A Heterogeneous Graph Node Classification Method Based on Relationship-Aware Label Propagation

Abstract

This invention relates to the fields of graph neural networks and heterogeneous graph representation learning technology, specifically to a heterogeneous graph node classification method based on relation-aware label propagation. The method includes the following steps: first, based on a given heterogeneous network, then performing type-specific linear transformations on nodes of different types and projecting their features onto a common feature space; second, designing a relation-aware label propagation algorithm for target nodes in the heterogeneous network, generating pseudo-labels through relation subgraphs; and third, employing a two-layer aggregation strategy based on type attention to transmit information and effectively fuse multi-level information of nodes, resulting in richer and more accurate node feature representations. This invention significantly improves classification performance and robustness by designing a relation-aware label propagation method to obtain pseudo-labels, utilizing two-layer aggregation based on type attention for heterogeneous message transmission, and combining it with a multi-objective optimization strategy.

Description

Technical Field

[0001] This invention relates to the fields of graph neural networks and heterogeneous graph representation learning technology, and in particular to a heterogeneous graph node classification method based on relation-aware label propagation. Background Technology

[0002] In recent years, although some research has begun to focus on the imbalance problem in homogeneous graphs, research on imbalanced node classification in heterogeneous graphs is relatively limited due to the inherent complexity and heterogeneity of heterogeneous graphs. Existing neural network methods for heterogeneous graphs mainly include: neighborhood aggregation-based HGNN methods, which learn node representations by aggregating the neighborhood information of nodes, such as GraphSAGE and GAT. However, these methods often ignore the features of minority class nodes when dealing with imbalanced data, resulting in poor classification performance. Metapath-based HGNN methods, which utilize predefined metapaths to learn node representations in heterogeneous graphs, such as HIN2Vec and Metapath2Vec. While these methods can capture structural information in heterogeneous graphs, they still have limitations when facing imbalanced node classification problems. Label propagation is a classic semi-supervised learning method that propagates information from known labeled nodes to unlabeled nodes. For example, traditional label propagation methods (LP) have some effect in homogeneous graphs, but their effectiveness is limited when dealing with heterogeneous graphs.

[0003] To further improve the classification performance of imbalanced nodes in heterogeneous graphs, researchers have proposed several novel methods. For example, Relationship-Aware Label Propagation (RLP) is an emerging approach that considers the relationships between nodes and enhances the classification performance of minority class nodes by propagating labels. However, existing RLP methods still have limitations in heterogeneous graphs because they do not fully utilize the type and structural information in the graphs. To address these issues, some studies have introduced a two-level aggregation strategy, which aggregates at both the micro and macro levels, better handling the information transfer and aggregation problems in heterogeneous graphs. For instance, some attention-based methods, such as GAT, while performing well in homogeneous graphs, still have limited application in heterogeneous graphs. In summary, existing methods have the following shortcomings when dealing with the classification problem of imbalanced nodes in heterogeneous graphs:

[0004] (1) Ignoring minority class nodes: Existing methods tend to favor majority class nodes, resulting in poor classification performance of minority class nodes.

[0005] (2) Incomplete use of heterogeneous graph information: Many methods do not make full use of the type and structural information in heterogeneous graphs.

[0006] (3) Limited to isomorphic graph methods: Some traditional label propagation and attention mechanisms are effective in isomorphic graphs, but have limited performance in heteromorphic graphs.

[0007] To address the problem of classifying imbalanced nodes in heterogeneous graphs, a heterogeneous graph node classification method based on relation-aware label propagation is proposed. Summary of the Invention

[0008] In view of this, the purpose of this invention is to propose a heterogeneous graph node classification method based on relation-aware label propagation, so as to solve the problems of imbalanced node classification and inability to utilize heterogeneous graph information in the prior art.

[0009] To achieve the above objectives, this invention provides a heterogeneous graph node classification method based on relation-aware label propagation, comprising the following steps:

[0010] S110. Based on the given heterogeneous network, perform type-specific linear transformations on nodes of different types. Nodes of different types have different semantic features and dimensions, and project their features into a common feature space.

