A network entity identification method based on heterogeneous graph learning

Abstract

The application discloses a network entity identification method based on a heterogeneous graph learning, relates to the technical field of network analysis, and comprises the following steps: constructing a heterogeneous graph structure model, performing feature fusion on each mode in an optimized network data set, and generating a node representation vector; performing multilayer propagation on the node representation vector and an aggregation weight by using a multilayer heterogeneous graph neural network mechanism, and outputting an updated node feature set; inputting the updated node feature set into a two-layer fully connected neural network classifier to perform entity category prediction, and generating a network entity identification report. By constructing the heterogeneous graph structure model and combining an edge type identification analysis method to distinguish heterogeneous edges and homogeneous edges, fine description of diversified connection relationships in a complex network is realized, the heterogeneous graph structure model can more accurately reflect actual interaction modes between nodes, and the expression capability of graph structure modeling is improved.

Description

Technical Field

[0001] This invention relates to the field of network analysis technology, and in particular to a network entity recognition method based on heterogeneous graph learning. Background Technology

[0002] In the fields of graph data modeling and network entity recognition, traditional recognition methods mainly rely on a combination of single-modal feature extraction and homogeneous graph structure modeling for node classification tasks. These methods typically begin by modeling the nodes and edges in the network structure, extracting their topological relationships, and then encoding features using multi-source information such as text and images. Graph neural networks or adjacency matrix-based propagation mechanisms are then used for node representation learning, updating the feature representation of the central node by aggregating information from neighboring nodes. Finally, a classifier is used to perform discriminative analysis on the node features, achieving the identification and prediction of entity categories in the network. These methods have been widely applied in areas such as social network analysis, recommender systems, and knowledge graph construction.

[0003] However, existing methods still have limitations in handling heterogeneous graph structures and multimodal information fusion. On the one hand, most models are based on the isomorphic graph assumption, that is, similar nodes tend to connect with each other, which makes it difficult to adapt to the heterogeneous structures that are widespread in real networks, leading to increased noise interference during information propagation. On the other hand, in the process of multimodal data fusion, simple splicing or weighted averaging strategies are usually adopted, lacking in-depth modeling of semantic differences and correlations between different modalities, which affects the expressive power and discriminative power of node representation vectors. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a network entity recognition method based on heterogeneous graph learning to solve the problems of insufficient heterogeneous graph modeling ability and weak multimodal information fusion and expression ability in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a network entity recognition method based on heterogeneous graph learning, comprising: collecting network structure data and multimodal data and preprocessing them to generate an optimized network dataset; constructing a heterogeneous graph structure model, performing feature fusion on each modality in the optimized network dataset to generate node representation vectors; obtaining the cosine similarity between the center node and associated nodes of the node representation vectors, and combining the decay coefficient of the edge type of the center node and associated nodes, generating aggregate weights through a dynamic threshold filtering mechanism; using a multi-layer heterogeneous graph neural network mechanism to propagate the node representation vectors and aggregate weights through multiple layers, outputting an updated node feature set; inputting the updated node feature set into a two-layer fully connected neural network classifier for entity category prediction, and generating a network entity recognition report.

[0007] As a preferred embodiment of the network entity recognition method based on heterogeneous graph learning described in this invention, the network structure data includes node data, edge data, and graph-level data; The multimodal data includes text data, image data, and time series data; The preprocessing includes data cleaning, normalization, and handling of missing values.

[0008] As a preferred embodiment of the network entity recognition method based on heterogeneous graph learning described in this invention, the generation of the optimized network dataset refers to the structured integration of preprocessed network structure data and multimodal data to generate an optimized network dataset containing complete node and edge information.

[0009] As a preferred embodiment of the network entity recognition method based on heterogeneous mapping graph learning described in this invention, the specific steps for constructing the heterogeneous mapping graph structure model are as follows: The graph initialization method is used to construct an initial graph structure for each node and each edge in the optimized network dataset; Use the edge type identifier parsing method to determine the type of each edge in the initial graph structure and distinguish between disjoint edges and homojoint edges; Based on the type of each edge, the weight and attenuation coefficient of the edge are set. The edge-guided fusion method is used to integrate the nodes, edge types, edge weights and attenuation coefficients in the optimized network dataset to form a heterogeneous graph structure model.

[0010] As a preferred embodiment of the network entity recognition method based on heterogeneous graph learning described in this invention, the specific steps for performing feature fusion on each modality in the optimized network dataset to generate node representation vectors are as follows: The heterogeneous graph structure model is used to extract data for each modality from the optimized network dataset, and feature encoding is performed to convert each multimodal data into a multi-source modal feature vector. The cross-modal fusion method is then used to integrate them into a unified node representation vector.

