Node classification method and system based on hybrid expert model, medium and product

Abstract

The invention discloses a node classification method and system based on a hybrid expert model.The node classification method comprises the steps that node data are input into a pre-trained hybrid expert model to obtain a classification result of the node data, and the hybrid expert model comprises a node mode extractor, a gating model and a plurality of classification models; the node mode extractor is used for extracting node mode information from given node data, and the gating model is used for receiving the node mode information and generating a model combination weight corresponding to each node; and dynamically selecting and combining the classification results of the node data by the plurality of different types of classification models according to the model combination weights corresponding to the nodes so as to obtain the final classification result of the nodes. The invention aims to solve the problem of node classification under a complex graph structure, accurately define a node mode under the condition that the node mode is complex, and properly combine different node predictors for different nodes.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis of social network platforms, and in particular to a node classification method, system, medium and product based on a hybrid expert model. Background Art

[0002] User nodes in social networking platforms have diverse and complex attributes and associated features. In addition to sharing text, pictures, videos and other content, users also form complex relationship networks through interactive behaviors such as following, liking, commenting, and forwarding. As the scale of social network users continues to expand, how to efficiently classify nodes (such as users, content, etc.) in such large and dynamic graph data has become a key issue that needs to be solved urgently. For example, on dating websites or microblog platforms, labeling and classifying users according to their gender, interests or behavioral characteristics can better achieve friend recommendations and information filtering; in content communities, distinguishing between high-influence users (KOLs) and ordinary users helps to explore potential business opportunities; in security review scenarios, it is necessary to identify junk users or fake accounts to create a healthy and trustworthy social environment.

[0003] Graph Neural Networks (GNN) have significant advantages in learning representations of graph data, and can learn embedded representations of nodes by aggregating neighbor node features. However, most existing GNN models assume that connected nodes are homogeneous in attributes or labels. For a large number of heterogeneous associations in social networks (for example, users may follow groups with different interests or backgrounds), such methods are difficult to take into account the node classification needs of diverse patterns. In order to cope with the diversity and complexity of node patterns in social networks, academia and industry have successively proposed a series of strategies to enhance GNN models. For example, some methods establish node associations by integrating wide-area neighborhood information to solve the correlation problem between originally unconnected nodes; other methods try to introduce heterogeneous structure modeling at the node level, considering that different nodes in the same graph may have different attributes and relationships. However, these methods usually rely on fixed feature integration strategies and lack adaptive capabilities; and when dealing with multiple attributes and heterogeneous edges, the generalization and robustness of the model are still insufficient. In addition, a single model architecture often cannot simultaneously take into account the significant differences between different nodes, and shows certain limitations in classification tasks.

[0004] These shortcomings have promoted the development of hybrid expert models. The hybrid expert model introduces multiple different classifiers and is equipped with a gating network to adaptively assign the most appropriate classifier to each node according to the specific pattern of the node and the overall context of the graph. This method can not only more flexibly cope with the diversity of node patterns, but also improve the generalization ability and classification accuracy of the model on complex heterogeneous graphs. Therefore, the hybrid expert model provides a more general and efficient solution to the node classification problem on heterogeneous graphs. However, for the node classification problem under complex graph structures, the use of hybrid expert models for node classification still faces huge challenges: (1) How to accurately define node patterns when the node patterns are complex. Node patterns not only involve the attribute information of the node, but also its structural position in the graph and its relationship with neighboring nodes. These patterns are often multidimensional and highly complex, and multiple aspects such as label homogeneity, structural information and attribute information need to be considered comprehensively. Therefore, how to effectively capture and represent this multimodal information to provide an accurate basis for subsequent classification tasks has become a difficult problem to be solved. (2) How to appropriately combine different node predictors for different nodes. Since nodes in the same graph may have different patterns and features, a single predictor is difficult to adapt to the needs of all nodes. How to dynamically select or combine the most suitable predictors based on the specific pattern and context information of each node is the key to improving classification performance. Summary of the invention

[0005] Technical problem to be solved by the present invention: In view of the above-mentioned problems in the prior art, a node classification method, system, medium and product based on a hybrid expert model are provided. The present invention aims to solve the node classification problem under complex graph structures, accurately define node patterns when the node patterns are complex, and appropriately combine different node predictors for different nodes.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A node classification method based on a hybrid expert model comprises the following steps: inputting node data into a pre-trained hybrid expert model to obtain a classification result of the node data, wherein the hybrid expert model comprises a node pattern extractor, a gating model and a plurality of classification models of different types, wherein the node pattern extractor is used to extract node pattern information from given node data, the gating model is used to receive the node pattern information and generate a model combination weight corresponding to each node, and the classification results of the node data by the plurality of classification models of different types are dynamically selected and combined according to the model combination weight corresponding to the node to obtain a final classification result of the node.

