Enterprise associated person identity recognition and relation graph construction method and system

Abstract

The invention relates to the technical field of data processing, in particular to an enterprise associated person identity recognition and relation graph construction method and system. The method comprises the following steps: accessing entity data of a person to be identified and enterprise data associated with the entity data from a heterogeneous data source; through a multi-channel candidate recall engine, a candidate person and object entity set with the same name as the to-be-recognized person is retrieved from the relation graph; for each candidate person, calculating a comprehensive similarity score between the candidate person and the person to be identified by using a multi-dimensional evidence weighting and similarity calculation model; according to a comparison result of the comprehensive similarity score and a preset threshold value, judging whether the to-be-recognized person and the candidate person are the same entity or not; if it is judged that the to-be-recognized persons are the same entity, a graph database node merging operation is triggered, the relation associated with the to-be-recognized persons is migrated to existing candidate person nodes, redundant nodes are deleted, and a relation graph is updated; therefore, the problem of figure identity disambiguation can be automatically and intelligently solved with high precision.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for identifying and constructing relationship graphs of related individuals within an enterprise. Background Technology

[0002] In today's big data business environment, the relationships between businesses and individuals, such as investment and employment, form complex networks. The data supporting these networks typically originates from multiple heterogeneous systems, resulting in inconsistent data quality and a serious problem of "entity identity ambiguity." For example, it is difficult to accurately determine whether individuals with the same name appearing in different business scenarios belong to the same legal entity.

[0003] In existing technologies, solving this problem typically relies on manual verification or rule-based matching methods based on a single dimension (such as exact name matching). Manual verification is inefficient and cannot handle massive amounts of data. On the other hand, single-rule matching methods have an extremely high false positive rate when faced with situations where common Chinese names have a high rate of duplicates and business relationship networks are complex and ever-changing, making it difficult to meet the stringent requirements for data accuracy in applications such as risk control and business intelligence.

[0004] Therefore, there is an urgent need in this field for a technical solution that can automatically, intelligently and with high precision solve the problem of disambiguation of person identities and can dynamically maintain the consistency of the relationship graph. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying and constructing relationship graphs of related persons in an enterprise, which can solve the problem of disambiguation of person identities automatically, intelligently and with high accuracy, and can dynamically maintain the consistency of the relationship graph.

[0006] To achieve the above objectives, the present invention employs a method for enterprise-related personnel identification and relationship graph construction, comprising the following steps: Access the entity data of the person to be identified and its associated enterprise data from heterogeneous data sources; The multi-channel candidate recall engine retrieves a set of candidate entities with the same name as the person to be identified from the relationship graph. For each candidate, a multidimensional evidence weighting and similarity calculation model is used to calculate its comprehensive similarity score with the person to be identified. Based on the comparison between the comprehensive similarity score and the preset threshold, it is determined whether the person to be identified and the candidate are the same entity. If they are determined to be the same entity, a graph database node merging operation is triggered to migrate the relationship associated with the person to be identified to the existing candidate node and delete redundant nodes to update the relationship graph.

[0007] In the step of retrieving a set of candidate entities with the same name as the person to be identified from the relationship graph using a multi-channel candidate retrieval engine: The multi-channel candidate recall engine employs three channels for recall: a strong association information channel based on unique contact methods, an indirect association channel based on corporate equity networks, and a deep association channel based on core collaboration network fingerprints.

[0008] In the step of calculating the comprehensive similarity score between each candidate and the person to be identified using a multidimensional evidence weighting and similarity calculation model: The calculation process integrates at least one of the following: dynamic name weight model, graph topology confidence suppression mechanism, and organization name matching engine.

[0009] The multi-channel candidate recall engine employs a deep association channel recall process based on core collaborative network fingerprints, which includes the following steps: Using the person to be identified as the center, identify all enterprises within their first-degree relationship; Identify other individuals who work or collaborate with the person to be identified across all enterprises, forming a collaborative network; Based on the frequency of co-occurrence between each person in the collaboration circle and the person to be identified, a list of core collaboration partners is determined, which serves as the collaboration network fingerprint of the person to be identified. Search for other people with the same name in the relationship graph and determine whether their collaborators overlap with the fingerprint of the collaboration network. If the overlap meets the condition, then include the person with the same name in the candidate entity set.

[0010] The application of the dynamic name weight model includes: The mapping relationship between names and weights is pre-stored, where the weight of a name is negatively correlated with its frequency of occurrence in the benchmark dataset; When the name of the person to be identified matches that of the candidate, the mapping table is queried to obtain the weight of the corresponding name, and this weight is used as a multiplier in the similarity calculation.

