Extensible and unsupervised heterogeneous information network community search method
By constructing isomorphic graphs and hyperedge mechanisms, combined with attention mechanisms and unsupervised learning, the scalability and adaptability issues of community search methods in heterogeneous information networks are solved, achieving efficient and flexible community search that can adapt to large-scale graph data and dynamic structures.
Patent Information
- Application Number
- CN202511509858.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-22
- Publication Date
- 2026-01-30
AI Technical Summary
Existing machine learning-based heterogeneous information network community search methods suffer from poor scalability and low computational efficiency when processing large-scale graph data. They also rely heavily on community tags and are difficult to adapt to dynamic changes in graph structures, resulting in high maintenance costs and poor adaptability.
We design an unsupervised community search method for heterogeneous information networks. By constructing isomorphic graphs, introducing hyperedge and attention mechanisms, and combining unsupervised contrastive learning and node classification tasks, we achieve efficient feature fusion and community representation learning, adapting to changes in graph structure.
It achieves efficient and flexible community search under unsupervised conditions, can adapt to large-scale graph data and dynamic graph structures, improves the ability to characterize the multi-scale semantics of community structure, and has good training and inference efficiency and generalization ability.
Smart Images

Figure CN121434449A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of community search, specifically, it relates to a scalable and unsupervised method for searching heterogeneous information network communities. Background Technology
[0002] With the widespread use of graph data in applications such as social networks, e-commerce platforms, and biomedicine, how to extract semantically connected communities from complex and diverse network structures—that is, community search—has become a research hotspot in the field of graph mining. Heterogeneous Information Networks (HINs) contain multiple types of nodes and edges, enabling them to more realistically represent complex semantic structures in the real world. Therefore, they are widely used in practical tasks such as paper recommendation, fraud detection, and gene function prediction.
[0003] In recent years, some studies have begun to extend community search tasks to Attributed Heterogeneous Information Networks (AHINs). By introducing diverse node types, edge types, and node attributes, AHINs can express richer semantic information, and their importance in practical applications is increasingly prominent. However, existing community search methods based on AHINs still mainly rely on traditional algorithms, resulting in relatively rigid and inflexible approaches. In contrast, research on machine learning-based methods remains limited, which contrasts sharply with the development of machine learning methods in isomorphic graph community search.
[0004] Currently, some methods attempt to model AHINs using supervised learning techniques such as Graph Neural Networks (GNNs) to achieve community search. Although they have achieved some success on certain datasets, several problems remain: First, these methods generally suffer from poor scalability and low computational efficiency when dealing with large-scale graph data. Second, they are highly dependent on community labels, making it difficult to generalize in real-world applications where labels are missing or inaccurate. Furthermore, GNN-type methods typically rely on fixed graph structures for propagation modeling. When the graph structure changes (e.g., adding nodes or edges), the model often needs to be retrained, resulting in high maintenance costs and poor adaptability.
[0005] Therefore, designing a community search method that is highly scalable and computationally efficient, can be trained under unsupervised or weakly supervised conditions, can flexibly adapt to dynamic changes in graph structure, and can effectively integrate the structural features and semantic information of the graph has become an urgent problem to be solved in the field of graph mining. Summary of the Invention
[0006] To address the problems of poor scalability, strong dependence on tags, and weak adaptability to dynamic graph structures in existing technologies, this invention provides a scalable and unsupervised heterogeneous information network community search method.
[0007] To address the aforementioned technical problems, this invention provides a scalable and unsupervised method for searching heterogeneous information network communities:
[0008] Step 1: Define the target node type and the meta-path based on the target type node, and construct isomorphic graphs under different meta-paths accordingly.
[0009] Step 101: Select the target node type from the heterogeneous information network And define a set of metapaths associated with the target type node. .
[0010] Step 102: Construct the corresponding isomorphic graph based on each meta-path P and the original heterogeneous graph. Each isomorphic graph contains only the target type node and its semantic neighbor relationships connected by meta-paths.
[0011] Step 103: Perform step 102 independently for each meta-path to form a set of isomorphic graphs under the meta-path set, which serves as the basis for subsequent feature construction and community modeling.
[0012] Step 2: Integrate the features of other types of nodes in the metapath into the target type node to construct new semantic features for the nodes in the isomorphic graph under each metapath.
[0013] Step 201: Use dimensionality reduction methods such as singular value decomposition to map the features of different types of nodes to a unified dimension, ensuring the fusion of cross-type semantic information.
