Industrial chain completion method based on double-tower structure and agent link

Abstract

The invention discloses an industrial chain completion method based on a double-tower structure and an intelligent agent link. The industrial chain completion method comprises the following steps: S1, reading original data of an enterprise related to industrial chain nodes; s2, carrying out characterization processing on the original data; s3, fusing the enterprise feature data with the industrial chain node data of the enterprise, and constructing a model training sample; s4, establishing a double-tower neural network structure composed of an enterprise feature tower and an industry chain node tower; s5, predicting an industry chain theme by using a double-tower model, and confirming by using a large model; and S6, the intelligent agent predicts industrial chain nodes. Under the condition that enterprise labels are scarce, a double-tower model capable of understanding semantics and structures at the same time is constructed by comprehensively utilizing the structural relation between multi-dimensional features such as the enterprise operation range, industry classification and patent information and industrial chain nodes, and hierarchical prediction from themes to nodes is achieved through agent links. And automatic and intelligent complementation of the enterprise industry chain nodes is realized.

Description

Technical Field

[0001] This invention relates to the field of supply chain mapping, and in particular to a supply chain completion method based on a dual-tower structure and intelligent agent links. Background Technology

[0002] In the construction of industry chain maps, the node to which a company belongs is core information. However, due to incomplete data disclosure and high costs of manual annotation, many company nodes lack industry chain affiliation, resulting in information gaps in the map and affecting the accuracy of industry analysis and investment research. Existing industry chain completion methods mainly rely on manual rules or keyword matching, which cannot identify semantic differences or emerging industry characteristics; while simply relying on industry classification cannot reflect the specific position of a company in the industry chain.

[0003] Currently, common solutions mainly rely on text similarity calculation or node embedding techniques, such as using TF-IDF, word vectors, or Node2Vec methods to calculate the similarity between enterprises and nodes in the industry chain, in order to assist in determining the attribution of industry chain nodes. While these methods have some effectiveness, they still mainly depend on shallow feature matching and lack the ability to jointly model enterprise semantics and industry chain structure.

[0004] Existing supply chain completion methods have the following main problems in application: It relies heavily on manual intervention and has poor generalization ability. Traditional rule-based or keyword-based matching methods heavily depend on manually building thesauri and rule templates, making it difficult to cope with the diversity and semantic ambiguity of enterprise business descriptions, and the update and maintenance costs are high; Insufficient utilization of structural information. Existing methods mostly rely solely on enterprise text or industry tags for judgment, failing to effectively integrate the upstream and downstream relationships and structural semantics between nodes in the industrial chain, resulting in prediction results lacking hierarchy and accuracy; The model's learning ability is limited. Machine learning or simple embedding methods are weak in handling high-dimensional heterogeneous features such as enterprise patents and industries, and their prediction performance is unstable when there is a severe imbalance between positive and negative samples. 4. Lack of intelligent reasoning and expansion mechanisms. Existing models mostly remain at the node level of prediction, unable to combine external knowledge or contextual information to achieve chain-like reasoning, making it difficult to automatically determine the specific position of an enterprise in the industrial chain system. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a supply chain completion method based on a dual-tower structure and intelligent agent link to overcome or at least partially solve the above problems.

[0006] According to one aspect of the present invention, a supply chain completion method based on a dual-tower structure and intelligent agent links is provided, the supply chain completion method comprising: Step S1: Read the raw data related to the enterprise and nodes in the industrial chain; Step S2: Perform feature processing on the original data; Step S3: Integrate enterprise feature data with enterprise supply chain node data to construct model training samples; Step S4: Establish a dual-tower neural network structure consisting of enterprise feature towers and industry chain node towers; Step S5: The dual-tower model predicts industry chain themes and confirms them using a large model; Step S6: The intelligent agent predicts the nodes of the industrial chain.