[0011] The applications of the graph topology confidence suppression mechanism include: For the associated node that serves as the source of matching evidence, obtain the degree of that node in the relationship graph; Based on the preset inhibition function, the inhibition coefficient is calculated based on the degree, where the inhibition coefficient decreases as the degree increases; The evidence score generated by the associated node is adjusted using the inhibition coefficient.

[0012] The execution process of the organization name matching engine includes the following steps: The input company name string is destructured and features at least one dimension from core business name, geographical location, industry characteristics and organizational form are extracted. Based on the TF-IDF algorithm, the features are transformed into weighted feature vectors; Calculate the cosine similarity between the weighted feature vectors of the two company names, and use it as the name matching score.

[0013] In the step of calculating the comprehensive similarity score between each candidate and the person to be identified using a multidimensional evidence weighting and similarity calculation model: The multidimensional evidence weighting and similarity calculation model directly assigns a base score higher than the highest value among preset thresholds to candidates recalled through strongly associated information channels based on unique contact methods. For candidates recalled through indirect association channels based on corporate equity networks and deep association channels based on core collaboration network fingerprints, the score calculation integrates multiple factors such as path depth, number of common partners, company association, dynamic name weight, and graph topology confidence suppression coefficient.

[0014] This invention also provides a system for identifying and constructing relationship graphs of related individuals within an enterprise, comprising a data access layer, a core computing layer, and a data persistence layer; wherein: The data access layer is used to access the entity data of the person to be identified and its associated enterprise data from heterogeneous data sources; The core computing layer is used to retrieve a set of candidate entities with the same name as the person to be identified from the relationship graph through a multi-channel candidate recall engine. For each candidate, a multi-dimensional evidence weighting and similarity calculation model is used to calculate its comprehensive similarity score with the person to be identified. Based on the comparison result of the comprehensive similarity score and a preset threshold, it is determined whether the person to be identified and the candidate are the same entity. If they are determined to be the same entity, a graph database node merging operation is triggered to migrate the relationship associated with the person to be identified to the existing candidate node and delete redundant nodes to update the relationship graph. The data persistence layer is used to store the relationship graph.

[0015] This invention discloses a method and system for enterprise-related person identification and relationship graph construction. The method comprises a data access layer, a core computing layer, and a data persistence layer, and includes the following steps: Accessing entity data of the person to be identified and its associated enterprise data from a heterogeneous data source; retrieving a set of candidate entities with the same name as the person to be identified from the relationship graph using a multi-channel candidate recall engine; calculating a comprehensive similarity score between each candidate entity and the person to be identified using a multi-dimensional evidence weighting and similarity calculation model; determining whether the person to be identified and the candidate entity are the same entity based on the comparison result of the comprehensive similarity score and a preset threshold; if they are determined to be the same entity, triggering a graph database node merging operation to migrate the relationships associated with the person to be identified to existing candidate entity nodes and deleting redundant nodes to update the relationship graph; through this method, the problem of disambiguation of person identification can be solved automatically, intelligently, and with high precision, and the consistency of the relationship graph can be dynamically maintained. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the enterprise-related personnel identification and relationship graph construction method of the present invention.

[0018] Figure 2 This is a flowchart illustrating the steps of the multi-channel candidate recall engine of the present invention, which employs deep association channel recall based on core collaborative network fingerprints.

[0019] Figure 3 This is a flowchart illustrating the execution process of the organization name matching engine of the present invention.

[0020] Figure 4 This is a structural principle diagram of the enterprise-related person identification and relationship graph construction system of the present invention.

[0021] Figure 5 This is a schematic diagram of the electronic device of the present invention.

[0022] 501 - Data Access Layer, 502 - Core Computing Layer, 503 - Data Persistence Layer. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0024] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0025] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0026] Please see Figures 1-3 This invention provides a method for identifying and constructing relationship graphs of related individuals within an enterprise, comprising the following steps: S100: Access the entity data of the person to be identified and its associated enterprise data from heterogeneous data sources; In this embodiment, the entity data of the person to be identified and the related enterprise data are accessed from multiple types of heterogeneous data sources, and the data is cleaned and standardized to provide a high-quality data foundation for subsequent calculations.

[0027] 1. Data source type: It covers commonly used relational databases (such as Doris and MySQL) and message queues (such as Kafka) in business systems, respectively corresponding to batch data storage and real-time incremental data transmission scenarios.

[0028] 2. Access method: Real-time access: Subscribe to specified topics through the Kafka consumer module to receive data change messages for people / enterprises pushed by upstream business systems.

[0029] Batch access: During system initialization or periodic full calibration, tens of thousands to hundreds of thousands of historical person-enterprise related data are retrieved from the relational database.

[0030] 3. Data preprocessing: Perform format validation and cleaning on the accessed data, remove invalid information (such as incorrectly formatted phone numbers and generic customer service emails), and standardize field formats (such as standardized company names and normalized contact information).

