An attribute completion and category balance integrated heterogeneous graph representation learning method
By employing a higher-order augmented class balancing and self-supervised heterogeneous attribute completion mechanism, the problems of attribute missingness and class imbalance in heterogeneous graphs are solved, improving the quality of node representation and classification performance, especially the performance of minority class nodes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INNER MONGOLIA UNIVERSITY
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-05
AI Technical Summary
Existing heterogeneous graph neural networks suffer from low representation accuracy and degraded classification performance when dealing with missing attributes and class imbalance. In particular, when missing attributes are coupled with class imbalance, existing methods have failed to effectively address the cross-class imbalance propagation effect.
We employ a high-order augmented class balancing mechanism and a self-supervised heterogeneous attribute completion mechanism. We learn node embeddings through random walks and skip word models, combine masked autoencoders and attention mechanisms to complete attributes and alleviate class imbalance, and finally learn node representations through heterogeneous graph neural networks.
It significantly improves the representation quality and classification accuracy of minority class nodes, solves the cross-type imbalance propagation effect, and enhances the performance of node classification and clustering tasks.
Smart Images
Figure CN122154756A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a heterogeneous graph representation learning method based on the integration of attribute completion and category balancing. Background Technology
[0002] In engineering applications, many systems can be represented using graph structures, such as social network analysis, recommender systems, bioinformatics, and knowledge graphs. Heterogeneous graphs, in particular, are widely used because they can contain various types of nodes and relationships, enabling more comprehensive feature representation. Heterogeneous graph neural networks are an effective method for learning heterogeneous graphs, capable of modeling complex graph structures and nodes to obtain node representations for downstream tasks.
[0003] Traditional heterogeneous graph neural networks (HGNNs) largely rely on complete node attributes, using message aggregation mechanisms to learn node representations. This approach often exhibits significant scenario dependence, with stringent requirements for attribute completeness and class distribution balance. For heterogeneous graphs with missing attributes, this can easily interfere with the attribute propagation process in HGNNs, reducing the accuracy of node representations. For class-imbalanced heterogeneous graphs, minority class nodes are prone to being misclassified as majority class nodes, degrading overall classification performance.
[0004] With the increasing demands for adaptability to complex scenarios and accuracy in representation, breakthroughs have been achieved in handling single problems. While attribute completion methods for missing attributes can reconstruct them using topological information or multi-view fusion, they are prone to biased representations of minority class nodes because they do not consider the impact of class imbalance. Methods addressing class imbalance all rely on the ideal assumption of complete node attributes and only focus on intra-class class imbalance, neglecting the propagation of imbalance across classes.
[0005] Studies have shown that attribute missing and class imbalance have a significant mutual amplification effect in heterogeneous graphs: attribute missing weakens the feature representation ability of minority class nodes, further exacerbating class bias; while class imbalance introduces biased guidance into the attribute completion process, causing the completion results to tilt towards majority class features. Summary of the Invention
[0006] In view of this, the purpose of this application is to provide a heterogeneous graph representation learning method based on the integration of attribute completion and class balance, which can simultaneously solve the problems of attribute missing and class imbalance in the heterogeneous graph learning process.
[0007] This application provides a heterogeneous graph representation learning method based on the integration of attribute completion and class balancing, including: A higher-order augmented class balancing mechanism is used to represent minority class nodes in a heterogeneous graph to obtain a class-balanced heterogeneous graph. Based on the self-supervised heterogeneous attribute completion mechanism, a masked autoencoder is introduced to learn better attribute embeddings for attribute-observable nodes, and an attention mechanism is introduced to complete attributes for nodes with missing attributes. The completed heterogeneous graph is then input into the heterogeneous graph neural network to learn the final node embeddings in order to perform node classification or node clustering tasks.
