Method and system for identifying industrial field of enterprise based on industrial chain

Abstract

This application relates to the field of industrial chain technology, providing a method and system for identifying the industry sector to which a company belongs based on an industrial chain, thus solving the problems of low accuracy and poor adaptability in existing technologies for identifying the industry sector in an industrial chain environment. The method includes: first, receiving multi-source data on new companies and related companies, processing it in parallel using a block model to form a set of company features, and mapping this set to an industry association rule base to generate a company knowledge graph; then, extracting the new company feature set and matching it with the graph to obtain a set of candidate industrial chain nodes, determining the location of new nodes based on the topological connection density between nodes, and constructing a dynamic industrial chain; finally, accurately identifying the industry sector to which the new company belongs by calculating the feature similarity between the new node and the core nodes of the sector, combined with the business flow intensity. This application improves the accuracy and adaptability of identifying the industry sector in an industrial chain environment.

Description

Technical Field

[0001] This application relates to the field of supply chain technology, and in particular to a method and system for determining the industry sector to which an enterprise belongs based on the supply chain. Background Technology

[0002] In the context of collaborative development across the industrial chain, it is necessary to accurately identify the industry sector affiliation of enterprises to support industrial strategy formulation and supply chain optimization. Existing technological needs are mainly reflected in three aspects: first, the need to process multi-source, heterogeneous enterprise data, including technology patents, product information, and supply chain relationships; second, the need to establish a dynamically updated industry knowledge system to reflect real-time changes in the industrial chain; and third, the need to quantify the strength of industrial linkages between enterprises to avoid biases caused by subjective judgments.

[0003] Currently, a representative solution for this requirement is an industry classification system based on multidimensional feature vector matching. This system first constructs a standard industry feature library, then uses feature engineering methods to extract the technological direction and product type features of enterprises. It calculates the feature similarity between enterprises and each industry category using a vector space model, and then combines this with the business relationship network between enterprises, using network node centrality indicators to assist in the judgment. Finally, it determines the industry classification based on the weighted scoring results. This system employs a periodic batch update mechanism, supporting automated processing of large-scale enterprise data.

[0004] This scheme has certain limitations in practical applications. The standard industry feature library relies on manual maintenance for updates, resulting in relatively fixed feature dimensions. The vector space model has limited ability to represent the characteristics of technological products, making it difficult to capture the complex relationships within the industry chain. Furthermore, the network node centrality index primarily reflects static topological characteristics, and its correspondence with the actual operational status of the industry is insufficient. These factors affect the accuracy and practicality of industry-specific judgments. Summary of the Invention

[0005] This application provides a method and system for identifying the industry sector to which an enterprise belongs based on the industrial chain, in order to solve the problems of low accuracy and poor adaptability in identifying the industry sector of an enterprise in the industrial chain environment in the prior art.

[0006] To address the aforementioned technical problems, firstly, this application provides a method for determining the industry sector to which an enterprise belongs based on the industrial chain, including:

[0007] Receive application data from new enterprises, as well as multi-source data from target enterprises and related enterprises;

[0008] The multi-source data is processed in parallel using a block model to form an enterprise feature set. The enterprise feature set is then mapped and associated with a preset industry association rule base to generate an enterprise knowledge graph. The enterprise feature set includes a technology feature set and a product feature set.

[0009] Extract the new enterprise feature set from the new enterprise application data, match the new enterprise feature set with the enterprise knowledge graph, and generate a candidate industrial chain node set based on the matching results;

[0010] Calculate the topological connection density between each pair of nodes in the candidate industry chain node set, select the node pair with the highest topological connection density from the candidate industry chain node set as the anchor node pair, extract features from the technology association edges and product circulation edges of the anchor node pair using a graph neural network to generate enhanced edge vectors, generate an offset coefficient for the new node addition position based on the enhanced edge vectors, and add the new enterprise as the new enterprise node. In the enterprise knowledge graph, the coordinates corresponding to the offset coefficient are used as the addition position, and the new enterprise node is added. Use a relational reasoning model to predict the potential interaction relationship between the new enterprise node and the anchor node pair to generate new technology association edges and new product circulation edges. Based on the new technology association edges and the new product circulation edges, generate a dynamic industry chain.

[0011] The feature similarity between new enterprise nodes and core nodes in the dynamic industrial chain is calculated in real time. Based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industrial chain node set, the industrial sector to which the new enterprise belongs is determined.

[0012] Optionally, the step of extracting features from the technical association edges and product circulation edges of the anchor node pairs using a graph neural network to generate enhanced edge vectors, and generating offset coefficients for the new node addition positions based on the enhanced edge vectors, includes:

[0013] The technical association edge of the anchor node pair is input into the first edge processing channel of the graph neural network, and the technical association feature vector is extracted from the technical association edge of the anchor node pair through the first edge processing channel.

[0014] The product flow edge of the anchor node pair is input into the second edge processing channel of the graph neural network, and the product flow feature vector is extracted from the product flow edge of the anchor node pair through the second edge processing channel.

[0015] The cross-channel attention fusion module of the graph neural network is used to weight and concatenate the technology association feature vector and the product circulation feature vector to generate an enhanced side vector.

[0016] The enhanced edge vector is input into the industrial chain spatial projection layer of the graph neural network, and the projection components of the technology-intensive direction and the circulation hub direction are output.

[0017] The directional fusion factor is calculated based on the ratio of the projection component of the technology-intensive direction to the projection component of the circulation hub direction.

[0018] Based on the direction fusion factor, the offset coefficient corresponding to the new node addition position is found in the preset offset mapping table.

[0019] Optionally, the cross-channel attention fusion module of the graph neural network performs weighted concatenation of the technology association feature vector and the product circulation feature vector to generate enhanced side vectors, including:

[0020] Based on the semantic diffusion path depth of the aforementioned technology-related feature vectors and the coverage of patent technology keywords, the dynamic impact factor of the technology is calculated.

[0021] Based on the number of logistics trajectory jumps and the spatiotemporal aggregation radius of the product circulation characteristic vector, calculate the circulation dynamic impact factor;

[0022] The technology dynamics impact factor and the circulation dynamics impact factor are normalized respectively to obtain the technology attention weight value and the circulation attention weight value.

[0023] Based on the technology attention weight value and the circulation attention weight value, the technology association feature vector and the product circulation feature vector are weighted and scaled respectively to generate weighted technology feature blocks and weighted circulation feature blocks;

[0024] A feature pyramid network is used to perform multi-scale feature fusion on the weighted technical feature block and the weighted flow feature block to generate a fused feature matrix.

[0025] The fused feature matrix is ​​subjected to dimensionality reduction and compression to output enhanced edge vectors.

[0026] Optionally, the step of using a feature pyramid network to perform multi-scale feature fusion on the weighted technical feature block and the weighted flow feature block to generate a fused feature matrix includes:

[0027] Construct a feature pyramid network comprising a first pyramid layer, a second pyramid layer, and a third pyramid layer;

[0028] The weighted technical feature block and the weighted circulation feature block are input into the feature pyramid network, and the output includes a technical multi-scale feature group and a circulation multi-scale feature group with three scales.

[0029] Cross-modal feature fusion is performed on the technical multi-scale feature group and the circulation multi-scale feature group, wherein the first pyramid layer adopts the element-wise addition method for fusion, the second pyramid layer adopts the channel attention weighted fusion, and the third pyramid layer adopts the spatial attention weighted fusion.

[0030] The fused multi-scale feature groups are upsampled to the original resolution to generate multi-scale fused features.

[0031] Based on the industry type configuration parameters, the multi-scale fusion features are rearranged into a two-dimensional matrix structure according to a preset dimension ratio to form a fusion feature matrix.

[0032] Optionally, the step of processing the multi-source data in parallel using a block model to form an enterprise feature set, mapping and associating the enterprise feature set with a preset industry association rule base, and generating an enterprise knowledge graph includes:

[0033] The multi-source data is segmented into benchmark technology feature blocks and benchmark product feature blocks using a block model;

[0034] Multiple patent technology keywords are extracted from the benchmark technology feature block to form a technology feature set, and multiple supply chain tags are extracted from the benchmark product feature block to form a product feature set;

[0035] Extract predefined industrial chain technology attribute sets and product attribute sets from the industrial association rule base, wherein the industrial chain technology attribute set contains key technology entries corresponding to each industrial field, and the product attribute set contains product circulation paths corresponding to each industrial field.

[0036] The technical feature set is matched with the key technical items for dependency relationship matching, and the semantic correlation between each patent technical keyword and the corresponding key technical item is calculated based on the matching results;

[0037] The product feature set is associated with the product distribution path, and the topological overlap between each supply chain tag and the corresponding product distribution path is calculated based on the association results.

[0038] When there are key technology items with semantic relevance greater than a preset relevance threshold and product circulation paths with topological overlap greater than a preset overlap threshold, technical relevance edges and product circulation edges are generated between enterprise entities and industry chain nodes.

[0039] Based on the technology-related edges and the product circulation edges, an enterprise knowledge graph is formed.

[0040] Optionally, the step of matching the new enterprise feature set with the enterprise knowledge graph and generating a candidate industry chain node set based on the matching result includes:

[0041] Separate the set of technology features to be matched and the set of product features to be matched from the new enterprise feature set;

[0042] The set of technical features to be matched, the set of product features to be matched, and the enterprise knowledge graph are input into a preset multimodal matching network. Through the technical matching sub-network of the multimodal matching network, the dependency strength between the set of technical features to be matched and the technical attributes of the nodes in the enterprise knowledge graph is calculated.

[0043] Through the product matching sub-network of the multimodal matching network, the overlap of circulation relationships between the feature set of the products to be matched and the product attributes of the nodes in the enterprise knowledge graph is calculated.