[0011] As a preferred embodiment of the network entity recognition method based on heterogeneous graph learning described in this invention, the specific steps of obtaining the cosine similarity between the center node and associated nodes of the node representation vector, and combining it with the decay coefficient of the edge type of the center node and associated nodes, are as follows: The cosine similarity between the center node and associated nodes of the node representation vector is obtained using the node similarity extraction method, forming a cosine similarity set; The cosine similarity set is weighted and attenuated by a heterogeneous edge weighting method to generate a weighted similarity set.

[0012] As a preferred embodiment of the network entity recognition method based on heterogeneous mapping graph learning described in this invention, the following steps are taken: An aggregation weight is generated through a dynamic threshold filtering mechanism. Using a dynamic threshold filtering mechanism, neighboring nodes with high similarity are selected from the weighted similarity set; The filtered neighboring nodes and their weighted similarity values ​​are normalized to generate aggregate weights.

[0013] As a preferred embodiment of the network entity recognition method based on heterogeneous graph learning described in this invention, the specific steps of using a multi-layer heterogeneous graph neural network mechanism to propagate node representation vectors and aggregation weights through multiple layers and output an updated node feature set are as follows: The node representation vector and aggregation weights are input into a multi-layer heterogeneous graph neural network mechanism; Multi-layer propagation includes a message generation layer, a message aggregation layer, and an update layer; The message generation layer uses adjacency feature extraction to extract the adjacency node feature vectors of associated nodes in the node representation vectors based on representation vectors and aggregation weights, and then weights and fuses them with the aggregation weights to form a message set. The message aggregation layer uses an edge-weighted average method to sum the message set according to the edge weights, generating an aggregation vector. The update layer uses the GRU feature update method to concatenate the center point feature vectors of the aggregate vector and the node representation vector and input them into the gated recurrent unit to generate the updated node feature set.

[0014] As a preferred embodiment of the network entity recognition method based on heterogeneous graph learning described in this invention, the step of inputting the updated node feature set into a two-layer fully connected neural network classifier for entity category prediction specifically includes the following steps: The first fully connected layer of the two-layer fully connected neural network classifier performs linear transformation and non-linear activation on the representation vector of each node in the updated node feature set to generate an intermediate representation vector set. The second fully connected layer performs a linear mapping on the intermediate representation vector set and outputs the node classification probability distribution.

[0015] As a preferred embodiment of the network entity recognition method based on heterogeneous graph learning described in this invention, the specific steps for generating the network entity recognition report are as follows: The predicted category label and confidence score of each node in the node classification probability distribution are extracted by the probabilistic category selection method. The information is then organized in a standardized format to statistically analyze the distribution of the number of nodes and the confidence score of different category labels, and a network entity recognition report is generated.

[0016] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the network entity recognition method based on heterogeneous graph learning as described in the first aspect of the present invention.

[0017] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the network entity recognition method based on heterogeneous graph learning as described in the first aspect of the present invention.

[0018] The beneficial effects of this invention are as follows: By constructing a heterogeneous graph structure model and combining it with the edge type identifier resolution method to distinguish between heterogeneous and homogeneous edges, a fine characterization of diverse connection relationships in complex networks is achieved, enabling the heterogeneous graph structure model to more accurately reflect the actual interaction patterns between nodes and improving the expressive power of graph structure modeling; by using a dynamic threshold screening mechanism to select highly similar adjacent nodes from the weighted similarity set and generating normalized aggregation weights, adaptive optimization selection of adjacency information is achieved, enhancing the focus and robustness of information propagation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Fig. 1 This is a flowchart of a network entity recognition method based on heterogeneous graph learning.

[0021] Fig. 2 This is a schematic diagram of a network entity recognition system architecture that integrates heterogeneous graph structures and multi-layer graph neural networks.

[0022] Fig. 3 A flowchart of the method for constructing a heterogeneous graph structure model.

[0023] Fig. 4This is a flowchart of a multi-layer heterogeneous graph neural network mechanism. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0027] Reference Figs. 1-4 This is one embodiment of the present invention, which provides a network entity recognition method based on heterogeneous mapping graph learning, including the following steps: S1. Collect network structure data and multimodal data and preprocess them to generate an optimized network dataset.