[0011] S120. For target nodes in heterogeneous networks, design a relation-aware label propagation algorithm and generate pseudo-labels through relation subgraphs.

[0012] S130. A two-layer aggregation strategy based on type attention mechanism is adopted for information transmission, and the multi-level information of nodes is effectively integrated to make the node feature representation richer and more accurate.

[0013] S140. An imbalanced loss function is used in the classifier for node classification tasks. The relationship-aware label propagation algorithm and the two-layer aggregation strategy are combined to form a multi-objective optimization problem, thereby achieving end-to-end training of the entire model.

[0014] Preferably, in S110, the linear transformation is defined as:

[0015]

[0016] in, , This indicates the node type in HIN. It represents the set of all node types. Indicates the node type is The original input features, Is a node type Learnable parameters, This represents the output characteristics after a linear transformation.

[0017] Preferably, in S120, given an initial label matrix The formula for iteratively propagating labels is as follows:

[0018]

[0019] in It is a layer pseudo-tags in It is an adjacency matrix. It is the corresponding degree matrix;

[0020] Furthermore, the process of generating pseudo-labels based on label propagation from relational subgraphs is as follows:

[0021]

[0022]

[0023] in, It is the node type of HIN. It is a relational subgraph The row normalized matrix, Indicates the type of the target node.

[0024] Preferably, in step S130, the two-level aggregation strategy includes micro-level aggregation and macro-level aggregation;

[0025] Among them, micro-level aggregation involves message passing and aggregation between nodes of the same type;

[0026] Among them, the macro-level aggregation utilizes the type attention mechanism to learn the weights of different types of neighbors of the target node, and performs information transfer and aggregation between different types of nodes.

[0027] Preferably, the aforementioned micro-level aggregation includes merging node features within the same relation, thereby creating a higher level of aggregation in the relation subgraph. Generate specific types of features;

[0028] Furthermore, we utilize simple mean aggregation to obtain specific type features of the central nodes and calculate the row normalization matrix. ,in It is a degree value matrix, where Representing relations The corresponding adjacency matrix;

[0029] Furthermore, heterogeneous graph convolution in relational subgraphs can be defined as:

[0030]

[0031] in, It is an instance of a relation type. It is a relation type in the subgraph Learnable parameters, yes The feature matrix of the type node.

[0032] Preferably, in the above-mentioned macro-level aggregation, the attention mechanism can be viewed as based on the query. , convert the key sequence Mapping to attention distribution These key sequences have one-to-one corresponding values. The definition of a specific type of attention is as follows:

[0033]

[0034]

[0035]

[0036] in The target node type is The hidden representation of the node, and The target node type is of The representation of neighboring nodes, to assess the importance of neighbor type representation, considers the type of the target node itself as the neighbor type, i.e. , These are the learnable parameters of the model. It is an activation function. It is by The attention coefficient calculated by the function;

[0037] Furthermore, the model uses the following formula to calculate the output characteristics of the target node after attention weighting:

[0038]

[0039] in It is a non-linear activation function. Is the target node type The neighbor set, output representation It is the input to the next layer block, and the final representation is generated by the last layer block.

[0040] Preferably, in S140, a single-layer linear transformation is used as the classifier, expressed as:

[0041]

[0042] in This is the final embedding matrix after classifier processing. Learn features from the output of the target type nodes in the last layer of the neural network. These are the learnable parameters in the classifier, used The function calculates the class probability of the target node and receives the final node representation from the heterogeneous message passing module, expressed as:

[0043]

[0044]

[0045] For the process Normalized target node label probability;

[0046] Furthermore, with The cross-entropy loss of the function and negative log-likelihood loss is:

[0047]

[0048] in This indicates the labeled subset of nodes. The target node The true label matrix, and The target node The predicted label distribution.

[0049] Preferably, for the label propagation algorithm, the loss function is calculated as the cross-entropy between the predicted label and the pseudo-label:

[0050]

[0051] in, The target node The pseudo-labels combine the above two steps into a multi-objective optimization problem, enabling end-to-end training of the entire model. Formally, the model is optimized using the following objective function:

[0052]

[0053] in, These are the balancing hyperparameters, which are further defined for classifying imbalanced nodes. , and as follows:

[0054]

[0055]

[0056]

[0057] in, , ,

[0058] as well as Is the type as The number of nodes. The output of the classifier will be used as the final prediction result of the model.

