A method for evaluating importance of nodes in an attribute-heterogeneous information network based on CMGI-GAT

Abstract

A kind of attribute heterogeneous information network node importance evaluation method based on CMGI-GAT. It relates to the field of computer service technology, in view of the problem that in the existing attribute heterogeneous information network node importance evaluation process, it is difficult to effectively handle the noise and redundant information in multi-source, multi-modal attribute characteristics, it is difficult to fully mine the high-order semantic relationship of network, leading to insufficient evaluation accuracy and robustness, the following scheme is proposed: differentially encode the category attribute, numerical attribute and text attribute, and combine cross-modal attribute interaction, multi-objective loss optimization, type-aware feature fusion, edge weight update, structure optimization, attribute embedding learning, structure embedding learning and cross-feature multi-scale fusion, finally obtain the node importance value through multi-path hierarchical prediction. The present application is suitable for key node identification and importance evaluation in social network analysis, recommendation system, knowledge graph and other attribute heterogeneous information network scenarios.

Description

Technical Field

[0001] This relates to the field of service computer technology and is a method for evaluating the importance of nodes in heterogeneous information networks based on CMGI-GAT. Background Technology

[0002] With the development of information technology and the continuous expansion of complex data scenarios, traditional homogeneous network models are increasingly unable to accurately describe the complex relational structures of real-world systems, characterized by "multiple types of nodes, multiple types of relationships, and nodes carrying multi-source attribute information." Against this backdrop, attribute-heterogeneous information networks have gradually become important modeling tools in fields such as service computing, recommender systems, knowledge graphs, and social network analysis. Existing research on the analysis of such networks typically focuses on network representation learning, node classification, link prediction, and node importance assessment. Among these, node importance assessment, as a fundamental problem for identifying key nodes, supporting resource allocation, risk warning, and decision optimization, has become a core component of complex network research. Relevant materials also clearly indicate that attribute-heterogeneous information networks have become an important modeling paradigm for characterizing complex real-world systems, and node importance assessment is a crucial method for identifying key nodes by mining node structural features, attribute features, and their interaction relationships.

[0003] Based on current research, methods for assessing node importance can be broadly categorized into two types. One type focuses on the network topology itself, using metrics such as node degree, betweenness, proximity centrality, path patterns, or subgraph statistics for evaluation. This type of method is relatively straightforward and suitable for networks with clear structural relationships. However, for heterogeneous information networks, relying solely on structural information often fails to reflect the true importance of nodes at the multi-source attribute semantic level. The other type of method introduces node attribute information, combining representation learning, graph neural networks, and attention mechanisms to jointly model node attributes and structural relationships. For example, it encodes category attributes, numerical attributes, and text attributes to participate in node representation learning, or aggregates neighborhood information through graph attention networks to improve the accuracy of node importance assessment.

[0004] However, existing technologies still have significant limitations when dealing with heterogeneous information networks. First, the node attributes in real-world networks are complex, often containing multiple modalities such as categories, values, and text. These modalities differ significantly in representation space, statistical characteristics, and semantic expressive capabilities. Existing methods typically lack differentiated encoding and targeted interaction mechanisms for different attribute categories, making it difficult to effectively remove noise and redundant information contained in the attribute data. Second, existing methods are insufficient for deep correlation mining between cross-modal attributes, often only achieving shallow splicing or simple fusion, failing to fully learn complementary relationships and higher-order semantics between attributes, thus affecting the effectiveness of node attribute representation. Third, in terms of structural modeling, existing methods mostly rely on direct propagation or aggregation based on original adjacency relationships, lacking the ability to complete missing attribute nodes, insufficiently characterizing the differentiated structural relationships between heterogeneous node types, and struggling to balance original structure preservation, degree distribution stability, and community structure consistency during graph optimization, thus limiting the overall information expressive capability of the network. Fourth, in the joint modeling of attribute features and structural features, existing methods are prone to introducing heterogeneous data bias, which distorts the fused node representation and further affects the accuracy and stability of node importance assessment results. These problems have been clearly summarized in relevant materials: on the one hand, existing methods cannot effectively handle attribute features from complex sources with severe noise interference, or fully explore the higher-order semantic relationships of the network; on the other hand, when modeling multi-source, multi-modal attribute features and the fusion of attribute and structural features, they are prone to introducing heterogeneous data bias, leading to inaccurate node feature representations and thus affecting the accuracy and stability of node importance assessment.

[0005] For example, in academic collaboration networks, social media networks, or knowledge-linked networks, a single node often simultaneously possesses identity category information, behavioral numerical information, and textual description information. If only a uniform approach is used for coarse encoding, or if only the original graph structure is used as the basis for propagation, problems such as insufficient semantic representation of attributes, inaccurate characterization of node relationships, and significant bias in the identification of key nodes can easily arise. Especially in scenarios with missing attributes, strong noise, diverse node types, and complex cross-type relationships, traditional methods based on a single structural centrality index or simple graph representation learning often struggle to simultaneously ensure attribute availability, structural expressiveness, and robustness of the final importance assessment results.

[0006] In summary, existing technologies suffer from several drawbacks: difficulty in effectively denoising and representing multi-source, multi-modal attribute features in heterogeneous information networks; difficulty in fully mining higher-order semantic relationships in the network; and the tendency to introduce heterogeneous data bias during the fusion modeling of attribute and structural features, resulting in insufficient accuracy and robustness in node importance assessment. Summary of the Invention

[0007] To address the shortcomings of existing technologies, such as difficulty in effectively denoising and representing multi-source, multi-modal attribute features in heterogeneous information networks, difficulty in fully mining higher-order semantic relationships in the network, and the tendency to introduce heterogeneous data bias during the fusion modeling of attribute and structural features, leading to insufficient accuracy and robustness in node importance assessment, the present invention provides the following technical solution: A method for evaluating the importance of nodes in attribute-heterogeneous information networks based on CMGI-GAT is proposed. This method evaluates node importance through multi-granularity optimization and learning of attribute features and structural features. The node importance evaluation method includes the following steps: Step 1: Perform differential encoding on multi-class attribute features with multiple semantic categories in the attribute heterogeneous information network; Step 2: Construct an initial adjacency matrix for nodes and edges in the heterogeneous information network to form structural features representing the network topology; Step 3: Perform cross-modal attribute interaction processing on the multi-class attribute features encoded in Step 1, reduce the dimension of the interacted attribute features by projection, and constrain the attribute feature expression from different dimensions through multi-objective loss optimization to obtain denoised and highly usable class attribute features. Step 4: Optimize the structural features obtained in Step 2. For nodes with missing attribute features, learn an average pattern based on their similar nodes. For each type of node, learn a type feature. Construct network structural features through multiple centrality indices. Update the edge weights of node pairs through a ternary gating mechanism. Optimize the network structure through edge weights to obtain a network topology with stronger information expression capabilities. Step 5: Enhance the expressive power of the multi-category attribute features obtained in Step 3 through cross-modal attribute interaction learning, realize the differentiated representation between different types of nodes through type interaction, retain the credible semantics in the original attribute features through residual compensation, and finally obtain the attribute embedding representing the node attributes through the multi-category attribute features. Step 6: The network topology obtained in Step 4 is used to obtain structural features through multiple centrality methods. Type-aware graph attention and residual modeling are used to model the heterogeneous relationships between nodes and alleviate information degradation. The structural features are recalibrated through SE-Block to suppress noise channels, and finally a highly discriminative and highly available structural embedding is obtained. Step 7: Perform cross-feature modality and multi-scale feature fusion on the attribute embeddings obtained in Step 5 and the structural embeddings obtained in Step 6 to obtain the optimal embedding representation of the node features; Step 8: The optimal node embedding representation obtained in Step 7 is used to obtain three different node importance evaluation results through three prediction paths of robust hierarchical importance prediction. These results are then dynamically weighted to obtain the final node importance value.

[0008] Furthermore, in a preferred embodiment, step 1 specifically involves dividing the initial node attribute information into three categories: category, numerical, and text. The discrete category information is mapped to a continuous space for encoding through an embedding layer. The numerical information is standardized and then encoded. The text information is encoded using BERT to maintain their respective semantic characteristics.