[0008] Furthermore, the method of representing minority class nodes in heterogeneous graphs using a higher-order enhanced class balancing mechanism includes: The node embeddings are learned using a classic method based on meta-path random walks and skip character models, resulting in the node topological embedding H:
[0009] In the formula, express function, These correspond to the node set, edge set, node type set, edge type set, and meta-path scheme, respectively. These represent the number of visits, neighborhood size, and visit length for each node, respectively. Given the set of minority class nodes and oversampling ratio is Then the node category The number of synthesis nodes is Indicates node category Quantity; To alleviate class imbalance, a dual-candidate set mechanism is used to obtain higher-order neighborhood information of minority class nodes, and then the embeddings of minority class nodes and edges are synthesized based on the higher-order neighborhood information.
[0010] Furthermore, the step of obtaining higher-order neighborhood information of minority class nodes based on the dual-candidate set mechanism, and then synthesizing the embeddings of minority class nodes and edges based on the higher-order neighborhood information, includes: Based on node topology embedding H, calculate minority class nodes. Its first-order neighbor nodes Cosine similarity:
[0011] In the formula, They are nodes The topological embedding vector, express The set of first-order neighbor nodes; The top M nodes with the highest similarity are selected to form the main candidate set. :
[0012] In the primary candidate set Introducing a higher-order mechanism in each candidate node The set of second-order neighbor nodes Random sampling is performed to construct an auxiliary candidate set. :
[0013] In the formula, This represents the union operation. This indicates a random sampling function without replacement. This represents the random sampling ratio; main candidate set and auxiliary candidate set The topological embedding vectors of the middle nodes are linearly aggregated to synthesize the embeddings of the minority class nodes. :
[0014] In the formula, To balance the weights, For nodes Topological embedding vector; Simultaneously, for each synthesis node, corresponding edges are generated by connecting selected nodes in its primary candidate set and auxiliary candidate set. :
[0015] Furthermore, the step of learning better attribute embeddings for attribute-observable nodes by introducing a masked autoencoder includes: With probability For observable properties Randomly set a zero mask to obtain The data is then input into a GAE with RGCN as the encoder, and the latent attribute representation of the observable nodes is calculated through relation-specific transformation matrices.
[0016] In the formula, For attribute observable nodes The mask attribute vector, This represents the activation function. It is a node In relation types The following is a collection of neighbors. Represents a node In the Layer characterization, For the first Hierarchical Relationship Type The corresponding weight matrix, For the first Layer self-connection weight matrix, This is a normalization constant; The latent attribute representations of all observable nodes are denoted as:
[0017] A multilayer perceptron (MLP) is used as the decoder to decode the latent attribute representations and obtain the reconstructed attributes. :
[0018] In the formula, For MLP decoders, Its parameters; By jointly optimizing attribute reconstruction loss and edge reconstruction loss, the potential attribute representation of attribute-observable nodes is improved. Optimize.
[0019] Furthermore, the attribute reconstruction loss is used to constrain the learning process of discriminative attribute information, specifically including: Attribute reconstruction loss is calculated based on normalized cosine similarity. :
[0020] In the formula, They are nodes Reconstructed attributes and original attributes, A set of nodes representing observable attributes; Furthermore, the edge reconstruction loss is used to preserve the topological relationships between nodes, specifically including: Calculate edge reconstruction loss based on mean square error. :
[0021] In the formula, The number of nodes with observable attributes. This is the adjacency matrix predicted for nodes containing observable attributes. Let v be its true adjacency matrix, where the subscript v, indicates that the (v)th row of the matrix is selected.
[0022] Furthermore, the method of introducing an attention mechanism to complete attributes for nodes with missing attributes includes: For nodes with missing attributes Obtain its first-order neighbor set. Based on nodes with missing attributes Its neighboring nodes Topological embedding vector Calculate the attention scores of both. :
[0023] In the formula, For learnable weight matrix, For activation function, ; Using the softmax function to obtain neighboring nodes nodes with missing attributes Normalized attention weights :
[0024] Employing a multi-head attention mechanism for nodes with missing attributes neighboring nodes Aggregate the attributes to obtain nodes with missing attributes. Attribute representation:
[0025] In the formula, Indicates to The average of the aggregated results of each attention head is taken. For neighboring nodes The mask attribute vector; The attribute representation for completing all missing attribute nodes is denoted as:
[0026] In the formula, This represents the number of nodes with missing attributes.