[0028] S1.1 Network structure data includes node data, edge data, and graph-level data; Specifically, network structure data is collected by crawling technologies or using existing APIs to extract node (e.g., users, products) and edge (e.g., friend relationships, purchasing behavior) information from social networks, web links, or other relationship graphs. Node data includes node identifiers, node types, and node attribute information. For example, a node identifier of user_001, a node type of user, and node attributes including registration time and activity level; edge data includes edge identifiers, starting node identifiers, ending node identifiers, edge direction, and edge type. For example, an edge identifier of edge_001, a starting node identifier of user_001, an ending node identifier of item_203, an edge direction of "yes," and an edge type of "purchase"; graph-level data includes graph identifiers and global graph attribute information. For example, a graph identifier of graph_2025, and global graph attributes including the platform and graph construction time; S1.2 Multimodal data includes text data, image data, and time series data; Specifically, for multimodal data, corresponding data types are acquired through channels such as text sources (e.g., comments, articles), image libraries (e.g., user-uploaded photos), and sensor data (e.g., location information). The text data includes user comments, node descriptions, label texts, and summary information; for example, a description field of a certain node being "electronics enthusiast, focusing on mobile phone evaluation" is a piece of text data; the image data includes node avatars, icons, screenshots, and charts; for example, the avatar file of a certain user is user_001.jpg; the time series data includes access records, click behaviors, and transaction logs, which refer to data arranged in chronological order; for example, the access records of a certain user within one week are 2025-05-14, 3 times; 2025-05-15, 5 times.

[0029] S1.3. The preprocessing includes data cleaning, normalization, and missing value processing; It should be noted that field identification is performed on node attributes and edge attributes in the network structure data to clarify the semantics and data type of each attribute, including three types of attributes: numerical, categorical, and textual. For example, in the node attributes, the user age is numerical, and the gender is categorical; in the edge attributes, the interaction frequency is numerical, and the edge type is categorical; Outliers beyond a reasonable range are detected and removed for numerical attributes, for example, records with a user age less than 0 or greater than 120 are marked as abnormal and deleted; the category name representations are unified for categorical attributes, for example, "Male" and "male" in the gender field are unified as "male"; meaningless characters and whitespace characters are removed for textual attributes, for example, consecutive spaces and special symbols in the comment content are cleaned; The min-max normalization method is adopted for all numerical attributes, and the values are mapped to the interval [0, 1]. For example, the "commodity rating" is linearly transformed from the original interval [1, 5] to [0, 1]; a filling strategy based on global statistical values is adopted for missing node attributes or edge attributes, the numerical attributes are filled with the mean value of the corresponding attribute in all samples, for example, the missing value is filled with the average value of all known "user ages"; the categorical attributes are filled with the category with the highest occurrence frequency, for example, the missing value is filled with "male", the most common value in the "gender" field; When the textual attribute of a certain node is missing, the textual attribute content corresponding to the edge that is closely connected to the node among adjacent nodes is searched, and the text fragment with the highest occurrence frequency is selected as the supplementary content of the missing value; for example, if the "user comment" attribute of the node with the node number 0 is missing, the "user comment" contents of adjacent node 1 and node 2 are searched; For text modal data, remove meaningless characters, duplicate content, and whitespace, such as cleaning up consecutive spaces and special symbols in comment content; for image modal data, detect and remove corrupted files or images with a resolution below a set threshold, such as removing images smaller than 64×64 pixels; use statistical methods to detect and remove outliers in time series; for example, in a time series of "daily login counts," if the number of logins on a certain day is much higher than the normal range (exceeding the mean plus three standard deviations), it is marked as an outlier and considered for deletion or replacement.

[0030] For multimodal data, a min-max normalization method is used to map the value range to the interval [0, 1]. For example, "product rating" is linearly transformed from the original interval [1, 5] to [0, 1]. For missing modal data, a filling strategy based on node association information is used. If a node is missing text modal data, the same modal content in neighboring nodes is searched, and the text fragment with the highest similarity is selected for filling. If a node is missing image modal data, the most common image modality representative sample under the node's category is used for replacement. For missing values ​​in the time series, an appropriate filling strategy is selected according to the specific situation. For continuous time series, interpolation methods (such as linear interpolation and spline interpolation) can be used. For example, when there are missing values ​​in the "user browsing time" time series, linear interpolation can be used to fill the gaps.

[0031] S1.4. The preprocessed network structure data and multimodal data are structurally integrated to generate an optimized network dataset containing complete node and edge information. Furthermore, based on the unique identifier of each node in the preprocessed network structure data, a mapping table is constructed between node identifiers and entity identifiers in the multimodal data. For example, node identifier 0 corresponds to user A, and node identifier 1 corresponds to product B, ensuring that each multimodal data item can be accurately associated with its corresponding node. Based on the node identifier mapping relationship, the multimodal data is aligned to the corresponding node identifiers. For example, user A's review text "The product is very good" is mapped to node identifier 0, and the image feature vector of product B is mapped to node identifier 1. The network structure data of each node is merged with the multimodal data of the corresponding node to form an enhanced set of node attributes. During the integration process, the edge set in the network structure data remains unchanged, including the start and end numbers of each edge and the edge type information; for example, the edge set records that the edge type from node number 0 to node number 1 is "purchase"; for example, node number 0 contains basic attributes (such as user age 25 years old) and multimodal attributes, such as the comment text "good product", and the browsing duration sequence [0.7, 0.8]. The enhanced set of node attributes and edge sets are organized in a unified format to generate an optimized network dataset containing complete node and edge information. For example, the output contains 3 nodes with node numbers 0, 1, and 2, corresponding to different types of modal information, and edge sets including 0, 1 (following) and 0, 2 (purchasing).

