Graph neural network architecture based on topology and semantic guidance

Abstract

The invention relates to the technical field of graph neural network architecture, in particular to a graph neural network architecture based on topology and semantic guidance. The system comprises a graph topology mining module S1 used for extracting topological features of a graph task; the graph semantic mining module S2 is used for extracting semantic features of the graph tasks; a feature extraction module S3, wherein the feature extraction module uses a full-connection network as a feature extraction operator; and a selection module and a crossover and variation module. According to the technical scheme, the problems that feature degradation is caused by excessive iteration of the graph neural network information extraction layer, overlapping and insufficiency exist in the aspect of graph mining, and mutual complementation is difficult are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of graph neural network architecture, and specifically to a graph neural network architecture based on topology and semantics guidance. Background Art

[0002] Graph neural networks (GNNs) represent a major breakthrough in the application of deep learning in non-Euclidean data. They are widely used in tasks such as social networks, biological systems, and transportation networks. Graph neural networks aggregate information of neighboring nodes based on graph topology to promote graph feature extraction and generate embeddings for downstream tasks. There are many practical applications of graph tasks, and the scale of graph structures is also very large. Graph neural networks with fixed structures often find it difficult to achieve optimal performance in different tasks. If we continue to rely on manual trial and error to select the best graph neural network architecture suitable for the target task, it will inevitably consume a lot of human resources.

[0003] In this context, the application of automated machine learning (AutoML) on GNN has attracted more and more attention. Automated machine learning aims to automatically select and combine existing GNN architectures for given graph tasks, and design targeted composite GNN models, thereby reducing the demand for human resources and enhancing the adaptability of the model; the application of automated machine learning on GNN mainly focuses on two aspects: hyperparameter optimization (HPO) and graph neural network architecture search (GNAS). Candidate GNN models are usually generated by sampling methods and then optimized by reinforcement learning or evolutionary algorithms. In terms of HPO, the hyperparameters required for GNN are the same as those of traditional deep learning, including learning rate, decay rate, number of convolution layers, number of neurons, etc. Regarding GNAS, due to the particularity of different types of graph tasks, various GNNs are usually selected as search units, and corresponding algorithms are designed to select graph convolution architectures.

[0004] Existing GNAS methods still face some challenges. On the one hand, despite the numerous existing graph neural network models, it is still very difficult to define a reasonable and effective search space. Simply using basic convolutional modules such as GCN, GAT, and GraphSAGE cannot produce ideal and explainable results. The GNAS process still lacks a deep understanding of the working principles of GNN. On the other hand, the existing architecture search object is a complete graph neural network model, covering the information aggregation layer and the feature extraction layer. When multiple iterations are performed to obtain high-order neighbor information, feature degradation problems may occur due to too many feature extraction processes, which limits the perception range of the searched graph neural network architecture.

[0005] Currently, the search space of most graph neural network architectures is built based on existing GNN models. For example, CTFGNAS uses GCN, GAT, and GraphSAGE to build the search space, and GCN-GA uses JKNet, ResGCN, and IncepGCN to build the search space. However, existing GNN models are designed independently, and there are overlaps or deficiencies in graph mining, making it difficult for them to complement each other. Using these models to build the search space results in inflexible architecture search and lack of interpretability, which greatly limits the expressive power of the model. Summary of the invention

[0006] In order to solve the problems in the related art, the present invention provides a graph neural network architecture based on topology and semantic guidance. The present invention solves the problem of feature degradation caused by too many iterations of the information extraction layer of the graph neural network, and the problems of overlap, deficiency and difficulty in complementing each other in graph mining.

[0007] In order to solve the above problems, the following technical solutions are provided: The topology- and semantic-guided graph neural network architecture of the present invention includes: Graph topology mining module S1 is used to extract the topological features of graph tasks. The initial state of the graph topology mining module is a first-order neighborhood topology mining operator. When K graph topology mining modules are selected continuously, the graph topology mining module extends the information aggregation process based on topological features to the K-order neighborhood topology feature mining module S1*(K); Graph semantic mining module S2 is used to extract the semantic features of graph tasks. The initial state of the graph semantic mining module is a first-order neighborhood semantic mining operator. When K graph semantic mining modules are selected continuously, the graph semantic mining module expands the information aggregation process based on semantic features into a K-order neighborhood semantic feature mining module S2*(K); A feature extraction module S3, wherein the feature extraction module uses a fully connected network as a feature extraction operator; The selection module sets the network length parameter n. The selection module randomly selects n modules from the graph topology mining module, graph semantic mining module and feature extraction module for sequence combination to form an initial model. The crossover and mutation module exchanges the initial models selected by the selection module, and randomly deletes or replaces the graph topology mining module, graph semantic mining module and feature extraction module to select the final model, which is the architecture of the graph neural network.