[0014] Step 202: Under the original path P, perform a weighted average of the dimensionality-reduced features of the semantic neighbors of the target type node in the isomorphic graph, and concatenate it with its own original features to form node features with fused semantics. .
[0015] Step 203: Repeat step 202 for all meta-paths to obtain the semantically enhanced feature representations of all target type nodes under each meta-path. .
[0016] Step 3: Introduce a hyperedge construction mechanism to perform structural cohesion screening on multi-hop neighbors, form a high-quality hyperedge set, and generate community feature tokens to represent different views.
[0017] Step 301: Set the maximum number of hops K and the structural cohesion screening threshold θ for each target type node to control the screening of multi-hop neighbors.
[0018] Step 302: Under the metapath P, with the target node v as the center, based on the isomorphic graph under the metapath In each hop, the set of neighboring nodes is expanded, and those nodes whose overlap with the set of neighbors in the previous hop is not less than θ are retained.
[0019] Step 303: Define the set of neighbors formed by each hop as a hyperedge, and construct a multi-hop hyperedge matrix. This represents the semantic connection relationships of nodes at different scales.
[0020] Step 304: Based on the hyperedge matrix The multi-hop semantic features of nodes are aggregated to generate feature tokens under multiple views. .
[0021] Step 305: Repeat steps 302-304 under all metapaths to obtain the feature token set. This is used for subsequent community representation learning.
[0022] At this point, the construction of feature tokens with no learnable parameters is complete, and the training phase begins.
[0023] Step 4: Input the feature tokens under each metapath into the attention-based model to extract low-order and high-order community representation information, and realize graph structure-independent feature fusion and representation learning.
[0024] Step 401: Under the meta-path P, transfer the features The input is first fed into an embedding layer, where it is transformed into a dense layer using a learnable linear transformation. Set-type characteristics.
[0025] Step 402: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The input is fed into the L-layer Transformer for further feature extraction, resulting in... .
[0026] Step 403: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The input is passed to a multi-view separation layer, and high-order features are extracted from it using an attention-based method. and low-order features .
[0027] Step 404: Repeat steps 401-403 under each meta-path to obtain the higher-order and lower-order representations under all original paths.
[0028] Step 405: Input all representations obtained in step 404 into the attention-based semantic aggregation layer, fuse the embedding representations under each meta-path, and obtain the global high-order representation. and lower-order representation .
[0029] Step 5: Through a dual-task training strategy, contrastive learning and auxiliary classification tasks are carried out simultaneously under the condition of no community label supervision, guiding the model to obtain a more reasonable community embedding representation.
[0030] Step 5, "Task 1: Unsupervised Comparative Learning," specifically includes:
[0031] Step 501: Calculate the semantic contrastive learning loss function under the metapath P. .
[0032] Step 502: For all training nodes, repeat step 501 under each meta-path, and sum the loss functions to obtain the semantic contrastive learning loss function. .
[0033] Step 503: Calculate the global contrastive learning loss function for the higher-order and lower-order representations after semantic aggregation. This guides the model to bring similar nodes closer together and distinguish unrelated nodes in the semantic space.
[0034] Step 502: Combine the two loss functions from steps 502 and 503 to construct the overall contrastive learning loss function. This achieves the training objective of unsupervised training without community labels.
[0035] Step 5, "Task 2: Node Classification," specifically includes:
[0036] Note: The node types mentioned in this invention refer to different types of nodes in a heterogeneous system, such as author-type nodes, paper-type nodes, site-type nodes, etc.; the node classification mentioned in this invention refers to the classification of a type of node based on features. For example, in author-type nodes, they can be classified according to research field into large language models, computer vision, and databases, etc.
[0037] Step 503: Represent the lower order The input is fed into the node classification layer to obtain the classification prediction result. .
[0038] Step 504: Based on the classification results and real node classification labels Construct the loss function for the node classification task. It is used to assist in model training.
[0039] Step 505: Finally, the contrastive learning and node splitting tasks are combined to obtain the overall training loss function. .
[0040] This concludes the training phase, which will be followed by the online community search phase.
[0041] Step 6: Use the trained model to generate community representations for target type nodes in the data graph, and predict the node classification results of these nodes. Store this as an auxiliary feature of the data graph.
[0042] Step 601: Using the trained model, generate high-order embedding representations for all target type nodes in the given data graph used for querying, as their community representations. It is stored as a lightweight auxiliary feature of the data graph.