[0032] S2. Construct a heterogeneous graph structure model, perform feature fusion on each modality in the optimized network dataset, and generate node representation vectors.

[0033] S2.1 Use graph initialization to construct an initial graph structure for each node and each edge in the optimized network dataset; Specifically, network structure data and multimodal data of all nodes in the optimized network dataset are extracted; based on the network structure data of the nodes, the start-point number, end-point number, and edge type of each edge are identified; An adjacency list is constructed to record the numbers of all adjacent nodes for each node. The set of all edges in the optimized network dataset is traversed. The adjacency list is constructed by reading the start and end numbers of each edge, adding the end number to the adjacency list corresponding to the start number. If the edge is undirected, the start number is also added to the adjacency list of the end number. For example, if there is an edge (0, 1, attention) in the edge set, then 1 is added to the adjacency list of node 0; if the edge is undirected, then 0 is added to the adjacency list of node 1. An edge index table is created to record the start-point number, end-point number, and edge type of each edge. This is achieved by extracting the start-point number, end-point number, and edge type for each edge and sequentially writing the triples (start-point number, end-point number, edge type) into the edge index table, ensuring the complete preservation of each edge's information. For example, when processing the edge set [(0, 1, Follow), (0, 2, Purchase)], each triple is stored as an independent entry in the edge index table, resulting in a final edge index table containing the connection relationships and type identifiers of all edges. The node set, edge set, and edge type information are organized into a standard graph structure format for output. For example, the output contains 3 nodes with node numbers 0, 1, and 2, and edge sets including 0,1,Follow, and 0,2,Purchase.

[0034] S2.2 Use the edge type identifier parsing method to determine the type of each edge in the initial graph structure and distinguish between disjoint edges and homojoint edges; It should be noted that a complete set of edges is extracted from the initial graph structure, including the start number, end number, and edge type of each edge; for example, the edge set contains (0, 1, Follow) and (0, 2, Purchase). Based on the basic attribute information of each node in the optimized network dataset, the category label of each node is determined. For example, nodes numbered 0 and 1 are classified as "user" based on the basic attribute, and node numbered 2 is classified as "product" based on the basic attribute. The edge type attribute is determined based on the category label of each node. If the starting node and the ending node belong to the same category (such as user-user, product-product), the edge is marked as a homologous edge; if the starting node and the ending node belong to different categories (such as user-product, product-user), the edge is marked as a heterologous edge. For example, the edge (0, 1, Follow) connects two nodes of the "user" category and is marked as a homologous edge; the edge (0, 2, Purchase) connects nodes of the "user" and "product" categories and is marked as a heterologous edge.

[0035] S2.3. Based on the type of each edge, set the weight and attenuation coefficient of the edge, and use the edge-guided fusion method to integrate the nodes, edge types, edge weights and attenuation coefficients in the optimized network dataset to form a heterogeneous graph structure model. Specifically, based on the result of each edge being marked as a homologous edge or a heterologous edge in the initial graph structure, corresponding edge weights and attenuation coefficients are assigned. Two new columns are added to the edge index table to record the edge weight and attenuation coefficient of each edge, forming a complete edge information enhancement table containing the start point number, end point number, edge type, edge attribute type, edge weight, and attenuation coefficient. For each node, the edge information enhancement table is traversed to collect information on all adjacent nodes, including adjacent node numbers, corresponding edge weights, attenuation coefficients, and modal characteristics of the adjacent nodes. For example, a node with node number 0 has two adjacent nodes: node 1 and node 2, corresponding to different edge weights, attenuation coefficients, and modal characteristics. For each node, an edge-guided fusion method is used. Based on the edge weights and attenuation coefficients of adjacent edges, the modal features of adjacent nodes are weighted and aggregated. The weighted modal features of all adjacent nodes are summed and concatenated with the modal features of each node itself to form an enhanced node representation vector. For example, the modal features of node number 0 are [0.3, 0.5], the weighted features of adjacent node 1 are [0.2, 0.4], and the weighted features of adjacent node 2 are [0.1, 0.3]. The final concatenated node representation vector is [0.3, 0.5, 0.2, 0.4, 0.1, 0.3]. The enhanced node representation vectors are organized into a unified format according to the node number order, while retaining the complete edge information enhancement table, together forming a heterogeneous graph structure model; for example, the output includes a heterogeneous graph structure model file and a node representation vector with node number 0 [0.3, 0.5, 0.2, 0.4, 0.1, 0.3].

