A knowledge graph-based professional influence evaluation and dynamic deduction method and system, and a storage medium

Abstract

The application provides a kind of professional influence evaluation and dynamic deduction method, system and storage medium based on knowledge graph, method includes: step one: build the structured knowledge set representing entity and relationship in professional field;Step two: based on the structured knowledge set of step one, calculate the global influence absolute score of target entity;Step three: based on the global influence absolute score of step two, carry out context-aware situational influence analysis;Step four: based on the perturbation simulation of the structured knowledge set of step one, carry out forward-looking influence deduction.The beneficial effects of the present application are: 1. Realize the holographic and objectivity of evaluation basis, fundamentally improve the credibility and fairness of evaluation results;2. Establish a unified, data-driven evaluation scale, for the first time to realize the fair benchmarking across organizations and regions;3. Provide multi-level analysis capability with context awareness, so that the evaluation has unprecedented depth and fineness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing and artificial intelligence, and particularly relates to a professional influence evaluation and dynamic deduction method and system based on a knowledge graph and a storage medium. BACKGROUND

[0002] In the era of knowledge economy, scientific, objective and comprehensive evaluation of professional talents (such as doctors, scientific researchers, engineers and lawyers) has become the cornerstone of talent management, resource allocation and strategic decision-making for various organizations. However, the current technical practice generally faces a core dilemma: the contradiction between the extreme dispersion of information required for evaluation and the serious one-dimensionality of evaluation method dimensions.

[0003] On the one hand, key data such as professional information, academic achievements, social reputation and network position that can comprehensively reflect the comprehensive ability of an individual are widely scattered in a large number of heterogeneous and disconnected data sources such as institutional websites, industry vertical platforms, academic databases, news media and social networks, forming an insurmountable "information island".

[0004] On the other hand, the existing evaluation system mostly relies on simple weighting of a few easily accessible explicit indicators (such as title, length of service and platform internal score), completely ignoring the consideration of the depth of individual skill combination, the pivotal role in the collaboration network and the actual reputation in the industry and other deeper implicit capital. This contradiction leads to the subjectivity of the existing evaluation results, poor comparability, and complete lack of context awareness and multi-level analysis capabilities.

[0005] More importantly, the existing technology is essentially a pure "backward-looking" static summary, and it is completely powerless for strategic problems facing the future, such as "how much quantitative impact will the core talent flow have on the organization?" or "how to form a team with maximum efficiency?" Therefore, there is an urgent need in the field for a new technical solution to fundamentally break through the above technical bottlenecks and upgrade the evaluation of professional personnel from static information display to dynamic and quantifiable decision support.

[0006] The most similar existing implementation scheme to the present application can be mainly divided into two categories: the first category is the online evaluation and ranking system widely used in the business field; the second category is the scientific metrology evaluation system long used in the academic field.

[0007] The first existing technical solution: online evaluation and ranking system based on internal platform data; the typical representative of this type of technology is various online medical, legal service or technical Q&A platforms. The technical solution can be decomposed into the following core modules in architecture:

[0008] 1. Data collection and preprocessing module: Obtain data through two ways, active registration submission (such as title, resume) and passive record of user behavior in the platform (such as consultation volume, score, text evaluation);

[0009] 2. Index quantification and feature engineering module: Convert the collected information into calculable numerical features. For example, map the text title "chief physician" to a numerical score through a pre-set mapping table, or aggregate and count user behavior data;

[0010] 3. Weighted scoring model module: Use a linear weighted model set by the platform operator, which can be expressed mathematically as: Where represents the i-th numerical feature, represents the pre-set weight of the feature;

[0011] 4. Sorting and display module: Sort the professionals in the platform in descending order according to the final calculated Score value and display it on the user interface;

[0012] Second type of existing technical solution: scientific metrology evaluation system based on citation network; this type of technology is mainly applied in the academic field, represented by "H index", "journal impact factor (JIF)" and others. The core logic of its technical solution is:

[0013] 1. Data collection and network construction: Collect structured data such as papers, authors, journals, and citations from professional academic databases (such as Web of Science, Scopus);

[0014] 2. Index calculation module: Build a citation network based on the collected data and run specific, fixed mathematical formulas or statistical indicators (such as H index algorithm) on this network to quantify the academic influence of scholars or institutions;

[0015] The above two types of existing technical solutions, although they meet the basic evaluation needs to some extent, but from a deeper technical perspective, they all have the following fundamental and difficult to overcome defects:

[0016] 1. One-sidedness of evaluation dimension and limitation of data basis: The data source of the first type of technology is limited to the platform, which is a typical "information island" and cannot reflect the individual's true value outside the platform. The data source of the second type of technology is limited to academic output, which is an "evaluation in the study", and cannot measure the individual's professional practice ability and social network value. Both of them fail to integrate the individual's multi-dimensional ability capital, leading to one-sided and distorted evaluation results;

[0017] 2. The evaluation dimension is one-sided and the data basis is limited: the data source of the first type of technology is limited to the platform and is a typical "information island" that cannot reflect the true value of individuals outside the platform. The data source of the second type of technology is limited to academic output, which is an "evaluation in the study" and cannot measure the professional practice ability and social network value of individuals. Both of them cannot integrate the multi-dimensional ability capital of individuals, resulting in one-sided and distorted evaluation results;

[0018] 3. Lack of unified evaluation scale and comparability: the scoring systems of different platforms are different, and the results are completely incomparable across platforms. Even the H-index is difficult to compare directly due to differences in subject areas.

[0019] 4. Static and retrospective nature of the system paradigm (most core defect): both types of existing technologies are essentially static data snapshots of past achievements of individuals or organizations. Their system architecture and algorithm design do not have the technical means to respond to virtual events such as core talent flow, simulate changes in system structure, and prospectively and quantitatively predict the possible impact of such events on the future. This makes such technologies completely useless when faced with future-oriented, dynamic, and predictive strategic problems, limiting their application value to the shallow level of "historical summary". SUMMARY

[0020] To solve the problems in the prior art, the present application provides a professional influence evaluation and dynamic deduction method based on a knowledge graph, comprising:

[0021] Step 1: Construct a structured knowledge set representing entities and relationships in the professional field;

[0022] Step 2: Based on the structured knowledge set of step 1, calculate the global influence absolute score of the target entity;

[0023] Step 3: Based on the global influence absolute score of step 2, perform context-aware situational influence analysis;

[0024] Step 4: Based on the perturbation simulation of the structured knowledge set of step 1, perform forward-looking influence deduction.

[0025] As a further improvement of the present application, the step 1 comprises:

[0026] Step S1, multi-source heterogeneous data acquisition: obtaining structured, semi-structured, and unstructured data about professionals and their related activities from multiple predefined online data sources;

[0027] Step S2, knowledge extraction: a unified extraction framework based on a pre-trained language model combined with specific prompt engineering is adopted, which identifies entities and extracts multi-dimensional relationship triples between them from long texts through a three-stage processing pipeline of summarization-identification-refinement, which conforms to the pre-defined mode;

[0028] Step S3, knowledge fusion and verification; for entities from different data sources pointing to the same real-world object, a multi-level conflict resolution framework based on signal priority is adopted for entity alignment and fact verification;

[0029] Step S4, knowledge storage: the structured knowledge after cleaning and fusion is loaded into a data storage system that can handle complex association relationships between processing entities.