[0044] Multiple candidate industry chain nodes with dependency strength greater than a preset strength threshold and circulation overlap greater than a preset circulation threshold are selected from the enterprise knowledge graph.

[0045] Aggregate all candidate industry chain nodes to form a candidate industry chain node set.

[0046] Optionally, the real-time calculation of the feature similarity between new enterprise nodes and core nodes in the dynamic industry chain, and the determination of the industry sector to which the new enterprise belongs based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industry chain node set, includes:

[0047] Calculate the similarity of technical features between the technical feature set of the new enterprise node and the technical attribute set of the domain core node;

[0048] Calculate the product feature similarity between the product feature set of the new enterprise node and the product attribute set of the domain core node;

[0049] Calculate the business flow density value between the new enterprise node and the associated nodes in the candidate industry chain node set;

[0050] The technical feature similarity and the business flow density value are first weighted and fused using an attention mechanism, and the product feature similarity and the business flow density value are second weighted and fused.

[0051] When the first weighted fusion result is greater than a preset first threshold and the second weighted fusion result is greater than a preset second threshold, the new enterprise is determined to belong to the industry sector corresponding to the core node of the field.

[0052] Secondly, this application provides a system for identifying the industry sector to which a company belongs based on the industrial chain, including:

[0053] The receiving module is used to receive application data from new enterprises, as well as multi-source data from target enterprises and related enterprises;

[0054] The association module is used to process the multi-source data in parallel through a block model to form an enterprise feature set, and to map and associate the enterprise feature set with a preset industry association rule base to generate an enterprise knowledge graph. The enterprise feature set includes a technology feature set and a product feature set.

[0055] The extraction module is used to extract the feature set of new enterprises from the new enterprise application data, match the feature set of new enterprises with the enterprise knowledge graph, and generate a candidate industrial chain node set based on the matching result;

[0056] The selection module is used to calculate the topological connection density between each pair of nodes in the candidate industry chain node set, select the node pair with the highest topological connection density from the candidate industry chain node set as the anchor node pair, extract features from the technology association edges and product circulation edges of the anchor node pair through a graph neural network to generate enhanced edge vectors, generate an offset coefficient for the new node addition position based on the enhanced edge vectors, and add the new enterprise as the new enterprise node. In the enterprise knowledge graph, the coordinates corresponding to the offset coefficient are used as the addition position, and the new enterprise node is added. The potential interaction relationship between the new enterprise node and the anchor node pair is predicted using a relational reasoning model to generate new technology association edges and new product circulation edges. Based on the new technology association edges and the new product circulation edges, a dynamic industry chain is generated.

[0057] The calculation module is used to calculate the feature similarity between new enterprise nodes and core nodes in the dynamic industrial chain in real time. Based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industrial chain node set, the industrial sector to which the new enterprise belongs is determined.

[0058] Thirdly, this application provides an electronic device, comprising:

[0059] Memory, used to store computer programs;

[0060] A processor, used to execute the computer program to implement the steps of the industry sector identification method based on the supply chain as described in the first aspect above.

[0061] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the method for determining the industry sector to which an enterprise belongs based on the industrial chain as described in the first aspect above.

[0062] This application provides a method for determining the industry sector to which an enterprise belongs based on an industry chain. The method includes: receiving application data from new enterprises, and multi-source data from target enterprises and related enterprises; performing parallel processing on the multi-source data using a block model to form an enterprise feature set; mapping and associating the enterprise feature set with a preset industry association rule base to generate an enterprise knowledge graph, wherein the enterprise feature set includes a technology feature set and a product feature set; extracting the new enterprise feature set from the new enterprise application data; matching the new enterprise feature set with the enterprise knowledge graph; generating a candidate industry chain node set based on the matching result; calculating the topological connection density between each pair of nodes in the candidate industry chain node set; selecting the node pair with the highest topological connection density from the candidate industry chain node set as an anchor node pair; and using a graph neural network to process the anchor node pair... Feature extraction is performed on the technology association edges and product circulation edges of the node pairs to generate enhanced edge vectors. Based on the enhanced edge vectors, an offset coefficient is generated for the new node addition position. The new enterprise is then designated as a new enterprise node. In the enterprise knowledge graph, the coordinates corresponding to the offset coefficient are used as the addition position, and the new enterprise node is added. A relational reasoning model is used to predict the potential interaction relationship between the new enterprise node and the anchor node pair, generating new technology association edges and new product circulation edges. Based on the new technology association edges and the new product circulation edges, a dynamic industrial chain is generated. The feature similarity between the new enterprise node in the dynamic industrial chain and the core nodes of the domain is calculated in real time. Based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industrial chain node set, the industry domain to which the new enterprise belongs is determined.

[0063] The technical solution provided in this application has the following beneficial effects:

[0064] This application achieves unified access to multi-source heterogeneous data, providing a complete data foundation for subsequent processing. It improves data processing efficiency, ensuring the effective extraction of technical and product characteristics. It constructs a structured industry knowledge system, intuitively showcasing the technological connections and product circulation relationships between enterprises. It quickly identifies the industry chain nodes with the highest relevance to new enterprises, narrowing the scope of judgment. It enables precise positioning of new enterprises within the industry chain, ensuring the rationality of the industry chain structure. By integrating static features and dynamic business data, it improves the accuracy and reliability of judgment.

[0065] Furthermore, this application also extracts technology association feature vectors and product circulation feature vectors through dual-channel processing of graph neural networks, generates enhanced edge vectors using cross-channel attention fusion modules, obtains projection components of technology-intensive direction and circulation hub direction through industrial chain spatial projection, calculates direction fusion factor and then queries offset mapping table to obtain offset coefficient of new node addition position.

[0066] Furthermore, this method achieves deep extraction and intelligent fusion of technical and circulation characteristics through graph neural networks, and then obtains accurate offset coefficients through spatial projection and direction factor calculation, ensuring that the addition position of new nodes in the industrial chain not only conforms to the technically intensive characteristics but also adapts to the distribution of circulation hubs, effectively improving the accuracy and rationality of new enterprises embedding into the industrial chain. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 A flowchart illustrating a method for determining the industry sector to which an enterprise belongs based on the industrial chain, provided in this application embodiment;

[0069] Figure 2 A schematic diagram illustrating a specific implementation of a method for determining the industry sector to which an enterprise belongs based on the industrial chain, provided in this application embodiment;

[0070] Figure 3 This is a schematic diagram of a system for identifying the industry sector of an enterprise based on the industrial chain, provided in an embodiment of this application. Detailed Implementation

[0071] While multi-dimensional feature matching-based industry classification systems can automate the current industry chain enterprise identification technology, they still have significant limitations. The standard industry feature database relies on manual maintenance for updates, making it difficult to adapt to rapidly changing industry environments. Vector space models have limited capacity to represent technological product features, failing to fully capture the complex relationships within the industry chain. Network node centrality indicators primarily reflect static topological characteristics, which differ from the dynamic characteristics of actual industry operations. These issues challenge the accuracy and practicality of existing methods in industry chain identification.

[0072] To address the aforementioned issues, this application proposes a method for identifying the industry sector of a company based on its industry chain. This method uses a block model to process multi-source company data in parallel, constructing a company knowledge graph that integrates technological and product characteristics. It employs topological connection density analysis to determine the optimal position of a new company within the industry chain and dynamically updates the industry chain structure by combining graph neural networks and relational reasoning models. Finally, it achieves accurate identification through a comprehensive calculation of feature similarity and business flow intensity. This method overcomes the limitations of traditional static feature matching. Through dynamic industry chain construction and real-time business flow analysis, it effectively solves the problems of lagging industry feature updates, insufficient capture of complex relationships, and lack of dynamic adaptability in existing technologies, thus improving the accuracy and practicality of industry sector identification.

[0073] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0074] The core of this application is to provide a method for determining the industry sector to which a company belongs based on the industrial chain. A flowchart of one specific implementation method is shown below. Figure 1 As shown, the method includes:

[0075] Step 101: Receive the application data of the new enterprise, as well as multi-source data of the target enterprise and related enterprises.

[0076] In step 101, the data submitted by new enterprises refers to the basic information submitted by newly registered or unclassified enterprises, including enterprise name, registered address, business scope, patent list, product catalog, etc. The target enterprise refers to the core enterprise requiring industry sector identification. Related enterprises refer to enterprises that have business dealings, investment relationships, or technological cooperation with the target enterprise. Multi-source data refers to heterogeneous data from different channels, including business registration information, patent applications, supply chain records, financial reports, etc.

[0077] In this embodiment of the application, the application data submitted by the new enterprise is received through the data interface, and multi-source data of the target enterprise and related enterprises are obtained from public databases or enterprise cooperation platforms, including business registration information, patent texts, product lists and supply chain records. All data are stored in a temporary database after being standardized in format, providing a unified data source for subsequent processing.

[0078] For example, taking the new energy vehicle industry chain as an example, the system receives the application data of new battery company D, including the company name, registered location, three lithium battery patents and two motor supply contracts; at the same time, it obtains the business registration information, five patents and annual product list of the target company battery manufacturer A, as well as the supply chain data and list of cooperative companies of the related company motor supplier B. All data is stored in a temporary database after being converted into a new format.

[0079] Step 102: The multi-source data is processed in parallel using a block model to form an enterprise feature set. The enterprise feature set is then mapped and associated with a preset industry association rule base to generate an enterprise knowledge graph. The enterprise feature set includes a technology feature set and a product feature set.

[0080] In step 102, the block model represents a data parallel processing architecture that divides data into independent modules based on type for parallel processing. The enterprise feature set includes a technology feature set (patent keywords, technical fields) and a product feature set (product type, supply chain tags). The industry association rule base represents a predefined set of industry chain rules, including technological dependencies and product circulation paths. The enterprise knowledge graph represents graph-structured data, where nodes represent enterprises or industry sectors, and edges represent technological associations or product circulation relationships.

