A node separation learning method and system based on discriminative difficulty perception

Abstract

The application discloses a node separation learning method and system based on discriminative difficulty perception, relates to the technical field of graph neural networks, and quantifies the discriminative difficulty of a node multi-hop neighborhood by calculating the label distribution relaxation entropy value of the node multi-hop neighborhood; a hybrid expert network architecture is constructed; the entropy value and node features are spliced and a threshold value for a Top-p expert selection mechanism is predicted; the threshold value is used to activate an expert subnetwork set and the outputs of the expert subnetwork set are weighted and combined to generate a node representation; finally, a prediction result is obtained through a classifier and the entropy value is dynamically updated during the training process. The application can adaptively allocate computing resources and improve the accuracy and robustness of heterogeneous graph node classification.

Description

Technical Field

[0001] This invention belongs to the field of graph neural network technology, and particularly relates to a node separation learning method and system based on discrimination difficulty perception. Background Technology

[0002] Graph data has wide applications in the real world, such as social networks, scientific literature, bioinformatics, and online recommendation systems. Accurate node classification within these complex and dynamic graph structures is fundamental to many downstream tasks, such as user interest identification, malicious account detection, literature topic classification, or product category prediction. However, in such massive and dynamic graph data, nodes often exhibit diverse and complex attributes and relationships. Therefore, how to efficiently classify nodes is a significant technical challenge in the field of graph data processing.

[0003] In traditional Graph Neural Network (GNN) node classification tasks, models are typically designed based on the "homophily" assumption, meaning that neighboring nodes have similar label features. However, in heteroophilic graphs, the labels of neighboring nodes differ significantly, causing the message passing mechanism based on the homoophily assumption to fail. For example, in social networks, users may be connected to neighboring nodes with significantly different interests; in academic citation networks, papers may cite literature from different research fields. In such scenarios, traditional GNN models struggle to effectively aggregate neighborhood information, leading to decreased classification accuracy. Existing strategies for enhancing GNNs to address heterogeneity, such as introducing heterogeneous edge modeling or integrating wide-area neighborhood information, are often insufficient. This is mainly because they lack adaptability and fixed feature integration strategies, performing poorly when handling multiple attributes and complex heterogeneous associations. Furthermore, a single model architecture struggles to adapt to the significant differences between nodes in the graph, limiting classification performance.

[0004] To address these challenges, the Mixture-of-Experts (MoE) architecture has gained attention in recent years. MoEs aim to improve model capacity and the ability to handle diverse information by introducing multiple parallel expert sub-networks and a gating network responsible for dynamically allocating tasks. However, applying MoEs to node classification in complex graph structures still faces significant challenges. First, accurately quantifying the inherent discrimination difficulty of nodes in the graph to provide a fine-grained expert selection criterion closely integrated with the model learning process is a technical problem that has not yet been fully solved in existing solutions. Second, how to achieve intelligent and adaptive selection and combination of different expert sub-networks by the gating network based on the perception of node discrimination difficulty, thereby optimizing the allocation of computing resources and improving classification performance, is also an aspect that needs further optimization when applying MoEs to graph node classification tasks. Existing MoE solutions typically lack a deep perception of the inherent discrimination difficulty of nodes, resulting in limited optimization space for expert routing mechanisms. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a node separation learning method and system based on discrimination difficulty perception, thereby resolving the issues present in the prior art.

[0006] Firstly, to achieve the above objectives, this invention provides a node separation learning method based on discrimination difficulty awareness, comprising the following steps:

[0007] S1. For each target node in the graph, analyze the distribution of node labels in its multi-hop neighborhood and calculate a label distribution relaxation entropy value.

[0008] S2. Construct a hybrid expert network architecture that includes M parallel expert subnetworks and one gated network.

[0009] S3. Concatenate the relaxed entropy value of the node's label distribution with the node's own features to form an input vector, and feed it into a prediction module to output a p threshold for the Top-p expert selection mechanism.

[0010] S4. Sort the expert weight vectors generated by the gated network in descending order, determine a set of activated expert subnetworks based on the p threshold, and then perform a weighted combination of the outputs of each expert subnetwork in the set to generate the final node representation.