[0043] Step 602: Using the trained model, predict the node classification result of the target type node. It is stored as a lightweight auxiliary feature of the data graph.
[0044] Step 7: For a query node, calculate the community score of the node and other nodes based on the community representation, and generate the community query result by combining the node classification consistency and structural connectivity constraints.
[0045] Step 701: Calculate the similarity score between the community representations of the query node q and other nodes based on the community representation set, and use this score as its community score, denoted as qq. .
[0046] Step 702: Based on the community score, select nodes with a similarity greater than or equal to a given threshold to form a preliminary community candidate set. .
[0047] Step 703: Based on the predicted node classification results, in the candidate set In the process, nodes with the same classification results as the target node are further selected.
[0048] Step 704: Finally, retain the set of nodes in the candidate set that are connected to the query node in the global isomorphic graph, as the result of the community search. .
[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0050] 1. This invention designs a semantic feature construction without learnable parameters, which can be used for graph-less and efficient feature learning.
[0051] 2. This invention designs an efficient, scalable semantic graph transformation model that does not rely on graph structure propagation. It can simultaneously model the low-order local features and high-order global features of query nodes, improve the ability to characterize the multi-scale semantics of community structure, and has good training and inference efficiency on large-scale graphs.
[0052] 3. This invention constructs an offline training strategy that does not rely on community labels, realizing effective learning of community representations under unsupervised conditions, and can better adapt to the actual situation where community labels are scarce or unavailable in reality.
[0053] 4. This invention further proposes a multi-constraint online query strategy that combines node similarity, semantic consistency and structural connectivity conditions to achieve efficient, accurate and robust community search of query nodes, and has strong practicality and generalization ability. Attached Figure Description
[0054] Figure 1 This is a flowchart of the method of the present invention;
[0055] Figure 2 This is a structural diagram of the method and model designed for this invention. Detailed Implementation
[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0057] like Figure 1 As shown, this invention proposes a scalable and unsupervised heterogeneous information network community search method. The method and model structure of the invention are as follows: Figure 2 As shown, it includes the following steps:
[0058] Step 1: Define the target node type and the meta-path based on the target type node, and construct isomorphic graphs under different meta-paths accordingly.
[0059] Step 1 specifically includes:
[0060] Step 101: Select the target node type from the heterogeneous information network (For example, in a DBLP heterogeneous information network, there are author-type nodes (A), paper-type nodes (P), location-type nodes (V), and research field-type nodes (T). Assume the target node type is author-type.) A set of meta-paths related to the target type node is defined. (For example, author class nodes can be connected through metapath APA, and the metapath set can include APA, APTPA, etc.).
[0061] Step 102: Construct the corresponding isomorphic graph based on each meta-path P and the original heterogeneous graph. Each isomorphic graph contains only nodes of the target type and their semantic neighbor relationships connected by metapaths. For example, based on the metapath P=APA, an isomorphic graph is constructed where all nodes are of the author class, and connected nodes represent co-author relationships.
[0062] Step 103: Perform step 102 independently for each meta-path to form a set of isomorphic graphs under the meta-path set, which serves as the basis for subsequent feature construction and community modeling.
[0063] Step 2: Integrate the features of other types of nodes in the metapath into the target type node to construct new semantic features for the nodes in the isomorphic graph under each metapath.
[0064] Step 2 specifically includes:
[0065] Step 201: Use singular value decomposition (SVD) to reduce the dimensionality of nodes of different types, mapping their features to a unified dimension to ensure the fusion of cross-type semantic information. For The feature matrix of all nodes of type is transformed into .in for The number of type nodes, The specified transformed dimension.
[0066] Step 202: Under the original path P, place the target type node in the isomorphic graph. The dimensionality-reduced features of semantic neighbors are weighted and averaged, and then concatenated with their original features to form node features with fused semantics. Where N is the number of target type nodes, The concatenated feature dimensions.
[0067] Step 203: Repeat step 202 for all meta-paths to obtain the semantically enhanced feature representations of all target type nodes under each meta-path. .
[0068] Step 3: Introduce a hyperedge construction mechanism to perform structural cohesion screening on multi-hop neighbors, form a high-quality hyperedge set, and generate community feature tokens to represent different views.
[0069] Step 3 specifically includes:
[0070] Step 301: Set the maximum number of hops K and the structural cohesion screening threshold θ for each target type node to control the screening of multi-hop neighbors.