[0081] In this embodiment, the block model divides multi-source data into technical data blocks and product data blocks. The technical data blocks extract patent keywords to form a technical feature set, and the product data blocks extract supply chain tags to form a product feature set. The technical feature set is matched with technical entries in the industry association rule base to calculate semantic relevance, and the product feature set is matched with product circulation paths to calculate topological overlap. When both semantic relevance and topological overlap exceed the threshold, technical association edges and product circulation edges between enterprise nodes and industry nodes are generated in the enterprise knowledge graph.

[0082] For example, the technical data block of battery manufacturer A extracts patent keywords such as "lithium-ion battery" and "energy density," while the product data block extracts supply chain tags such as "Yangtze River Delta delivery" and "48-hour delivery." The semantic correlation between the technical feature set and the rule base for "battery energy density ≥300Wh / kg" is 0.92, and the topological overlap between the product feature set and the "battery-motor-vehicle manufacturer" path is 0.85. Since both values ​​exceed the threshold of 0.8, technical correlation edges and product circulation edges between A and new energy vehicle industry nodes are generated in the knowledge graph.

[0083] Step 103: Extract the new enterprise feature set from the new enterprise application data, match the new enterprise feature set with the enterprise knowledge graph, and generate a candidate industrial chain node set based on the matching result.

[0084] In step 103, the new enterprise feature set represents the technical and product features extracted from the new enterprise data. The candidate industry chain node set represents the set of industry nodes that have a high degree of matching with the enterprise knowledge graph.

[0085] In this embodiment, technical keywords and supply chain tags are extracted from the new enterprise application data to form a new enterprise feature set. The technical keywords are compared with the technical attributes of nodes in the knowledge graph to obtain the technical matching degree, and the supply chain tags are compared with the product attributes of nodes to obtain the product matching degree. Nodes with both technical matching degree and product matching degree exceeding the threshold are selected and aggregated into a candidate industrial chain node set.

[0086] For example, new battery company D extracts the technical characteristics "high-nickel cathode" and "silicon-carbon anode" and the product characteristics "South China warehousing" and "72-hour delivery". The technical matching degree of D with battery manufacturer A node is calculated to be 0.88 and the product matching degree is 0.70, and the technical matching degree with vehicle manufacturer C node is 0.86 and the product matching degree is 0.78. A threshold of 0.8 is set. Node C's double matching degree exceeds the threshold and is selected into the candidate node set.

[0087] Step 104: Calculate the topological connection density between each pair of nodes in the candidate industry chain node set. Select the node pair with the highest topological connection density from the candidate industry chain node set as the anchor node pair. Extract features from the technology association edges and product circulation edges of the anchor node pair using a graph neural network to generate enhanced edge vectors. Based on the enhanced edge vectors, generate the offset coefficient for the new node addition position and add the new enterprise as the new enterprise node. In the enterprise knowledge graph, use the coordinates corresponding to the offset coefficient as the addition position and add the new enterprise node. Use a relational reasoning model to predict the potential interaction relationship between the new enterprise node and the anchor node pair to generate new technology association edges and new product circulation edges. Based on the new technology association edges and the new product circulation edges, generate a dynamic industry chain.

[0088] In step 104, topological connectivity density represents the comprehensive strength index of technological associations and product circulation relationships between nodes. Anchor node pairs are the node pairs with the highest topological connectivity density in the candidate node set, serving as a reference benchmark for adding new enterprise nodes. Their selection is based on the strength of technological and product-related associations between nodes. Enhanced edge vectors are technology and product feature fusion vectors extracted through a graph neural network. Offset coefficients are position adjustment parameters calculated based on the enhanced edge vectors, used to determine the addition position of the new node relative to the anchor node pair. The addition position is the coordinate point in the enterprise knowledge graph determined by the offset coefficient. The new enterprise node is the newly added node entity at this coordinate point. The relational reasoning model is an algorithmic model used to predict potential relationships between nodes. New technology association edges and new product circulation edges are new edge relationships generated based on the prediction results. The dynamic industry chain represents the real-time updated industry chain structure, including newly added enterprise nodes and their associations.

[0089] In this embodiment, the topological connection density between each pair of nodes in the candidate node set is calculated, and the node pair with the highest density is selected as the anchor node pair. The technical association edge and product circulation edge features of the anchor node pair are extracted by graph neural network to generate enhanced edge vectors, and the offset coefficient of the new node addition position is calculated. The new enterprise node is added to the position corresponding to the offset coefficient in the knowledge graph, and the interaction relationship between the new node and the anchor node pair is predicted by relational reasoning model to generate new technology association edge and new product circulation edge, and the knowledge graph is updated to form a dynamic industrial chain.

[0090] For example, the candidate node pair with the highest topological connection density (joint patents × annual transaction volume = 8 × 100,000 = 800,000) is selected as the anchor node pair. The graph neural network extracts the technical association edge features (patent diffusion path depth 3) and product circulation edge features (logistics hop count 4), generates an enhanced edge vector, and calculates the offset coefficient +0.3. At the position where the connection between A and C is offset by 30%, the enterprise node D is added. The relationship reasoning model predicts that the technical association edge weight between D and A is 0.8 and the product circulation edge weight between D and C is 0.6, generating a dynamic industrial chain. Here is a specific example: In the new energy vehicle industry chain scenario, the candidate industry chain node set includes battery manufacturer A, motor supplier B, and vehicle manufacturer C. When calculating the topological connection density between each pair of nodes, battery manufacturer A and motor supplier B have 6 shared patents and an annual transaction volume of 100,000, resulting in a topological connection density of 6 × 100,000 = 600,000. Motor supplier B and vehicle manufacturer C have 4 shared patents and an annual transaction volume of 80,000, resulting in a topological connection density of 4 × 80,000 = 320,000. Battery manufacturer A and vehicle manufacturer C have 8 shared patents and an annual transaction volume of 100,000, resulting in a topological connection density of 100,000 = 600,000. With a quantity of 100,000 units and a topological connection density of 8 × 100,000 = 800,000, battery manufacturer A and vehicle manufacturer C, which have the highest topological connection density, are selected as the anchor node pair. The technical association edges of the anchor node pair are processed using a graph neural network. The patent technology keyword coverage is extracted to be 0.7 and the semantic diffusion path depth is 3, resulting in a technology dynamic impact factor of 0.7 × 3 = 2.1. The product circulation edges are processed to extract the circulation trajectory hop count to be 4 and the spatiotemporal aggregation radius to be 200 kilometers, resulting in a circulation dynamic impact factor of 4 ÷ 200 = 0.02. The two factors are then input into a normalization layer to obtain the technology attention weight value. The weight is approximately 0.89, and the circulation attention weight is 0.11. A weighted technical feature block is obtained by weighting the technology association feature vector by 0.89, and a weighted circulation feature block is obtained by weighting the product circulation feature vector by 0.11. These are then concatenated at a 7:3 ratio and the resulting enhanced edge vector is output after dimensionality reduction. This vector is input into the industry chain spatial projection layer to obtain a technology-intensive direction projection component of 0.85 and a circulation hub direction projection component of 0.15. The direction fusion factor is calculated to be 0.85 ÷ 0.15 ≈ 5.67, and the offset coefficient is determined to be +0.3 based on the offset mapping table. A new battery company node D is added at a coordinate offset of 30% from the line connecting battery manufacturer A and vehicle manufacturer C. The potential interaction relationship between company D and the anchor node pair is analyzed using a relational reasoning model. The predicted weight of the technology association edge with battery manufacturer A is 0.8, representing the probability of technology cooperation, and the weight of the product circulation edge with vehicle manufacturer C is 0.6, representing the probability of product supply. Based on the generated new technology association edge and new product circulation edge, the enterprise knowledge graph is updated to form a dynamic industry chain, completing the precise embedding of the new enterprise in the industry chain.

[0091] Step 105: Calculate the feature similarity between the new enterprise node and the core node of the domain in the dynamic industrial chain in real time. Based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industrial chain node set, determine the industrial domain to which the new enterprise belongs.

[0092] In step 105, the domain core node represents the central node in the knowledge graph that represents a specific industry domain. Business flow intensity represents the density of business interactions between new enterprises and nodes associated with the candidate node set.

[0093] In this embodiment, the similarity of technical features between the technical feature set of the new enterprise node and the technical attribute set of the domain core node is calculated, and the similarity of product features between the product feature set and the product attribute set is calculated. The business flow density value of the new enterprise and the associated nodes in the candidate node set is calculated. The technical feature similarity and the business flow density value are weighted and fused, and the product feature similarity and the business flow density value are weighted and fused. When both fusion results exceed the threshold, the new enterprise is determined to belong to the industry domain.

[0094] For example, the technical feature similarity between company D and core node A in the domain is calculated to be 0.85 (cosine similarity) and the product feature similarity is 0.75 (Jaccard similarity); the business flow density value between D and candidate node C is calculated to be 0.8 (quarterly order volume / industry benchmark); the weighted fusion value of technical similarity and business flow density is 0.83 (weights 0.6 and 0.4), and the weighted fusion value of product similarity and business flow density is 0.775 (weights 0.5 and 0.5); both values ​​exceed the threshold of 0.7, so D is determined to belong to the new energy vehicle industry sector.

[0095] This method constructs an enterprise knowledge graph through parallel processing of multi-source data, uses topological connection density analysis to achieve accurate positioning of new enterprises in the industrial chain, and dynamically identifies the industry sector by combining feature similarity and business flow intensity, effectively improving the integrity of the industrial chain structure and the accuracy of industry sector identification, and adapting to the rapidly changing industrial environment.

