Heterogeneous graph node classification method based on relation perception label propagation

Abstract

The invention relates to the technical field of graph neural network and heterogeneous graph representation learning, in particular to a heterogeneous graph node classification method based on relation perception label propagation, which comprises the following steps of: firstly, according to a given heterogeneous network, carrying out type-specific linear transformation on different types of nodes, and then carrying out type-specific linear transformation on different types of nodes; the method comprises the following steps of: designing a relation sensing label propagation algorithm for target nodes in a heterogeneous network, generating a pseudo label through a relation sub-graph, performing information transmission by adopting a double-layer aggregation strategy based on a type attention mechanism, and effectively fusing multi-layer information of the nodes, so as to obtain a multi-layer information fusion model; and the node feature representation is richer and more accurate. According to the method, the pseudo labels are obtained by designing a relation perception label propagation method, heterogeneous message passing is carried out by utilizing double-layer aggregation based on a type attention mechanism, and the classification performance and robustness are remarkably improved in combination with a multi-objective optimization strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of graph neural networks and heterogeneous graph representation learning, and in particular to a heterogeneous graph node classification method based on relationship-aware label propagation. Background Art

[0002] In recent years, although some studies have begun to focus on the imbalance problem in homogeneous graphs, due to the inherent complexity and heterogeneity of heterogeneous graphs, there are relatively few studies on imbalanced node classification in heterogeneous graphs. The existing heterogeneous graph neural network methods mainly include the following: HGNN methods based on neighborhood aggregation, such as GraphSAGE and GAT, learn node representations by aggregating node neighborhood information. However, these methods often ignore the characteristics of minority class nodes when dealing with imbalanced data, resulting in poor classification performance. HGNN methods based on meta-paths, such as HIN2Vec and Metapath2Vec, learn node representations in heterogeneous graphs using predefined meta-paths. Although these methods can capture structural information in heterogeneous graphs, they still have limitations when facing the problem of imbalanced node classification. Label propagation is a classic semi-supervised learning method that propagates the information of known label nodes to unlabeled nodes. For example, the traditional label propagation method (LP) has a certain effect in homogeneous graphs, but has limited effect when dealing with heterogeneous graphs.

[0003] In order to further improve the classification performance of unbalanced nodes in heterogeneous graphs, researchers have proposed some new methods. For example, relationship-aware label propagation (RLP) is an emerging method that considers the relationship 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 information and structural information in heterogeneous graphs. To solve the above problems, some studies have introduced a two-layer aggregation strategy, which can better handle the information transfer and aggregation problems in heterogeneous graphs by aggregating at the micro level and the macro level. For example, some attention-based methods, such as GAT, although they perform well in homogeneous graphs, are still limited in their application in heterogeneous graphs. In summary, existing methods have the following shortcomings in dealing with the problem of unbalanced node classification in heterogeneous graphs:

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

[0005] (2) Underutilization of heterogeneous graph information: Many methods do not fully utilize the type information and structure information in heterogeneous graphs.

[0006] (3) Methods limited to homogeneous graphs: Some traditional label propagation and attention mechanism methods are effective in homogeneous graphs, but have limited performance in heterogeneous graphs.

[0007] In order to solve the problem of unbalanced node classification in heterogeneous graphs, a heterogeneous graph node classification method based on relationship-aware label propagation was proposed. Summary of the invention

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

[0009] Based on the above purpose, the present invention provides a heterogeneous graph node classification method based on relationship-aware label propagation, comprising the following steps:

[0010] S110, according to a given heterogeneous network, performing type-specific linear transformations on different types of nodes, and projecting their features into a common feature space;

[0011] S120, for target nodes in heterogeneous networks, design a relationship-aware label propagation algorithm and generate pseudo labels through relationship subgraphs;

[0012] S130, adopt a two-layer aggregation strategy based on type attention mechanism to transfer information, and effectively integrate the multi-level information of nodes to make the node feature representation richer and more accurate;

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

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

[0015] H o =X o Θ o , o∈O

[0016] Among them, o∈O, o represents the node type in HIN, O represents the set of all node types, X o represents the original input feature of node type o, Θ o is a learnable parameter of node type o, H o Represents the output features after linear transformation.

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

[0018]

[0019] in are the pseudo labels in layer l, A is the adjacency matrix, and D is the corresponding degree matrix;

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

[0021]

[0022]

[0023] Among them, s∈O is the node attribute of HIN, is the relation subgraph G r is a row normalized matrix of , where t represents the target node type.