[0031] S200: Through a multi-channel candidate recall engine, retrieve a set of candidate entities with the same name as the person to be identified from the relationship graph; In this embodiment, the multi-channel candidate recall engine uses three channels for recall: a strong association information channel based on unique contact information, an indirect association channel based on corporate equity network, and a deep association channel based on core collaboration network fingerprint.

[0032] The multi-channel candidate recall engine employs a deep association channel recall process based on core collaborative network fingerprints, including the following steps: S201: Identify all companies within a first-degree relationship of the person to be identified; S202: Identify other individuals who work or collaborate with the person to be identified across all enterprises, forming a collaborative network; S203: Determine a list of core collaborating partners based on the frequency of co-occurrence between each person in the collaborating circle and the person to be identified, so as to serve as the collaborating network fingerprint of the person to be identified; S204: Search for other people with the same name in the relationship graph and determine whether their collaborators overlap with the fingerprint of the collaboration network. If the overlap meets the condition, then include the person with the same name in the candidate entity set.

[0033] In the above process, based on the information of the person to be identified, a multi-channel parallel recall mechanism is initiated to accurately locate the set of candidate entities with the same name from the Neo4j relationship graph, balancing recall rate and retrieval efficiency. The implementation logic of the three core channels is as follows: Channel 1: Strongly Associated Information Channel Based on Unique Contact Information Using highly confident unique identifiers as the core, we can quickly filter candidates with a high probability of matching.

[0034] 1. Extract unique contact information such as phone numbers and email addresses from the data of individuals to be identified, and perform validity checks (such as removing non-standard mobile phone numbers with fewer than 11 digits).

[0035] 2. Use a graph database to query all business entities that use this contact method.

[0036] 3. The filtering criteria require that a certain contact method be associated with only ≤N companies (N is a configurable threshold, default is 5). This is to avoid situations where the contact method reported by a company is an accounting firm.

[0037] Channel 2: Indirect Linkage Channel Based on Corporate Equity Network By exploring the connections between companies, we can uncover indirectly related candidates with the same name. The core logic is that "the closer the connection between companies, the higher the probability that the individuals share the same origin."

[0038] 1. Starting from company A, where the person to be identified is located, perform a breadth-first or shortest path search in the relationship graph to retrieve upstream and downstream companies connected by relationships such as "investment" and "shareholder". The search depth is limited to 5 layers (to balance computational overhead and recall).

[0039] 2. If there is a person with the same name in company B, the endpoint of the path, include that person in the candidate set and record the path length (depth) between companies A and B; the shorter the path, the stronger the evidence.

[0040] 3. Optimize with dynamic name weight model: For high-frequency names such as "Zhang Wei" and "Li Ming", automatically shrink the search depth (e.g., reduce from 5 layers to 2 layers) to avoid "candidate explosion"; for rare names, maintain the maximum search depth to ensure recall.

[0041] Channel 3: Deep Association Channel Based on Core Collaborative Network Fingerprint The innovative approach recalls candidates by leveraging the "collaboration circle characteristics" of individuals, with the core assumption that "people with the same name and a high degree of overlap in core partners are likely to be the same entity."

[0042] 1. Construct the core collaboration circle of the person to be identified, A: First, identify all the companies where A works within the first degree of relationship, and then extract the people who work / cooperate with A in these companies to form a preliminary collaboration circle.

[0043] 2. Screening key partners: Count the number of times that people in the collaboration circle have served in the same position as A, sort them and select the top 50 (configurable) as A's "core collaboration partners (Top Partners)", which constitute A's unique "collaboration network fingerprint".

[0044] 3. Fingerprint matching and candidate screening: Search for all persons with the same name (B, C, D...) in the graph and analyze their core collaborators; if the overlap between a person B's core collaborators and A reaches a preset ratio (e.g., 30%), B will be included in the candidate set.

[0045] S300: For each candidate, use a multidimensional evidence weighting and similarity calculation model to calculate its comprehensive similarity score with the person to be identified. In this embodiment, the calculation process integrates at least one of the following: a dynamic name weight model, a graph topology confidence suppression mechanism, and an organization name matching engine.

[0046] The application of the dynamic name weight model includes: The mapping relationship between names and weights is pre-stored, where the weight of a name is negatively correlated with its frequency of occurrence in the benchmark dataset; When the name of the person to be identified matches that of the candidate, the mapping table is queried to obtain the weight of the corresponding name, and this weight is used as a multiplier in the similarity calculation.

[0047] The applications of the graph topology confidence suppression mechanism include: For the associated node that serves as the source of matching evidence, obtain the degree of that node in the relationship graph; Based on the preset inhibition function, the inhibition coefficient is calculated based on the degree, where the inhibition coefficient decreases as the degree increases; The evidence score generated by the associated node is adjusted using the inhibition coefficient.