[0011] S5. Input the final node representation into the classifier to obtain the prediction result, and use the cross-entropy loss function to perform end-to-end optimization training on the entire model; when the validation set performance reaches the preset optimal standard, recalculate and update the label distribution relaxation entropy value of all nodes.

[0012] Optionally, in S1, the process of calculating the label distribution relaxation entropy value includes: for nodes that do not yet have real labels, their label distribution relaxation entropy value is initialized using pseudo labels predicted by the pre-trained model or default values.

[0013] Optionally, in S2, the expert sub-network is a graph convolutional network or a graph attention network; the gated network is a multilayer perceptron, used to receive the self-features of the target node and output an M-dimensional expert weight vector normalized by the Softmax function.

[0014] Optionally, in S3, the prediction module is a multilayer perceptron network, and a sigmoid activation function is used in its output layer to map the output to the (0,1) interval to obtain the p threshold.

[0015] Optionally, in S4, the process of determining the set of activated expert subnetworks based on the p threshold includes: after sorting the expert weight vectors in descending order, determining the minimum number of experts that satisfy the cumulative weight and not less than the p threshold, and including the corresponding number of top weight experts into the set.

[0016] Optionally, in S5, the process of recalculating and updating the label distribution relaxation entropy value of all nodes includes: regenerating pseudo-labels for all nodes without real labels in the graph using the current model, and performing the step of calculating the label distribution relaxation entropy value based on the updated pseudo-labels.

[0017] Secondly, the present invention also provides a node separation learning system based on discrimination difficulty awareness, for implementing a node separation learning method based on discrimination difficulty awareness, the system comprising:

[0018] The difficulty quantification module is used to analyze the distribution of node labels in the multi-hop neighborhood of each target node in the graph and calculate a label distribution relaxation entropy value.

[0019] The hybrid expert network module is used to construct a hybrid expert network architecture that includes M parallel expert subnetworks and a gated network.

[0020] The dynamic expert routing module is used to concatenate the relaxed entropy value of the node's label distribution with the node's own features into an input vector, and then feed it into a prediction module to output a p threshold for the Top-p expert selection mechanism.

[0021] The node representation and classification module is used to sort the expert weight vectors generated by the gated network in descending order, determine a set of activated expert sub-networks based on the p threshold, and then perform a weighted combination of the outputs of each expert sub-network in the set to generate the final node representation. The final node representation is then input into the classifier to obtain the prediction result.

[0022] The model training and update module is used to perform end-to-end optimization training of the entire model using the cross-entropy loss function, and when the validation set performance reaches the preset optimal standard, it triggers the discrimination difficulty quantization module to recalculate and update the label distribution relaxation entropy value of all nodes.

[0023] Thirdly, the present invention also provides a computer terminal device, comprising:

[0024] One or more processors;

[0025] A memory, coupled to the processor, for storing one or more programs;

[0026] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the node separation learning method based on discrimination difficulty awareness in the first aspect described above.

[0027] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the node separation learning method based on discrimination difficulty perception in the first aspect described above.

[0028] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the node separation learning method based on discrimination difficulty perception described in the first aspect above.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] This invention provides a node separation learning method and system based on discriminative difficulty awareness. The invention uses Label Distribution Relaxation Entropy (LDRE) to provide a fine-grained measure of discriminative difficulty for each node. Based on this difficulty awareness, the model dynamically adjusts the Top-p expert routing mechanism, allocating more computational resources to high-difficulty nodes. Combined with a dynamic update mechanism during training, a virtuous cycle of self-optimization is formed, effectively improving the accuracy and robustness of heterogeneous graph node classification. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0032] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0033] Figure 2 This is a detailed flowchart illustrating the dynamic expert routing and model optimization process according to an embodiment of the present invention.

[0034] Figure 3 This is a framework diagram of a node separation learning system based on discrimination difficulty perception, according to an embodiment of the present invention. Detailed Implementation

[0035] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0037] Example 1

[0038] This embodiment provides a node separation learning method based on discrimination difficulty awareness, including:

[0039] S1. For each target node in the graph, analyze the distribution of node labels in its multi-hop neighborhood and calculate a label distribution relaxation entropy value.

[0040] S2. Construct a hybrid expert network architecture that includes M parallel expert subnetworks and one gated network.

