Personnel file hidden relationship mining and compliance risk early warning method based on graph flow linkage

Abstract

The present application provides a kind of personnel file implicit relationship mining and compliance risk early warning method based on graph flow linkage, it is related to knowledge graph and archive management and other technical fields.The method constructs the personnel social relationship knowledge graph covering blood relationship, academic origin, geographical origin, industry origin, and converts compliance requirements into structured executable rules;Through Louvain community discovery, double-end BFS and Pregel iterative reasoning algorithm, implicit relationship is mined, and chain attenuation confidence model is used to control false positive rate;Through broadcast state mechanism, graph calculation result is injected into flow calculation channel, and real-time compliance verification is realized.The present application solves the problems of low manual review coverage, difficult identification of implicit relationship and risk discovery lag.

Description

Technical Field

[0001] This invention relates to the technical fields of personnel file management, big data, information processing and knowledge graphs, specifically a method and system for mining implicit relationships in personnel files and providing early warning of compliance risks based on graph flow linkage. Background Technology

[0002] The core requirements for personnel file management are authenticity and compliance—the files must be strictly reviewed at key points such as job promotion, onboarding, and transfer; current regulations have listed informatization as a core function of daily file management, requiring the use of modern information technologies such as big data to establish a scientific management mechanism for personnel files.

[0003] However, current personnel file management faces the following technical bottlenecks: First, manual review suffers from a severe scalability bottleneck: a medium-sized province typically has tens of thousands to hundreds of thousands of personnel files, each containing more than 42 core fields, covering age, work experience, and identity (date of birth, date of employment, educational background, work history, and personal identification information). Under current technological conditions, organizational departments mainly rely on manual file review, with a sampling rate typically below 5%, making it difficult to guarantee comprehensive compliance supervision. Second, implicit relationships are difficult to identify manually: the avoidance system for personnel (including but not limited to civil servants, public institution staff, and state-owned enterprise managers, etc., individuals included in the personnel file management system) requires that spouses, direct blood relatives, collateral blood relatives within three generations, and close relatives by marriage generally cannot hold positions with direct superior-subordinate relationships or supervisory and restraining relationships in the same unit. In complex multi-layered relationship networks, manual methods can only focus on direct kinship relationships at most, and the review of other indirect relationships is almost infeasible. Traditional relational databases face the JOIN operation combinatorial explosion problem when handling such multi-hop relationship queries—in Search for all individuals within the entity. Relationship chains within a certain degree need to be Level JOIN operation, when Calculations become infeasible when the number reaches tens of thousands. Third, compliance risk discovery suffers from significant time lag: compliance checks such as recusal and job rotation currently rely primarily on pre-approval and post-audit, often only being discovered after promotion or transfer. Fourth, the data silo effect severely restricts the completeness of risk profiling: personnel information is scattered across multiple related departmental information systems, lacking unified data exchange standards and entity alignment mechanisms.

[0004] Therefore, there is an urgent need for a technical solution that can systematically mine the implicit relationships in personnel files and monitor and warn of compliance risks in real time. Summary of the Invention

[0005] The purpose of this invention is to provide a solution for mining implicit relationships in personnel files and providing early warning of compliance risks based on graph-stream linkage. It comprehensively utilizes data governance, graph databases, rule engines, graph computing algorithms (Louvain community discovery, two-end BFS shortest path analysis, Pregel iterative reasoning), and stream computing engines to achieve full-scale automated scanning of personnel file risks, intelligent mining of implicit relationships, and real-time dynamic monitoring of compliance risks. Specifically, this invention proposes a linkage mechanism between graph computing results and stream computing channels: high-confidence implicit relationship results generated by iterative reasoning are injected into the broadcast state through a message queue, while all relationship data is loaded into the state backend using employee ID as the key, enabling compliance verification triggered by appointment and dismissal events to be completed in milliseconds.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, this invention provides a method for mining implicit relationships in personnel files and providing early warning of compliance risks based on graph-flow linkage. The method includes the following steps: S1: Data Acquisition and Governance – This involves collecting multi-source heterogeneous data. Data sources may include, but are not limited to, personnel file data from personnel management information systems, personnel file management systems, compliance supervision information systems, population basic information systems, or academic qualification certification systems. After ETL data acquisition, cleaning, entity alignment, and standardization, a standardized archive data warehouse is formed.

