Policy recommendation-oriented multi-dimensional graph data recall strategy system and method

Abstract

The invention belongs to the technical field of multi-dimensional data processing, and particularly relates to a policy recommendation-oriented multi-dimensional graph data recall policy system and method.The multi-dimensional graph data is cleaned and structured through data preprocessing, and the multi-dimensional graph data comprises enterprise portrait data, policy data and other related data; feature extraction: encoding the policy text by adopting a deep semantic analysis model, and extracting a feature vector containing context semantics; dynamic feature interaction: adjusting the importance of the user and policy features through a dynamic weight mechanism, and optimizing the feature interaction effect; multi-dimensional task collaboration: based on a multi-task learning framework, processing a plurality of policy objectives at the same time, and generating a comprehensive matching result; a recall strategy is generated, a multi-dimensional graph data recall strategy suitable for policy recommendation is generated in combination with the processing result, and real-time synchronization between a recommendation system and enterprise requirements and policy environment changes is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of multidimensional data processing, and in particular to a multidimensional graph data recall strategy system and method for policy recommendation. Background Art

[0002] Traditional policy recommendation systems rely on a single data source or simple rule matching, exposing four core shortcomings in modeling complex government-enterprise relationships:

[0003] Information silos and shallow association analysis. Existing methods rely on static rules (such as industry classification matching) or simple vector matching, failing to integrate heterogeneous data from multiple sources, including enterprise (size, tax payment, grade, R&D investment, and other dimensions), policy (validity period, support amount, application conditions), industry, and region. For example, recommending policies based solely on "industry = manufacturing" ignores the deeper mapping between "tax grade A + 20% R&D investment" and "R&D expense additional deduction policy," resulting in a potential association mining coverage rate of less than 40%.

[0004] Dynamic adaptability lags behind changing demand, relying on static company profiles (registered industry / establishment date) and historical interaction data, lacking real-time dynamic tracking capabilities. When key events such as obtaining an invention patent or expanding into new regions occur, the model must be manually updated, with an average lag of 48 hours. This results in a disconnect between recommended policies and actual demand exceeding 60%.

[0005] Due to the lack of high-order relationship modeling capabilities, traditional methods only support surface feature matching and are unable to handle indirect associations from “policy → industrial chain → enterprise” (such as “new energy subsidy policy → power battery industry chain → upstream raw material enterprise”). It is difficult to capture high-order semantic paths in graph structures such as “tax preferential policies → enterprise tax levels → industrial chain cost transmission”. The matching accuracy of complex policy terms is less than 35%.

[0006] The recommendation efficiency in cold start scenarios is low. For new enterprises (no policy interaction records) or new policies (no implementation feedback), it only relies on limited basic attribute matching. For example, the accuracy of recommending "innovation fund subsidies" for "establishment time < 1 year + technology industry" is less than 30%, and the exposure rate of new policies in the first week is less than 20%, resulting in low policy resource utilization.

[0007] In response to the above technical defects, a solution of a multi-dimensional graph data recall strategy system and method for policy recommendation is proposed. Summary of the Invention

[0008] In order to solve the above problems, the present invention provides the following technical solutions:

[0009] A multi-dimensional graph data recall strategy method for policy recommendation, including:

[0010] Data preprocessing: cleaning and structuring multi-dimensional graph data, including enterprise profile data, policy data, and other relevant data;

[0011] Feature extraction: A deep semantic parsing model is used to encode the policy text and extract feature vectors containing contextual semantics;

[0012] Dynamic feature interaction: adjust the importance of user and policy features through a dynamic weight mechanism to optimize the effect of feature interaction;

[0013] Multi-dimensional task collaboration, based on a multi-task learning framework, simultaneously processes multiple policy objectives and generates comprehensive matching results;

[0014] Recall strategy generation,combining the above processing results, generates a multi-dimensional graph data recall strategy suitable for,policy recommendation.

[0015] Furthermore, the steps of cleaning and structuring the multi-dimensional graph data include checking and deleting duplicate enterprise records or policy entries to ensure the uniqueness of each piece of data, identifying and processing abnormal data points for missing fields, such as the enterprise's registered capital or policy support measures, and processing excessively high R&D investment or excessively low operating income by deleting, correcting or marking outliers;

[0016] Integrate enterprise portrait data, policy data, and other relevant data into a unified dataset, ensure that the field names and meanings in different data sources are consistent, perform field mapping and conversion when necessary, convert field types into a format suitable for analysis, perform natural language processing on the text description of the policy, extract keywords and topic tags, and convert them into structured feature vectors, build a relationship map between enterprises and policies, including the association between enterprise types and policy scope of application, to facilitate subsequent graph data processing and analysis.