[0041] S3. Concatenate the relaxed entropy value of the node's label distribution with the node's own features to form an input vector, and feed it into a prediction module to output a p threshold for the Top-p expert selection mechanism.

[0042] S4. Sort the expert weight vectors generated by the gated network in descending order, determine a set of activated expert subnetworks based on the p threshold, and then perform a weighted combination of the outputs of each expert subnetwork in the set to generate the final node representation.

[0043] S5. Input the final node representation into the classifier to obtain the prediction result, and use the cross-entropy loss function to perform end-to-end optimization training on the entire model; when the validation set performance reaches the preset optimal standard, recalculate and update the label distribution relaxation entropy value of all nodes.

[0044] Furthermore, the above content specifically includes:

[0045] S1: For each target node in the target graph, analyze the distribution of node labels in its multi-hop neighborhood, and calculate and obtain a label distribution relaxation entropy (LDRE) value.

[0046] S2: Construct a hybrid expert (MoE) network architecture, which consists of M parallel expert subnetworks for extracting node features, and a gated network for generating an M-dimensional expert weight vector for each node.

[0047] S3: Concatenate the relaxed entropy value of the node's label distribution with the node's own features to form an input vector, and feed this vector into a prediction module. This module outputs a p-threshold for the Top-p expert selection mechanism.

[0048] S4: Sort the expert weight vectors generated by the gating network in descending order, and determine an active expert subnetwork set based on the p threshold; then, use the corresponding expert weights generated by the gating network to weight and combine the outputs of each expert subnetwork in the activated expert subnetwork set to generate the final node representation.

[0049] S5: Input the final node representation into a classifier to obtain the prediction result, and use the cross-entropy loss function to perform end-to-end optimization training on all learnable parameters of the entire model; during the training process, when the performance of the validation set reaches the preset optimal standard, recalculate and update the relaxation entropy value of the label distribution of all nodes.

[0050] As one implementation method in this embodiment, in S1, the process of calculating the label distribution relaxation entropy value includes: for nodes that do not yet have real labels, their label distribution relaxation entropy value is initialized using pseudo labels predicted by the pre-trained model or default values.

[0051] Further, refer to the appendix. Figure 1 The specific steps for inputting graph data and calculating the LDRE value of each node are as follows:

[0052] S1.1: The target graph package mentioned in this embodiment is a set of nodes. Edge set Node feature matrix The graph data structure includes partial node labels. The graph data can originate from various application scenarios and is not limited to a specific type. For example, the target graph can be a scientific literature network (such as Cora or CiteSeer), where nodes represent papers and edges represent citations; it can also be a social network, where nodes represent users and edges represent follow relationships; or it can be a bioinformatics network (such as a protein-protein interaction network). In this embodiment, a general graph dataset is used as an example, whose node features are represented by feature vectors. The topological connections between nodes are represented by the adjacency matrix. It is indicated that the class labels of the training set nodes are known, denoted as . It is used for supervised training of the model.

[0053] S1.2: In order for the model to perceive the classification difficulty of each node in the graph and provide key prior knowledge for subsequent adaptive processing, it is first necessary to process each node in the graph. Calculate its Relaxed entropy of label distribution within skip neighborhood Value. In this embodiment, it is set. =1, meaning only one-hop neighbors of each node are considered. The formula for calculating LDRE is as follows:

[0054] ;

[0055] in,

[0056] ;

[0057] in Represents the set of node label categories, where Represents a node of Skip Neighborhood. A higher value indicates a greater degree of label confusion within the target neighborhood.

[0058] in It is a node of The most prominent label category within the 2-hop neighborhood. For example, for a paper A in the field of "artificial intelligence," its 2-hop neighborhood contains 10 papers, of which 3 belong to "artificial intelligence," 5 to "databases," and 2 to "cybersecurity." Therefore, its primary label category is... This is referred to as the "database". The LDRE value quantifies the degree of disorder in these labels; the higher the value, the more "mismatched" the neighborhood, and the greater the difficulty in classification. For nodes in the test set without real labels, their LDRE values ​​are uniformly set to the default value in the initial stage of this embodiment.

[0059] As one implementation method in this embodiment, in S2, the expert sub-network is a graph convolutional network or a graph attention network; the gated network is a multilayer perceptron, used to receive the self-features of the target node and output an M-dimensional expert weight vector normalized by the Softmax function.

