Industrial competitive power scoring method based on graph neural network and data fusion

Abstract

The present application relates to the technical field of industrial competitiveness evaluation, and discloses an industrial competitiveness quantitative evaluation method based on a graph neural network and data fusion, comprising: constructing a heterogeneous graph; collecting real-time data and updating it to the heterogeneous graph to obtain a real-time heterogeneous graph, training and feature learning on the real-time heterogeneous graph based on a graph neural network model to obtain a competitiveness feature vector of a target node; constructing an evaluation mapping system, mapping the competitiveness feature vector of the target node to a quantitative evaluation result based on the evaluation mapping system, effectively integrating multi-source heterogeneous data in real time, capturing complex network structures and dynamic relationships within an industry through a graph neural network, and realizing accurate and dynamic quantitative evaluation of industrial competitiveness, thereby providing strong quantitative support for enterprise strategic decision-making and industrial policy formulation.

Description

Technical Field

[0001] This application relates to the field of industrial competitiveness assessment technology, and in particular to a quantitative scoring method for industrial competitiveness based on graph neural networks and data fusion. Background Technology

[0002] Industrial competitiveness is influenced by a variety of factors, including upstream and downstream enterprise relationships, market dynamics, and policy environment. Traditional industrial competitiveness assessment methods mainly rely on static indicator systems and expert judgment. While these methods can reflect the basic situation of an industry to some extent, they are difficult to effectively capture and quantify the complex network effects and structural characteristics between enterprises within an industry and between different industries. Moreover, market environment data changes rapidly, making it difficult to accurately match and update entities and respond in real time to the dynamic evolution of the competitive landscape of an industry. This results in assessment results that are lagging and one-sided, and cannot provide accurate and dynamic quantitative basis for enterprise strategic decision-making and industrial policy formulation. Summary of the Invention

[0003] To address the aforementioned technical challenges, this application provides a quantitative scoring method for industrial competitiveness based on graph neural networks and data fusion. This method constructs a heterogeneous graph, collects real-time data, and updates the graph to obtain a real-time heterogeneous graph. A graph neural network model is then used to train and learn representations of the real-time heterogeneous graph, yielding the competitiveness feature vector of the target node. A scoring mapping system is then constructed to map this competitiveness feature vector to a quantitative scoring result. This method effectively integrates multi-source heterogeneous data in real time, capturing the complex network structure and dynamic relationships within the industry through graph neural networks. This enables accurate and dynamic quantitative assessment of industrial competitiveness, providing strong quantitative support for corporate strategic decision-making and industrial policy formulation.

[0004] In some embodiments of this application, a method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion is provided, including:

[0005] Construct a heterogeneous graph; Real-time data is collected and updated to the heterogeneous graph to obtain a real-time heterogeneous graph. The real-time heterogeneous graph is then trained and its representation is learned based on a graph neural network model to obtain the competitive feature vector of the target node. A scoring mapping system is constructed, and the competitiveness feature vector of the target node is mapped to a quantitative scoring result based on the scoring mapping system.

[0006] In some embodiments of this application, constructing a heterogeneous graph includes: Define the node type system, edge type system, and attribute patterns of heterogeneous graphs; Build an industry database; Entities and relationships are extracted from the industry database. Based on the extraction results and definition results, graph nodes and edges are created in batches. Each graph node has a corresponding unique entity. Populate each graph node with basic attributes obtained from the data source; The initial weight of each edge with a relationship is calculated based on the strength of the relational evidence and the multi-source calculation. Construct a heterogeneous graph based on several graph nodes, edges, basic attributes, and initial weights.

[0007] In some embodiments of this application, the method further includes: Several quality assessment indicators are pre-defined, including pattern assessment indicators, data assessment indicators, logic assessment indicators, and structural assessment indicators. Based on quality assessment metrics, collect the current heterogeneous graph and a validation dataset that matches each assessment metric; Generate quality assessment values ​​for the corresponding quality assessment indicators based on the validation dataset; The overall quality assessment value of the current heterogeneous graph is obtained by weighting the quality assessment values ​​of all quality assessment indicators and the corresponding weight coefficients of the indicators. Determine whether to correct the heterogeneous graph based on the comprehensive quality assessment value. If so, identify the abnormal graph data and generate the corresponding correction instructions based on the preset correction strategy library. The corrected heterogeneous graph is re-evaluated until the overall quality evaluation value is greater than the preset threshold. Then the corrected heterogeneous graph is replaced with the current heterogeneous graph.

[0008] In some embodiments of this application, real-time data is collected and updated to the heterogeneous graph to obtain a real-time heterogeneous graph, including: Real-time data on enterprises, industries, and market environment is collected, and the real-time data is preprocessed, including format parsing, outlier filtering, and data standardization. Entity recognition and event extraction are performed on the preprocessed real-time data, and the extracted entities are matched with the nodes in the current heterogeneous graph in multiple dimensions to obtain the matching degree. Based on the matching degree, the associated nodes and associated edges of each extracted entity are obtained, and an update data packet containing the node attribute update of the corresponding associated node, the edge weight adjustment or addition of the corresponding associated edge, and the newly added edge is generated. The update data packet is submitted to the graph database to perform basic data updates and trigger a cascading update process of the nodes and edges associated with the updated content, resulting in a real-time heterogeneous graph.