[0017] Furthermore, the step of encoding the policy text using a deep semantic parsing model and extracting a feature vector containing contextual semantics includes capturing the contextual relationship of the text, using a pre-trained model as a base model, fine-tuning the model according to the characteristics of the policy text to optimize its performance in a specific field, converting the preprocessed text into an input format acceptable to the model, inputting the processed text into the deep semantic parsing model, aggregating the vectors of words to generate a feature vector for the entire sentence or text paragraph, ensuring that the aggregated vector can retain the contextual information and semantic structure of the original text, standardizing the feature vector to ensure comparability between different texts, verifying the quality of the feature vector through similarity calculation and classification task methods, and storing the extracted feature vector in a database or vector index to facilitate subsequent retrieval and matching.

[0018] Furthermore, the steps of adjusting the importance of user and policy features through a dynamic weight mechanism and optimizing the feature interaction effect include assigning an initial weight to each feature, collecting real-time user data and the latest policy information, dynamically adjusting feature weights using a machine learning algorithm, and recalculating the feature vectors of users and policies based on the adjusted weights so that the importance of each feature is fully reflected. Based on the adjusted feature weights and similarity calculation results, the recommendation model is optimized to generate a more accurate policy recommendation list, and the dynamic weight mechanism and the optimized recommendation model are integrated into the policy recommendation system to ensure efficient operation of the entire process.

[0019] Furthermore, the steps of simultaneously processing multiple policy objectives based on a multi-task learning framework, generating comprehensive matching results, and dynamically fusing multi-source data include collecting data from multiple dimensions including business administration, finance, and R&D, including enterprise registration information and change records; financial data including revenue, profit, and balance sheet; and R&D data including patent applications and R&D investment;

[0020] Use real-time ETL tools to clean the collected data, removing noise and redundancy. Convert data in different formats to a unified standard format. Based on the cleaned and standardized data, construct a corporate profile encompassing 12 core dimensions, including company size, industry classification, founding date, registered capital, and credit rating. This corporate profile supports full lifecycle feature management, such as feature analysis at different stages of the startup, growth, and maturity stages.

[0021] Identify and define key events, including successful financing, patent authorization, and business expansion. When a key event is triggered, the system automatically updates the node attributes and edge weights in the graph model, including the company's latest financial data and R&D results; the edge weights reflect the strength of the relationship between the company and policies, industrial chains, and regions. By optimizing the data processing process and using efficient technical means, ensure that the data update delay is less than 5 minutes.

[0022] Furthermore, the dynamic fusion of multi-source data includes learning a low-dimensional representation of the data through a neural network for feature extraction and dimensionality reduction;

[0023] Encoder calculation:

[0024] h=f(W e x+b e ),

[0025] Among them, h is the hidden layer display after encoding, x is the input data vector, b e is the bias vector of the encoder;

[0026] Decoder calculation:

[0027]

[0028] Among them, h is the hidden layer display after encoding, is the reconstructed data vector, b d is the bias vector of the decoder;

[0029] By minimizing the loss function, the weight matrix W is optimized e , W d and the bias vector b e , b d , calculated as follows:

[0030]

[0031] Among them, x i is the i-th input data, N is the total number of samples;

[0032] By aggregating the information of neighboring nodes, a vector representation of the node is generated, which is suitable for processing graph structure data. The calculation is as follows:

[0033]

[0034] in, is the representation of node v in the kth layer, AGG is the aggregation function, and σ is the activation function.

[0035] According to one aspect of the present invention, a multi-dimensional graph data recall strategy system for policy recommendation is provided, comprising:

[0036] Data collection and preprocessing: collecting policy-related information from multiple data sources and performing cleaning and standardization;

[0037] Feature extraction and data fusion, using principal component analysis and autoencoders to extract key features and construct multidimensional graph data structures;

[0038] Graph embedding and knowledge graph construction: applying graph embedding technology to generate vector representations of nodes and relationships and constructing policy knowledge graphs;

[0039] Recall strategy and recommendation model: design multi-dimensional graph data recall strategy based on knowledge graph and use graph neural network for policy recommendation;

[0040] Model optimization and evaluation: Use optimization algorithms to train the recommendation model and design evaluation indicators to evaluate system performance.

