Self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity
By constructing a multi-hop path similarity matrix and a self-expressive solver to capture homogeneity, and combining a path encoder to model long-range heterogeneity, the performance degradation problem in self-supervised heterogeneous knowledge graph learning is solved, achieving significant improvements on various datasets and enhancing the performance of recommendation systems and semantic search.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HAINAN UNIV
- Filing Date
- 2026-03-31
- Publication Date
- 2026-07-03
AI Technical Summary
Existing self-supervised heterogeneous knowledge graph learning methods suffer from performance degradation when faced with sparse meta-path structures or noisy node features, and fail to effectively capture long-range heterogeneity, resulting in poor model performance in diverse applications.
By constructing a multi-hop path similarity matrix and a self-expressive solver to capture homogeneity, combining a path encoder to model long-range heterogeneity, and suppressing noise through adaptive fusion, a self-supervised heterogeneous knowledge graph learning method is designed.
It significantly improves model performance on various datasets, captures the complementary homogeneity of node features and meta-paths, reduces noise interference, captures more task-related information, and improves the accuracy of recommendation systems and semantic search.
Smart Images
Figure CN122334418A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of text data analysis technology, and more specifically, relates to a self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity. Background Technology
[0002] Heterogeneous knowledge graph learning aims to model the complex relationships between different types of nodes in a text description, thereby generating semantically rich representations. To avoid time-consuming manual annotation, self-supervised heterogeneous knowledge graph learning has emerged as a promising new paradigm, achieving significant success in diverse practical applications such as recommender systems and semantic search. Existing self-supervised heterogeneous knowledge graph learning methods can be broadly categorized into two types: meta-path-based methods and feature-based methods. The former typically utilizes predefined meta-paths to explicitly capture structural homogeneity, aggregating information about connected nodes with the same labels. In contrast, feature-based methods focus on directly inferring potential homogeneous structures from node features.
[0003] Despite the promising results achieved by existing methods, several limitations remain to be addressed. Firstly, most methods for capturing homogeneity rely solely on meta-paths or node features, neglecting the inherent complementarity between knowledge graph structure and node features. This can lead to significant performance degradation when faced with sparse meta-path structures or noisy node features. Secondly, existing methods are often limited to interactions between nodes of the same type. When meta-paths exhibit low homogeneity, this can lead to confusion between adjacency information from different categories, resulting in noise accumulation. Furthermore, these methods fail to capture long-range heterogeneity (i.e., cross-type interactions), which may carry task-related information.
[0004] Based on the above analysis, the effectiveness of self-supervised heterogeneous knowledge graph learning can be improved by simultaneously utilizing the homogeneity in meta-paths and node features, while explicitly modeling long-range heterogeneity. This requires addressing two key challenges: first, how to extract complementary homogeneity from meta-paths and node features, as there may be potential conflicts between the two; and second, how to effectively suppress noise introduced by meta-paths while capturing long-range heterogeneity in heterogeneous knowledge graphs. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity. It captures homogeneity from meta-path and node features and obtains homogeneous representations to avoid the potential significant performance degradation of the model when facing sparse meta-path structures or noisy node features. At the same time, it models multi-hop neighbors and uses adaptive fusion to suppress noise and capture long-range heterogeneity to integrate homogeneity and heterogeneity, thereby introducing more task-related information.