[0009] In some embodiments of this application, the matching degree includes: One randomly selected entity is chosen as the target entity; Extract the feature attributes of the target entity and compare them with the corresponding attributes of each node in the current heterogeneous graph to obtain the attribute matching degree; Set the topological distance between the target entity and each node, and extract the corresponding node's comparative network structure diagram according to the topological distance; The target entity is placed in its local relation, and several network structure diagrams of the target entity are generated by combining the comparison network structure diagram of each node. The structure matching degree is obtained by comparing each of these diagrams with the corresponding comparison network structure diagram. Extract several behavioral features of the target entity and construct a sequence of behavioral features of the target entity; Set the number of comparison features between the target entity and each node, filter out several behavioral features to be compared with each node by combining the behavioral feature sequence, and compare them with the behavioral features of the corresponding node at the same time node to obtain the time behavior matching degree. The matching degree between the target entity and the corresponding node is generated by comparing the attribute matching degree, structural matching degree, and temporal behavior matching degree between the target entity and the same node. The matching degree between each extracted entity and each node is generated sequentially.

[0010] In some embodiments of this application, the graph neural network model includes: An initial heterogeneous graph neural network model is constructed, which includes a node-level heterogeneous attention layer, a relationship-aware message passing layer, and a temporal feature fusion layer. The initial heterogeneous graph neural network model is trained using a hybrid task learning objective, which includes a self-supervised learning task and a supervised learning task. The self-supervised learning task includes several preset training tasks, and the supervised learning task includes learning discriminative features related to competitiveness. Adversarial samples are introduced during training, and the corresponding model's robustness coefficient is calculated. When the training result accuracy is greater than a preset accuracy threshold and the robustness coefficient is greater than a preset coefficient threshold, the trained model is set as a graph neural network model.

[0011] In some embodiments of this application, the competitive feature vector of the target node is obtained, including: The real-time heterogeneous graph is input into the graph neural network model. The node-level heterogeneous attention layer based on the model obtains the attribute features of the target node, the relationship-aware message passing layer based on the model obtains the structural association features of the target node, and the temporal feature fusion layer based on the model obtains the temporal behavior features of the target node. The target node's attribute features, structural association features, and temporal behavior features are concatenated to form the target node's competitive feature vector.

[0012] In some embodiments of this application, a scoring mapping system is constructed, and the competitiveness feature vector of the target node is mapped to a quantitative scoring result based on the scoring mapping system, including: Based on the multi-dimensional needs of industrial competitiveness assessment, a scoring mapping system is constructed, which includes a basic scoring mapping table, a dynamic weighting mechanism, and quantitative scoring calibration. The competitive feature vector of the target node is input into the scoring mapping system. The quantitative scoring result of the target node is output through three steps: looking up the basic scoring mapping table, dynamically adjusting the feature weights, and calibrating the quantitative score.

[0013] In some embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the industrial competitiveness quantitative scoring method based on graph neural networks and data fusion described above.

[0014] In some embodiments of this application, a non-transitory computer-readable storage medium is also provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the above-described method for quantitative scoring of industrial competitiveness based on graph neural networks and data fusion.

[0015] The industrial competitiveness quantitative scoring method based on graph neural networks and data fusion in this application has the following advantages compared with the prior art: By constructing a heterogeneous graph, collecting real-time data, and updating the graph, a real-time heterogeneous graph is obtained. Based on a graph neural network model, the real-time heterogeneous graph is trained and its representation is learned to obtain the competitiveness feature vector of the target node. A scoring mapping system is then constructed to map the competitiveness feature vector into a quantitative scoring result. This approach effectively integrates multi-source heterogeneous data in real time and captures the complex network structure and dynamic relationships within the industry through graph neural networks. This enables accurate and dynamic quantitative evaluation of industry competitiveness, providing strong quantitative support for corporate strategic decision-making and industrial policy formulation. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the quantitative scoring method for industrial competitiveness based on graph neural networks and data fusion in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0017] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0018] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0019] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] like Figure 1 As shown in the embodiment of this application, the method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion includes: Step S101: Construct a heterogeneous graph; Step S102: Collect real-time data and update it to the heterogeneous graph to obtain a real-time heterogeneous graph. Train and learn the representation of the real-time heterogeneous graph based on the graph neural network model to obtain the competitive feature vector of the target node. Step S103: Construct a scoring mapping system, and map the competitiveness feature vector of the target node to a quantitative scoring result based on the scoring mapping system.

[0022] In this embodiment, the target node is pre-defined and refers to a specific entity that needs to be quantitatively scored in the industry competitiveness assessment scenario. This entity can be a company in the industry, a specific sub-sector of the industry, or a specific link in the industrial chain, etc.