[0041] Furthermore, the system includes using principal component analysis to reduce the dimensionality of high-dimensional policy data, extracting key features, learning low-dimensional representations of data through autoencoders, optimizing the feature extraction process, fusing multi-source data, and constructing a multi-dimensional graph data structure, including policy nodes, association relationships and attribute information, applying graph embedding technology to process multi-dimensional graph data, generating vector representations of nodes and relationships, constructing a policy knowledge graph, integrating policy nodes and their association relationships, forming a structured knowledge system, designing a multi-dimensional graph data recall strategy based on the constructed knowledge graph, achieving accurate policy matching, using graph neural networks for policy recommendations, and improving the accuracy and personalization of recommendations by aggregating node information and relationship information, using optimization algorithms to train recommendation models, optimizing model parameters, designing evaluation indicators, and evaluating and verifying the performance of the recommendation system.

[0042] According to one aspect of the present invention, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program and the processor implements the steps of the above-mentioned policy recommendation-oriented multi-dimensional graph data recall strategy method when executing the computer program.

[0043] According to one aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned multi-dimensional graph data recall strategy method for policy recommendation are implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] In a multi-dimensional graph data recall strategy method for policy recommendation of the present invention, the multi-dimensional graph data is cleaned and structured through data preprocessing, and the multi-dimensional graph data includes enterprise portrait data, policy data and other related data; feature extraction, a deep semantic parsing model is used to encode the policy text, and feature vectors containing contextual semantics are extracted; dynamic feature interaction, the importance of user and policy features is adjusted through a dynamic weight mechanism, and the feature interaction effect is optimized; multi-dimensional task collaboration, based on a multi-task learning framework, simultaneously processes multiple policy goals and generates comprehensive matching results; recall strategy generation, combined with the above processing results, generates a multi-dimensional graph data recall strategy suitable for policy recommendation, which ensures real-time synchronization between the recommendation system and changes in enterprise needs and policy environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0047] Figure 1 This is an overall schematic diagram of a multi-dimensional graph data recall strategy method for policy recommendation according to the present invention;

[0048] Figure 2 This is a schematic diagram of a framework of a policy recommendation-oriented multi-dimensional graph data recall strategy system according to the present invention;

[0049] Figure 3 This is a schematic diagram of the computer structure in a multi-dimensional graph data recall strategy system for policy recommendation according to the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] like Figure 1-Figure 3 As shown, this application provides a multi-dimensional graph data recall strategy method for policy recommendation, including:

[0052] S1: Data preprocessing, cleaning and structuring multi-dimensional graph data, which includes enterprise profile data, policy data, and other relevant data;

[0053] S2: Feature extraction: Use a deep semantic parsing model to encode the policy text and extract feature vectors containing contextual semantics;

[0054] S3: Dynamic feature interaction, which adjusts the importance of user and policy features through a dynamic weight mechanism to optimize the feature interaction effect;

[0055] S4: Multi-dimensional task collaboration, based on a multi-task learning framework, simultaneously processes multiple policy objectives and generates comprehensive matching results;

[0056] S5: Recall strategy generation: Combined with the above processing results, a multi-dimensional graph data recall strategy suitable for policy recommendation is generated.

[0057] In one embodiment, the data source collects policy-related information from multiple data sources, including but not limited to: government-issued policy documents (e.g., PDFs, Word documents), corporate strategic plans (e.g., annual reports, press releases), public discussions on social media platforms (e.g., Weibo, forums), and relevant reports from news media (e.g., news websites, news APIs).

[0058] Data cleaning involves cleaning the collected raw data to remove noise and redundant information. For example, this includes removing duplicate data (such as duplicate policy documents or news reports), removing irrelevant data (such as social media comments unrelated to the policy), and handling missing values ​​(such as filling in missing policy effective dates).