[0030] The beneficial effects of the present application are: 1. The holographic and objective nature of the evaluation basis is realized, which fundamentally improves the credibility and fairness of the evaluation results; specific embodiment: the data basis of the prior art is "islandization" and "one-sidedness", or limited to the internal platform, or limited to a single academic dimension. The structured knowledge aggregation unit of the present application constructs a holographic knowledge graph containing individual academic, professional, social network and other multi-dimensional ability capital by automatically extracting and fusing information from the whole network multi-source heterogeneous data. The evaluation process is completely driven by data and algorithms, especially the "signal priority" multi-level conflict resolution framework, which ensures the accuracy of the facts and eliminates artificial intervention and subjective bias; this makes the evaluation conclusion of the present application no longer based on one-sided information, but on the accurate portrait based on panoramic data, its objectivity, comprehensiveness and credibility are completely beyond the reach of the prior art; 2. A unified, data-driven evaluation scale is creatively established, which realizes fair benchmarking across organizations and regions for the first time; specific embodiment: the scoring system of the prior art relies on subjective weighting or uses a rigid formula, resulting in non-transparent and non-comparable results. The individual ability quantification unit of the present application creatively calculates a "global influence absolute score (GIA)" that is not limited by any external context through a pre-trained comprehensive evaluation model (preferably GNN) that can capture complex nonlinear relationships. This score provides a unified, comprehensive ability-based "ability measurement" for every professional in the world. This makes it possible to make fair and objective horizontal comparisons between professionals in different institutions, different regions, and even different countries, solving the fundamental problem of chaotic evaluation standards and the inability to benchmark in the industry; 3. Provide context-aware multi-level analysis capability, making the evaluation have unprecedented depth and fineness; specific embodiment: the prior art can only provide a general, single-level ranking. The context influence analysis unit of the present application can provide a multi-level ranking based on the unified ability scale, which can be used to analyze the influence of different contexts on the evaluation results, and to provide a more accurate and comprehensive evaluation conclusion. ​The ruler dynamically performs penetrating analysis according to an arbitrary range (from a microscopic department to a macroscopic country, or even a specific technical field) specified by a user, and adopts a pre-calculation and cache strategy to ensure instantaneous response of the query; this capability enables evaluation to be no longer one-size-fits-all, but can accurately answer "the person is an absolute core in the hospital, but what level is he in the city?" or "in the 'AI pharmaceutical' sub-track, what is his industry status?" Such complex questions with depth and context provide users with more insightful, multi-dimensional decision-making basis; 4. The paradigm breakthrough from static evaluation to dynamic deduction is realized, and the system is endowed with forward-looking decision support capability (core beneficial effect); The paradigm breakthrough from static evaluation to dynamic deduction is realized, and the system is endowed with forward-looking decision support capability (core beneficial effect); This makes the present application not limited to "evaluating the present situation", but can "predict the future". It changes the evaluation system from a static information display tool to a dynamic "sand table deduction" engine that can support talent introduction risk assessment, team optimization, organizational structure adjustment and other major strategic decisions. This leap from "descriptive analysis" to "predictive and instructive analysis" represents a generational progress in technology application, providing organizational managers with unprecedented data-driven decision-making weapons. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a whole function module architecture diagram of the system of the present application;

[0032] Figure 2 is a workflow schematic diagram of the structured knowledge aggregation unit of the present application;

[0033] Figure 3 is a calculation flow schematic diagram of the individual ability quantification unit of the present application;

[0034] Figure 4 is a logic schematic diagram of the situational influence analysis unit of the present application;

[0035] Figure 5 is a core flow schematic diagram of the system disturbance deduction unit of the present application;

[0036] Figure 6 is a simplified medical field knowledge graph Schema example diagram of the present application. DETAILED DESCRIPTION

[0037] In view of the technical problem in the background art that the evaluation system can only make static and retrospective summary of historical data and cannot make prospective and quantifiable prediction of future events, thereby resulting in limited application value, the purpose of the present application is to provide a professional influence evaluation and dynamic deduction system and method, which aims to break through the paradigm limitation of the prior art that can only make static evaluation, and endows the evaluation system with prospective and quantifiable dynamic deduction and decision support capability.

[0038] 1. The present application aims to provide a technical solution capable of automatically collecting, extracting and fusing information from multi-source heterogeneous data, and structurally representing it as a unified and machine-readable knowledge network, thereby laying a comprehensive and reliable data foundation for subsequent objective evaluation and dynamic deduction;

[0039] 2. In view of the technical problem in the prior art that the evaluation result is inaccurate and not objective due to the simple linearity or rigidity of the evaluation model, the present application aims to provide a technical solution capable of combining high-dimensional and heterogeneous feature vectors into an absolute influence power index that is not limited by external context and can be globally compared, based on the aforementioned knowledge network, using a comprehensive evaluation model (such as the preferred graph neural network model) that can effectively learn network structure information, thereby establishing a unified and objective evaluation scale;

[0040] 3. In view of the technical problem in the prior art that the evaluation result is inaccurate and not objective due to the simple linearity or rigidity of the evaluation model, the present application aims to provide a technical solution capable of combining high-dimensional and heterogeneous feature vectors into an absolute influence power index that is not limited by external context and can be globally compared, based on the aforementioned knowledge network, using a comprehensive evaluation model (such as the preferred graph neural network model) that can effectively learn network structure information, thereby establishing a unified and objective evaluation scale;

[0041] 4. In view of the core technical defect that the prior art is essentially a summary of historical data and cannot make prospective and quantifiable prediction, the present application aims to provide a core technical solution based on structured knowledge set perturbation simulation, which can prospectively and quantitatively deduce the specific influence of a key event on related organizations in multiple macro dimensions by simulating the structural or parameter changes caused by the event on a virtual copy of the knowledge network, thereby transforming the evaluation system from a static information display tool to a dynamic deduction engine that can support major strategic decisions.

[0042] The present application discloses a professional influence evaluation and dynamic deduction method based on a knowledge graph, comprising:

[0043] Step one: Construct a structured knowledge set representing entities and relationships in the professional field; this step aims to automatically convert multi-source, heterogeneous raw data into a high-quality structured knowledge model that provides data support for subsequent calculations. In a preferred embodiment, the structured knowledge set is embodied as a domain knowledge graph, and its construction process includes:

[0044] Step S1, multi-source heterogeneous data collection: through the configuration of network crawlers or API interfaces, unstructured or semi-structured data about professionals and their related activities is obtained from multiple predefined online data sources (including but not limited to official websites, academic databases, patent databases, news media, etc.).

[0045] Step S2, knowledge extraction: a unified extraction framework based on a pre-trained language model (preferably a large language model LLM) combined with specific prompt engineering is used. This framework uses a three-stage processing pipeline of summarization-recognition-refinement to accurately identify entities (such as [professional], [organization], [academic achievement], etc.) that meet the pre-defined schema from long text and extract multi-dimensional relationship triples (for example, <head entity, relationship type, tail entity>).

[0046] Step S3, knowledge fusion and verification: for entities from different data sources pointing to the same real-world object, a multi-level conflict resolution framework based on "signal priority" is used for entity alignment and fact verification. This framework comprehensively utilizes metadata (such as timestamps, data source authority), internal graph topology consistency, and LLM-based context reasoning to ensure the uniqueness, accuracy, and completeness of the final knowledge;

[0047] Step S4, knowledge storage: the cleaned and fused structured knowledge is loaded into a data storage system (such as a graph database or a relational database through multi-table association) that can efficiently handle complex relationships between entities.

[0048] Step two: based on the structured knowledge set of step one, calculate the absolute score of the global influence of the target entity; this step aims to calculate a quantitative ability index for each professional entity in the knowledge graph that is not limited by external context, objective, and globally comparable. Its calculation process includes:

[0049] Step 1, multi-dimensional ability capital variable calculation: based on the structured knowledge set, a pre-defined multi-dimensional ability capital feature vector is calculated for each professional entity through structured queries (such as Cypher queries) or graph algorithms. This vector contains at least the following three-dimensional variables:

[0050] • Academic / Knowledge Capital Variables: e.g. high-impact paper scores weighted by journal impact factor and citation count, advanced research project lead scores, etc.

[0051] • Professional / Skill Capital Variables: e.g. integrated experience score, maximum difficulty coefficient of problems handled, degree of pioneering technical contribution, etc.

[0052] • Social / Network Capital Variables: e.g. influence score in authoritative academic organizations, centrality (preferably Betweenness Centrality) in collaboration network, etc.

[0053] Step 2, Score Synthesis: input the normalized multi-dimensional capability feature vector into a pre-trained comprehensive evaluation regression model for end-to-end non-linear mapping, and finally output a scalar value, i.e. global influence absolute score (S) ). In a preferred embodiment, the model is a graph neural network (GNN, such as Graph Attention Network, GAT), which takes the feature vector as the initial input of the node, and through information propagation and aggregation on the collaboration network subgraph of the professional, it can effectively capture the network structure information, so as to obtain more accurate scores.

[0054] Step Three: Context-aware Situational Influence Analysis Based on Global Influence Absolute Score in Step Two; this step aims to achieve multi-level and penetrating analysis of individual influence based on a unified evaluation scale. Its implementation includes:

[0055] Step y1, Dynamic Definition of Evaluation Range: receive one or more context parameters input by the user, which are used to define a specific evaluation range (for example, organizational range "XX Hospital XX Department", or geographical range "XX City").

[0056] Step y2, Evaluation Group Screening: convert the context parameters into a structured query to the structured knowledge set to obtain all professional entities that meet the range conditions, forming a temporary evaluation group.

[0057] Step y3, Relative Position Calculation: obtain the global influence absolute score (S ) of all members in the evaluation group, and calculate the ranking (Rank) and percentile (Percentile) of the target professional entity in the group, thereby outputting its relative influence in the specific context.