[0023] In some embodiments of this application, constructing a heterogeneous graph includes: Define the node type system, edge type system, and attribute patterns of heterogeneous graphs; Build an industry database; Entities and relationships are extracted from the industry database. Based on the extraction results and definition results, graph nodes and edges are created in batches. Each graph node has a corresponding unique entity. Populate each graph node with basic attributes obtained from the data source; The initial weight of each edge with a relationship is calculated based on the strength of the relational evidence and the multi-source calculation. Construct a heterogeneous graph based on several graph nodes, edges, basic attributes, and initial weights.

[0024] In this embodiment, the node type system includes different types of nodes such as enterprise nodes, industry nodes, regional nodes, policy nodes, technology nodes, and market nodes, and each type of node is further divided into subtypes according to business type.

[0025] In this embodiment, the edge type system includes edges used to characterize various relationships, including "belonging to" and "located in" edges that characterize organizational structure relationships, "supplying to" and "competing with" edges that characterize economic transaction relationships, "owning patents" and "using technology" edges that characterize technological associations, and "affected by" edges that characterize environmental impacts, etc.

[0026] In this embodiment, the attribute system includes a set of attributes defined for each type of node and edge, including basic attributes, temporal attributes, and network structure attributes calculated through graph computation.

[0027] In this embodiment, the industry database includes structured databases, semi-structured documents, and unstructured text.

[0028] In this embodiment, entity and relation extraction refers to mapping entities in multi-source data to a defined node type system through entity recognition and disambiguation technology to obtain several graph nodes and their corresponding node types. Then, through a rule-based and machine learning-based relation extraction engine, various relationships between entities are identified from the data, and attributes and confidence scores are assigned.

[0029] In this embodiment, calculating the initial weight based on the strength and multi-source nature of relational evidence refers to assessing the reliability of the source data for each relation and assigning evidence weights to different data sources; counting the frequency of the same relation in different data sources, with higher frequencies resulting in larger initial weight base values ​​for the relation; and combining the evidence weights and frequency base values ​​to calculate the initial weight of each edge using a weighted summation formula, thereby reflecting the true reliability and importance of the relation.

[0030] In this embodiment, the construction process of the heterogeneous graph achieves accurate modeling of complex entities and relationships in the industrial competitive environment through multi-dimensional definition, multi-source data integration and structured processing, providing a structured foundation for subsequent dynamic updates and deep learning.

[0031] In some embodiments of this application, the method further includes: Several quality assessment indicators are pre-defined, including pattern assessment indicators, data assessment indicators, logic assessment indicators, and structural assessment indicators. Based on quality assessment metrics, collect the current heterogeneous graph and a validation dataset that matches each assessment metric; Generate quality assessment values ​​for the corresponding quality assessment indicators based on the validation dataset; The overall quality assessment value of the current heterogeneous graph is obtained by weighting the quality assessment values ​​of all quality assessment indicators and the corresponding weight coefficients of the indicators. Determine whether to correct the heterogeneous graph based on the comprehensive quality assessment value. If so, identify the abnormal graph data and generate the corresponding correction instructions based on the preset correction strategy library. The corrected heterogeneous graph is re-evaluated until the overall quality evaluation value is greater than the preset threshold. Then the corrected heterogeneous graph is replaced with the current heterogeneous graph.

[0032] In this embodiment, the pattern evaluation index is used to check whether the node type and edge type of the heterogeneous graph conform to the defined node type system, whether the attribute fields are complete and whether the data format is matched. If the type conforms, the attribute fields are complete and the format is matched, the quality evaluation value of the pattern evaluation index is larger, and vice versa.

[0033] In this embodiment, statistical analysis is performed on the numerical range, outlier ratio, and missing rate of node and edge attributes using data evaluation metrics. When the values ​​exceed a reasonable range, the outlier ratio, and the missing rate are higher, the corresponding quality evaluation value is lower, and vice versa.

[0034] In this embodiment, the consistency of relationships between entities is detected by logical evaluation indicators. The greater the consistency, the higher the corresponding quality evaluation value, and vice versa. For example, the industry node pointed to by the "belongs to" edge of an enterprise node must match the enterprise's main business classification, and the enterprise node connected by the "compete with" edge must be in the same market segment.

[0035] In this embodiment, structural evaluation indicators are used to measure the overall topological rationality and information transmission efficiency of heterogeneous graphs. Specifically, these indicators include the average path length of the computation graph, clustering coefficient, node degree distribution characteristics, and the proportion of core-edge structures. When all parameters of the structural evaluation indicators are within the optimal threshold range, the corresponding quality evaluation value is larger, and vice versa.

[0036] In this embodiment, the quality assessment values ​​of the quality assessment indicators are all in the range of 0-1, which is achieved through normalization processing. The original statistical values ​​of each indicator are mapped to this range, and the comprehensive quality assessment value is calculated by weighted summation formula. The weight coefficients are preset based on the degree of influence of each indicator on the modeling accuracy of heterogeneous graphs.