[0071] Step 302: Under the metapath P, with the target node v as the center, based on the isomorphic graph under the metapath In each hop, the set of neighboring nodes is expanded, and those nodes whose overlap with the set of neighbors in the previous hop is not less than θ are retained.
[0072] Step 303: Define the set of neighbors formed by each hop as a hyperedge, and construct a multi-hop hyperedge matrix. This represents the semantic connection relationships of nodes at different scales.
[0073] Step 304: Based on the hyperedge matrix The multi-hop semantic features of nodes are aggregated to generate feature tokens under multiple views. The calculation formula is:
[0074]
[0075] Step 305: Repeat steps 302-304 under all metapaths to obtain the feature token set. This is used for subsequent community representation learning.
[0076] At this point, the construction of feature tokens with no learnable parameters is complete, and the training phase begins.
[0077] Step 4: Input the feature tokens under each meta-path into the attention-based model to extract low-order and high-order community representation information, and realize graph structure-independent feature fusion and representation learning.
[0078] Step 4 specifically includes:
[0079] Step 401: Under the meta-path P, transfer the features The input is first fed into an embedding layer, where it is transformed into a dense layer using a learnable linear transformation. Set type characteristics, calculated using the following formula:
[0080]
[0081] in This is a learnable weight matrix.
[0082] Step 402: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The input is fed into the L-layer Transformer for further feature extraction. The propagation formula for the L-layer Transformer during the propagation process is as follows:
[0083]
[0084]
[0085] MSA stands for Multi-Head Self-Attention Mechanism, and LN stands for Layer Regularization.
[0086] Ultimately, we obtained .
[0087] Step 403: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The input is passed to a multi-view separation layer, and high-order features are extracted from it using an attention-based method. and low-order features The formulas for calculating higher-order aggregation weights and aggregation methods are as follows:
[0088]
[0089]
[0090] Where || represents the vector concatenation operation. This is a learnable weight matrix.
[0091] The formulas for calculating low-order aggregation weights and aggregation methods are as follows:
[0092]
[0093]
[0094] Step 404: Repeat steps 401-403 under each meta-path to obtain the higher-order and lower-order representations under all original paths.
[0095] Step 405: Input all representations obtained in step 404 into the attention-based semantic aggregation layer. The weights of the semantic aggregation layer are calculated as follows:
[0096]
[0097] in , , , which are learnable parameters.
[0098] Based on the semantic aggregation weights, the embedding representations under each meta-path are fused to obtain the global high-order representation. and lower-order representation The calculation formula is as follows:
[0099]
[0100] Step 5: Using a dual-task training strategy, two training tasks, "Task 1: Unsupervised contrastive learning" and "Task 2: Node classification", are carried out simultaneously under the condition of no community label supervision, guiding the model to obtain a more reasonable community embedding representation.
[0101] Step 5, "Task 1: Unsupervised Comparative Learning," specifically includes:
[0102] Step 501: Calculate the semantic contrastive learning loss function under the metapath P. :
[0103]
[0104] in It refers to the set of positive samples of query node v under metapath P, which are contained in the isomorphic graph based on metapath and are neighboring nodes that have the same node classification label as node v. This refers to the set of negative samples, which are nodes that are more than K hops away from the query node or have different node classification labels, contained in the metapath-based isomorphic graph. sim(.) is the cosine similarity. It is the margin parameter in the margin triplet loss.
[0105] Step 502: For all training nodes, repeat step 501 under each meta-path, and sum the loss functions to obtain the semantic contrastive learning loss function. :
[0106]
[0107] in This represents the number of training nodes.
[0108] Step 503: Calculate the global contrastive learning loss function for the higher-order and lower-order representations after semantic aggregation. This guides the model to bring similar nodes closer together and distinguish unrelated nodes in the semantic space. The global contrastive learning loss function is calculated as follows:
[0109]
[0110] in The set of positive samples representing the global model is contained in the global isomorphism graph. In a globally isomorphic graph, nodes that are connected and share the same node category label are considered to be connected. It is based on the isomorphic graph under the metapath The constructed condition is that if two nodes are in any... If connected in the middle, then in Connected in the middle. Represents the global set of negative samples, contained in Query nodes that are more than K hops away from the mid-range node or nodes that have different node category labels.