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

[0025] Among them, micro-level aggregation carries out message transmission and aggregation between nodes of the same type;

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

[0027] Preferably, the above micro-level aggregation includes merging node features in the same relationship, so that Generate specific types of features in

[0028] Furthermore, we use simple mean aggregation to obtain the specific type of features of the central node and calculate the row normalization matrix in is a degree matrix, where A r Represents the adjacency matrix corresponding to the relationship r;

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

[0030]

[0031] where r∈R is a relation type instance, Θ r is a learnable parameter of the relation type r in the subgraph, H s is the feature matrix of the s-type node.

[0032] Preferably, in the above macro-level aggregation, the attention mechanism can be viewed as mapping key sequences K to attention distributions α according to the query Q, where these key sequences have one-to-one corresponding values ​​V, and the definition of a specific type of attention is as follows:

[0033] Q t =H t ΘQ , K r =H r Θ K

[0034] e=ELU([K r ‖Q t ]Θ att )

[0035] α=softmax(e r )

[0036] Among them, H t is the hidden representation of the target node type t∈O, and H r is the representation of the neighbor nodes of r∈R∪{tt} whose target node type is t. To evaluate the importance of neighbor type representation, the target node type itself is regarded as the neighbor type, that is, {tt}, Θ Q’ Θ K， Θ att is the learnable parameter of the model, ELU is the activation function, and α is the attention coefficient calculated by the softmax function;

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

[0038]

[0039] where σ(·) is a nonlinear activation function, Γ t is the neighbor relationship set of the target node type t, and the output represents H′ t It is the input of the next layer of blocks, and the final representation is produced by the blocks of the last layer.

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

[0041]

[0042] in is the final embedding matrix after classifier processing, Z t The output features of the target type node in the last layer of the neural network are learned, Θ c is a learnable parameter in the classifier, and the category probability of the target node is calculated using the Softmax function. The final node representation received from the heterogeneous message passing module is expressed as:

[0043]

[0044]

[0045] is the target node label probability after Softmax normalization;

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

[0047]

[0048] in Represents some nodes with labels, y ic ∈Y is the target node v i The true label matrix is is the target node v i The predicted label distribution of .

[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, is the target node v i The pseudo-labels of the above two steps are combined into a multi-objective optimization problem, which enables the entire model to be trained in an end-to-end manner. Formally, the model is optimized by the following objective function:

[0052]

[0053] Among them, λ is a balancing hyperparameter. For unbalanced node classification, Re-Weight, BalancedSoftmax and Post-Compensated Softmax are further defined as follows:

[0054]

[0055]

[0056]

[0057] in, bs=(logn1,logn2,...,logn c ),

[0058] pc=(-logn1,-logn2,…,-logn c ) and n c is the number of nodes of type c. 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 relationship-aware label propagation, including 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 a specific type of linear transformation, ensuring that the feature vector dimensions of all nodes are the same;

[0061] The relationship-aware label propagation module is used to generate pseudo labels through a relationship-aware mechanism, consider the relationship between nodes, and use the label propagation algorithm to propagate known label information to unlabeled nodes to enhance the classification performance of minority class nodes;

[0062] A heterogeneous message passing module that leverages the pseudo-labels generated by the relation-aware label propagation module and works on top of the unified feature representation generated by the heterogeneous linear transformation module;

[0063] The classifier design module is used to perform classification or prediction based on the final node representation.

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

[0065] 1. It can improve the classification performance of minority class nodes: Through the relationship-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 uses the information of known label nodes and extends the label information to nodes of the same and different types through the relationship-aware mechanism, which makes up for the lack of labels of minority class nodes in the training data and significantly improves the classification accuracy of minority class nodes.

[0066] 2. Capable of multi-level information aggregation: The heterogeneous message passing module adopts a two-layer aggregation strategy to perform information transmission and aggregation at the micro level (within the type) and macro level (between types). In micro-level aggregation, information is exchanged between nodes of the same type to ensure the integrity of the internal features of the nodes; in macro-level aggregation, the type attention mechanism is used to learn the weights of neighbors of different types and optimize information fusion across nodes of different types. This multi-level aggregation method enables node features to represent their semantic information more comprehensively and accurately, improving the robustness and classification effect of the model.