[0008] S2: Constructing a multi-dimensional knowledge graph for personnel social relations—Constructing a personnel social relations knowledge graph based on a graph database, which includes basic information entities of target personnel, work experience entities, family member entities, and social relationship entities; the relation types of the knowledge graph cover four dimensions: blood relations, academic relations, geographical relations, and professional relations.

[0009] S3: Construct a dedicated structured rule base for the HR compliance domain—transforming HR management compliance requirements into a machine-executable structured rule base; including four major categories of executable rules: age, experience, and identity logic verification rules, appointment avoidance rules, tenure rotation rules, and compliance review rules; using the domain-specific language of the rule engine for formal rule definition, each rule consists of a condition part and an action part; the rule base supports hot deployment mechanism and Git-based version management.

[0010] S4: Execute graph computation algorithms for latent relationship mining—Perform three types of graph computation operations on the knowledge graph: (a) Community detection and dense subgraph detection—Execute the Louvain community detection algorithm, divide the graph into several communities through modularity optimization, calculate the edge density within the community, and mark a community as a densely connected group when the edge density of a community is significantly higher than the average level of the entire graph and its size exceeds a preset threshold; (b) Shortest path analysis—Perform bidirectional BFS search on person entity pairs to calculate the length of the relationship chain, and mark a suspected latent relationship when the number of hops in the shortest path does not exceed a preset threshold and the path contains a specific relationship edge combination pattern; (c) Iterative reasoning based on the Pregel model—Define the family attribute vector F = {father node set, mother node set, spouse node set, ... In each iteration, the nodes propagate family attribute vectors to their neighbors along the family member relationship edges. The receiver merges all neighbor messages and updates its own vector. Based on the propagated family attribute vectors, four inference rules (sibling inference rule, uncle / nephew inference rule, cousin inference rule, and brother-in-law / sister-in-law inference rule) are applied to discover collateral blood relatives within three generations that are not directly recorded and potential relationships. For each inferred relationship, a chain-decay confidence model is performed, and the confidence is classified into high, medium, and low confidence intervals. The high-confidence results are written back to the knowledge graph.

[0011] S5: Construct a real-time risk calculation channel linking graph computing and stream computing—Based on the stream computing engine, a risk calculation channel is constructed. The high-confidence implicit relationship results output from step S4 are written to the broadcast state of the stream computing engine through a message queue, so that all computing nodes hold the latest implicit relationship edge set in their local memory. At the same time, the full knowledge graph relationship data—including the original explicit relationship edges and the implicit relationship edges output from S4—is loaded into the state backend with the personnel identifier as the key and the relationship subgraph of that personnel as the value. When a personnel appointment or removal event triggers risk calculation, the rule condition matching is first completed in the local memory, and then the relationship subgraph is queried through the state backend to retrieve the associated personnel with appointment avoidance risks.

[0012] S6: Tiered early warning and push notification based on risk score - The warning is weighted by three dimensions: rule matching weight, relationship mining confidence and historical risk record. The warning information is divided into three levels: high risk, medium risk and low risk according to the preset threshold and distributed to the corresponding responsible entities and regulatory nodes through the message push mechanism.

[0013] Furthermore, entity alignment in S1 adopts a hybrid strategy based on rule matching and similarity calculation, using the ID number as the exact matching key and the combination key of name, date of birth and gender as the fuzzy matching key; records with the same relationship existing in two or more authoritative data sources are given high credibility weights, while relationship edges existing only in a single data source are given basic credibility weights.

