Industrial association degree measuring and calculating method and system based on enterprise multi-source data fusion

Abstract

The invention discloses an industrial chain association degree quantitative analysis method and system based on multi-source enterprise big data. The method comprises the following steps: collecting multi-dimensional data such as enterprise transaction, capital, innovation, talents and the like, establishing industry association through a knowledge graph and a semantic embedding technology, and constructing a basic traffic matrix; constructing and carrying out weighted fusion on the contribution matrix and the dependency matrix of each dimension to form a composite industry contribution matrix and a dependency matrix; the weight is dynamically determined by adopting an adjusting coefficient entropy weight method, and objectivity is enhanced in combination with network topology and priori knowledge; and further calculating a full-order indirect correlation effect through a matrix power series, and finally synthesizing an industrial total correlation matrix simultaneously covering forward supply driving and backward demand pulling. The method can efficiently and accurately quantify the direct and multiplication conduction effect of the performance change between the industries, and provides a reliable basis for industrial policy making, industrial financial risk management and enterprise strategic decision making.

Description

BACKGROUND

[0001] Industrial correlation analysis is an important method to study the input-output interdependence relationship between various industrial sectors of the national economy. The core of the traditional industrial correlation analysis tool is the input-output table, which is based on macro value-type data. It has inherent defects such as coarse granularity of sector classification, price bias in tabulation, and mixed survey objects. First, the preparation cycle is long (usually in years, updated every five years), and the data update is severely lagging, which cannot capture the dynamic changes under emerging industries, sub-sectors, and sudden economic events. Second, it mainly reflects the technical and economic links (value flow) of products and services, and it is difficult to depict the capital control, innovation collaboration and knowledge diffusion, talent flow and other important correlation dimensions in modern economy. Finally, the sector classification has coarse granularity and the survey objects are often mixed, which makes it difficult to accurately distinguish the correlation strength of enterprises of different sizes, regional differences or specific industry chain links, thereby affecting the accuracy of the analysis and the value of policy guidance.

[0002] In recent years, big data technology has provided new enterprise micro data sources for industrial analysis, such as enterprise bidding, equity relations, senior management positions, and patent applications. Existing research has attempted to use these data to build industrial networks, but most have limitations. Patent CN110443489A uses neural networks to predict target industries and their associated industries by filtering customer history and real-time data. This method focuses on feature extraction and does not consider semantic embedding and multi-source weighted fusion, resulting in low correlation accuracy. Patent CN109377378A stores financial indicators and compares floating information to determine correlation and assess risk. This method ensures data security but is limited to the financial dimension and cannot quantify the full transmission effect between industries. Patent CN113792158A uses transaction flow to construct a capital graph and uses PrefixSpan to mine industry sequences. This method emphasizes capital paths but has a single preprocessing method and does not integrate capital and innovation, making it difficult to capture structural correlations. Patent CN119919056A assesses supply chain risk by calculating static and dynamic risk. This method introduces a graph but focuses on risk edge weights and does not extend to multi-source matrix fusion and full correlation coefficients, making it suitable for specific risk correlation calculations.

[0003] In summary, existing technology routes either rely on a single data source (such as only using patent cooperation to represent innovation correlations) or simply sum the data without addressing data sparsity, noise, and dimension inconsistencies. More importantly, there is a lack of a unified framework that can integrate multi-dimensional flows (transaction flow, capital flow, innovation flow, and knowledge flow) and quantify direct and indirect transmission effects. There is an urgent need for a method that can utilize multi-source heterogeneous big data, overcome the lag of traditional methods, and accurately and comprehensively measure the complex correlation between industries. SUMMARY

[0004] (a) Purpose of the invention

[0005] To address the problems of existing technologies, this invention provides a method and system for measuring industrial correlation based on multi-source data fusion. This invention emphasizes two core dimensions: first, the method for calculating industrial correlation, which constructs a flow matrix from multi-source data, fuses a composite coefficient matrix, and calculates the total correlation coefficient, achieving a comprehensive quantification of direct and indirect correlation effects; second, the big data processing system employing this method, which utilizes a distributed computing framework to efficiently process massive amounts of data, supporting real-time analysis and large-scale applications. This method and system are applicable to real-time / monthly / quarterly / semi-annual / annual calculations of correlation between first-, second-, and third-level sub-sectors of any industry, significantly improving the accuracy and timeliness of updates.

[0006] (II) Technical Solution

[0007] A method for measuring industrial relevance based on multi-source data fusion includes the following steps:

[0008] First, we legally collect multi-dimensional open-source big data from the internet and third-party databases, including transaction data (bidding records, supplier-customer relationships), capital data (equity relationships, financing events), innovation data (patents, software copyrights, technical standards), and talent data (changes in legal representatives, appointments of senior executives). We then create a knowledge graph of the industry chain and generate keywords using a Large Language Model (LLM). Vector search technology is applied to map enterprises onto the industry chain knowledge graph. Data collection covers over 50 million enterprise-related big data points nationwide, ensuring timeliness and completeness.

[0009] Secondly, a transaction flow matrix Z is constructed based on the transaction data, which quantifies the intensity of inter-industry transactions; then, the contribution coefficient matrix Z is calculated. * (Characterizing supply contribution) and dependency coefficient matrix Z ** (Reflecting procurement dependence). Similarly, a capital flow matrix M is constructed based on capital data, and M is calculated. * and M ** Construct an innovation flow matrix T based on innovation data, and calculate T. * and T ** Construct a talent flow matrix K based on talent data, and calculate K. * and K ** Semantic embedding matrices G (binary matrices representing structural associations) are constructed based on semantic data. The construction process of these matrices involves time-decay weighting and fusion mechanisms to capture dynamic changes.