[0037] In this embodiment, the preset correction strategy library contains correction rules for different anomaly types. For example, when a node attribute missing rate is detected to be too high, the K-nearest neighbor filling method based on graph structure similarity is adopted; when there is a logical conflict in the relationship, a manual review process is triggered and correction is made in conjunction with the industry knowledge graph; when the topology structure evaluation is not up to standard, the network structure is optimized by adjusting the edge weight allocation or supplementing key connection nodes.

[0038] In this embodiment, the aforementioned quality assessment and correction mechanism ensures that the heterogeneous graph accurately reflects the true state of the industrial competitive environment, laying a reliable data foundation for the subsequent extraction of competitive feature vectors.

[0039] In some embodiments of this application, real-time data is collected and updated to the heterogeneous graph to obtain a real-time heterogeneous graph, including: Real-time data on enterprises, industries, and market environment is collected, and the real-time data is preprocessed, including format parsing, outlier filtering, and data standardization. Entity recognition and event extraction are performed on the preprocessed real-time data, and the extracted entities are matched with the nodes in the current heterogeneous graph in multiple dimensions to obtain the matching degree. Based on the matching degree, the associated nodes and associated edges of each extracted entity are obtained, and an update data packet containing the node attribute update of the corresponding associated node, the edge weight adjustment or addition of the corresponding associated edge, and the newly added edge is generated. The update data packet is submitted to the graph database to perform basic data updates and trigger a cascading update process of the nodes and edges associated with the updated content, resulting in a real-time heterogeneous graph.

[0040] In this embodiment, real-time data is continuously or periodically collected from designated data sources on the Internet, based on web crawlers or by calling public data interfaces, and is related to industrial competitiveness. Designated data sources include, but are not limited to: publicly available annual reports of enterprises, stock exchange announcements, government statistical data platforms, industry research report websites, news and information websites, patent databases, and recruitment websites. The collected real-time data includes at least the latest operational data (such as enterprise financial reports and production data) and market data.

[0041] In this embodiment, the associated nodes and associated edges of each extracted entity are obtained based on the matching degree. This means that when the matching degree is greater than or equal to the preset matching threshold (0.85 in this application), it is determined to be the same entity; when the matching degree is less than the preset matching threshold but greater than or equal to the minimum association threshold (0.5 in this application), it is determined to be an associated node and a possible associated edge; when the matching degree is less than the minimum association threshold, it is determined to be a new entity, and an update request for the new node is generated.

[0042] In this embodiment, the updated data packet includes: for the same entity or related nodes, if the real-time data contains newly added or changed attributes of the node (such as the latest revenue data of the enterprise, changes in the number of patents), a corresponding node attribute update instruction is generated; for related edges, if the real-time data reflects changes in the strength of the original relationship (such as an increase or decrease in the frequency of cooperation between enterprises), an edge weight adjustment instruction is generated based on a preset weight update algorithm (such as dynamically adjusting the edge weight by combining the frequency and reliability of new evidence); if a new relationship between entities is identified in the real-time data, an instruction to add an edge is generated; for cases determined to be new entities, an instruction to add a node containing node type and basic attributes is generated.

[0043] In this embodiment, a graph database is a database management system specifically designed for storing and processing graph structure data. It uses nodes, edges, and attributes as basic building blocks to directly and naturally represent entities (nodes) and various relationships (edges) between entities.

[0044] In this embodiment, after the update data packet is submitted to the graph database, the graph database will first perform a validity check on the update data packet, including whether the data format conforms to the preset specifications, whether the associated nodes and edges are within the type system of the current heterogeneous graph, and whether the attribute values ​​meet the data type constraints.

[0045] In this embodiment, if the verification passes, the graph database performs basic data updates according to the update data packet. Specifically, for node attribute updates and edge weight adjustments, or for adding nodes and edges, node attribute updates refer to directly locating the target node and replacing or appending the values ​​of the corresponding attribute fields, while recording the timestamp and data source of the attribute update. Edge weight adjustments refer to updating the weight attributes of the corresponding edge according to the edge identifier and new weight value specified in the update data packet, and triggering the recalculation of the network structure attributes related to the edge weight. For adding edges or nodes, it is first checked whether there are duplicate entities or relationships. If not, a new node or edge instance is created in the graph database and assigned basic attributes and initial weights.

[0046] In this embodiment, after the basic data update is completed, the graph database automatically initiates a cascading update process, identifies nodes and edges that are directly or indirectly related to the updated content, and automatically calculates the influence weight of the node on its directly connected nodes based on the message passing mechanism of the graph neural network when a node's attributes change, and updates the weight values ​​of the relevant edges. For newly added entities or relationships, the optimal insertion position in the heterogeneous graph is determined by calculating the similarity with the existing graph structure, while adjusting the connection relationships of surrounding nodes to maintain the integrity of the graph structure. For market dynamic data with high time sensitivity (such as price fluctuations and policy changes), a time decay factor is set to dynamically adjust the weight of historical data to ensure that the heterogeneous graph can reflect the latest changes in the current industry competitive landscape in real time.