[0059] Data standardization standardizes the cleaned data and unifies the data format and unit. For example: convert text data into a unified encoding format (such as UTF-8). Unify the date format to "YYYY-MM-DD". Convert numerical data into a unified unit (such as unifying the amount into RMB). Feature extraction and data fusion use principal component analysis (PCA) to reduce the dimensionality of high-dimensional policy data and extract key features. For example: extract keywords (such as "tax incentives" and "industrial upgrading") from policy texts. Extract key indicators (such as "annual growth rate" and "market share") from corporate data. Extract sentiment tendencies (such as "support" and "oppose") from social media data. Further use autoencoders to reduce the dimensionality of the data and optimize the feature extraction process. For example: map high-dimensional feature vectors to low-dimensional space to generate a more compact feature representation.

[0060] Data fusion combines data from multiple sources to construct a multidimensional graph data structure. For example, a policy node (such as "tax incentives") is integrated with an associated relationship (such as "applicable to manufacturing"). Enterprise data (such as "annual growth rate of a company") is associated with a policy node. Social media data (such as "public support for a policy") is associated with a policy node. The resulting multidimensional graph data structure includes:

[0061] Policy nodes represent specific policy content. Relationships represent the relationship between policies and other entities (e.g., enterprises, the public). Attribute information represents additional information about nodes and relationships (e.g., timestamps, geographic locations).

[0062] Graph embedding and knowledge graph construction: Graph embedding techniques (such as GraphSAGE) are applied to multidimensional graph data to generate vector representations of nodes and relationships. For example, policy nodes can be converted into low-dimensional vector representations (e.g., 128-dimensional vectors). Relationships can be converted into vector representations to capture the semantic relationships between nodes. Knowledge graph construction builds a policy knowledge graph based on the graph embedding results. For example, policy nodes and their relationships can be integrated to form a structured knowledge system. Policy nodes can be associated with entities such as businesses and the public to form a complete knowledge graph.

[0063] Recall strategies and recommendation models are designed based on the constructed knowledge graph. A multi-dimensional graph data recall strategy is designed. For example, based on the user's query intent (e.g., "Search for tax incentives suitable for small and medium-sized enterprises"), relevant policy nodes are retrieved from the knowledge graph. The recall results are further optimized by incorporating user contextual information (e.g., industry, location). The recommendation model utilizes graph neural networks (GNNs) for policy recommendations. For example, a graph convolutional network (GCN) is used to aggregate node and relationship information to generate the final recommendation results. Based on user preferences (e.g., historical click history and feedback), personalized policy recommendations are made.

[0064] Model Optimization and Evaluation: Use optimization algorithms (such as the Adam optimizer) to train the recommendation model and optimize model parameters. For example, adjust hyperparameters such as the learning rate and batch size to improve model performance. Design evaluation metrics (such as accuracy, recall, and F1 score) to evaluate and validate the performance of the recommendation system. For example, use real user data to test and assess the accuracy and relevance of recommendation results. Use A / B testing to compare the performance of different recall strategies and recommendation models.

[0065] Through the above-mentioned implementation steps, the multi-dimensional graph data recall strategy method for policy recommendation of the present invention can achieve the following effects: Efficient data integration constructs a complete multi-dimensional graph data structure through the cleaning, standardization and fusion of multi-source heterogeneous data, laying the foundation for subsequent policy recommendations. Accurate recall strategy is based on the multi-dimensional graph data recall strategy of the knowledge graph, which can accurately recall relevant policies according to the user's query intention and context information. Personalized recommendation results use graph neural networks (GNN) for policy recommendations, which can aggregate node information and relationship information to generate personalized and accurate recommendation results. Significant performance improvement is achieved through the combination of optimization algorithms and evaluation indicators, which significantly improves the accuracy, recall rate and user satisfaction of the recommendation system.

[0066] Specifically, the steps of cleaning and structuring the multi-dimensional graph data include checking and deleting duplicate enterprise records or policy entries to ensure the uniqueness of each piece of data, identifying and processing abnormal data points for missing fields, such as the enterprise's registered capital or policy support measures, and processing excessively high R&D investment or excessively low operating income by deleting, correcting, or marking outliers.

[0067] Integrate enterprise portrait data, policy data, and other relevant data into a unified dataset, ensure that the field names and meanings in different data sources are consistent, perform field mapping and conversion when necessary, convert field types into a format suitable for analysis, perform natural language processing on the text description of the policy, extract keywords and topic tags, and convert them into structured feature vectors, build a relationship map between enterprises and policies, including the association between enterprise types and policy scope of application, to facilitate subsequent graph data processing and analysis.