[0008] In the above scheme, through the setting of graph topology mining module, graph semantic mining module and feature extraction module, the graph topology mining module and graph semantic mining module are used to extract the topological and semantic features of graph tasks. The adaptive combination of graph topology mining module and graph semantic mining module can comprehensively and flexibly mine graph features, can be adjusted according to the target graph task, and effectively solves the problems of too many candidate modules and rigid module selection; when K graph topology mining modules are selected continuously, it is adaptively expanded into K-order neighborhood topology feature mining module; when K graph semantic mining modules are selected continuously, it is adaptively expanded into K-order neighborhood semantic feature mining module, which can adaptively iterate the information aggregation layer according to the task, and perform feature extraction after meeting the information requirements, thereby reducing the number of feature extraction layers, thereby solving the problem of feature degradation caused by excessive iterations of the information extraction layer of the graph neural network.

[0009] The calculation process of the first-order neighborhood topology mining operator in the graph topology mining module S1 is as follows: , in, and represents learnable static weights, which are used to weight themselves and the first-order neighbor set respectively; represents the input of layer i, represents the output of layer i, and They are all structures of a certain layer in the graph neural network architecture; Representation Node The first-order neighbor set of .

[0010] The K-order neighborhood topology feature mining module S1*(K) in the graph topology mining module S1 has an information aggregation process based on topology features as follows: , in, Used to weight the K-order neighbor set; Representation Node The K-order neighbor set of ; Represents a slave node To Node The set of nodes passed through.

[0011] The graph semantic mining module S2 first calculates the semantic associations between nodes through the attention mechanism, and then realizes information aggregation based on the semantic associations. The calculation process of the first-order neighborhood semantic mining operator is: , , in, is a learnable weight vector, W is a learnable weight matrix; the LeakyReLU function is used to introduce nonlinearity in the semantic-based propagation weight calculation; the Softmax function is used to facilitate the comparison of the propagation weights of adjacent nodes and ensure that the target node The sum of the semantic contributions of all neighboring nodes is 1.

[0012] The K-order neighborhood semantic feature mining module S2*(K) in the graph semantic mining module S2 has an information aggregation process based on semantic features as follows: , , , in, is the learnable weight vector and W is the learnable weight matrix.

[0013] The calculation process of the feature extraction operator in the feature extraction module S3 is: , Among them, W is a learnable weight matrix, and the Relu function is used to perform nonlinear transformation on the extracted features.

[0014] The crossover and mutation module is divided into a crossover module and a mutation module. The crossover module is used to exchange the positions of the graph topology mining module, the graph semantic mining module and the feature extraction module to form a new model.

[0015] The mutation module randomly selects any one of the graph topology mining module, graph semantic mining module and feature extraction module in the initial model, replaces it with a different one of the graph topology mining module, graph semantic mining module and feature extraction module, and then randomly replaces the positions of the graph topology mining module, graph semantic mining module and feature extraction module.