[0009] Furthermore, in a preferred embodiment, step 3 further includes the following steps: Multimodal attributes are interactively enhanced by learning the correlation between different modal attributes through an attention mechanism to obtain enhanced attribute representations. Each enhanced modal attribute is then subjected to projection transformation to reduce feature dimensionality and unify the representation space. Based on the attribute interaction results and dimensionality reduction results, optimization constraints are constructed to jointly optimize the attribute representations. The optimization constraints include at least the reconstruction constraints for preserving the original semantic information, the contrast constraints for enhancing the distinguishability of similar nodes, the specificity constraints for preserving the independent expressive ability of different attribute modalities, and the preservation constraints for controlling the degree of deviation of attribute representations before and after optimization. Different types of attributes are constrained using a difference metric method adapted to their data features.

[0010] Furthermore, in a preferred embodiment, step 4 further includes the following steps: Based on node type, node features are completed and fused to learn common features related to each type of node and fill in the corresponding type features for nodes with missing attributes, so as to form a node representation containing type semantics. Based on the original graph structure, learnable parameters and node attribute features, the edge association strength between nodes is calculated and the edge weights are updated accordingly. Based on the updated edge weights, structural optimization constraints are constructed. The structural optimization constraints include at least structural preservation constraints to maintain the original graph connection relationship, degree distribution constraints to limit the change of node connection distribution, and community preservation constraints to maintain the original community organization relationship. The connection relationship in the graph is adjusted based on the edge weights to obtain the optimized network structure.

[0011] Furthermore, in a preferred embodiment, step 5 specifically involves: realizing the interaction between multi-category attribute features based on multi-head self-attention; using adaptive gating to learn the attribute fusion strategy of each category at the node level to obtain the node attribute embedding; then learning the features of each type of node through two layers of nonlinear transformation and adaptively fusing them with the node attribute embedding; finally fusing the original attributes through residual processing to obtain the attribute embedding.

[0012] Furthermore, in a preferred embodiment, step 6 specifically involves constructing structural features on the optimized graph based on degree, path pattern, subgraph statistics, etc., and normalizing them. Then, an attention module for the type of node to be focused on is added to the graph attention network to aggregate multi-level neighborhoods. At the same time, the feature channels are recalibrated based on SE-Block. Finally, the original structural features are fused through residual processing to obtain the structural embedding.

[0013] Furthermore, in a preferred embodiment, step 7 specifically involves performing internal semantic enhancement on the attribute and structure embeddings based on multi-head self-attention, then constructing three nonlinear layers with different fusion scales to perform multi-scale fusion of the attribute and structure embeddings, and finally obtaining the optimal embedding representation of the node.

[0014] Furthermore, in a preferred embodiment, step 8 specifically involves constructing three nonlinear importance prediction paths—single-layer, double-layer, and triple-layer—based on MLP to obtain three importance evaluation results, and then weighting and fusing the three results to obtain the final node importance value.

[0015] A node importance assessment device for attribute-heterogeneous information networks based on CMGI-GAT evaluates node importance through multi-granularity optimization and learning of attribute features and structural features. The node importance assessment device includes the following modules: Module 1: Differentiated encoding of multi-class attribute features with multiple semantic categories in heterogeneous attribute information networks; Module 2: Construct an initial adjacency matrix for nodes and edges in a heterogeneous information network to form structural features representing the network topology; Module 3: Perform cross-modal attribute interaction processing on the multi-class attribute features encoded in Step 1, reduce the dimension of the interacted attribute features by projection, and constrain the attribute feature expression from different dimensions through multi-objective loss optimization to obtain denoised and highly available class attribute features. Module 4: Optimize the structural features obtained in step 2. Learn an average pattern for nodes with missing attribute features based on their similar nodes. Learn a type feature for each type of node. Construct network structural features through multiple centrality indices. Update the edge weights of node pairs through a ternary gating mechanism. Optimize the network structure through edge weights to obtain a network topology with stronger information expression capabilities. Module 5: Enhance the expressive power of the multi-category attribute features obtained in step 3 through cross-modal attribute interaction learning, realize the differentiated representation between different types of nodes through type interaction, retain the credible semantics in the original attribute features through residual compensation, and finally obtain the attribute embedding representing the node attributes through the multi-category attribute features; Module 6: The network topology obtained in step 4 is used to obtain structural features through various centrality methods. Type-aware graph attention and residual modeling are used to model the heterogeneous relationships between nodes and alleviate information degradation. The structural features are recalibrated through SE-Block to suppress noise channels, and finally a highly discriminative and highly available structural embedding is obtained. Module 7: Perform cross-feature modality and multi-scale feature fusion on the attribute embeddings obtained in step 5 and the structural embeddings obtained in step 6 to obtain the optimal embedding representation of the node features; Module 8: The optimal node embedding representation obtained in step 7 is used to obtain three different node importance evaluation results through three prediction paths of robust hierarchical importance prediction. These results are then dynamically weighted to obtain the final node importance value.

[0016] A computer, including a processor and a storage medium, executes the method when the processor reads a computer program stored in the storage medium.

[0017] Compared with the prior art, the advantages of the technical solution provided by the present invention are as follows: The initial node attribute information is divided into three categories: category, numerical, and text. The discrete category information is mapped to a continuous space through an embedding layer, the numerical information is standardized and encoded, and the text information is encoded through BERT. This allows for differentiated processing based on the expression methods and statistical features of different attribute data. It avoids the problems of semantic compression distortion, numerical scale confusion, and insufficient utilization of text semantics caused by using a unified encoding method for multiple heterogeneous attributes in existing technologies. This improves the usability and distinguishability of various attribute information from the source, and provides higher quality input features for subsequent attribute interaction, fusion, and importance assessment.

[0018] Constructing the nodes, edges, and edge weights of the initial network into an adjacency matrix to represent the current network allows for a unified expression of the fundamental connectivity, interaction strength, and topological distribution in heterogeneous information networks. This provides a clear structural framework for subsequent structural optimization, edge weight updates, structural embedding learning, and node importance propagation. Compared to research methods that only model node attributes while ignoring the original connectivity, this feature enables node importance assessment to not only rely on attribute semantics but also be built upon the real network topology, thereby enhancing the ability of the assessment results to reflect the overall organizational structure of the network.

[0019] By assigning attention heads to each modal attribute through multi-head attention, cross-modal attribute interactions are performed on each information sequence. Combined with learning dimensionality reduction projection matrices for each modal attribute and four types of constraints—semantic reconstruction loss, contrastive loss, specificity loss, and preservation loss—this approach fully explores complementary relationships between different attribute modalities while suppressing noise and redundant information. It avoids the modal conflicts and information pollution problems commonly found in existing studies caused by simple splicing and coarse fusion. Specifically, multi-head attention delivers more comprehensive modal association learning results, differential dimensionality reduction compresses redundancy and improves representation compactness, and multi-objective losses constrain attribute expressions from the perspectives of semantic preservation, category differentiation, feature uniqueness, and information fidelity. Ultimately, the resulting categorical attribute features possess stronger discriminative power, higher usability, and better robustness.

[0020] For nodes lacking partial attribute features, an average pattern is learned based on nodes of the same type. A type feature is also learned for each type of node. These, combined with multiple centrality indicators, constitute the network structural features. This effectively alleviates the common problems in attribute-heterogeneous information networks, such as missing attributes, significant type differences, and insufficient representational power of single structural indicators. The benefits of this approach are twofold: firstly, by supplementing missing attributes with type-related common features, the information representation ability of incomplete nodes is improved; secondly, by introducing multiple centrality information rather than relying on a single degree value or a single path indicator, the structural features can more comprehensively reflect the different roles of nodes in local connectivity, global propagation, and community organization, thus providing a richer and more stable structural basis for subsequent edge weight optimization and structural embedding learning.

[0021] Based on the original graph structure representation, learnable weight matrix, and attribute features, a ternary gating mechanism is used to update the edge weights of node pairs. This is combined with structure preservation loss, degree distribution loss, and community preservation loss to constrain the edge weights. Then, edges are added and deleted based on these weights. This approach overcomes the limitations of existing methods that directly use the original graph structure and are difficult to adapt to task requirements. The benefits of this part are: the gating mechanism dynamically determines the strength of relationships between nodes, ensuring that edge weight updates consider both original structural information and attribute similarity, thus improving the rationality of edge connections; the structure preservation loss prevents excessive deviation from the original topology after optimization; the degree distribution loss maintains the overall network connectivity; and the community preservation loss preserves the original clustering and community organization within the network. Therefore, without disrupting the basic organizational rules of the network, the information representation capability of the network topology is improved, making the optimized network more suitable for node importance assessment tasks.