[0096] To further improve the accuracy of adding new nodes, in some embodiments, step 104 involves: extracting features from the technical association edges and product circulation edges of the anchor node pairs using a graph neural network to generate enhanced edge vectors; and generating offset coefficients for the new node addition positions based on these enhanced edge vectors. Figure 2 As shown, it includes:

[0097] Step 201: Input the technical association edge of the anchor node pair into the first edge processing channel of the graph neural network, and extract the technical association feature vector from the technical association edge of the anchor node pair through the first edge processing channel.

[0098] In step 201, the technology association feature vector is a set of features extracted from the technology association edges that represent the characteristics of technology association, including data such as patent technology keyword coverage and semantic diffusion path depth.

[0099] In this embodiment, the technical association edges of the anchor node pairs are input into the first edge processing channel of the graph neural network. The distribution features of patented technical keywords on the edges are extracted through the convolutional neural network layer, and the technical diffusion path features are analyzed through the recurrent neural network layer. Finally, the technical association feature vector containing multi-dimensional technical characteristics is output.

[0100] Step 202: Input the product flow edge of the anchor node pair into the second edge processing channel of the graph neural network, and extract the product flow feature vector from the product flow edge of the anchor node pair through the second edge processing channel.

[0101] In step 202, the product circulation feature vector is a set of features representing the characteristics of product circulation extracted from the product circulation edge, including dimensional data such as the number of hops in the logistics trajectory and the spatiotemporal aggregation radius.

[0102] In this embodiment, the product circulation edge of the anchor node pair is input into the second edge processing channel of the graph neural network. The logistics trajectory features on the edge are extracted through the graph attention network layer, and the product circulation range features are analyzed through the spatiotemporal coding layer. Finally, the product circulation feature vector containing multi-dimensional circulation characteristics is output.

[0103] Step 203: Through the cross-channel attention fusion module of the graph neural network, the technology association feature vector and the product circulation feature vector are weighted and concatenated to generate an enhanced side vector.

[0104] In this embodiment, the attention weights of the associated feature vector and the product circulation feature vector are calculated by the cross-channel attention fusion module. The two feature vectors are weighted and summed according to the weights, and then the weighted feature vectors are concatenated to generate an enhanced side vector.

[0105] Step 204: Input the enhanced edge vector into the industrial chain spatial projection layer of the graph neural network, and output the projection components of the technology-intensive direction and the circulation hub direction.

[0106] In step 204, the projection components of the technology-intensive direction and the projection components of the logistics hub direction are the decomposition features of the enhanced edge vector in the industrial chain space, representing the degree of technology agglomeration and the degree of logistics concentration, respectively.

[0107] In this embodiment, the enhanced edge vector is input into the industrial chain spatial projection layer, and the projection values ​​of the vector in the technology-intensive direction and the circulation hub direction are calculated by a fully connected neural network, and the feature components in the two directions are output.

[0108] Step 205: Calculate the direction fusion factor based on the ratio of the projection component of the technology-intensive direction to the projection component of the circulation hub direction.

[0109] In step 205, the directional fusion factor is the ratio of the projection component of the technology-intensive direction to the projection component of the circulation hub direction, reflecting the relative importance of the two dimensions of technology and logistics.

[0110] In this embodiment of the application, the ratio of the projection component of the technology-intensive direction to the projection component of the circulation hub direction is calculated to obtain the direction fusion factor, which is used to determine the addition direction tendency of the new node.

[0111] Step 206: Based on the direction fusion factor, find the offset coefficient of the corresponding new node addition position in the preset offset mapping table.

[0112] In this embodiment of the application, the corresponding offset coefficient is looked up in a preset offset mapping table based on the calculated direction fusion factor. The mapping table defines the position offset corresponding to different factor ranges.

[0113] Here is a specific example:

[0114] In the new energy vehicle industry chain scenario, battery manufacturer A and vehicle manufacturer C are used as anchor node pairs. The technology association edge is input into the first processing channel of a graph neural network to extract a technology association feature vector containing a patent technology keyword coverage of 0.7 and a semantic diffusion path depth of 3. The coverage is calculated by statistically analyzing the proportion of repeated keywords in the shared patents of A and C, and the path depth is determined by analyzing the number of levels traversed by the technology from A to C. The product circulation edge is input into the second processing channel to extract a product circulation feature vector containing a logistics trajectory hop count of 4 and a spatiotemporal aggregation radius of 200 kilometers. The hop count is obtained by statistically analyzing the number of warehousing and transportation links experienced by the product from A to C, and the aggregation radius is determined by calculating the average distance of all delivery paths from A to C. The technology attention weight and circulation attention weight are calculated through a cross-channel attention fusion module, specifically using the formula... Calculate the technical weights, where Indicates technical attention weight, The impact factor of technological dynamics is 2.1 × (0.7 × 3). The circulation dynamics impact factor is 0.02 (4 ÷ 200), resulting in... , The technology-related feature vector is weighted by 0.89 to obtain a weighted technology feature block, and the product circulation feature vector is weighted by 0.11 to obtain a weighted circulation feature block. These are then concatenated at a 7:3 ratio to reduce the dimensionality and output a 128-dimensional enhanced edge vector. This vector is input into the industry chain spatial projection layer, and a fully connected network is used to calculate the technology-intensive direction projection component (0.85) and the circulation hub direction projection component (0.15). The directional fusion factor is then calculated. ,in Indicates the direction fusion factor. For the projection component in the technology-intensive direction. This represents the projection component in the direction of the circulation hub. According to the preset offset mapping table, when R is greater than 5, the corresponding offset coefficient is +0.3, determining that the new node is offset by 30% in the direction of high technology density. The final offset coefficient +0.3 is used for calculating the location of the new node.

[0115] In this embodiment, enhanced edge vectors are generated by dual-channel feature extraction and attention fusion, and precise offset coefficients are calculated by spatial projection and orientation fusion. This achieves accurate quantification of the location for adding new nodes and improves the accuracy and adaptability of dynamic industrial chain construction.

[0116] To further improve the accuracy of feature fusion, in some embodiments, step 203: the weighted concatenation of the technology-related feature vector and the product circulation feature vector through the cross-channel attention fusion module of the graph neural network to generate an enhanced side vector includes:

[0117] Step 301: Calculate the dynamic impact factor of the technology based on the semantic diffusion path depth of the technology-related feature vector and the coverage of patent technology keywords.

[0118] In step 301, the semantic diffusion path depth refers to the number of transmission levels of patented technology in the industry chain, obtained by statistically analyzing the number of inter-enterprise transfers from R&D to application. For example, the path depth of battery technology from material R&D to cell design to system integration to vehicle application is 4. Patent technology keyword coverage refers to the distribution ratio of a patent keyword among related enterprises, obtained by statistically analyzing the percentage of enterprises using the keyword out of the total number of related enterprises in the industry chain. For example, if the keyword "high-nickel cathode" is used by 8 out of 10 battery companies, the coverage is 80%. The technology dynamic impact factor is a quantitative indicator measuring the importance of technological features, obtained by multiplying the semantic diffusion path depth in the technology association feature vector by the patent technology keyword coverage.

[0119] In this embodiment, two parameters, semantic diffusion path depth and patent technology keyword coverage, are obtained from the technology association feature vector. These two parameters are multiplied to calculate the technology dynamic influence factor. The larger the factor value, the more important the technology feature.

[0120] Step 302: Calculate the dynamic impact factor of circulation based on the number of hops in the logistics trajectory and the spatiotemporal aggregation radius of the product circulation feature vector.

[0121] In step 302, the number of hops in the logistics trajectory refers to the number of intermediate links in the product's flow from production to consumption. This is obtained by statistically analyzing the number of nodes such as warehousing, transportation, and quality inspection recorded in supply chain documents. For example, if a battery needs to pass through 3 warehouses and 2 transfers from the factory to vehicle assembly, the number of hops is 5. The spatiotemporal aggregation radius refers to the coverage area of ​​the product distribution network, obtained by calculating the average distance from regional warehousing centers to each distribution point. For example, if the average distance from a battery company's East China warehousing center to all 4S stores is 150 kilometers, then the aggregation radius is 150 kilometers. The circulation dynamic impact factor is a quantitative indicator that measures the importance of product circulation characteristics. It is obtained by dividing the number of hops in the logistics trajectory in the product circulation characteristic vector by the spatiotemporal aggregation radius.

[0122] In this embodiment, two parameters, the number of hops in the logistics trajectory and the spatiotemporal aggregation radius, are obtained from the product circulation feature vector. The number of hops in the logistics trajectory is divided by the spatiotemporal aggregation radius to calculate the circulation dynamic impact factor. The larger the factor value, the more important the circulation feature.

[0123] Step 303: Normalize the technology dynamics impact factor and the circulation dynamics impact factor to obtain the technology attention weight value and the circulation attention weight value.

[0124] In step 303, the technology attention weight value and the circulation attention weight value are weight coefficients obtained through normalization, which are used to measure the relative importance of technology and circulation characteristics.

[0125] In this embodiment, the technology dynamics impact factor and the circulation dynamics impact factor are input into the normalization processing layer, and two weight values ​​are calculated by the softmax function to ensure that the sum of the two weight values ​​is a fixed value.

[0126] Step 304: Based on the technology attention weight value and the circulation attention weight value, the technology association feature vector and the product circulation feature vector are weighted and scaled respectively to generate a weighted technology feature block and a weighted circulation feature block.

[0127] In step 304, the weighted technical feature block and the weighted flow feature block are feature blocks obtained by adjusting the weights of the original feature vectors respectively.

[0128] In this embodiment, the technology attention weight value is used to perform element-wise multiplication on the technology-related feature vector to generate a weighted technology feature block, and the circulation attention weight value is used to perform element-wise multiplication on the product circulation feature vector to generate a weighted circulation feature block.

[0129] Step 305: Use a feature pyramid network to perform multi-scale feature fusion on the weighted technical feature block and the weighted flow feature block to generate a fused feature matrix.