[0047] In this embodiment, after performing basic data updates and cascading updates based on the graph database according to the update data packets, the dynamic evolution of heterogeneous graphs can be realized, so that the graph structure and attribute information always keep in sync with the actual changes in the industrial competitive environment, thereby obtaining real-time heterogeneous graphs and providing timely and accurate input data for the subsequent training of graph neural networks.

[0048] In some embodiments of this application, the matching degree includes: One randomly selected entity is chosen as the target entity; Extract the feature attributes of the target entity and compare them with the corresponding attributes of each node in the current heterogeneous graph to obtain the attribute matching degree; Set the topological distance between the target entity and each node, and extract the corresponding node's comparative network structure diagram according to the topological distance; The target entity is placed in its local relation, and several network structure diagrams of the target entity are generated by combining the comparison network structure diagram of each node. The structure matching degree is obtained by comparing each of these diagrams with the corresponding comparison network structure diagram. Extract several behavioral features of the target entity and construct a sequence of behavioral features of the target entity; Set the number of comparison features between the target entity and each node, filter out several behavioral features to be compared with each node by combining the behavioral feature sequence, and compare them with the behavioral features of the corresponding node at the same time node to obtain the time behavior matching degree. The matching degree between the target entity and the corresponding node is generated by comparing the attribute matching degree, structural matching degree, and temporal behavior matching degree between the target entity and the same node. The matching degree between each extracted entity and each node is generated sequentially.

[0049] In this embodiment, the feature attributes include name, code (such as unified social credit code of enterprise, stock code) and core value (such as registered address, legal representative). The attribute matching degree is generated based on the consistency of each attribute. When the consistency of all attributes reaches 100%, the attribute matching degree is 1, and the closer it is to 0, the better.

[0050] In this embodiment, the topological distance refers to the path length of a node in the current heterogeneous graph within the graph structure. The topological distance is set based on the attribute matching degree. In this application, when the attribute matching degree is between 0.8-1, 0.6-0.8, 0.4-0.6, 0.2-0.4, and 0-0.2, the topological distances are set to 1, 2, 3, 4, and 5, respectively. This corresponds to extracting the local network structure graphs of a node's direct neighbor nodes (distance 1), neighbor's neighbor nodes (distance 2), ... as a comparison network structure graph.

[0051] In this embodiment, the structural matching degree is determined by calculating the overlap of neighbor types and the similarity of relationship paths between the network structure graph generated by the target entity and the comparison network structure graph of the nodes. The overlap of neighbor types refers to the matching ratio of the types of each neighbor node in the target entity's network structure graph with the types of the corresponding neighbor nodes in the comparison network structure graph. The similarity of relationship paths is calculated by comparing the edit distance of the type sequences of relationship paths between entities in the two structure graphs (such as "supply to → located in" and "own patent → use technology"). The higher the ratio and the smaller the edit distance (i.e., the more similar the paths), the closer the structural matching degree is to 1, and vice versa.

[0052] In this embodiment, behavioral characteristics include, but are not limited to, dynamic indicators that change over time, such as the company's annual revenue growth rate, R&D investment ratio, market share changes, and patent application number trends. The behavioral characteristic sequence is a vector sequence formed by arranging the behavioral characteristic values ​​of the target entity at multiple consecutive time points in chronological order.

[0053] In this embodiment, the number of comparison features is set based on the mean processing result of attribute matching degree and structure matching degree. The closer the mean processing result is to 1, the fewer the comparison features are, and vice versa. For example, when the mean processing result is 0.8-1, 0.6-0.8, 0.4-0.6, 0.2-0.4, 0-0.2, the corresponding number of comparison features is 1, 3, 5, 7, 9. The number of comparison features is set according to the number of features in the actual behavioral feature sequence. This dynamically adjusts the number of dimensions used to measure the similarity of behavioral features between the target entity and the comparison entity, ensuring that core features are focused when the attributes and structure are highly matched, and that the comparison dimensions are expanded to fully capture differences when the matching degree is low. This improves the accuracy of extracting the competitive feature vectors of the subsequent real-time heterogeneous graph and the target node.

[0054] In this embodiment, the temporal behavior matching degree is determined by calculating the cosine similarity or Euclidean distance of the behavioral feature values ​​of the target entity and the comparison node at the same time node. The larger the cosine similarity or the smaller the Euclidean distance, the closer the temporal behavior matching degree is to 1, and vice versa.

[0055] In this embodiment, the matching degree between the target entity and the corresponding node is obtained by weighted summation of attribute matching degree, structural matching degree and temporal behavior matching degree. The weights of attribute matching degree, structural matching degree and temporal behavior matching degree are 0.4, 0.3 and 0.3 respectively, and the final matching degree is also normalized to the 0-1 range.