[0022] This approach leverages multi-head self-attention to facilitate interaction between multi-category attribute features. It utilizes adaptive gating to learn node-level attribute fusion strategies, then employs two layers of nonlinear transformations to learn features for each node type and adaptively fuse them with node attribute embeddings. Finally, residual processing is used to fuse the original attributes, further enhancing attribute representation capabilities at both the node and type levels. The benefits include: multi-head self-attention strengthens deep semantic connections between different attribute categories; adaptive gating prevents different nodes from being forced to use the same fusion ratio; two layers of nonlinear transformations enable different types of nodes to form more differentiated type semantic expressions; and residual processing prevents excessive weakening of original effective attributes during complex fusion processes. Compared to existing methods that only perform uniform fusion or single-layer linear transformations, this approach is more conducive to preserving reliable semantics and improving the task adaptability and representational accuracy of attribute embeddings.

[0023] On the optimized graph, structural features are constructed and normalized based on degree, path pattern, and subgraph statistics. Then, an attention module focusing on node types is added to the graph attention network to aggregate multi-level neighborhoods. Simultaneously, SE-Block is used to recalibrate the structural feature channels, and residual processing is used to fuse the original structural features, significantly enhancing the discriminative power of the structural embedding. The benefits of this part are that the combination of multiple centralities and structural statistics allows structural representation to move beyond a single topological perspective; type-aware graph attention enables the differentiation of heterogeneous relationships between different types of nodes; multi-level neighborhood aggregation improves the ability to mine higher-order structural semantics; SE-Block highlights key structural channels and suppresses noisy channels; and residual fusion mitigates the information degradation problem caused by deep propagation. Therefore, a more discriminative and usable structural embedding can be obtained.

[0024] This approach leverages multi-head self-attention to enhance the internal semantics of attribute embeddings and structural embeddings separately. Furthermore, it constructs three nonlinear layers with different fusion scales to perform multi-scale fusion of these two types of embeddings. This effectively addresses the problems of simple concatenation and insufficient deep coupling between attribute and structural information in existing technologies. The benefits of this approach lie in first enhancing the internal semantics of each type of embedding, which helps release the effective semantics within each modality. Then, through a nonlinear fusion mechanism at different scales, it gradually establishes a collaborative relationship between attributes and structure from shallow to deep associations. This results in a more comprehensive and stable optimal node embedding representation, ensuring that the final node features retain both the semantic advantages of attributes and the structural topological advantages.

[0025] Based on MLP, three nonlinear importance prediction paths—single-layer, two-layer, and three-layer—are constructed to obtain three different importance evaluation results. These three results are then dynamically weighted and fused, avoiding the problem of a single prediction path being overly sensitive to a particular feature pattern. The benefit of this approach is that different levels of prediction paths adapt to feature mapping relationships of varying complexity. Single-layer paths are beneficial for preserving direct discriminative information, while deeper paths are better at characterizing more complex nonlinear relationships. Dynamic weighted fusion can comprehensively form the final importance value based on the effectiveness of the prediction results, thereby improving the accuracy, stability, and robustness of the node importance evaluation results.

[0026] The aforementioned features are not isolated from each other, but rather form a complete chain of action, from multi-source heterogeneous attribute preprocessing, attribute denoising and enhancement, network topology optimization, attribute embedding learning, structural embedding learning, cross-modal fusion to hierarchical prediction output. Compared to existing research that typically only enhances one side of the information or only performs shallow joint modeling of attributes and structure, this scheme improves the effectiveness of attribute features, the expressive power of structural features, and the quality of the final fused representation through multi-granularity collaborative modeling. Therefore, it can more effectively address the problems of insufficient accuracy, weak stability, and low robustness in node importance assessment in attribute heterogeneous information networks. Attached Figure Description

[0027] Figure 1 This is a comprehensive architecture diagram.

[0028] Figure 2 This is a flowchart for attribute feature optimization.

[0029] Figure 3 This is a flowchart for structural feature optimization.

[0030] Figure 4 This is a flowchart of the attribute embedding representation learning process.

[0031] Figure 5 This is a flowchart of the structure embedding representation learning process.

[0032] Figure 6 This is a flowchart for importance assessment. Detailed Implementation

[0033] To make the advantages and benefits of the technical solution provided by the present invention clearer, the technical solution provided by the present invention will now be described in further detail with reference to the accompanying drawings, specifically: Implementation Method 1: This implementation method provides a node importance assessment method for attribute-heterogeneous information networks based on CMGI-GAT. It assesses node importance through multi-granularity optimization and learning of attribute features and structural features. The node importance assessment method includes the following steps: Step 1: Perform differential encoding on multi-class attribute features with multiple semantic categories in the attribute heterogeneous information network; Step 2: Construct an initial adjacency matrix for nodes and edges in the heterogeneous information network to form structural features representing the network topology; Step 3: Perform cross-modal attribute interaction processing on the multi-class attribute features encoded in Step 1, reduce the dimension of the interacted attribute features by projection, and constrain the attribute feature expression from different dimensions through multi-objective loss optimization to obtain denoised and highly usable class attribute features. Step 4: Optimize the structural features obtained in Step 2. For nodes with missing attribute features, learn an average pattern based on their similar nodes. For each type of node, learn a type feature. Construct network structural features through multiple centrality indices. Update the edge weights of node pairs through a ternary gating mechanism. Optimize the network structure through edge weights to obtain a network topology with stronger information expression capabilities. Step 5: Enhance the expressive power of the multi-category attribute features obtained in Step 3 through cross-modal attribute interaction learning, realize the differentiated representation between different types of nodes through type interaction, retain the credible semantics in the original attribute features through residual compensation, and finally obtain the attribute embedding representing the node attributes through the multi-category attribute features. Step 6: The network topology obtained in Step 4 is used to obtain structural features through multiple centrality methods. Type-aware graph attention and residual modeling are used to model the heterogeneous relationships between nodes and alleviate information degradation. The structural features are recalibrated through SE-Block to suppress noise channels, and finally a highly discriminative and highly available structural embedding is obtained. Step 7: Perform cross-feature modality and multi-scale feature fusion on the attribute embeddings obtained in Step 5 and the structural embeddings obtained in Step 6 to obtain the optimal embedding representation of the node features; Step 8: The optimal node embedding representation obtained in Step 7 is used to obtain three different node importance evaluation results through three prediction paths of robust hierarchical importance prediction. These results are then dynamically weighted to obtain the final node importance value.

[0034] Furthermore, in a preferred embodiment, step 1 specifically involves dividing the initial node attribute information into three categories: category, numerical, and text. The discrete category information is mapped to a continuous space for encoding through an embedding layer. The numerical information is standardized and then encoded. The text information is encoded using BERT to maintain their respective semantic characteristics.

[0035] Furthermore, in a preferred embodiment, step 3 further includes the following steps: Multimodal attributes are interactively enhanced by learning the correlation between different modal attributes through an attention mechanism to obtain enhanced attribute representations. Each enhanced modal attribute is then subjected to projection transformation to reduce feature dimensionality and unify the representation space. Based on the attribute interaction results and dimensionality reduction results, optimization constraints are constructed to jointly optimize the attribute representations. The optimization constraints include at least the reconstruction constraints for preserving the original semantic information, the contrast constraints for enhancing the distinguishability of similar nodes, the specificity constraints for preserving the independent expressive ability of different attribute modalities, and the preservation constraints for controlling the degree of deviation of attribute representations before and after optimization. Different types of attributes are constrained using a difference metric method adapted to their data features.