[0058] Step Four: Prospective Influence Derivation Based on Disturbance Simulation of the Structured Knowledge Set in Step One; this step is the core innovation point of the present application, aiming to realize the leap from static evaluation to dynamic prediction. The derivation process includes:

[0059] Step s1, organization macroscopic index definition and baseline calculation: a set of macroscopic variable system capable of measuring the comprehensive ability of the organization (such as department, hospital) (for example, total amount of human capital, proportion of top-notch talents, internal collaboration network density, etc.) is defined in advance. Before the deduction starts, first calculate all the macroscopic variable values of the organization to be analyzed in the current state as the baseline (Baseline).

[0060] Step s2, non-persistent disturbance simulation: in response to a simulated event (such as "a core personnel flows from A organization to B organization"), a non-persistent, virtual modification is made to the structured knowledge set in a temporary, isolated computing environment (such as memory graph copy). The modification includes at least one of the following:

[0061] First, topological structure modification: temporarily increase, delete or modify the edge representing the relationship between entities (for example, disconnect the-[employed]-> relationship between the personnel and A organization, and establish the-[employed]-> relationship between the personnel and B organization).

[0062] Second, attribute / weight modification: temporarily modify the numerical attribute or weight parameter associated with the relevant entity or relationship to simulate the change in its state in the calculation.

[0063] Third, reachability / flow logic modification: modification at the algorithm level. For example, when calculating the internal network density of organization B after disturbance, the system temporarily includes personnel P in the node set of organization B in the program logic of the graph algorithm, even if the edge of its affiliation has not been established at the data level. This way directly simulates the change of affiliation in the calculation logic.

[0064] Step s3, post-disturbance recalculation and impact quantification analysis: based on the disturbed structured knowledge set, recalculate all the macroscopic variables of the organization to get the new value (New_Value). Finally, by calculating the change rate , a clear and quantitative impact analysis report is generated to accurately show the predicted impact of the simulated event on the relevant parties in multiple dimensions.

[0065] Through the cooperative work of the above three steps, the present application completely constructs a technical closed loop from data collection, knowledge construction to static evaluation, dynamic deduction, and the technical scheme is logically rigorous, the steps are clear, the implementability is strong, and significant beneficial effects can be produced.

[0066] As shown in Figure 1 , as the second embodiment of the present application, this embodiment is illustrated in the medical field, but this should not be understood as limiting the application field of the present application.

[0067] Module 1: Structured Knowledge Aggregation Unit, as shown in Figure 2 .

[0068] The first step in the method of the present application, i.e. "constructing or obtaining a structured knowledge set representing entities and their relationships in a professional field", this unit is the cornerstone of the entire system, and its core task is to automatically convert multi-source, heterogeneous raw data into a high-quality, high-accuracy, high-completeness structured knowledge model, providing a solid and reliable data foundation for the effective operation of all subsequent evaluation and deduction modules. Its specific implementation scheme includes the following key technical steps and considerations.

[0069] Step S1, Multi-source Heterogeneous Data Collection: This step aims to obtain as comprehensive raw data as possible from the Internet and internal databases.

[0070] (1) Definition and classification of data sources

[0071] The system pre-configures and maintains an extensible data source list, which at least includes:

[0072] • High-value structured / semi-structured data sources: including but not limited to domestic and foreign academic literature databases (such as PubMed, Web of Science, Google Scholar, China Knowledge Network), patent databases (such as USPTO, EPO, CNIPA), clinical trial registration centers (such as ClinicalTrials.gov), official agency directories (such as the National Health Commission physician registration information query system).

[0073] • High timeliness unstructured data sources: including but not limited to official websites, department homepages, news portals, industry vertical media, professional personal blogs or public social media pages of major hospitals and research institutes.

[0074] (2) Adaptation and execution of collection strategy: different collection strategies are adopted for different types of data sources:

[0075] In step S1, for API interface: for data sources that provide API (such as PubMed), the system uses the API client written with preset query keywords (such as personnel name, institution name, disease name) and update frequency to perform periodic and incremental API calls to obtain structured data.

[0076] In step S1, for web data: a configurable distributed web crawler cluster based on Scrapy or similar frameworks is used. To achieve efficient and stable collection, the crawler cluster has the following technical features:

[0077] Dynamic IP proxy pool: Integrate multiple commercial or self-built IP proxy services to achieve millisecond-level switching of request IPs to cope with IP-based access frequency limits;

[0078] Dynamic User-Agent rotation: Maintain a list of hundreds of mainstream browser and mobile device User-Agents, randomly select one for each request to simulate real user behavior;

[0079] Adaptive request delay and session maintenance: Adjust request rate dynamically based on target server response status codes (such as 200, 429, 503) and response time, and simulate login process to maintain session if necessary.

[0080] 2. Step S2, knowledge extraction: This step uses a unified extraction framework based on a large language model (LLM) combined with specific prompt engineering to achieve high-precision and efficient entity and relationship extraction.

[0081] (1) Long text preprocessing pipeline:

[0082] To solve the LLM input length limit, reduce the call cost and improve the extraction accuracy, all collected long texts (such as biographies, news long texts) will go through a "summary-identification-refinement" three-stage preprocessing pipeline before entering the extraction process:

[0083] Step a, coarse-grained summary and key information area positioning: First, use an unsupervised extractive summary algorithm (such as TextRank) or a lightweight generative summary model (such as T5-small) to compress long texts into "long summaries" containing core information. In this process, the system will record the location index of each selected sentence in the original text for subsequent backtracking verification.

[0084] Step b, semi-structured knowledge extraction based on LLM: Embed the generated "long summary" text into a pre-designed unified Prompt template for entity and relationship extraction. The Prompt uses "In-Context Learning" and "Chain-of-Thought (CoT)" design patterns to guide the LLM to output all entities and relationships in the text that meet the pre-defined Schema in JSON format at once.

[0085] Step c, backtracking the refined and context-verified original text: for each knowledge point (entity or relation) extracted by LLM, the system quickly locates its context fragment in the original text using the position index recorded in the first step. Then, through a verification prompt, it asks LLM to combine more rich original context to confirm, complete or correct the extraction results, thereby greatly improving the accuracy of extraction.

[0086] (2) Extensible knowledge graph Schema design, as shown in Figure 6 : The system pre-defines a hierarchical and extensible domain knowledge graph schema (Schema), which not only specifies the types of entities and relations, but also defines their attributes, data types and constraints.

[0087] Entity layer (Nodes): For example, in the medical field, [doctor], [medical institution], [paper], [journal], [disease], [treatment / technology], [research project], [society / association] and other core entity types are defined, and for each entity type, its must-have or optional properties (Properties) are specified, for example, [journal] entity must have “impact factor” (floating point type) and “partition” (character type) attributes.

[0088] Relationship layer (Edges): Detailed definition of relationship types between entities, such as [employed by], [graduated from], [published], [co-author is], [good at], [hosted], etc. Some relationships can also have attributes, for example, the [published] relationship has “author order” (such as 'first author', 'corresponding') and “publication year” attributes.

[0089] Step S3, knowledge fusion and verification: This step aims to solve the information conflict and redundancy problem from different data sources, and ensure the uniqueness, accuracy and integrity of knowledge, including:

[0090] Step c1, Entity Alignment: For entities from different data sources that may point to the same real-world object (e.g. “XX University Third Hospital” vs. “Beijing Medical Third Hospital”), the system adopts a hybrid alignment strategy:

[0091] Step c10, rule-based fast matching: First, a pre-set synonym dictionary (such as institution aliases, common name writing methods) and string similarity algorithm (such as Jaro-Winkler) are used for a round of fast and high-confidence matching;

[0092] Step c11, similarity calculation based on representation learning: for ambiguous cases that cannot be resolved by rules, the system calculates the cosine similarity between the attribute vector (based on textual attributes) and the graph embedding vector (based on its neighbor structure in the graph, calculated by models such as GraphSAGE) of the entities to be matched, respectively;

[0093] Step c12, pairwise judgment based on pre-trained language model (LLM): input the descriptive information of the two entities into a "pairwise judgment" Prompt, and make the final judgment similar to that of a human expert by LLM.

[0094] Step c2, "signal priority" multi-level resolution framework for fact conflicts: when there are multiple conflicting facts about the same attribute of the same entity (for example, source A says a doctor is "chief physician", and source B says "deputy chief physician"), the system starts a multi-level resolution framework, including:

[0095] Level 1, priority rules based on metadata: first apply hard rules. Prefer information with the latest timestamp; if the timestamps are similar, prefer information from data sources with higher preset authority scores.