[0006] To achieve the above-mentioned objectives, this invention provides a self-supervised heterogeneous knowledge graph learning method based on the homogeneity of feature structures and long-range heterogeneity, characterized by comprising the following steps:
[0007] (1) Constructing heterogeneous knowledge graphs based on real-world knowledge text data
[0008] We perform entity sorting and relation extraction on real-world knowledge text data used in recommendation systems and semantic search, constructing a heterogeneous knowledge graph with entities as nodes and relations as edges. ,in, and Representing the node set and edge set respectively. This is a set of feature matrices for each type of node. Each type of node feature matrix corresponds to a node type, and is composed of feature vectors obtained from the knowledge text data of all nodes of that node type through a word embedding model. and Represent the node type set and the edge type set respectively;
[0009] (2) Constructing multi-hop intra-class paths and inter-class paths
[0010] According to the heterogeneity diagram Define multi-hop paths : ,in, For multi-hop paths Number of nodes For multi-hop paths Upper The node type of the node. For multi-hop paths Upper node to the The relationship type of each node;
[0011] For multi-hop paths Its subgraph matrix for:
[0012]
[0013] in, Indicates node type All nodes and node types The row-normalized adjacency matrix of all nodes;
[0014] Multi-hop paths are divided into two categories: if the first node type Then multi-hop path For intra-class paths, if the first node type Then multi-hop path For inter-class paths;
[0015] (3) Construct a self-expressive solver to capture homogeneity
[0016] First, construct a multi-hop path similarity matrix. To capture the structural homogeneity of intra-class path subgraphs, its The position value is the cosine similarity. :
[0017]
[0018] in, The target node type is the node type. For target node type The first path in the class Subgraph matrix of the bar, matrix The average subgraph for all intra-class paths, , Represent matrices respectively The line, number OK, Cosine similarity function, cosine similarity The larger the value, the higher the number of nodes in the target node type. Nodes With the Nodes There are more neighbor nodes shared across all intra-class paths;
[0019] Then, the multi-hop path similarity matrix By incorporating a self-expressive solver, the constraint matrix is obtained. :
[0020]
[0021] in, It represents the Hadamah accumulation. This indicates selecting the first row from each row. Find the largest element, assign the selected element a value of 1, and set the rest of the elements to 0. The feature similarity matrix is obtained by calculating the cosine similarity based on the node features of all target node types. It is the identity matrix;
[0022] Then, design a self-expressive solver:
[0023]
[0024] in, A matrix consisting of the node features of all target node types. This is a self-expression matrix, whose elements are... The value reflects the node and The probability of belonging to the same category. and The parameter is non-negative, while the constraint condition is... Preventing nodes from rebuilding themselves is achieved by setting diagonal elements to zero. This indicates a search for the norm.
[0025] Finally, the self-expression solver is solved to obtain the self-expression matrix. Homogeneous representation is obtained :
[0026]
[0027] (4) Construct a path encoder to capture heterogeneity
[0028] First, a multilayer perceptron is used as the encoder. In each subgraph matrix of a multi-hop path Path-specific representation obtained above ,Right now:
[0029]
[0030] in, For activation function, The target node type is the node type. For target node type The first multi-hop path The subgraph matrix of the bar, Indicates encoder Trainable parameters;
[0031] Then, the path fusion representation is obtained through adaptive fusion. :
[0032]
[0033] in, These are learnable weights;
[0034] (5) Obtain the final representation
[0035] First, learn about projector heads. and To represent homogeneity Path fusion representation Mapped to the same latent space, i.e.:
[0036]
[0037] in, It is represented by a homogeneous projection. For path fusion projection representation
[0038] Then, calculate the consistency loss. To maximize the invariance between the two, that is:
[0039]
[0040] in, The first homogeneous projection is represented by the A vector formed by rows, For path fusion projection representation The A vector formed by rows, The number of rows;
[0041] Calculate contrast loss Represented by separating nodes belonging to different categories:
[0042]
[0043] in, For temperature parameters, Represents a node of The set of nearest neighbors and These are assigned to nodes. The non-negative parameters of neighbors and non-neighbors are set by... Contrast loss will focus more on separating node representations belonging to different categories, thereby increasing inter-class distance;
[0044] Then, the consistency loss Compared with loss After integration, the objective function is obtained. :
[0045]
[0046] in, and The parameter is non-negative.
[0047] In the objective function After optimization under guidance, the final homogeneous representation will be obtained. Path fusion representation The data is then concatenated to obtain the final representation used for downstream tasks. ;
[0048] (6) Use the final representation for downstream tasks
[0049] Obtain the final representation of each target node type It is used for semantic search and recommendation systems.
[0050] The objective of this invention is achieved as follows.