[0027] Furthermore, the step of inputting the completed heterogeneous graph into the heterogeneous graph neural network to learn the final node embedding includes: The optimal attribute representation of observable nodes and the attribute representation of missing nodes are fused together to obtain the final attribute representation of the nodes. This final attribute representation is then fed into the MAGNN model along with the topology of the heterogeneous graph to learn the final node embedding and predict the node category. An overall loss function is constructed based on prediction loss, attribute reconstruction loss, and edge reconstruction loss to optimize the node representation learning process.
[0028] Furthermore, the prediction loss is used to optimize the node classification task, specifically including:
[0029] In the formula, For loss function, For the predicted node category, Represents the actual node labels; The node category is predicted using the following methods. :
[0030] In the formula, This represents the topological structure of a heterogeneous graph. This represents the final attribute representation of the node; The overall loss function is defined as:
[0031] In the formula, These are the weighting coefficients for each loss term.
[0032] Furthermore, when the method is applied to academic network scenarios, the node types of the heterogeneous graph include: papers, authors, and disciplines. Among them, discipline nodes are a typical minority class, and paper nodes have native attributes, while author and discipline nodes have missing attributes.
[0033] Furthermore, when the method is applied to a movie recommendation scenario, the node types of the heterogeneous graph include: movie, director, and actor. Among them, the director node is a typical minority class and has native attributes, while the movie and actor nodes have missing attributes.
[0034] The heterogeneous graph representation learning method based on attribute completion and class balancing provided in this application has the following beneficial effects: it solves the technical defects of existing technologies in handling cross-type class imbalance problems in heterogeneous graphs, overcomes the representation bias and classification accuracy decline caused by the cross-type imbalance propagation effect; it solves the defects of existing technologies in ignoring attribute missing problems in class imbalance heterogeneous graphs, breaks the negative coupling between attribute missing and class imbalance, and improves the representation quality of minority class nodes. Attached Figure Description
[0035] Figure 1 The flowchart illustrates a heterogeneous graph representation learning method based on the integration of attribute completion and category balancing provided in an embodiment of this application. Detailed Implementation
[0036] To make the objectives, technical solutions, and advantages of this technical solution clearer, the following detailed description, in conjunction with specific embodiments, further illustrates this technical solution. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this technical solution.
[0037] First, we introduce the application scenarios of this application. This application is applicable to heterogeneous graph scenarios with class imbalance and missing attributes, including academic network scenarios and movie recommendation scenarios.
[0038] Please see as follows Figure 1 The flowchart shown is for a heterogeneous graph representation learning method based on the integration of attribute completion and class balancing. Figure 1 As shown, the method includes; S101. Use a higher-order enhanced class balancing mechanism to represent the minority class nodes in the heterogeneous graph to obtain a class-balanced heterogeneous graph.
[0039] In practical implementation, the minority class nodes in the heterogeneous graph are represented using a higher-order augmented class balancing mechanism in the following way: Step 1011: Learn the node embeddings using the classic method of random walk and skipping word model based on metapaths, and obtain the node topological embedding H:
[0040] In the formula, express function, These correspond to the node set, edge set, node type set, edge type set, and meta-path scheme, respectively. These represent the number of visits, neighborhood size, and visit length for each node, respectively. Given the set of minority class nodes and oversampling ratio is Then the composite node of the node category The quantity is; Indicates node category Quantity; Step 1012: To alleviate class imbalance, high-order neighborhood information of minority class nodes is obtained based on a dual-candidate set mechanism, and then the embeddings of minority class nodes and edges are synthesized based on the high-order neighborhood information. Specifically, this includes: Based on node topology embedding H, calculate minority class nodes. Its first-order neighbor nodes Cosine similarity:
[0041] In the formula, They are nodes The topological embedding vector, express The set of first-order neighbor nodes; The top M nodes with the highest similarity are selected to form the main candidate set. :
[0042] In the primary candidate set Introducing a higher-order mechanism in each candidate node The set of second-order neighbor nodes Random sampling is performed to construct an auxiliary candidate set. :
[0043] In the formula, This represents the union operation. This indicates a random sampling function without replacement. This represents the random sampling ratio; main candidate set and auxiliary candidate set The topological embedding vectors of the middle nodes are linearly aggregated to synthesize the embeddings of the minority class nodes. :
[0044] In the formula, To balance the weights, For nodes Topological embedding vector; Simultaneously, for each synthesis node, corresponding edges are generated by connecting selected nodes in its primary candidate set and auxiliary candidate set. :
[0045] S102. Based on the self-supervised heterogeneous attribute completion mechanism, a masked autoencoder is introduced to learn better attribute embeddings for attribute-observable nodes, and an attention mechanism is introduced to complete attributes for nodes with missing attributes.