[0056] In this embodiment, by calculating the matching degree between each extracted entity and the existing nodes in the heterogeneous graph, the relationship between entities can be accurately identified, providing a reliable matching basis for the subsequent fusion of real-time data and heterogeneous graph, and ensuring that the real-time heterogeneous graph accurately depicts the dynamic changes in the industrial competitive environment.

[0057] In some embodiments of this application, the graph neural network model includes: An initial heterogeneous graph neural network model is constructed, which includes a node-level heterogeneous attention layer, a relationship-aware message passing layer, and a temporal feature fusion layer. The initial heterogeneous graph neural network model is trained using a hybrid task learning objective, which includes a self-supervised learning task and a supervised learning task. The self-supervised learning task includes several preset training tasks, and the supervised learning task includes learning discriminative features related to competitiveness. Adversarial samples are introduced during training, and the corresponding model's robustness coefficient is calculated. When the training result accuracy is greater than a preset accuracy threshold and the robustness coefficient is greater than a preset coefficient threshold, the trained model is set as a graph neural network model.

[0058] In this embodiment, the node-level heterogeneous attention layer is used to process the feature differences of different types of nodes in the heterogeneous graph, and to assign learnable attention weights to different types of nodes, so that the model can adaptively focus on the node types and their attribute information that are more important to the competitiveness assessment.

[0059] In this embodiment, the relationship-aware message passing layer focuses on modeling complex relationships between entities and designs differentiated message passing mechanisms for different types of edges to ensure that relational semantic information can be effectively encoded into node features.

[0060] In this embodiment, the temporal feature fusion layer integrates the dynamic behavioral features of entities that change over time. By introducing a Long Short-Term Memory (LSTM) network, it captures the evolution of corporate behavior patterns at different time points, thereby enhancing the model's ability to perceive dynamic changes in competitiveness.

[0061] In this embodiment, several preset training tasks include, but are not limited to, node-level attribute prediction tasks, edge-level relationship prediction tasks, and graph-level industry trend prediction tasks. The node-level attribute prediction task aims to predict the future key financial indicators of enterprise nodes; the edge-level relationship prediction task aims to predict whether there is a potential supply chain or competitive relationship between any two enterprise nodes; and the graph-level industry trend prediction task aims to predict the comprehensive development trend index of the entire industry subgraph in the future time period.

[0062] In this embodiment, the supervised learning task introduces labeled enterprise competitiveness score data (such as industry ranking and financial health index) to guide the model to learn discriminative features directly related to competitiveness. Discriminative features include, but are not limited to, the correlation pattern between high R&D investment ratio, number of core patents and market share growth, so that the model can focus on dimensions that have a key impact on competitiveness assessment when extracting node features.

[0063] In this embodiment, by constructing an initial heterogeneous graph neural network model, it is possible to fully capture the comprehensive impact of different types of nodes, diverse relationships, and dynamic time-series information on enterprise competitiveness in the industrial competitive environment.

[0064] In this embodiment, adversarial samples refer to input samples generated by adding minor perturbations to the original entity features.

[0065] In this embodiment, the robustness coefficient refers to the degree of difference between the competitive feature vectors of the adversarial samples and the original samples output by the model. It is quantified by calculating the cosine similarity between the competitive feature vectors of the adversarial samples and the original samples. The higher the robustness coefficient, the stronger the model's robustness to small perturbations and the more stable the output results.

[0066] In this embodiment, the preset accuracy threshold is set to 0.9, and the preset coefficient threshold is set according to the volatility characteristics of industry data. In this application, for traditional manufacturing industries, since their industry data is relatively stable, the preset coefficient threshold is 0.85, and for emerging industries such as the Internet, the preset coefficient threshold is 0.8, in order to balance the model's adaptability and anti-disturbance capability to the rapidly changing market environment.

[0067] In this embodiment, a graph neural network model is obtained by combining hybrid task learning with adversarial training. This model not only has a powerful feature extraction capability, but also maintains the stability of the output when faced with data noise or small perturbations, providing reliable model support for the quantitative scoring of industrial competitiveness.

[0068] In some embodiments of this application, the competitive feature vector of the target node is obtained, including: The real-time heterogeneous graph is input into the graph neural network model. The node-level heterogeneous attention layer based on the model obtains the attribute features of the target node, the relationship-aware message passing layer based on the model obtains the structural association features of the target node, and the temporal feature fusion layer based on the model obtains the temporal behavior features of the target node. The target node's attribute features, structural association features, and temporal behavior features are concatenated to form the target node's competitive feature vector.

[0069] In this embodiment, the node-level heterogeneous attention layer performs weighted aggregation on various attributes of the target node (such as basic attributes like enterprise revenue, number of patents, and employee size, as well as dynamically updated real-time data attributes) to obtain attribute features. The weights are dynamically adjusted according to the importance of the attributes to competitiveness assessment.

