Innovative resource recommendation method and system based on industrial chain knowledge graph

Abstract

The invention discloses an innovation resource recommendation method and system based on an industrial chain knowledge graph, and relates to the technical field of industrial innovation decision support, and the method comprises the steps: collecting multi-source industrial chain innovation ecological data, and carrying out the preprocessing; performing entity recognition and relation extraction on the preprocessed multi-source industrial chain innovation ecological data, and constructing a static industrial chain knowledge graph; performing policy and regulation information analysis on the dynamic industry chain knowledge graph, extracting mandatory compliance terms and converting the mandatory compliance terms into a filtering rule set, filtering entity nodes and relation edges in the dynamic industry chain knowledge graph by using the filtering rule set, identifying encouraging resources, endowing policies with favorable identifiers, and forming a security resource pool. According to the method, entity nodes and relation edges in the dynamic industry chain knowledge graph are filtered to form a security resource pool, so that deep fusion of compliance screening and opportunity discovery is realized, and commercial feasibility and strategic guidance of a recommendation result are ensured.

Description

Technical Field

[0001] This invention relates to the field of industrial innovation decision support technology, and in particular to an innovation resource recommendation method and system based on industrial chain knowledge graph. Background Technology

[0002] In the field of technology transfer and innovation resource recommendation, traditional methods mainly rely on constructing knowledge graphs to integrate multi-source heterogeneous data such as patents, academic papers, corporate information, and policies and regulations, thereby forming a static description of the industrial chain structure. Traditional methods aim to provide support for information query and preliminary matching by depicting the network of connections between technology, industry, and innovation entities, and provide visualization tools for understanding complex innovation ecosystems. This is an important stage in the digitalization process of resource recommendation.

[0003] However, traditional methods still have certain limitations; the constructed knowledge graphs are usually static, and the weights of entity nodes and relationship edges fail to be effectively linked with the real-time data stream of continuously updated business events and policy adjustments, resulting in a delay in response to market dynamics and policy adjustments; traditional methods often treat compliance requirements as an independent filtering link and fail to deeply internalize them into the recommendation logic, affecting the accuracy and compliance of decision-making. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an innovative resource recommendation method based on industry chain knowledge graph to solve the problems of insufficient dynamic adaptability and poor decision-making accuracy caused by static graphs and externalized compliance rules.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, this invention provides an innovation resource recommendation method based on an industry chain knowledge graph. The method includes: collecting and preprocessing multi-source industry chain innovation ecosystem data; performing entity identification and relationship extraction on the preprocessed multi-source industry chain innovation ecosystem data to construct a static industry chain knowledge graph; dynamically adjusting the weights of entity nodes and relationship edges in the static industry chain knowledge graph based on continuously updated preprocessed multi-source industry chain innovation ecosystem data to generate a dynamic industry chain knowledge graph; parsing policy and regulatory information in the dynamic industry chain knowledge graph, extracting mandatory compliance clauses and converting them into a filtering rule set; using the filtering rule set to filter entity nodes and relationship edges in the dynamic industry chain knowledge graph, identifying encouraged resources and assigning policy benefit labels to them, forming a secure resource pool; collecting user query requests, performing comprehensive ranking calculations in the secure resource pool based on the dynamic weights of entity nodes and relationship edges, policy benefit labels, and the matching degree of user query requests, and outputting a business decision support report.

[0007] As a preferred embodiment of the innovation resource recommendation method based on the industrial chain knowledge graph described in this invention, the multi-source industrial chain innovation ecosystem data includes patent data, policy and regulatory data, enterprise operation data, market intelligence data, and academic research data. The preprocessing includes data cleaning, data denoising, and formatting.

[0008] As a preferred embodiment of the innovation resource recommendation method based on industry chain knowledge graph described in this invention, the steps for constructing a static industry chain knowledge graph are as follows: Scan and match patterns on the preprocessed multi-source industrial chain innovation ecosystem data to identify entity names and generate a set of entity names. Using syntactic dependency parsing and co-occurrence statistics, semantic relations between entities in the entity name set are extracted to generate a set of relation triples. A static industry chain knowledge graph is constructed by using the set of entity names as nodes and the set of relation triples as edges.

[0009] As a preferred embodiment of the innovation resource recommendation method based on industry chain knowledge graph described in this invention, the steps for generating a dynamic industry chain knowledge graph are as follows: Continuously monitor newly incoming preprocessed multi-source industrial chain innovation ecosystem data, identify and extract data fragments containing event information, and generate a set of event data fragments to be processed; Perform event feature analysis and classification on a set of event data fragments to be processed, and generate event classification results; The event classification results are analyzed to assess the scope and intensity of impact, and a list of event impacts is generated. Based on the event impact list, identify the entity nodes and relationship edges whose weights need to be adjusted, calculate the weight adjustment amount, and generate a weight adjustment instruction set; Execute the weight adjustment instruction set to update the weights of corresponding entity nodes and relationship edges in the static industry chain knowledge graph, and generate a dynamic industry chain knowledge graph.

[0010] As a preferred embodiment of the innovation resource recommendation method based on industry chain knowledge graphs described in this invention, the steps of parsing policy and regulatory information from the dynamic industry chain knowledge graph, extracting mandatory compliance clauses, and converting them into a set of filtering rules are as follows: Extract text content and integrate related information from entity nodes related to policies and regulations in the dynamic industrial chain knowledge graph to generate a set of policy clause information; Semantic analysis is performed on each policy clause in the policy clause information set to identify clauses containing mandatory requirements and generate a list of mandatory compliance clauses. The list of mandatory compliance clauses is transformed into conditional judgment logic statements that can be executed on a dynamic industry chain knowledge graph, creating a set of filtering rules.