[0059] A heterogeneous graph node classification system based on relation-aware label propagation, which implements the above method, includes the following modules:

[0060] The heterogeneous linear transformation module is used to project the attribute features of heterogeneous nodes into a common low-dimensional feature space through specific types of linear transformations, ensuring that the feature vectors of all nodes have the same dimension.

[0061] The relation-aware label propagation module is used to generate pseudo-labels through a relation-aware mechanism, taking into account the relationships between nodes, and using a label propagation algorithm to propagate known label information to unlabeled nodes, thereby enhancing the classification performance of minority class nodes.

[0062] The heterogeneous message passing module is used to utilize the pseudo-tags generated by the relation-aware tag propagation module and works on the basis of the unified feature representation generated by the heterogeneous linear transformation module.

[0063] The classifier design module is used to classify or predict based on the final node representation.

[0064] The beneficial effects of this invention are as follows:

[0065] 1. Improves the classification performance of minority class nodes: By using a relation-aware label propagation method, pseudo-labels are generated and propagated to unlabeled nodes, thereby enhancing the label information of minority class nodes. Specifically, label propagation utilizes information from known labeled nodes and extends label information to nodes of the same and different types through a relation-aware mechanism, compensating for the lack of labels for minority class nodes in the training data and significantly improving the classification accuracy of minority class nodes.

[0066] 2. Capable of multi-level information aggregation: The heterogeneous message passing module employs a two-layer aggregation strategy, performing information passing and aggregation at both the micro-level (within-type) and the macro-level (between-type) respectively. In micro-level aggregation, nodes of the same type exchange information, ensuring the integrity of features within each node. In macro-level aggregation, a type attention mechanism is used to learn the weights of neighbors of different types, optimizing information fusion across types of nodes. This multi-level aggregation approach allows node features to represent their semantic information more comprehensively and accurately, improving the model's robustness and classification performance.

[0067] 3. Capable of handling data imbalance issues: Employs multiple imbalance strategies, including reweighting ( ),balance ( ) and post-compensation ( This addresses the issue of imbalanced class distribution. Specifically, the reweighting strategy adjusts the class weights in the loss function, ensuring that minority class nodes receive more attention during training; balancing... and subsequent compensation The strategies adjust the class probabilities at the classifier level to ensure a more balanced distribution of classes in the model's output. These strategies effectively mitigate the bias caused by class imbalance and improve overall classification performance. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0070] Figure 2 This is a schematic diagram of the module structure according to an embodiment of the present invention. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0072] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this invention should have the ordinary meaning understood by those skilled in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0073] like Figure 1 , Figure 2 As shown, a heterogeneous graph node classification method based on relation-aware label propagation includes the following steps:

[0074] S110. Based on the given heterogeneous network, perform type-specific linear transformations on nodes of different types and project their features onto a common feature space.

[0075] The Heterogeneous Linear Transformation module addresses the inherent heterogeneity of nodes in Heterogeneous Information Networks (HINs), where different types of nodes possess different semantic features and dimensions. Through specific types of linear transformations, the module projects the attribute features of heterogeneous nodes into a common low-dimensional feature space, ensuring that the feature vectors of all nodes have the same dimension.

[0076] Furthermore, the specific definition of heterogeneous linear transformation is:

[0077]

[0078] in, , This indicates the node type in HIN. It represents the set of all node types. Indicates the node type is The original input features, Is a node type Learnable parameters, This represents the output characteristics after a linear transformation.

[0079] The heterogeneous linear transformation module provides a unified feature representation for subsequent relation-aware label propagation and heterogeneous message passing, ensuring that features of different types of nodes can be effectively propagated and aggregated in the same feature space.

[0080] S120. For target nodes in heterogeneous networks, design a relation-aware label propagation algorithm and generate pseudo-labels through relation subgraphs.