[0046] In practical implementation, the step of introducing a masked autoencoder to learn better attribute embeddings for attribute-observable nodes specifically includes: Step 1021, with probability For observable properties Randomly set a zero mask to obtain The data is then input into a GAE with RGCN as the encoder, and the latent attribute representation of the observable nodes is calculated through relation-specific transformation matrices.
[0047] In the formula, For attribute observable nodes The mask attribute vector, This represents the activation function. It is a node In relation types The following is a collection of neighbors. Represents a node In the Layer characterization, For the first Hierarchical Relationship Type The corresponding weight matrix, For the first Layer self-connection weight matrix, This is a normalization constant; The latent attribute representations of all observable nodes are denoted as:
[0048] A multilayer perceptron (MLP) is used as the decoder to decode the latent attribute representations and obtain the reconstructed attributes. :
[0049] In the formula, For MLP decoders, Its parameters; By jointly optimizing attribute reconstruction loss and edge reconstruction loss, the potential attribute representation of attribute-observable nodes is improved. Optimize.
[0050] The attribute reconstruction loss is used to constrain the learning process of discriminative attribute information, and specifically includes: Attribute reconstruction loss is calculated based on normalized cosine similarity. :
[0051] In the formula, They are nodes Reconstructed attributes and original attributes, A set of nodes representing observable attributes; Furthermore, the edge reconstruction loss is used to preserve the topological relationships between nodes, specifically including: Calculate edge reconstruction loss based on mean square error. :
[0052] In the formula, The number of nodes with observable attributes. This is the adjacency matrix predicted for nodes containing observable attributes. Let v be its true adjacency matrix, where the subscript v, indicates that the (v)th row of the matrix is selected.
[0053] In practical implementation, the method of introducing an attention mechanism to complete the attributes of nodes with missing attributes specifically includes: For nodes with missing attributes Obtain its first-order neighbor set. Based on nodes with missing attributes Its neighboring nodes Topological embedding vector Calculate the attention scores of both. :
[0054] In the formula, For learnable weight matrix, For activation function, ; Using the softmax function to obtain neighboring nodes nodes with missing attributes Normalized attention weights :
[0055] Employing a multi-head attention mechanism for nodes with missing attributes neighboring nodes Aggregate the attributes to obtain nodes with missing attributes. Attribute representation:
[0056] In the formula, Indicates to The average of the aggregated results of each attention head is taken. For neighboring nodes The mask attribute vector; The attribute representation for completing all missing attribute nodes is denoted as:
[0057] In the formula, This represents the number of nodes with missing attributes.
[0058] S103. Input the completed heterogeneous graph into the heterogeneous graph neural network to learn the final node embedding, so as to perform node classification task or node clustering task.
[0059] In practice, the completed heterogeneous graph is input into the heterogeneous graph neural network to learn the final node embeddings in the following way: Step 1031: Fuse the optimal attribute representation of the observable node and the attribute representation of the missing node to obtain the final attribute representation of the node. Then, input the attribute representation of the node and the topology of the heterogeneous graph into the MAGNN model to learn the final node embedding in order to predict the node category.
[0060] Step 1032: Construct an overall loss function based on prediction loss, attribute reconstruction loss, and edge reconstruction loss to optimize the node representation learning process.