[0130] In step 305, the fused feature matrix is ​​a composite feature representation formed by splicing two weighted feature blocks according to a preset ratio.

[0131] In this embodiment, the weighted technical feature blocks and the weighted circulation feature blocks are spliced ​​together according to a pre-set dimensional ratio, with the technical feature blocks occupying the main proportion and the circulation feature blocks occupying the secondary proportion, to form a fused feature matrix.

[0132] Step 306: Perform dimensionality reduction and compression processing on the fused feature matrix to output enhanced edge vectors.

[0133] In this embodiment, principal component analysis is performed on the fused feature matrix to reduce its dimensionality, extract the most important feature components, and output an enhanced edge vector with a fixed dimension.

[0134] Here is a specific example:

[0135] In the new energy vehicle industry chain scenario, based on the technology association feature vector between battery manufacturer A and vehicle manufacturer C, which includes patent technology keyword coverage of 0.7 and semantic diffusion path depth of 3, the technology dynamic impact factor is calculated to be 0.7 × 3 = 2.1. The coverage is obtained by statistically analyzing the proportion of repeated keywords in the shared patents of A and C, and the path depth is determined by analyzing the three levels of R&D, production, and testing that the technology passes through from A to C. Based on the product circulation feature vector, which includes the number of logistics trajectory hops of 4 and the spatiotemporal aggregation radius of 200 kilometers, the circulation dynamic impact factor is calculated to be 4 ÷ 200 = 0.02. The number of hops is obtained by statistically analyzing the four stages of product circulation from A to C: factory warehousing, trunk transportation, regional distribution, and last-mile delivery. The aggregation radius is determined by calculating the average distance of all delivery paths from A to C, which is 200 kilometers. The technology dynamic impact factor of 2.1 and the circulation dynamic impact factor of 0.02 are normalized and then processed using the formula... Calculate the technical attention weight value to obtain Attention weight value The technology-related feature vectors are weighted and scaled using a technology attention weight of 0.89, multiplying each dimension value by 0.89 to generate weighted technology feature blocks. Similarly, the product circulation feature vectors are weighted and scaled using a circulation attention weight of 0.11 to generate weighted circulation feature blocks. These weighted technology and circulation feature blocks are then concatenated into a fusion feature matrix with a preset dimensionality ratio of 7:3, where technology features occupy 70% of the dimension and circulation features occupy 30%. Principal component analysis (PCA) is performed on the fusion feature matrix to reduce dimensionality, extracting the top 128 key feature components and outputting a 128-dimensional enhanced edge vector. This vector comprehensively retains the core information of both the technology and circulation features.

[0136] In this embodiment, the intelligent fusion of technology and circulation features is achieved through dynamic impact factor calculation and attention weight allocation. Then, enhanced edge vectors are generated by proportional splicing and dimensionality reduction compression, which effectively improves the accuracy and computational efficiency of feature representation and provides a reliable feature basis for subsequent offset coefficient calculation.

[0137] To further improve the effect of multi-scale feature fusion, in some embodiments, step 305: the multi-scale feature fusion of the weighted technical feature block and the weighted circulating feature block using a feature pyramid network to generate a fused feature matrix includes:

[0138] Step 401: Construct a feature pyramid network containing a first pyramid layer, a second pyramid layer, and a third pyramid layer.

[0139] In step 401, the feature pyramid network is a multi-scale feature extraction architecture. The first pyramid layer is a feature layer that maintains the original resolution, the second pyramid layer is a feature layer that has undergone one downsampling, and the third pyramid layer is a feature layer that has undergone two downsamplings.

[0140] In this embodiment, a feature pyramid network structure with three pyramid layers is constructed. The first pyramid layer directly processes the original resolution features. The second pyramid layer halves the feature size through convolution and downsampling operations. The third pyramid layer performs convolution and downsampling operations on the basis of the second layer to reduce the feature size to one-quarter of the original size.

[0141] Step 402: Input the weighted technical feature block and the weighted circulation feature block into the feature pyramid network, and output a technical multi-scale feature group with three scales and a circulation multi-scale feature group with three scales.

[0142] In step 402, the technology multi-scale feature group is a set of three different scales of technology features obtained after the weighted technology feature blocks are processed by the feature pyramid network, and the circulation multi-scale feature group is a set of three different scales of circulation features obtained after the weighted circulation feature blocks are processed by the feature pyramid network.

[0143] In this embodiment, weighted technical feature blocks are input into a feature pyramid network. Through convolution and downsampling operations at each level, a technical multi-scale feature set containing the original scale, half scale, and quarter scale is output. At the same time, weighted flow feature blocks are input into the same network structure to output flow multi-scale feature sets of the corresponding three scales.

[0144] Step 403: Perform cross-modal feature fusion on the technical multi-scale feature group and the circulation multi-scale feature group, wherein the first pyramid layer adopts the element-wise addition method for fusion, the second pyramid layer adopts the channel attention weighted fusion, and the third pyramid layer adopts the spatial attention weighted fusion.

[0145] In step 403, cross-modal feature fusion is the process of combining technical multi-scale feature groups and circulation multi-scale feature groups at different scales. Element-by-element addition is the direct addition of feature values ​​at corresponding positions. Channel attention weighting is the weighted fusion based on the importance of feature channels. Spatial attention weighting is the weighted fusion based on the importance of feature spatial positions.

[0146] In this embodiment, the technical multi-scale feature group and the circulation multi-scale feature group of the first pyramid layer are fused by element-wise addition. The feature group of the second pyramid layer is fused by weighted fusion after calculating the weight of each channel using a channel attention mechanism. The feature group of the third pyramid layer is fused by weighted fusion after calculating the weight of each position using a spatial attention mechanism.

[0147] Step 404: Unify the fused multi-scale feature groups to the original resolution through an upsampling operation to generate multi-scale fused features.

[0148] In step 404, the multi-scale fusion feature is a comprehensive feature representation obtained by unifying the fusion features of different scales to the original resolution.

[0149] In this embodiment, the features after the three pyramid layers are fused are upsampled to restore all features to their original resolution size, and then stitched together to generate multi-scale fused features containing multi-scale information.

[0150] Step 405: Based on the industry type configuration parameters, rearrange the multi-scale fusion features into a two-dimensional matrix structure according to a preset dimension ratio to form a fusion feature matrix.

[0151] In step 405, the industry type configuration parameter is a dimension ratio parameter set according to the characteristics of different industries.

[0152] In this embodiment, the corresponding configuration parameters are selected according to the industry type, and the multi-scale fusion features are reorganized and arranged according to the preset ratio of technical features and circulation features to form a standardized two-dimensional matrix structure as the fusion feature matrix.

[0153] Here is a specific example:

[0154] In the new energy vehicle industry chain scenario, based on the weighted technical feature block obtained from the aforementioned embodiment, which contains four dimensions of data [0.623, 0.623, 0.267, 2.67], and the weighted circulation feature block, which contains three dimensions of data [1.1, 0.44, 22], a feature pyramid network is constructed, consisting of a first pyramid layer maintaining the original resolution of 4×3, a second pyramid layer downsampled to 2×2, and a third pyramid layer downsampled to 1×1. The weighted technical feature block and the weighted circulation feature block are input into the feature pyramid network, and feature extraction is performed through convolution operations with a kernel size of 3×3. The first pyramid layer outputs a technical multi-scale feature group [0.62, 0.62, 0.27, 2.67] and a circulation multi-scale feature group [1.10, 0.44, 22.00]. The second pyramid layer outputs the downsampled feature group through max pooling with a stride of 2, and the third pyramid layer continues downsampling to output the coarsest-grained feature group. Cross-modal feature fusion is performed on the three pyramid layers respectively. The first pyramid layer uses an element-wise addition formula. Perform calculations, where The fused feature matrix represents the first... Line number Column elements, Representation of multi-scale feature groups Line number Column elements, Represents the multi-scale characteristic group of circulation. Line number The column elements are used to calculate fusion values ​​such as [1.72, 1.06, 22.27]. The second pyramid layer uses a channel attention-weighted fusion formula. ,in Indicates the first The fusion characteristics of each channel Indicates the technical feature number Attention weights for each channel, Indicates the technical feature number The value of each channel, Indicating circulation characteristics Attention weights for each channel, Indicating circulation characteristics The values ​​of each channel are used to calculate attention weights using the softmax function. The third pyramid layer employs a spatial attention weighted fusion formula. ,in Indicates position The fusion characteristics Indicates the technical features in location Attention weights Indicates the technical features in location The value, Indicates the circulation characteristics in location Attention weights Indicates the circulation characteristics in location The value of the feature group after the fusion of the three pyramid layers is uniformly restored to the original resolution of 4×3 through bilinear upsampling operation, generating a multi-scale fusion feature containing multi-scale information; based on the characteristics of the new energy vehicle industry as a technology-intensive industry, the ratio of the technical feature dimension is configured as 0.7 and the ratio of the circulation feature dimension is 0.3. The multi-scale fusion feature is rearranged into a two-dimensional matrix structure of 4 rows and 3 columns according to this ratio, in which the technical feature related data occupies the first 2 rows of the matrix and the circulation feature related data occupies the last 2 rows of the matrix, forming the final fusion feature matrix for subsequent processing.

[0155] In this embodiment, multi-scale feature extraction is achieved through a feature pyramid network, different fusion strategies are used to process features at different levels, and a fusion feature matrix is ​​generated through industry-adaptive dimensional configuration, which effectively improves the comprehensiveness and industry adaptability of feature fusion.

[0156] To further improve the accuracy and efficiency of enterprise knowledge graph construction, in some embodiments, step 102: The process of parallel processing of the multi-source data using a block model to form an enterprise feature set, and mapping and associating the enterprise feature set with a preset industry association rule base to generate an enterprise knowledge graph, includes:

[0157] Step 501: Divide the multi-source data into benchmark technology feature blocks and benchmark product feature blocks using a block model.