[0036] S2.4. Extract data for each modality from the optimized network dataset using the heterogeneous graph structure model, and perform feature encoding to convert the modality data into multi-source modality feature vectors. Furthermore, complete node and edge set information is read from the constructed heterogeneous graph structure model, including the node number, basic attributes, adjacent node list, and the type, weight, and decay coefficient of each edge; for example, the basic attribute of node number 0 is "user", the adjacent nodes are 1 (product) and 2 (user), and the corresponding edges are heterogeneous edges and homogeneous edges, respectively; based on the node number, all modal data associated with the node number are extracted from the optimized network dataset; for example, the node with node number 0 contains different types of modal data such as text modality (comment content), image modality (avatar image), and numerical sequence modality (browsing time sequence); Standard feature encoding methods are used to process different types of modal data: text modal data uses word vector models (such as Word2Vec or BERT) to map comment text into fixed-dimensional feature vectors; image modal data uses convolutional neural networks (such as ResNet-18) to extract global feature vectors of images; numerical sequence modal data generates fixed-length feature representations through statistical feature extraction (such as mean, variance) or sliding window concatenation; for example, a comment text is encoded as a 768-dimensional BERT vector, and a profile picture is encoded as a 512-dimensional ResNet feature vector; The various modal feature vectors of each node are concatenated in a unified format to form a multi-source modal feature vector. The multi-source modal feature vector refers to extracting all associated modal data (such as text, images, numerical sequences, etc.) for each node, and using the corresponding encoding method to convert each modal data into a fixed-dimensional feature vector. Then, the feature vectors of different modalities are concatenated together in a certain order to form a unified feature vector containing multi-modal information.

[0037] S2.5. Employ a cross-modal fusion method to integrate multi-source modal feature vectors into a unified node representation vector; It should be noted that normalization is performed on the multi-source modality feature vectors to make the data from different modalities comparable during fusion. The feature vector of each modality in the multi-source modality feature vectors is mapped to the [0,1] region. For example, the original value range of the image modality feature vector is [0.1, 0.9]. After normalization, it still maintains the same distribution but the values ​​are more stable. Fixed weights are set according to the importance of the modalities and used for weighted fusion of feature vectors from different modalities. The importance of a modality refers to the contribution and relevance of each modality's data to achieving the goal in a specific task. This is determined based on the amount of information carried by the modality, its unique role in the task, its impact on model performance, and domain expertise. Through a weighted fusion mechanism, the feature vectors of each modality are integrated according to their contribution based on the corresponding fixed weights to obtain a weighted fusion vector. The weighted fusion intermediate vector is input into a fixed-dimensional fully connected layer and compressed into a node representation vector of uniform length. For example, a fully connected layer is used to compress a 128-dimensional fusion vector into a 128-dimensional unified node representation vector. The node representation vectors generated after cross-modal fusion of all nodes are organized into a uniform format according to the node number order.

[0038] S3. Obtain the cosine similarity between the center node and the associated nodes of the node representation vector, and combine it with the decay coefficient of the edge type of the center node and the associated nodes to generate aggregate weights through a dynamic threshold filtering mechanism.

[0039] S3.1. Use the node similarity extraction method to obtain the cosine similarity between the center node and the associated nodes of the node representation vector, and form a cosine similarity set; Specifically, obtain the node representation vectors of each central node and its associated nodes in the adjacency list from the unified node representation vector; Calculate the cosine similarity for the center node and related nodes respectively, using the following expression:

[0040] in, The representative node represents the center node of the vector. Representative and Related nodes Indicates the central node The node represents a vector. Indicates associated nodes The nodes represent vectors; It is the central node Associated nodes The nodes represent the cosine similarity between vectors; Organize the cosine similarity between the central node and all associated nodes into a set according to the order of the associated nodes. For example, the cosine similarity set of the central node with number 0 is [0.85, 0.62], which corresponds to the similarity values ​​of the adjacent nodes with numbers 1 and 2, respectively. S3.2. The cosine similarity set is weighted and attenuated using the heterogeneous edge weighting method to generate a weighted similarity set; Furthermore, the cosine similarity between each central node and its associated nodes is read from the cosine similarity set, and the attenuation coefficient corresponding to each edge is extracted from the edge information enhancement table. The cosine similarity value corresponding to each associated node of the central node is multiplied element by element with the attenuation coefficient of the edge to obtain the weighted similarity value. All the weighted similarity values ​​are organized into a set according to the order of the associated nodes to form a weighted similarity set.