[0051] This invention proposes a self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity. First, a self-expressive solver is proposed, which captures the complementary homogeneity between meta-paths and node features, thus obtaining a homogeneous representation. Simultaneously, this invention designs a path encoder to model various interaction relationships. Through adaptive fusion, it explicitly incorporates cross-type interaction relationships while reducing noise interference and capturing long-range heterogeneity. Theoretical analysis verifies that the homogeneous representation has a higher-order grouping effect, effectively capturing complementary homogeneity; the path encoder possesses adaptive smoothing capabilities to filter noise; and cross-type interaction modeling, integrating homogeneity and heterogeneity, can introduce more task-related information. Experiments conducted on various datasets, including large-scale datasets, fully verify the superiority of the proposed method.
[0052] Compared with the prior art, the present invention has the following advantages:
[0053] 1. The first work to build a unified framework for self-supervised heterogeneous knowledge graph learning, while taking into account both feature-structure homogeneity and long-range heterogeneity;
[0054] 2. This invention improves the self-expressive solver to capture the complementary homogeneity between meta-paths and node features. Furthermore, through path encoders and adaptive fusion, long-range heterogeneity is explicitly modeled and noise is filtered out.
[0055] 3. Theoretical analysis shows that homogeneity representation has a higher-order grouping effect, which can simultaneously capture the homogeneity of meta-path and node features; at the same time, the path encoder filters noise through adaptive fusion, and can introduce more task-related information by taking into account both homogeneity and heterogeneity.
[0056] 4. Experiments show that, compared with many existing methods, this invention demonstrates significant advantages on diverse datasets, including large-scale heterogeneous knowledge graphs. Attached Figure Description
[0057] Figure 1 This is a flowchart of a specific implementation of the self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity of the present invention;
[0058] Figure 2 This is a schematic diagram illustrating the principle of a specific implementation of the self-supervised heterogeneous knowledge graph learning method based on the homogeneity of feature structures and long-range heterogeneity of the present invention. Detailed Implementation
[0059] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.
[0060] Figure 1 , 2 These are flowcharts and schematic diagrams illustrating a specific implementation of the self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity of the present invention.
[0061] In this embodiment, as Figure 1 As shown, the self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity of the present invention includes the following steps:
[0062] Step S1: Construct a heterogeneous knowledge graph based on real-world knowledge text data
[0063] We perform entity sorting and relation extraction on real-world knowledge text data used in recommendation systems and semantic search, constructing a heterogeneous knowledge graph with entities as nodes and relations as edges. ,in, and Representing the node set and edge set respectively. This is a set of feature matrices for each type of node. Each type of node feature matrix corresponds to a node type, and is composed of feature vectors obtained from the knowledge text data of all nodes of that node type through a word embedding model. and These represent the node type set and the edge type set, respectively.
[0064] Taking academic data as knowledge text data as an example, by using relevant knowledge, we can sort out and obtain multiple types of node sets such as "paper", "author", and "conference". , These are subsets of paper nodes, author nodes, and conference nodes, from which nodes are extracted. , Multiple types of relations as edge sets Read the natural language text of the paper From text data Filtering keywords to construct a thesaurus To form a paper-keyword mapping , Using word embedding model From paper-keyword mapping The feature vectors generated for the paper are the node features, which are the node features of all target node types. Construct a matrix It can generate feature vectors for other types of nodes; note that this invention is not limited to academic data, and can be applied to any data that can be processed using the described process.
[0065] Step S2: Construct multi-hop intra-class paths and inter-class paths
[0066] like Figure 2 As shown, according to the heterogeneity diagram Define multi-hop paths : ,in, For multi-hop paths Number of nodes For multi-hop paths Upper The node type of the node. For multi-hop paths Upper node to the The relationship type of each node.
[0067] For multi-hop paths Its subgraph matrix for:
[0068]
[0069] in, Indicates node type All nodes and node types The row-normalized adjacency matrix of all nodes.
[0070] Unlike previous meta-path methods that typically only focus on connections between nodes of the same type, multi-hop paths extend this concept to connections between nodes of different types. This invention categorizes multi-hop paths into two types: if the first node type... Then multi-hop path For intra-class paths, if the first node type Then multi-hop path These are inter-class paths. Therefore, the subgraph matrix induced by intra-class paths can be used to capture structural homogeneity, while the subgraph matrix induced by inter-class paths can be used to capture heterogeneity.