[0036] Furthermore, in a preferred embodiment, step 4 further includes the following steps: Based on node type, node features are completed and fused to learn common features related to each type of node and fill in the corresponding type features for nodes with missing attributes, so as to form a node representation containing type semantics. Based on the original graph structure, learnable parameters and node attribute features, the edge association strength between nodes is calculated and the edge weights are updated accordingly. Based on the updated edge weights, structural optimization constraints are constructed. The structural optimization constraints include at least structural preservation constraints to maintain the original graph connection relationship, degree distribution constraints to limit the change of node connection distribution, and community preservation constraints to maintain the original community organization relationship. The connection relationship in the graph is adjusted based on the edge weights to obtain the optimized network structure.

[0037] Furthermore, in a preferred embodiment, step 5 specifically involves: realizing the interaction between multi-category attribute features based on multi-head self-attention; using adaptive gating to learn the attribute fusion strategy of each category at the node level to obtain the node attribute embedding; then learning the features of each type of node through two layers of nonlinear transformation and adaptively fusing them with the node attribute embedding; finally fusing the original attributes through residual processing to obtain the attribute embedding.

[0038] Furthermore, in a preferred embodiment, step 6 specifically involves constructing structural features on the optimized graph based on degree, path pattern, subgraph statistics, etc., and normalizing them. Then, an attention module for the type of node to be focused on is added to the graph attention network to aggregate multi-level neighborhoods. At the same time, the feature channels are recalibrated based on SE-Block. Finally, the original structural features are fused through residual processing to obtain the structural embedding.

[0039] Furthermore, in a preferred embodiment, step 7 specifically involves performing internal semantic enhancement on the attribute and structure embeddings based on multi-head self-attention, then constructing three nonlinear layers with different fusion scales to perform multi-scale fusion of the attribute and structure embeddings, and finally obtaining the optimal embedding representation of the node.

[0040] Furthermore, in a preferred embodiment, step 8 specifically involves constructing three nonlinear importance prediction paths—single-layer, double-layer, and triple-layer—based on MLP to obtain three importance evaluation results, and then weighting and fusing the three results to obtain the final node importance value.

[0041] A node importance assessment device for attribute-heterogeneous information networks based on CMGI-GAT evaluates node importance through multi-granularity optimization and learning of attribute features and structural features. The node importance assessment device includes the following modules: Module 1: Differentiated encoding of multi-class attribute features with multiple semantic categories in heterogeneous attribute information networks; Module 2: Construct an initial adjacency matrix for nodes and edges in a heterogeneous information network to form structural features representing the network topology; Module 3: Perform cross-modal attribute interaction processing on the multi-class attribute features encoded in Step 1, reduce the dimension of the interacted attribute features by projection, and constrain the attribute feature expression from different dimensions through multi-objective loss optimization to obtain denoised and highly available class attribute features. Module 4: Optimize the structural features obtained in step 2. Learn an average pattern for nodes with missing attribute features based on their similar nodes. Learn a type feature for each type of node. Construct network structural features through multiple centrality indices. Update the edge weights of node pairs through a ternary gating mechanism. Optimize the network structure through edge weights to obtain a network topology with stronger information expression capabilities. Module 5: Enhance the expressive power of the multi-category attribute features obtained in step 3 through cross-modal attribute interaction learning, realize the differentiated representation between different types of nodes through type interaction, retain the credible semantics in the original attribute features through residual compensation, and finally obtain the attribute embedding representing the node attributes through the multi-category attribute features; Module 6: The network topology obtained in step 4 is used to obtain structural features through various centrality methods. Type-aware graph attention and residual modeling are used to model the heterogeneous relationships between nodes and alleviate information degradation. The structural features are recalibrated through SE-Block to suppress noise channels, and finally a highly discriminative and highly available structural embedding is obtained. Module 7: Perform cross-feature modality and multi-scale feature fusion on the attribute embeddings obtained in step 5 and the structural embeddings obtained in step 6 to obtain the optimal embedding representation of the node features; Module 8: The optimal node embedding representation obtained in step 7 is used to obtain three different node importance evaluation results through three prediction paths of robust hierarchical importance prediction. These results are then dynamically weighted to obtain the final node importance value.

[0042] A computer, including a processor and a storage medium, executes the method when the processor reads a computer program stored in the storage medium.

[0043] Implementation Method Two: This implementation method is a further detailed description of the technical solution provided in Implementation Method One, specifically: The first step is to organize the original heterogeneous information network to form unified network input data.

[0044] A heterogeneous attribute information network to be processed is obtained, which includes at least a node set, an edge set, and a node attribute set. The nodes in the node set can be classified into multiple node types, the edges in the edge set can represent the connection relationships between nodes of different types or between nodes of the same type, and the node attribute set can include category attributes, numerical attributes, and text attributes. To ensure consistency in subsequent calculations, the original data is first standardized. A unique identifier is assigned to each node to record its type; the starting node, ending node, and corresponding edge weight are recorded for each edge; an attribute record is established for each node, and all attributes corresponding to that node are uniformly organized into the same attribute container. For missing, abnormal, or inconsistently formatted data, cleaning and correction are performed first. For tags with few missing items, the missing state can be retained for use by the subsequent type-aware feature completion module; for obviously erroneous values, they can be corrected or removed according to preset rules; for invalid characters, duplicate tags, or garbled content in text attributes, they can be filtered first. After processing, we obtain a node table, an edge table, and an attribute table. The node table serves as the basis for network type division and node indexing, the edge table serves as the basis for adjacency relationship construction, and the attribute table serves as the input for subsequent attribute encoding.

[0045] The second step is to perform differential encoding on the categorical attributes, numerical attributes, and text attributes to obtain the initial multi-category attribute features.

[0046] The attribute table obtained in the first step is split according to the attribute data type, forming subsets of categorical attributes, numerical attributes, and text attributes. For categorical attributes, a categorical dictionary is first established, mapping discrete categorical values ​​to categorical indices. Then, an embedding layer transforms the categorical indices into continuous vectors, enabling categorical information to participate in subsequent representation learning in a trainable manner. If a node corresponds to multiple categorical labels, each label is embedded separately and then merged according to a preset aggregation rule. This aggregation rule can be summation, averaging, or weighted summation to obtain the categorical attribute representation for that node. For numerical attributes, values ​​with different dimensions and ranges are first standardized to ensure all values ​​fall into a uniform scale, preventing any single high-dimensional attribute from unreasonably dominating the overall representation. The standardized values ​​are then input into a numerical encoding layer for mapping to obtain a numerical attribute representation suitable for subsequent interactive learning. For text attributes, word segmentation, truncation, padding, and text length alignment are first performed, then the text is input into a pre-trained language model for encoding, preferably using the BERT model to extract contextual semantic information from the text attributes. Finally, categorical attribute representation, numerical attribute representation, and text attribute representation are obtained for each node. These three types of attributes together constitute the initial multi-category attribute features of the node and serve as input for the subsequent attribute feature optimization module.

[0047] The third step is to construct an adjacency matrix based on nodes, edges, and edge weights to form the initial network structure features.

[0048] Using the edge table and node index relationship obtained in the first step, the node connection relationships in the original network are mapped to an adjacency matrix. If the edge is directed, the connection value at the corresponding position in the adjacency matrix is ​​recorded according to the direction; if the edge is undirected, the connection value is recorded at the symmetrical position simultaneously. If edge weights exist in the original network, the edge weights are filled into the corresponding positions in the adjacency matrix; if there are no edge weights, a unified initial value is used to represent the existence of a connection relationship. To ensure the stability of subsequent graph structure optimization and graph neural network propagation, the adjacency matrix can be initially normalized to eliminate numerical fluctuations caused by excessive differences in node degree distribution. After the adjacency matrix is ​​constructed, the basic topological connection relationships of each node in the network are uniformly represented, forming the initial structural feature input. This initial structural feature will serve as the basis for subsequent type-aware fusion, edge weight update, and structural embedding learning.

[0049] The fourth step involves performing cross-modal interaction, differential dimensionality reduction, and multi-objective optimization on the multi-category attribute features to obtain denoised, highly usable category attribute features.