[0111] Step 504: Combine the two loss functions from steps 502 and 503 to construct the overall contrastive learning loss function. To achieve the training objective without community label supervision, the calculation method is as follows:
[0112]
[0113] in For weight hyperparameters.
[0114] Step 5, "Task 2: Node Classification," specifically includes:
[0115] Note: Node type refers to different types of nodes in a heterogeneous system, such as author-type nodes, paper-type nodes, site-type nodes, etc.; the node classification mentioned in this invention refers to the classification of a type of node based on features. For example, in author-type nodes, they can be classified according to research field as large language models, computer vision, and databases, etc.
[0116] Step 505: Represent the lower order The input is fed into the node classification layer to obtain the classification prediction result. The calculation method is as follows:
[0117]
[0118] in and These are learnable parameters. It is the sigmod activation function.
[0119] Step 506: Based on the classification results and real node classification labels Construct the loss function for the node classification task. This is used to assist in model training. The loss function is calculated as follows:
[0120]
[0121] in The number of categories for classifying nodes of the target type.
[0122] Step 507: Finally, the contrastive learning and node splitting tasks are combined to obtain the overall training loss function. The calculation method is as follows:
[0123]
[0124] in For weight hyperparameters.
[0125] This concludes the training phase, which will be followed by the online community search phase.
[0126] Step 6: Use the trained model to generate community representations for target type nodes in the data graph, and predict the node classification results of these nodes. Store this as an auxiliary feature of the data graph.
[0127] Step 6 specifically includes:
[0128] Step 601: Using the trained model, generate high-order embedding representations for all target type nodes in the given data graph used for querying, as their community representations. It is stored as a lightweight auxiliary feature of the data graph.
[0129] Step 602: Using the trained model, predict the node classification result of the target type node. It is stored as a lightweight auxiliary feature of the data graph.
[0130] Step 7: For a query node, calculate the community score of the node and other nodes based on the community representation, and generate the community query result by combining the node classification consistency and structural connectivity constraints.
[0131] Step 7 specifically includes:
[0132] Step 701: Calculate the similarity score between the community representations of the query node q and other nodes based on the community representation set, and use this score as its community score, denoted as qq. The calculation method is as follows:
[0133]
[0134] Step 702: Based on the community score, select nodes with a similarity greater than or equal to a given threshold to form a preliminary community candidate set. .
[0135] Step 703: Based on the predicted node classification results, in the candidate set In the process, nodes with the same classification results as the target node are further selected.
[0136] Step 704: Finally, retain the candidate nodes that are isomorphic to the query nodes in the global graph. The set of nodes connected in the middle is used as the result of the community search. .
[0137] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.
[0138] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. An extensible and unsupervised heterogeneous information network community search method, characterized in that: Step 1: setting the target node type, and based on the meta-path of the target type node, and constructing the isomorphic graph under different meta-paths; Step 2: the characteristics of other type nodes in the meta-path are integrated into the target type node, and new semantic features are constructed for the nodes in the isomorphic graph under each meta-path; Step 3: introducing a super-edge construction mechanism to perform structural cohesion screening on multi-hop neighbors to form a super-edge set, and generating a community feature Token representing different views; Step 4: inputting the feature Token under each meta-path into a model based on the attention mechanism to extract low-order and high-order community representation information, realizing feature fusion and representation learning independent of the graph structure; Step 5: through a double-task training strategy, simultaneously performing two training tasks of "task 1: unsupervised contrastive learning" and "task 2: node classification" under the condition of no community label supervision, guiding the model to obtain community embedding representation; Step 6: using the trained model to generate community representation for the target type nodes in the data graph, and predicting the node classification results of these nodes, which are stored as auxiliary features of the data graph; Step 7: for the query node, calculating the community score of the node and other nodes according to the community representation, and combining the node classification consistency and structural connectivity constraint conditions to generate the community query result. In step 1, the specific method is: Step 103: independently performing step 102 for each meta-path to finally form an isomorphic graph set under the meta-path set, which serves as the basis for subsequent feature construction and community modeling. In step 2, the specific method is: Step 201: using dimensionality reduction methods such as singular value decomposition to map the features of different types of nodes to a unified dimension to ensure the fusion of cross-type semantic information; In step 3, the specific method is: Step 301: setting the maximum hop number K and the structural cohesion screening threshold θ for each target type node to control the screening of multi-hop neighbors; At this point, the feature Token construction without learnable parameters is completed, and then the training phase follows. In step 4, the specific method is: Step 404: repeating steps 401-403 under each meta-path to obtain high-order representation and low-order representation under all original paths; In step 5, "task 1: unsupervised contrastive learning" specifically includes: In step 5, "task 2: node classification" specifically includes: Node type refers to different types of nodes in a heterogeneous body, and node classification refers to classification according to features in a type of node; At this point, the training phase is completed, and then the online community search phase follows. In step 6, it specifically includes: In step 7, it specifically includes: 2.The scalable and unsupervised community search method of heterogeneous information networks of claim 1, wherein, Step 101: Select target node type from heterogeneous information network and set a set of meta-path sets related to the target type node ; Step 102: constructing a corresponding isomorphic graph according to each meta-path P and the original heterogeneous graph wherein each isomorphic graph only contains target type nodes and their semantic neighbor relations connected by the meta-path; 3. The scalable and unsupervised heterogeneous information network community search method of claim 1, wherein, Step 202: under the original path P, the semantic neighbors of the target type node in the homogeneous graph are weighted and averaged, and the original features of the target type node itself are spliced to form the fused semantic node features ; Step 203: repeat step 202 for all meta-paths to obtain the semantic enhanced feature representation of all target type nodes under each meta-path .
