolutional neural network is mainly composed of three parts: feature transformation, information aggregation, and feature updating.

3. The distributed robust optimization method for out-of-distribution generalization of a graph neural network according to claim 2, characterized in that the graph roll-up neural network is a Graph Convolutional Networks (GCN) model:

wherein W is ^l Representing a feature transformation matrix of a first layer; h ^l Representation of features of an individual at layer I, H ⁰ Is an original feature of an individual; d represents a degree matrix of the nodes in the graph and represents the number of neighbors of each node; a represents the adjacency matrix of the graph,then normalizing the adjacency matrix; sigma represents an activation function; the characteristic matrix of each node in the graph and the adjacency matrix representing the edge relation between the nodes are input into the graph convolution neural network, and finally the classification result of each node is obtained.

4. The distributed robust optimization method for the outward generalization of the graph neural network distribution according to claim 1, wherein the current graph data has N individuals, the model classifies each individual, and there are M risk categories, and the model parameters are optimized by using the cross entropy loss calculated by the labels, and the calculation formula is as follows:

wherein N is the number of training individuals, L _i For loss corresponding to individual i, p _ic Representing the model's predicted probability of an individual i over category c, y _ic Then it is the corresponding label from which the final total loss L is derived.

5. The distribution robust optimization method suitable for the distribution generalization of the graph neural network according to claim 1, wherein a distribution robust optimization algorithm based on KL divergence is used to calculate the loss weight of all current individuals according to the obtained loss value, and the loss weight distribution is the worst distribution found by the current algorithm.

6. The distribution robust optimization method suitable for off-distribution generalization of a graph neural network according to claim 1, wherein the optimization objective is as follows:

where θ is a parameter of the graph model, θ represents a set of possible parameters of the model, p represents a characteristic distribution of the future possible individual data, and p represents an uncertainty set of all possible distributions, whose uncertainty range is defined byDetermining that D (P, Q) represents a distance function between the distributions P and Q, taking KL divergence as an example, when D (P, Q) represents KL divergence between P and Q, D (P, Q) is expressed as +.> f (X, θ) represents the model's predicted outcome for each individual,

l (f (X, theta)) represents a loss value corresponding to each individual, wherein L uses a cross entropy loss function, samples (X, y) to p are randomly sampled from data of distribution p, X represents individual information and comprises individual characteristic information and sub-graph structure information, the sub-graph structure information comprises association information between the current individual and other individuals,it is the expectation of calculating the individual loss L under distribution p,/->Data distribution representing training set, the worst distribution is limited to be +.>Is centered and has a radius rho in the sphere range.

7. The distribution robust optimization method for extradistribution generalization of a graph neural network according to claim 6, wherein the distance function is a wasperstein distance or KL divergence.

8. The distribution robust optimization method suitable for the outward generalization of the graph neural network distribution according to claim 1, wherein the individual loss weights obtained in the step 3 are smoothed, so that individuals with adjacent or similar local neighbor structures can have more similar weight values, and graph structure information is introduced.

9. The method for distribution robustness optimization for use in extradistribution generalization of a graph neural network according to claim 8, wherein the smoothing operation is interpolation smoothing, the smoothing operation is as follows,

wherein W is _s ⁿ Represents the weight after smoothing n times, W _s ⁰ For the initial weights calculated in step 3,representing a smoothing matrix, wherein 1 represents the importance degree of the initial weight, and lambda is utilized to enable the smoothed weight to retain part of the weight information obtained in the step 3.

10. A system for implementing a distributed robust optimization method for off-distribution generalization of a graph neural network as recited in claim 1, comprising:

the operation module: inputting the graph data into a graph convolution neural network, obtaining psychological risk level prediction of each individual, and calculating a corresponding loss value by using a label;

and a weight calculation module: calculating the loss weight of each individual by using a distribution robust optimization algorithm based on KL divergence;

and a weight smoothing module: smoothing the obtained loss weight to enable individuals with adjacent or similar local neighbor structures to have more similar weight values;

training module: calculating the total weighted loss by using the smoothed weight and the loss of each individual, and training the graph convolution neural network;

the execution module: when the psychological questionnaire survey results of students in a certain area are obtained, the psychological questionnaire survey results are constructed into graph data, and the graph data are input into a trained model to conduct psychological disease risk prediction.