[0081] The relation-aware label propagation module generates pseudo-labels through a relation-aware mechanism. It considers the relationships between nodes and uses a label propagation algorithm to spread known label information to unlabeled nodes, enhancing the classification performance of minority class nodes. This module works on the unified feature representation generated by the heterogeneous linear transformation module.

[0082] Furthermore, given an initial label matrix The formula for iteratively propagating labels is as follows:

[0083]

[0084] in It is a layer pseudo-tags in It is an adjacency matrix. It is the corresponding degree matrix;

[0085] Furthermore, the process of generating pseudo-labels based on label propagation from relational subgraphs is as follows:

[0086]

[0087]

[0088] in, It is the node type of HIN. It is a relational subgraph The row normalized matrix, Indicates the type of the target node.

[0089] This invention presents a relation-aware label propagation method that generates pseudo-labels and propagates them to unlabeled nodes by considering the relationships and connections between nodes. Compared with traditional label propagation methods, relation-aware label propagation can better utilize the complex relationships between nodes in heterogeneous graphs, especially minority class nodes, thereby improving the classification performance of minority class nodes.

[0090] S130. A two-layer aggregation strategy based on type attention mechanism is adopted for information transmission, and the multi-level information of nodes is effectively integrated to make the node feature representation richer and more accurate.

[0091] The heterogeneous message passing module includes micro-level (intra-type) aggregation and macro-level (inter-type) aggregation. Micro-level aggregation performs message passing and aggregation between nodes of the same type; macro-level aggregation utilizes a type attention mechanism to learn the weights of the target node's neighbors of different types, enabling information passing and aggregation between nodes of different types. This module uses pseudo-labels generated by the relation-aware label propagation module and operates based on the unified feature representation generated by the heterogeneous linear transformation module.

[0092] The aforementioned micro-level aggregation includes merging node features within the same relation, thereby creating a subgraph of the relation. Generate specific types of features;

[0093] Furthermore, we utilize simple mean aggregation to obtain specific type features of the central nodes and calculate the row normalization matrix. ,in It is a degree value matrix, where Representing relations The corresponding adjacency matrix;

[0094] Furthermore, heterogeneous graph convolution in relational subgraphs can be defined as:

[0095]

[0096] in, It is an instance of a relation type. It is a relation type in the subgraph Learnable parameters, yes The feature matrix of the type node.

[0097] Among them, the macro-level aggregation utilizes the type attention mechanism to learn the weights of different types of neighbors of the target node, and performs information transfer and aggregation between different types of nodes.

[0098] In the aforementioned macro-level aggregation, the attention mechanism can be viewed as based on the query... , convert the key sequence Mapping to attention distribution These key sequences have one-to-one corresponding values. The definition of a specific type of attention is as follows:

[0099]

[0100]

[0101]

[0102] in The target node type is The hidden representation of the node, and The target node type is of The representation of neighboring nodes, to assess the importance of neighbor type representation, considers the type of the target node itself as the neighbor type, i.e. , These are the learnable parameters of the model. It is an activation function. It is by The attention coefficient calculated by the function;

[0103] Furthermore, the model uses the following formula to calculate the output characteristics of the target node after attention weighting:

[0104]

[0105] in It is a non-linear activation function. Is the target node type The neighbor set, output representation It is the input to the next layer block, and the final representation is generated by the last layer block.

[0106] The proposed dual-layer aggregation strategy includes micro-level (intra-type) aggregation and macro-level (inter-type) aggregation. Micro-level aggregation facilitates information transfer between nodes of the same type, ensuring the integrity of intra-type features; macro-level aggregation utilizes type attention mechanisms to learn the weights of neighboring nodes of different types, thereby optimizing information fusion across types of nodes. This dual-layer aggregation strategy results in richer and more accurate node feature representations, improving the model's robustness and classification performance.

[0107] S140. An imbalanced loss function is used in the classifier for node classification tasks. The relationship-aware label propagation algorithm and the two-layer aggregation strategy are combined to form a multi-objective optimization problem, thereby achieving end-to-end training of the entire model.