[0060] Furthermore, as attached Figure 2 The specific steps of S2, which constructs the expert routing mechanism, and S3 and S4, which are based on the current difficulty-aware dynamic expert routing process, include:

[0061] S2.1: In this embodiment, M=8 parallel expert subnetworks are set. Each expert subnetwork... Each uses an independent, learnable parameter A two-layer graph convolutional network (GCN). Each expert network receives a complete graph adjacency matrix. and node features As input, these expert networks can learn node representations from different feature subspaces and output their respective understandings of the nodes. The output node representation matrix is ​​as follows:

[0062] ;

[0063] S2.2: To dynamically assign the most suitable expert to different nodes, a gated network G needs to be set up. This network is implemented as a network with learnable parameters. Multilayer perceptrons (MLPs) or similar feedforward neural networks. Gated networks accept input from a single node. As input, it outputs an M-dimensional (M=8) weight vector normalized by the Softmax function. It is expressed as follows:

[0064] ;

[0065] in The Each component This indicates that the j-th expert has a view on the node. Applicability or contribution weight.

[0066] As one implementation method in this embodiment, in S3, the prediction module is a multilayer perceptron network, and the Sigmoid activation function is used in its output layer to map the output to the (0,1) interval to obtain the p threshold.

[0067] Furthermore, in step S3, the difficulty-aware dynamic expert routing mechanism is implemented through a p-value prediction network, and its specific steps are as follows:

[0068] S3.1: For each node Construct an input vector that integrates its own features and discrimination difficulty information. This vector is derived from the original features of the node. The relaxed entropy of the label distribution calculated in S1.2 It is assembled from various parts, as shown below:

[0069] ;

[0070] S3.2: Using the obtained concatenation vector The model can dynamically predict a p-threshold for each node, which directly determines the number of experts activated. The prediction network described in this embodiment is a network with learnable parameters. Multilayer perceptron network Or a similar feedforward neural network, using a sigmoid activation function in its output layer to ensure the p-threshold of the output. Within the interval (0, 1), it is represented as follows:

[0071] ;

[0072] Nodes that are difficult to identify (high LDRE value) will tend to have a higher p threshold.

[0073] As one implementation method in this embodiment, in S4, the process of determining the set of activated expert subnetworks based on the p threshold includes: after arranging the expert weight vectors in descending order, determining the minimum number of experts that satisfy the cumulative weight and not less than the p threshold, and including the corresponding number of top weight experts into the set.

[0074] Furthermore, in step S4, the adaptive activation of the expert sub-network's Top-p selection and weighted combination process specifically includes the following sub-steps:

[0075] S4.1: For each node Here, the expert weight vector obtained in S2.2 will be used. Sort in descending order to obtain an ordered sequence of expert indexes. Then determine whether the sum of the accumulated weights is not less than the p threshold. Minimum number of experts , means as follows:

[0076] ;

[0077] This determines the set of indices for the activated expert subnetworks. .

[0078] S4.2: To obtain the final node representation, only the outputs of the activated expert subnetworks are weighted and combined here to obtain the node representation. Final expression The formula is as follows:

[0079] ;

[0080] in Indicates the first The node representation matrix output by each activated expert subnetwork. Refers to the nodes in the matrix The corresponding node representation vector. Assign nodes to the gated network The The weights of each activated expert subnetwork. In this way, the nodes... The final expression It is a weighted aggregation of its representations from the perspectives of each activated expert.

[0081] As one implementation method in this embodiment, in S5, the process of recalculating and updating the label distribution relaxation entropy value of all nodes includes: using the current model to regenerate pseudo labels for all nodes in the graph that have no real labels, and performing the step of calculating the label distribution relaxation entropy value based on the updated pseudo labels.

[0082] Furthermore, in step S5, the model optimization and dynamic update strategy specifically includes:

[0083] S5.1: Represent the nodes obtained in S4.2 The node representation matrix is ​​composed of The input is fed into a downstream classifier, which is typically implemented as a linear layer. This linear layer has a learnable weight matrix. and bias vector (The parameters can be collectively referred to as) This linear layer represents the input nodes. The raw scores (logits) are mapped to the corresponding categories. These raw scores are then transformed into the final classification prediction probability matrix B using a softmax activation function. Therefore, the calculation process of the classification prediction probability matrix B can be expressed as:

[0084] ;

[0085] H is represented by all nodes. The nodes that make up the matrix represent the matrix. and These are the learnable parameters of the linear layer.