[0096] Level 2, topological verification based on internal consistency of the graph: if the rules cannot be resolved, check which fact can form a logical closed loop with more existing high-confidence knowledge in the graph. For example, if the author's unit of the doctor's recent published papers is A hospital, the confidence of the fact that "he is employed by A hospital" will be much higher than "he is employed by B hospital".

[0097] Level 3, context-based comprehensive reasoning and resolution by LLM: for the most complex conflicts, the system organizes all conflicting facts and their metadata, related context evidence into a structured text, and passes it to an LLM in the role of a "fact investigator" for comprehensive reasoning, and requires it to output the final resolution result and detailed reasoning process.

[0098] Step S4, knowledge storage: this step aims to persistently store the high-quality knowledge after final cleaning and fusion.

[0099] (1) Definition of logical data model: the core of the invention is a logical data model that can represent a set of entities (Entities) and the many-to-many relationships (N-ary Relationships) between them.

[0100] (2) Implementation of physical storage: those skilled in the art should understand that the above logical data model can be physically implemented as (but not limited to):

[0101] Step S40: A Property Graph Database: This is a preferred embodiment of the invention, such as using Neo4j, JanusGraph, etc. Entities are mapped to nodes, and relationships are mapped to edges with direction and attributes, which can most efficiently support subsequent graph queries and graph algorithms.

[0102] Step S41: An RDF Triple Store: For example, using Apache Jena.

[0103] Step S42: A set of data tables that are related to each other in a relational database (such as MySQL, PostgreSQL) through foreign keys: for example, through a "personnel table", a "paper table" and a "personnel-paper relationship table".

[0104] Step S43: One or more JSON or XML documents: stored in a document database (such as MongoDB), representing the relationships between entities through nesting or ID references.

[0105] Step S44: A set of object instances in computer memory: interconnected by pointers or references, for high-performance real-time computing scenarios.

[0106] Through the five tightly coupled steps described above, the structured knowledge aggregation unit of this invention can construct and maintain a high-quality set of structured knowledge that can be used for serious evaluation and deduction, and provides a solid and reliable data foundation for the effective operation of all subsequent modules.

[0107] Module Two: Individual Ability Quantification Units, such as Figure 3 As shown.

[0108] Step two in the method of this invention, namely "calculating the absolute global influence score of the target entity based on the structured knowledge set in step one," is the core computing engine of the entire system. Its task is to calculate one or more objective, globally comparable quantitative capability indicators for each professional entity in the knowledge graph, unaffected by any external context (such as organizational affiliation or geographical location). In this embodiment, we will calculate a comprehensive "absolute global influence score" (…). This will be illustrated using the example of "". The calculation process for this score is logically rigorous and data-driven, forming the technological foundation for all subsequent evaluation and deduction functions. Its specific implementation plan includes the following three key technical steps:

[0109] Step 1, Refined Calculation of the Multidimensional Capability Capital Variable System:

[0110] This step aims to convert the discrete, multi-modal information stored in the knowledge graph into a comprehensive, computable, high-dimensional feature vector that reflects the individual's overall ability. This feature vector is composed of a pre-defined multi-dimensional ability capital variable system.

[0111] (1) Definition of variable system:

[0112] This system contains at least the following three capital dimensions, each of which governs multiple specific variables that can be accurately calculated by structured knowledge sets.

[0113] Dimension One: Academic / Knowledge Capital: Measures the individual's ability in knowledge creation, academic contribution, and frontier exploration.

[0114] Variable V_Paper_Impact (High Impact Paper Score): Through structured queries, traverse all [academic achievements] entities with the target entity as the core author (such as author order 'first author' or 'corresponding author'), and perform weighted summation according to the preset impact score of the [journal] entity (such as impact factor, partition, citation number, etc.). Its formula can be expressed as:

[0115] .

[0116] Where, is the author order weight, is the citation number weight.

[0117] Variable V_Project_Lead (High Level Project Lead Score): Through structured queries, filter out [scientific research projects] entities with the target entity as the core role (such as 'chief scientist' and 'project leader'), and perform weighted summation according to the project level (such as 'national level' and 'provincial and ministerial level') and funding size.

[0118] Variable V_Patent_Value (Patent Value Score): Count the number of [patent] entities with the target entity as the inventor, and perform comprehensive weighting according to the patent type (such as invention patent and utility model), patent family size, citation number, and whether there is technology transfer, etc.

[0119] Dimension Two: Professional / Skill Capital: Measures the individual's practical experience, technical level, and ability to solve complex problems in the professional field.

[0120] Variable V_Experience_Comprehensive (Comprehensive Experience Score): A composite variable composed of multiple sub-attributes, calculated by the formula Computed. Where YearsOfPractice and TitleLevelScore are directly obtained from the attributes of the [Professional] entity, and the weights Can be calibrated by Analytic Hierarchy Process (AHP) or by a committee of domain experts.

[0121] YearsOfPractice and TitleLevelScore are directly obtained from the attributes of the [Professional] entity, and the weights Can be calibrated by Analytic Hierarchy Process (AHP) or by a committee of domain experts.

[0122] Variable V_Skill_Pioneering (Technical Pioneering Contribution): Count the number of [Therapy / Technology] entities that the target entity is the pioneer or first introducer of. This is a high-weighted sparse variable to identify the technology leaders in the field.

[0123] Dimension Three: Social / Network Capital: Measures the individual's reputation, status, and ability to connect and mobilize resources within the industry.

[0124] Variable V_Network_Appointment_Influence (Academic Appointment Influence Score): Through structured query, traverse the [Academic Appointment] relationships of the target entity in various [Society / Association] entities, and add up the weights according to the level of the society (such as international, national, provincial) and the weight of the position (such as chairman, vice-chairman, member).

[0125] Variable V_Network_Centrality (Cooperation Network Centrality): This is a pure graph algorithm driven variable. The system first constructs a cooperation network subgraph in the structured knowledge set, containing only [Professional] entities and their [Cooperator is]-> relationships. Then, on this subgraph, calculate the Betweenness Centrality for each node. This index measures to what extent a node is a "bridge" in the shortest path between other nodes in the network, the higher the score, the stronger the node's role as a hub in knowledge dissemination and resource connection. Those skilled in the art will understand that other graph centrality algorithms such as Eigenvector Centrality or Degree Centrality can also be used as supplements or alternatives.

[0126] Step 2, construction and normalization of multi-variable feature vector, including:

[0127] Step b1, constructing feature vector: after the calculation of all the above variables is completed, the system will generate a unified dimension high-dimensional static ability feature vector composed of the calculation results of these variables for each professional entity in the knowledge graph .

[0128] Step b2, data normalization: due to the huge difference in the dimension and numerical range of different variables, before being fed into the final synthetic model, each dimension of the feature vector must be normalized to eliminate the influence of dimensional differences and improve the training efficiency and stability of the model. The preferred embodiment of the present application adopts the Z-Score standardization method to convert the values of all variables into a standard normal distribution with a mean of 0 and a standard deviation of 1. Those skilled in the art can also select other normalization methods as needed, such as Min-Max Scaling.

[0129] Global influence absolute score (S) Synthetic model of global influence absolute score (S) This step is to synthesize the normalized multi-dimensional feature vector into a single, comparable scalar score

[0130] (1) Model selection and architecture: preferred embodiment: graph neural network (GNN) model

[0131] Technical route: the preferred embodiment of the present application adopts a graph attention network (Graph Attention Network, GAT). The model takes the feature vector generated in the previous step as the initial embedding representation of each node in the graph. Through multi-layer information propagation on the collaboration network subgraph, each node can selectively and weightedly aggregate the information of its neighbor nodes to update its own representation. This attention mechanism enables the model to automatically learn the deep network structure effect of “who to cooperate with is more important than how many times to cooperate”.

[0132] Model architecture: the model contains several graph attention layers, each followed by a nonlinear activation function (such as LeakyReLU) and a Dropout layer. Finally, through a global pooling layer (Global Pooling Layer) and one or two fully connected layers (Fully Connected Layers), the final node embedding representation is mapped to a scalar score .

[0133] Alternative embodiment: other machine learning regression models

[0134] Those skilled in the art should understand that in the simplified scenario without using network structure information, other traditional machine learning regression models such as gradient boosting decision tree (XGBoost, LightGBM) or a multi-layer perceptron (Multi-Layer Perceptron, MLP) can also be used. These models will map the feature vector X i As a single data point, a mapping function from features to scores is fitted by supervised learning.