[0050] The categorical, numerical, and textual attribute representations obtained in the second step are input into the attribute feature optimization module. First, a multi-head attention mechanism is used to establish interaction relationships between attributes of different modalities. At least one attention head is assigned to each attribute class, enabling categorical, numerical, and textual attributes to perceive each other and learn the correlation, complementarity, and consistency information between modalities from different perspectives. Through this interaction process, each attribute class not only retains its original modality information but also introduces effective supplements from other modalities, resulting in a cross-modal enhanced attribute representation. Subsequently, dimensionality reduction projection layers are constructed for different modalities to differentially compress the interacted attribute representations. Since the original dimensions of categorical, numerical, and textual attributes are usually different, direct concatenation can easily lead to dimensionality redundancy and modality imbalance. Therefore, corresponding projection matrices are learned separately to compress the three attributes into a feature space suitable for fusion. To ensure that key information is not lost during attribute enhancement and dimensionality reduction, while suppressing noise and redundancy, multi-objective loss constraints need to be applied to the optimization process.

[0051] The multi-objective loss includes at least semantic reconstruction loss, contrastive loss, specificity loss, and preservation loss. Semantic reconstruction loss constrains the dimensionality-reduced features to still recover or correspond to the original semantics. Specifically, category attributes are constrained based on distribution differences to ensure the probability distribution of the reduced category is as consistent as possible with the original category representation; numerical attributes are constrained with a strong tolerance for abnormal fluctuations to ensure the overall trend of the reduced numerical representation is consistent with the original numerical value; and text attributes are constrained based on vector direction consistency to ensure the reduced text semantics are close to the original text semantics. Contrastive loss narrows the attribute representation distance between semantically similar nodes and widens the attribute representation distance between nodes with significant semantic differences. Specifically, it can be understood as first constructing positive and negative sample pairs based on node attribute similarity, then making similar samples closer in the embedding space and dissimilar samples more separated. Specificity loss ensures that different modal attributes retain their unique information expression capabilities while sharing interactions, preventing excessive homogenization between different attributes after multimodal interaction. The preservation loss is used to constrain the semantics of attributes before and after the interaction from uncontrolled shifts. For text attributes, cosine distance is used to measure the consistency of semantic direction before and after the interaction; for categorical attributes, a distribution-based difference-based approach is used to measure the degree of preservation; and for numerical attributes, a smoothing error approach is used to constrain their variation. Through this multi-objective optimization, highly usable categorical attribute features with lower noise, less redundancy, and more complete semantic preservation are obtained and used as input for subsequent attribute embedding learning.

[0052] The fifth step involves performing type-aware completion, edge weight update, and graph structure optimization on the initial structural features to obtain a network topology with stronger information representation capabilities.

[0053] The adjacency matrix obtained in step three and the denoised attribute features obtained in step four are jointly input into the structural feature optimization module. First, for nodes in the network with missing attributes, a type-aware feature completion mechanism is established. This mechanism learns common attribute patterns corresponding to each node type. This can be understood as extracting average semantics and stable features from nodes of the same type and using them as the default pattern for that type of node. When a node has some missing attributes, the common features corresponding to its node type are used as supplementary inputs, together with the node's existing attributes, to form a more complete node feature. Simultaneously, a type feature vector is learned for each node type to represent the overall attribute semantics and structural role of that type of node in the network. After completion, the node features no longer rely solely on the original attributes but also include type-level prior information.

[0054] Based on this, structural features are constructed. These features include not only the original adjacency relationships but also node topological information extracted from multiple centrality indicators. These centrality indicators may include node degree information, neighbor distribution information, path correlation statistics, subgraph statistics, and other indicators that characterize the role of nodes in the community or propagation structure. After unifying and normalizing these indicators, a structural feature representation reflecting the topological role of nodes is obtained.

[0055] Subsequently, based on the original graph structure representation, the learnable weight matrix, and node attribute features, a ternary gating mechanism is used to calculate the edge weights between node pairs. This ternary gating mechanism can be understood as simultaneously considering the original connection state, the strength of structural effects, and the degree of attribute matching to comprehensively judge whether each pair of nodes should strengthen, weaken, or maintain its connection. For node pairs with existing edges, if the gating output indicates that the two nodes have high consistency in task-related attributes and structural roles, the edge weight is increased accordingly; if the gating output indicates low correlation, the edge weight is decreased. For node pairs without existing edges, if the gating output indicates a potentially effective connection, a connection can be added in subsequent graph optimization. To prevent excessive modification of the graph structure, graph optimization constraints need to be applied to the edge weight update process.

[0056] The graph optimization constraints include at least structure preservation loss, degree distribution loss, and community preservation loss. Structure preservation loss limits the difference between the optimized and original network topologies to a reasonable range, ensuring that the original connection patterns are not completely destroyed. Degree distribution loss constrains the changes in the connection scale of each node before and after optimization, preventing abnormal expansion or sudden drop in the degree of local nodes. Community preservation loss ensures that the community organizational structure in the original network remains largely intact after optimization, avoiding meaningless destruction of community boundaries due to edge reconstruction. After completing edge weight updates and constraint optimization, edges with lower weights are deleted according to a preset graph optimization ratio, and potential edges with higher weights that were not originally present are added, resulting in the optimized network topology. This optimized network topology better reflects the effective connection relationships relevant to the task compared to the original graph and serves as input for subsequent structural embedding learning.

[0057] The sixth step involves joint modeling of node-level, type-level, and residual compensation based on multi-category attribute features to obtain attribute embeddings.

[0058] The highly available category attribute features obtained in step four are input into the attribute embedding learning module. First, based on a multi-head self-attention mechanism, the internal relationships between category attributes, numerical attributes, and text attributes are further learned, enabling different attribute modalities to interact more specifically around the importance assessment task of the current node. Compared to the aforementioned cross-modal attribute interaction, this approach emphasizes centering on the node itself, reorganizing each category attribute into a unified expression suitable for node representation learning. After completing the multi-head self-attention interaction, for each node, an adaptive gating mechanism is used to learn the fusion weights of its different category attributes. That is, for different nodes, the same category attribute fusion ratio is not enforced; instead, the contribution of category attributes, numerical attributes, and text attributes to the current node's attribute embedding is dynamically determined based on the node's own feature state. After this node-level fusion, the initial representation of the node attribute embedding is obtained.

[0059] Building upon node-level fusion, type-level feature learning is further introduced. Specifically, a two-layer nonlinear transformation structure is set for each type of node, enabling the system to learn the common attribute semantics from samples of that type of node and generate a type-level representation corresponding to that type. Then, the type-level representation is fused with the node-level attribute embedding in an adaptive manner. This fusion ensures that the attribute representation of each node reflects both its individual attribute features and the common patterns of its node type. Finally, to avoid the suppression or loss of original effective attribute information during multiple rounds of interaction and fusion, the original attribute features are introduced into the final attribute representation through a residual path. After the above processing, an attribute embedding that takes into account individual semantics, type semantics, and original reliable information is obtained and serves as one of the inputs for subsequent cross-feature fusion.

[0060] Step 7: Learn structural embeddings based on the optimized network topology and multi-class structural statistics.

[0061] The optimized network topology obtained in step five is then embedded into the learning module. First, various structural statistical features are extracted from the optimized graph. These features include at least node degree, the number of neighbors at several orders, path pattern frequency, local subgraph statistics, and other centrality indicators that characterize the propagation role and connectivity of nodes in heterogeneous networks. After normalizing the extracted structural statistical features, an initial node structure representation is formed. This initial structure representation is then input into a graph attention network for multi-order neighborhood aggregation. To differentiate the heterogeneous relationships between different types of nodes, a type-aware attention module is added to the graph attention network. When calculating the attention coefficients between node pairs, this module considers not only the similarity of node features but also the combination of node types, allowing for differentiated weighting of the influence of different types of node pairs in message propagation. In this way, the interaction mechanisms between different types of nodes in attribute-heterogeneous information networks can be more accurately characterized.

[0062] In the multi-layer graph attention propagation process, residual connections are used to preserve the structural information of previous layers, preventing the structural features from becoming overly smooth or important patterns from degrading as the propagation depth increases. After completing multi-order neighborhood aggregation, SE-Block is used to recalibrate each channel of the structural embedding. Specifically, the global responses of each channel are first summarized, then the importance of different channels is calculated, and finally, key channels are given greater weight while noisy channels are suppressed. This process further improves the sensitivity of the structural embedding to key topological patterns. Finally, the original structural features are reintroduced into the structural embedding using a residual approach, so that the final structural embedding retains both basic topological information and high-order structural semantics after deep aggregation. After the above processing, a structural embedding with strong discriminative power and high usability is obtained, which serves as another input for subsequent cross-feature fusion.