[0010] Next, the weights of each submatrix are determined using the entropy weighting method with adjustment coefficients. This method combines entropy (information measure), average degree (network density), and importance (preset factor) to ensure the objectivity and adaptability of the weights; then, the contribution coefficient matrix Z is fused.* M * T * The industrial composite contribution matrix L is obtained from K* and G. * (Quantify forward influence), and fuse the dependency coefficient matrix Z ** M ** T ** K ** G obtains the industry composite dependency matrix L ** (Quantifying backward effects). Composite matrix fusion enables the organic integration of multi-dimensional data.

[0011] Finally, based on the aforementioned industry composite contribution matrix L * Calculate the forward total contribution correlation coefficient matrix R * (Including direct and all-order indirect effects), based on the aforementioned industry composite dependency matrix L ** Calculate the backward full dependency correlation coefficient matrix R ** The forward total contribution correlation coefficient matrix R * and the backward full dependency correlation coefficient matrix R ** The summation yields the overall industry correlation matrix R. This calculation considers the two-way transmission effect of supply-driven and demand-pull factors, quantifying the multiplicative impact of changes in industry performance.

[0012] Furthermore, constructing the transaction flow matrix Z based on the transaction data includes: collecting bidding data and supplier-customer relationship data of enterprises in the industry; and calculating the number of transactions Z1 between industry i and industry j. ij The cumulative weighted average of new transaction volume over the years, with the weight decreasing exponentially over time; Z2 ij Let Z be the target number of transactions between industry i and industry j. The transaction flow matrix Z combines Z1 and Z2. .

[0013] Furthermore, the contribution coefficient matrix Z * element Z ij * =Z ij / Z oi Z oi Let be the total number of output transactions in industry i, representing the supply contribution level of industry i to j. This coefficient indicates that a one-unit change in the performance of industry i directly leads to a change of Z in j. * ij Unit; the dependency coefficient matrix Z ** element Z ij ** =Z ij / Z Ij Z IjThe total number of transactions for industry j represents the degree to which industry j depends on demand from industry i. This coefficient indicates that a one-unit change in the performance of industry j directly leads to a change of Z in industry i. ** ij Units. Similar to calculating M. * M ** T * T ** K * K ** Where G is a 0-1 matrix representing upstream and downstream semantic relationships, defined by the knowledge graph and expert experience, for example G ij =1 indicates that industry i is upstream of industry j.

[0014] Furthermore, the determination of the weights of each submatrix using the adjustment coefficient entropy weight method includes: calculating the entropy value based on the contribution sequence. p is the normalization element, which yields the initial weights. Here, m represents the number of elements, and this weight reflects information redundancy. The average degree of the sub-network, D = (total degree of nodes) / number of nodes, is calculated, and the weights are adjusted to reflect connection density. An additional multiplication by the sub-network importance I (preset values, such as transaction I = 1.2, capital I = 1.0, innovation I = 1.1, semantic I = 0.8) yields the comprehensive weight α = W × D × I. This method is superior to subjective weighting and avoids bias. * =a1Z * +a2M * +a3T * +a4K * +a5G

[0015] L ** Similarly. This fusion ensures L * and L ** The economic explanatory power.

[0016] Furthermore, the industrial composite contribution matrix L... * Calculate the forward total contribution correlation coefficient matrix R * The formula is: , where [L * (T)] k L* is a power of k matrix, representing the k-th order indirect effect. For example, k=0 represents itself (identity matrix), k=1 represents direct influence, and k=2 represents second-order transmission (e.g., i affects k in the middle, which then affects j); R ** Similar calculation. Total industry linkage coefficient matrix R = R * + (R ** ) T , where T represents matrix transpose. This calculation considers both supply-driven (forward) and demand-pull (backward) effects, quantifies the transmission of performance changes, and assumes synchronous labor force changes to ensure accurate multiplication of the total effect.

[0017] ② This invention also provides an industry correlation measurement system based on multi-source data fusion. This system uses the above method to achieve big data processing and real-time calculation and updating of correlation, including:

[0018] The data acquisition module, adhering to the fundamental principle of data security, legally collects nationwide enterprise-related big data from the "National Enterprise Credit Information Publicity System," the "National Bidding and Tendering Website," and third-party databases, covering dimensions such as industry and commerce, intellectual property, bidding and tendering, news and public opinion, and investment and financing. This module supports distributed crawling and storage, utilizing the Hadoop or Spark framework to process massive amounts of data, ensuring data real-time performance and integrity.

[0019] The knowledge modeling module defines first-, second-, and third-level industry entities in the industrial chain by designing large model prompts. It generates 10-15 core domain-specific thesaurus terms around each industry entity node. These thesaurus terms should accurately describe the products, technologies, and services in that domain, generating a structured industry knowledge system. From the collected raw data, key textual information for each enterprise, such as business scope, product descriptions, patent names, and winning project names, is extracted and merged into a "company document" representing the enterprise's business characteristics. BERT is used to convert each "industry-specific thesaurus group" and each "company document" generated in the previous step into high-dimensional vectors. All "company document vectors" are stored in an Elasticsearch (ES) index. When matching is required, an "industry thesaurus vector" is used as the query vector, and ES's cosine similarity function is used for searching. Through aggregation of multiple query results, the industry to which each enterprise belongs is finally determined.