[0014] Furthermore, the knowledge graph in S2 is constructed using a graph database as the storage engine, specifically including: Entity extraction—Extracting personnel entities, unit entities, family member entities, and social relationship entities from standardized archival data, and assigning a globally unique identifier to each entity; Relationship extraction—Establish direct relationship edges based on explicit association information in the archives, including blood relationship edges between the target person and their spouse and children, academic relationship edges between the target person and their graduating institution, professional relationship edges between the target person and their former employer, and geographical relationship edges between the target person and their place of origin. Graph fusion—resolves and merges the same entity from different data sources. Target personnel entities are precisely merged using their ID card numbers as anchors, while unit entities are merged based on their unified social credit codes and standard names.

[0015] Furthermore, the structured rule base in S3 adopts the DSL syntax definition of the rule engine, and the four types of executable rules are as follows: The first category of logical verification rules for age, work experience and identity is defined as follows: the logical consistency constraints between age, work experience and education level are defined, including the constraints that the reasonable age range between the date of birth and the time of starting work must be earlier than the specific time of employment, and the minimum matching constraints between the education level and the job level. The second category of avoidance rules defines that spouses, direct blood relatives, collateral blood relatives within three generations, and close relatives by marriage shall not hold positions in the same unit that are directly superior to or have a supervisory and restraining relationship. The third type of job rotation rule is defined by the maximum continuous service period threshold based on the risk level of the position. The maximum continuous service period for high-risk positions is 5 years, and for general positions it is 8 years. The fourth category of compliance review rules defines risk triggering conditions based on risk clue record counting thresholds, promotion restriction periods after disciplinary action records, and the rectification status of compliance verification issues.

[0016] Furthermore, the Louvain community detection algorithm in S4 employs a two-stage iteration: Stage 1 is modularity optimization—for each node, calculate the modularity increment after removing it from the current community and adding it to the community of a neighboring node, select the neighboring community with the largest increment to move, and repeat until a local optimum is reached; Stage 2 is graph compression—shrink each community obtained in Stage 1 into a super node, aggregate the edge weights between communities into the edge weights between super nodes in the new graph, and repeat Stage 1 for the new graph; the two stages are executed alternately until the modularity no longer increases; for each detected community, when the ratio of internal edge density to outward edge density exceeds a preset threshold and the community size exceeds a preset minimum size, the community is marked as a densely connected suspicious group.

[0017] Furthermore, the four inference rules of Pregel iterative inference in S4 are as follows: (1) Sibling inference - if two different entity nodes share at least one common parent node or mother node after propagation, and there is no direct blood relationship edge between the two nodes in the knowledge graph, then it is inferred that the two nodes are siblings; (2) Uncle / nephew inference - if the parent node of node A and node B have a confirmed sibling relationship, and the shortest path between A and B does not contain a direct blood relationship edge, then it is inferred that A and B are uncle / nephew; (3) Cousin / Cousin Sibling inference—If the father node of node A and the father node of node B have a confirmed sibling relationship, or if each mother node has a confirmed sibling relationship, then it is inferred that A and B are cousins / fathers, belonging to collateral blood relatives within three generations; (4) Brother-in-law / Sister-in-law inference—If the spouse node of node A and the spouse node of node B have a confirmed sibling relationship, then it is inferred that A and B are brothers-in-law / sisters-in-law; The above four inference rules can cover all collateral blood relative relationships within three generations through no more than three rounds of Pregel iterations, and the algorithm time complexity is ,in For iteration rounds and , The total number of nodes. Let be the total number of edges.