[0048] The execution process of the organization name matching engine includes the following steps: S301: Deconstruct the input company name string and extract features from at least one dimension of core business name, geographical location, industry characteristics, and organizational form; S302: Based on the TF-IDF algorithm, features are transformed into weighted feature vectors; S303: Calculate the cosine similarity between the weighted feature vectors of two company names as the name matching score.

[0049] The multidimensional evidence weighting and similarity calculation model directly assigns a base score higher than the highest value among preset thresholds to candidates recalled through strongly associated information channels based on unique contact methods. For candidates recalled through indirect association channels based on corporate equity networks and deep association channels based on core collaboration network fingerprints, the score calculation integrates multiple factors such as path depth, number of common partners, company association, dynamic name weight, and graph topology confidence suppression coefficient.

[0050] In the above process, for each candidate, a comprehensive similarity score is calculated between the person to be identified and the candidate. This includes: (1) Dynamic Name Weighting Model The evidentiary value of a name is quantified by its frequency of occurrence, with rare names given higher weight, to avoid "misjudging high-frequency names and omitting low-frequency names".

[0051] Data source: Names collected from databases of executives, shareholders, legal representatives, etc., depending on the system.

[0052] Calculation logic: Count the number of times the name P_name appears in the dataset (Count(P_name)) and record the number of times the most frequent name appears (N_max).

[0053] Weight calculation formula: W_name = log( N_max / (Count(P_name) + 1) ) (+1 is used for smoothing to avoid the denominator being 0).

[0054] Normalization: The original weights are mapped to the range of 0.1-1.0 and stored in the MySQL name weight database for real-time querying during calculation.

[0055] (2) Core collaboration circle size suppression weight Based on the fact that the larger the population size, the greater the probability of encountering individuals with the same name (but not the same person), a suppression weight based on population size is calculated to avoid false matching. The weight calculation logic is as follows: 1. Node Degree Calculation: Obtain the degree d (number of direct relationships) of related nodes (such as jointly employed companies or partners) in real time through Neo4j.

[0056] 2. Suppression function: f(d) = 1 / [1 + (d / 100)^1.2], with the following characteristics: when d approaches 0, f(d)≈1.0 (the evidence of small nodes is credible); when d increases, f(d) continuously decreases (the credibility of the evidence of large nodes decreases).

[0057] 3. Score adjustment: S_adjusted = S_evidence * f(d). For example, a co-partner discovered through a company with 800 employees has a basic evidence score of 10 points, and the adjusted score is 10 * 0.7 = 7 points.

[0058] Suppression weights are used to calculate two types of data: the size weight of the target person's partners (partners_weight) and the company size weight of the candidate to be matched (company_size_weight).

[0059] (3) Organization name matching engine Achieve high-precision matching of company names in a non-LLM framework to provide contextual evidence for person-to-person association.

[0060] 1. Structural deconstruction and normalization Word segmentation and part-of-speech tagging: The jieba word segmentation library and a custom dictionary for the financial enterprise field are used to segment company names (e.g., “Xinyang / ns Wanguo / nz Real Estate / n Co., Ltd. / n”).

[0061] Feature extraction: Classified by part of speech into four dimensions: core business name (K, such as "Wanguo"), geographical location (T, such as "Xinyang"), industry characteristics (I, such as "real estate"), and organizational form (S, such as "Limited Company").

[0062] Normalization: Remove commonly used stop words such as "development", "limited", and "company".

[0063] 2. Feature vectorization and TF-IDF weighting Construct a global dictionary: Based on a directory of tens of millions of enterprises, build a dictionary containing all K / T / I features and define vector dimensions.

[0064] TF-IDF calculation: W(t, d) = TF(t, d) * IDF(t), where TF(t, d) = 1 (the word in the company name usually appears only once), and IDF(t) = log( N / (DF(t) + 1) ) (N is the total number of companies, and DF(t) is the number of companies containing the word t).

[0065] 3. Company name similarity calculation: For any two company names to be compared, the similarity calculation steps are as follows: 3.1 Feature subset extraction: Based on the word segmentation and part-of-speech tagging results described in step 1, feature word sets corresponding to the three types of features, namely core business name (K), geographical location (T), and industry features (I), are extracted from the two company names respectively, and are denoted as K1, T1, I1 (first company) and K2, T2, I2 (second company).

[0066] 3.2 Feature Class Intersection Over Union (IoU) Calculation: For each feature class, calculate the Intersection Over Union (IoU) of the feature word sets corresponding to the two companies, as shown in the following formula: The intersection-union ratio of core merchants is: IoU(K) = |K1 ∩ K2| / |K1 ∪ K2|, where |·| represents the number of elements in the set.