4. The scalable and unsupervised heterogeneous information network community search method of claim 1, wherein, Step 302: under the meta-path P, centering on the target node v, based on the isomorphic graph under the meta-path The neighbor node set is expanded in each hop, and those nodes with an overlap degree not less than θ with the last-hop neighbor set are reserved; Step 303: define the neighbor set formed by each hop as a super-edge, and construct a multi-hop super-edge matrix to represent the semantic connection relationship of nodes at different scales; Step 304: based on the hyper-edge matrix , the multi-hop semantic features of the nodes are aggregated to generate feature tokens under multi-view ; Step 305: repeat steps 302-304 under all meta-paths to obtain a token set of features for subsequent community representation learning; 5. The scalable and unsupervised heterogeneous information network community search method of claim 1, wherein, Step 401 : Under the meta-path P, features are first input to an embedding layer, which transforms them into dense type features using a learnable linear transformation; Step 402: inputting the feature vector into the L-layer Transformer to further extract features to obtain ; and ; Step 403: obtaining the high-order feature and the low-order feature from the input multi-view separation layer by using the attention-based method ; Step 405: transmit all representations obtained in step 404 to an attention-based semantic aggregation layer, fuse the embedding representations under each meta-path to obtain a global high-order representation and low-order representation .
6. The scalable and unsupervised heterogeneous information network community search method of claim 1, wherein, Step 501: Under the meta-path P, calculate the semantic contrastive learning loss function ; Step 502: For all training nodes, repeat step 501 under each meta-path and accumulate the loss function to obtain a semantic contrast learning loss function ; Step 503: calculating a global contrastive learning loss function for the high-order representation and the low-order representation after semantic aggregation , guiding the model to pull similar nodes closer and distinguish irrelevant nodes in the semantic space; Step 502: fuse the two loss functions in steps 502 and 503 to construct an overall contrastive learning loss function , and achieve the training goal without community label supervision.
7. The scalable and unsupervised heterogeneous information network community search method of claim 1, wherein, Step 503: inputting the low-order representation to the node classification layer to obtain a classification prediction result ; Step 504: Based on the classification results and real node classification labels Construct the loss function for the node classification task. Used to assist in model training; Step 505: Finally, the contrastive learning and node split tasks are fused to obtain an overall training loss function ; 8. The scalable and unsupervised heterogeneous information network community search method of claim 1, wherein, Step 601: using the trained model, generating high-order embedding representation of all target type nodes in the given data graph for query as their community representation and storing as a lightweight auxiliary feature of the data graph; Step 602: using the trained model, predicting the node classification result of the target type node and stored as a lightweight auxiliary feature of the data graph.
9. The scalable and unsupervised heterogeneous information network community search method of claim 1, wherein, Step 701: Compute the similarity scores of the query node q with the community representations of other nodes according to the community representation set, as its community scores, denoted as ; Step 702: According to the community score, select nodes with similarity greater than or equal to a given threshold to form a preliminary community candidate set ; Step 703: According to the predicted node classification result, further filter out the nodes with the same classification result as the target node from the candidate set Step 704: Finally, retain the set of nodes in the candidate set that are connected to the query node in the global isomorphic graph as the result of the community search .