[0020] The manual verification module features a user-friendly web-based management interface. Industry experts can batch review and view automatically generated company tags by industry or random sampling. For erroneous tags, experts can directly select the correct industry category or add notes explaining the reason. They can also add, delete, and modify industry terms generated by the LLM (Local Management Model) to better reflect actual business understanding. All manual verification results are recorded by the system, forming a high-quality "verification dataset." This dataset is used for model fine-tuning, allowing the model to learn the expert's judgment logic and continuously optimize the accuracy of the next automatic labeling. The corrected, accurate data is then fed back into the knowledge base and Elasticsearch index, overwriting previous erroneous data and ensuring that downstream computing modules use high-quality data.

[0021] The industry correlation calculation module is used to process data according to the above-mentioned industry correlation measurement method based on multi-source data fusion. It performs distributed ETL cleaning on the raw data through Spark SQL to generate an industry flow matrix that spans years and departments. It calls the BLAS library of MLlib to realize distributed matrix inversion, eigenvalue calculation and coefficient derivation, and calculates contribution coefficient, dependency coefficient, forward total contribution correlation coefficient, backward total dependency correlation coefficient and total industry correlation coefficient. It uses Spark Streaming to perform micro-batch processing on incremental data (such as real-time bidding data) and combines caching mechanisms (such as Alluxio) to accelerate iterative calculation and meet the minute-level response requirements for TB-level data.

[0022] The results display module showcases the historical correlation matrices of primary, secondary, and tertiary industries. This module utilizes visualization tools (such as Echarts or Tableau) and supports heatmaps, network diagrams, and time-series views for user-friendly interactive analysis. The heatmap displays the coefficient strength of the primary, secondary, and tertiary industry correlation matrices for any given year; the network diagram uses a force-directed graph to dynamically present the correlation topology between industry sectors; and the time-series view allows for sliding time axes to compare the evolution trends of correlation over multiple years.

[0023] This system is built on a big data framework (such as Apache Spark), supports cloud deployment and real-time updates, and is suitable for calculating the correlation of massive amounts of industrial data, ensuring computational efficiency and scalability. The system's logical flow is as follows: data acquisition (source input) → knowledge modeling (entity definition and linking, providing a basis for classification) → manual verification (accuracy assurance, correcting automation errors) → correlation calculation (core operations, generating matrices) → result display (output applications, supporting decision-making), forming a closed-loop optimization to ensure the logical coherence and reliability from data to results.

[0024] (III) Beneficial Effects

[0025] Compared with the prior art, the significant advantages and innovations of this invention are as follows:

[0026] ① Universality and Precision: Applicable to the correlation measurement of any industry (e.g., embodied intelligence, brain-like intelligence, low-altitude economy, digital economy), unlike the input-output table (method) which is limited to the standard national economic industry classification. It integrates four-dimensional micro-enterprise big data of transactions, capital, innovation and talent within a unified framework, supports multi-dimensional correlation analysis of value transfer, capital investment, technological innovation and talent factors between industries, and aggregates and calculates the correlation of sub-nodes in the industrial chain through micro-enterprise behavior, making the correlation of industries more granular.

[0027] ② Dynamism and Foresight: Employing a high-frequency internet data time decay model, historical and recent data are organically combined, enabling the calculation results to more quickly reflect the dynamic changes in the industrial structure. This addresses the pressing issue of the current five-year update cycle for my country's input-output tables being inefficient and lagging in terms of updates. It not only provides historical data but also allows for prediction and simulation based on data trends, greatly enhancing the forward-looking nature and practicality of the analysis. Multi-source data complementarity and equity penetration techniques effectively combat the incompleteness and noise problems of single data sources, improving the model's stability and accuracy.

[0028] ③ Scientific Rigor and Objectivity: By automatically determining weights using the entropy weight method combined with network topology characteristics, the bias of subjective judgment is greatly reduced. The entire weight determination process is driven by algorithms and data, requiring little or no human intervention, fundamentally eliminating subjective biases caused by personal experience, preferences, or cognitive limitations. When new input data changes over time, the entropy weights and topology weights are automatically recalculated and adjusted, enabling the model to dynamically evolve and continuously ensuring the objectivity of the results.

[0029] ④ Strong Explanatory Power and Practicality: This invention not only captures the direct consumption relationships between industrial sectors, but more importantly, it fully quantifies the intricate indirect transmission effects. It clearly reveals the amplifying and cyclical ripple effects of changes in final demand within a particular sector on the entire economic system. This characteristic enables the matrix to realistically and comprehensively reflect the deep-seated and interdependent relationships between various links in the modern industrial chain network. It has extremely high practical value in assisting in the formulation of precise industrial policies, systematically assessing and preventing supply chain disruption risks, and monitoring and controlling risks within the industrial chain. It can provide in-depth decision-making support for governments, research institutions, and enterprises.

[0030] Instruction manual illustrations

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Figure 1 A flowchart illustrating an exemplary embodiment of the present disclosure for calculating industry relevance based on multi-source data fusion is provided.

[0033] Figure 2 The diagram illustrates the system architecture of an industry correlation measurement system based on multi-source data fusion according to an exemplary embodiment of the present disclosure.