[0135] (2) The training and deployment of the model include:

[0136] Training data step: The training of the model is not performed at each query, but is completed during the system deployment or periodic update phase. The training data comes from a high-quality dataset annotated by multiple senior experts in the field using the back-to-back scoring method to take the average value. The dataset contains hundreds to thousands of feature vectors of professional personnel and their corresponding "expert consensus scores".

[0137] Training process: Use mean squared error (Mean Squared Error, MSE) or smooth L1 loss (Smooth L1Loss) as the loss function, and use Adam or AdamW optimizer to train and optimize the model parameters until the performance of the model on the validation set converges.

[0138] Running and output: In actual operation, the system only needs to send the normalized feature vector of the person to be evaluated (and its neighbor information if GNN is used) into this trained model for an efficient forward propagation, and the final global influence absolute score can be calculated. Its score range can be normalized to 0 to 100.

[0139] Through the rigorous design and implementation of the above three steps, the individual ability quantification unit of the present application can condense the rich and multi-dimensional information in the knowledge graph into a single score that is objective, fair and has a high degree of discrimination, laying a solid and reliable quantitative foundation for all subsequent multi-level evaluation and dynamic deduction applications.

[0140] Module three: context influence analysis unit, as shown in Figure 4

[0141] Step three in the method of the present application, i.e. "context-aware context influence analysis based on the global influence absolute score", is realized by the context influence analysis unit in the system. The core task of this unit is to perform context influence analysis on the "global influence absolute score"​ ) as a unified, invariant evaluation scale, enabling individual influence, context-aware, multi-level, and penetrating analysis. To improve system response speed and user experience, the embodiment preferably adopts a "pre-computation and caching" strategy rather than real-time computation at each query. The specific implementation includes the following key technical steps.

[0142] 1. Pre-definition and configuration of multi-level evaluation system:

[0143] This step aims to establish a structured and flexible configurable evaluation hierarchy system as the basis for subsequent batch computation.

[0144] Step y1, dynamic definition of evaluation range:

[0145] Step y10, definition of evaluation hierarchy system:

[0146] The system pre-defines a set of standard evaluation tiers that can be tree-like or network-like structures. In the medical field embodiment of the invention, the hierarchy system at least includes:

[0147] Organizational Tiers:

[0148] Tier 1 (micro level): specific department in the hospital (e.g., XX Hospital - Department of Cardiology)

[0149] Tier 2 (mesoscopic level): single medical institution (e.g., XX Hospital)

[0150] Geographical Tiers:

[0151] Tier 3 (city level): the city where the target entity is located (e.g., XX City)

[0152] Tier 4 (regional / provincial level): the province or defined economic / geographical region where the target entity is located (e.g., North China)

[0153] Tier 5 (national level): nationwide (e.g., China)

[0154] Domain-specific Tiers:

[0155] Tier 6 (sub-field): specific professional field defined based on the relationship between [professional] and [disease / technology] (e.g., all professional groups proficient in "coronary heart disease intervention").

[0156] Step y11, configuration management:

[0157] The evaluation hierarchy is stored in a configuration file (such as YAML or JSON file) or database table that can be accessed and modified by the system administrator. This design allows flexible addition (such as adding a "global" level), modification or disablement of specific levels according to business needs, without modifying the core calculation code of the system, ensuring the scalability and maintainability of the system.

[0158] Step y2, efficient group data extraction based on batch graph query and cache mechanism:

[0159] The core of this step is how to efficiently obtain the benchmark group data of all personnel in all preset levels in the knowledge graph at one time.

[0160] Step y20, batch processing flow:

[0161] This module uses an offline periodic batch calculation task triggered by a timing task (such as Cron Job) to execute. The task will traverse all entity nodes of type [professional personnel] in the knowledge graph.

[0162] Step y21, query cache and optimization (key performance guarantee):

[0163] To avoid huge calculation redundancy and database load caused by repeated queries of the same range of group data, this embodiment introduces a query result cache mechanism based on key-value storage (such as Redis, Memcached).

[0164] Technical route:

[0165] Step y210, unique level instance extraction: when the batch task starts, first traverse all [professional personnel] entities to extract all unique level instance lists they belong to (for example, all unique department names, all unique city names, etc.).

[0166] Step y211, batch query and cache of group data: the system only performs a structured query on each of these unique level instances once. The purpose of this query is to obtain the ID of all members under this level and their corresponding global influence absolute score ( ) list. The query result is serialized and stored in the cache with the unique identifier of the level instance as the key (Key). For example, Key: "tier:department:XX Hospital-Cardiology", Value: "[{ 'id': 'p1','s_abs': 98.5}, { 'id': 'p2','s_abs': 95.4}, …]".

[0167] Step y212, cache life cycle management: the life cycle (TTL, Time-To-Live) of the cache should match the execution cycle of the batch task to ensure the timeliness of the data. For example, if the batch task is executed at 4 a.m. every day, the validity period of the cache can be set to 24 hours.

[0168] Step y3, calculation and storage of global relative influence index:

[0169] After obtaining all the relevant group data, the system calculates the complete and multi-level relative influence portrait of each professional and stores it as a node attribute.

[0170] Step y30, batch calculation process: the system traverses each [professional] entity p in the knowledge graph:

[0171] Step y301: create a JSON object for entity p to store its complete influence portrait, denoted as InfluenceProfile.

[0172] Step y302: traverse all predefined evaluation tiers (Tier 1 to Tier 6).

[0173] Step y303: for each tier, first determine the specific instance to which the entity p belongs (e.g., the department A it belongs to, the city B, etc.).

[0174] Step y304: for each tier, first determine the specific instance to which the entity p belongs (e.g., the department A it belongs to, the city B, etc.).

[0175] Step y305: in this list, locate the position of entity p through efficient algorithms such as binary search, to calculate its rank (Rank) and percentile (Percentile) in that tier. The percentile can be calculated by the formula to ensure that the percentile of the highest score is 100. The structured results containing range name, rank, total number of people in the group, percentile, etc. are stored as a sub-object in InfluenceProfile.

[0176] Step y31, data storage and update (pre-computation strategy):

[0177] Technical route: after calculation, the InfluenceProfile JSON object containing all tier influence indexes will be written directly as a new attribute to the node of the [professional] entity p.

[0178] Storage example (p.InfluenceProfile property added on a doctor node):

[0179] {

[0180] "department": {

[0181] "scope": "XX Hospital - Cardiology Department",

[0182] "rank": 1,

[0183] "total": 25,

[0184] "percentile": 100.0,

[0185] "s_abs": 98.5

[0186] },

[0187] "hospital": {

[0188] "scope": "XX Hospital",

[0189] "rank": 5,

[0190] "total": 850,

[0191] "percentile": 99.41,

[0192] "s_abs": 98.5

[0193] },

[0194] "city": {

[0195] "scope": "XX City",

[0196] "rank": 15,

[0197] "total": 18500,

[0198] "percentile": 99.92,

[0199] "s_abs": 98.5

[0200] },

[0201] "last_updated": "2025-07-18T04:00:00Z"

[0202] }

[0203] Advantages and precautions: this "pre-computed and stored" strategy is a typical space-time trade-off solution. It places complex and time-consuming batch computing tasks in the background offline, which is not felt by the user. When the upper application (such as user interface or API) needs to query the complete influence portrait of a professional, only a simple attribute reading operation is needed for the node of the person, and the evaluation results of all levels can be obtained instantly in milliseconds, greatly improving the response speed of the front-end application and user experience, and significantly reducing the real-time query pressure of the database.

[0204] Through the rigorous design and implementation of the above three steps, the context influence analysis unit of the application can systematically and efficiently build a complete and multi-level relative influence portrait for each member in the knowledge graph, greatly improving the depth and breadth of evaluation, and providing strong performance guarantee for the instant call of the upper application.

[0205] Module four: system disturbance driving unit, as shown in Figure 5

[0206] Step four in the method of the application, that is, "based on the perturbation simulation of the structured knowledge set, the forward-looking influence deduction", is realized by the system disturbance deduction unit in the system. This step is the most creative part of the application, and its core task is to realize the paradigm breakthrough from "static evaluation" to "dynamic prediction". It accurately and quantitatively deduces how a simulated event (such as the inter-organizational flow of core professionals) will have a multi-dimensional impact on related parties through controllable and non-destructive perturbation simulation (Perturbation Simulation) on the structured knowledge set. This module is a key bridge connecting data analysis and high-level strategic decision-making, and its specific implementation scheme includes the following key technical steps.