[0016] The above solution has the following advantages: The present invention sets a graph topology mining module, a graph semantic mining module and a feature extraction module. The graph topology mining module and the graph semantic mining module are used to extract the topological and semantic features of graph tasks. The adaptive combination of the graph topology mining module and the graph semantic mining module can comprehensively and flexibly mine graph features, can be adjusted according to the target graph task, and effectively solves the problems of too many candidate modules and rigid module selection; when K graph topology mining modules are selected continuously, they are adaptively expanded into a K-order neighborhood topology feature mining module; when K graph semantic mining modules are selected continuously, they are adaptively expanded into a K-order neighborhood semantic feature mining module, which can adaptively iterate the information aggregation layer according to the task, and perform feature extraction after meeting the information requirements, thereby reducing the number of feature extraction layers, thereby solving the problem of feature degradation caused by excessive iterations of the graph neural network information extraction layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein: Figure 1 It is an information aggregation process diagram of a graph topology mining module in the present invention; Figure 2 It is an information aggregation process diagram of stacking K graph topology mining modules in the present invention; Figure 3 It is an information aggregation process diagram of a graph semantic mining module in the present invention; Figure 4 This is a diagram of the information aggregation process of stacking K graph semantic mining modules in the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The topology- and semantic-guided graph neural network architecture of the present invention includes: like Figure 1 and Figure 2 As shown, the graph topology mining module S1 is used to extract the topological features of the graph task. The initial state of the graph topology mining module is a first-order neighborhood topology mining operator, and the calculation process is: , in, and represents learnable static weights, which are used to weight themselves and the first-order neighbor set respectively; represents the input of layer i, represents the output of layer i, and They are all structures of a certain layer in the graph neural network architecture; Representation Node The first-order neighbor set of ; When K graph topology mining modules are selected continuously, the graph topology mining module will be adaptively expanded into a K-order neighborhood topology feature mining module S1*(K). The information aggregation process based on topology features is: , in, Used to weight the K-order neighbor set; Representation Node The K-order neighbor set of ; Represents a slave node To Node The set of nodes passed through.

[0020] like Figure 3 and Figure 4 As shown in the figure, the graph semantic mining module S2 is used to extract the semantic features of the graph task. The initial state of the graph semantic mining module is the first-order neighborhood semantic mining operator. The graph semantic mining module first calculates the semantic association between nodes through the attention mechanism, and then realizes information aggregation based on the semantic association. The specific calculation process is: , , in, is a learnable weight vector, W is a learnable weight matrix; the LeakyReLU function is used to introduce nonlinearity in the semantic-based propagation weight calculation; the Softmax function is used to facilitate the comparison of the propagation weights of adjacent nodes and ensure that the target node The sum of the semantic contributions of all adjacent nodes is 1; When K graph semantic mining modules are selected continuously, the graph semantic mining module will be adaptively expanded into a K-order neighborhood semantic feature mining module S2*(K). The information aggregation process based on semantic features is: , , , in, is the learnable weight vector and W is the learnable weight matrix.

[0021] Feature extraction module S3: The feature extraction module uses a fully connected network as a feature extraction operator, and its calculation process is: , Among them, W is a learnable weight matrix, and the Relu function is used to perform nonlinear transformation on the extracted features.

[0022] The selection module sets the network length parameter n. The selection module randomly selects n modules from the graph topology mining module, graph semantic mining module and feature extraction module for sequence combination to form an initial model. Set the network length parameters .

[0023] The crossover and mutation module exchanges the initial models selected by the selection module through the crossover and mutation module, and randomly deletes or replaces the graph topology mining module, the graph semantic mining module and the feature extraction module to select the final model, which is the architecture of the graph neural network; the crossover and mutation module is divided into a crossover module and a mutation module. The crossover module is used to exchange the positions of the graph topology mining module, the graph semantic mining module and the feature extraction module to form a new model; the mutation module randomly selects any one of the graph topology mining module, the graph semantic mining module and the feature extraction module in the initial model, replaces it with a different one of the graph topology mining module, the graph semantic mining module and the feature extraction module, and then randomly replaces the positions of the graph topology mining module, the graph semantic mining module and the feature extraction module.

[0024] In Example 1, when the network length parameter n is set to 6, six architectures for constructing a graph neural network need to be selected from the graph topology mining module, the graph semantic mining module, and the feature extraction module. When the selected model sequence is "S1-S1-S2-S3-S2-S2", the repeated units are first merged into one module in order, and the model sequence becomes "S1*(2)-S2-S3-S2*(2)", corresponding to the graph topology mining module, the graph semantic mining module, and the feature extraction module. The architecture of the initial model is: "second-order neighborhood topology information mining, first-order neighborhood semantic feature mining, feature extraction, second-order neighborhood semantic feature mining"; The algorithm process of the crossover module, at the beginning of each crossover, two modules are randomly selected for crossover. For the crossover of hyperparameters, a random value is first generated. If the generated random number is less than the crossover probability, the crossover operation of the hyperparameter is performed, otherwise the initial sample is retained; when performing the crossover of hyperparameters, two random crossover points are set in the parameter sequence of the initial sample, and then the two modules are exchanged. For the crossover of the network architecture, considering the different lengths of each individual architecture, first determine whether the architecture needs to be crossed by a random value; if the architecture needs to be crossed, select the crossover points of any length in the variable-length architecture sequence of each module, and then exchange between their crossover points to obtain two variable-length architecture sequences until the final architecture sequence is generated, which is then saved.