[0018] Furthermore, in S4, a chain-decay confidence model is performed for each inferred implicit relation r, and the confidence calculation formula is as follows: ,in, To infer the implicit relationship The overall confidence level, and its range. This indicates the degree of credibility of the present invention regarding the implicit relationship; The implicit relation triples obtained through Pregel iterative reasoning consist of two personnel nodes and the relationship type between them; To infer the relationship The set of ordered edges contained in the propagation path, that is, the ordered sequence formed along the edges of the knowledge graph from the starting point personnel to the ending point personnel; Number the edge sequentially within the path. = [1, 2, …, L], where L is the propagation path length (the number of edges contained in the propagation path); For the first in the propagation path The data quality score (source credibility) of each edge is calculated, with edges directly sourced from archival data receiving a score of 1.0 and edges generated through inference receiving a score of 0.7. For the first The relationship type weights (semantic reliability) of the edges are as follows: spouse relationship is 1.0, parent relationship is 1.0, child relationship is 0.95, and sibling relationship is 0.85. is the confidence decay factor per hop, and 0 < <1; =0.85 was determined through parameter sweep experiments — in In the six comparative experiments with values ​​of 0.70, 0.75, 0.80, 0.85, 0.90, and 0.95, =0.85 is optimal in terms of the harmonic mean of F1 for precision and recall; three confidence thresholds are set: the high confidence interval [0.8, 1.0] is directly used as the basis for warning, the medium confidence interval [0.5, 0.8) is marked as a suspected latent relationship and pushed to manual verification, and the low confidence interval [0, 0.5) is archived but does not trigger a warning.

[0019] Furthermore, the linkage mechanism between graph computation and stream computation in S5 is as follows: the high-confidence implicit relationship results output by the Pregel iterative inference in step S4 are serialized into incremental relation edge data, and written to the broadcast state of the stream computation engine through a message queue, so that all computing nodes hold the latest implicit relation edge set in their local memory; at the same time, the full knowledge graph relation data—including the original explicit relation edges and the implicit relation edges output by S4—is loaded into the state backend with the personnel identifier as the key and the relation subgraph of that personnel as the value; when the personnel appointment and removal event triggers risk calculation, the rule condition matching is first completed in the broadcast state with a time of less than 1 millisecond, and after the match is hit, the relation subgraph of that personnel is queried through the state backend with a time of less than 5 milliseconds, and the deepest 3-hop graph traversal is performed on the relation subgraph to retrieve the associated personnel with appointment avoidance risks. The median latency of a single end-to-end compliance verification is 340 milliseconds and the 99th percentile latency is 1.2 seconds.

[0020] Furthermore, the risk calculation channel in S5 supports dual-mode operation: near real-time batch processing mode—utilizing the batch processing capabilities of the stream computing engine, it performs a complete rule scan and graph implicit relationship mining on all files from 11 PM to 6 AM the next day, generating a full risk list report; real-time event-driven mode—it continuously monitors the data change event stream 24 hours a day, and completes a full rule scan and graph association query of the person and its related entities within milliseconds when the personnel appointment and dismissal approval form is submitted.

[0021] Furthermore, the risk score calculation formula in S6 is as follows: ,in The total risk score for a single warning event, with a range of values. It is used to drive the three-level early warning classification; Matching weights to rules constitutes a serious violation. General violations Minor risk ; The relationship mining confidence score output in step S4 of this invention is expressed in the above formula. The value of is [0, 1]. To meet the historical risk score based on the number of risk events recorded by the same target personnel within the past 12 months. , The number of risk events recorded involving the target personnel within the past 12 months; weighting coefficient. The analytic hierarchy process (AHP) is used to determine and support adjustments via configuration files.

[0022] Secondly, embodiments of the present invention also provide a system for mining implicit relationships in personnel files and providing early warning of compliance risks based on graph-flow linkage. Using the above-described method, the system achieves the mining of implicit relationships in personnel files and early warning of compliance risks. The system includes: The data governance module is used to perform the collection, cleaning, and standardization of multi-source heterogeneous data, and output a standardized archival data warehouse. The knowledge graph module is based on a graph database to build and store a knowledge graph of social relationships of people covering four dimensions: kinship, education, geography, and profession. It provides graph query and graph computing analysis interfaces. The rules engine module configures and manages a structured rule base for HR compliance, providing online editing, compilation and deployment, hot reloading, and version management functions for rules. The graph computation module performs Louvain community discovery algorithm, two-end BFS shortest path analysis, and Pregel model-based iterative inference algorithm on the knowledge graph to discover implicit relationships not directly recorded in the archives, and evaluates the confidence of the inferred relationships according to the chain decay confidence model. Risk calculation module – Based on the stream computing engine, a risk calculation channel is built. The rule set is loaded through broadcast state, and knowledge graph relationship data is loaded through the state backend. Real-time rule matching and graph association query are performed to realize the linkage between graph calculation results and stream computing channel. Early warning push module – Based on a comprehensive weighted risk score, early warnings are issued in a tiered manner, and the early warning information is distributed to the corresponding responsible parties through a multi-channel message push mechanism; The visualization module provides visualizations of risk heatmaps, relationship network topology, and early warning trend analysis.