[0011] As a preferred embodiment of the innovation resource recommendation method based on industry chain knowledge graphs described in this invention, the steps for forming a secure resource pool are as follows: The filtering rule set is used to match and verify the entity nodes and relationship edges in the dynamic industrial chain knowledge graph, and resources that meet the mandatory requirements are selected to generate a preliminary filtering result set. Based on the policy clause information set, resources that meet the mandatory requirements in the preliminary filtering result set are subjected to policy orientation matching analysis and policy benefit labels are added to form a safe resource pool.

[0012] As a preferred embodiment of the innovation resource recommendation method based on industry chain knowledge graph described in this invention, the following steps are taken: Collecting user query requests and performing a comprehensive ranking calculation in a secure resource pool based on the dynamic weights of entity nodes and relationship edges, policy benefit identifiers, and the matching degree of user query requests. Collect user query requests and perform semantic parsing and query element extraction to generate semantic query objects; The semantic query object is matched and calculated with the entity nodes in the security resource pool in multiple dimensions to generate a preliminary matching score set. A comprehensive score calculation and ranking of each security resource in the initial matching score set is performed to generate a personalized innovative resource recommendation list.

[0013] As a preferred embodiment of the innovation resource recommendation method based on industry chain knowledge graph described in this invention, the steps of performing multi-dimensional matching calculations between semantic query objects and entity nodes in the security resource pool to generate a preliminary matching score set are as follows: Extract query elements from the semantic query object and match them with entity nodes in the security resource pool to generate a multi-dimensional matching element set. The set of multi-dimensional matching elements is compared with the text content, dynamic weights and policy benefit identifiers of entity nodes to calculate similarity, and a preliminary matching score set is generated.

[0014] As a preferred embodiment of the innovation resource recommendation method based on the industry chain knowledge graph described in this invention, the business decision support report is generated by extracting the dynamic weight, policy benefit identifier, and comprehensive score of each security resource from the personalized innovation resource recommendation list, and conducting a multi-dimensional business value assessment.

[0015] Secondly, the present invention provides an innovation resource recommendation system based on the knowledge graph of the industrial chain, including a data acquisition module for collecting multi-source industrial chain innovation ecosystem data and performing preprocessing. The graph construction module is used to perform entity recognition and relationship extraction on preprocessed multi-source industrial chain innovation ecosystem data to construct a static industrial chain knowledge graph. The graph dynamic update module is used to dynamically adjust the weights of entity nodes and relationship edges in the static industrial chain knowledge graph based on the continuously updated preprocessed multi-source industrial chain innovation ecosystem data, and generate a dynamic industrial chain knowledge graph. The resource filtering module is used to analyze policy and regulatory information in the dynamic industrial chain knowledge graph, extract mandatory compliance clauses and convert them into a set of filtering rules. The set of filtering rules is used to filter entity nodes and relationship edges in the dynamic industrial chain knowledge graph, identify encouraged resources and assign them policy benefit labels to form a safe resource pool. The resource recommendation module collects user query requests and performs comprehensive ranking calculations in the secure resource pool based on the dynamic weights of entity nodes and relation edges, policy benefit indicators, and the matching degree of user query requests, and outputs a business decision support report.

[0016] The beneficial effects of this invention are as follows: By dynamically adjusting the weights of entity nodes and relationship edges in the static industrial chain knowledge graph, a dynamic industrial chain knowledge graph is generated, which realizes real-time perception and quantitative response to changes in market popularity and policy orientation, and improves the timeliness and foresight of recommended resources; by filtering entity nodes and relationship edges in the dynamic industrial chain knowledge graph, a safe resource pool is formed, which realizes the deep integration of compliance screening and opportunity discovery, and can ensure the commercial feasibility and strategic guidance of the recommendation results. Attached Figure Description

[0017] 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.

[0018] Figure 1 This is a flowchart of an innovation resource recommendation method based on industry chain knowledge graphs.

[0019] Figure 2 This is a schematic diagram of an innovation resource recommendation system based on an industry chain knowledge graph.

[0020] Figure 3 The flowchart for entity relation extraction.

[0021] Figure 4The flowchart for generating the filtering rule set. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an innovation resource recommendation method based on an industry chain knowledge graph, including the following steps: S1. Collect multi-source industrial chain innovation ecosystem data and perform preprocessing.

[0026] S1.1 Multi-source industrial chain innovation ecosystem data includes patent data, policy and regulatory data, enterprise operation data, market intelligence data, and academic research data.

[0027] It should be noted that patent data refers to authorized patents, patent application documents, and legal status information published by patent offices of various countries, obtained through API interfaces accessing commercial patent databases; policy and regulatory data refers to industrial policies, laws and regulations, industry standards, and normative documents issued by governments and regulatory agencies at all levels; enterprise operation data refers to information reflecting the operating status of enterprises, including business registration, annual reports, bidding, product launches, and financing dynamics, obtained through enterprise information query platforms, stock exchange announcements, and public reports in the news media; market intelligence data refers to data reflecting market supply and demand, competitive landscape, and technological trends, such as industry research reports, market analysis, price indices, and competitor dynamics, obtained through subscriptions to market research institution reports and industry information platforms; and academic research data refers to research results generated by the academic community, including journal articles, conference papers, dissertations, and research project information, obtained through interfaces of academic databases.

[0028] S1.2 Preprocessing includes data cleaning, data denoising, and formatting.