[0025] The algorithm process of the mutation module, for the mutation of hyperparameters, only consider changing the module; for the mutation of the architecture, select one of the graph neural network architecture sequence and change it to another one of the graph topology mining module, graph semantic mining module and feature extraction module; randomly select a layer of the graph neural network architecture and randomly insert a module; select modules at any two positions in the architecture for exchange.

[0026] Use the cross entropy loss function to calculate the individual prediction results With the true label The gap between them is used as the criterion for individual evaluation. The process of calculating the cross entropy loss is: , The lower it is, the more adaptable the individual is to the target graph task.

[0027] This paper evaluates the performance of a topology- and semantics-guided graph neural network architecture model (TSGNAS) by comparing it with mainstream hand-designed GNN and AutoML methods in a node classification task. The experimental setup is first described, including the dataset, baseline models for comparison, and implementation details. The experimental results on different real datasets are then presented and analyzed. All experiments are conducted on a machine equipped with an NVIDIA GeForce RTX 4090 (24G RAM), a 13th generation Intel(R) Core(TM) i9-13900K CPU (3.00 GHz), and 32G RAM.

[0028] In order to comprehensively evaluate the adaptability of the model in different scenarios, node classification experiments were conducted in three homogeneous datasets and three heterogeneous datasets, covering four task scenarios: citation network, Amazon shopping network, actor co-occurrence network, and WebKB web link network. In homogeneous datasets, connected nodes tend to share the same labels, including: (1) citation network: Citeseer and Pubmed; (2) Amazon shopping network: Photo; in heterogeneous datasets, connected nodes tend to have different labels, including: (1) actor co-occurrence network: Actor; (2) WebKB web link network: Cornell and Texas. Table 1 shows the statistical information of these two types of datasets. H(G) is an indicator to measure the homogeneity of the dataset. The larger the H(G) value, the stronger the homogeneity.

[0029] Table 1 Datasets used in node classification experiments

[0030] To demonstrate the effectiveness of TSGNAS, it is compared with five representative hand-designed GNNs and four AutoML methods. The hand-designed GNNs include: GCN, GAT, DirGNN, A-DGNS, AF-GCN and M2M-GNN; the AutoML methods include: Genetic-GNN, F2GNN, GCN-GA and CTFGNA.

[0031] In the hyperparameter search space, the following five types of hyperparameters are mainly optimized. The number of model layers is introduced as one of the hyperparameters to enable TSGNAS to achieve task-specific expansion. In addition, TSGNAS is a general framework that can introduce other important parameters as needed.

[0032] Learning rate (P1): The learning rate determines the update amplitude of the parameters when training the model based on the gradient descent algorithm. A larger learning rate may cause the model to diverge, while a smaller learning rate may cause the optimization to converge slowly.

[0033] Weight decay (P2): Weight decay introduces additional penalties to model parameters, pushing them toward smaller values. This strategy reduces the complexity of the model and improves its generalization ability.

[0034] Dropout (P3): The purpose of Dropout is to prevent overfitting of the neural network. In each training iteration, Dropout randomly discards some neurons in the network with a certain probability, thereby improving the robustness of the network.

[0035] Model layers (P4): By increasing the number of neural network layers, the representation of data is gradually mapped from the original space to a more abstract and advanced representation space, enabling the network to learn more complex features and patterns. However, this also brings challenges such as gradient disappearance and gradient explosion.

[0036] Number of neurons (P5): Neurons are the basic units of a neural network layer, responsible for processing input data and generating outputs. More neurons enhance the learning ability of the network, enabling it to better fit the training data and improve the generalization ability of the network. However, increasing the number of neurons also increases the computational complexity of the network and brings the risk of overfitting.