[0071] Step S3: Construct a self-expressive solver to capture homogeneity
[0072] A traditional self-expressive solver can be represented as follows:
[0073]
[0074] in, A matrix consisting of the node features of all target node types. This is a self-expression matrix, whose elements are... The value reflects the node and The probability of belonging to the same category. and The parameter is non-negative, while the constraint condition is... Preventing nodes from rebuilding themselves is achieved by setting diagonal elements to zero. To find the norm, This is a feature similarity matrix calculated based on the cosine similarity of node features across all target node types. The first term linearly reconstructs each node from the other nodes; the second term ensures that nodes with similar features contribute significantly; and the third term is a regularization term that promotes sparsity, based on the assumption that nodes with similar features may belong to the same category. The value reflects the node and The probability of belonging to the same category is used to capture the homogeneity effect from node features. However, traditional self-expression solvers ignore the structural homogeneity of intra-class paths. To address this issue, this invention imposes constraints on the self-expression matrix to fully utilize structural homogeneity. Specifically, it includes the following steps:
[0075] like Figure 2 As shown, firstly, a multi-hop path similarity matrix is constructed. To capture the structural homogeneity of intra-class path subgraphs, its The position value is the cosine similarity. :
[0076]
[0077] in, The target node type is the node type. For target node type The first path in the class Subgraph matrix of the bar, matrix The average subgraph for all intra-class paths, , Represent matrices respectively The line, number OK, Cosine similarity function, cosine similarity The larger the value, the higher the number of nodes in the target node type. Nodes With the Nodes The paths within the same class share more neighbor nodes. Since nodes sharing more neighbors usually belong to the same class, the path similarity matrix... This reflects the homogeneity of paths exported within a class.
[0078] Then, the multi-hop path similarity matrix By incorporating a self-expressive solver, the constraint matrix is obtained. :
[0079]
[0080] in, It represents the Hadamah accumulation. This indicates selecting the first row from each row. Find the largest element, assign the selected element a value of 1, and set the rest of the elements to 0. The feature similarity matrix is obtained by calculating the cosine similarity based on the node features of all target node types. It is an identity matrix.
[0081] pass right Weighting is applied to incorporate path similarity while considering the original constraints, thus creating a new constraint matrix. Therefore, this invention presents a novel self-expressive solver designed to capture the complementary homogeneity between node features and intra-class paths:
[0082]
[0083] in, A matrix consisting of the node features of all target node types. This is a self-expression matrix, whose elements are... The value reflects the node and The probability of belonging to the same category. and The parameter is non-negative, while the constraint condition is... Preventing nodes from rebuilding themselves is achieved by setting diagonal elements to zero. The expression represents the norm calculation. The first term utilizes the homogeneity in node features, while the second term utilizes the homogeneity in intra-class paths. Furthermore, the second term penalizes the weights between nodes of different classes induced by the first term, while the first term captures intra-class connections ignored by the second term, thus forming a complement. Therefore, the self-expressive solver captures this complementarity by balancing the two.
[0084] Finally, the self-expression solver is solved to obtain the self-expression matrix. Homogeneous representation is obtained :
[0085]
[0086] Therefore, homogeneity representation can simultaneously capture the inherent complementarity between node features and similar paths, thereby enabling information transfer between nodes that may belong to the same category.
[0087] Step S4: Build a path encoder to capture heterogeneity
[0088] This invention employs a path encoder to filter noise along intra-class paths and captures long-range heterogeneity to incorporate more task-related information. Specifically, it includes the following steps:
[0089] like Figure 2 As shown, firstly, a multilayer perceptron is used as the encoder. In each subgraph matrix of a multi-hop path Path-specific representation obtained above ,Right now:
[0090]
[0091] in, For activation function, The target node type is the node type. For target node type The first multi-hop path The subgraph matrix of the bar, Indicates encoder Trainable parameters.