[0007] Optionally, step S1: reading the original data related to the enterprise and the nodes in the industrial chain specifically includes: Enterprise data includes the company name, industry, and patent information; The supply chain data includes the names of supply chain nodes and the upstream and downstream structural relationships between parent and child nodes; The enterprise information table and the industry chain node table are recorded separately in the form of data tables, which provide basic data for subsequent feature conversion and model training.

[0008] Optionally, step S2: performing feature processing on the original data specifically includes: Convert the industry field to an integer number, and count the number of patents according to the main classification number and generate a fixed-length vector. The nodes in the industrial chain are trained using the Node2Vec algorithm based on the parent-child relationship structure to obtain the structural embedding vector of each node, which is used to represent the topological semantic features between nodes. This forms a structured table of enterprise characteristics and a table of node vectors.

[0009] Optionally, step S3: fusing enterprise feature data with enterprise supply chain node data to construct model training samples specifically includes: The positive sample consists of enterprises and their actual corresponding nodes in the industry chain, while the negative sample consists of enterprises and unrelated nodes. Negative samples are generated through topic exclusion and random sampling; A fusion data table containing enterprise characteristics, industry codes, patent vectors, node numbers, and sample labels is generated to provide input data for model training.

[0010] Optionally, step S4: establishing a dual-tower neural network structure composed of enterprise feature towers and industry chain node towers specifically includes: Enterprise Tower is used to extract semantic features of enterprises. It includes a patent vector compression layer, an industry embedding layer, and a fusion layer. The enterprise vector is obtained after the fusion layer. ; The industry chain node tower is used to extract the structural features of nodes. It maps the pre-trained node embedding vectors to the same-dimensional space through a linear mapping layer to obtain the industry chain node vectors. ; The model uses the similarity between enterprise vectors and node vectors as its learning objective, and calculates the similarity using the dot product method. The dot product calculation is as follows: And it is optimized using the cross-entropy loss function, the calculation formula is as follows: Where y is the label, which is 1 for associated nodes and 0 for others; through the training process, the model automatically learns the semantic and structural correspondence between enterprises and nodes, and realizes high-dimensional feature modeling of industry chain affiliation.

[0011] Optionally, step S5: the dual-tower model predicts industry chain themes, and the large model is used to confirm the specifics, including: In the model prediction phase, the trained dual-tower model is first used to calculate the similarity between the enterprise and each industry chain theme, and the theme with the highest similarity is selected as the preliminary prediction result. The intelligent agent system accesses this result and, through the network search function, obtains publicly available information, technical fields, and news materials related to the company's industrial chain by using the company name and industrial chain as search criteria. The large model performs secondary confirmation of the predicted topics based on the business scope, patent information and search results, thereby completing the multi-source information on the industrial chain topics involved by the enterprise.

[0012] Optionally, step S6: the intelligent agent predicts the nodes in the industry chain specifically includes: After determining the theme, all nodes under the theme are extracted as a candidate set. The large model uses the business scope, patent field and network search of enterprises to output several nodes that each enterprise is most likely to belong to.

[0013] This invention provides a supply chain completion method based on a dual-tower structure and intelligent agent links. The supply chain completion method includes: Step S1: Reading raw data related to enterprises and supply chain nodes; Step S2: Performing feature processing on the raw data; Step S3: Fusing enterprise feature data with enterprise supply chain node data to construct model training samples; Step S4: Establishing a dual-tower neural network structure composed of enterprise feature towers and supply chain node towers; Step S5: The dual-tower model predicts supply chain themes and confirms them using a large model; Step S6: Intelligent agents predict supply chain nodes. In situations where enterprise tags are scarce, this method comprehensively utilizes multi-dimensional features such as enterprise business scope, industry classification, and patent information, along with the structural relationship between these features and supply chain nodes, to construct a dual-tower model that simultaneously understands semantics and structure. Furthermore, it achieves hierarchical prediction from theme to node through intelligent agent links, realizing automated and intelligent completion of enterprise supply chain nodes.