[0037] Each layer of the graph neural network can apply different P1, P2, P3, and P5, but this will extend the complexity of the hyperparameter search to , which greatly increases the complexity of the model and leads to overfitting problems. Therefore, we set these parameters to be globally shared, reducing the complexity of the hyperparameter search to .

[0038] Table 2 Hyperparameter search space

[0039] All comparison models are reproduced based on the Pytorch framework, and the optimal parameters are selected. For the manually designed GNN, the hyperparameters are set to 500 epochs, the hidden layer size is 64, the learning rate is selected from {0.001, 0.002, 0.01,0.03, 0.1}, the weight decay is selected from {0, 0.0005, 0.00005}, the dropout is selected from {0.0, 0.5, 0.6, 0.7, 0.8, 0.9}, and the early stopping range is [10, 200]; for the AutoML method, the model is first run to obtain the optimal hyperparameters and model architecture, and then the GNN model is generated based on the optimal hyperparameters and model architecture for comparison experiments.

[0040] A fully supervised approach was used to learn and compare all models. The dataset was randomly divided into training, test, and validation sets with proportions of 60%, 20%, and 20%, respectively. To ensure the fairness of the experiment, 10 dataset partitioning seeds were randomly generated, and all models were evaluated under the same dataset partitioning. After calculating the classification accuracy of the 10 partitions, a bootstrap algorithm was used to generate a normal distribution of the model classification accuracy. The number of random samples was 1000, and the average accuracy (%) of the normal distribution ±95% confidence interval was used as the evaluation indicator to evaluate the accuracy and error bar of the model.

[0041] Following the above experimental settings, the TSGNAS model is compared with nine baseline models on the node classification dataset. Table 3 summarizes the relevant results, where boldfaced words represent the best results on the dataset.

[0042] Table 3 Performance study of node classification benchmark dataset

[0043] Whether comparing handwritten GNNs or AutoML-based GNNs, TSGNAS achieved the best results on all six datasets. Compared with the second-ranked AF-GCN model, TSGNAS improved by nearly 4.5% on the Actor dataset. Traditional graph neural networks (such as GCN and GAT) are not very suitable for heterophilic datasets. In subsequent development, handwritten GNNs make the model perform well on heterophilic datasets by constructing graph distant connections in the spatial domain (such as A-DGN) or introducing high-frequency information in the spectral domain (such as AF-GCN). As for other AutoML methods, their basic components are simple network models, such as GCN, GAT, etc. These substructures perform poorly on heterophilic datasets, and the combined GNNs also perform poorly. TSGNAS introduces scalable topology and semantic modules, allowing nodes to adaptively adjust their dependence on different neighbors according to task requirements, thereby achieving ideal results on different types of datasets.

[0044] To comprehensively evaluate the efficiency of TSGNAS, we record the number of parameters and running time of the models generated by TSGNAS and the baseline models, as shown in Table 4.

[0045] Table 4. Efficiency study of node classification benchmark dataset

[0046] In general, the time and space cost of AutoML methods is higher than that of hand-designed GNNs. Hand-designed GNNs are already optimized models that can achieve good performance with only a small number of layers; while AutoML models need to continuously stack network layers to adapt to the task, which leads to more parameters and running time. Genetic-GNN changes the feature extraction layer to operations such as summation, maximum value and mean, which greatly reduces the number of linear transformations, so it takes up the least storage and time consumption in the AutoML method; TSGNAS does not advocate over-stacking of feature extraction layers, so its number of parameters is the second smallest compared to other AutoML methods.

[0047] In summary, the compared AutoML methods are inferior to the most advanced manually designed GNN in terms of performance and application efficiency. In fact, a model composed of stacked modules will be limited by the learning ability of the modules themselves, and may not always achieve the expected results under normal circumstances; the advantage of AutoML is that it does not require manual model design, and only requires the input of reasonable modules to generate high-quality models, which greatly reduces the threshold for model design and application.

[0048] The architectures and hyperparameters automatically searched by TSGNAS on different datasets are intuitively displayed, as shown in Table 5.

[0049] Table 5. TSGNAS searches for GNN architecture for each dataset

[0050] Among the best architectures searched, there are both single GTM modules and combined GSM modules, which shows that the architecture search space designed by the present invention is practical; the search structure lengths of different datasets are different. Compared with fixed layer models, TSGNAS is more flexible in terms of perception range; in addition, each dataset can arbitrarily select the required information aggregation module according to the task without forcing mixed use. For example, the Citeseer dataset only uses GTM, while the Texas and Cornell datasets only use GSM. In short, TSGNAS is flexible and task adaptable.