[0063] The eighth step involves internal enhancement and multi-scale fusion of attribute embedding and structure embedding to obtain the optimal embedding representation of the node.

[0064] The attribute embeddings obtained in step 6 and the structural embeddings obtained in step 7 are jointly input into the cross-feature fusion module. First, internal semantic enhancement is performed on both the attribute embeddings and the structural embeddings. For attribute embeddings, multi-head self-attention is used to further strengthen the collaborative relationships between different attribute dimensions, making the attribute representation more focused on the semantic parts related to node importance. For structural embeddings, the same mechanism is used to strengthen the cooperative relationships between different structural dimensions, making the structural representation more prominent in terms of patterns related to propagation, connection, and bridging. After completing the internal enhancement, enhanced attribute representations and enhanced structural representations are obtained, respectively.

[0065] Subsequently, three nonlinear mapping layers with different fusion scales are set up to perform multi-scale fusion of the enhanced attribute representation and the enhanced structural representation. The shallower-scale fusion layer preserves the direct coupling relationship between attributes and structure, the medium-scale fusion layer mines the complementary relationships between them in local combinations, and the deeper-scale fusion layer extracts more complex high-order cooperative semantics. The fusion results obtained from the three scales are integrated to form a unified representation of the node. This unified representation includes both the node's own multimodal attribute semantics and the multi-layer structural semantics of the network topology, thus constituting the optimal embedding representation of the node. This optimal embedding representation of the node serves as the input to the subsequent importance prediction module.

[0066] Step 9: Predict the importance value of output nodes based on multi-path hierarchical prediction.

[0067] The optimal embedding representation of the nodes obtained in step eight is input into the importance prediction module. This importance prediction module includes three parallel prediction paths: a single-layer nonlinear prediction path, a two-layer nonlinear prediction path, and a three-layer nonlinear prediction path. The single-layer path directly extracts significant discriminative information from the node representation; the two-layer path models moderately complex nonlinear relationships; and the three-layer path characterizes deeper and more complex importance mapping relationships. Each of the three paths outputs a node importance evaluation result. To ensure the final result takes into account the expressive advantages of different paths, the outputs of the three prediction paths are dynamically weighted and fused. Dynamic weighting can be adjusted based on the stability of each path's current output, its discriminative ability, or the weight parameters learned during training, so that the contribution of different paths to the final result is not fixed but adaptively adjusted according to the sample state and task requirements. After dynamic weighting, the final importance value of each node is obtained. This importance value can be used for node ranking, key node identification, key object screening, or subsequent decision analysis.

[0068] In one embodiment, the training and inference of the method can be completed within the same model framework. During the training phase, heterogeneous attribute information network samples with known importance labels or constructible supervision signals are input into the above process, jointly optimizing the parameters in the attribute feature optimization module, structural feature optimization module, attribute embedding learning module, structural embedding learning module, cross-feature fusion module, and importance prediction module. During the inference phase, only the new heterogeneous attribute information network needs to be input into the trained model to output the importance values ​​of each node. In another embodiment, if some nodes in the target network lack a certain type of attribute, only the missing label needs to be retained in the first step, and the default mode needs to be supplemented in the fifth step through a type-aware feature completion mechanism, without prior external manual filling of the missing attributes. In another embodiment, if the text attributes in the network are long, the text content can be truncated to a preset length before encoding; if there are too many types of category attributes, the impact of category sparsity can be reduced through low-dimensional embedding and regularization training; if the network size is large, batch sampling, neighborhood sampling, or subgraph training methods can be used to perform the above-mentioned calculations.

[0069] Through the above process, the multi-source heterogeneous attributes in the original network are first differentially encoded and denoised. The original topology is updated with edge weights and optimized with the help of attribute and type information. Then, attribute embeddings and structural embeddings are learned respectively. The optimal embedding representation of the node is obtained through internal enhancement and multi-scale fusion. Finally, a stable node importance value is obtained through multi-path hierarchical prediction, thereby achieving accurate and robust evaluation of the importance of nodes in attribute heterogeneous information networks.

[0070] Implementation Method 3, in conjunction with Appendix Figure 1-6 This embodiment describes the technical solution provided above in further detail through specific examples. Specifically: Example 1 The main approach is to construct CMGI-GAT to evaluate the importance of nodes in heterogeneous attribute information networks. An attribute and structural feature optimization model is adopted to encode different types of attribute features. Noise and redundant attribute information are reduced through feature interaction and multi-objective optimization. High-order structural semantics are mined through ternary gating to enhance the network's information expression ability. The optimal embedding representation of attributes and structures is learned at multiple granularities at the node, type, and feature levels, and multi-scale fusion is performed to obtain the optimal node representation. Through multi-level importance evaluation, an accurate and robust characterization of node importance is achieved.

[0071] A method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT, the method comprising the following steps: Step 1: Perform differential encoding on multi-class attribute features with multiple semantic categories in the attribute heterogeneous information network; Step 2: Construct an initial adjacency matrix for nodes and edges in the heterogeneous information network to form structural features representing the network topology; Step 3: Perform cross-modal attribute interaction processing on the multi-class attribute features encoded in Step 1, reduce the dimension of the interacted attribute features by projection, and constrain the attribute feature expression from different dimensions through multi-objective loss optimization to obtain denoised and highly usable class attribute features. Step 4: Optimize the structural features obtained in Step 2. For nodes with missing attribute features, learn an average pattern based on their similar nodes. For each type of node, learn a type feature. Construct network structural features through multiple centrality indices. Update the edge weights of node pairs through a ternary gating mechanism. Optimize the network structure through edge weights to obtain a network topology with stronger information expression capabilities. Step 5: Enhance the expressive power of the multi-category attribute features obtained in Step 3 through cross-modal attribute interaction learning, realize the differentiated representation between different types of nodes through type interaction, retain the credible semantics in the original attribute features through residual compensation, and finally obtain the attribute embedding representing the node attributes through the multi-category attribute features. Step 6: The network topology obtained in Step 4 is used to obtain structural features through multiple centrality methods. Type-aware graph attention and residual modeling are used to model the heterogeneous relationships between nodes and alleviate information degradation. The structural features are recalibrated through SE-Block to suppress noise channels, and finally a highly discriminative and highly available structural embedding is obtained. Step 7: Perform cross-feature modality and multi-scale feature fusion on the attribute embeddings obtained in Step 5 and the structural embeddings obtained in Step 6 to obtain the optimal embedding representation of the node features; Step 8: The optimal node embedding representation obtained in Step 7 is used to obtain three different node importance evaluation results through three prediction paths of robust hierarchical importance prediction. These results are then dynamically weighted to obtain the final node importance value.

[0072] Furthermore, step 1 specifically includes: The initial node attribute information is divided into three categories: category, numerical, and text. The discrete category information is mapped to a continuous space for encoding through the embedding layer, the numerical information is standardized and encoded, and the text information is encoded through BERT to maintain their respective semantic characteristics.

[0073] Furthermore, step 2 specifically includes: The nodes, edges, and edge weights in the initial network are used to construct an adjacency matrix representing the current network.

[0074] Furthermore, step 3 also includes the following steps: Step 3.1: First, perform cross-modal attribute interaction: Assign an attention head to each modal attribute through multi-head attention, and perform interaction between information sequences to learn a more comprehensive representation of each modal attribute; Step 3.2: Learn a dimension-reduced projection matrix for each modal attribute to perform differentiated dimension reduction for each modal attribute; Step 3.3: Calculate the optimization loss based on the results of Step 3.1 and Step 3.2: Define four loss functions: semantic reconstruction loss, contrast loss, specificity loss, and preservation loss. The preservation loss is based on Step 3.1, and the semantic reconstruction loss is based on Step 3.2. In the semantic reconstruction loss, categorical attributes are based on KL divergence, numerical attributes are based on Huber loss, and text attributes are based on cosine similarity. The comparative loss is calculated using the following formula:

[0075] in, The attribute similarity between nodes ij; The specificity loss is calculated using the following formula:

[0076] in, The representative represents the characteristic representation of attribute s; The retention loss is calculated using the cosine distance metric for text attributes, the KL divergence metric for categorical attributes, and the smooth L1 metric for numerical attributes.