[0029] It should be noted that data cleaning refers to identifying and correcting format errors, duplicate records, and missing fields in the collected multi-source industrial chain innovation ecosystem data to ensure data integrity and consistency; data denoising refers to identifying and filtering irrelevant information, advertising content, HTML tags, and special characters in the collected multi-source industrial chain innovation ecosystem data to eliminate interference components; formatting refers to structural transformation and field mapping of data from different sources with heterogeneous formats in the collected multi-source industrial chain innovation ecosystem data according to unified specifications and standards: for patent data, the original field patent_title is mapped to the target field patent title, and if the field is empty, it is automatically filled with "unknown title"; for policy data, the original field pub_date is unified to YYYY-MM-DD format through format conversion rules, and if an unparseable date format is encountered, the original value is retained and an exception mark is added; for the reg_capital field in enterprise data, the unit is uniformly converted to "ten thousand yuan" during mapping, and if the field value contains non-numeric characters, the numeric part is automatically truncated and logged, thus achieving standardization of multi-source data.

[0030] S2. Entity identification and relationship extraction are performed on the preprocessed multi-source industrial chain innovation ecosystem data to construct a static industrial chain knowledge graph.

[0031] S2.1 Scan and pattern match the preprocessed multi-source industrial chain innovation ecosystem data, identify entity names, and generate a set of entity names.

[0032] Furthermore, based on a predefined entity dictionary, the text content in the preprocessed multi-source industrial chain innovation ecosystem data is scanned line by line. Specific patterns (such as organization codes and patent number formats) are matched using regular expressions, and candidate strings are segmented from the text content using a maximum positive matching algorithm. The similarity between the candidate strings and standard entries in the entity dictionary is calculated and verified (e.g., company names must contain the keywords "company" or "group," and technical terms must conform to the definitions in the professional terminology database). Strings that conform to the naming conventions of company names, technical terms, patent names, and policy and regulatory names are selected. All successfully identified strings are then aggregated to generate a set of entity names.

[0033] It should be noted that the entity dictionary is formed by extracting data and collecting terms from the National Economic Industry Classification, the International Patent Classification, and industry chain-related databases, and then cleaning and classifying them through string standardization and term filtering. It includes company names, technical terms, patent names, and policy and regulation names.

[0034] S2.2 Using syntactic dependency parsing and co-occurrence statistics, extract the semantic relationships between entities in the entity name set and generate a set of relation triples.

[0035] Furthermore, syntactic dependency analysis is performed on the contextual statements in the entity name set to parse out the syntactic structural relationships (such as subject-verb-object, attributive-head-adverbial) between entities. In the preprocessed multi-source industrial chain innovation ecosystem data, the co-occurrence frequency between pairs of entities in the entity name set is counted (the context window size is set to sentence units, and the word distance is defined as the number of words between two entities, with adjacent words counted as 1 and n words apart counted as n+1), and the co-occurrence strength is calculated. The syntactic structural relationship and the co-occurrence strength are fused. When a pair of entities has a specific syntactic structural relationship and the co-occurrence strength exceeds the preset co-occurrence strength threshold, it is determined that there is a valid semantic structural relationship between the entity pair. All the determined semantic structural relationships are standardized in the form of "entity-relationship-entity" to generate a set of relation triples.

[0036] The expression for calculating the co-occurrence intensity is: ; in, Co-occurrence strength represents the strength of the association between entities; Co-occurrence frequency represents the total number of times an entity pair co-occurs within a specific context window (such as a sentence or paragraph) across the entire corpus; It is the average co-occurrence distance, which represents the average number of words between two entities across all co-occurrence windows; It is a smoothing factor, a constant parameter used to prevent the denominator from being zero and to weaken the influence of accidental co-occurrence. An example initial value is set to 1.0. The combined effect of different values ​​(such as 0.1, 0.5, 1.0, 1.5, 2.0) in the range of 0.1 to 2.0 on the precision and recall of relation extraction is tested using a grid search method (evaluated using F1 score). The specific value corresponding to the highest F1 score is selected as the official value.

[0037] It should be noted that the co-occurrence intensity threshold is obtained by calculating the co-occurrence intensity of all entity pairs in the preprocessed multi-source industrial chain innovation ecosystem data to form a co-occurrence intensity distribution, and determining the critical point of the co-occurrence intensity distribution through statistical methods (such as the 75th percentile calculation), and setting it in the middle to high range of the distribution. An exemplary value range is [0.6, 0.9]. If it is higher than 0.9, it will lead to overly strict filtering conditions and missed reports of valid associations. If it is lower than 0.6, it will lead to overly wide filtering conditions and introduce a large amount of noisy associations.

[0038] S2.3. Construct a static industry chain knowledge graph by using the set of entity names as nodes and the set of relation triples as edges.

[0039] Furthermore, each entity name in the entity name set is instantiated as a node with a unique identifier; each "subject-verb-object" triple in the relation triple set is instantiated as a directed edge, connecting the node corresponding to the subject entity name to the node corresponding to the object entity name, and the verb is used as the type attribute of the directed edge; all nodes created from the entity name set and edges created from the relation triple set are associated and integrated to form a static industry chain knowledge graph.

[0040] S3. Based on the continuously updated preprocessed multi-source industrial chain innovation ecosystem data, dynamically adjust the weights of entity nodes and relationship edges in the static industrial chain knowledge graph to generate a dynamic industrial chain knowledge graph.

[0041] S3.1 Continuously monitor newly incoming preprocessed multi-source industrial chain innovation ecosystem data, identify and extract data fragments containing event information, and generate a set of event data fragments to be processed.