[0070] In this embodiment, the relationship-aware message passing layer collects the feature information of the target node's directly and indirectly connected nodes based on the topological position of the target node in the real-time heterogeneous graph, and performs differentiated message passing and fusion according to the edge type (such as cooperative relationship, competitive relationship, supply chain relationship, etc.) and weight to obtain the structural association feature vector.

[0071] In this embodiment, the temporal feature fusion layer uses a long short-term memory network (LSTM) to model the dynamic behavioral feature sequence of the target node under multiple time windows (such as the quarterly changes in revenue growth rate, the annual fluctuations in R&D investment, the monthly evolution of market share, etc.), capture its long-term development trend and short-term sudden change pattern, and obtain the temporal behavioral feature vector.

[0072] In this embodiment, based on the hybrid task learning objective of the graph neural network model, the feature weights are dynamically allocated during the splicing process. The attribute features, structural association features, and temporal behavior features are weighted and spliced ​​together. The resulting competitive feature vector can comprehensively integrate the static attributes, network structure position, and dynamic behavior patterns of the target node, providing multi-dimensional and in-depth feature support for subsequent quantitative scoring.

[0073] In some embodiments of this application, a scoring mapping system is constructed, and the competitiveness feature vector of the target node is mapped to a quantitative scoring result based on the scoring mapping system, including: Based on the multi-dimensional needs of industrial competitiveness assessment, a scoring mapping system is constructed, which includes a basic scoring mapping table, a dynamic weighting mechanism, and quantitative scoring calibration. The competitive feature vector of the target node is input into the scoring mapping system. The quantitative scoring result of the target node is output through three steps: looking up the basic scoring mapping table, dynamically adjusting the feature weights, and calibrating the quantitative score.

[0074] In this embodiment, by using historical competitiveness feature vectors and corresponding actual scoring data, the quantile regression method is employed to determine the baseline threshold range for different feature dimensions in different competitiveness level intervals, and a basic scoring mapping table is established.

[0075] In this embodiment, a dynamic weighting mechanism is introduced to adjust the weights of attribute features, structural correlation features, and temporal behavior features in real time based on the characteristics of the industry to which the target node belongs and the current economic cycle stage.

[0076] In this embodiment, an initial quantitative score is generated based on a basic scoring mapping table. The calibration of the quantitative score involves introducing an industry environment adjustment factor and outlier correction rules to further optimize the scoring results. The industry environment adjustment factor comprehensively considers the impact of macroeconomic indicators (such as GDP growth rate and industry prosperity index), policy and regulatory changes (such as tax policy adjustments and changes in industry entry barriers), and sudden public events (such as pandemics and supply chain disruptions) on the overall industry competitiveness assessment standards. By establishing a non-linear mapping relationship between the adjustment factor and the score, the initial quantitative score is dynamically calibrated. The outlier correction rules target extreme feature values ​​at target nodes, using Z-score standardization to identify outliers. Based on the degree of anomaly, the contribution of the corresponding feature dimension to the score is attenuated or enhanced to avoid excessive interference from a single extreme feature on the overall score. Finally, a multi-dimensional calibrated quantitative score result that conforms to the actual industry competitive landscape is output.

[0077] In this embodiment, the quantitative scoring results are in percentage form, with a value range of 0-100. The higher the score, the stronger the industrial competitiveness of the target node. It also supports outputting sub-item scores by feature dimension, making it easy for users to intuitively understand the competitive advantages and disadvantages.

[0078] like Figure 2As shown, the electronic device may include a processor 201, a communications interface 202, a memory 203, and a communication bus 204. The processor 201, communications interface 202, and memory 203 communicate with each other via the communication bus 204. The processor 201 can call logical instructions from the memory 203 to execute a quantitative scoring method for industrial competitiveness based on graph neural networks and data fusion.

[0079] Furthermore, the logical instructions in the aforementioned memory 203 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0080] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the industrial competitiveness quantitative scoring method based on graph neural networks and data fusion provided by the above methods.

[0081] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the industrial competitiveness quantitative scoring method based on graph neural networks and data fusion provided by the above methods.

[0082] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0083] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A quantitative scoring method for industrial competitiveness based on graph neural networks and data fusion, characterized in that, include: Construct a heterogeneous graph; Real-time data is collected and updated to the heterogeneous graph to obtain a real-time heterogeneous graph. The real-time heterogeneous graph is then trained and its representation is learned based on a graph neural network model to obtain the competitive feature vector of the target node. A scoring mapping system is constructed, and the competitiveness feature vector of the target node is mapped to a quantitative scoring result based on the scoring mapping system.

2. The method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion as described in claim 1, characterized in that, Constructing a heterogeneous graph includes: Define the node type system, edge type system, and attribute patterns of heterogeneous graphs; Build an industry database; Entities and relationships are extracted from the industry database. Based on the extraction results and definition results, graph nodes and edges are created in batches. Each graph node has a corresponding unique entity. Populate each graph node with basic attributes obtained from the data source; The initial weight of each edge with a relationship is calculated based on the strength of the relational evidence and the multi-source calculation. Construct a heterogeneous graph based on several graph nodes, edges, basic attributes, and initial weights.