[0014] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart of a supply chain completion method based on a dual-tower structure and intelligent agent link provided for embodiments of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0020] like Figure 1 As shown, a supply chain completion method based on a dual-tower structure and intelligent agent links includes: Step S1: Retrieve raw data related to enterprises and industry chain nodes from the database. Enterprise data includes enterprise name, industry, patent information, etc.; industry chain data includes the names of industry chain nodes and the upstream and downstream structural relationships between parent and child nodes. Record the enterprise information table and industry chain node table separately in the form of data tables to provide basic data for subsequent feature transformation and model training; Step S2: Perform feature processing on the raw data. Convert the industry field into integer numbers, and count the number of patents according to the main classification number to generate a fixed-length vector. Based on their parent-child relationship structure, the nodes in the industry chain are trained using the Node2Vec algorithm to obtain the structural embedding vector of each node, which is used to represent the topological semantic features between nodes. After this step, a structured enterprise feature table and node vector table can be formed; Step S3: Integrate enterprise feature data with enterprise supply chain node data to construct model training samples. Positive samples consist of enterprises and their actual supply chain nodes, while negative samples consist of enterprises and unrelated nodes. Negative samples are generated through topic exclusion and random sampling. The final result is a fused data table containing enterprise features, industry codes, patent vectors, node numbers, and sample labels, providing input data for model training. Step S4: Establish a dual-tower neural network structure consisting of an enterprise feature tower and a supply chain node tower. The enterprise tower is used to extract semantic features of enterprises and includes a patent vector compression layer, an industry embedding layer, and a fusion layer. After the fusion layer, the enterprise vector is obtained. The industry chain node tower is used to extract the structural features of nodes. A linear mapping layer maps the pre-trained node embedding vectors to a space of the same dimension, resulting in industry chain node vectors. The model uses the similarity between enterprise vectors and node vectors as its learning objective, and calculates the similarity using the dot product method. The dot product calculation is as follows: And it is optimized using the cross-entropy loss function, the calculation formula is as follows: Where y is the label, which is 1 for associated nodes and not 0; through the training process, the model automatically learns the semantic and structural correspondence between enterprises and nodes, and realizes high-dimensional feature modeling of industry chain affiliation. Step S5: In the model prediction stage, the trained dual-tower model is first used to calculate the similarity between the enterprise and each industry chain theme, and the theme with the highest similarity is selected as the preliminary prediction result. Subsequently, the intelligent agent system accesses this result and, through the network search function, uses "enterprise name + industry chain" as the search criteria to obtain public information, technical fields, and news materials related to the enterprise's industry chain. The large model performs a secondary confirmation of the predicted theme based on the business scope, patent information, and search results, thereby achieving multi-source information completion for the industry chain themes involved in the enterprise. Step S6: After determining the theme, extract all nodes under that theme as a candidate set. The large model uses the company's business scope, patent field, and online search to output several nodes to which each company is most likely to belong. The output of this step can not only be directly used to update the industry chain map, but also provide structured support for subsequent risk identification, investment and financing analysis, etc.

[0021] A supply chain completion method based on a dual-tower structure and intelligent agent links is proposed, and the above steps are implemented using an example.

[0022] Based on the database data, step S1 yields the following raw data on enterprise characteristics and industry chain node data: The company's industry information and patent information are shown in the following two tables: The industry chain data is shown in the table below: For data on nodes in the industrial chain, in order to ensure the uniqueness of nodes under a theme, a "-" is used as a separator. The part before the "-" is the theme name and the part after the "-" is the node name. If it is a theme, its name is used directly without the "-".

[0023] Step S2 performs feature transformation on the original enterprise data. Based on their parent-child relationship structure, the nodes in the industry chain are trained using the Node2Vec algorithm to obtain fixed-dimensional vector data, which is stored in a file. The vector of a certain industry chain node is shown below: [1.0189, 0.2874, ..., -0.8905].

[0024] After the enterprise performs feature transformation, the feature data is shown in the table below: The four-character code represents the main patent classification number, and the corresponding value represents the number of patents held by the company under that classification number. The industry code adopts a discrete coding format.