[0034] Figure 3A schematic diagram illustrating a multi-level equity penetration diagram according to an exemplary embodiment of this disclosure; Detailed Implementation Plan

[0035] To achieve the above objectives, the present invention adopts the following technical implementation scheme:

[0036] (I) Process Overview

[0037] A method for measuring industry correlation based on multi-source data fusion of enterprises, executed by a system deployed on a big data computing platform, is proposed. Its core idea is to aggregate micro-enterprise behavioral data into an inter-industry flow matrix, form a composite industry network through standardization and weighted fusion, and analyze the direct and indirect correlation effects within the network using matrix operations. The following section uses the lithium industry as an example to illustrate this method, specifically including the following steps:

[0038] S1 Data Preparation: Collect and preprocess enterprise-related data on the Internet, and complete the modeling and extraction of the industry chain knowledge graph;

[0039] S2 Basic Matrix Construction: Complete complex correlation calculations of multi-dimensional data to construct transaction, capital, innovation, and talent flow matrices;

[0040] S3 coefficient matrix calculation: The basic flow matrix is ​​standardized into a contribution coefficient matrix and a dependency coefficient matrix to eliminate the influence of scale and reveal the strength of the relatively direct correlation.

[0041] S4 Semantic Knowledge Embedding: Introduces a semantic matrix based on industry knowledge graphs to quantify expert experience and make up for the shortcomings of pure data-driven approaches;

[0042] S5 Adaptive Matrix Fusion: Employs an improved entropy weighting method, comprehensively considering information content and network topology, assigning appropriate weights to each sub-matrix, and fusing them to generate a composite matrix;

[0043] S6 Full Correlation Calculation: By solving the matrix series (i.e., the analogous form of the Leontief inverse matrix), the forward full contribution and backward full dependence correlation coefficient matrix containing all indirect effects are calculated, and finally the total industry correlation matrix is ​​synthesized.

[0044] S7 Data Visualization: This solution is based on Echarts to develop heatmaps, force-directed network diagrams, and time axes, enabling dynamic, interactive visualization analysis and trend comparison of the primary, secondary, and tertiary industry correlation matrices.

[0045] (II) Detailed Steps

[0046] S1: Multi-source data acquisition and knowledge modeling

[0047] The system receives raw data from an internal database or a third-party data provider. Preprocessing includes:

[0048] ① Data cleaning: Remove duplicate and invalid records, and unify company names, industry codes, and basic information.

[0049] ② Knowledge Modeling: Through consultation with industry experts and interviews with representative companies, experts manually defined a lithium industry chain knowledge graph including first-level, second-level, and third-level sub-industry nodes. The graph consists of 8 first-level industries, including lithium mining and beneficiation, lithium smelting, lithium salt processing, and lithium batteries; 24 second-level industries, including lithium ore, basic lithium salts, secondary lithium salts, and electrolyte lithium salts; and 68 third-level industries, including spodumene, lithium feldspar, lithium carbonate, and lithium hydroxide. The graph also uses LLM (Large Language Model) prompts to generate keywords for sub-industry nodes. For example, synonyms or near-synonyms for lithium hydroxide include lithium hydroxide, lithium hydroxide compound, lithium alkali, LiOH, and lithium hydroxide.

[0050] ③ Industry Mapping: Utilizing ElasticSearch's vector semantic search technology, precise links are established between specific industry nodes and enterprise entities. First, a BERT+CRF model is used to extract key textual information for each enterprise, such as: business scope (e.g., lithium mining and beneficiation), main products (e.g., lithium carbonate and lithium hydroxide), relevant patent names (e.g., "a lithium extraction process from spodumene"), and names of winning bids. This textual information is then merged to generate a comprehensive document representing the enterprise's business characteristics. Second, the BERT model is used to convert predefined "industry-specific thesaurus" (e.g., "lithium mining," "lithium salt processing," "cathode materials") and each business characteristic document into high-dimensional vectors. All document vectors generated for each enterprise are then built into a dedicated index in ElasticSearch (ES). When it is necessary to determine the specific industry to which an enterprise belongs, the thesaurus vectors such as "lithium mining and refining" are used as query vectors. ES's cosine similarity function is used for semantic search. Through multiple queries and result aggregation, the specific industry node to which each lithium industry chain enterprise belongs is ultimately accurately determined. Ultimately, we obtained data from 23,464 lithium upstream and downstream enterprises nationwide, totaling 36,301 enterprise benchmark data entries.

[0051] ④ Information Extraction: The system accurately extracts four key fields—"Bidding Party - Successful Bidder - Time - Amount"—from bidding transaction data to analyze business cooperation dynamics; it analyzes "Investor - Investee - Capital Contribution - Shareholding Ratio - Establishment Time" from equity data to clearly depict corporate capital structure and investment relationships; it identifies "Applicant / Copyright Holder / Formulator - Partner - Type - Time" from intellectual property data to reveal innovation activities and technology collaboration networks; and it extracts "Former Company - New Company - Time - Position" from personnel employment data to accurately track the flow and career trajectory of core talent, providing solid data support for competitive intelligence and risk monitoring.

[0052] ⑤ Manual Sampling Inspection: To ensure the quality and reliability of the company's product labeling system, after the system automatically completes the initial mapping and labeling, analysts will review each business and product label generated by the system. Based on established classification standards and business rules, they will verify the consistency between the labels and the actual business operations, focusing on correcting mislabeling, omissions, and classification discrepancies. This manual review mechanism aims to completely eliminate any ambiguity or errors that may exist in machine recognition, ensuring that each label accurately reflects the company's true business scope, thereby providing a high-quality and reliable data foundation for subsequent calculations.