[0007] Optionally, the node pattern extractor extracts a node pattern from given node data including: S101, for each node in the given node data , for nodes The neighborhood of the node is randomly sampled to obtain The local subgraph contains several nodes Nodes in the neighborhood; S102, for each node in the given node data , the node Any node in its local subgraph A node pair , the node pair Input into the edge discriminator composed of multi-layer perceptron to obtain the measurement node Any node in its local subgraph The edge discriminator scores the degree of association in the feature space; the nodes The number of connections to its local subgraph or to all local subgraphs as a node Node degree information of S103, for each node in the given node data , the node The edge discriminator score of the node is concatenated with the node degree information as Local node mode information; S104, aggregating the local node mode information of all nodes in a specified aggregation manner to obtain global node mode information, and outputting the local node mode information and the global node mode information as the finally obtained node mode information.

[0008] Optionally, the aggregation method specified in step S104 refers to average aggregation.

[0009] Optionally, the gating model receives the node pattern and generates a model combination weight corresponding to each node including: S201, using a first multi-layer perceptron to perform nonlinear mapping on local node pattern information of a node to obtain a local pattern representation, and using a second multi-layer perceptron to perform nonlinear mapping on global node pattern information to obtain a global pattern representation; S202, concatenating the local pattern representation and the global pattern representation of the node; S203, input the concatenated result into the third multi-layer perceptron, and then output a set of weight vectors of the node through the softmax function ,in ~ are the model combination weights of the M classification models corresponding to the node, and M is the number of classification models.

[0010] Optionally, the classification results of the node data by the multiple different types of classification models are dynamically selected and combined according to the model combination weights corresponding to the nodes to obtain the function expression of the final classification result of the node: , in, is the final classification result of the i-th node, It means summation, is the model combination weight of the j-th classification model corresponding to the i-th node, is the classification result of the node data of the i-th node by the j-th classification model.

[0011] Optionally, before the classification results of the node data by the multiple different types of classification models are dynamically selected and combined according to the model combination weights corresponding to the nodes to obtain the final classification results of the nodes, the step of constructing and training a hybrid expert model is also included: S301, select multiple different types of graph neural networks (GNNs) with K layers of learnable parameters, and use source data containing node-label pairs for separate pre-training, so that they have preliminary classification capabilities by minimizing the main task loss function; S302, incorporating the pre-trained graph neural network GNN as a classification model into the hybrid expert model, so that it is constructed together with the node pattern extractor and the gating model to obtain the hybrid expert model; S303, taking all nodes with true labels as supervisory signals, for each node, defining the loss function as the cross entropy between its prediction result and the true label, performing end-to-end training on the hybrid expert model to optimize and update the parameters of the classification model, the parameters of the edge discriminator in the node pattern extractor, and the parameters of the three multi-layer perceptrons in the gating model, and finally obtaining a hybrid expert model that has completed training.

[0012] Optionally, the node data is the data of a user node in a social networking platform, and the data of the user node includes the user's basic information, user sharing information and the user's interactive behavior, the user sharing information includes part or all of the text, pictures and videos shared by the user, and the user's interactive behavior includes part or all of the user's following, liking, commenting and forwarding.

[0013] In addition, the present invention also provides a node classification system based on a hybrid expert model, comprising a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the node classification method based on the hybrid expert model.

[0014] In addition, the present invention also provides a computer-readable storage medium, in which a computer program or instruction is stored. The computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor.

[0015] In addition, the present invention also provides a computer program product, including a computer program or an instruction, wherein the computer program or the instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor.