[0042] Furthermore, the text content in the continuously flowing preprocessed multi-source industrial chain innovation ecosystem data is scanned in real time. By analyzing the pattern features of verb phrases and temporal markers (such as dates and time adverbs) in the text content, potential event descriptions are identified (e.g., when the text content contains "Company A released a new generation of water treatment technology B in 2023", it is identified as a potential event of "technology release" by identifying the verb phrase "released" and the temporal marker "2023"). When a text pattern containing actions and a clear time context is identified, a complete sentence or semantically coherent paragraph that matches the text pattern is extracted as a candidate event data fragment. All extracted candidate event data fragments are collected to generate a set of event data fragments to be processed.

[0043] S3.2 Perform event feature analysis and classification on the set of event data fragments to be processed, and generate event classification results.

[0044] Furthermore, each event data fragment in the set of event data fragments to be processed undergoes dependency parsing and semantic role labeling to identify and extract predicates and associated agent and patient semantic roles in the sentences, forming event element features that include action type, action subject, and target. The entities identified in the event data fragments to be processed are associated with corresponding nodes in the static industry chain knowledge graph to determine the industry links involved in the event (such as R&D, production, and sales). The event element features and industry links are matched with preset typical event templates to obtain the optimal matching pattern. Based on the optimal matching pattern, each event data fragment to be processed is labeled with the corresponding event type (such as policy release, technology financing, and corporate cooperation), generating event classification results.

[0045] It should be noted that the typical event template is made by extracting a large number of event instances and feature elements from historical news, announcements and reports, and automatically classifying event instances with similar feature combinations through clustering algorithms (such as K-means). The most representative event element pattern is extracted from each cluster as the typical event template, which includes constraints on the types of action subjects, action types and target types.

[0046] S3.3 Analyze the event content of the event classification results, assess the scope and intensity of the impact, and generate an event impact list.

[0047] Furthermore, based on each classified event in the event classification results, the entities associated with the corresponding event data fragments (such as involved enterprises, technologies, and industrial chain links) are extracted. By querying the static industrial chain knowledge graph, all entity nodes directly connected to the entities through upstream and downstream relationships, competitive relationships, and collaborative relationships are identified. The number of different industrial chain links affected is counted to obtain the breadth index of affected industrial chain links. The type attributes of the events are extracted, such as the level and binding strength of the issuing agency of policies and regulations, the disclosed amount of financing events, and the technology maturity level of technology events. Combined with the node centrality index of the event subject in the static industrial chain knowledge graph, the type attributes are quantified into a comprehensive impact intensity score through a weighted calculation formula. The breadth index of affected industrial chain links and the comprehensive impact intensity score of each event are combined to generate an event impact list.

[0048] The expression for calculating the comprehensive impact intensity score is: ; in, It is a comprehensive impact intensity score; It refers to the level of the issuing agency, such as district level = 1, city level = 2, provincial level = 3, ministerial level = 4, and national level = 5; It refers to the binding force, indicating the degree of mandatory nature of policy provisions, such as encouragement = 1, recommendation = 2, and mandatory = 3; It is the amount to be disclosed; It is a technology maturity level, which indicates the maturity level of the relevant technology, such as laboratory stage = 1, pilot production = 2, industrialization = 3, large-scale production = 4, and phase-out = 5; It is the Min-Max normalization function; , , and The weighting coefficients representing the issuing organization level, binding force strength, disclosure amount, and technology maturity level are obtained by collecting historical event data and calculating the Pearson correlation coefficient between each event attribute (such as issuing organization level, binding force strength, disclosure amount, and technology maturity level) and the actual impact of the event. This yields an initial importance measure for each attribute, which is then adjusted through normalization to meet the following requirements. + + + =1, ensuring that the relative importance of each attribute in the overall score is objective and reasonable.

[0049] S3.4. Based on the event impact list, determine the entity nodes and relationship edges that need to have their weights adjusted, calculate the weight adjustment amount, and generate a weight adjustment instruction set.

[0050] Furthermore, each event entry in the event influence list is traversed, and based on the different industry chain links affected, the corresponding entity nodes in the static industry chain knowledge graph are marked as target entity nodes whose weights need to be adjusted, and the direct relationship edges between target entity nodes are marked as target relationship edges whose weights need to be adjusted. Based on the comprehensive influence intensity score of the event entries in the event influence list, the weight adjustment amount is calculated, and an operation instruction containing a unique identifier and the corresponding weight adjustment amount is generated for each target entity node and target relationship edge, and a weight adjustment instruction set is compiled.

[0051] The expression for calculating the weight adjustment is: ; in, It is the weight adjustment amount; It is the overall impact strength score of the event item; It is the current weight value of the target entity node or relation edge; It is the current maximum weight value of all entity nodes or relation edges; It is the adjustment coefficient for the current weight influence; It is a moderating coefficient for the influence of online dissemination, used to adjust the degree of influence of "the online dissemination effect of the event's impact" on the weight adjustment amount; It is a control factor affecting the decay rate; It is the base of the natural logarithm; It is the shortest path distance from the node to the target node in the static industrial chain knowledge graph that the event directly affects (the shortest path hop count between two nodes in the static industrial chain knowledge graph (unweighted graph), the distance between adjacent nodes is 1).

[0052] It should be noted that, , and This involves initializing a set of parameter values ​​on a knowledge graph dataset of historical events and their corresponding industry chains, and then using a grid search method within a fixed range (e.g., ...). , Between 0 and 1, Multiple parameter combinations are generated (between 1 and 5), and each parameter combination is applied to historical events to simulate the effect of adjusting the weights of the static industry chain knowledge graph. The recommendation accuracy of the dynamic industry chain knowledge graph after weight adjustment is used as the evaluation index, and the parameter combination with the highest recommendation accuracy is selected as the final setting value.