[0023] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the above-described method.

[0024] The beneficial effects of the present invention are as follows (see the detailed implementation section for specific quantitative data): (1) Significantly improved audit coverage: This method achieves automated compliance check of all files, achieving full coverage of files without relying on manual spot checks, effectively overcoming the inherent limitations of manual audit in terms of coverage scale.

[0025] (2) Implicit relationship discovery capability: Based on the multidimensional relationship fusion of knowledge graph and collaborative reasoning of graph computing algorithm, this method can discover implicit kinship relationships that are not directly recorded in the archives and are difficult to identify by manual review. It can also automatically detect possible violations of the avoidance of office rules, making up for the blind spot of traditional manual review which can only make judgments based on explicit archive records.

[0026] (3) Real-time compliance verification capability: Through the linkage mechanism between graph computing results and stream computing channels, the delay of compliance verification triggered by appointment and removal events and the waiting time for approval are significantly shortened, which can meet the requirements of the appointment and removal approval process for immediate compliance verification.

[0027] (4) Effectiveness of early warning: This method classifies the inferred relationship through a chain decay confidence model and implements graded early warning based on a comprehensive weighted risk score. This expands the coverage of early warning to violations that are difficult to detect through traditional manual risk clue recording channels. It can significantly improve the breadth and accuracy of risk detection while ensuring a high verification rate of early warning information.

[0028] (5) Compatible with domestic software and hardware: Fully adaptable to domestic software and hardware environments, with core components being independently controllable.

[0029] Other features and advantages of the present invention will be set forth in the following description, and some of them may be learned by practicing the invention.

[0030] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0031] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0032] Figure 1 This is a schematic diagram of the overall method flow provided by the present invention, showing the complete processing flow from S1 data acquisition to S6 hierarchical early warning; wherein, the left side is the multi-source heterogeneous data input source, and the right side is the key technical components and final output results corresponding to each step.

[0033] Figure 2 This is a schematic diagram of the overall system architecture provided by the present invention, showing the core modules and their data flow relationships; Specifically: The infrastructure layer, located at the bottom of the architecture, provides all business modules above it with a unified operating system, CPU / NPU computing resources, database storage, and message queues, forming the physical and system support layer of the entire system. The data governance module, knowledge graph module, and rule engine module form the data preparation layer, operating horizontally and serving as inputs to each other. The data governance module collects and standardizes multi-source heterogeneous archival data; the knowledge graph module constructs a social relationship graph covering four dimensions—kinship, education, geography, and profession—based on standardized data; and the rule engine module transforms compliance requirements into a machine-executable structured rule base. Together, these three modules supply the downstream analysis and computing modules with data, graphs, and rules. The graph computing module, located above the data preparation layer, receives inputs from the knowledge graph module's graph and the rule engine module's inference rule set, and outputs high-level risk calculations to the risk calculation module. The confidence level implicit relationship edge set is the key computational link in the current system's transformation from "explicit archival data" to "implicit relationships not directly recorded." The risk calculation module sits above the graph calculation module, consuming the implicit relationship edge set, rule set, and real-time personnel appointment and dismissal events output by the graph calculation module. It completes rule condition matching through broadcast status, queries the relationship subgraph through the status backend, and outputs risk event records downwards, realizing the coordinated linkage between graph calculation results and stream calculation channels. The early warning push module and the visualization display module are located side by side at the top of the architecture: the former consumes risk event records downwards, classifies early warning levels according to preset thresholds, and distributes them to the corresponding responsible entities through multiple channels such as SMS, government WeChat, email, and in-system messages; the latter further consumes early warning messages and full risk records, presenting visualization results such as risk heatmaps, relationship network topology diagrams, early warning trend time series, and early warning handling dashboards to regulatory personnel.