[0067] 3. Ability to handle data imbalance: A variety of imbalance strategies are adopted, including re-weighting, balanced softmax, and post-compensated softmax, to deal with the problem of imbalanced category distribution. Specifically, the re-weighting strategy adjusts the category weights in the loss function so that minority class nodes receive more attention during training; the balanced softmax and post-compensated softmax strategies adjust the category probabilities at the classifier level to ensure that the category distribution of the model output is more balanced. These strategies effectively alleviate the bias caused by category imbalance and improve the overall classification performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0069] Figure 1 A schematic diagram of a flow chart of an embodiment of the present invention;

[0070] Figure 2 Schematic diagram of the module structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0071] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0072] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not represent any order, quantity or importance, but are only used to distinguish different components. "Including" or "comprising" and similar words mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connecting" or "connected" and similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used 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 2As shown, a heterogeneous graph node classification method based on relationship-aware label propagation includes the following steps:

[0074] S110, according to a given heterogeneous network, performing type-specific linear transformations on different types of nodes, and projecting their features into a common feature space;

[0075] Heterogeneous Linear Transformation Module,Due to the inherent heterogeneity of nodes in heterogeneous information networks (HINs),,different types of nodes have different semantic features and dimensions.,The heterogeneous linear transformation module projects the attribute features of,heterogeneous nodes into a common low-dimensional feature space through a,specific type of linear transformation, ensuring that the feature vectors of all nodes have,the same dimension.

[0076] Furthermore, the specific heterogeneous linear transformation is defined as:

[0077] H o =X o Θ o , o∈O

[0078] Among them, o∈O, o represents the node type in HIN, O represents the set of all node types, X o represents the original input feature of node type o, Θ o is a learnable parameter of node type o, H o Represents the output features after linear transformation.

[0079] The heterogeneous linear transformation module provides a unified feature representation for subsequent relationship-aware label propagation and heterogeneous message passing, ensuring that the 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 relationship-aware label propagation algorithm and generate pseudo labels through relationship subgraphs;

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

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

[0083]

[0084] in are the pseudo labels in layer l, A is the adjacency matrix, and D is the corresponding degree matrix;

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

[0086]

[0087]

[0088] Among them, s∈O is the node attribute of HIN, is the relation subgraph G r is a row normalized matrix of , where t represents the target node type.

[0089] This paper designs a relationship-aware label propagation method, which generates pseudo labels and propagates them to unlabeled nodes by considering the relationships and connections between nodes. Compared with traditional label propagation methods, relationship-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, adopt a two-layer aggregation strategy based on type attention mechanism to transfer information, and effectively integrate the multi-level information of nodes to make the node feature representation richer and more accurate;

[0091] Heterogeneous message passing module, including 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 uses the type attention mechanism to learn the weights of different types of neighbors of the target node, and performs information passing and aggregation between nodes of different types. This module uses the pseudo-labels generated by the relation-aware label propagation module and works on the unified feature representation generated by the heterogeneous linear transformation module.

[0092] The above micro-level aggregation involves merging node features in the same relationship, thereby Generate specific types of features in

[0093] Furthermore, we use simple mean aggregation to obtain the specific type of features of the central node and calculate the row normalization matrix in is a degree matrix, where A r Represents the adjacency matrix corresponding to the relationship r;

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

[0095]

[0096] Among them, r∈R is a relation type instance, Θ r is a learnable parameter of the relation type r in the subgraph, H sis the feature matrix of the s-type node.

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

[0098] In the above macro-level aggregation, the attention mechanism can be seen as mapping the key sequence K to the attention distribution α according to the query Q. These key sequences have a one-to-one corresponding value V. The definition of a specific type of attention is as follows:

[0099] Q t =H t Θ Q , K r =H r Θ K

[0100] e=ELU([K r ‖Q t ]Θ att )

[0101] α=softmax(e r )

[0102] Among them, H t is the hidden representation of the target node type t∈O, and H r is the representation of the neighbor nodes of r∈R∪{tt} whose target node type is t. To evaluate the importance of neighbor type representation, the target node type itself is regarded as the neighbor type, that is, {tt}, Θ Q , Θ K , Θ att is the learnable parameter of the model, ELU is the activation function, and α is the attention coefficient calculated by the softmax function;

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

[0104]

[0105] where σ(·) is a nonlinear activation function, Γt t is the neighbor relationship set of the target node type t, and the output represents H′ t It is the input of the next layer of blocks, and the final representation is produced by the blocks of the last layer.

[0106] The two-layer aggregation strategy proposed in this invention includes micro-level (within-type) aggregation and macro-level (between-type) aggregation. Micro-level aggregation transfers information between nodes of the same type to ensure the integrity of features within the type; macro-level aggregation uses the type attention mechanism to learn the weights of neighboring nodes of different types, thereby optimizing information fusion across types of nodes. The two-layer aggregation strategy makes the node feature representation richer and more accurate, improving the robustness and classification effect of the model.