[0016] Compared with the prior art, the present invention mainly has the following advantages: 1. The present invention proposes a hybrid expert model for the node classification problem under complex graph structures. The model includes multiple diversified (different types) classification models and a specially designed node pattern extractor and a gating network. The node pattern extractor is used to extract node pattern information from given node data. The gating network can adaptively assign appropriate classifier weights to each node according to the node pattern information of the node and the context of the overall graph, so as to better cope with the multimodal patterns of nodes in complex graphs, thereby achieving accurate classification of different nodes.

[0017] 2. The present invention has the advantages of flexible expansion in both node pattern extraction and gating model structure. It can add or replace classification models according to actual needs, and further combine methods such as self-supervised training to effectively cope with complex and changeable graph node classification tasks. Through the organic connection of the above steps, this embodiment can demonstrate excellent performance and wide applicability in multiple practical application scenarios (including social analysis, recommendation systems, traffic prediction, etc.). BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the working principle of the hybrid expert model in an embodiment of the present invention.

[0019] Figure 2 Schematic diagram of the working process of the node pattern extractor in the embodiment of the present invention.

[0020] Figure 3 Schematic diagram of the working process of the gating model in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0022] Embodiment 1: like Figure 1As shown, the node classification method based on the hybrid expert model of this embodiment includes the following steps: inputting the node data into a pre-trained hybrid expert model (named MoE-NP in this embodiment) to obtain the classification result of the node data, the hybrid expert model includes a node pattern extractor (Node Pattern Extractor), a gating model (The Design of the Gating Model) and multiple different types of classification models, the node pattern extractor is used to extract node pattern information from given node data, the gating model is used to receive node pattern information and generate a model combination weight corresponding to each node, and the classification results of the node data of multiple different types of classification models are dynamically selected and combined according to the model combination weights corresponding to the nodes to obtain the final classification results of the nodes. Through the hybrid expert model, the multimodal patterns of nodes in complex graphs can be better coped with, thereby achieving accurate classification of different nodes.

[0023] In this embodiment, node data refers to user node data in a social networking platform, which includes not only basic information about users, but also texts, pictures or videos shared by users, as well as interactive behaviors such as following, liking, commenting and forwarding. In this scenario, user nodes can be divided into active users with frequent interactions, influential users with strong diffusion capabilities, potential users with less interactions but high potential value, and ordinary users with medium activity and influence based on dimensions such as user activity, scale and dissemination. These classification results can provide more refined decision support for content operations on social networking platforms.

[0024] The node pattern extractor of this embodiment is used to extract node pattern information from the multimodal information of a given node (for example, local structural context and node features). Since nodes in a graph often present multimodal features (such as degree distribution, label homogeneity, local structure, etc.), the goal of this step is to fully capture these multimodal features to support subsequent accurate classification. In this embodiment, each node and its neighborhood are first randomly sampled to objectively reflect the local structure of the node. Then, the node pattern is extracted based on the sampling results, which mainly includes the following contents: (1) Structural information: The structural pattern of the node is characterized by recording the degree distribution, connection relationship, etc. of the neighborhood nodes. (2) Feature information: An edge discriminator in the form of a multi-layer perceptron (MLP) is used to predict whether there is an association between the target node and its neighborhood nodes, so as to learn more effective node feature similarity without real labels. Finally, the degree information of the node is cascaded with the output of the above-mentioned edge discriminator to obtain a complete node pattern representation, which lays the foundation for subsequent gating model operations. Specifically, Figure 2 As shown, in this embodiment, the node pattern extractor extracts the node pattern from the given node data including: S101, for each node in the given node data , for nodes The neighborhood of the node is randomly sampled to obtain The local subgraph contains several nodes Nodes in the neighborhood; S102, for each node in the given node data , the node Any node in its local subgraph A node pair , the node pair Input into the edge discriminator composed of multi-layer perceptron to obtain the measurement node Any node in its local subgraph The edge discriminator scores the degree of association in the feature space; the nodes The number of connections to its local subgraph or to all local subgraphs as a node Node degree information of S103, for each node in the given node data , the node The edge discriminator score of the node is concatenated with the node degree information as Local node mode information; S104, aggregating the local node mode information of all nodes in a specified aggregation manner to obtain global node mode information, and outputting the local node mode information and the global node mode information as the finally obtained node mode information.

[0025] See also Figure 2 In this embodiment, the aggregation method specified in step S104 refers to average aggregation. Undoubtedly, other aggregation methods may be used as needed, such as clustering or machine learning methods.