[0025] Step S3 integrates enterprise feature data with enterprise supply chain node data to construct model training samples, as shown in the following table: If the supply chain is related to a company, the label is 1; otherwise, it is 0. Negative samples are generated through topic exclusion and random sampling.

[0026] Step S4 inputs the above samples into the dual-tower model. The enterprise tower extracts vectors (integrating industry and patent features), and the node tower inputs Node2Vec vectors. After linear mapping, dot product similarity is calculated in the common space. The model is trained using cross-entropy loss to gradually learn the matching relationship between enterprises and nodes.

[0027] For the enterprise tower, the number of patents under the main patent classification number is first logarithmically transformed to reduce the impact of extreme values. Since there are many main patent classification numbers, a fully connected layer of a neural network is used for linear mapping to achieve a lower dimensionality. The industry embedding vector is in one-hot form. Then, the professional and industry vectors are fused. Finally, the fused vector is mapped to the specified dimension. A specific example is shown below: For the supply chain pyramid, a linear mapping layer maps the pre-trained supply chain vectors to a specified dimension, as shown in the following example: By defining dimensions that make the output dimensions of the enterprise tower and the supply chain tower the same, similarity can be calculated using dot product and optimized using the cross-entropy loss function.

[0028] Step S5: After training is complete, for company A, the trained dual-tower model is used to calculate the similarity between the company and each industry chain theme. The theme with the highest similarity is selected as the preliminary prediction result. For example, the output result is: "New software and services, computer, cloud computing".

[0029] The agent receives the company name and the predicted theme. It then calls code nodes to obtain data on the company's business scope and core technology areas. Simultaneously, it searches through network nodes, retrieving information on the company's industrial chain using the format "company name + industrial chain". The large model then analyzes all industrial chain themes based on this information, identifying the most likely themes, such as "new software and services, computers, and the Internet".

[0030] Step S6: After confirming the industry chain theme, select all nodes under the corresponding theme as backup nodes. At the same time, combined with the relevant data from step S5, the large model gives several most likely industry chain nodes, such as "new software and services - financial technology, computer - artificial intelligence software products, computer - private cloud services...".

[0031] This invention proposes a method for using a dual-tower neural network structure (enterprise tower and industry chain node tower) for industry chain completion tasks. By separately encoding the semantic features of enterprises and the structural features of industry chain nodes within each tower, and performing similarity matching and cross-entropy training in a unified vector space, high-dimensional semantic association learning between enterprises and nodes is achieved. This method maintains stable predictive capabilities even under conditions of scarce labels and heterogeneous features, and constructs a scalable and transferable modeling approach for industry chain node prediction.

[0032] This invention proposes a hierarchical completion method that combines the semantic reasoning capabilities of a large-scale model. Based on the topic prediction output of a dual-tower model, a large-scale model is introduced to realize a multi-level reasoning chain from topic identification to node confirmation. The large-scale model automatically parses corporate text, patent descriptions, and external information, identifies potential technological directions, and performs semantic comparison, achieving a leap from numerical computation to semantic logical reasoning. This forms an integrated approach to semantic reasoning and node prediction that can be applied across different industry sectors.

[0033] Beneficial Effects: This invention proposes a supply chain completion method based on a dual-tower structure and intelligent agent links. This method constructs a dual-tower neural network consisting of an enterprise feature tower and a supply chain node tower to achieve high-dimensional matching and learning of semantic vectors between enterprises and nodes. Combined with an intelligent agent link mechanism, it introduces business scope, patent fields, and online retrieval information to complete layer-by-layer reasoning from the topic level to the node level. Thus, even when enterprise data is incomplete or tags are scarce, it automatically and accurately completes the supply chain nodes to which an enterprise belongs, achieving intelligent updating and expansion of the supply chain map.