[0053] S2: Constructing the basic traffic matrix

[0054] ①Industry transaction flow matrix

[0055] Use a i Let b represent the a-th company in industry i. j This represents the b-th company in industry j. ij(a,b) Indicates a i Selling goods and services to b j The number of transactions or customers, the total number of transactions (total number of customers) of goods and services sold from industry i to industry j is: Z Oi (O represents output out) represents the total number of output transactions (total number of customers) for industry i across all industries; Z Ij (I represents input) represents the total number of transactions (total number of customers) between all industries and industry j. Transaction flow Z ij The calculation is based on the weighted cumulative value of new transaction volume over the years (T is the winning bid time), with the weight decaying exponentially over time (see Formula 1-1). Finally, the transaction count matrix Z1 and the transaction target (number of customers) matrix Z2 are obtained.

[0056] (T is the current time point, t is the historical time point) (Equation 1-1)

[0057] The industry chain knowledge graph was processed into an industry tree structure, and a joint query was performed with enterprise bidding big data. The query results included more than ten fields such as "industry category i, bidding party (customer), industry category j, winning bidder (supplier), bidding time, and amount", totaling 518,804 enterprise data. Through classification and summary statistics, the total number of bids from industry i to industry j was obtained as Z1. ij Then, a consolidated query was performed with the "supplier-customer" data extracted from court documents, financial reports, and news and public opinion from a third-party data source. The query results contained a total of 474,626 data entries, including fields such as "industry i, customer, industry j, supplier, reporting period, and relationship". Through classification, summarization, and statistics, the total number of customers of industry i to industry j, Z2, was obtained. ijFinally, the matrix Z1 of the number of bidding transactions and the matrix Z2 of the transaction targets (number of customers) are obtained.

[0058] Table 1-1 Transaction Count and Flow Matrix Z1 Unit: Transactions

[0059]

[0060] Table 1-2 Transaction Target Flow Matrix Z2 (Unit: Quantities)

[0061]

[0062] ②Industry Capital Flow Matrix

[0063] m ij For a i The sum of the equivalent capital contributions to b through all direct and indirect paths is calculated by adding up the capital contributions subscribed by a for first-level direct shareholding in b and the indirect capital contributions from second to fifth levels; for details of the five-level equity penetration path, see [link to details]. Figure 3 Investment flow from industry i to industry j M Oi Industry i represents outflows of funds across all industries; M represents outflows of funds across all industries. Ij This represents the total inflow of funds into industry j across all industries. Among these, capital flow M... ij (T) The data is calculated based on the weighted cumulative value of new investment over the years, with the weights decreasing exponentially over time (T represents the investment time; this article uses the registration and establishment time of the investing company, not the subscribed capital time). After jointly querying the industry tree structure data and equity investment data, the following results were obtained: 159,453 entries for Level 1 equity penetration (including both upward and downward penetration, with duplicate entries obtained after deduplication); 166,361 entries for Level 2 equity penetration; 122,309 entries for Level 3 equity penetration; 86,835 entries for Level 4 equity penetration; and 84,697 entries for Level 5 equity penetration.

[0064] (Equation 1-2)

[0065] Note: f Aij Let p0 be the total subscribed capital contribution of a to b, and p0 be the shareholding ratio of a to b; f Bij Let f be the subscribed capital contribution of c1 to b, p1 be the shareholding ratio of a in c, and dp1 be the shareholding ratio of c1 in b; Cij p2 is the subscribed capital contribution of c2 to b, p2 is the shareholding ratio of c1 to c2, and dp2 is the shareholding ratio of c2 to b; f Dij f represents the subscribed capital contribution of c3 to b, p3 represents the shareholding ratio of c2 to c3, and dp3 represents the shareholding ratio of c3 to b; Eijdp4 represents the subscribed capital contribution of c4 to b, p4 represents the shareholding ratio of c3 to c4, and dp4 represents the shareholding ratio of c4 to b.

[0066] (T is the current time point, t is the historical time point) (Equation 1-3)

[0067] Table 1-3 Capital Flow Matrix M (Unit: RMB 10,000)

[0068]

[0069] ③Industry Innovation Flow Matrix

[0070] p ij For a i and b j The total number of jointly filed patents, the total number of jointly filed patents for industry i and industry j is P ij (t) equals P per year ij The weighted cumulative decay value; P Oi P represents the number of patents jointly filed by industry i and all industries. Ii This represents the number of patent applications jointly filed by all industries in industry i. During data processing, a total of 1,048,576 patent application (grant) records containing industry classifications were extracted, yielding 4,731,815 patent entries containing fields such as "industry classification i, applicant, industry classification j, collaborator, patent name, application date, application number / grant number, and inventor". i j Represented as a i and b j The number of copyrights obtained through collaboration, and the total number of copyrights obtained through patent collaboration between industry i and industry j are: W ij (t) equals W per year ij Weighted attenuation cumulative value; P Oi P represents the number of copyrights obtained by industry i in cooperation with all industries. Ii This represents the number of copyrights acquired through collaborations between all industries and industry i; since lithium companies have acquired relatively few copyrights through collaborations, they are ignored in this calculation of the traffic matrix. ij Represented as a i and b j The number of jointly constructed standards, the number of jointly constructed standards for industry i and industry j is: S ij (t) is S ij Weighted attenuation cumulative value; S Oi S represents the number of standards jointly built by industry i and all industries. IiThis represents the number of standards jointly built by all industries and industry i; a total of 1,524,191 standard application (authorization) data containing industry classifications were extracted, and 4,478,114 patent records were obtained, including fields such as "industry classification i, applicant, industry classification j, partner, standard name, promulgation time, and standard number".