[0026] The gating model of this embodiment is used to receive node patterns and generate expert weights corresponding to each node, so as to dynamically select and combine among multiple node classifiers (experts). After obtaining the local pattern of the node, the distribution information of the overall graph must also be considered to make more reasonable expert assignments to the nodes. Therefore, this step combines local and global patterns in the following ways: (1) Calculate the global pattern of the graph: average the local patterns of all nodes to obtain an overall distribution representation at the graph level. (2) Pattern fusion: Use a learnable MLP to map the local pattern and the global pattern to the same vector space, and cascade the two to obtain a fusion vector. (3) Gating network calculation: Execute a multi-layer perceptron (MLP) on the fusion vector and pass it through a softmax function to output an expert weight vector for each node, which is used to weightedly fuse the prediction results of all experts. In this way, while taking into account the local pattern of the node and the context of the overall graph, the gating network can dynamically assign the most appropriate classifier weight to each node. Specifically, Figure 3 As shown, in this embodiment, the gating model receives the node pattern and generates the model combination weight corresponding to each node, including: S201, using a first multi-layer perceptron to perform nonlinear mapping on local node pattern information of a node to obtain a local pattern representation, and using a second multi-layer perceptron to perform nonlinear mapping on global node pattern information to obtain a global pattern representation; S202, concatenating the local pattern representation and the global pattern representation of the node; S203, input the concatenated result into the third multi-layer perceptron, and then output a set of weight vectors of the node through the softmax function ,in ~ They are the model combination weights of the M classification models corresponding to the node, and M is the number of classification models. Since the combination weights of some models may be 0, it is equivalent to removing the classification results of the corresponding classification models, thereby achieving the purpose of dynamically selecting and combining the model combination weights corresponding to the node to obtain the final classification result of the node.

[0027] In this embodiment, the classification results of node data by multiple different types of classification models are dynamically selected and combined according to the model combination weights corresponding to the nodes to obtain the function expression of the final classification result of the node: , in, is the final classification result of the i-th node, It means summation, is the model combination weight of the j-th classification model corresponding to the i-th node, is the classification result of the node data of the i-th node by the j-th classification model. Through this dynamic weighted fusion strategy, each node can absorb the most discriminative features from different expert models based on its own local structure, local features and global distribution features, providing more accurate classification for different types of nodes.

[0028] In this embodiment, multiple different types of classification models can adopt graph neural network classification models such as GCN, GAT, GraphSAGE, etc. as needed. In addition, more classifiers can be flexibly introduced as needed to achieve better classification performance in diversified scenarios.

[0029] In this embodiment, before the classification results of node data by multiple different types of classification models are dynamically selected and combined according to the model combination weights corresponding to the nodes to obtain the final classification results of the nodes, the steps of constructing and training a hybrid expert model are also included: S301, select multiple different types of graph neural networks (GNNs) with K layers of learnable parameters, and use source data containing node-label pairs for separate pre-training, so that they have preliminary classification capabilities by minimizing the main task loss function; then use these optimized pre-training parameters (including gated model parameters and expert model parameters) as the starting point for the subsequent adaptation stage, improve training efficiency and ensure that the model has a certain classification capability at the initial stage; S302, incorporate the pre-trained graph neural network GNN as a classification model into the hybrid expert model, so that it is constructed together with the node pattern extractor and the gating model to obtain a hybrid expert model; after the above training is completed, each GNN model has a certain node feature extraction and classification capability. This batch of pre-trained GNN models are incorporated into the subsequent hybrid expert model (MoE-NP) framework to serve as "expert model" modules (different types of classification models). Due to the diversity of "expert model" modules, the overall system can better cover and adapt to nodes with different structures or feature distributions.

[0030] S303, taking all nodes with true labels as supervisory signals, for each node, defining the loss function as the cross entropy between its prediction result and the true label, performing end-to-end training on the hybrid expert model to optimize and update the parameters of the classification model, the parameters of the edge discriminator in the node pattern extractor, and the parameters of the three multi-layer perceptrons in the gating model, and finally obtaining a hybrid expert model that has completed training.