[0068] Furthermore, the step of encoding the policy text using a deep semantic parsing model and extracting a feature vector containing contextual semantics includes capturing the contextual relationship of the text, using a pre-trained model as a base model, fine-tuning the model according to the characteristics of the policy text to optimize its performance in a specific field, converting the preprocessed text into an input format acceptable to the model, inputting the processed text into the deep semantic parsing model, aggregating the vectors of words to generate a feature vector for the entire sentence or text paragraph, ensuring that the aggregated vector can retain the contextual information and semantic structure of the original text, standardizing the feature vector to ensure comparability between different texts, verifying the quality of the feature vector through similarity calculation and classification task methods, and storing the extracted feature vector in a database or vector index to facilitate subsequent retrieval and matching.

[0069] Furthermore, the steps of adjusting the importance of user and policy features through a dynamic weight mechanism and optimizing the feature interaction effect include assigning an initial weight to each feature, collecting real-time user data and the latest policy information, dynamically adjusting feature weights using a machine learning algorithm, and recalculating the feature vectors of users and policies based on the adjusted weights so that the importance of each feature is fully reflected. Based on the adjusted feature weights and similarity calculation results, the recommendation model is optimized to generate a more accurate policy recommendation list, and the dynamic weight mechanism and the optimized recommendation model are integrated into the policy recommendation system to ensure efficient operation of the entire process.

[0070] Furthermore, the steps of simultaneously processing multiple policy objectives based on a multi-task learning framework, generating comprehensive matching results, and dynamically fusing multi-source data include collecting data from multiple dimensions including business administration, finance, and R&D, including enterprise registration information and change records; financial data including revenue, profit, and balance sheet; and R&D data including patent applications and R&D investment;

[0071] Use real-time ETL tools to clean the collected data, removing noise and redundancy. Convert data in different formats to a unified standard format. Based on the cleaned and standardized data, construct a corporate profile encompassing 12 core dimensions, including company size, industry classification, founding date, registered capital, and credit rating. This corporate profile supports full lifecycle feature management, such as feature analysis at different stages of the startup, growth, and maturity stages.

[0072] Identify and define key events, including successful financing, patent authorization, and business expansion. When a key event is triggered, the system automatically updates the node attributes and edge weights in the graph model, including the company's latest financial data and R&D results; the edge weights reflect the strength of the relationship between the company and policies, industrial chains, and regions. By optimizing the data processing process and using efficient technical means, ensure that the data update delay is less than 5 minutes.

[0073] Specifically, the dynamic fusion of multi-source data includes learning a low-dimensional representation of the data through a neural network for feature extraction and dimensionality reduction;

[0074] Encoder calculation:

[0075] h=f(W e x+b e ),

[0076] Among them, h is the hidden layer display after encoding, x is the input data vector, b e is the bias vector of the encoder;

[0077] Decoder calculation:

[0078]

[0079] Among them, h is the hidden layer display after encoding, is the reconstructed data vector, b d is the bias vector of the decoder;

[0080] By minimizing the loss function, the weight matrix W is optimized e , W d and the bias vector b e , b d , calculated as follows:

[0081]

[0082] Among them, x i is the i-th input data, N is the total number of samples;

[0083] By aggregating the information of neighboring nodes, a vector representation of the node is generated, which is suitable for processing graph structure data. The calculation is as follows:

[0084]

[0085] in, is the representation of node v in the kth layer, AGG is the aggregation function, and σ is the activation function.

[0086] According to one aspect of the present invention, a multi-dimensional graph data recall strategy system for policy recommendation is provided, comprising:

[0087] Data collection and preprocessing: collecting policy-related information from multiple data sources and performing cleaning and standardization;

[0088] Feature extraction and data fusion, using principal component analysis and autoencoders to extract key features and construct multidimensional graph data structures;

[0089] Graph embedding and knowledge graph construction: applying graph embedding technology to generate vector representations of nodes and relationships and constructing policy knowledge graphs;

[0090] Recall strategy and recommendation model: design multi-dimensional graph data recall strategy based on knowledge graph and use graph neural network for policy recommendation;

[0091] Model optimization and evaluation: Use optimization algorithms to train the recommendation model and design evaluation indicators to evaluate system performance.