[0071] Table 1-4 Patent Flow Matrix P Unit: Items

[0072]

[0073] Table 1-5 Standard Flow Matrix S Unit: pieces

[0074]

[0075] ④ Industry talent flow matrix

[0076] k ij For a i Joining B j The total number of senior talents, including legal representatives, senior executives (directors / supervisors / managers), and research and development personnel, and the total number of senior talents flowing from industry i to industry j is [missing information]. K ij (t) equals K per year ij The weighted cumulative decay value; K Oi K represents the total number of senior talents flowing from industry i to all industries. Ii This represents the total number of senior talents flowing from all industries to industry i; a total of 361,876 talent flow data points with industry classifications were extracted, and 578,114 micro-level talent flow data points were obtained, including fields such as "industry classification i, original company, new company, time, position (legal representative), and industry classification j".

[0077] Table 1-6 Talent Flow Matrix K Unit: Items

[0078]

[0079] S3: Calculate the standardized coefficient matrix

[0080] Contribution coefficient matrix (X) * ): Represents the strength of the influence of output industry i on input industry j. * ij = X ij / X Oi , where X Oi = Σj X ij (Total output of industry i). The sum of all rows in this matrix is ​​1.

[0081] - Dependency coefficient matrix (X) **): Represents the degree of dependence of input industry j on output industry i. X ** ij = X ij / X Ij , where X Ij = Σi X ij (Total input for industry j). The sum of all columns in this matrix is ​​1.

[0082] This step is applied to the transaction (Z), capital (M), innovation (T), and talent (G) matrices respectively to obtain Z. * Z ** M * M ** , T * , T ** , K * ,K ** .

[0083] Elements of the contribution coefficient and dependency coefficient matrix of transaction number matrix Z1: Z1 ij * =Z1 ij / Z1 Oi Z1 ij ** =Z1 ij / Z1 Ij

[0084] Contribution and dependency coefficients of the transaction target matrix Z2: Z2 ij * =Z2 ij / Z1 Oi Z2 ij ** =Z2 ij / Z1 Ij

[0085] Table 1-7 Transaction Flow Contribution Coefficient Matrix Unit: Dimensionless

[0086]

[0087] Table 1-8 Transaction Flow Dependency Coefficient Matrix Unit: Dimensionless

[0088]

[0089] Table 1-9 Transaction Target Contribution Coefficient Matrix Unit: Dimensionless

[0090]

[0091] Table 1-10 Transaction Target Dependency Coefficient Matrix Unit: Dimensionless

[0092]

[0093] Contribution and dependence coefficients of capital flow matrix M. Matrix elements: M ij * =M ij / M Oi M ij ** =M ij / M Ij (Same as above)

[0094] Contribution and dependency coefficients of the patent flow matrix P: Matrix elements: P ij * =P ij / P Oi P ij ** =P ij / P Ij (Same as above)

[0095] Copyright traffic matrix W: Contribution coefficient and dependency coefficient matrix elements: W ij * =W ij / W Oi W ij ** =W ij / W Ij (Same as above)

[0096] Contribution and dependency coefficients of the standard flow matrix S. Matrix elements: S ij * =S ij / S Oi S ij ** =S ij / S Ij (Same as above)

[0097] Contribution and dependency coefficients of talent flow matrix K. Matrix elements: K ij * =K ij / K Oi K ij ** =K ij / K Ij (Same as above)

[0098] Industry transaction contribution and dependency matrix Z * and Z ** Z * =αZ1 * +βZ2 * The dependency coefficient matrix Z = αZ1 ** +βZ2 **α and β are determined by different industry attributes; because the number of transactions better reflects the intensity of enterprise transactions, they are given a higher weight; while the supplier-customer relationship is counted as 1 as long as there is a transaction relationship between enterprises by default, mainly to make up for the lack of effective data on the number of transactions, so it should not be given too high a weight.

[0099] Industry innovation contribution and dependency matrix T * and T ** It is composed of three sub-networks: patents, software copyrights, and standards: T * =b1P * +b2W * +b3S * T ** =b1P ** +b2W ** +b3S ** The weights are set according to the relative importance of the three types of intellectual property.

[0100] Table 1-11 Innovation Dependence Coefficient Matrix Unit: Dimensionless

[0101]

[0102] Table 1-12 Innovation Contribution Coefficient Matrix Unit: Dimensionless

[0103]

[0104] S4: Semantic Knowledge Embedding

[0105] The industry semantic embedding matrix G consists of 0s and 1s. If G ij =1, then there are upstream and downstream relationships between industry i and j, such as supply chain, value chain, and product chain, with j being the downstream of i. Specifically, the lithium industry chain knowledge graph is transformed into a network structure, embedding "industry knowledge and experience" into the network; semantics extracted from the knowledge graph and the actual experience of industry experts are used to jointly define the "upstream and downstream" relationships between any two nodes in the first, second, and third levels of the industry. For example, lithium carbonate is related to all industries in the transaction, capital, and innovation matrix, but from the industry consensus perspective, it is the basic raw material for the preparation of secondary lithium salts and has the highest correlation with lithium hydroxide, lithium bromide, lithium manganese oxide, and lithium iron phosphate, requiring artificial enhancement of the upstream and downstream relationships within the industry.

[0106] S5: Adaptive Matrix Fusion

[0107] The weights of each submatrix are determined using an objective weighting method based on information entropy (entropy weighting method), as follows:

[0108] ① Calculate the entropy value ek: For each coefficient matrix to be fused (e.g., Z...) * Z ** M * M** T * T ** K * K ** (G), treating it as an evaluation metric, and calculating its information entropy. The smaller the entropy value, the greater the difference in the matrix values, the more information it provides, and the greater the weight should be.