[0077] Furthermore, step 4 also includes the following steps: Step 4.1: First, perform type-aware feature fusion: learn common features of different attributes for each type of node, fill in type-related features for nodes with missing attributes, and fuse type semantics with existing features; Step 4.2: Perform gated edge weight calculation: Calculate edge weights using ternary gating based on the original graph structure representation, learnable weight matrix, and attribute features; Step 4.3: Calculate the optimization loss based on the results of Steps 4.1 and 4.2: Define three loss functions: structure preservation loss, degree distribution loss, and community preservation loss. Use these three loss functions to constrain the edge weights in Step 4.2. Based on these edge weights, add or delete edges in the graph to optimize the graph structure. The structural retention loss is calculated using the following formula:

[0078] in, Represents the edge set of the original graph. Represents the adjacency matrix of the original graph. The degree distribution loss is calculated using the following formula:

[0079] in, This represents the degree of node i before optimization. Represents the optimized degree; The community retention loss is calculated using the following formula:

[0080]

[0081] Where K represents the number of communities, This represents the number of community nodes.

[0082] Furthermore, step 5 specifically includes: The interaction between multi-class attribute features is realized based on multi-head self-attention. Adaptive gating is used to learn the node-level attribute fusion strategy to obtain the node attribute embedding. Then, the features of each type of node are learned through two layers of nonlinear transformation and adaptively fused with the node attribute embedding. Finally, the original attributes are fused through residual processing to obtain the attribute embedding.

[0083] Furthermore, step 6 specifically includes: Structural features are constructed on the optimized graph based on degree, path pattern, subgraph statistics, etc., and normalized. Then, attention modules for node types are added to the graph attention network to aggregate multi-level neighborhoods. Meanwhile, feature channels are recalibrated based on SE-Block. Finally, the original structural features are fused through residual processing to obtain the structural embedding.

[0084] Furthermore, step 7 specifically includes: Based on multi-head self-attention, internal semantic enhancement is performed on attribute and structure embeddings respectively. Then, three nonlinear layers with different fusion scales are constructed to perform multi-scale fusion of attribute and structure embeddings, and finally the optimal embedding representation of the node is obtained.

[0085] Furthermore, step 8 specifically includes: Based on MLP, three nonlinear importance prediction paths—single-layer, two-layer, and three-layer—are constructed to obtain three importance evaluation results. The three results are then weighted and fused to obtain the final node importance value.

[0086] Example 2 from Figure 1As can be seen, the method presented in this embodiment evaluates the importance of nodes in multi-source heterogeneous data in attribute heterogeneous information networks, ultimately constructing a CMGI-GAT with high-quality feature representation, sufficient structure-attribute synergy, and robust importance evaluation. First, differential encoding is performed on categorical, numerical, and textual attributes in the attribute heterogeneous information network. Then, cross-modal attribute interaction, dimensionality reduction projection, and multi-objective loss optimization are combined to denoise and enhance the usability of multi-class attribute features. Next, initial structural features are constructed and extracted. The network structure is optimized through type-aware feature fusion and ternary gating to obtain a network topology with stronger information representation capabilities. Then, the optimal attribute embedding and structural embedding for the current task are learned from the optimized multi-class attribute features and network topology, respectively. The optimal embedding representation of nodes is obtained through cross-feature modality and multi-scale fusion strategies. Finally, based on multiple prediction paths of robust hierarchical importance prediction, the importance of nodes is evaluated and dynamically weighted to obtain the final node importance value, thereby achieving accurate evaluation of node importance in attribute heterogeneous information networks, which is consistent with the expected goal of this embodiment.

[0087] Step 1: Preprocessing of multi-source heterogeneous features The initial node attribute information is divided into three categories: category, numerical, and text. The discrete category information is mapped to a continuous space for encoding through an embedding layer. The numerical information is standardized and then encoded. The text information is encoded through BERT to preserve their respective semantic characteristics. Then, based on the nodes, edges, and edge weights in the initial network, an adjacency matrix representing the current network is constructed, and finally, the attribute feature representation and network topology representation are obtained.

[0088] Step 2: Attribute Feature Optimization like Figure 2 As shown, firstly, the multi-type attribute features obtained in the first step are used to enhance the expressive power of cross-modal attribute interaction. Multi-head attention is used to learn the correlation and complementary information between attributes from different perspectives, resulting in attribute feature representations consistent with the initial feature dimensions. Then, dimensionality reduction projection is used to learn a projection function for each attribute. To ensure effective improvement in attribute quality and preservation of effective information in dimensionality reduction during cross-modal attribute interaction, a set of loss functions is used to achieve multi-objective optimization. In the semantic reconstruction loss calculation, categorical attributes are based on KL divergence, numerical attributes on Huber loss, and text attributes on cosine similarity. The contrastive loss is calculated using the following formula:

[0089] in, This represents the attribute similarity between nodes ij.

[0090] The specificity loss is calculated using the following formula:

[0091] in, The representative represents the characteristic representation of attribute s.

[0092] The retention loss is calculated using the cosine distance metric for text attributes, the KL divergence metric for categorical attributes, and the smooth L1 metric for numerical attributes.

[0093] Step 3: Structural Feature Optimization like Figure 3 As shown, a learnable default feature is first maintained for the missing attributes of nodes. During training, this default feature captures common attribute features related to the node type, i.e., the average pattern, which, together with the original attribute features, constitutes a complete attribute feature set. The original weighted adjacency matrix is ​​used as the structural feature to obtain the type feature, structural feature, and attribute feature. Using the obtained three features, an edge weight matrix is ​​learned through ternary gating. Based on the set graph optimization ratio and edge weights, the graph is optimized by deleting and adding edges. A set of loss functions is used to implement graph optimization constraints. The structure preservation loss is calculated using the following formula:

[0094] in, Represents the edge set of the original graph. This represents the adjacency matrix of the original graph.

[0095] The degree distribution loss is calculated using the following formula:

[0096] in, This represents the degree of node i before optimization. This represents the optimized degree.

[0097] The community retention loss is calculated using the following formula:

[0098]

[0099] Where K represents the number of communities, This represents the number of community nodes.

[0100] Step 4: Attribute Embedding Representation Learning like Figure 4 As shown, the multimodal attribute features obtained in the second step are used to perform task-oriented type-level cross-modal interaction learning, and a node-level attribute fusion strategy is learned. The node-level attribute features are calculated using the following formula:

[0101] in, The weights representing modal attributes, This represents the embedding of node attributes after interaction.

[0102] Then, task-oriented learning is used to learn the common attribute semantics of each type of node. The common attribute semantics are then weighted and fused with the node attribute embeddings after interaction using learnable parameters at the node level to obtain differentiated attribute embeddings. Finally, residual compensation is used to prevent important information from being suppressed or lost, ultimately yielding the optimal node attribute embedding.

[0103] Step 5: Structural Embedding Representation Learning like Figure 5 As shown, using the optimized network topology obtained in the third step, an initial structural embedding is obtained through normalization of various centrality features such as node degree, k-order neighbor count, and path pattern frequency. Then, a graph attention network is used to aggregate the structural features of multi-order neighbor nodes, and an attention weight is learned for each node type to obtain a perceptual type structural embedding. During the multi-order aggregation process, an adaptive residual connection is used to connect multiple layers of the network to prevent the disappearance of key structural features. Finally, SE-Block is used to learn the weights of each dimension of the structural embedding, highlighting key channels and suppressing noisy channels, ultimately obtaining the optimal structural embedding of the node.

[0104] Step 6: Node Importance Assessment like Figure 6 As shown, using the attribute embeddings obtained in step 5 and the structural embeddings obtained in step 6, node semantics at different scales are projected onto three scales. Then, a graph attention network is used to learn the multi-level heterogeneous features of nodes, focusing on the differentiated contributions of attribute features and structural features in importance propagation. Residual compensation is used to avoid the loss of important features, and node-level feature fusion is performed to obtain the optimal node feature representation. Finally, based on MLP, three nonlinear importance prediction paths (single-layer, two-layer, and three-layer) are constructed to obtain three importance evaluation results. The three results are weighted and fused to obtain the final node importance value.