[0158] In step 501, the benchmark technical feature block and the benchmark product feature block are data units obtained by classifying multi-source data through a block model, and respectively contain technical-related data and product-related data.

[0159] In this embodiment, multi-source data is divided into technical feature blocks and product feature blocks according to data type using a block model. The technical feature blocks contain patent text and technical document data, while the product feature blocks contain supply chain records and product information data, thereby enabling parallel processing of data classification.

[0160] Step 502: Extract multiple patent technology keywords from the benchmark technology feature block to form a technology feature set, and extract multiple supply chain tags from the benchmark product feature block to form a product feature set.

[0161] In step 502, the technical feature set is a set of patent technology keywords extracted from the benchmark technical feature block. The product feature set is a set of supply chain tags extracted from the benchmark product feature block.

[0162] In this embodiment of the application, patent technology keywords are extracted from the technical feature block using text mining technology to form a technical feature set, and supply chain tags are extracted from the product feature block using pattern recognition to form a product feature set, providing a feature basis for subsequent association matching.

[0163] Step 503: Extract predefined industrial chain technology attribute sets and product attribute sets from the industrial association rule base, wherein the industrial chain technology attribute sets contain key technology entries corresponding to each industrial field, and the product attribute sets contain product circulation paths corresponding to each industrial field.

[0164] In step 503, the industrial chain technology attribute set is a predefined set of key technology entries for each industrial sector in the industrial association rule base. The product attribute set is a predefined set of product distribution paths.

[0165] In this embodiment of the application, a predefined set of technical attributes and a set of product attributes are extracted from the industry association rule base. The set of technical attributes includes the technical standards and requirements of various industry fields, and the set of product attributes includes typical product circulation and supply chain paths.

[0166] Step 504: Perform dependency relationship matching between the technical feature set and the key technical entries, and calculate the semantic correlation between each patent technical keyword and the corresponding key technical entry based on the matching results.

[0167] In step 504, the semantic relevance is the degree of semantic similarity between patent technology keywords and key technology entries, which is calculated using natural language processing technology.

[0168] In this embodiment of the application, each patent technology keyword in the technical feature set is semantically matched with the key technology item in the industrial chain technical attribute set, and the semantic similarity is calculated using a word vector model to obtain the degree of association between each keyword and the corresponding item.

[0169] Step 505: Associate the product feature set with the product distribution path, and calculate the topological overlap between each supply chain tag and the corresponding product distribution path based on the association results.

[0170] In step 505, the topological overlap is the degree of matching between the supply chain label and the product circulation path, which is calculated by the path similarity algorithm.

[0171] In this embodiment of the application, each supply chain tag in the product feature set is matched with the product circulation path in the product attribute set, and the degree of overlap between the business process described by the tag and the standard path is calculated to obtain the topological overlap.

[0172] Step 506: When there are key technology entries with semantic relevance greater than the preset relevance threshold and product circulation paths with topological overlap greater than the preset overlap threshold, generate technology relevance edges and product circulation edges between enterprise entities and industry chain nodes.

[0173] In step 506, an enterprise entity refers to a data node in the enterprise knowledge graph that represents a specific enterprise. It contains the enterprise's set of technical characteristics and set of product characteristics, and is connected to industry chain nodes through technical association edges and product circulation edges. An industry chain node refers to a data node in the enterprise knowledge graph that represents a specific industry sector. It contains a predefined set of industry chain technical attributes (key technical items) and a set of product attributes (typical circulation paths), used to characterize the technical and product characteristics of the industry sector.

[0174] In this application embodiment, when the semantic relevance between the keywords in the technical feature set and the technical entries exceeds a set threshold, and the topological overlap between the tags in the product feature set and the product path exceeds a set threshold, a technical association edge and a product circulation edge are established between the enterprise entity and the industry chain node.

[0175] Step 507: Based on the technology association edge and the product circulation edge, form an enterprise knowledge graph.

[0176] In this embodiment of the application, a graph structure containing all enterprise entities and industry chain nodes is constructed based on the generated technology association edges and product circulation edges, forming a complete enterprise knowledge graph.

[0177] Here is a specific example:

[0178] In the new energy vehicle industry chain scenario, for the multi-source data processing of battery manufacturer A, the data is first divided into a benchmark technical feature block containing five patent texts and a benchmark product feature block containing an annual product list and supply chain records, using a block model. Patent technology keywords such as lithium-ion battery and energy density are extracted from the technical feature blocks to form a technical feature set, and supply chain tags such as Yangtze River Delta delivery and 48-hour delivery are extracted from the product feature blocks to form a product feature set. From the industry association rule base, a new energy vehicle industry chain technical attribute set containing key technical items such as battery energy density greater than or equal to 300 Wh / kg is extracted, and a product attribute set containing product distribution paths such as batteries, motors, and vehicle manufacturers is extracted. The energy density keyword in the technical feature set is matched with the battery energy density greater than or equal to 300 Wh / kg item in the technical attribute set, and the semantic relevance is calculated using a word vector model, with the formula: Where u and v represent the word vectors of keywords and entries, respectively, the semantic relevance is calculated to be 0.92. The distribution relationships between the Yangtze River Delta delivery tags in the product feature set and the battery, motor, and vehicle manufacturing paths in the product attribute set are correlated, and the topological overlap is calculated using the formula: The topological overlap was found to be 0.85. With both the relevance threshold and overlap threshold set to 0.8, and given that both the semantic relevance (0.92) and topological overlap (0.85) are greater than the thresholds, technological association edges and product circulation edges are generated between battery manufacturer A and nodes in the new energy vehicle industry chain. Based on these edge relationships, a corporate knowledge graph is constructed, where nodes represent companies and industry sectors, and edges represent technological associations and product circulation relationships, forming a complete knowledge graph structure that provides a foundation for subsequent analysis.

[0179] In this embodiment, a block model is used to achieve parallel processing and feature extraction of multi-source data. Then, an enterprise knowledge graph is generated by mapping and associating with the industry rule base, which effectively improves the accuracy and completeness of the knowledge graph construction and provides a reliable knowledge foundation for subsequent industry chain analysis.

[0180] To further improve the accuracy of matching new enterprises with industry chain nodes, in some embodiments, step 103: matching the new enterprise feature set with the enterprise knowledge graph, and generating a candidate industry chain node set based on the matching results, includes:

[0181] Step 601: Separate the set of technical features to be matched and the set of product features to be matched from the new enterprise feature set.

[0182] In step 601, the set of technical features to be matched and the set of product features to be matched are two sets of features separated from the new enterprise feature set, which respectively contain technology-related features and product-related features.

[0183] In this embodiment of the application, technical features and product features are separated from the new enterprise feature set. Technical features include technical attributes such as patent technology keywords, and product features include product attributes such as supply chain tags, providing classification feature input for subsequent matching.

[0184] Step 602: Input the set of technical features to be matched, the set of product features to be matched, and the enterprise knowledge graph into a preset multimodal matching network. Through the technical matching sub-network of the multimodal matching network, calculate the dependency strength between the set of technical features to be matched and the technical attributes of the nodes in the enterprise knowledge graph.

[0185] In step 602, the dependency strength is the degree of association between the feature set of the technology to be matched and the node technology attributes in the enterprise knowledge graph, which is obtained by semantic similarity calculation.

[0186] In this embodiment of the application, the technical matching sub-network of the multimodal matching network is used to calculate the similarity between each technical feature in the set of technical features to be matched and the technical attributes of the nodes in the knowledge graph, so as to obtain the association strength value between each feature and the node technical attributes.

[0187] Step 603: Through the product matching sub-network of the multimodal matching network, calculate the overlap of circulation relationships between the feature set of the product to be matched and the product attributes of the nodes in the enterprise knowledge graph.

[0188] In step 603, the overlap of circulation relationships is the degree of matching between the feature set of the product to be matched and the node product attributes in the enterprise knowledge graph, which is calculated through path similarity.

[0189] In this embodiment of the application, the matching degree of each product feature in the feature set of products to be matched is calculated with the product attributes of the nodes in the knowledge graph through the product matching sub-network of the multimodal matching network, so as to obtain the degree of overlap between each feature and the product attributes of the nodes.

[0190] Step 604: Select multiple candidate industry chain nodes from the enterprise knowledge graph whose dependency strength is greater than a preset strength threshold and whose circulation overlap is greater than a preset circulation threshold.

[0191] In step 604, candidate industry chain nodes are nodes in the knowledge graph that simultaneously meet the requirements of technology matching and product matching, representing industry nodes with a high degree of relevance to the new enterprise.

[0192] In this embodiment of the application, nodes that have a technical matching strength exceeding a set threshold and a product matching overlap exceeding a set threshold are selected from the enterprise knowledge graph, and these nodes constitute a candidate node set.

[0193] Step 605: Aggregate all candidate industry chain nodes to form a candidate industry chain node set.

[0194] In this embodiment of the application, all the selected candidate industry chain nodes are aggregated to form a complete node set, which provides a basis for the subsequent industry chain positioning of new enterprises.