[0207] Step s1, definition and baseline calculation of organizational macro capability variable system:

[0208] Before any deduction, a set of macro index system capable of measuring the capability of the organization (such as hospital, department, research team) must be established.

[0209] (1) Definition of macro variable:

[0210] The system predefines a set of configurable macro variables that can fully reflect the comprehensive strength of the organization. These variables can be obtained by aggregating queries or graph algorithm calculations on the structured knowledge set. The core variables include at least:

[0211] Talent capital index:

[0212] ​V_Org_Talent_Sum: Talent capital total. Calculate the sum of the absolute scores of all members in the organization. Reflects the "total firepower" or overall strength of the organization.

[0213] V_Org_Talent_Mean: Talent capital mean. Calculate the standard deviation of all members in the organization. Reflects the dispersion or balance of talent levels within the organization.

[0214] V_Org_Talent_TopRatio: Top talent ratio. Calculate the proportion of the number of people whose scores are in the top 1% (or other configurable threshold) in the global to the total number of people in the organization. Measures the enrichment of core talent within the organization.

[0215] Output and prestige indicators:

[0216] V_Org_Output_HighImpact: High-impact output total. Aggregate and calculate the weighted sum of high-impact papers, high-level projects, high-value patents, etc. produced by all members in the organization in the past period, reflecting the academic or technical prestige of the organization.

[0217] Network structure indicators (core):

[0218] V_Org_Network_Density: Internal collaboration network density. Calculate the density (Density) of the graph on the internal collaboration subgraph of the organization, which is the ratio of the actual number of edges to the theoretically maximum number of edges. Measures the tightness of cooperation between internal members, and is a reflection of team cohesion and knowledge sharing efficiency.

[0219] V_Org_Network_Cohesion: Internal network cohesion. Calculate the number of connected components (Connected Components) of the internal collaboration subgraph of the organization. If the number is greater than 1, it indicates that the organization is divided into "small groups" that do not cooperate with each other, and this indicator is a key measure of whether the organization has the risk of internal friction.

[0220] V_Org_Network_BridgeCount: External connection bridge count. Count how many members in the organization have cooperative relationships in other important organizations, and these members are the "bridges" for the organization to obtain external information.

[0221] Baseline calculation process in step s1:

[0222] ​​​When a scenario to be simulated is received, which contains at least one simulated event (e.g. person P flows from organization A to organization B), the system performs the following operations:

[0223] Step s10: On the current unmodified, real structured knowledge set, perform a graph query for each of organization A and organization B respectively, to filter out their respective member groups.

[0224] Step s11: Based on the two groups, calculate the current values of all the macro variables mentioned above.

[0225] Step s12: Store the calculated macro variable values of organization A and organization B as baseline (Baseline) in memory or temporary database for subsequent comparison.

[0226] Step s2: Non-destructive perturbation simulation based on in-memory graph copy.

[0227] To ensure that the simulation process does not affect the online real knowledge graph data and can achieve high-performance real-time calculation, all simulation operations are performed in a temporary, isolated in-memory environment.

[0228] (1) Technical route: In-Memory Graph Copy

[0229] When a simulation request arrives, the system will not directly manipulate the physically stored graph database. Instead, it will perform an efficient graph query to load the subgraph related to the simulation (i.e. the local network containing the entity to be perturbed, related organization entities, and their respective members and first / second-order neighbor relationships) into the server's memory. This forms a lightweight, temporary graph copy. Those skilled in the art will understand that for very large-scale graphs, more advanced graph sampling techniques (such as GraphSAINT) can also be used to obtain a representative and controllable-scale calculation subgraph.

[0230] (2) Implementation of perturbation operations (core protection point):

[0231] The system performs non-persistent, virtual modifications on the structured knowledge on this in-memory graph copy to simulate events. This modification includes at least one or a combination of the following ways to block all possible circumvention paths:

[0232] First, topological structure modification: directly modify the adjacency relationship of the graph. For example, to simulate the flow of personnel P, the system will temporarily delete the edge representing the membership relationship between P and the outflow organization A ((P)-[r: employed by]->(A)), and temporarily create a new edge representing the membership relationship between P and the inflow organization B ((P)-[r': employed by]->(B)).

[0233] Second, attribute / weight modification: Instead of changing the topology, the system simulates the state change by modifying the parameters. For example, to simulate a person "leaving the organization but still cooperating", the system can temporarily change the "state" attribute of the edge between the person and the original organization from "in-service" to "consultant", and adjust the weight factor of the person in the calculation of the organizational macro variable from 1.0 to 0.2. To simulate "relationship break", the weight of the person can be temporarily set to 0.

[0234] Third, reachability / flow logic modification: modification at the algorithm level. For example, when calculating the internal network density of organization B after the disturbance, the system temporarily includes person P in the node set of organization B in the program logic of the graph algorithm, even if the membership edge has not been established at the data level. This way directly simulates the change of affiliation in the calculation logic.

[0235] Step s3, re-calculation of post-disturbance indicators and quantification of impact analysis:

[0236] After the copy of the graph is successfully "disturbed", the system will re-calculate the macro-ability of the organization based on this simulated "future" graph state.

[0237] Step s30: Post-Perturbation Recalculation Process:

[0238] The system re-calculates all the macro variables defined earlier on the disturbed in-memory graph copy for the member groups of organization A (which now does not contain or has changed the state of person P) and organization B (which now contains or has changed the state of person P), respectively, to obtain a new set of values (New_Value).

[0239] Step s31, impact quantification and analysis:

[0240] The system compares the re-calculated New_Value with the Baseline stored before the deduction begins one by one, and calculates the absolute change and the relative change rate to accurately quantify the change of each macro variable.

[0241] Step s32, generation and output of structured impact analysis report:

[0242] The last step is to present all the calculated quantitative results in a clear, intuitive and easy-to-understand way for decision-makers.

[0243] Output data structure step:

[0244] The system encapsulates all the calculated changes into a structured JSON object. The object is clearly divided into two parts: "impact on outflow organization" and "impact on inflow organization", each containing the baseline value, new value, absolute change and relative change rate of all macro variables.

[0245] Visualization and interpretation output steps:

[0246] In the front-end application that interacts with the user, these data can be rendered into intuitive comparison charts. For example: use the comparison radar chart to show the comprehensive changes of the organization's various capability indicators (such as the total number of talents, network density, output, etc.) before and after the disturbance.

[0247] • Use the waterfall chart to clearly show how the "total amount of human capital" increases or decreases due to personnel flow.

[0248] • For the changes in key indicators (such as the network cohesion changing from 1 to 2), the system can automatically generate a natural language explanation: "Warning: the outflow of core personnel P may cause the internal cooperation network of A organization to split into 2 independent subgroups, there is a risk of team fragmentation."

[0249] Through the rigorous design and implementation of the above four steps, the system's disturbance deduction unit of the present application can transform the evaluation system from a static information display tool into a data-driven dynamic "sand table deduction" engine that can support major strategic decisions such as talent introduction, risk control, and team building, thereby producing a huge beneficial effect that is incomparable to the prior art.

[0250] For several core modules in the technical solutions described in the present application, those skilled in the art can conceive some technical alternatives. However, through in-depth technical comparison and analysis, it can be found that these alternative solutions, although they can achieve some functions to a certain extent on the surface, cannot reach the height of the selected technical solution of the present application in terms of final technical effect, system performance, intelligence level or information utilization rate. The technical defects of these alternatives also prove the necessity and creativity of the selected solution of the present application.

[0251] 1. Analysis of alternative solutions for the score synthesis method in module two (individual capability quantification unit)

[0252] Preferred solution of the present application: adopt the graph neural network (GNN, such as GAT) model. This solution takes the calculated multi-dimensional variables as the initial features of the nodes, automatically aggregates neighbor node information and updates its own representation through end-to-end learning on the professional cooperation network subgraph, and finally maps to the score.

[0253] Alternative: Traditional machine learning regression models such as gradient boosting decision trees (XGBoost, LightGBM) or support vector regression (SVR) can be used. This approach requires all calculated variables as a feature set (X) and relies on a small-scale dataset pre-labeled by domain experts as training labels (Y) to train a regressor for score prediction through supervised learning.

[0254] Comparative analysis and conclusion: Although the alternative solution can also achieve the purpose of synthesizing multiple dimensions into a single score, it has two major defects:

[0255] 1. Inherent deficiency of information utilization: The alternative solution treats each professional as an isolated data point, completely losing the valuable "relationship" information in the structured knowledge set. It cannot automatically learn and capture deep, non-local network structure effects like "who you work with matters more than how many times you work together" and "the value of being a bridge in the knowledge dissemination network" from the data, as GNN does. Therefore, the accuracy and depth of its evaluation results are far inferior to the GNN solution.