[0031] In this embodiment, when constructing and training the hybrid expert model, the training of the hybrid expert model includes two main stages: the pre-training stage and the training stage, striving to make full use of complex graph information under the constraint of limited labeled nodes to achieve high-precision classification of nodes. In the pre-training stage, this embodiment first assumes that there is a part of source data containing sufficient node-label pairs, and uses it to train a variety of graph neural networks (GNNs) with K-layer learnable parameters. These GNNs include but are not limited to GCN, GAT, GraphSAGE or other optional models. By minimizing the main task loss function, each GNN is trained separately to obtain a set of pre-trained models with preliminary classification capabilities. In the training stage, in order to make full use of the synergy between the gated network and each expert classifier, they need to be jointly optimized. Specifically, the "expert model" module (different types of classification models): the initialization parameters are derived from the best parameters learned in the pre-training stage, so that it has a certain recognition ability at the starting moment. Gated network: Its parameters are randomly initialized, which is used to gradually learn the dynamic weighting strategy for each expert model during the training process. Edge discriminator and multi-layer perceptron (MLP): In the embodiment of this embodiment, the edge discriminator and the subsequent multi-layer perceptron (MLP) are also randomly initialized to gradually learn effective representations of local and global patterns of nodes during training. For any node in the training set (or on the target graph) containing the annotation In this embodiment, at the beginning of the training process, random walk sampling is performed to obtain the node A local subgraph of Adjacent or close nodes. This can more objectively retain The local structural information in the graph provides a basis for subsequent pattern extraction. In the local subgraph, for the central node With any neighboring node , all form a node pair ( ).Will( ) is input into the "edge discriminator" in the form of a multi-layer perceptron, and a score is output to measure the degree of their association in the feature space. This score can be regarded as a deep measure of the node-node adjacency relationship without the constraint of the true label. In addition, for each node in the local subgraph, its degree (the number of connections of the node in the subgraph or in the global graph) is also obtained separately. The edge discriminator score corresponding to each node is spliced ​​with the node degree information to obtain a preliminary "local node pattern". In order to further take into account the overall distribution information, after obtaining the local node patterns of all nodes, this embodiment obtains the global node pattern by average aggregation (or other aggregation methods). Then: use the first multi-layer perceptron (MLP_local) to perform nonlinear mapping on the local node pattern to obtain a local pattern representation. Use the second multi-layer perceptron (MLP_global) to perform the same nonlinear transformation on the global node pattern to obtain a global pattern representation. The two are feature concatenated (concatenation), and then input into the third multi-layer perceptron (MLP_gating), and a set of weight vectors are output through the softmax function. , where M is the number of expert models. In the training phase, this embodiment uses all nodes with real labels as supervisory signals and uses cross entropy loss to update parameters. For each node, the loss function is defined as the cross entropy between its prediction result and the real label; after accumulating and summing the losses of all nodes, the following are updated through gradient descent and back propagation algorithms: the parameters of all "expert model" modules (different types of classification models) (fine-tuning the existing pre-trained parameters); the parameters of three multi-layer perceptrons (MLP_local, MLP_global, MLP_gating); the parameters of the edge discriminator. This end-to-end training method enables the gating network to learn how to optimally combine the expert models in different node modes, while enabling the expert model to further adapt to the gating mechanism and make targeted optimizations on new training data. In addition, as an optional implementation, in order to further improve the generalization ability of the hybrid expert model on large-scale graphs, unlabeled nodes can be self-trained. Specifically, multiple augmented views can be generated, the difference in predicted distribution under weak augmentation and strong augmentation views can be calculated, and weighted cross entropy training can be performed in combination with prediction confidence, so as to make full use of unlabeled data while ensuring the model's vigilance against uncertain samples. After all the above components are optimized jointly or step by step and completed training, the updated MoE-NP model is used to infer the nodes in the test graph. At this point, the gating model will dynamically assign expert weights to each node based on the final pattern representation of each node, and output the final prediction that integrates multiple expert results, thereby achieving accurate classification of nodes in complex heterogeneous graphs.

[0032] In order to verify the node classification method based on the hybrid expert model of this embodiment, seven data sets selected from public data sets are used in this embodiment, namely Cora, PubMed, Texas, Cornell, Wisconsin, Chameleon and Actor. During the experiment, each data set is tuned to meet its optimal settings, and then different initial states are set to test ten times and the average accuracy index is calculated. The models compared with the hybrid expert model (MoE-NP) used in the method of this embodiment include MLP (multi-layer perceptron that only relies on node feature data), GCN, HighPass GCN, ACMGCN, LINK, LSGNN, GloGNN, and the results are shown in Table 1. At the same time, the performance difference between the hybrid expert model (MoE-NP) using the gated network and the hybrid expert model (no gated network) without the gated network is tested as shown in Table 2.