[0195] Here is a specific example:

[0196] In the new energy vehicle industry chain scenario, for the feature set processing of new battery company D, the following steps are taken: First, the matching technology feature set is extracted from its feature set, which includes two patent technology keywords: high-nickel cathode and silicon-carbon anode. The matching product feature set includes two supply chain tags: South China warehousing and 72-hour delivery. These two feature sets and the company knowledge graph are input into a multimodal matching network. The technology matching sub-network calculates the dependency strength between the matching technology feature set and the node technology attributes in the knowledge graph. The cosine similarity formula for the technology attributes of high-nickel cathode and battery manufacturer A is as follows: Where u and v represent the word vectors of keywords and node technical attributes, respectively, the dependency strength is calculated to be 0.88, and the dependency strength between silicon-carbon anode and node A technical attributes is 0.92. The product matching subnetwork calculates the overlap of circulation relationships between the feature set of products to be matched and the node product attributes. The Jaccard similarity calculation formula between South China Warehousing and node A product attributes is as follows: ,in The degree of overlap in distribution relationships is represented by A and B, which represent the sets of tag features and node product attribute features, respectively. The calculated degree of overlap in distribution relationships is 0.65, and the degree of overlap between 72-hour delivery and node A's product attributes is 0.70. A strength threshold of 0.8 and a distribution threshold of 0.75 are set. Although node A has indicators exceeding the thresholds in technical matching, its product matching does not meet the standards. Continuing to match node C (the vehicle manufacturer), the calculated strength of the dependency relationship between high-nickel cathode and node C's technical attributes is 0.86, the dependency strength between silicon-carbon anode and node C's technical attributes is 0.82, the degree of overlap in distribution relationships between South China warehousing and node C's product attributes is 0.78, and the degree of overlap between 72-hour delivery and node C's product attributes is 0.75. Since node C has indicators exceeding the thresholds in both technical and product matching, node C is selected into the candidate industry chain node set. Finally, all nodes that meet the conditions are aggregated to form a candidate node set, providing a foundation for subsequent industry chain positioning.

[0197] In this embodiment, a multimodal matching network is used to achieve feature matching in both technology and product dimensions, and to screen out the industrial chain nodes with the highest relevance to the new enterprise. This effectively improves the accuracy and completeness of the candidate node set and lays a solid foundation for subsequent industrial chain positioning.

[0198] To further improve the accuracy of industry sector identification, in some embodiments, step 105: the real-time calculation of the feature similarity between the new enterprise node and the core node of the sector in the dynamic industry chain, and based on the feature similarity, combined with the business flow intensity between the new enterprise node and the associated nodes in the candidate industry chain node set, to determine the industry sector to which the new enterprise belongs, includes:

[0199] Step 701: Calculate the similarity of technical features between the technical feature set of the new enterprise node and the technical attribute set of the domain core node.

[0200] In step 701, the technical feature similarity is the degree of matching between the new enterprise node and the core node of the domain in terms of technical features, reflecting the similarity of their technical directions.

[0201] In this embodiment of the application, the similarity between the technical feature set of the new enterprise node and the technical attribute set of the domain core node is calculated. By comparing the distribution characteristics of the patent technical keywords, a similarity algorithm is used to quantify the degree of technical matching between the two.

[0202] Step 702: Calculate the product feature similarity between the product feature set of the new enterprise node and the product attribute set of the domain core node.

[0203] In step 702, product feature similarity is the degree of matching between the new enterprise node and the domain core node in terms of product features, reflecting the similarity of their product characteristics.

[0204] In this embodiment of the application, the similarity between the product feature set of the new enterprise node and the product attribute set of the domain core node is calculated. By comparing the matching of supply chain tags and product characteristics, a similarity algorithm is used to quantify the degree of product matching between the two.

[0205] Step 703: Calculate the business flow density value between the new enterprise node and the associated nodes in the candidate industry chain node set.

[0206] In step 703, the business flow density value is the intensity of business interactions between the new enterprise node and the candidate industry chain node cluster, reflecting the actual degree of business association.

[0207] In this embodiment of the application, the business transaction data between the new enterprise node and each associated node in the candidate node set are statistically analyzed, including indicators such as order quantity and transaction frequency, and the business flow density value is obtained through quantitative calculation.

[0208] Step 704: Use an attention mechanism to perform a first weighted fusion of the technical feature similarity and the business flow density value, and perform a second weighted fusion of the product feature similarity and the business flow density value.

[0209] In step 704, the first weighted fusion and the second weighted fusion are processes that organically combine feature similarity with business flow density through an attention mechanism.

[0210] In this embodiment, an attention mechanism is used to assign weights to the technical feature similarity and business flow density values, and weighted fusion is performed to obtain a first fusion result. At the same time, weights are assigned to the product feature similarity and business flow density values, and weighted fusion is performed to obtain a second fusion result.

[0211] Step 705: When the first weighted fusion result is greater than the preset first threshold and the second weighted fusion result is greater than the preset second threshold, it is determined that the new enterprise belongs to the industry sector corresponding to the core node of the field.

[0212] In step 705, the first threshold and the second threshold are preset discrimination criteria used to determine the final industry sector affiliation.

[0213] In this embodiment of the application, the first weighted fusion result is compared with the first threshold, and the second weighted fusion result is compared with the second threshold. When both fusion results exceed the corresponding threshold, it is determined that the new enterprise belongs to the industry field corresponding to the core node of the field.

[0214] Here is a specific example:

[0215] In the new energy vehicle industry chain scenario, the technical feature set of new battery company D (containing two patented technology keywords: high-nickel cathode and silicon-carbon anode) and the technical attribute set of core node battery manufacturer A (containing two keywords: lithium-ion battery and energy density) are used to calculate the similarity of technical features. The cosine similarity formula is employed. ,in Indicates the similarity of technical features. Indicates the number of feature dimensions. The first element representing the technical feature vector of the new enterprise node One portion, The first element representing the domain core node's technical attribute vector. The technical feature similarity was calculated to be 0.85. The product feature similarity between company D (containing two supply chain tags: South China warehousing and 72-hour delivery) and node A (containing two tags: Yangtze River Delta delivery and 48-hour arrival) was calculated using the Jaccard similarity formula, yielding a similarity of 0.75. Business transaction data between company D and the associated node C (a vehicle manufacturer) in the candidate node set was analyzed. The quarterly order volume of 500,000 units divided by the industry benchmark of 625,000 units yielded a business flow density value of 0.8. A weighted fusion was performed using an attention mechanism. Technical feature similarity was assigned a weight of 0.6, and business flow density value was assigned a weight of 0.4, resulting in a first weighted fusion result of 0.85 × 0.6 + 0.8 × 0.4 = 0.83. Product feature similarity was assigned a weight of 0.5, and business flow density value was assigned a weight of 0.5, resulting in a second weighted fusion result of 0.75 × 0.5 + 0.8 × 0.5 = 0.775. With a first threshold of 0.8 and a second threshold of 0.7, the first weighted fusion result of 0.83 is greater than the first threshold and the second weighted fusion result of 0.775 is greater than the second threshold. Therefore, it is determined that the new battery company D belongs to the new energy vehicle industry field corresponding to the core node battery manufacturer A.

[0216] In this embodiment, by comprehensively calculating feature similarity and business flow density and using an attention mechanism for weighted fusion, multiple considerations of both technology and product dimensions and actual business are achieved, effectively improving the accuracy and reliability of industry-specific judgments and ensuring that the judgment results reflect both the characteristics of technology and products and the actual business relationships.

[0217] Figure 3 A schematic diagram of a system for determining the industry sector of an enterprise based on the industrial chain, provided in this application embodiment, is described in the detailed implementation section as follows:

[0218] The receiving module 31 is used to receive the application data of the new enterprise, as well as multi-source data of the target enterprise and related enterprises.

[0219] The association module 32 is used to process the multi-source data in parallel through a block model to form an enterprise feature set, and to map and associate the enterprise feature set with a preset industry association rule base to generate an enterprise knowledge graph. The enterprise feature set includes a technology feature set and a product feature set.

[0220] Extraction module 33 is used to extract the new enterprise feature set from the new enterprise application data, match the new enterprise feature set with the enterprise knowledge graph, and generate a candidate industrial chain node set based on the matching result.

[0221] Selection module 34 is used to calculate the topological connection density between each pair of nodes in the candidate industry chain node set, select the node pair with the highest topological connection density from the candidate industry chain node set as the anchor node pair, extract features from the technology association edges and product circulation edges of the anchor node pair through a graph neural network to generate enhanced edge vectors, generate an offset coefficient for the new node addition position based on the enhanced edge vectors, and add the new enterprise as a new enterprise node. In the enterprise knowledge graph, the coordinates corresponding to the offset coefficient are used as the addition position, and the new enterprise node is added. The potential interaction relationship between the new enterprise node and the anchor node pair is predicted using a relational reasoning model to generate new technology association edges and new product circulation edges. Based on the new technology association edges and the new product circulation edges, a dynamic industry chain is generated.

[0222] The calculation module 35 is used to calculate the feature similarity between new enterprise nodes and core nodes in the dynamic industrial chain in real time. Based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industrial chain node set, the industrial sector to which the new enterprise belongs is determined.

[0223] The industry sector identification system based on the industrial chain of this application is used to implement the aforementioned industry sector identification method based on the industrial chain. Therefore, the specific implementation of the industry sector identification system based on the industrial chain can be found in the embodiment section of the industry sector identification method based on the industrial chain above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0224] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described methods for determining the industry sector to which an enterprise belongs based on the industrial chain.

[0225] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for determining the industry sector to which an enterprise belongs based on the industrial chain.

[0226] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0227] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the method for determining the industry sector to which an enterprise belongs based on the industrial chain.