[0086] Then, the cross-entropy loss function is used. End-to-end optimization of all learnable parameters of the entire model is represented as follows:

[0087] ;

[0088] in This represents the label matrix of the training set, and its transpose is then applied. Represents the prediction probability matrix Take the natural logarithm of each element; Represents the trace of a matrix.

[0089] S5.2: During training, when the model's accuracy on the validation set reaches a new peak, a dynamic update mechanism is triggered. This mechanism uses the current model to regenerate pseudo-labels for all nodes in the graph that lack true labels. Based on these updated pseudo-labels, it returns to step S1.2 to globally update the LDRE values ​​and related routing parameters (i.e., p-values) of all nodes, thereby improving the model's adaptability and robustness in subsequent training.

[0090] To demonstrate the effectiveness of this invention, it is compared with the following four common models to analyze the performance of classification tasks.

[0091] Table 1. Comparison of Results between the Method of This Embodiment and Existing Models

[0092]

[0093] Table 1 shows a comparison of the node classification accuracy (mean ± standard deviation, based on 10 tests) of a method implemented according to an embodiment of the present invention with several other models on the Computers, Wiki-CS, Actor, and Roman-empire datasets. Bold text indicates the highest accuracy on this dataset. As can be seen from the data in Table 1, the method implemented according to an embodiment of the present invention exhibits high classification accuracy on the tested datasets. This result demonstrates that the technical solution proposed in this invention is effective in handling heterogeneous graph data and improving the model's adaptive capability, providing a technical solution for graph node classification tasks.

[0094] Based on this, the present invention provides a node separation learning method based on discrimination difficulty perception. The advantages and effects of the present invention compared with the prior art are as follows:

[0095] (1) This invention provides a fine-grained, dynamically updatable measure of discrimination difficulty for each node in the graph by defining the label distribution relaxation entropy (LDRE). This enables the model to accurately perceive which nodes are "easy to discriminate" and which are "fuzzy and difficult to distinguish" for the first time, providing crucial prior knowledge for subsequent adaptive processing.

[0096] (2) This invention proposes a dynamic expert routing mechanism that is aware of the difficulty of discrimination. This mechanism takes the LDRE value as input and dynamically predicts the p threshold of each node in the Top-p routing strategy. This allows the model to intelligently allocate more computing resources (i.e., activate more experts) to nodes with higher discrimination difficulty, while using lighter computing for nodes with lower discrimination difficulty, which helps to achieve a balance between computational efficiency and model performance.

[0097] (3) This invention introduces a dynamic update mechanism during the training process. When the model's performance improves on the validation set, the LDRE values ​​of all nodes are recalculated. This constitutes a virtuous cycle: the improved model capability makes its judgment of node difficulty more accurate, and the more accurate difficulty judgment, in turn, guides the model to perform more effective training, thereby continuously improving the model's adaptability and robustness.

[0098] Example 2

[0099] See attached document Figure 3 Based on the same general inventive concept, this invention also provides a node separation learning system based on discrimination difficulty awareness. The node separation learning system based on discrimination difficulty awareness provided by this invention is described below. The node separation learning system based on discrimination difficulty awareness described below can be referred to in correspondence with the node separation learning method based on discrimination difficulty awareness described above. The system includes:

[0100] The difficulty quantification module analyzes the local neighborhood environment of each node in the graph and calculates the label distribution relaxation entropy value, which serves as an indicator to quantify its classification ambiguity or difficulty. This module provides crucial difficulty perception basis for subsequent dynamic routing.

[0101] The hybrid expert network module consists of multiple parallel expert sub-networks and a gating network. Each expert sub-network is responsible for learning and extracting node features from different perspectives, while the gating network initially evaluates the applicability of all experts for each node and generates a set of basic weights.

[0102] The dynamic expert routing module is the core of the system's decision-making process. It receives the node's features and the LDRE value output by the discrimination difficulty quantification module, and intelligently and dynamically determines the number of experts required to activate for each node's decision. Subsequently, based on the weights of the gating network, it determines which activated expert subnetworks will participate in the current node's processing.