[0033] Table 1: Comparison of the results of the mixed expert model (MoE-NP) used in this embodiment and the existing model

[0034] Table 2: Comparison of results of the mixture of experts model (MoE-NP) with and without the gating network

[0035] As can be seen from Tables 1 and 2, the hybrid expert model (MoE-NP) used in the method of this embodiment is superior to other traditional models listed and multiple heterogeneous graph node classification methods on average in terms of accuracy indicators. It can be seen that the hybrid expert model (MoE-NP) used in the method of this embodiment realizes adaptive processing of node diversity and uneven distribution by considering multi-dimensional features (structure, features, neighborhood relationships, etc.) in the node pattern extraction link and integrating local patterns and global patterns in the gating model. Different from the idea of ​​a traditional single model, this embodiment introduces multiple "expert" classifiers in the framework, and then uses the gating network to assign appropriate classifier weights to the specific patterns of the nodes, thereby further alleviating the generalization problem caused by heterogeneity and maintaining higher classification accuracy and robustness in complex graph environments. In addition, the hybrid expert model (MoE-NP) used in the method of this embodiment is highly scalable in construction. On the one hand, multiple classification models can be flexibly added or replaced to adapt to different scenarios. On the other hand, it can also be combined with self-supervised training to perform self-training on unlabeled nodes, thereby improving the generalization and adaptability of the model on large-scale graphs.

[0036] In addition, the present embodiment also provides a node classification system based on a hybrid expert model, including a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the node classification method based on the hybrid expert model. In addition, the present embodiment also provides a computer-readable storage medium, wherein a computer program or instruction is stored in the computer-readable storage medium, wherein the computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor. In addition, the present embodiment also provides a computer program product, including a computer program or instruction, wherein the computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor.

[0037] Embodiment 2: This embodiment is basically the same as the first embodiment, with the main difference being that the scenario applied in this embodiment is an item recommendation scenario, specifically, a recommendation system applied to implement item recommendation, in which the node data covers the user's browsing, clicking, collecting, purchasing and other preference behaviors as well as user portraits. By analyzing the consumption capacity and interest distribution, users can be segmented into high-value users with strong willingness to pay, potential interested users whose interest tags are not yet clear but have potential purchasing motivations, and multi-field active users who are active in multiple content or product categories. These classification results can provide more refined decision support for the personalized push of the recommendation system.

[0038] In addition, the present embodiment also provides a node classification system based on a hybrid expert model, including a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the node classification method based on the hybrid expert model. In addition, the present embodiment also provides a computer-readable storage medium, wherein a computer program or instruction is stored in the computer-readable storage medium, wherein the computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor. In addition, the present embodiment also provides a computer program product, including a computer program or instruction, wherein the computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor.

[0039] Embodiment three: This embodiment is basically the same as the first embodiment, the main difference being that the scenario used in this embodiment is a traffic prediction scenario. Node data may refer to the geographical location, historical traffic volume, time period characteristics, weather conditions, etc. around urban road intersections or transportation hubs, thereby classifying nodes into congested nodes prone to congestion during peak hours, key nodes that have a key impact on the overall traffic flow, and stable nodes with relatively stable traffic that can provide diversion or buffering effects. These classification results can provide more refined decision support for congestion warning and capacity scheduling in urban traffic management.

[0040] In addition, the present embodiment also provides a node classification system based on a hybrid expert model, including a microprocessor and a memory connected to each other, wherein the microprocessor is programmed or configured to execute the node classification method based on the hybrid expert model. In addition, the present embodiment also provides a computer-readable storage medium, wherein a computer program or instruction is stored in the computer-readable storage medium, wherein the computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor. In addition, the present embodiment also provides a computer program product, including a computer program or instruction, wherein the computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model through a processor.