[0053] S3.5 Execute the weight adjustment instruction set to update the weights of the corresponding entity nodes and relationship edges in the static industry chain knowledge graph and generate a dynamic industry chain knowledge graph.

[0054] Furthermore, each operation instruction in the weight adjustment instruction set is read sequentially, and the unique identifier of the target entity node or target relationship edge and the corresponding weight adjustment amount are extracted. Based on the unique identifier, the specific entity node or relationship edge is located in the static industry chain knowledge graph, and the current weight value is combined with the weight adjustment amount to obtain the updated weight value. The weight values ​​of all target entity nodes and target relationship edges in the static industry chain knowledge graph are replaced with the corresponding updated weight values ​​to generate a dynamic industry chain knowledge graph.

[0055] It should be noted that batch updates of multiple events are processed sequentially according to the time of occurrence of the events. The weight update values ​​of the same entity are accumulated, but the upper limit of the accumulated weight is truncated to 1.0. After all updates are completed, all weights of the dynamic industrial chain knowledge graph are normalized by maximum-min to maintain the relative stability of the weight distribution.

[0056] S4. Analyze policy and regulatory information in the dynamic industrial chain knowledge graph, extract mandatory compliance clauses and transform them into a set of filtering rules. Use the set of filtering rules to filter entity nodes and relationship edges in the dynamic industrial chain knowledge graph, identify encouraged resources and assign them policy benefit labels to form a safe resource pool.

[0057] S4.1 Extract text content and integrate related information from entity nodes related to policies and regulations in the dynamic industrial chain knowledge graph to generate a set of policy clause information.

[0058] Furthermore, the system traverses all entity nodes of type "policy and regulation" in the dynamic industry chain knowledge graph, accesses and extracts the corresponding stored text content (such as policy title, main text, and detailed clauses) of the entity nodes; queries the relationship edges directly connected to the policy and regulation entity nodes in the dynamic industry chain knowledge graph, locates and extracts the supplementary text of related attachments, interpretation documents, and revision instructions; concatenates and deduplicates the text content of each policy and regulation entity node with the related supplementary text to form a policy document; and collects the policy documents corresponding to all policy and regulation entity nodes to generate a set of policy clause information.

[0059] S4.2 Perform semantic analysis on each policy clause in the policy clause information set, identify clauses containing mandatory requirements, and generate a list of mandatory compliance clauses.

[0060] Furthermore, the syntactic structure and semantic relationships of each policy clause text in the policy clause information set are analyzed to identify the predicate verbs, modal verbs (such as "must," "shall," "prohibited," "must not"), negative words, and the objects or complements they govern. If the semantic structure of the policy clause text contains modal verbs expressing "obligation" or "prohibition," and clearly indicates the actor, the bound behavior or state, and the potential negative consequences (such as "punishment" or "assuming responsibility"), then the policy clause is determined to contain mandatory requirements. All policy clause texts determined to contain mandatory requirements, the identified actors, the bound behaviors, and the consequences are structured, recorded, and stored to generate a list of mandatory compliance clauses.

[0061] S4.3. Transform the list of mandatory compliance clauses into conditional judgment logic statements that can be executed on the dynamic industry chain knowledge graph, and create a set of filtering rules.

[0062] Furthermore, the types of actors (e.g., "high water-consuming enterprises"), the constrained behaviors (e.g., "using specific technologies"), and the constraints (e.g., "phase-out within a specified period") are extracted from the list of mandatory compliance clauses and mapped to the entity types, attributes, and relationships in the dynamic industry chain knowledge graph. For example, the actor type "high water-consuming enterprise" is mapped to an enterprise entity with the "high water consumption" tag in the dynamic industry chain knowledge graph, and the constrained behavior "using specific technologies" is mapped to the "use" relationship between enterprises and technologies in the dynamic industry chain knowledge graph. Based on the mapping results, the constraint logic of the mandatory compliance clauses is constructed into conditional judgment statements using standardized query languages ​​(e.g., Cypher, SPARQL). The set of all such conditional judgment statements constitutes the filtering rule set.

[0063] It should be noted that the specific process of constructing conditional judgment statements is as follows: the semantic constraints of mandatory compliance clauses are transformed into standard graph database query logic, that is, using the syntax rules of Cypher or SPARQL graph query languages, the mapped entity types, attributes, and relationships are combined into a complete query pattern. The query pattern is used to accurately match entity nodes and relationship edges that violate the clause constraints in the dynamic industry chain knowledge graph. For example, the clause logic of "prohibiting high water-consuming enterprises from using outdated technology" is constructed into a query condition, the function of which is "to find all entity nodes of type 'high water-consuming enterprise' in the dynamic industry chain knowledge graph, and whether there is a relationship edge of type 'use' connected to the 'outdated technology' node". The execution result of this conditional judgment statement is the violation item. The set of all such conditional judgment statements constructed for different clauses constitutes the filtering rule set.

[0064] S4.4. Use the filtering rule set to match and verify the entity nodes and relationship edges in the dynamic industry chain knowledge graph, filter out the resources that meet the mandatory requirements, and generate a preliminary filtering result set.

[0065] Furthermore, each conditional statement in the filtering rule set is traversed and submitted as an independent query command to the graph database storing the dynamic industry chain knowledge graph for execution. After executing the query, the graph database returns a set of all entity nodes and relational edges that satisfy the pattern described by the conditional statement, i.e., the set of non-compliant resources that do not conform to the current mandatory clauses. For all entity nodes and relational edges in the dynamic industry chain knowledge graph, it is checked whether they are not included in the query results of any conditional statement in the filtering rule set, i.e., it is determined that they do not violate any mandatory clauses. All entity nodes and relational edges that pass the verification and do not violate any mandatory clauses are aggregated to generate a preliminary filtering result set.