[0092] Specifically, the system includes using principal component analysis to reduce the dimensionality of high-dimensional policy data, extracting key features, learning low-dimensional representations of data through autoencoders, optimizing the feature extraction process, fusing multi-source data, and constructing a multi-dimensional graph data structure, including policy nodes, association relationships and attribute information, applying graph embedding technology to process multi-dimensional graph data, generating vector representations of nodes and relationships, constructing a policy knowledge graph, integrating policy nodes and their association relationships, forming a structured knowledge system, designing a multi-dimensional graph data recall strategy based on the constructed knowledge graph, achieving accurate policy matching, using graph neural networks for policy recommendations, and improving the accuracy and personalization of recommendations by aggregating node information and relationship information, using optimization algorithms to train recommendation models, optimizing model parameters, designing evaluation indicators, and evaluating and verifying the performance of the recommendation system.

[0093] The system architecture is designed as follows:

[0094] Multi-dimensional graph model building module and three-tier architecture realize digital modeling of government-enterprise relations:

[0095] Physical layer (data cornerstone):

[0096] Enterprise node: Integrates over 80 attributes, including industrial and commercial registration (registered capital / equity structure), finance (revenue / R&D expenditure), and R&D (number of patents / technical fields). For example, the "High-tech Enterprise" node includes fine-grained features such as "R&D investment accounts for 25% + 12 invention patents."

[0097] Policy node: stores policy text (goals / measures / scope of application) and attributes (issuing department / validity period / subsidy amount). For example, the "R&D Subsidy Policy" node specifies the application conditions of "annual revenue ≤ 50 million + R&D personnel ratio ≥ 30%".

[0098] Industrial chain nodes: define the relationship between "upstream suppliers-midstream manufacturers-downstream distributors", and regional nodes mark policy clusters such as the "Beijing-Tianjin-Hebei Coordinated Development Zone".

[0099] Relationship layer (association hub):

[0100] Policy applicability condition edge: an explicit matching condition marked in both directions as “policy → enterprise”.

[0101] Enterprise historical behavior edge: records "application time / results / benefit amount" to form a profile of the enterprise's policy preferences. For example, for an enterprise that "successfully applied for the Science and Technology-based Small and Medium Enterprises Innovation Fund in 2023", the matching weight of similar policies will be automatically increased by 15%.

[0102] Industry-related edge: captures the transmission relationship of "new energy vehicle industry → power battery raw material suppliers → lithium mining subsidy policy", and the regional coordination edge is related to "local industrial park policies → enterprises registered in the park".

[0103] Attribute layer (dynamic features):

[0104] Node attributes: real-time enterprise tax data (updated monthly), remaining validity period of the policy (accurate to days);

[0105] Edge attributes: frequency of government-enterprise interaction (number of consultations in the last 30 days) and historical matching effect (policy application approval rate).

[0106] Dynamic weight update module and dual-drive mechanism ensure real-time accuracy:

[0107] Multi-factor weight formula: W e (t) = α·f + β·g + γ·h

[0108] Among them, f, g, and h are interaction frequency, policy timeliness, and historical matching, respectively. α, β, and γ are dynamically adjusted through the attention mechanism (for example, the weight of β in emergency policy scenarios is increased to 0.6). The interaction frequency adopts a time decay function (the interaction weight in the last month = twice that of 3 months ago). The policy timeliness decays by 1% per day for policies with a remaining validity period of <30 days.

[0109] Real-time update architecture:

[0110] Based on the Kafka message queue, it monitors enterprise change events (tax grade upgrade / new patent authorization) and policy dynamics (extension / new clause), triggering the SparkGraphX ​​distributed computing framework to recalculate weights. The update delay is less than 2 minutes, ensuring global graph data consistency.

[0111] Hybrid recall algorithm module, three-level recall process covers all scenarios:

[0112] Rule recall sub-module: predefines 200+ policy tags (such as "industrial upgrading-intelligent manufacturing" and "environmental protection compliance-carbon emission subsidies"), performs Boolean matching with enterprise qualifications (high-tech enterprises / specialized and new small giants), quickly filters out obviously irrelevant policies, and outputs a preliminary candidate set (about 1,000 items).