[0041] S3.3 Utilize a dynamic threshold filtering mechanism to filter out adjacent nodes with high similarity from the weighted similarity set; Specifically, setting a dynamic threshold refers to determining the threshold based on the mean and standard deviation of the weighted similarity set. The dynamic threshold is usually set as the mean plus a certain multiple of the standard deviation to achieve the filtering effect. The mean and standard deviation are calculated for the weighted similarity set. For example, for the weighted similarity set [0.765, 0.372], the calculated mean is 0.5685 and the standard deviation is 0.1965. The dynamic threshold is set to the mean plus 0.5 times the standard deviation. For example, the dynamic threshold is 0.6668. Only 0.765 in the weighted similarity set is greater than the dynamic threshold, so it is retained. Based on the filtering results, the adjacent node numbers corresponding to the weighted similarity items that meet the conditions are retained. The corresponding adjacent nodes refer to the nodes that are directly connected by edges in the graph structure. For example, in a social network graph, if user A follows user B, then user A and user B are a pair of adjacent nodes, where user A is an adjacent node of user B.

[0042] S3.4 Normalize the filtered adjacent nodes and weighted similarity values ​​to generate aggregate weights.

[0043] It should be noted that the neighboring node numbers and corresponding weighted similarity values ​​retained for each central node are read from the output of the dynamic threshold filtering mechanism. The corresponding weighted similarity value refers to the weighted result of the cosine similarity between the central node and each neighboring node after processing with the heterogeneous edge weighting method; All the weighted similarity values ​​retained under the central node are grouped into a local set; The weighted similarity set is processed using the Softmax normalization method to ensure that all similarity values ​​are converted into a probability distribution form with a sum of 1. The normalized results are then bound one by one to the corresponding adjacent node numbers to form an adjacent node-aggregate weight mapping relationship. S4. Utilize a multi-layer heterogeneous graph neural network mechanism to propagate the node representation vector and aggregation weights through multiple layers, and output the updated node feature set.

[0044] S4.1 Input the node representation vector and aggregation weight into the multilayer heterogeneous graph neural network mechanism; Furthermore, the node representation vector is input as the node feature into the graph convolutional layer of the message generation layer, and the aggregated weights are used as edge weight parameters in the information propagation process. For example, the node feature with node number 0 obtains weighted feature information from its neighbor 1 through the aggregated weights {1:1.0}, and combines it with its own features to generate a new intermediate node representation; S4.2 Multi-layer propagation includes a message generation layer, a message aggregation layer, and an update layer. The message generation layer, based on the representation vector and aggregation weight, uses the adjacency feature extraction method to extract the adjacency node feature vectors of the associated nodes in the node representation vector and multiplies them with the aggregation weight to form a message set. Specifically, the current node representation vector of each central node is read from the node representation vector, and all adjacent node numbers and their corresponding aggregate weight values ​​are extracted from the aggregate weight structure. For each adjacent node number, the corresponding node representation vector is searched and extracted from the node representation vector set as the adjacent node feature vector. The adjacent node feature vector refers to the node representation vector corresponding to the adjacent node directly connected to the current central node in the graph structure. Each central node obtains information from its adjacent nodes, and the node representation vectors of the adjacent nodes are the adjacent node feature vectors. The node representation vector of each adjacent node is multiplied element by element with the aggregate weight value of the adjacent node to obtain the weighted adjacency feature vector; all the weighted adjacency feature vectors are organized into a set according to the order of the adjacent nodes, which serves as the message set of the central node.

[0045] S4.3 The message aggregation layer uses the edge-weighted average method to sum the message set according to the edge weights and generate an aggregation vector; Furthermore, the feature vectors of adjacent nodes corresponding to each central node are obtained from the message set, and the aggregation weights of the edges corresponding to the adjacent nodes are extracted. The aggregation weight of the corresponding edge refers to a value assigned to each edge connecting the central node and the adjacent node in the graph structure according to its importance or relevance. For example, if the aggregation weight of the edge between the central node numbered 0 and the adjacent node numbered 1 is 0.7, when generating the aggregation vector, the feature vector of adjacent node 1 will contribute 0.7 to the aggregation vector calculation of the central node 0. Each element in the feature vector of each adjacent node is multiplied by the aggregate weight of the edge to obtain a weighted feature vector of the adjacent node; all weighted feature vectors of the adjacent nodes are added element by element at corresponding positions to form a fixed-dimensional aggregate vector.

[0046] S4.4 The update layer uses the GRU feature update method to concatenate the center point feature vectors of the aggregate vector and the node representation vector and input them into the gated recurrent unit to generate the updated node feature set.

[0047] It should be noted that the node representation vector of each central node is read from the set of node representation vectors, and the corresponding aggregation vector is obtained from the output of the message aggregation layer. The node representation vector of the central node and the corresponding aggregation vector are concatenated along the feature dimension to form a high-dimensional input vector; for example, concatenating a 128-dimensional node representation vector with a 128-dimensional aggregation vector yields a 256-dimensional concatenated vector. Gated Recurrent Units (GRUs) are used to perform nonlinear transformations on the concatenated vectors. Internal update and reset gates control the flow and fusion of information, outputting a fixed-dimensional hidden state vector. For example, a GRU model outputs a 128-dimensional hidden state vector, serving as the updated node feature of the current center node. These hidden state vectors are then organized in node number order into a unified format of updated node feature sets. Each node feature set contains a set of feature vectors calculated by each node based on its own attributes and relationships with neighboring nodes, expressing the node's role, attributes, and relationships with other nodes in the graph. After one update layer of the graph neural network, each node has a new feature vector, and the set of these feature vectors constitutes the updated node feature set.