[0105] Depend on Figure 1 It can be seen that, compared with traditional methods for evaluating the importance of nodes in heterogeneous attribute information networks, this method has the advantages of strong adaptability, accurate importance evaluation, and robustness. Compared with the original attributes and structural features, the introduction of attribute and structural optimization reduces the noise of attribute information and improves the information expression ability of the network. Compared with traditional node importance evaluation methods, the multi-granularity interactive processing makes the task of evaluating the importance of nodes in heterogeneous attribute information networks oriented towards multi-source heterogeneous information more accurate and robust.

[0106] The above description of several specific embodiments further details the technical solution provided by the present invention in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, combinations of embodiments, and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the importance of nodes in attribute-heterogeneous information networks based on CMGI-GAT, characterized by evaluating node importance through multi-granularity optimization and learning of attribute features and structural features, wherein... The node importance assessment method includes the following steps: Step 1: Perform differential encoding on multi-class attribute features with multiple semantic categories in the attribute heterogeneous information network; Step 2: Construct an initial adjacency matrix for nodes and edges in the heterogeneous information network to form structural features representing the network topology; Step 3: Perform cross-modal attribute interaction processing on the multi-class attribute features encoded in Step 1, reduce the dimension of the interacted attribute features by projection, and constrain the attribute feature expression from different dimensions through multi-objective loss optimization to obtain denoised and highly usable class attribute features. Step 4: Optimize the structural features obtained in Step 2. For nodes with missing attribute features, learn an average pattern based on their similar nodes. For each type of node, learn a type feature. Construct network structural features through multiple centrality indices. Update the edge weights of node pairs through a ternary gating mechanism. Optimize the network structure through edge weights to obtain a network topology with stronger information expression capabilities. Step 5: Enhance the expressive power of the multi-category attribute features obtained in Step 3 through cross-modal attribute interaction learning, realize the differentiated representation between different types of nodes through type interaction, retain the credible semantics in the original attribute features through residual compensation, and finally obtain the attribute embedding representing the node attributes through the multi-category attribute features. Step 6: The network topology obtained in Step 4 is used to obtain structural features through multiple centrality methods. Type-aware graph attention and residual modeling are used to model the heterogeneous relationships between nodes and alleviate information degradation. The structural features are recalibrated through SE-Block to suppress noise channels, and finally a highly discriminative and highly available structural embedding is obtained. Step 7: Perform cross-feature modality and multi-scale feature fusion on the attribute embeddings obtained in Step 5 and the structural embeddings obtained in Step 6 to obtain the optimal embedding representation of the node features; Step 8: The optimal node embedding representation obtained in Step 7 is used to obtain three different node importance evaluation results through three prediction paths of robust hierarchical importance prediction. These results are then dynamically weighted to obtain the final node importance value.

2. The method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT according to claim 1, characterized in that, Specifically, step 1 involves dividing the initial node attribute information into three categories: category, numerical, and text. The discrete category information is mapped to a continuous space for encoding through an embedding layer. The numerical information is standardized and then encoded. The text information is encoded using BERT to preserve their respective semantic characteristics.

3. The method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT according to claim 1, characterized in that, Step 3 also Includes the following steps, Multimodal attributes are interactively enhanced by learning the relationships between different modal attributes through an attention mechanism to obtain enhanced attribute representations; Each enhanced modal attribute is subjected to projection transformation to reduce feature dimension and unify representation space; Based on the attribute interaction results and dimensionality reduction results, optimization constraints are constructed to jointly optimize the attribute representation. The optimization constraints include at least the reconstruction constraints for preserving the original semantic information, the contrast constraints for enhancing the ability to distinguish similar nodes, the specific constraints for preserving the independent expressive ability of different attribute modalities, and the preservation constraints for controlling the degree of deviation of attribute representation before and after optimization. Among them, different types of attributes are constrained by using a difference measurement method adapted to their data features.

4. The method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT according to claim 1, characterized in that, Step 4 also includes the following steps: Based on node type, node features are completed and fused to learn common features related to each type of node and fill in the corresponding type features for nodes with missing attributes, so as to form a node representation containing type semantics. Based on the original graph structure, learnable parameters and node attribute features, the edge association strength between nodes is calculated and the edge weights are updated accordingly. Based on the updated edge weights, structural optimization constraints are constructed. The structural optimization constraints include at least structural preservation constraints to maintain the original graph connection relationship, degree distribution constraints to limit the change of node connection distribution, and community preservation constraints to maintain the original community organization relationship. The connection relationship in the graph is adjusted based on the edge weights to obtain the optimized network structure.

5. The method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT according to claim 1, characterized in that, Specifically, step 5 involves: realizing the interaction between multi-category attribute features based on multi-head self-attention; using adaptive gating to learn the node-level attribute fusion strategy to obtain node attribute embeddings; then learning the features of each type of node through two layers of nonlinear transformations and adaptively fusing them with the node attribute embeddings; finally fusing the original attributes through residual processing to obtain attribute embeddings.

6. The method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT according to claim 1, characterized in that, Specifically, step 6 involves constructing structural features on the optimized graph based on degree, path pattern, subgraph statistics, etc., and normalizing them. Then, attention modules for the types of nodes to be focused on are added to the graph attention network to aggregate multi-level neighborhoods. At the same time, feature channels are recalibrated based on SE-Block. Finally, the original structural features are fused through residual processing to obtain the structural embedding.

7. The method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT according to claim 1, characterized in that, Specifically, step 7 involves performing internal semantic enhancement on attribute and structural embeddings based on multi-head self-attention, then constructing three nonlinear layers with different fusion scales to perform multi-scale fusion of attribute and structural embeddings, ultimately obtaining the optimal embedding representation of the node.

8. The method for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT according to claim 1, characterized in that, Specifically, step 8 involves constructing three nonlinear importance prediction paths—single-layer, two-layer, and three-layer—based on MLP, obtaining three importance evaluation results, and then weighting and fusing the three results to obtain the final node importance value.

9. A device for evaluating the importance of nodes in a heterogeneous information network based on CMGI-GAT, which evaluates node importance through multi-granularity optimization and learning of attribute features and structural features, characterized in that... The node importance assessment device includes the following modules: Module 1: Differentiated encoding of multi-class attribute features with multiple semantic categories in heterogeneous attribute information networks; Module 2: Construct an initial adjacency matrix for nodes and edges in a heterogeneous information network to form structural features representing the network topology; Module 3: Perform cross-modal attribute interaction processing on the multi-class attribute features encoded in Step 1, reduce the dimension of the interacted attribute features by projection, and constrain the attribute feature expression from different dimensions through multi-objective loss optimization to obtain denoised and highly available class attribute features. Module 4: Optimize the structural features obtained in step 2. Learn an average pattern for nodes with missing attribute features based on their similar nodes. Learn a type feature for each type of node. Construct network structural features through multiple centrality indices. Update the edge weights of node pairs through a ternary gating mechanism. Optimize the network structure through edge weights to obtain a network topology with stronger information expression capabilities. Module 5: Enhance the expressive power of the multi-category attribute features obtained in step 3 through cross-modal attribute interaction learning, realize the differentiated representation between different types of nodes through type interaction, retain the credible semantics in the original attribute features through residual compensation, and finally obtain the attribute embedding representing the node attributes through the multi-category attribute features; Module 6: The network topology obtained in step 4 is used to obtain structural features through various centrality methods. Type-aware graph attention and residual modeling are used to model the heterogeneous relationships between nodes and alleviate information degradation. The structural features are recalibrated through SE-Block to suppress noise channels, and finally a highly discriminative and highly available structural embedding is obtained. Module 7: Perform cross-feature modality and multi-scale feature fusion on the attribute embeddings obtained in step 5 and the structural embeddings obtained in step 6 to obtain the optimal embedding representation of the node features; Module 8: The optimal node embedding representation obtained in step 7 is used to obtain three different node importance evaluation results through three prediction paths of robust hierarchical importance prediction. These results are then dynamically weighted to obtain the final node importance value.

10. A computer, comprising a processor and a storage medium, characterized in that, When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1.

Images