[0067] Intersection over union ratio (IoU(T)) = |T1 ∩ T2| / |T1 ∪ T2|.

[0068] Industry-specific intersection-union ratio: IoU(I) = |I1 ∩ I2| / |I1 ∪ I2|.

[0069] 3.3 Weighting coefficient assignment: For each feature word in the intersection (K1 ∩ K2, T1 ∩ T2, I1 ∩ I2) of the two types of enterprises, the TF-IDF value calculated in the steps is used as the weighting coefficient of the feature word; if there are no feature words in the intersection, the weighting coefficient of the corresponding category is 0.

[0070] 3.4 Weighted Cross-Union Ratio (CUIR) Calculation: Multiply the CUIR of each feature class by the average TF-IDF value of the intersection of features of that class to obtain the weighted CUIR of that feature class. Weighted intersection-union ratio of core firms: S(K) = IoU(K) × (Σw k / |K1 ∩ K2|), where w k Let S(K) be the TF-IDF value of the feature words in K1 ∩ K2. When |K1 ∩ K2|=0, S(K)=0.

[0071] Geographically weighted intersection-union ratio: S(T) = IoU(T) × (Σw t / |T1 ∩ T2|), where w t Let S(T) be the TF-IDF value of the feature words in T1 ∩ T2. When |T1 ∩ T2|=0, S(T)=0.

[0072] Industry-feature weighted intersection-union ratio: S(I) = IoU(I) × (Σwᵢ / |I1 ∩ I2|), where wᵢ is the TF-IDF value of the feature words in I1 ∩ I2. When |I1 ∩ I2|=0, S(I)=0.

[0073] 3.5 Comprehensive Similarity Calculation: The weighted intersection-union ratios of the three categories—core business name, geographical location, and industry characteristics—are summed to obtain the final similarity score between the two company names. S = S(K) + S(T) + S(I); The similarity score S ranges from [0, +∞), and a larger S value indicates a higher similarity between the two company names.

[0074] Overall score calculation logic: Based on the strength of evidence gathered during the candidate recall process, a differentiated scoring strategy is employed, and the final score is calculated as total_score: Score for strongly related information channels based on unique contact methods: If the information is retrieved via phone / email, it is a strongly related information and is directly assigned 3 points (meeting the matching conditions), which can be considered as related.

[0075] Indirect association channel score based on corporate equity network: Score = BaseScore(depth) * W_name, where BaseScore(depth) decreases with path depth (e.g., depth=1 gets 2.0 points, depth=3 gets 1.0 points), and rare names can further increase the score.

[0076] Deep association channel score based on core collaborative network fingerprint: Final_Score = SUM((matched_name_score + relation_score) * partners_weight * company_size_weight) matched_name_score: Determined by the number of common core partners and the weight of partner names (the rarer the partner and the more overlap, the higher the score).

[0077] relation_score: If there is a short path relationship between the person's company and the company, extra points will be awarded based on the path length.

[0078] partners_weight / company_size_weight: Coefficients based on graph topology suppression mechanism (the larger the partner circle / company size, the lower the coefficient).

[0079] Edge case refinement: If the score is in the "fuzzy range" of 0.7-1.5, call the enterprise name matching engine and add the enterprise name similarity, geographical location matching and industry matching to the total score with a weight of 0.1-0.3.

[0080] Heuristic rule penalty: If the company name contains "cooperative", "commercial bank", "investment company" etc. (due to regulatory requirements, financial companies have a higher minimum number of members), the score is multiplied by a penalty coefficient of 0.5 to reduce the false positive rate.

[0081] S400: Based on the comparison result of the comprehensive similarity score and the preset threshold, determine whether the person to be identified and the candidate are the same entity; if they are determined to be the same entity, trigger the graph database node merging operation, migrate the relationship associated with the person to be identified to the existing candidate node, and delete redundant nodes to update the relationship graph.

[0082] In the above process, the determination of entity homology is as follows: Based on the comparison between total_score and preset threshold, a clear judgment result is output, balancing automation and business flexibility: Score > 3.0: Strong evidence (such as matching contact information) triggers the determination of "same entity" without manual intervention.

[0083] 1.6 < Score ≤ 3.0: Moderately strong evidence combination (such as short-path corporate association + rare name) is judged as "highly suspected, confirmed as the same entity".

[0084] 0.9 < Score ≤ 1.6: The strength of evidence is moderate, supporting two decision-making modes: in business-sensitive scenarios, it is judged as "suspected and requires manual review"; in non-sensitive scenarios, it can be automatically judged as "not the same entity".

[0085] Score ≤ 0.9: Insufficient evidence, determined to be "not the same entity", the person to be identified is added to the graph as a new node.