[0107] S140. An unbalanced loss function is used in the classifier to perform node classification tasks, and the relationship-aware label propagation algorithm and the two-layer aggregation strategy are combined to form a multi-objective optimization problem to achieve 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, which is expressed as:

[0109]

[0110] in is the final embedding matrix after classifier processing, Z t The output features of the target type node in the last layer of the neural network are learned, Θ c is a learnable parameter in the classifier, and the category probability of the target node is calculated using the Softmax function. The final node representation received from the heterogeneous message passing module is expressed as:

[0111]

[0112]

[0113] is the target node label probability after Softmax normalization;

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

[0115]

[0116] in Represents some nodes with labels, y ic ∈Y is the target node v i The true label matrix is is the target node v i The predicted label distribution of .

[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, is the target node v i The pseudo-labels of the above two steps are combined into a multi-objective optimization problem, which enables the entire model to be trained in an end-to-end manner. Formally, the model is optimized by the following objective function:

[0120]

[0121] Among them, λ is a balancing hyperparameter. For unbalanced node classification, Re-Weight, BalancedSoftmax and Post-Compensated Softmax are further defined as follows:

[0122]

[0123]

[0124]

[0125] in, bs=(logn1,logn2,...,logn c ),

[0126] pc=(-logn1,-logn2,...,-logn c ) and n c is the number of nodes of type c, and the output of the classifier will be used as the final prediction result of the model.

[0127] In order to optimize label propagation and classification tasks simultaneously, this paper adopts a multi-objective optimization strategy. By combining label propagation loss and classification loss, a comprehensive loss function is designed to achieve end-to-end model training. The multi-objective optimization strategy not only balances the weights of the two tasks, but also improves the performance of the overall model.

[0128] In order to solve the problem of unbalanced node classification in heterogeneous graphs, this paper introduces unbalanced processing strategies such as re-weighting, balanced softmax and post-compensated softmax. The re-weighting strategy adjusts the category weights in the loss function so that minority class nodes get more attention during the training process; the balanced softmax and post-compensated softmax strategies adjust the category probability at the classifier level to ensure that the category distribution of the model output is more balanced, thereby effectively dealing with the problem of unbalanced category distribution and improving classification accuracy.

[0129] A heterogeneous graph node classification system based on relationship-aware label propagation, including 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 a specific type of linear transformation, ensuring that the feature vector dimensions of all nodes are the same;

[0131] The relationship-aware label propagation module is used to generate pseudo labels through a relationship-aware mechanism, consider the relationship between nodes, and use the label propagation algorithm to propagate known label information to unlabeled nodes to enhance the classification performance of minority class nodes;

[0132] A heterogeneous message passing module that leverages the pseudo-labels generated by the relation-aware label propagation module and works on top of the unified feature representation generated by the heterogeneous linear transformation module;

[0133] The classifier design module is used to perform classification or prediction based on the final node representation.

[0134] In summary, the present invention effectively improves the performance and robustness of unbalanced node classification in heterogeneous graphs by combining relationship-aware label propagation and a two-layer aggregation strategy, solves the shortcomings of existing methods in processing unbalanced node classification in heterogeneous graphs, and has significant application value and broad application prospects. Other features and advantages of the present invention will be described in the subsequent description, and some will become apparent from the description or be understood by implementing the present invention. The objects and other advantages of the present invention can be achieved and obtained through 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 illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity.

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

Claims

1. A heterogeneous graph node classification method based on relationship-aware label propagation, characterized in that: The following steps are involved: S110, according to a given heterogeneous network, performing type-specific linear transformations on different types of nodes, and projecting their features into a common feature space; S120, for target nodes in heterogeneous networks, design a relationship-aware label propagation algorithm and generate pseudo labels through relationship subgraphs; S130, adopt a two-layer aggregation strategy based on type attention mechanism to transfer information, and effectively integrate the multi-level information of nodes to make the node feature representation richer and more accurate; S140. An unbalanced loss function is used in the classifier to perform node classification tasks, and the relationship-aware label propagation algorithm and the two-layer aggregation strategy are combined to form a multi-objective optimization problem to achieve end-to-end training of the entire model.

2. According to claim 1, a heterogeneous graph node classification method based on relationship-aware label propagation is characterized in that: In S110 the linear transformation is defined as: H o =X o I o ,o∈O Among them, o∈O, o represents the node type in HIN, O represents the set of all node types, X o represents the original input feature of node type o, Θ o is a learnable parameter of node type o, H o Represents the output features after linear transformation.