[0113] Collaborative filtering submodule: Construct an enterprise-policy interaction matrix (row = enterprise, column = policy, value = interaction intensity), calculate enterprise similarity through graph structure similarity algorithms (such as SimRank), recommend policies that similar enterprises are interested in, and expand the candidate set to 2,000 items, covering "unpopular policies declared by the top 10 enterprises in the same industry."

[0114] Graph neural network submodule: Graph structure data and node attributes are input. The GCN model aggregates neighbor node information through the Laplacian matrix to capture the indirect relationship between "enterprise-industry chain-policy". The GAT model assigns higher weights to key neighbors through the attention mechanism (such as the policy weight associated with leading enterprises +30%), outputs a 128-dimensional vector, calculates the cosine similarity, and screens the top-50 policies with a matching score >0.7.

[0115] In a multi-dimensional graph data recall strategy method for policy recommendation of the present invention, the multi-dimensional graph data is cleaned and structured through data preprocessing, and the multi-dimensional graph data includes enterprise portrait data, policy data and other related data; feature extraction, a deep semantic parsing model is used to encode the policy text, and feature vectors containing contextual semantics are extracted; dynamic feature interaction, the importance of user and policy features is adjusted through a dynamic weight mechanism, and the feature interaction effect is optimized; multi-dimensional task collaboration, based on a multi-task learning framework, simultaneously processes multiple policy goals and generates comprehensive matching results; recall strategy generation, combined with the above processing results, generates a multi-dimensional graph data recall strategy suitable for policy recommendation, which ensures real-time synchronization between the recommendation system and changes in enterprise needs and policy environment.

[0116] The present invention also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned policy recommendation-oriented multi-dimensional graph data recall strategy method are implemented.

[0117] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned policy-recommendation-oriented multi-dimensional graph data recall strategy method.

[0118] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0119] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0120] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A multi-dimensional graph data recall strategy method for policy recommendation, characterized by: include: Data preprocessing: cleaning and structuring multi-dimensional graph data, including enterprise profile data, policy data, and other relevant data; Feature extraction: A deep semantic parsing model is used to encode the policy text and extract feature vectors containing contextual semantics; Dynamic feature interaction: adjust the importance of user and policy features through a dynamic weight mechanism to optimize the effect of feature interaction; Multi-dimensional task collaboration, based on a multi-task learning framework, simultaneously processes multiple policy objectives and generates comprehensive matching results; Recall strategy generation,combining the above processing results, generates a multi-dimensional graph data recall strategy suitable for,policy recommendation.

2. The multi-dimensional graph data recall strategy method for policy recommendation according to claim 1, characterized in that: The steps of cleaning and structuring the multi-dimensional graph data include checking and deleting duplicate enterprise records or policy entries to ensure the uniqueness of each piece of data, identifying and processing abnormal data points for missing fields, such as the enterprise's registered capital or policy support measures, and processing excessively high R&D investment or excessively low operating income by deleting, correcting or marking outliers; Integrate enterprise portrait data, policy data, and other relevant data into a unified dataset, ensure that the field names and meanings in different data sources are consistent, perform field mapping and conversion when necessary, convert field types into a format suitable for analysis, perform natural language processing on the text description of the policy, extract keywords and topic tags, and convert them into structured feature vectors, build a relationship map between enterprises and policies, including the association between enterprise types and policy scope of application, to facilitate subsequent graph data processing and analysis.

3. The multi-dimensional graph data recall strategy method for policy recommendation according to claim 1, characterized in that: The steps of encoding the policy text using a deep semantic parsing model and extracting a feature vector containing contextual semantics include capturing the contextual relationship of the text, using a pre-trained model as a base model, fine-tuning the model according to the characteristics of the policy text to optimize its performance in a specific field, converting the pre-processed text into an input format acceptable to the model, inputting the processed text into the deep semantic parsing model, aggregating word vectors to generate feature vectors for the entire sentence or text paragraph, ensuring that the aggregated vectors can retain the contextual information and semantic structure of the original text, standardizing the feature vectors to ensure comparability between different texts, verifying the quality of the feature vectors through similarity calculation and classification task methods, and storing the extracted feature vectors in a database or vector index to facilitate subsequent retrieval and matching.