[0086] Atomic update of relation graph: If the nodes are determined to be the same entity, the graph database node merging module is triggered to perform atomic operations in Neo4j, ensuring that the graph is free of redundancy and the data is consistent.

[0087] 1. Technology dependency: Based on the Neo4j APOC extension library, calling the graph reconstruction process.

[0088] 2. Execution process: Relationship migration: seamlessly migrate all relationships (such as "employment" and "investment") of the person node to be identified (new_id) to the existing candidate node (old_id).

[0089] Redundancy removal: Delete the new_id node to clear redundant data in the graph.

[0090] 3. Atomicity guarantee: All operations are encapsulated in update transactions, which either all succeed or roll back the whole if a step fails, avoiding intermediate error states such as "relationship migration is halfway done but node has not been deleted".

[0091] Corresponding to the aforementioned embodiments of the enterprise-related person identification and relationship graph construction method, this application also provides embodiments of the enterprise-related person identification and relationship graph construction system.

[0092] Figure 4 This is a block diagram illustrating a system for identifying and constructing relationship graphs of related individuals within an enterprise, based on an exemplary embodiment. (Refer to...) Figure 4 The system may include a data access layer 501, a core computing layer 502, and a data persistence layer 503; wherein: The data access layer 501 is used to access the entity data of the person to be identified and its associated enterprise data from heterogeneous data sources; The core computing layer 502 is used to retrieve a set of candidate entities with the same name as the person to be identified from the relationship graph through a multi-channel candidate recall engine; for each candidate entity, a multi-dimensional evidence weighting and similarity calculation model is used to calculate its comprehensive similarity score with the person to be identified; based on the comparison result of the comprehensive similarity score and a preset threshold, it is determined whether the person to be identified and the candidate entity are the same entity; if they are determined to be the same entity, a graph database node merging operation is triggered to migrate the relationship associated with the person to be identified to the existing candidate entity nodes and delete redundant nodes to update the relationship graph; The data persistence layer 503 is used to store the relationship graph.

[0093] In this embodiment, the data access layer 501 is configured to access personnel and enterprise data from heterogeneous data sources; the data access layer 501 includes a Kafka consumer module for receiving incremental data change messages pushed by upstream business systems in real time; and / or a MySQL batch read module for batch pulling data during system initialization or calibration. The core computing layer 502, coupled to the data access layer 501, includes: Multi-channel candidate recall engine, configured to execute multi-channel candidate recall steps; The similarity calculation and decision engine integrates a dynamic name weight model, a graph topology confidence suppression mechanism, an organization name matching engine, and a multi-dimensional evidence weighting and similarity calculation model, configured to perform similarity calculation and decision steps; The graph operation module is used to encapsulate access to and manipulation of the relationship graph; The data persistence layer 503, coupled to the core computing layer 502, includes a graph database and a relational database.

[0094] The above methods can automatically, intelligently, and with high precision solve the problem of disambiguation of person identities, and can dynamically maintain the consistency of the relationship graph.

[0095] The overall system architecture of this invention includes a data source layer, a data access layer 501, a core computing layer 502, and a data persistence layer 503. The data source layer may include a Doris database, a MySQL database, and a Kafka message queue. The data access layer 501 acquires data through a Kafka consumer module and a MySQL batch read module. The core computing layer 502 is the core of the system, containing an NLP and feature engineering module, a similarity calculation and decision engine, a graph operation module (DAO), and a runner. The data persistence layer 503 uses the Neo4j graph database to store the relational graph and uses MySQL to store configuration and intermediate data. The runner connects and coordinates the workflows of the data access layer 501, the core computing layer 502, and the data persistence layer 503.

[0096] The data processing flow is divided into real-time incremental streams and batch processing streams. The real-time stream begins when the upstream business system pushes data changes to Kafka. After consuming the message, the system sequentially performs candidate retrieval, feature extraction and comparison, similarity calculation, decision-making, and graph merging. The batch stream is triggered by the scheduler, reading data in batches from MySQL and processing it in parallel.

[0097] Regarding the specific implementation of the core algorithm: 1. Multi-channel candidate recall engine: Strongly correlated information channels based on unique contact information: Extract and clean unique contact information such as phone numbers and email addresses, and use graph database reverse indexes to query companies and people with the same name who share the contact information.

[0098] Indirect association channels based on corporate equity networks: Starting from the company where the person to be matched is located, a graph path search (such as BFS) with a limited depth is performed to locate the person with the same name under the company at the end of the path as a candidate. The path depth can be adaptively adjusted by combining dynamic name weights.

[0099] A deep association channel based on core collaborative network fingerprints: Centered on the person to be matched, a set of companies within their first-degree relationship is constructed, and other persons appearing in these companies are counted. The top N most frequent individuals are selected as their "core collaborative partners" list (i.e., network fingerprints). A search is conducted across the entire network for persons with the same name, and their network fingerprints are compared. Those with high overlap are listed as candidates.