[0108] The classifier design module performs classification or prediction based on the final node representation. A single-layer linear transformation is used as the classifier, represented as:

[0109]

[0110] in This is the final embedding matrix after classifier processing. Learn features from the output of the target type nodes in the last layer of the neural network. These are the learnable parameters in the classifier, used The function calculates the class probability of the target node and receives the final node representation from the heterogeneous message passing module, expressed as:

[0111]

[0112]

[0113] For the process Normalized target node label probability;

[0114] Furthermore, with The cross-entropy loss of the function and negative log-likelihood loss is:

[0115]

[0116] in This indicates the labeled subset of nodes. The target node The true label matrix, and The target node The predicted label distribution.

[0117] For the label propagation algorithm, the loss function is calculated as the cross-entropy between the predicted label and the pseudo-label:

[0118]

[0119] in, The target node The pseudo-labels combine the above two steps into a multi-objective optimization problem, enabling end-to-end training of the entire model. Formally, the model is optimized using the following objective function:

[0120]

[0121] in, These are the balancing hyperparameters, which are further defined for classifying imbalanced nodes. , and As follows:

[0122]

[0123]

[0124]

[0125] in, , ,

[0126] as well as Is the type as The number of nodes is determined, and the classifier's output will be used as the model's final prediction result.

[0127] To simultaneously optimize label propagation and classification tasks, this invention employs a multi-objective optimization strategy. A comprehensive loss function is designed by combining the label propagation loss and classification loss, enabling end-to-end model training. This multi-objective optimization strategy not only balances the weights of the two tasks but also improves the overall model performance.

[0128] To address the problem of classifying unbalanced nodes in heterogeneous graphs, this invention introduces a reweighting method. ),balance ( ) and post-compensation ( Imbalanced handling strategies include reweighting, which adjusts the class weights in the loss function to give more attention to minority class nodes during training; and balancing. and subsequent compensation The strategy adjusts the class probabilities at the classifier level to ensure a more balanced distribution of classes in the model output, thereby effectively addressing the problem of imbalanced class distribution and improving classification accuracy.

[0129] A heterogeneous graph node classification system based on relation-aware label propagation, which implements the above method, includes the following modules:

[0130] The heterogeneous linear transformation module is used to project the attribute features of heterogeneous nodes into a common low-dimensional feature space through specific types of linear transformations, ensuring that the feature vectors of all nodes have the same dimension.

[0131] The relation-aware label propagation module is used to generate pseudo-labels through a relation-aware mechanism, taking into account the relationships between nodes, and using a label propagation algorithm to propagate known label information to unlabeled nodes, thereby enhancing the classification performance of minority class nodes.

[0132] The heterogeneous message passing module is used to utilize the pseudo-tags generated by the relation-aware tag propagation module and works on the basis of the unified feature representation generated by the heterogeneous linear transformation module.

[0133] The classifier design module is used to classify or predict based on the final node representation.

[0134] In summary, this invention, by combining relation-aware label propagation and a two-layer aggregation strategy, effectively improves the performance and robustness of imbalanced node classification in heterogeneous graphs, overcoming the shortcomings of existing methods in handling imbalanced node classification in heterogeneous graphs. It has significant application value and broad application prospects. Other features and advantages of this invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.