[0048] S5. Input the updated node feature set into a two-layer fully connected neural network classifier to predict the entity category and generate a network entity recognition report.

[0049] S5.1 The first fully connected layer of the two-layer fully connected neural network classifier performs linear transformation and non-linear activation on the representation vector of each node in the updated node feature set to generate an intermediate representation vector set; Specifically, the representation vector of each node in the updated node feature set is read, and the weight matrix and bias vector of the first fully connected layer are configured. For example, the weight matrix is ​​set to a 128×64 matrix, and the bias vector is a one-dimensional vector of length 64. The bias vector refers to a set of learnable parameters used to translate and adjust the linear transformation result in the fully connected layer of the neural network. In the first fully connected layer, each output dimension corresponds to a bias term, and each bias term is combined to form a one-dimensional vector called the bias vector. The representation vector of each node is multiplied by the weight matrix, and a bias vector is added to obtain the intermediate feature vector after linear transformation. The intermediate feature vector after linear transformation is then subjected to nonlinear transformation using the ReLU activation function. For example, the negative values ​​in the 64-dimensional linear intermediate feature vector are set to 0, while the positive values ​​remain unchanged. The nonlinear activation results of all nodes after processing by the first fully connected layer are organized into a set of intermediate representation vectors in a unified format according to the node number order.

[0050] S5.2 The second fully connected layer performs a linear mapping on the intermediate representation vector set and outputs the node classification probability distribution; Furthermore, the intermediate representation vectors of all nodes are obtained from the intermediate representation vector set, with each node corresponding to a vector of fixed dimensions; the weight matrix and bias vector of the second fully connected layer are configured. For example, if the task is a three-class classification task, the weight matrix is ​​set to a 64×3 matrix and the bias vector is a one-dimensional vector of length 3; the intermediate representation vectors in the intermediate representation vector set are linearly mapped to the weight matrix of the second layer to obtain the output score vector. The output score vector after linear mapping is normalized using the Softmax function, and converted into a probability distribution vector with a sum of 1, representing the probability that a node belongs to each category; The expression for the second fully connected linear mapping layer is:

[0051] in, Representative node The original classification score vector generated after linear mapping This represents the weight matrix of the second fully connected layer. Representative node The intermediate representation vector after processing by the first fully connected layer This represents the bias vector of the second fully connected layer. The category score vector generated by the second-layer fully connected linear mapping expression is converted into a probability distribution by the Softmax function, transforming the original scores into probability values ​​for each category, which facilitates category prediction.

[0052] S5.3 Extract the predicted category label and confidence level of each node in the node classification probability distribution using the probability category selection method, and organize the information according to the standardized format to statistically analyze the distribution of the number of nodes and the confidence level of different category labels, and generate a network entity recognition report.

[0053] It should be noted that for each node's classification probability distribution, the category corresponding to the highest probability value is selected as the node's predicted category label, and the highest probability value is used as the confidence score. All predicted category labels and confidence scores for all nodes are organized into a set with a uniform format according to node number order.

[0054] The predicted category labels of all nodes are categorized and statistically analyzed to calculate the number of nodes contained in each category. For example, category 0 contains 120 nodes, category 1 contains 85 nodes, and category 2 contains 95 nodes. Based on the nodes in each category, statistical indicators such as the mean, median, and standard deviation of the confidence scores are calculated to form confidence score distribution information. Confidence score distribution information refers to the information obtained in the node classification task by statistically analyzing the distribution of classification confidence scores for all nodes in each predicted category, including statistical indicators such as the mean, median, and standard deviation of the confidence scores for all nodes in the predicted category.

[0055] The predicted category label and confidence score for each node, along with the distribution of node numbers and confidence scores for different category labels, are compiled into structured text or tables to generate a network entity recognition report. For example, the final output might include: a total of 300 nodes, with category 0 accounting for 40%, category 1 for 28.3%, and category 2 for 31.7%, along with detailed distribution statistics for the confidence scores of each category. The detailed distribution statistics of confidence scores refer to the multi-dimensional statistical description of the node confidence scores under each predicted category in the network entity recognition report, demonstrating the reliability and concentration of the predicted category recognition results.

[0056] This embodiment also provides a computer device applicable to the network entity recognition method based on heterogeneous graph learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the network entity recognition method based on heterogeneous graph learning as proposed in the above embodiment.