[0100] 2. Organization Name Matching Engine: First, the input company name string is segmented and tagged with parts of speech (e.g., using `jieba` segmentation).

[0101] Secondly, based on a predefined rule base, words are categorized into four structured dimensions: core business name (K), geographical location (T), industry characteristics (I), and organizational form (S).

[0102] Then, a feature space is constructed based on a large-scale corpus of company names, and the TF-IDF algorithm is used to transform the structured features into weighted feature vectors.

[0103] Finally, the cosine similarity between the feature vectors of the two company names is calculated as the name matching score.

[0104] 3. Dynamic Name Weighting Model: Based on a large-scale benchmark population name dataset, count the frequency of occurrence of each name `P_name` using `Count(P_name)`.

[0105] The original weights are calculated using the formula `W_name = log( N_max / (Count(P_name) + 1) )`, and then linearly normalized to obtain the final weight values ​​(e.g., mapped to the 0.1-1.0 range), which are then persistently stored.

[0106] 4. Graph topological confidence suppression mechanism: For related nodes that serve as sources of evidence (such as jointly employed companies or common partners), obtain their degree `d` in the graph.

[0107] The inhibition coefficient is calculated using the inhibition function `f(d) = 1 / [1 + (d / 100)^1.2]`.

[0108] The inhibition coefficient is used as a multiplier and applied to the original score generated from this evidence to obtain the adjusted evidence score.

[0109] 5. Multidimensional evidence weighting and similarity calculation model: Evidence from various recall channels is integrated, and different base scores are assigned to evidence of different strengths (e.g., strong connection matching is directly assigned a high score).

[0110] The evidence score is adjusted by comprehensively applying the dynamic name weight `W_name` as the multiplier and the graph topological suppression coefficient `f(d)`.

[0111] For candidates whose scores fall within the ambiguous range, a refinement process is initiated, incorporating the matching degree of company name, geographical location, and industry category for fine-tuning.

[0112] The decision is made based on the final total score and a preset threshold (e.g., >3.0 directly indicates the same person; 0.9~1.6 is suspected and requires review; <=0.9 indicates a different person).

[0113] 6. Graph Database Node Merging Module: When two person nodes are determined to be the same entity, an atomic transaction operation is performed using Neo4j's APOC extension library. The core Cypher statement logic of this operation is: migrate all relations of the node to be merged (`new_id`) to the target node (`old_id`), and then delete the node to be merged. This process ensures data consistency.

[0114] This invention aims to solve the following problems existing in the prior art: 1. In complex business networks, the accuracy and recall rates of identifying individuals with the same name are low; 2. There is a lack of effective quantification and suppression mechanisms for factors such as the prevalence of names and network topology noise; 3. The identity recognition results are disconnected from the updates of the backend relationship graph, making it impossible to achieve automated and atomic graph data maintenance; 4. When faced with massive incremental data, the processing efficiency is low, and the timeliness of the relationship graph cannot be guaranteed; Compared with the prior art, the present invention has the following significant advantages: 1. High-precision identity recognition: Through multi-channel recall and a weighted model that integrates multiple innovative algorithms, it can achieve an accuracy rate of over 95% in hundreds of millions of nodes, which is significantly better than traditional rule-based methods.

[0115] 2. Deep Relationship Mining Capability: The innovative "Core Collaborative Network Fingerprint" channel can go beyond direct association and find matching evidence in a person's second-degree social network, greatly improving the recognition capability under sparse relationships.

[0116] 3. Strong noise resistance: Through dynamic name weight and graph topology confidence suppression mechanism, data noise caused by name prevalence and network hub nodes is scientifically quantified and suppressed, reducing the false judgment rate.

[0117] 4. Automation and efficiency: It realizes full-process automation from identity recognition to graph merging, supports near real-time incremental processing and efficient batch processing, and ensures the timeliness and consistency of the relationship graph.

[0118] 5. Flexible and scalable architecture: The system adopts a modular design, which is easy to expand and maintain, and can adapt to application scenarios of different scales and data sources.

[0119] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0120] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0121] Accordingly, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the enterprise-related person identification and relationship graph construction method described above. Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, used in an embodiment of the present invention to provide a system for identifying and constructing a relationship graph of related individuals within an enterprise. (Except for...) Figure 5 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0122] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the enterprise-related person identification and relationship graph construction method described above. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0123] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0124] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for identifying and constructing relationship graphs of related individuals within an enterprise, characterized in that, Includes the following steps: Access the entity data of the person to be identified and its associated enterprise data from heterogeneous data sources; The multi-channel candidate recall engine retrieves a set of candidate entities with the same name as the person to be identified from the relationship graph. For each candidate, a multidimensional evidence weighting and similarity calculation model is used to calculate its comprehensive similarity score with the person to be identified. Based on the comparison between the comprehensive similarity score and the preset threshold, it is determined whether the person to be identified and the candidate are the same entity. If they are determined to be the same entity, a graph database node merging operation is triggered to migrate the relationship associated with the person to be identified to the existing candidate node and delete redundant nodes to update the relationship graph.