[0092] when When it is an intra-class path subgraph, It can capture information about neighbors of the same type. Unlike predefined meta-paths, which only focus on a few symmetric paths between nodes of the same type, intra-class paths can flexibly cover more paths, including asymmetric paths. Meanwhile, When it is a subgraph of inter-class paths, It can capture heterogeneity and aggregate information from different types of nodes.
[0093] To effectively utilize the representations of intra-class paths and inter-class paths, this invention obtains a path fusion representation through adaptive fusion. :
[0094]
[0095] in, These are learnable weights. Path fusion representations are derived by assigning different weights to the paths. It can integrate information from both similar and cross-type neighborhoods, and effectively suppress noise interference.
[0096] Step S5: Obtain the final representation
[0097] For homogeneous representation Path fusion representation Although derived from different perspectives, the two methods share the same original node features, thus intuitively containing consistent information. Therefore, this invention designs a consistency loss to extract the consistent information between the two. Specifically,
[0098] like Figure 2 As shown, first, learn about the projection head. and To represent homogeneity Path fusion representation Mapped to the same latent space, i.e.:
[0099]
[0100] in, It is represented by a homogeneous projection. For path fusion projection representation
[0101] Then, calculate the consistency loss. To maximize the invariance between the two, that is:
[0102]
[0103] in, The first homogeneous projection is represented by the A vector formed by rows, For path fusion projection representation The A vector formed by rows, The number of rows.
[0104] Consistency loss facilitates homogeneous projection representation Path-fused projection representation Alignment between them captures consistency information. However, the consistency loss lacks constraints on different categories of nodes, which may lead to difficulty in distinguishing the representations of different categories of nodes, thereby degrading the performance of downstream tasks.
[0105] To address this issue, based on the assumption that neighboring nodes may belong to the same category, this invention further calculates the contrastive loss. Represented by separating nodes belonging to different categories:
[0106]
[0107] in, For temperature parameters, Represents a node of The set of nearest neighbors and These are assigned to nodes. The non-negative parameters of neighbors and non-neighbors are set by... Contrast loss will focus more on separating node representations belonging to different categories, thereby increasing inter-class distance.
[0108] Then, the consistency loss Compared with loss After integration, the objective function is obtained. :
[0109]
[0110] in, and It is a non-negative parameter.
[0111] In the objective function After optimization under guidance, the final homogeneous representation will be obtained. Path fusion representation The data is then concatenated to obtain the final representation used for downstream tasks. Therefore, the final expression It contains both homogeneous information about node features and intra-class paths, and can also model heterogeneity.
[0112] Step S6: Perform downstream tasks using the final representation.
[0113] Obtain the final representation of each target node type It is used for semantic search and recommendation systems.
[0114] To better illustrate the technical effects of this invention, specific examples are used to conduct experimental verification. In this experimental verification, four benchmark datasets in the field of data analysis were used to evaluate the effectiveness of this invention on a node classification task, including three citation datasets (i.e., ACM, DBLP, and Aminer) and one commercial dataset (i.e., Yelp).
[0115] This experiment verified 15 comparison methods, including 12 heterogeneous graph methods (Mp2vec, HAN, HGT, DMGI, DMGIattn, HDMI, HeCo, HGCML, CPIM, HGMAE, HERO, MGHC, D2CMG) and 3 isogeneous graph methods (DeepWalk, GCN, GAT).
[0116] Table 1 is a statistical table of the classification F1 scores of the present invention and the comparison method in this embodiment.
[0117]
[0118] Table 1
[0119] As shown in Table 1, the results show that the present invention achieved the best results on all four datasets. This indicates that the final representation generated by the present invention brings semantically similar nodes closer together and semantically dissimilar nodes apart, thus possessing rich semantic information. This is beneficial for solving the technical problem of improving inference accuracy in text embedding and semantic search processes. The performance on the Yelp commercial dataset shows that the present invention helps to solve the technical problem of improving the relevance of goods and services when recommending goods and services to users in recommendation systems.
[0120] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.