[0228] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0229] The foregoing has provided a detailed description of the method, system, electronic device, and storage medium for determining the industry sector of an enterprise based on the industrial chain, as provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for determining the industry sector to which a company belongs based on the industrial chain, characterized in that, include: Receive application data from new enterprises, as well as multi-source data from target enterprises and related enterprises; The multi-source data is processed in parallel using a block model to form an enterprise feature set. The enterprise feature set is then mapped and associated with a preset industry association rule base to generate an enterprise knowledge graph. The enterprise feature set includes a technology feature set and a product feature set. Extract the new enterprise feature set from the new enterprise application data, match the new enterprise feature set with the enterprise knowledge graph, and generate a candidate industrial chain node set based on the matching results; The topological connection density between each pair of nodes in the candidate industry chain node set is calculated. The node pair with the highest topological connection density is selected as the anchor node pair. Feature extraction of the technical association edges and product circulation edges of the anchor node pair is performed using a graph neural network to generate enhanced edge vectors. Based on these enhanced edge vectors, an offset coefficient is generated for the new node's addition position. The new enterprise is then designated as a new enterprise node. In the enterprise knowledge graph, the coordinates corresponding to the offset coefficient are used as the addition position, and the new enterprise node is added. A relational reasoning model is used to investigate the potential interaction relationships between the new enterprise node and the anchor node pair. The process involves predicting and generating new technology association edges and new product circulation edges. Based on these edges, a dynamic industrial chain is generated. The process of generating offset coefficients for adding new nodes based on the enhanced edge vectors includes: inputting the enhanced edge vectors into the industrial chain spatial projection layer of the graph neural network, outputting a technology-intensive direction projection component and a circulation hub direction projection component; calculating a direction fusion factor based on the ratio of the technology-intensive direction projection component to the circulation hub direction projection component; and searching for the offset coefficient for the corresponding new node addition position in a preset offset mapping table based on the direction fusion factor. The feature similarity between new enterprise nodes and core nodes in the dynamic industrial chain is calculated in real time. Based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industrial chain node set, the industrial sector to which the new enterprise belongs is determined.

2. The method for determining the industry sector to which an enterprise belongs based on the industrial chain, as described in claim 1, is characterized in that... The step of extracting features from the technical association edges and product circulation edges of the anchor node pairs using a graph neural network to generate enhanced edge vectors includes: The technical association edge of the anchor node pair is input into the first edge processing channel of the graph neural network, and the technical association feature vector is extracted from the technical association edge of the anchor node pair through the first edge processing channel. The product flow edge of the anchor node pair is input into the second edge processing channel of the graph neural network, and the product flow feature vector is extracted from the product flow edge of the anchor node pair through the second edge processing channel. The cross-channel attention fusion module of the graph neural network is used to weight and concatenate the technology association feature vector and the product circulation feature vector to generate an enhanced side vector.

3. The method for determining the industry sector to which an enterprise belongs based on the industrial chain, as described in claim 2, is characterized in that... The cross-channel attention fusion module of the graph neural network performs weighted concatenation of the technology association feature vector and the product circulation feature vector to generate enhanced side vectors, including: Based on the semantic diffusion path depth of the aforementioned technology-related feature vectors and the coverage of patent technology keywords, the dynamic impact factor of the technology is calculated. Based on the number of logistics trajectory jumps and the spatiotemporal aggregation radius of the product circulation characteristic vector, calculate the circulation dynamic impact factor; The technology dynamics impact factor and the circulation dynamics impact factor are normalized respectively to obtain the technology attention weight value and the circulation attention weight value. Based on the technology attention weight value and the circulation attention weight value, the technology association feature vector and the product circulation feature vector are weighted and scaled respectively to generate weighted technology feature blocks and weighted circulation feature blocks; A feature pyramid network is used to perform multi-scale feature fusion on the weighted technical feature block and the weighted flow feature block to generate a fused feature matrix. The fused feature matrix is ​​subjected to dimensionality reduction and compression to output enhanced edge vectors.

4. The method for determining the industry sector to which an enterprise belongs based on the industrial chain, as described in claim 3, is characterized in that... The step of using a feature pyramid network to perform multi-scale feature fusion on the weighted technical feature block and the weighted flow feature block to generate a fused feature matrix includes: Construct a feature pyramid network comprising a first pyramid layer, a second pyramid layer, and a third pyramid layer; The weighted technical feature block and the weighted circulation feature block are input into the feature pyramid network, and the output includes a technical multi-scale feature group and a circulation multi-scale feature group with three scales. Cross-modal feature fusion is performed on the technical multi-scale feature group and the circulation multi-scale feature group, wherein the first pyramid layer adopts the element-wise addition method for fusion, the second pyramid layer adopts the channel attention weighted fusion, and the third pyramid layer adopts the spatial attention weighted fusion. The fused multi-scale feature groups are upsampled to the original resolution to generate multi-scale fused features. Based on the industry type configuration parameters, the multi-scale fusion features are rearranged into a two-dimensional matrix structure according to a preset dimension ratio to form a fusion feature matrix.

5. The method for determining the industry sector to which an enterprise belongs based on the industrial chain, as described in claim 1, is characterized in that... The process of parallel processing of the multi-source data using a block model to form an enterprise feature set, mapping and associating the enterprise feature set with a preset industry association rule base to generate an enterprise knowledge graph, includes: The multi-source data is segmented into benchmark technology feature blocks and benchmark product feature blocks using a block model; Multiple patent technology keywords are extracted from the benchmark technology feature block to form a technology feature set, and multiple supply chain tags are extracted from the benchmark product feature block to form a product feature set; Extract predefined industrial chain technology attribute sets and product attribute sets from the industrial association rule base, wherein the industrial chain technology attribute set contains key technology entries corresponding to each industrial field, and the product attribute set contains product circulation paths corresponding to each industrial field. The technical feature set is matched with the key technical items for dependency relationship matching, and the semantic correlation between each patent technical keyword and the corresponding key technical item is calculated based on the matching results; The product feature set is associated with the product distribution path, and the topological overlap between each supply chain tag and the corresponding product distribution path is calculated based on the association results. When there are key technology items with semantic relevance greater than a preset relevance threshold and product circulation paths with topological overlap greater than a preset overlap threshold, technical relevance edges and product circulation edges are generated between enterprise entities and industry chain nodes. Based on the technology-related edges and the product circulation edges, an enterprise knowledge graph is formed.

6. The method for determining the industry sector to which an enterprise belongs based on the industrial chain, as described in claim 1, is characterized in that... The step of matching the new enterprise feature set with the enterprise knowledge graph and generating a candidate industry chain node set based on the matching results includes: Separate the set of technology features to be matched and the set of product features to be matched from the new enterprise feature set; The set of technical features to be matched, the set of product features to be matched, and the enterprise knowledge graph are input into a preset multimodal matching network. Through the technical matching sub-network of the multimodal matching network, the dependency strength between the set of technical features to be matched and the technical attributes of the nodes in the enterprise knowledge graph is calculated. Through the product matching sub-network of the multimodal matching network, the overlap of circulation relationships between the feature set of the products to be matched and the product attributes of the nodes in the enterprise knowledge graph is calculated. Multiple candidate industry chain nodes with dependency strength greater than a preset strength threshold and circulation overlap greater than a preset circulation threshold are selected from the enterprise knowledge graph. Aggregate all candidate industry chain nodes to form a candidate industry chain node set.

7. The method for determining the industry sector to which an enterprise belongs based on the industrial chain, as described in claim 1, is characterized in that... The real-time calculation of the feature similarity between new enterprise nodes and core nodes in the dynamic industry chain, and based on the feature similarity, combined with the business flow intensity between the new enterprise node and the associated nodes in the candidate industry chain node set, determines the industry sector to which the new enterprise belongs, including: Calculate the similarity of technical features between the technical feature set of the new enterprise node and the technical attribute set of the domain core node; Calculate the product feature similarity between the product feature set of the new enterprise node and the product attribute set of the domain core node; Calculate the business flow density value between the new enterprise node and the associated nodes in the candidate industry chain node set; The technical feature similarity and the business flow density value are first weighted and fused using an attention mechanism, and the product feature similarity and the business flow density value are second weighted and fused. When the first weighted fusion result is greater than a preset first threshold and the second weighted fusion result is greater than a preset second threshold, the new enterprise is determined to belong to the industry sector corresponding to the core node of the field.

8. A system for identifying the industry sector to which an enterprise belongs based on the industrial chain, characterized in that, include: The receiving module is used to receive application data from new enterprises, as well as multi-source data from target enterprises and related enterprises; The association module is used to process the multi-source data in parallel through a block model to form an enterprise feature set, and to map and associate the enterprise feature set with a preset industry association rule base to generate an enterprise knowledge graph. The enterprise feature set includes a technology feature set and a product feature set. The extraction module is used to extract the feature set of new enterprises from the new enterprise application data, match the feature set of new enterprises with the enterprise knowledge graph, and generate a candidate industrial chain node set based on the matching results; The selection module is used to calculate the topological connection density between each pair of nodes in the candidate industry chain node set, select the node pair with the highest topological connection density from the candidate industry chain node set as the anchor node pair, extract features from the technical association edges and product circulation edges of the anchor node pair using a graph neural network to generate enhanced edge vectors, generate an offset coefficient for the new node addition position based on the enhanced edge vectors, and add the new enterprise as a new enterprise node. In the enterprise knowledge graph, the coordinates corresponding to the offset coefficient are used as the addition position, and the new enterprise node is added. The relational reasoning model is used to analyze the potential interactions between the new enterprise node and the anchor node pair. Interrelationships are predicted to generate new technology association edges and new product circulation edges. Based on these edges, a dynamic industrial chain is generated. The process of generating offset coefficients for adding new nodes based on the enhanced edge vectors includes: inputting the enhanced edge vectors into the industrial chain spatial projection layer of the graph neural network, outputting a technology-intensive direction projection component and a circulation hub direction projection component; calculating a direction fusion factor based on the ratio of the technology-intensive direction projection component to the circulation hub direction projection component; and searching for the offset coefficient for the corresponding new node addition position in a preset offset mapping table based on the direction fusion factor. The calculation module is used to calculate the feature similarity between new enterprise nodes and core nodes in the dynamic industrial chain in real time. Based on the feature similarity and the business flow intensity between the new enterprise node and the associated nodes in the candidate industrial chain node set, the industrial sector to which the new enterprise belongs is determined.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the method for determining the industry sector to which an enterprise belongs based on the industrial chain as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the method for determining the industry sector to which an enterprise belongs based on the industrial chain as described in any one of claims 1 to 7.

Images