[0057] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0058] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the network entity recognition method based on heterogeneous mapping learning as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0059] In summary, this invention achieves a fine characterization of diverse connection relationships in complex networks by constructing a heterogeneous graph structure model and combining it with an edge type identifier resolution method to distinguish between heterogeneous and homogeneous edges. This enables the heterogeneous graph structure model to more accurately reflect the actual interaction patterns between nodes, thus improving the expressive power of graph structure modeling. Furthermore, by using a dynamic threshold filtering mechanism to select highly similar neighboring nodes from a weighted similarity set and generating normalized aggregation weights, this invention achieves adaptive optimization selection of adjacency information, enhancing the focus and robustness of information propagation.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A network entity recognition method based on heterogeneous mapping graph learning, characterized in that: include, Collect network structure data and multimodal data and preprocess them to generate an optimized network dataset; Construct a heterogeneous graph structure model, perform feature fusion on each modality in the optimized network dataset, and generate node representation vectors; The cosine similarity between the center node and associated nodes of the node representation vector is obtained, and combined with the decay coefficient of the edge type of the center node and associated nodes, an aggregate weight is generated through a dynamic threshold filtering mechanism. A multi-layer heterogeneous graph neural network mechanism is used to propagate the node representation vector and aggregation weights through multiple layers, and the updated node feature set is output. The updated node feature set is input into a two-layer fully connected neural network classifier to predict entity categories and generate a network entity recognition report.

2. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The network structure data includes node data, edge data, and graph-level data; The multimodal data includes text data, image data, and time series data; The preprocessing includes data cleaning, normalization, and handling of missing values.

3. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The generation of optimized network datasets refers to the structured integration of preprocessed network structure data and multimodal data to generate optimized network datasets containing complete node and edge information.

4. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The specific steps for constructing the heterogeneous graph structure model are as follows: The graph initialization method is used to construct an initial graph structure for each node and each edge in the optimized network dataset; The edge type identifier resolution method is used to determine the type of each edge in the initial graph structure and to distinguish between disjoint edges and homojoint edges; Based on the type of each edge, the weight and attenuation coefficient of the edge are set. The edge-guided fusion method is used to integrate the nodes, edge types, edge weights and attenuation coefficients in the optimized network dataset to form a heterogeneous graph structure model.

5. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The specific steps for performing feature fusion on each modality in the optimized network dataset to generate node representation vectors are as follows: The heterogeneous graph structure model is used to extract data for each modality from the optimized network dataset, and feature encoding is performed to convert each multimodal data into a multi-source modal feature vector. The cross-modal fusion method is then used to integrate them into a unified node representation vector.

6. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The specific steps for obtaining the cosine similarity between the center node and associated nodes of the node representation vector, and combining it with the decay coefficient of the edge type of the center node and associated nodes, are as follows: The cosine similarity between the center node and associated nodes of the node representation vector is obtained using the node similarity extraction method, forming a cosine similarity set; The cosine similarity set is weighted and attenuated by a heterogeneous edge weighting method to generate a weighted similarity set.

7. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The process of generating aggregate weights through a dynamic threshold filtering mechanism involves the following steps: Using a dynamic threshold filtering mechanism, neighboring nodes with high similarity are selected from the weighted similarity set; The filtered neighboring nodes and their weighted similarity values ​​are normalized to generate aggregate weights.

8. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The method utilizes a multi-layer heterogeneous graph neural network mechanism to propagate node representation vectors and aggregation weights through multiple layers, outputting an updated node feature set. The specific steps are as follows: The node representation vector and aggregation weights are input into a multi-layer heterogeneous graph neural network mechanism; Multi-layer propagation includes a message generation layer, a message aggregation layer, and an update layer; The message generation layer uses adjacency feature extraction to extract the adjacency node feature vectors of associated nodes in the node representation vectors based on representation vectors and aggregation weights, and then weights and fuses them with the aggregation weights to form a message set. The message aggregation layer uses an edge-weighted average method to sum the message set according to the edge weights, generating an aggregation vector; The update layer uses the GRU feature update method to concatenate the center point feature vectors of the aggregate vector and the node representation vector and input them into the gated recurrent unit to generate the updated node feature set.

9. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The specific steps for inputting the updated node feature set into a two-layer fully connected neural network classifier for entity category prediction are as follows: The first fully connected layer of the two-layer fully connected neural network classifier performs linear transformation and non-linear activation on the representation vector of each node in the updated node feature set to generate an intermediate representation vector set. The second fully connected layer performs a linear mapping on the intermediate representation vector set and outputs the node classification probability distribution.

10. The network entity recognition method based on heterogeneous graph learning as described in claim 1, characterized in that: The specific steps for generating the network entity identification report are as follows: The predicted category label and confidence score of each node in the node classification probability distribution are extracted by the probabilistic category selection method. The information is then organized in a standardized format to statistically analyze the distribution of the number of nodes and the confidence score of different category labels, and a network entity recognition report is generated.

Images