[0256] 2. Limitations of model capabilities: Traditional regression models essentially fit a fixed feature space, while GNN learns in a dynamic information propagation space determined by the graph structure. GNN can better handle sparse, high-dimensional, and complex dependent data, and its model expression and generalization capabilities are usually superior to traditional models under the same data volume.

[0257] Therefore, the GGNN solution selected by the present application can produce more accurate, profound, and objective evaluation results by end-to-end fusion learning of node attribute information and graph topology information, and its technical effects are difficult to achieve by traditional machine learning alternative solutions.

[0258] 2. Alternative solution analysis for module four (system disturbance deduction unit) implementation method

[0259] The present application: Adopt the method based on knowledge graph disturbance simulation. This solution simulates events by modifying the relationships (or their attributes, weights) between entities directly on the in-memory copy of the graph structure and recalculates the macroscopic graph indicators of the affected organizations to quantify the impact.

[0260] Alternative: A complex rule-based expert system can be constructed. This solution requires a large number of field experts to be invited in advance, and hundreds of "IF-THEN-ELSE" rules are defined manually through knowledge engineering. For example: "IF the title of the leaver = 'chief physician' AND S_abs>95 AND the department level = 'national key discipline' THEN the academic prestige of the organization = the academic prestige of the organization - 20".

[0261] Comparative analysis and conclusion: The rule-based alternative solution seems feasible in scenarios with simple logic and few influencing factors, but it has three fundamental flaws that make it impossible to truly achieve the purpose of the invention:

[0262] 1) Unable to capture systemic chain reactions: Expert rules are essentially linear, local, and based on experience. They cannot capture the non-linear, complex network structure chain reactions triggered by personnel changes. For example, the departure of a key "bridge" figure may cause the originally tight internal cooperation network to split into several "small groups", resulting in a sharp decline in the organization's "internal network density" and "network cohesion" indicators. This systemic and emergent change cannot be predicted by any limited and pre-set rules.

[0263] 2) Poor scalability and maintainability: The construction and maintenance of the rule base is extremely costly, requiring a large amount of expert time and effort. When new evaluation indicators or new situations need to be added, the entire rule base may need to be modified and verified on a large scale to avoid conflicts between rules, which is extremely difficult and fragile in practice.

[0264] 3) Lack of dynamic adaptability: Rule-based systems are static, with knowledge fixed in rules. The graph-based simulation method of the invention is data-driven and dynamically adaptive. When the knowledge graph itself is updated due to the addition of new data, the results of the simulation will automatically and correspondingly change without any human intervention.

[0265] In summary, the graph perturbation simulation method of the invention can truly and dynamically reflect the systemic changes triggered by events by directly operating at the data structure level of the system. Its depth, accuracy, and automation far exceed the rigid, rule-based alternative solution.

[0266] Summary: Although there are the above-mentioned alternative technical paths, they either have inherent deficiencies in information utilization or have obvious short boards in model complexity and dynamic adaptability. The technical solution combination set forth in the present application, by organically combining the leading technologies such as graph neural network and graph perturbation simulation, is the best technical choice to achieve the purpose of the present application, and the beneficial effects brought by it are difficult to achieve by other alternative schemes.

[0267] The application further discloses a professional influence evaluation and dynamic deduction system based on a knowledge graph, which comprises a memory, a processor and a computer program stored in the memory, and the computer program is configured to realize the steps of the method of the application when called by the processor.

[0268] The application further discloses a computer readable storage medium, which stores a computer program configured to realize the steps of the method of the application when called by a processor.

[0269] The above is a further detailed description of the application in combination with specific preferred embodiments, and the specific implementation of the application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the application belongs, a number of simple deductions or substitutions can be made without departing from the concept of the application, and all of them should be regarded as falling within the protection scope of the application.

Claims

1. A knowledge graph-based professional influence evaluation and dynamic deduction method, characterized in that, Comprise: Step one: build a structured knowledge set representing entities and relationships in the professional field; Step two: based on the structured knowledge set of step one, calculate the global influence absolute score of the target entity; Step three: based on the global influence absolute score of step two, conduct context-aware situational influence analysis; Step four: based on the perturbation simulation of the structured knowledge set of step one, conduct forward-looking influence deduction; The step two comprises: Step 1, multi-dimensional ability capital variable calculation: based on the structured knowledge set of step one, calculate the pre-defined multi-dimensional ability capital feature vector for professional entities through structured query or graph algorithm; Step 2, score synthesis: input the normalized multi-dimensional ability capital feature vector into the pre-trained comprehensive evaluation regression model for end-to-end nonlinear mapping, and finally output a scalar value, i.e. global influence absolute score; The step three comprises: Step y1, dynamic definition of evaluation range: receive one or more user input context parameters for defining the evaluation range; Step y2, evaluation group screening: convert the context parameters into a structured query of the structured knowledge set to obtain all professional entities that meet the conditions, forming a temporary evaluation group; Step y3, relative position calculation: obtain the global influence absolute score of all members in the evaluation group, calculate the ranking and percentile of the target professional entity in the group, and output its relative influence; In the step four, the deduction process comprises: Step s1, organization macro indicator definition and baseline calculation: predefine the macro variable system for measuring the comprehensive ability of the organization, and calculate all macro variable values of the organization in the current state before analysis as the baseline; Step s2, non-persistent perturbation simulation: make non-persistent and virtual modifications to the structured knowledge set in a temporary and isolated computing environment; Step s3, Post-disturbance Recalculation and Impact Quantification Analysis: Based on the structured knowledge set after the disturbance, recalculate all macro variables of the organization to get new values, and calculate the change rate Generate a quantified impact analysis report to show the predicted impact of the simulated event on the relevant parties in multiple dimensions.

2. The professional influence evaluation and dynamic deduction method according to claim 1, characterized in that, The step one comprises: Step S1, multi-source heterogeneous data collection: obtain structured data, semi-structured data and unstructured data about professional personnel and their related activities from multiple pre-defined online data sources; Step S2, knowledge extraction: use a unified extraction framework based on pre-trained language models and combined with specific prompt engineering, which identifies entities and extracts multi-dimensional relationship triples between them from long text through a three-stage processing pipeline of abstract-recognition-refinement; Step S3, knowledge fusion and verification: for entities from different data sources pointing to the same real-world object, use a multi-level conflict resolution framework based on signal priority for entity alignment and fact verification; Step S4, knowledge storage: load the cleaned and fused structured knowledge into a data storage system that can handle complex relationships between processing entities.

3. The professional influence evaluation and dynamic deduction method according to claim 2, characterized in that, In the step S1, the structured data, semi-structured data includes domestic and foreign academic literature database, patent database, clinical trial registration center, official agency directory; unstructured data includes official website of hospitals and research institutes, department homepage, science / health block of news portal website, industry vertical media, personal blog or public social media page of professionals; In the step S1, the structured data is obtained by periodically calling the data source providing API through the API client with preset keywords and update frequency; In the step S1, the web page data is collected by a distributed web crawler cluster; The crawler cluster has the following technical features: Dynamic IP proxy pool: integrate multiple commercial or self-built IP proxy services to achieve millisecond-level switching of request IP to respond to IP-based access frequency restrictions; Dynamic User-Agent rotation: maintain a list containing multiple mainstream browser and mobile device User-Agents, randomly select one for each request to simulate real user behavior; Adaptive request delay and session maintenance: dynamically adjust the request rate according to the target server's response status code and response time, and maintain the session during the simulated login process.

4. The professional influence evaluation and dynamic deduction method according to claim 2, characterized in that, In the step S2, the three-stage processing pipeline includes: Step a, coarse-grained summary and key information area positioning: use unsupervised extraction-based summary algorithm or lightweight generative summary model to compress long text into long summary containing core information, and record the position index of selected sentences in the original text for subsequent backtracking verification; Step b, semi-structured knowledge extraction based on pre-trained language model: embed the long summary text into a pre-designed unified Prompt template for entity and relationship extraction, which uses context learning and thought chain design patterns to guide the pre-trained language model to output all entities and relationships in JSON format that meet the predefined Schema; Step c, refine and context verification of the original text: for the knowledge points extracted by the pre-trained language model, use the position index recorded in step a to locate the context segment in the original text, and through a verification Prompt, require the pre-trained language model to combine the original text context to confirm, complete or modify the extraction results; In the step S2, the pre-defined pattern is a pre-defined domain knowledge graph pattern, which is used to specify the types, attributes, data types and constraint conditions of entities and relationships.