[0066] S4.5 Based on the policy clause information set, conduct policy-oriented matching analysis on the resources that meet the mandatory requirements in the preliminary filtering result set and add policy benefit labels to form a safe resource pool.

[0067] Furthermore, dependency parsing and semantic role labeling are performed on the policy text content in the policy clause information set to identify sentences containing encouraging semantic frameworks and extract the encouraged subject (e.g., "water-saving technology"), encouragement conditions (e.g., "meeting specific standards"), and encouragement measures. Semantic similarity calculation is performed between the encouraged subject, encouragement conditions, and encouragement measures and the attribute descriptions and relational contexts of entity nodes in the preliminary filtering result set. When the similarity between the encouraged subject and the entity node type, and the matching degree between the encouragement conditions and the entity node attributes both exceed the preset semantic matching threshold, the entity node is determined to be highly consistent with the policy orientation. For entity nodes determined to be highly consistent, a policy benefit identifier is added to the dynamic industrial chain knowledge graph, and all entity nodes with policy benefit identifiers and their relationships are integrated to form a security resource pool.

[0068] It should be noted that the semantic matching threshold is determined by collecting historical matching data containing policy clauses and resource entities, labeling correct and incorrect matching pairs, calculating the semantic similarity score of all sample pairs (using a word vector model (Word2Vec) to generate text vector representations, and calculating the matching degree through cosine similarity), forming a score distribution, and simultaneously calculating the precision and recall of different classification results. The optimal critical point is determined by evaluating the F1 score, with an exemplary value range of 0.7-0.9. If it is higher than 0.9, the conditions will be too strict, and potential policy-beneficial resources will be missed. If it is lower than 0.7, the conditions will be too lenient, and resources that do not conform to the policy orientation will be misclassified as beneficial resources. The final semantic matching threshold is 0.8, which is determined based on the principle of maximizing the F1 score of historical data (corresponding to a precision of 0.85 and a recall of 0.82).

[0069] S5. Collect user query requests, perform comprehensive sorting and calculation in the security resource pool based on the dynamic weights of entity nodes and relation edges, policy benefit indicators, and the matching degree of user query requests, and output a business decision support report.

[0070] S5.1 Collect user query requests and perform semantic parsing and query element extraction to generate semantic query objects.

[0071] Furthermore, the system receives query request text input by users in natural language form and performs word segmentation, part-of-speech tagging, and dependency parsing on the query request text to identify verbs, noun phrases, and modifiers in the query request. Using semantic role labeling, the query request is classified into intent expression (composed of predicate verbs and nouns, such as "searching for technology"), target object type (composed of object noun phrases, such as "membrane separation technology"), and limiting conditions (composed of adjectives and prepositional phrases, such as "efficient"). The classification results are then encapsulated into a triple structure of "intent expression - target object type - limiting conditions" to generate a semantic query object.

[0072] S5.2 Decompose the semantic query object into query elements of each dimension that match the entity nodes in the security resource pool, and generate a multi-dimensional matching element set.

[0073] Furthermore, the intent expressed in the semantic query object is mapped to the matching requirements for the type of entity node in the security resource pool (e.g., mapping "finding technology" to matching entities of the "technology" class), the specific description of the target object type (e.g., "membrane separation") is mapped to the text matching requirements for the entity node name, and the limiting conditions (e.g., "efficient") are mapped to the matching requirements for the specific attribute value of the entity node (e.g., "efficiency level"). A multi-dimensional matching element set is then generated by organizing the matching requirements for entity type, text content, and specific attribute value respectively.

[0074] S5.3 Calculate the similarity between the multi-dimensional matching element set and the text content, dynamic weight, and policy benefit identifier of the entity node, respectively, to generate a preliminary matching score set.

[0075] Furthermore, for the text content dimension, the cosine similarity algorithm is used to calculate the similarity between the vector representations of the target object type description and the entity node name and attribute description text in the multi-dimensional matching element set, generating a text matching score. For the dynamic weight dimension, the constraints in the multi-dimensional matching element set (such as "efficient") are transformed into numerical range constraints, and the degree of conformity between the dynamic weight value of the entity node and the numerical range constraints is calculated to generate a weight matching score. For the policy benefit identifier dimension, it is checked whether there are policy-oriented constraints in the multi-dimensional matching element set. If so, a Boolean match is performed with the policy benefit identifier of the entity node to generate a policy matching score. The text matching score, weight matching score, and policy matching score corresponding to each entity node are combined into a preliminary matching score, and the preliminary matching scores of each entity node are integrated to generate a preliminary matching score set.

[0076] S5.4 Perform comprehensive scoring and ranking of each security resource in the preliminary matching score set to generate a personalized innovative resource recommendation list.

[0077] Furthermore, historical user click and adoption data on recommendation results are collected to construct a labeled dataset containing text matching scores, weighted matching scores, policy matching scores, and user adoption status for historical recommendation results. Logistic regression analysis is used, with text matching scores, weighted matching scores, and policy matching scores from the labeled dataset as independent variables and user adoption status as the dependent variable, to fit the data and obtain regression coefficients for each scoring dimension. These regression coefficients are then normalized to a sum of 1, yielding the optimal weight coefficients for each scoring dimension. The text matching score, weighted matching score, and policy matching score of each security resource in the initial matching score set are weighted and summed with their corresponding optimal weight coefficients to obtain a comprehensive score for each security resource. All security resources are then sorted from highest to lowest comprehensive score to generate a security resource list. Based on the user-defined recommendation quantity requirements, the top-ranked security resources are selected from the security resource list to generate a personalized innovative resource recommendation list.