[0061] The prediction loss is used to optimize the node classification task, and specifically includes:
[0062] In the formula, For loss function, For the predicted node category, Represents the actual node labels; The node category is predicted using the following methods. :
[0063] In the formula, This represents the topological structure of a heterogeneous graph. This represents the final attribute representation of the node; The overall loss function is defined as:
[0064] In the formula, These are the weighting coefficients for each loss term.
[0065] Next, the application of the method described in this application will be introduced: Application Example 1: When the method is applied to academic network scenarios, the node types of the heterogeneous graph include: papers, authors, and disciplines. Among them, discipline nodes are a typical minority class, and paper nodes have native attributes, while author and discipline nodes have missing attributes.
[0066] In academic network scenarios, heterogeneous graphs commonly suffer from a coupling problem of missing attributes and class imbalance: on the one hand, nodes corresponding to popular research topics account for a significantly higher proportion in number, resulting in a clear skewed class distribution; on the other hand, some nodes (such as author nodes) often have missing attributes due to data collection limitations. Taking the classic ACM dataset as an example, this dataset contains three types of nodes: papers (P), authors (A), and subjects (S). Among them, subject (S) nodes are a typical minority class, and only paper nodes have native attributes, while author (A) and subject (S) nodes both have missing attributes. To address this issue, this application uses a higher-order augmented class balancing module to synthesize and sample minority class S nodes to alleviate class imbalance, and then uses a heterogeneous attribute completion module to accurately complete the missing attributes of A and S nodes; subsequently, the completed heterogeneous graph is input into an HGNN model for node representation learning, ultimately achieving significant performance improvements in both node classification and clustering tasks.
[0067] Application Example 2: When the method is applied to a movie recommendation scenario, the node types of the heterogeneous graph include: movie, director, and actor. Among them, the director node is a typical minority class and has native attributes, while the movie and actor nodes have missing attributes.
[0068] Here, in movie recommendation systems, missing attributes and class imbalance are also prevalent. For example, in the IMDB dataset, the graph consists of three types of nodes: Movie (M), Director (D), and Actor (A). The number of Director nodes is relatively small, and only Movie nodes contain native attributes, leading to uneven distribution of feature information and susceptibility of node representations to bias. To alleviate these problems, we first utilize a higher-order augmented class balancing module to oversample minority class nodes D, generating semantically consistent synthetic nodes to mitigate the class imbalance of Director nodes. Subsequently, a heterogeneous attribute completion module is used to complete the missing attributes of Movie nodes M and Actor nodes, resulting in a feature-complete heterogeneous graph. Finally, the completed graph is input into a Heterogeneous Graph Neural Network (HGNN) to learn node representations and perform node classification and clustering tasks. Experimental results show that this method achieves significant improvements on multiple evaluation metrics.
[0069] Comparative experiment: The ACM, DBLP, and IMDB datasets were selected, and the dataset composition is shown in Table 1. Comparative experiments were conducted on HGNN-AC, HetReGAT-FC, HeGAE-AC, and the method proposed in this application (named GraphHOAC).
[0070] Table 1. Dataset Composition
[0071] For the node classification task, Table 2 shows the performance results of the graph neural network oriented towards attribute missing, and Table 3 shows the performance results of the graph neural network oriented towards class imbalance.
[0072] Table 2. Performance results of graph neural networks for node classification tasks with missing attributes.
[0073] As shown in Table 2, GraphHOAC significantly outperforms the baseline attribute completion model. For example, on the ACM dataset, compared to the suboptimal results, GraphHOAC improves the Macro-F1 and Micro-F1 scores by 0.3% and 0.22%, respectively; on the DBLP dataset, these two scores improve by 0.04% and 0.03%, respectively; and on the IMDB dataset, the Macro-F1 and Micro-F1 scores improve by 0.01% and 0.17%, respectively. This is because the proposed method synthesizes a small number of nodes with rich information quality through a higher-order augmented class balancing module, significantly improving the quality of information transmitted to the target node, thereby effectively improving classification performance.
[0074] Table 3. Performance results of graph neural networks for class imbalance on node classification tasks.