3. The method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion as described in claim 2, characterized in that, Also includes: Several quality assessment indicators are pre-defined, including pattern assessment indicators, data assessment indicators, logic assessment indicators, and structural assessment indicators. Based on quality assessment metrics, collect the current heterogeneous graph and a validation dataset that matches each assessment metric; Generate quality assessment values ​​for the corresponding quality assessment indicators based on the validation dataset; The overall quality assessment value of the current heterogeneous graph is obtained by weighting the quality assessment values ​​of all quality assessment indicators and the corresponding weight coefficients of the indicators. Determine whether to correct the heterogeneous graph based on the comprehensive quality assessment value. If so, identify the abnormal graph data and generate the corresponding correction instructions based on the preset correction strategy library. The corrected heterogeneous graph is re-evaluated until the overall quality evaluation value is greater than the preset threshold. Then the corrected heterogeneous graph is replaced with the current heterogeneous graph.

4. The method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion as described in claim 3, characterized in that, Collect real-time data and update it to the heterogeneous graph to obtain a real-time heterogeneous graph, including: Real-time data on enterprises, industries, and market environment is collected, and the real-time data is preprocessed, including format parsing, outlier filtering, and data standardization. Entity recognition and event extraction are performed on the preprocessed real-time data, and the extracted entities are matched with the nodes in the current heterogeneous graph in multiple dimensions to obtain the matching degree. Based on the matching degree, the associated nodes and associated edges of each extracted entity are obtained, and an update data packet containing the node attribute update of the corresponding associated node, the edge weight adjustment or addition of the corresponding associated edge, and the newly added edge is generated. The update data packet is submitted to the graph database to perform basic data updates and trigger a cascading update process of the nodes and edges associated with the updated content, resulting in a real-time heterogeneous graph.

5. The method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion as described in claim 4, characterized in that, The matching degree includes: One randomly selected entity is chosen as the target entity; Extract the feature attributes of the target entity and compare them with the corresponding attributes of each node in the current heterogeneous graph to obtain the attribute matching degree; Set the topological distance between the target entity and each node, and extract the corresponding node's comparative network structure diagram according to the topological distance; The target entity is placed in its local relation, and several network structure diagrams of the target entity are generated by combining the comparison network structure diagram of each node. The structure matching degree is obtained by comparing each of these diagrams with the corresponding comparison network structure diagram. Extract several behavioral features of the target entity and construct a sequence of behavioral features of the target entity; Set the number of comparison features between the target entity and each node, filter out several behavioral features to be compared with each node by combining the behavioral feature sequence, and compare them with the behavioral features of the corresponding node at the same time node to obtain the time behavior matching degree. The matching degree between the target entity and the corresponding node is generated by comparing the attribute matching degree, structural matching degree, and temporal behavior matching degree between the target entity and the same node. The matching degree between each extracted entity and each node is generated sequentially.

6. The method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion as described in claim 1, characterized in that, The graph neural network model includes: An initial heterogeneous graph neural network model is constructed, which includes a node-level heterogeneous attention layer, a relationship-aware message passing layer, and a temporal feature fusion layer. The initial heterogeneous graph neural network model is trained using a hybrid task learning objective, which includes a self-supervised learning task and a supervised learning task. The self-supervised learning task includes several preset training tasks, and the supervised learning task includes learning discriminative features related to competitiveness. Adversarial samples are introduced during training, and the corresponding model's robustness coefficient is calculated. When the training result accuracy is greater than a preset accuracy threshold and the robustness coefficient is greater than a preset coefficient threshold, the trained model is set as a graph neural network model.

7. The method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion as described in claim 6, characterized in that, The competitive feature vector of the target node is obtained, including: The real-time heterogeneous graph is input into the graph neural network model. The node-level heterogeneous attention layer based on the model obtains the attribute features of the target node, the relationship-aware message passing layer based on the model obtains the structural association features of the target node, and the temporal feature fusion layer based on the model obtains the temporal behavior features of the target node. The target node's attribute features, structural association features, and temporal behavior features are concatenated to form the target node's competitive feature vector.

8. The method for quantitatively scoring industrial competitiveness based on graph neural networks and data fusion as described in claim 7, characterized in that, Construct a scoring mapping system, and map the competitiveness feature vectors of target nodes to quantitative scoring results based on the scoring mapping system, including: Based on the multi-dimensional needs of industrial competitiveness assessment, a scoring mapping system is constructed, which includes a basic scoring mapping table, a dynamic weighting mechanism, and quantitative scoring calibration. The competitive feature vector of the target node is input into the scoring mapping system. The quantitative scoring result of the target node is output through three steps: looking up the basic scoring mapping table, dynamically adjusting the feature weights, and calibrating the quantitative score.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the quantitative scoring method for industrial competitiveness based on graph neural networks and data fusion as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the quantitative scoring method for industrial competitiveness based on graph neural networks and data fusion as described in any one of claims 1 to 8.

Images