[0078] It should be noted that the logistic regression model is trained using historical user behavior data, with a sample size of ≥1000 valid records. During training, 5-fold cross-validation is used to evaluate the stability of the logistic regression model, and L2 regularization (regularization coefficient λ=0.01) is used to control overfitting. For cold-start users without historical data, a default weight combination based on resource popularity (such as dynamic weight 0.4, policy identifier weight 0.3, and matching degree weight 0.3) is used, and the weight is incrementally updated after users generate interaction data.

[0079] S5.5 Extract the dynamic weight, policy benefit indicator, and comprehensive score of each security resource from the personalized innovation resource recommendation list, and conduct a multi-dimensional business value assessment to generate a business decision support report.

[0080] Furthermore, the system iterates through each security resource in the personalized innovation resource recommendation list, reading its dynamic weight, policy benefit indicator, and comprehensive score. Based on a preset dynamic weight threshold, it assesses market value: a dynamic weight greater than the threshold indicates high market value, a dynamic weight within the threshold indicates medium market value, and a dynamic weight less than the threshold indicates basic market value, generating a market value assessment description. It also generates a policy compliance assessment description based on policy benefit indicators (e.g., "A certain technology complies with the latest energy efficiency policy guidelines and enjoys tax incentives"). Finally, it assesses technology applicability based on a preset comprehensive threshold: a comprehensive score greater than the threshold indicates high applicability, a comprehensive score within the threshold indicates moderate applicability, and a comprehensive score less than the threshold indicates general applicability, generating a technology applicability assessment description. The system then organizes all security resource market value assessment descriptions, policy compliance assessment descriptions, and technology applicability assessment descriptions in the recommendation list according to the recommendation order, integrating them into a business decision support report containing a security resource list, assessment analysis, and recommendations.

[0081] It should be noted that the dynamic weight threshold is obtained by collecting the historical dynamic weights of all entity nodes in the security resource pool over a continuous time period, forming a time-series dataset of historical dynamic weights. A clustering algorithm is then used to cluster all historical dynamic weights, identifying naturally formed dense intervals in the time-series dataset. Simultaneously, based on the business requirements for defining high, medium, and low market value, these naturally formed dense intervals are mapped to three levels. An exemplary value range is [0.5, 0.8]. A value higher than 0.8 would result in too few high-value resources, missing potential high-quality resources; a value lower than 0.5 would result in too many medium- and high-value resources, reducing recommendation differentiation. An exemplary value for the dynamic weight threshold is 0.65 (determined through K-means cluster centers). The comprehensive score threshold is obtained by collecting historical recommendations... The data records the overall score of each recommended resource and the user's adoption behavior for each recommended resource. Using the overall score as the independent variable and user adoption as the dependent variable, an ROC curve is plotted. The optimal classification boundary is obtained by calculating the true positive rate and false positive rate corresponding to each threshold point on the ROC curve, finding the threshold point that maximizes the Youden index. An exemplary value range is [70%, 90%]. A value higher than 90% will result in too few recommended resources and missed potential applicable resources, while a value lower than 70% will result in too many recommended resources, including a large number of resources with low user adoption rates. An exemplary value for the overall score threshold is 85% (determined based on maximizing the Youden index). The confusion matrix for the test set (200 samples) shows: precision 0.88, recall 0.82, and F1 score 0.85.

[0082] This embodiment also provides an innovation resource recommendation system based on an industry chain knowledge graph, including: a data acquisition module for collecting and preprocessing multi-source industry chain innovation ecosystem data; a graph construction module for performing entity recognition and relationship extraction on the preprocessed multi-source industry chain innovation ecosystem data to construct a static industry chain knowledge graph; a graph dynamic update module for dynamically adjusting the weights of entity nodes and relationship edges in the static industry chain knowledge graph based on continuously updated preprocessed multi-source industry chain innovation ecosystem data to generate a dynamic industry chain knowledge graph; a resource filtering module for parsing policy and regulatory information in the dynamic industry chain knowledge graph, extracting mandatory compliance clauses and converting them into a filtering rule set, using the filtering rule set to filter entity nodes and relationship edges in the dynamic industry chain knowledge graph, identifying encouraged resources and assigning policy benefit labels to them to form a safe resource pool; and a resource recommendation module for collecting user query requests, performing comprehensive ranking calculations in the safe resource pool based on the dynamic weights of entity nodes and relationship edges, policy benefit labels, and the matching degree of user query requests, and outputting a business decision support report.

[0083] In summary, this invention generates a dynamic industry chain knowledge graph by dynamically adjusting the weights of entity nodes and relationship edges in a static industry chain knowledge graph. This enables real-time perception and quantitative response to changes in market trends and policy guidance, improving the timeliness and foresight of recommended resources. Furthermore, by filtering entity nodes and relationship edges in the dynamic industry chain knowledge graph to form a secure resource pool, it achieves a deep integration of compliance screening and opportunity discovery, ensuring the commercial feasibility and strategic guidance of the recommended results.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An innovative resource recommendation method based on industry chain knowledge graph, characterized in that: include, Collect and preprocess multi-source industrial chain innovation ecosystem data; Entity identification and relationship extraction are performed on the preprocessed multi-source industrial chain innovation ecosystem data to construct a static industrial chain knowledge graph; Based on the continuously updated preprocessed multi-source industrial chain innovation ecosystem data, the weights of entity nodes and relationship edges in the static industrial chain knowledge graph are dynamically adjusted to generate a dynamic industrial chain knowledge graph. The dynamic industrial chain knowledge graph is analyzed for policy and regulatory information, mandatory compliance clauses are extracted and transformed into a set of filtering rules, and the set of filtering rules is used to filter entity nodes and relationship edges in the dynamic industrial chain knowledge graph, identify incentive resources and assign policy benefit labels to them, and form a safe resource pool. Collect user query requests, perform comprehensive sorting and calculation in the secure resource pool based on the dynamic weights of entity nodes and relation edges, policy benefit indicators, and the matching degree of user query requests, and output a business decision support report.

2. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 1, characterized in that: The multi-source industrial chain innovation ecosystem data includes patent data, policy and regulatory data, enterprise operation data, market intelligence data, and academic research data. The preprocessing includes data cleaning, data denoising, and formatting.

3. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 1, characterized in that: The steps for constructing a static industry chain knowledge graph are as follows: Scan and match patterns on the preprocessed multi-source industrial chain innovation ecosystem data to identify entity names and generate a set of entity names. Using syntactic dependency parsing and co-occurrence statistics, semantic relations between entities in the entity name set are extracted to generate a set of relation triples. A static industry chain knowledge graph is constructed by using the set of entity names as nodes and the set of relation triples as edges.

4. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 1, characterized in that: The steps for generating a dynamic industry chain knowledge graph are as follows: Continuously monitor newly incoming preprocessed multi-source industrial chain innovation ecosystem data, identify and extract data fragments containing event information, and generate a set of event data fragments to be processed; Perform event feature analysis and classification on a set of event data fragments to be processed, and generate event classification results; The event classification results are analyzed to assess the scope and intensity of impact, and a list of event impacts is generated. Based on the event impact list, identify the entity nodes and relationship edges whose weights need to be adjusted, calculate the weight adjustment amount, and generate a weight adjustment instruction set; Execute the weight adjustment instruction set to update the weights of corresponding entity nodes and relationship edges in the static industry chain knowledge graph, and generate a dynamic industry chain knowledge graph.

5. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 1, characterized in that: The steps for parsing policy and regulatory information from the dynamic industry chain knowledge graph, extracting mandatory compliance clauses, and converting them into a set of filtering rules are as follows. Extract text content and integrate related information from entity nodes related to policies and regulations in the dynamic industrial chain knowledge graph to generate a set of policy clause information; Semantic analysis is performed on each policy clause in the policy clause information set to identify clauses containing mandatory requirements and generate a list of mandatory compliance clauses. The list of mandatory compliance clauses is transformed into conditional judgment logic statements that can be executed on a dynamic industry chain knowledge graph, creating a set of filtering rules.

6. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 1, characterized in that: The steps for forming a secure resource pool are as follows: The filtering rule set is used to match and verify the entity nodes and relationship edges in the dynamic industrial chain knowledge graph, and resources that meet the mandatory requirements are selected to generate a preliminary filtering result set. Based on the policy clause information set, resources that meet the mandatory requirements in the preliminary filtering result set are subjected to policy orientation matching analysis and policy benefit labels are added to form a safe resource pool.

7. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 1, characterized in that: The process of collecting user query requests involves a comprehensive ranking calculation within the security resource pool based on the dynamic weights of entity nodes and relation edges, policy benefit indicators, and the matching degree of the user query requests. The steps are as follows: Collect user query requests and perform semantic parsing and query element extraction to generate semantic query objects; The semantic query object is matched and calculated with the entity nodes in the security resource pool in multiple dimensions to generate a preliminary matching score set. A comprehensive score calculation and ranking of each security resource in the initial matching score set is performed to generate a personalized innovative resource recommendation list.

8. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 7, characterized in that: The steps for performing multi-dimensional matching calculations between the semantic query object and entity nodes in the security resource pool to generate a preliminary matching score set are as follows. Extract query elements from the semantic query object and match them with entity nodes in the security resource pool to generate a multi-dimensional matching element set. The set of multi-dimensional matching elements is compared with the text content, dynamic weights and policy benefit identifiers of entity nodes to calculate similarity, and a preliminary matching score set is generated.

9. The innovation resource recommendation method based on industry chain knowledge graph as described in claim 1, characterized in that: The business decision support report is generated by extracting the dynamic weight, policy benefit indicators, and comprehensive score of each security resource from the personalized innovation resource recommendation list, and conducting a multi-dimensional business value assessment.

10. An innovation resource recommendation system based on an industry chain knowledge graph, based on the innovation resource recommendation method based on an industry chain knowledge graph as described in any one of claims 1 to 9, characterized in that: include, The data acquisition module is used to collect and preprocess multi-source industrial chain innovation ecosystem data. The graph construction module is used to perform entity recognition and relationship extraction on preprocessed multi-source industrial chain innovation ecosystem data to construct a static industrial chain knowledge graph. The graph dynamic update module is used to dynamically adjust the weights of entity nodes and relationship edges in the static industrial chain knowledge graph based on the continuously updated preprocessed multi-source industrial chain innovation ecosystem data, and generate a dynamic industrial chain knowledge graph. The resource filtering module is used to analyze policy and regulatory information in the dynamic industrial chain knowledge graph, extract mandatory compliance clauses and convert them into a set of filtering rules. The set of filtering rules is used to filter entity nodes and relationship edges in the dynamic industrial chain knowledge graph, identify encouraged resources and assign them policy benefit labels to form a safe resource pool. The resource recommendation module collects user query requests and performs comprehensive ranking calculations in the secure resource pool based on the dynamic weights of entity nodes and relation edges, policy benefit indicators, and the matching degree of user query requests, and outputs a business decision support report.

Images