[0075] As shown in Table 3, GraphHOAC demonstrates significant advantages over baseline models specifically designed for class imbalance. Taking the ACM dataset as an example, GraphHOAC outperforms the best baseline model by 3.00% and 3.04% on the Macro-F1 and Micro-F1 metrics, respectively. This is attributed to the method's method of filling in missing attributes before addressing class imbalance, thereby obtaining richer attribute information.
[0076] For the node clustering task, we obtained the graph neural network clustering results for attribute-missing tasks as shown in Table 4 and the graph neural network clustering results for class imbalance tasks as shown in Table 5.
[0077] Table 4. Clustering results of graph neural networks for missing attributes
[0078] As shown in Table 4, GraphHOAC equipped with the attribute completion module outperforms the version without it. For example, on the ACM dataset, GraphHOAC improves the Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) metrics by 0.62% and 2.82%, respectively, compared to the suboptimal results. This indicates that the improved class balancing module enables target nodes to learn more discriminative node representations compared to baseline models designed for attribute-missing graphs.
[0079] Table 5. Clustering results of graph neural networks for class imbalance
[0080] As shown in Table 5, GraphHOAC equipped with the class imbalance module outperforms the model without it. For example, on the ACM dataset, compared to the suboptimal results, the proposed method improves the Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI) metrics by 2.63% and 5.10%, respectively; on the DBLP dataset, these two metrics are improved by 3.97% and 3.02%, respectively. This indicates that, compared to the baseline model designed for the class imbalance problem, the proposed self-supervised heterogeneous attribute completion module can learn more informative node representations by completing missing attributes.
[0081] The above content is only a preferred embodiment of the present invention. For those skilled in the art, many changes can be made in the specific implementation and application scope based on the ideas of the present invention. As long as these changes do not depart from the concept of the present invention, they all fall within the protection scope of the present invention.
Claims
1. A heterogeneous graph representation learning method based on the integration of attribute completion and class balancing, characterized in that, The method includes: A higher-order augmented class balancing mechanism is used to represent minority class nodes in a heterogeneous graph to obtain a class-balanced heterogeneous graph. Based on the self-supervised heterogeneous attribute completion mechanism, a masked autoencoder is introduced to learn better attribute embeddings for attribute-observable nodes, and an attention mechanism is introduced to complete attributes for nodes with missing attributes. The completed heterogeneous graph is then input into the heterogeneous graph neural network to learn the final node embeddings in order to perform node classification or node clustering tasks.
2. The method as described in claim 1, characterized in that, The method of using a higher-order augmented class balancing mechanism to represent minority class nodes in heterogeneous graphs includes: The node embeddings are learned using a classic method based on metapath-based random walks and skip character models, resulting in node topological embeddings. H : In the formula, express function, These correspond to the node set, edge set, node type set, edge type set, and meta-path scheme, respectively. These represent the number of visits, neighborhood size, and visit length for each node, respectively. Given the minority class node set as and oversampling ratio is Then the node category The number of synthesis nodes is ;in, Indicates node category Quantity; To alleviate class imbalance, a dual-candidate set mechanism is used to obtain higher-order neighborhood information of minority class nodes, and then the embeddings of minority class nodes and edges are synthesized based on the higher-order neighborhood information.