[0109] ② Calculate the weight base value ωk: ωk = 1-ek.

[0110] ③ Introduce network topology adjustment: Adjust the weights using the average node degree Dk (a measure of the density of network connections) of the network corresponding to each submatrix: ωk' = ωk × Dk.

[0111] ④ Introduce subjective importance adjustment: For a specific matrix (such as the transaction matrix Z), it can be multiplied by an importance factor Ik (such as Ik=1.1), and the final weight αk = ωk' × Ik.

[0112] Finally, the formula for calculating the composite matrix is:

[0113] L * (T)=a1Z * (T)+a2M * (T)+a3T * (T)+a4K * (T)+a5G(T)(α=W×D×I)(Equation 1-4)

[0114] L ** (T)=a1Z ** (T)+a2M ** (T)+a3T ** (T)+a4K ** (T)+a5G(T)(α=W×D×I)(Equation 1-5)

[0115] Table 1-13 Direct Contribution Coefficient Matrix Unit: Dimensionless

[0116]

[0117] Table 1-14 Direct Dependency Coefficient Matrix (Unit: Dimensionless)

[0118]

[0119] S6: Full Association Calculation

[0120] ① Calculate the forward total contribution correlation coefficient matrix R *This matrix measures the sum of the direct and indirect pulling effects of a change in final demand in industry i on industry j through forward supply relationships. Its mathematical essence is a composite contribution coefficient matrix L. * The series sum:

[0121] R * = L * + L * ·L * + L * ·L * ·L * + ... = (I - L * ) -1 - I (where I is the identity matrix) (Equation 1-6)

[0122] Formula 1-6 R * (T) = Σ k=1 n R * ik (T) · L * kj (T) + L * ij (T) is its iterative calculation form.

[0123] ② Calculate the backward full dependency correlation coefficient matrix R ** This matrix measures the sum of the direct and indirect dependency effects of a change in supply in industry j on another industry i through backward demand relationships. Its mathematical essence is the composite dependency coefficient matrix L. ** The series sum:

[0124] R ** = L ** + L ** ·L ** + L ** ·L ** ·L ** + ... = (I - L ** ) -1 - I (Formula 1-7)

[0125] Formula 1-7 R ** (T) = Σ k=1 n L ** ik (T) · R ** kj (T) + L ** ij (T) is its iterative calculation form.

[0126] ③ Synthesize the overall correlation matrix R: Add the forward and backward influences to obtain the overall correlation matrix.

[0127] R = R * + (R ** ) T (Equation 1-8)

[0128] Where T denotes matrix transpose. The elements R in matrix R... ij This comprehensively reflects the strength of the correlation between industry i and industry j through all paths and dimensions. Taking the lithium primary industry as an example, the full correlation matrix is ​​shown below:

[0129] Table 1-15 First-level industry correlation in the lithium industry in 2024 (Unit: Dimensionless)

[0130]

[0131] S7: Data Visualization

[0132] This step aims to build an efficient and intuitive visualization and analysis interface for industry correlation matrices. Its core is a deeply customized development based on the powerful open-source visualization library Apache Echarts. This platform will completely change the traditional static analysis mode of two-dimensional data tables, providing users with a dynamic and interactive data exploration experience through three core views.

[0133] First, the heatmap view will serve as the core display method for matrix data. The system will allow users to freely select the year (e.g., 2010-2022) and industry level (first, second, and third-level categories) via drop-down menus. After the front-end obtains the corresponding JSON-formatted matrix data by calling the back-end RESTful API interface, Echarts will render a color-mapped matrix chart. The X and Y axes represent the target and source industries, respectively, and the color depth of each cell precisely corresponds to the magnitude of the correlation coefficient. Users can quickly view the specific values ​​between any two industries by hovering the mouse over them (tooltip) and can click on row or column labels to highlight all relevant correlations, quickly identifying key influencing factors.

[0134] Secondly, the network diagram view will employ a force-directed graph model to dynamically reveal the complex topological relationships between industries. In this view, each industry sector becomes a draggable node, and the thickness of the lines connecting nodes represents the strength of the association. The system will incorporate intelligent algorithms to automatically cluster and arrange industry sectors, allowing users to interact by dragging, zooming, and clicking. Clicking on a specific node will highlight all its directly related industries and hide secondary connections, enabling users to clearly understand core industry clusters and key pivot industries, and comprehend the overall structure and vulnerabilities of the economic system.

[0135] Finally, the time series view focuses on revealing the dynamic evolution of industrial linkages. This view provides an intuitive timeline slider, allowing users to select multiple industry sectors or linkages of interest. As the timeline is slid, the chart dynamically displays the coefficient changes or popularity evolution of these selected objects over many years, supporting cross-sectional comparative analysis. This feature is extremely valuable for assessing the long-term effects of industrial policies, identifying emerging leading industries, and predicting future changes in economic structure.

[0136] The entire platform will adopt a responsive design to ensure good display on different devices, and through a modular front-end architecture (such as Vue components) and a clear back-end API design, it will ensure the maintainability and scalability of the system, ultimately becoming a powerful data visualization tool to support macro-level decision-making.