[0103] The node representation and classification module integrates the processing results of experts activated by the dynamic routing module. It intelligently weights and combines the outputs of these selected experts to form a final node representation that integrates multiple aspects of information, and performs the final classification prediction based on this representation.

[0104] The model training and update module is responsible for optimizing the entire system. It guides the learning of all module parameters through a loss function. Simultaneously, this module includes a unique feedback mechanism that dynamically triggers and updates the calculation basis of the discrimination difficulty quantification module based on the model's performance during training, thereby achieving continuous self-improvement and evolution of the overall system performance.

[0105] It should be understood that the node separation learning system based on discrimination difficulty awareness provided in the embodiments of the present invention has all the advantages of the node separation learning method based on discrimination difficulty awareness provided in the above embodiments.

[0106] Example 3

[0107] In this embodiment, a computer terminal device is provided, including:

[0108] One or more processors;

[0109] A memory, coupled to the processor, for storing one or more programs;

[0110] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described node separation learning method based on discrimination difficulty awareness.

[0111] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described node separation learning method based on discrimination difficulty perception.

[0112] In this embodiment, an electronic device is also provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps of the above-described node separation learning method based on discrimination difficulty perception.

[0113] In this embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the above-described node separation learning method based on discrimination difficulty perception.

[0114] The aforementioned program can run on a processor or be stored in memory (or a computer-readable medium). Computer-readable media includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random-access memory (SRAM), dynamic random-access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0115] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented by different modules for different steps.

[0116] This embodiment provides such a device or system. The system, referred to as a node separation learning system based on discrimination difficulty awareness, includes:

[0117] The difficulty quantification module is used to analyze the distribution of node labels in the multi-hop neighborhood of each target node in the graph and calculate a label distribution relaxation entropy value.

[0118] The hybrid expert network module is used to construct a hybrid expert network architecture that includes M parallel expert subnetworks and a gated network.

[0119] The dynamic expert routing module is used to concatenate the relaxed entropy value of the node's label distribution with the node's own features into an input vector, and then feed it into a prediction module to output a p threshold for the Top-p expert selection mechanism.

[0120] The node representation and classification module is used to sort the expert weight vectors generated by the gated network in descending order, determine a set of activated expert sub-networks based on the p threshold, and then perform a weighted combination of the outputs of each expert sub-network in the set to generate the final node representation. The final node representation is then input into the classifier to obtain the prediction result.

[0121] The model training and update module is used to perform end-to-end optimization training of the entire model using the cross-entropy loss function, and when the validation set performance reaches the preset optimal standard, it triggers the discrimination difficulty quantization module to recalculate and update the label distribution relaxation entropy value of all nodes.

[0122] As one implementation method in this embodiment, the difficulty quantification module includes:

[0123] An initialization unit is used to initialize the label distribution relaxation entropy value of nodes that do not yet have real labels using pseudo-labels predicted by a pre-trained model or default values.

[0124] As one implementation method in this embodiment, the hybrid expert network module includes:

[0125] Multiple expert sub-network units are composed of graph convolutional networks or graph attention networks;

[0126] A gated network unit, composed of multilayer sensing mechanisms, is used to receive the self-features of the target node and output an M-dimensional expert weight vector normalized by the Softmax function.

[0127] As one implementation method in this embodiment, the dynamic expert routing module includes:

[0128] A prediction unit, consisting of a multilayer perceptron network, uses a sigmoid activation function in its output layer to map the output to the (0,1) interval to obtain the p threshold.

[0129] As one implementation method in this embodiment, the node representation and classification module includes:

[0130] An expert selection unit is used to sort the expert weight vectors in descending order, determine the minimum number of experts that satisfy the cumulative weight and not less than the p threshold, and include the corresponding number of top weight experts into the set of activated expert subnetworks.

[0131] As one implementation method in this embodiment, the model training and update module includes:

[0132] A feedback update unit is used to regenerate pseudo-labels for all nodes without real labels in the graph using the current model, and instructs the discrimination difficulty quantization module to recalculate the label distribution relaxation entropy value based on the updated pseudo-labels.