3. The method as described in claim 2, characterized in that, The process of obtaining high-order neighborhood information of minority class nodes based on a dual-candidate set mechanism, and then synthesizing the embeddings of minority class nodes and edges based on the high-order neighborhood information, includes: Based on node topology embedding H Calculate minority class nodes Its first-order neighbor nodes Cosine similarity: In the formula, They are nodes The topological embedding vector, , express The set of first-order neighbor nodes; Select the front with high similarity M The nodes constitute the primary candidate set. : In the primary candidate set Introducing a higher-order mechanism in each candidate node The set of second-order neighbor nodes Random sampling is performed to construct an auxiliary candidate set. : In the formula, This represents the union operation. This indicates a random sampling function without replacement. This represents the random sampling ratio; main candidate set and auxiliary candidate set The topological embedding vectors of the middle nodes are linearly aggregated to synthesize the embeddings of the minority class nodes. : In the formula, To balance the weights, For nodes Topological embedding vector; Simultaneously, for each synthesis node, corresponding edges are generated by connecting selected nodes in its primary candidate set and auxiliary candidate set. : 。 4. The method as described in claim 1, characterized in that, The method of introducing a masked autoencoder to learn better attribute embeddings for attribute-observable nodes includes: With probability For observable properties Randomly set a zero mask to obtain The data is then input into a GAE with RGCN as the encoder, and the latent attribute representation of the observable nodes is calculated through relation-specific transformation matrices. ； In the formula, For attribute observable nodes The mask attribute vector, This represents the activation function. It is a node In relation types The following is a collection of neighbors. Represents a node In the Layer representation, For the first Hierarchical Relationship Type The corresponding weight matrix, For the first Layer self-connection weight matrix, This is a normalization constant; The latent attribute representations of all observable nodes are denoted as: A multilayer perceptron (MLP) is used as the decoder to decode the latent attribute representations and obtain the reconstructed attributes. : In the formula, For MLP decoders, Its parameters; By jointly optimizing attribute reconstruction loss and edge reconstruction loss, the potential attribute representation of attribute-observable nodes is improved. Optimize.
5. The method as described in claim 4, characterized in that, The attribute reconstruction loss is used to constrain the learning process of discriminative attribute information, specifically including: Attribute reconstruction loss is calculated based on normalized cosine similarity. : In the formula, They are nodes Reconstructed attributes and original attributes, A set of nodes representing observable attributes; Furthermore, the edge reconstruction loss is used to preserve the topological relationships between nodes, specifically including: Calculate edge reconstruction loss based on mean square error. : In the formula, The number of nodes with observable attributes. This is the adjacency matrix predicted for nodes containing observable attributes. Let v be its true adjacency matrix, where the subscript v, indicates that the (v)th row of the matrix is selected.
6. The method as described in claim 1, characterized in that, The method of introducing an attention mechanism to complete attributes for nodes with missing attributes includes: For nodes with missing attributes Obtain its first-order neighbor set. Based on nodes with missing attributes Its neighboring nodes Topological embedding vector Calculate the attention scores of both. : In the formula, For learnable weight matrix, For activation function, Using the softmax function to obtain neighboring nodes nodes with missing attributes Normalized attention weights : Employing a multi-head attention mechanism for nodes with missing attributes neighboring nodes Aggregate the attributes to obtain nodes with missing attributes. Attribute representation: In the formula, Indicates to The average of the aggregated results of each attention head is taken. Neighboring nodes The mask attribute vector; The attribute representation for completing all missing attribute nodes is denoted as: In the formula, This represents the number of nodes with missing attributes.
7. The method as described in claim 1, characterized in that, The step of inputting the completed heterogeneous graph into the heterogeneous graph neural network to learn the final node embedding includes: The optimal attribute representation of observable nodes and the attribute representation of missing nodes are fused together to obtain the final attribute representation of the nodes. This final attribute representation is then fed into the MAGNN model along with the topology of the heterogeneous graph to learn the final node embedding and predict the node category. An overall loss function is constructed based on prediction loss, attribute reconstruction loss, and edge reconstruction loss to optimize the node representation learning process.
8. The method as described in claim 7, characterized in that, The prediction loss is used to optimize the node classification task, specifically including: In the formula, For loss function, For the predicted node category, Represents the actual node labels; The node category is predicted using the following methods. : In the formula, This represents the topological structure of a heterogeneous graph. This represents the final attribute representation of the node; The overall loss function is defined as: In the formula, These are the weighting coefficients for each loss term.
9. The method as described in claim 1, characterized in that, When the method is applied to academic network scenarios, the node types of the heterogeneous graph include: papers, authors, and disciplines. Among them, discipline nodes are a typical minority class, and paper nodes have native attributes, while author and discipline nodes have missing attributes.
10. The method as described in claim 1, characterized in that, When the method is applied to a movie recommendation scenario, the node types of the heterogeneous graph include: movie, director, and actor. Among them, the director node is a typical minority class and has native attributes, while the movie and actor nodes have missing attributes.