2. The method for enterprise-related person identification and relationship graph construction as described in claim 1, characterized in that, In the step of retrieving a set of candidate entities with the same name as the person to be identified from the relationship graph using a multi-channel candidate retrieval engine: The multi-channel candidate recall engine employs three channels for recall: a strong association information channel based on unique contact methods, an indirect association channel based on corporate equity networks, and a deep association channel based on core collaboration network fingerprints.

3. The method for enterprise-related person identification and relationship graph construction as described in claim 2, characterized in that, In the step of calculating the comprehensive similarity score between each candidate and the person to be identified using a multidimensional evidence weighting and similarity calculation model: The calculation process integrates at least one of the following: dynamic name weight model, graph topology confidence suppression mechanism, and organization name matching engine.

4. The method for enterprise-related person identification and relationship graph construction as described in claim 3, characterized in that, The multi-channel candidate recall engine employs a deep association channel recall process based on core collaborative network fingerprints, including the following steps: Using the person to be identified as the center, identify all enterprises within their first-degree relationship; Identify other individuals who work or collaborate with the person to be identified across all enterprises, forming a collaborative network; Based on the frequency of co-occurrence between each person in the collaboration circle and the person to be identified, a list of core collaboration partners is determined, which serves as the collaboration network fingerprint of the person to be identified. Search for other people with the same name in the relationship graph and determine whether their collaborators overlap with the fingerprint of the collaboration network. If the overlap meets the condition, then include the person with the same name in the candidate entity set.

5. The method for enterprise-related personnel identification and relationship graph construction as described in claim 4, characterized in that, The application of the dynamic name weight model includes: The mapping relationship between names and weights is pre-stored, where the weight of a name is negatively correlated with its frequency of occurrence in the benchmark dataset; When the name of the person to be identified matches that of the candidate, the mapping table is queried to obtain the weight of the corresponding name, and this weight is used as a multiplier in the similarity calculation.

6. The method for enterprise-related person identification and relationship graph construction as described in claim 5, characterized in that, The applications of the graph topology confidence suppression mechanism include: For the associated node that serves as the source of matching evidence, obtain the degree of that node in the relationship graph; Based on the preset inhibition function, the inhibition coefficient is calculated based on the degree, where the inhibition coefficient decreases as the degree increases; The evidence score generated by the associated node is adjusted using the inhibition coefficient.

7. The method for enterprise-related person identification and relationship graph construction as described in claim 6, characterized in that, The execution process of the organization name matching engine includes the following steps: The input company name string is destructured and features at least one dimension from core business name, geographical location, industry characteristics and organizational form are extracted. Based on the TF-IDF algorithm, the features are transformed into weighted feature vectors; Calculate the cosine similarity between the weighted feature vectors of the two company names, and use it as the name matching score.

8. The method for enterprise-related person identification and relationship graph construction as described in claim 7, characterized in that, In the step of calculating the comprehensive similarity score between each candidate and the person to be identified using a multidimensional evidence weighting and similarity calculation model: The multidimensional evidence weighting and similarity calculation model directly assigns a base score higher than the highest value among preset thresholds to candidates recalled through strongly associated information channels based on unique contact methods. For candidates recalled through indirect association channels based on corporate equity networks and deep association channels based on core collaboration network fingerprints, the score calculation integrates multiple factors such as path depth, number of common partners, company association, dynamic name weight, and graph topology confidence suppression coefficient.

9. A system for identifying and constructing a relationship graph of related individuals within an enterprise, employing the method for identifying and constructing a relationship graph of related individuals within an enterprise as described in claim 1, characterized in that, It includes a data access layer, a core computing layer, and a data persistence layer; among which: The data access layer is used to access the entity data of the person to be identified and its associated enterprise data from heterogeneous data sources; The core computing layer is used to retrieve a set of candidate entities with the same name as the person to be identified from the relationship graph through a multi-channel candidate recall engine. For each candidate, a multi-dimensional evidence weighting and similarity calculation model is used to calculate its comprehensive similarity score with the person to be identified. Based on the comparison result of the comprehensive similarity score and a preset threshold, it is determined whether the person to be identified and the candidate are the same entity. If they are determined to be the same entity, a graph database node merging operation is triggered to migrate the relationship associated with the person to be identified to the existing candidate node and delete redundant nodes to update the relationship graph. The data persistence layer is used to store the relationship graph.

Images