Claims

1. A method for measuring industrial relevance based on multi-source data fusion, characterized in that, The method includes the following steps: Data collection steps: Collect multidimensional heterogeneous open-source data from the Internet and third-party databases, including transaction data, equity data, innovation data, and talent data; Matrix construction steps: Construct a transaction flow matrix Z based on the transaction data, construct a capital flow matrix M based on the equity data, construct an innovation flow matrix T based on the innovation data, and construct a talent flow matrix K based on the talent data; Coefficient calculation steps: Calculate the contribution coefficient matrix Z corresponding to the transaction flow matrix Z. * and dependency coefficient matrix Z ** Calculate the contribution coefficient matrix M corresponding to the capital flow matrix M. * and dependency coefficient matrix M ** Calculate the contribution coefficient matrix T corresponding to the innovation flow matrix T. * and dependency coefficient matrix T ** Calculate the contribution coefficient matrix K corresponding to the talent flow matrix K. * and dependency coefficient matrix K ** ; Semantic embedding steps: Construct a semantic embedding matrix G, which is a binary matrix used to represent the upstream and downstream semantic relationships between industries based on knowledge graphs; Fusion steps: Use the adjustment coefficient entropy weight method to calculate the contribution coefficient matrix Z. * M * T * K * The semantic embedding matrix G is weighted and then fused to obtain the industry composite contribution matrix L. * ; is the dependency coefficient matrix Z ** M ** T ** K ** The semantic embedding matrix G is weighted and then fused to obtain the industry composite dependency matrix L. ** ; Correlation calculation steps: Based on the aforementioned industry composite contribution matrix L * The forward total contribution correlation coefficient matrix R is obtained by summing the matrix series. * Based on the aforementioned industry composite dependency matrix L ** The backward full dependency correlation coefficient matrix R is obtained by calculating the sum of matrix series. ** ; Total correlation synthesis step: The forward total contribution correlation coefficient moments R... * With the backward full dependency correlation coefficient matrix R ** The transposes of the matrices are added together to obtain the total industry correlation matrix R.

2. The method as described in claim 1, characterized in that, The transaction flow matrix Z is obtained by fusing the transaction count matrix Z1 and the transaction target count matrix Z2. The fusion formula is: Z = αZ1 + βZ2, where α and β are preset weight coefficients. The elements in matrices Z1 and Z2 are weighted and accumulated using a time exponential decay function to reflect the time sensitivity of the data.

3. The method as described in claim 1, characterized in that, The contribution coefficient matrix Z * element Z ij * The calculation formula is: Z ij * =Z ij / Z oi ,in (n is the total number of industries), representing the total output of industry i; the dependency coefficient matrix Z ** Medium element Z ij ** The calculation formula is: Z ij ** =Z ij / Z Ij ,in (n is the total number of industries), representing the total input of industry j; the contribution coefficient matrix and dependence coefficient matrix corresponding to the capital flow matrix M, innovation flow matrix T, and talent flow matrix K are all calculated according to this rule.

4. The method as described in claim 1, characterized in that, The adjustment coefficient entropy weight method includes the following sub-steps: Calculate the entropy: For the matrix to be weighted, calculate its information entropy. , where p is the matrix element normalization; Calculate initial weights: Calculate initial weights based on entropy values. , where m is the total number of matrix elements; Network density adjustment: Calculate the average degree D of the network corresponding to this matrix, and adjust the initial weights to W' = W×D; Importance adjustment: Introduce a preset importance factor I to obtain the final comprehensive weight α = W'×I.

5. The method as described in claim 4, characterized in that, The industrial composite contribution matrix L * It is obtained through linear weighted fusion, and its calculation formula is: L * =a1Z * +a2M * +a3T * +a4K * +a5G, where α1 to a5 are the comprehensive weights determined for each matrix by the aforementioned adjustment coefficient entropy weighting method; the industry composite dependency matrix L ** The calculation formula is L ** =a1Z ** +a2M ** +a3T ** +a4K ** +a5G.

6. The method as described in claim 1, characterized in that, The forward full contribution correlation coefficient matrix R * It is obtained by solving the matrix series, and its calculation formula is as follows: Where I is the identity matrix; the backward full dependency correlation coefficient matrix R ** The calculation formula is: .

7. The method as described in claim 1, characterized in that, The innovation flow matrix T is a linear superposition of the patent cooperation matrix P, the software copyright cooperation matrix W, and the technology standard cooperation matrix S, with the formula: T = b1P + b2W + b3S Among them, b1, b2, and b3 are weighting coefficients pre-set based on the importance of the three types of innovation activities and the quality of data.

8. The method as described in claim 1, characterized in that, The construction process of the semantic embedding matrix G includes: extracting upstream and downstream relationships between industry nodes based on a predefined industry chain knowledge graph; if industry i is an upstream supplier of industry j, then G is defined. {ij} = 1, otherwise 0; the industry node relationships in the knowledge graph are jointly defined by domain expert knowledge and industry topic terms generated by the large language model.

9. An industry correlation measurement system based on multi-source data fusion, used to implement the method of any one of claims 1-8, characterized in that, The system includes: The data acquisition module is used to collect enterprise data from multiple source databases; The knowledge modeling module is used to construct a knowledge graph of the industry chain and realize the mapping between enterprises and industry entities; The manual proofreading module provides an interactive interface for experts to correct the automatically labeled results. The industry correlation calculation module is used to construct various flow matrices, calculate coefficient matrices, and fusion matrices; The results display module is used to visually display the industry association matrix.

10. The system as described in claim 9, characterized in that, The knowledge modeling module uses the BERT model to generate vector representations of industry keywords and enterprise documents, and uses ElasticSearch's cosine similarity search function to achieve accurate matching between enterprises and subdivided industry nodes.

Images