[0034] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A supply chain completion method based on a dual-tower structure and intelligent agent links, characterized in that, The methods for completing the industrial chain include: Step S1: Read the raw data related to the enterprise and nodes in the industrial chain; Step S2: Perform feature processing on the original data; Step S3: Integrate enterprise feature data with enterprise supply chain node data to construct model training samples; Step S4: Establish a dual-tower neural network structure consisting of enterprise feature towers and industry chain node towers; Step S5: The dual-tower model predicts industry chain themes and confirms them using a large model; Step S6: The intelligent agent predicts the nodes of the industrial chain.

2. The supply chain completion method based on a dual-tower structure and intelligent agent link as described in claim 1, characterized in that, Step S1: Reading the raw data related to enterprises and nodes in the industrial chain specifically includes: Enterprise data includes the company name, industry, and patent information; The supply chain data includes the names of supply chain nodes and the upstream and downstream structural relationships between parent and child nodes; The enterprise information table and the industry chain node table are recorded separately in the form of data tables, which provide basic data for subsequent feature conversion and model training.

3. The supply chain completion method based on a dual-tower structure and intelligent agent link as described in claim 1, characterized in that, Step S2: The feature processing of the original data specifically includes: Convert the industry field to an integer number, and count the number of patents according to the main classification number and generate a fixed-length vector. The nodes in the industrial chain are trained using the Node2Vec algorithm based on the parent-child relationship structure to obtain the structural embedding vector of each node, which is used to represent the topological semantic features between nodes. This forms a structured table of enterprise characteristics and a table of node vectors.

4. The supply chain completion method based on a dual-tower structure and intelligent agent link as described in claim 1, characterized in that, Step S3, which involves fusing enterprise feature data with enterprise supply chain node data to construct model training samples, specifically includes: The positive sample consists of enterprises and their actual corresponding nodes in the industry chain, while the negative sample consists of enterprises and unrelated nodes. Negative samples are generated through topic exclusion and random sampling; A fusion data table containing enterprise characteristics, industry codes, patent vectors, node numbers, and sample labels is generated to provide input data for model training.

5. The supply chain completion method based on a dual-tower structure and intelligent agent link as described in claim 1, characterized in that, Step S4: Establishing a dual-tower neural network structure composed of enterprise feature towers and industry chain node towers specifically includes: Enterprise Tower is used to extract semantic features of enterprises. It includes a patent vector compression layer, an industry embedding layer, and a fusion layer. The enterprise vector is obtained after the fusion layer. ; The industry chain node tower is used to extract the structural features of nodes. It maps the pre-trained node embedding vectors to the same-dimensional space through a linear mapping layer to obtain the industry chain node vectors. ; The model uses the similarity between enterprise vectors and node vectors as its learning objective, and calculates the similarity using the dot product method. The dot product calculation is as follows: And it is optimized using the cross-entropy loss function, the calculation formula is as follows: Where y is the label, which is 1 for associated nodes and 0 for others; through the training process, the model automatically learns the semantic and structural correspondence between enterprises and nodes, and realizes high-dimensional feature modeling of industry chain affiliation.

6. The supply chain completion method based on a dual-tower structure and intelligent agent link as described in claim 1, characterized in that, Step S5: The dual-tower model predicts industry chain themes, and the large model is used to confirm the specifics, including: In the model prediction phase, the trained dual-tower model is first used to calculate the similarity between the enterprise and each industry chain theme, and the theme with the highest similarity is selected as the preliminary prediction result. The intelligent agent system accesses this result and, through the network search function, obtains publicly available information, technical fields, and news materials related to the company's industrial chain by using the company name and industrial chain as search criteria. The large model performs secondary confirmation of the predicted topics based on the business scope, patent information and search results, thereby completing the multi-source information on the industrial chain topics involved by the enterprise.

7. The supply chain completion method based on a dual-tower structure and intelligent agent link as described in claim 1, characterized in that, Step S6: The intelligent agent predicts the nodes of the industrial chain, specifically including: After determining the theme, all nodes under the theme are extracted as a candidate set. The large model uses the business scope, patent field and network search of enterprises to output several nodes that each enterprise is most likely to belong to.

Images