Claims
1. A self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity, characterized in that, Includes the following steps: (1) Construct heterogeneous knowledge graphs based on real-world knowledge text data; We perform entity sorting and relation extraction on real-world knowledge text data used in recommendation systems and semantic search, constructing a heterogeneous knowledge graph with entities as nodes and relations as edges. ,in, and Representing the node set and edge set respectively. This is a set of feature matrices for each type of node. Each type of node feature matrix corresponds to a node type, and is composed of feature vectors obtained from the knowledge text data of all nodes of that node type through a word embedding model. and Represent the node type set and the edge type set respectively; (2) Construct multi-hop intra-class paths and inter-class paths; According to the heterogeneity diagram Define multi-hop paths : ,in, For multi-hop paths Number of nodes For multi-hop paths Upper The node type of the node. For multi-hop paths Upper Node to the The relationship type of each node; For multi-hop paths whose subgraph matrix is: ； in, Indicates node type All nodes and node types The row-normalized adjacency matrix of all nodes; Multi-hop paths are divided into two categories: if the first node type Then multi-hop path For intra-class paths, if the first node type Then multi-hop path For inter-class paths; (3) Construct a self-expressive solver to capture homogeneity; First, a multi-hop path similarity matrix is constructed to capture structural homophily in intra-class path subgraphs, with the value of a position being cosine similarity : ； in, The target node type is the node type. For target node type The first path in the class Subgraph matrix of the bar, matrix The average subgraph for all intra-class paths, , Represent matrices respectively The line, number OK, Cosine similarity function, cosine similarity The larger the value, the higher the number of nodes in the target node type. Nodes With the Nodes There are more neighbor nodes shared across all intra-class paths; Then, the multi-hop path similarity matrix is incorporated into the self-expression solver to obtain the constraint matrix : ； in, It represents the Hadamah accumulation. This indicates selecting the first row from each row. Find the largest element, assign the selected element a value of 1, and set the rest of the elements to 0. The feature similarity matrix is obtained by calculating the cosine similarity based on the node features of all target node types. It is the identity matrix; Then, design a self-expressive solver: ； in, A matrix consisting of the node features of all target node types. This is a self-expression matrix, whose elements are... The value reflects the node and The probability of belonging to the same category. and The parameter is non-negative, while the constraint condition is... Preventing nodes from rebuilding themselves is achieved by setting diagonal elements to zero. This indicates a search for the norm. Finally, the self-expression solver is solved to obtain the self-expression matrix. Homogeneous representation is obtained : ； (4) Construct a path encoder to capture heterogeneity; First, a multilayer perceptron is used as an encoder At each subgraph matrix of the multi-hop path to obtain a path-specific representation That is: ； wherein, is an activation function, is a target node type, i.e. a node type is a target node type is a subgraph matrix of the th hop of the path, denotes trainable parameters of the encoder denotes trainable parameters of the encoder Then, the path fusion representation is obtained by adaptive fusion : ； wherein, are learnable weights; (5) Obtain the final representation; First, learn about projector heads. and To represent homogeneity Path fusion representation Mapped to the same latent space, i.e.: ； in, It is represented by a homogeneous projection. This is a path fusion projection representation; Then, calculate the consistency loss. To maximize the invariance between the two, that is: ； in, The first homogeneous projection is represented by the A vector formed by rows, For path fusion projection representation The A vector formed by rows, The number of rows; Calculate contrast loss Represented by separating nodes belonging to different categories: ； in, For temperature parameters, Represents a node of The set of nearest neighbors and These are assigned to nodes. The non-negative parameters of neighbors and non-neighbors are set by... Contrast loss will focus more on separating node representations belonging to different categories, thereby increasing inter-class distance; Then, the consistency loss Compared with loss After integration, the objective function is obtained. : ； in, and The parameter is non-negative. In the objective function After optimization under guidance, the final homogeneous representation will be obtained. Path fusion representation The data is then concatenated to obtain the final representation used for downstream tasks. ; (6) Use the final representation to perform downstream tasks; Obtain the final representation of each target node type It is used for semantic search and recommendation systems.
2. The self-supervised heterogeneous knowledge graph learning method based on feature structure homogeneity and long-range heterogeneity according to claim 1, characterized in that, The knowledge text data mentioned is academic data.