[0133] The system or apparatus is used to implement the functions of the methods in the above embodiments. Each module in the system or apparatus corresponds to each step in the method, as has been described in the method and will not be repeated here.

[0134] The above implementation method solves the problem of node separation learning based on discrimination difficulty perception in related technologies, thereby ensuring that the problems existing in the prior art are resolved.

[0135] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A discriminative difficulty-aware based node disentangled learning method, characterized in that, The method comprises the following steps: S1, for each target node in the graph, analyze the node label distribution in its multi-hop neighborhood, and calculate a label distribution relaxation entropy value, wherein the graph is a scientific literature network, the target node is a paper, the edge is a citation, and the scientific literature network comprises Cora or CiteSeer; S2, a hybrid expert network architecture comprising M parallel expert subnetworks and a gating network is constructed; S3, the label distribution relaxation entropy value of the node and the node's own features are spliced into an input vector, which is input into a prediction module to output a p threshold value for a Top-p expert selection mechanism; S4, the expert weight vector generated by the gating network is arranged in descending order, and an activated expert subnetwork set is determined according to the p threshold value, and the outputs of the expert subnetworks in the set are combined to generate the final node representation; S5, the final node representation is input into a classifier to obtain a prediction result, and an cross-entropy loss function is used to optimize and train the whole model in an end-to-end manner; When the performance of the validation set reaches the preset optimal standard, the label distribution relaxation entropy values of all nodes are recalculated and updated.

2. The method of claim 1, wherein, In S1, the process of calculating the label distribution relaxation entropy value includes: for the nodes without real labels, the label distribution relaxation entropy value is initialized by using the pseudo-label predicted by the pre-trained model or a default value.

3. The method of claim 1, wherein, In S2, the expert subnetwork is a graph convolutional network or a graph attention network; the gating network is a multilayer perceptron, which is used to receive the node's own features and output an M-dimensional expert weight vector normalized by a Softmax function.

4. The method of claim 1, wherein, In S3, the prediction module is a multilayer perceptron network, and a Sigmoid activation function is used in the output layer to map the output to the interval (0, 1) to obtain the p threshold value.

5. The method of claim 1, wherein, In S4, the process of determining the activated expert subnetwork set according to the p threshold value includes: after arranging the expert weight vector in descending order, the minimum number of experts satisfying the cumulative weight sum not less than the p threshold value is determined, and the corresponding number of top weight experts are included in the set.

6. The method of claim 1, wherein, In S5, the process of recalculating and updating the label distribution relaxation entropy values of all nodes includes: using the current model to regenerate the pseudo-labels for all nodes without real labels in the graph, and performing the step of calculating the label distribution relaxation entropy value based on the updated pseudo-labels.

7. A discriminative difficulty-aware based node disentangled learning system, characterized in that, The system comprises: a difficulty quantification module, configured to analyze the node label distribution in the multi-hop neighborhood of each target node in the graph, and calculate a label distribution relaxation entropy value, wherein the graph is a scientific literature network, the target node is a paper, the edge is a citation, and the scientific literature network comprises Cora or CiteSeer; a hybrid expert network module, configured to construct a hybrid expert network architecture comprising M parallel expert subnetworks and a gating network; a dynamic expert routing module, configured to splice the label distribution relaxation entropy value of the node and the node's own features into an input vector, and input the input vector into a prediction module to output a p threshold value for a Top-p expert selection mechanism; The node representation and classification module is configured to arrange the expert weight vectors generated by the gating network in descending order, determine an activated expert sub-network set according to a p threshold, combine the outputs of the expert sub-networks in the set to generate a final node representation, and input the final node representation into a classifier to obtain a prediction result. The model training and updating module is configured to perform end-to-end optimization training on the entire model by using a cross-entropy loss function, and trigger the difficulty quantification module to recalculate and update the label distribution relaxation entropy values of all nodes when the performance of a verification set reaches a preset optimal standard.

8. A computer terminal device, characterized by The computer program product comprises: one or more processors; a memory coupled to the processors for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program product comprises:

10. A computer program product comprising a computer program, characterized in that, one or more processors; a memory coupled to the processors for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method according to any one of claims 1-6. The computer program product comprises: one or more processors; a memory coupled to the processors for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method according to any one of claims 1-6.

Images