3. According to claim 2, a heterogeneous graph node classification method based on relationship-aware label propagation is characterized in that: In S120, given the initial label matrix Y t (0) , the formula for iterative label propagation is as follows: in are the pseudo labels in layer l, A is the adjacency matrix, and D is the corresponding degree matrix; Furthermore, the process of generating pseudo labels based on label propagation of relational subgraphs is as follows: Among them, s∈O is the node attribute of HIN, is the relation subgraph G r is a row normalized matrix of , where t represents the target node type.

4. According to claim 3, a heterogeneous graph node classification method based on relationship-aware label propagation is characterized in that: In step S130, the two-layer aggregation strategy includes micro-level aggregation and macro-level aggregation; Among them, micro-level aggregation carries out message transmission and aggregation between nodes of the same type; Among them, the macro-level aggregation uses the type attention mechanism to learn the weights of different types of neighbors of the target node and perform information transfer and aggregation between nodes of different types.

5. According to claim 4, a heterogeneous graph node classification method based on relationship-aware label propagation is characterized in that: The above micro-level aggregation involves merging node features in the same relationship, thereby Generate specific types of features in Furthermore, we use simple mean aggregation to obtain the specific type of features of the central node and calculate the row normalization matrix in is a degree matrix, where A r Represents the adjacency matrix corresponding to the relationship r; Furthermore, the heterogeneous graph convolution in the relational subgraph can be defined as: Among them, r∈R is a relation type instance, Θ r is a learnable parameter of the relation type r in the subgraph, H s is the feature matrix of the s-type node.

6. According to claim 4, a heterogeneous graph node classification method based on relationship-aware label propagation is characterized in that: In the above macro-level aggregation, the attention mechanism can be seen as mapping the key sequence K to the attention distribution α according to the query Q. These key sequences have a one-to-one corresponding value V. The definition of a specific type of attention is as follows: Q t =H t I Q ,K r =H r I K e=LIFE([K r ‖Q t ]Θ att ) α=softmax(e r ) Among them, H t is the hidden representation of the target node type t∈O, and H r is the representation of the neighbor nodes of r∈R∪{tt} whose target node type is t. To evaluate the importance of neighbor type representation, the target node type itself is regarded as the neighbor type, that is, {tt}, Θ Q , Θ K , Θ att is the learnable parameter of the model, ELU is the activation function, and α is the attention coefficient calculated by the softmax function; Furthermore, the model uses the following formula to calculate the output characteristics of the target node after attention weighting: where σ(·) is a nonlinear activation function, Γ t is the neighbor relationship set of the target node type t, and the output represents H′ t It is the input of the next layer of blocks, and the final representation is produced by the blocks of the last layer.

7. The heterogeneous graph node classification method based on relationship-aware label propagation according to claim 6 is characterized in that: In S140, a single-layer linear transformation is used as a classifier, which is expressed as: in is the final embedding matrix after classifier processing, Z t The output features of the target type node in the last layer of the neural network are learned, Θ c is a learnable parameter in the classifier, and the category probability of the target node is calculated using the Softmax function. The final node representation received from the heterogeneous message passing module is expressed as: is the target node label probability after Softmax normalization; Furthermore, the cross entropy loss with the Softmax function and negative log-likelihood loss is: in Represents some nodes with labels, y ic ∈Y is the target node v i The true label matrix is is the target node v i The predicted label distribution of .

8. The heterogeneous graph node classification method based on relationship-aware label propagation according to claim 7, 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, is the target node v i The pseudo-labels of the above two steps are combined into a multi-objective optimization problem, which enables the entire model to be trained in an end-to-end manner. Formally, the model is optimized by the following objective function: Among them, λ is a balancing hyperparameter. For unbalanced node classification, Re-Weight, BalancedSoftmax and Post-Compensated Softmax are further defined as follows: in, bs=(logn1,logn2,...,logn c ), pc=(-logn1,-logn2,...,-logn c ) and n c is the number of nodes of type c, and the output of the classifier will be used as the final prediction result of the model.

9. A heterogeneous graph node classification system based on relationship-aware label propagation, characterized by: Includes the following modules: 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 relationship-aware label propagation module is used to generate pseudo labels through a relationship-aware mechanism, consider the relationship between nodes, and use the label propagation algorithm to propagate known label information to unlabeled nodes to enhance the classification performance of minority class nodes; A heterogeneous message passing module that leverages the pseudo-labels generated by the relation-aware label propagation module and works on top of the unified feature representation generated by the heterogeneous linear transformation module; The classifier design module is used to perform classification or prediction based on the final node representation.

Images