[0051] Obviously, the above embodiments are merely examples for clear explanation and are not limitations on the implementation methods. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made on the basis of the above description. It is not necessary and impossible to list all the implementation methods here, and the obvious changes or modifications derived therefrom are still within the protection scope of the invention.

Claims

1. A graph neural network architecture based on topology and semantics guidance, characterized by: include: Graph topology mining module S1 is used to extract the topological features of graph tasks. The initial state of the graph topology mining module is a first-order neighborhood topology mining operator. When K graph topology mining modules are selected continuously, the graph topology mining module extends the information aggregation process based on topological features to the K-order neighborhood topology feature mining module S1*(K); Graph semantic mining module S2 is used to extract the semantic features of graph tasks. The initial state of the graph semantic mining module is a first-order neighborhood semantic mining operator. When K graph semantic mining modules are selected continuously, the graph semantic mining module expands the information aggregation process based on semantic features into a K-order neighborhood semantic feature mining module S2*(K); A feature extraction module S3, wherein the feature extraction module uses a fully connected network as a feature extraction operator; The selection module sets the network length parameter n. The selection module randomly selects n modules from the graph topology mining module, graph semantic mining module and feature extraction module for sequence combination to form an initial model. The crossover and mutation module exchanges the initial models selected by the selection module, and randomly deletes or replaces the graph topology mining module, graph semantic mining module and feature extraction module to select the final model, which is the architecture of the graph neural network.

2. The topology- and semantics-guided graph neural network architecture according to claim 1, characterized in that: The calculation process of the first-order neighborhood topology mining operator in the graph topology mining module S1 is as follows: ， in, and represents learnable static weights, which are used to weight themselves and the first-order neighbor set respectively; represents the input of layer i, represents the output of layer i, and They are all structures of a certain layer in the graph neural network architecture; Representation Node The first-order neighbor set of .

3. The topology- and semantics-guided graph neural network architecture according to claim 2, characterized in that: The K-order neighborhood topology feature mining module S1*(K) in the graph topology mining module S1 has an information aggregation process based on topology features as follows: ， in, Used to weight the K-order neighbor set; Representation Node The K-order neighbor set of ; Represents a slave node To Node The set of nodes passed through.

4. The topology- and semantic-guided graph neural network architecture according to claim 1, characterized in that: The graph semantic mining module S2 first calculates the semantic associations between nodes through the attention mechanism, and then realizes information aggregation based on the semantic associations. The calculation process of the first-order neighborhood semantic mining operator is: ， ， in, is a learnable weight vector, W is a learnable weight matrix; the LeakyReLU function is used to introduce nonlinearity in the semantic-based propagation weight calculation; the Softmax function is used to facilitate the comparison of the propagation weights of adjacent nodes and ensure that the target node The sum of the semantic contributions of all neighboring nodes is 1.

5. The topology- and semantic-guided graph neural network architecture according to claim 4, characterized in that: The K-order neighborhood semantic feature mining module S2*(K) in the graph semantic mining module S2 has an information aggregation process based on semantic features as follows: ， ， ， in, is the learnable weight vector and W is the learnable weight matrix.

6. The topology- and semantics-guided graph neural network architecture according to claim 1, characterized in that: The calculation process of the feature extraction operator in the feature extraction module S3 is: ， Among them, W is a learnable weight matrix, and the Relu function is used to perform nonlinear transformation on the extracted features.

7. The topology- and semantic-guided graph neural network architecture according to claim 1, characterized in that: The crossover and mutation module is divided into a crossover module and a mutation module. The crossover module is used to exchange the positions of the graph topology mining module, the graph semantic mining module and the feature extraction module to form a new model.

8. The topology- and semantics-guided graph neural network architecture according to claim 7, characterized in that: The mutation module randomly selects any one of the graph topology mining module, graph semantic mining module and feature extraction module in the initial model, replaces it with a different one of the graph topology mining module, graph semantic mining module and feature extraction module, and then randomly replaces the positions of the graph topology mining module, graph semantic mining module and feature extraction module.

Images