4. The multi-dimensional graph data recall strategy method for policy recommendation according to claim 1, characterized in that: The steps of adjusting the importance of user and policy features through a dynamic weight mechanism and optimizing the feature interaction effect include assigning an initial weight to each feature, collecting real-time user data and the latest policy information, dynamically adjusting feature weights using a machine learning algorithm, and recalculating the feature vectors of users and policies based on the adjusted weights so that the importance of each feature is fully reflected. Based on the adjusted feature weights and similarity calculation results, the recommendation model is optimized to generate a more accurate policy recommendation list, and the dynamic weight mechanism and the optimized recommendation model are integrated into the policy recommendation system to ensure the efficient operation of the entire process.

5. The multi-dimensional graph data recall strategy method for policy recommendation according to claim 1, characterized in that: The steps of the multi-task learning framework, which simultaneously processes multiple policy objectives, generates comprehensive matching results, and dynamically integrates multi-source data include collecting data from multiple dimensions, including business, finance, and R&D, including enterprise registration information and change records; financial data includes revenue, profit, and balance sheets; and R&D data includes patent applications and R&D investment. Use real-time ETL tools to clean the collected data, removing noise and redundancy. Convert data in different formats to a unified standard format. Based on the cleaned and standardized data, construct a corporate profile encompassing 12 core dimensions, including company size, industry classification, founding date, registered capital, and credit rating. This corporate profile supports full lifecycle feature management, such as feature analysis at different stages of the startup, growth, and maturity stages. Identify and define key events, including successful financing, patent authorization, and business expansion. When key events are triggered, the system automatically updates the node attributes and edge weights in the graph model, including the company's latest financial data and R&D results. Edge weights reflect the strength of the relationship between enterprises and policies, industrial chains, and regions. By optimizing data processing procedures and using efficient technical means, we ensure that data update delays are less than 5 minutes.

6. The multi-dimensional graph data recall strategy method for policy recommendation according to claim 5, characterized in that: Dynamic fusion of multi-source data involves learning low-dimensional representations of data through neural networks for feature extraction and dimensionality reduction; Encoder calculation: h=f(W e x+b e ), Among them, h is the hidden layer display after encoding, x is the input data vector, b e is the bias vector of the encoder; Decoder calculation: Among them, h is the hidden layer display after encoding, is the reconstructed data vector, b d is the bias vector of the decoder; By minimizing the loss function, the weight matrix W is optimized e , W d and the bias vector b e , b d , calculated as follows: Among them, x i is the i-th input data, N is the total number of samples; By aggregating the information of neighboring nodes, a vector representation of the node is generated, which is suitable for processing graph structure data. The calculation is as follows: in, is the representation of node v in the kth layer, AGG is the aggregation function, and σ is the activation function.

7. A multi-dimensional graph data recall strategy system for policy recommendation, characterized by: include: Data collection and preprocessing: collecting policy-related information from multiple data sources and performing cleaning and standardization; Feature extraction and data fusion, using principal component analysis and autoencoders to extract key features and construct multidimensional graph data structures; Graph embedding and knowledge graph construction: applying graph embedding technology to generate vector representations of nodes and relationships and constructing policy knowledge graphs; Recall strategy and recommendation model: design multi-dimensional graph data recall strategy based on knowledge graph and use graph neural network for policy recommendation; Model optimization and evaluation: Use optimization algorithms to train the recommendation model and design evaluation indicators to evaluate system performance.

8. The multi-dimensional graph data recall strategy system for policy recommendation according to claim 7, characterized in that: The system includes using principal component analysis to reduce the dimensionality of high-dimensional policy data, extracting key features, learning low-dimensional representations of data through autoencoders, optimizing the feature extraction process, fusing multi-source data, and constructing a multi-dimensional graph data structure, including policy nodes, association relationships and attribute information. The system applies graph embedding technology to process multi-dimensional graph data, generates vector representations of nodes and relationships, constructs a policy knowledge graph, integrates policy nodes and their association relationships, and forms a structured knowledge system. Based on the constructed knowledge graph, a multi-dimensional graph data recall strategy is designed to achieve accurate policy matching. Graph neural networks are used for policy recommendations, and the accuracy and personalization of recommendations are improved by aggregating node information and relationship information. An optimization algorithm is used to train the recommendation model, optimize model parameters, design evaluation indicators, and evaluate and verify the performance of the recommendation system.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of a multi-dimensional graph data recall strategy method for policy recommendation according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a multi-dimensional graph data recall strategy method for policy recommendation according to any one of claims 1 to 6 are implemented.

Images