[0041] Those skilled in the art should understand that the technical solutions provided by the embodiments of the present application may be in the form of methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0042] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A node classification method based on a hybrid expert model, characterized in that: The method comprises the following steps: inputting node data into a pre-trained hybrid expert model to obtain a classification result of the node data, wherein the hybrid expert model comprises a node pattern extractor, a gating model and a plurality of classification models of different types, wherein the node pattern extractor is used to extract node pattern information from given node data, the gating model is used to receive node pattern information and generate a model combination weight corresponding to each node, and the classification results of the node data by the plurality of classification models of different types are dynamically selected and combined according to the model combination weight corresponding to the node to obtain the final classification result of the node.

2. The node classification method based on the hybrid expert model according to claim 1 is characterized in that: The node pattern extractor extracts node patterns from given node data including: S101, for each node in the given node data , for nodes The neighborhood of the node is randomly sampled to obtain The local subgraph contains several nodes Nodes in the neighborhood; S102, for each node in the given node data , the node Any node in its local subgraph A node pair , the node pair Input into the edge discriminator composed of multi-layer perceptron to obtain the measurement node Any node in its local subgraph The edge discriminator scores the degree of association in the feature space; The number of connections to its local subgraph or to all local subgraphs as a node Node degree information of S103, for each node in the given node data , the node The edge discriminator score of the node is concatenated with the node degree information as Local node mode information; S104, aggregating the local node mode information of all nodes in a specified aggregation manner to obtain global node mode information, and outputting the local node mode information and the global node mode information as the finally obtained node mode information.

3. The node classification method based on the hybrid expert model according to claim 2 is characterized in that: The aggregation method specified in step S104 is average aggregation.

4. The node classification method based on the hybrid expert model according to claim 2 is characterized in that: The gating model receives the node pattern and generates the model combination weights corresponding to each node including: S201, using a first multi-layer perceptron to perform nonlinear mapping on local node pattern information of a node to obtain a local pattern representation, and using a second multi-layer perceptron to perform nonlinear mapping on global node pattern information to obtain a global pattern representation; S202, concatenating the local pattern representation and the global pattern representation of the node; S203, input the concatenated result into the third multi-layer perceptron, and then output a set of weight vectors of the node through the softmax function ,in ~ are the model combination weights of the M classification models corresponding to the node, and M is the number of classification models.

5. The node classification method based on the hybrid expert model according to claim 4 is characterized in that: The classification results of the node data by the multiple different types of classification models are dynamically selected and combined according to the model combination weights corresponding to the nodes to obtain the function expression of the final classification result of the node: ， in, is the final classification result of the i-th node, It means summation, is the model combination weight of the j-th classification model corresponding to the i-th node, is the classification result of the node data of the i-th node by the j-th classification model.

6. The node classification method based on the hybrid expert model according to claim 5 is characterized in that: Before the classification results of the node data by the multiple different types of classification models are dynamically selected and combined according to the model combination weights corresponding to the nodes to obtain the final classification results of the nodes, the step of constructing and training a hybrid expert model is also included: S301, select multiple different types of graph neural networks (GNNs) with K layers of learnable parameters, and use source data containing node-label pairs for separate pre-training, so that they have preliminary classification capabilities by minimizing the main task loss function; S302, incorporating the pre-trained graph neural network GNN as a classification model into the hybrid expert model, so that it is constructed together with the node pattern extractor and the gating model to obtain the hybrid expert model; S303, taking all nodes with true labels as supervisory signals, for each node, defining the loss function as the cross entropy between its prediction result and the true label, performing end-to-end training on the hybrid expert model to optimize and update the parameters of the classification model, the parameters of the edge discriminator in the node pattern extractor, and the parameters of the three multi-layer perceptrons in the gating model, and finally obtaining a hybrid expert model that has completed training.

7. The node classification method based on hybrid expert model according to claim 1 is characterized in that: The node data is the data of the user node in the social networking platform. The data of the user node includes the user's basic information, user sharing information and user's interactive behavior. The user sharing information includes part or all of the text, pictures and videos shared by the user. The user's interactive behavior includes part or all of the user's following, liking, commenting and forwarding.

8. A node classification system based on a hybrid expert model, comprising a microprocessor and a memory connected to each other, characterized in that: The microprocessor is programmed or configured to execute the node classification method based on the hybrid expert model as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model described in any one of claims 1 to 7 through a processor.

10. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the node classification method based on the hybrid expert model described in any one of claims 1 to 7 through a processor.

Images