[0135] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0136] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A heterogeneous graph node classification method based on relation-aware label propagation, characterized in that, Includes the following steps: S110. Based on the given heterogeneous network, perform type-specific linear transformations on nodes of different types. Nodes of different types have different semantic features and dimensions, and project their features into a common feature space. S120. For target nodes in heterogeneous networks, design a relation-aware label propagation algorithm and generate pseudo-labels through relation subgraphs. S130. A two-layer aggregation strategy based on type attention mechanism is adopted for information transmission, and the multi-level information of nodes is effectively integrated to make the node feature representation richer and more accurate. In step S130, the two-level aggregation strategy includes micro-level aggregation and macro-level aggregation; Among them, micro-level aggregation involves message passing and aggregation between nodes of the same type; Micro-level aggregation involves merging node features within the same relation, thereby creating a subgraph of the relation. Specific types of features are generated; furthermore, simple mean aggregation is used to obtain specific types of features of the central nodes, and the row normalization matrix is ​​calculated. ,in It is a degree value matrix, where Representing relations The corresponding adjacency matrix; Furthermore, heterogeneous graph convolution in relational subgraphs is defined as: in, It is an instance of a relation type. It is a relation type in the subgraph Learnable parameters yes Feature matrix of type nodes; Among them, the macro-level aggregation utilizes the type attention mechanism to learn the weights of different types of neighbors of the target node, and performs information transfer and aggregation between different types of nodes; In macro-level aggregation, the attention mechanism is viewed as being based on the query. , convert the key sequence Mapping to attention distribution These key sequences have one-to-one corresponding values. The definition of a specific type of attention is as follows: in The target node type is The hidden representation of the node, and The target node type is of The representation of neighboring nodes, to assess the importance of neighbor type representation, considers the type of the target node itself as the neighbor type, i.e. , These are the learnable parameters of the model. It is an activation function. It is by The attention coefficient calculated by the function; Furthermore, the model uses the following formula to calculate the output characteristics of the target node after attention weighting: in It is a non-linear activation function. Is the target node type The neighbor set, output representation It is the input to the next layer block, and the final representation is generated by the last layer block; S140. An imbalanced loss function is used in the classifier for node classification tasks. The relationship-aware label propagation algorithm and the two-layer aggregation strategy are combined to form a multi-objective optimization problem, thereby achieving end-to-end training of the entire model.

2. The heterogeneous graph node classification method based on relation-aware label propagation according to claim 1, characterized in that, In S110, the linear transformation is defined as: in, , This indicates the node type in HIN. It represents the set of all node types. Indicates the node type is The original input features, Is a node type Learnable parameters This represents the output characteristics after a linear transformation.

3. The heterogeneous graph node classification method based on relation-aware label propagation according to claim 2, characterized in that, In S120, given the initial label matrix The formula for iteratively propagating labels is as follows: in It is a layer pseudo-tags in It is an adjacency matrix. It is the corresponding degree matrix; Furthermore, the process of generating pseudo-labels based on label propagation from relational subgraphs is as follows: in, It is the node type of HIN. It is a relational subgraph The row normalized matrix, Indicates the type of the target node.

4. The heterogeneous graph node classification method based on relation-aware label propagation according to claim 3, characterized in that, In S140, a single-layer linear transformation is used as the classifier, as shown below: in This is the final embedding matrix after classifier processing. Learn features from the output of the target type nodes in the last layer of the neural network. These are the learnable parameters in the classifier, used The function calculates the class probability of the target node and receives the final node representation from the heterogeneous message passing module, expressed as: For the process Normalized target node label probability; Furthermore, with The cross-entropy loss of the function and negative log-likelihood loss is: in This indicates the labeled subset of nodes. The target node The true label matrix, and The target node The predicted label distribution.

5. A heterogeneous graph node classification method based on relation-aware label propagation according to claim 4, characterized in that, For the label propagation algorithm, the loss function is calculated as the cross-entropy between the predicted label and the pseudo-label: in, The target node The pseudo-labels, combined with the above steps, form a multi-objective optimization problem, enabling end-to-end training of the entire model. Formally, the model is optimized using the following objective function: in, These are the balancing hyperparameters, which are further defined for classifying imbalanced nodes. , and as follows: in, , , as well as Is the type as The number of nodes is determined, and the classifier's output will be used as the model's final prediction result.

6. A heterogeneous graph node classification system based on relation-aware label propagation, the system being used to implement the method described in any one of claims 1-5, characterized in that: Includes the following modules: Since different types of nodes have different semantic features and dimensions, the heterogeneous linear transformation module is used to project the attribute features of heterogeneous nodes into a common low-dimensional feature space through a specific type of linear transformation, ensuring that the feature vector dimensions of all nodes are the same. The relation-aware label propagation module is used to generate pseudo-labels through a relation-aware mechanism, taking into account the relationships between nodes, and using a label propagation algorithm to propagate known label information to unlabeled nodes, thereby enhancing the classification performance of minority class nodes. The heterogeneous message passing module is used to utilize the pseudo-tags generated by the relation-aware tag propagation module and works on the basis of the unified feature representation generated by the heterogeneous linear transformation module. The classifier design module is used to classify or predict based on the final node representation.

Images