5. The professional influence evaluation and dynamic deduction method according to claim 2, characterized in that, In the step S3, the multi-level conflict resolution framework uses the internal topological consistency of the metadata graph and the context reasoning based on the pre-trained language model to ensure the uniqueness, accuracy and completeness of the final knowledge; The step S3 specifically includes: Step c1, entity alignment: for entities from different data sources that may refer to the same real-world object, use a hybrid alignment strategy, the specific steps are as follows: Step c10, rule-based fast matching: high-confidence matching through pre-set synonym dictionary and string similarity algorithm; Step c11, similarity calculation based on representation learning: for ambiguous cases that cannot be solved by rules, the cosine similarity of the attribute vector and the graph embedding vector of the entity to be matched is calculated respectively; Step c12, pairwise judgment based on pre-trained language model: input the descriptive information of the two entities into the pairwise judgment Prompt, and make the final judgment similar to that of human experts by the pre-trained language model; Step c2, signal priority multi-level decision framework for fact conflicts: when there are multiple conflicting facts about the same attribute of the same entity, a multi-level decision framework is started, and the specific steps are as follows: Level one, priority rule based on metadata: apply hard rules to preferentially adopt information with the latest timestamp, and if the timestamps are similar, preferentially adopt information from a data source with a higher preset authority score; Level two, topological verification based on internal consistency of the graph: if the rule cannot be decided, the existing high-confidence knowledge in the graph is verified to form a logical closed loop; Level three, context comprehensive reasoning decision based on pre-trained language model: for complex conflicts, all conflicting facts and their metadata, related context evidence are organized into structured text, which is input into the pre-trained language model of the fact checker role for comprehensive reasoning, and the final decision result and detailed reasoning process are required to be output.

6. The professional influence evaluation and dynamic deduction method according to claim 2, characterized in that, In the step S4, the multi-to-multi relationship between the logical data model entities is represented by defining a logical data model, and the physical implementation process includes: Step S40: attribute graph database, in which entities are mapped to nodes and relationships are mapped to edges with direction and attributes; Step S41: RDF triple storage; Step S42: multiple data tables associated by foreign keys in a relational database; Step S43: one or more documents stored in a document database in JSON or XML format, representing the relationship between entities through nesting or ID reference; Step S44: a collection of object instances connected to each other by pointers or references in computer memory.

7. The professional influence evaluation and dynamic deduction method according to claim 1, characterized in that, In the step 1, the pre-defined multi-dimensional capability capital feature vector at least includes variables of the following three dimensions: academic / knowledge capital variable, professional / skill capital variable, and social / network capital variable; In the step 2, it also includes: Step b1, constructing a feature vector: for a professional entity in the knowledge graph, a high-dimensional static ability feature vector of uniform dimension composed of variable calculation results is generated X i ; Step b2, data normalization: using a standardization method based on statistical quantities, the values of all variables are converted into a standard normal distribution with a mean of 0 and a standard deviation of 1; In the step 2, the training and deployment of the comprehensive evaluation regression model are also included, and the specific steps are as follows: Training data step: the training of the model is completed in the deployment or periodic update stage, not in each query; the training data comes from multiple experienced experts in the field, and the data set is manually annotated by taking the mean value after back-to-back scoring, which contains the feature vectors of multiple professionals and their corresponding expert consensus scores; Training process step: use mean square error or smooth L1 loss as the loss function, and use an adaptive learning rate optimizer to train and optimize the model parameters until the performance of the validation set converges; Running and outputting step: in actual operation, only the normalized feature vector of the person to be evaluated is sent into the trained model for forward propagation, and the final global influence absolute score can be calculated, with the score range normalized to 0-100.

8. The professional influence evaluation and dynamic deduction method according to claim 1, characterized in that, In the step y1, further comprising: Step y10: defining a standard evaluation hierarchy in a tree or network structure, which at least includes an organizational level, a geographical level, and a professional field level; Step y11: storing the evaluation hierarchy in a configuration file or database table accessed and modified by a system administrator, allowing specific levels to be added, modified, or disabled according to business needs without changing the core calculation code of the system.

9. The professional influence evaluation and dynamic deduction method according to claim 8, characterized in that, In the step y2, further comprising: Step y20, batch processing flow: using an offline periodic batch calculation task triggered by a timing task to execute, which traverses all entity nodes of type professional personnel in the knowledge graph; Step y21, query caching and optimization: using a key-value storage-based mechanism to cache query results, the specific steps are as follows: Step y210, unique level instance extraction: when the batch task starts, traverse all professional personnel entities to extract a list of all unique level instances they belong to; Step y211, batch query and cache of population data: only one structured query is performed for each unique hierarchical instance to obtain the IDs of all members under the hierarchical instance and their corresponding global absolute influence scores List, and the query result is serialized and stored in the cache with the unique identifier of the hierarchical instance as the key. Step y212, cache life cycle management: the life cycle of the cache matches the execution cycle of the batch task to ensure the timeliness of the data.

10. The professional influence evaluation and dynamic deduction method according to claim 9, characterized in that, In the step y3, further comprising: Step y30, batch calculation flow: traversing each professional personnel entity p in the knowledge graph, specifically including: Step y301: creating a JSON object for entity p to store its complete influence profile, denoted as InfluenceProfile; Step y302: traversing all predefined evaluation levels; Step y303: for each level, determining the specific instance that entity p belongs to; Step y304: Read the sorted list of scores for this instance directly from the cache. Step y304: Read the sorted list of scores for this instance directly from the cache. Step y305: From the score list in step y304, locate the position of entity p by binary search or other efficient algorithm, calculate its rank and percentile in this level, percentile is calculated by formula , make sure the percentile of the highest score is 100, store the structured result including range name, rank, total number of the group, percentile information as a sub-object in InfluenceProfile; Step y31, data storage and update: after calculation, the InfluenceProfile JSON object containing all level influence indicators is written directly as a new attribute to the node of the professional personnel entity p.

11. The professional influence evaluation and dynamic deduction method according to claim 1, characterized in that, In the step s1, the macro variable is obtained by aggregating query or graph algorithm calculation on the structured knowledge set, and the core variable at least includes talent capital index, output and prestige index, and network structure index, the talent capital index includes talent capital total, talent capital mean, talent capital mean, the output and prestige index includes high influence output total, and the network structure index includes internal collaboration network density, internal network cohesion, and external connection bridge number; In the step s1, the baseline calculation flow is as follows: Step s10: on the current unmodified and real structured knowledge set, respectively for organization A and organization B, performing graph query to filter out their respective member groups; Step s11: based on the two groups in the step s10, calculating the current values of all macro variables; Step s12: storing the calculated macro variable values of organization A and organization B as baseline in memory or temporary database.

12. The professional influence evaluation and dynamic deduction method according to claim 1, characterized in that, In the step s2, the modification at least includes one of the following ways: First, topology modification: temporarily increase, delete or modify the edges representing the relationship between entities; Second, attribute / weight modification: temporarily modify the numerical attributes or weight parameters associated with related entities or relationships to simulate the change of their state in calculation; Third, reachability / flow logic modification: at the algorithm level, by temporarily modifying the calculation logic of the graph algorithm to simulate the change of entity affiliation; In the step s3, it also includes: Step s30, post-disturbance recalculation process: on the disturbed memory graph copy, re-calculate all macro variables defined before for the member groups of organization A and organization B respectively, to get a new set of values; Step s31, impact quantification and analysis: compare the recalculated new values with the baseline stored before the deduction starts, and accurately quantify the change of each macro variable by calculating the absolute change and relative change rate; Step s32, generation and output of structured impact analysis report: present all the calculated quantification results in a way that decision makers can understand, including: Output data structure step: encapsulate all the calculated change into a structured JSON object, which is clearly divided into two parts: the impact on the outflow organization and the impact on the inflow organization, each part contains the baseline value, new value, absolute change and relative change rate of all macro variables; Visual presentation and explanatory output step: in the front-end application interacting with the user, the data is rendered into comparison charts.

13. A knowledge graph-based professional influence evaluation and dynamic deduction system, characterized in that, Comprise: A memory, a processor and a computer program stored on the memory, the computer program is configured to be called by the processor to realize the steps of the method in any one of claims 1-12.

14. A computer-readable storage medium, characterized in that: The computer readable storage medium stores a computer program, the computer program is configured to be called by the processor to realize the steps of the method in any one of claims 1-12.

Images