[0034] Figure 3 This is a schematic diagram of the knowledge graph construction and implicit relationship mining process provided by the present invention, showing the complete process from entity extraction to implicit relationship discovery based on Pregel iterative reasoning.

[0035] Figure 4 This is a schematic diagram of the rule base organization structure provided by the present invention, showing the hierarchical organization structure of four types of compliance rules.

[0036] Figure 5 This is a schematic diagram of the stream computing risk processing channel provided by the present invention, showing the data flow of event access, rule matching, graph association query, and risk event output.

[0037] Figure 6 The diagram below illustrates the three-level early warning and graded push process provided by this invention, showing the complete decision-making chain from risk score calculation to closed-loop feedback. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0039] See Figure 1 As shown in the figure, this invention discloses a method for mining implicit relationships in personnel files and for early warning of compliance risks based on graph flow linkage. The method mainly includes the following steps: Construct a multi-dimensional knowledge graph oriented towards personnel social relationships. Using personnel file data from one or more data sources (including but not limited to personnel management information systems, personnel file management systems, compliance supervision information systems, population basic information systems, or academic qualification verification systems) as the data source, after ETL data collection, cleaning, and standardization, a personnel social relationship knowledge graph is constructed based on a graph database, covering four dimensions: kinship (spouse, parents-child, siblings), educational ties (classmates, alumni), geographical ties (same place of origin), and professional ties (colleagues, superiors and subordinates, former colleagues). Referencing the same entity across different data sources is resolved and entities are merged; multi-source cross-validation is performed on the same relationship—relationships existing in two or more authoritative data sources are assigned high credibility weights, while relationships existing only in a single data source are assigned basic credibility weights and marked as requiring further verification.

[0040] A dedicated structured rule base for HR compliance has been established. HR management compliance requirements are systematically categorized into four main types of structured, executable rules: rules for age, work experience, and identity verification; rules for recusal; rules for tenure rotation; and rules for compliance review. The rule engine uses a DSL syntax for formal rule definition, with each rule consisting of a condition part (LHS) and an action part (RHS). The rule base supports hot deployment (new rules are compiled into a Rete network and dynamically loaded into the running rule engine) and Git-based version management (using semantic versioning and Git-flow strategies). Rule changes require a three-tiered approval process: business expert review, technical expert review, and administrator approval. Verification is conducted using anonymized historical data in a sandbox environment, and only rules with a false positive rate below 5% are released.

[0041] The algorithm performs implicit relation mining using graph computation. It also performs collaborative processing of three types of graph computation operations on the knowledge graph. (a) Community Detection and Dense Subgraph Detection – The Louvain community detection algorithm is executed to divide the graph into several communities through modularity optimization, and the edge density within each community is calculated. When the edge density of a community is significantly higher than the average level of the entire graph and the community size exceeds a preset threshold, the community is marked as a suspected group with dense relationships. The unique value of the Louvain algorithm in human relationship graphs lies in its ability to identify close interpersonal networks formed by the superposition of multiple relationships such as kinship, education, geography, and profession. These networks are often undetectable by traditional single-dimensional review (such as reviewing only direct kinship).

[0042] (b) Shortest Path Analysis and Relationship Chain Length Calculation—A bidirectional BFS search is performed on person entity pairs to calculate the relationship chain length. Since kinship edges (approximately 65%) and colleague edges (approximately 25%) dominate the person relationship graph, the search prioritizes these two types of high-weight relationship edges. When the shortest path has no more than 3 hops and contains a specific kinship edge combination pattern, the path is marked as a suspected implicit relationship path. Utilizing the graph database's native index, the response time for a single shortest path query is controlled within 100 milliseconds.

[0043] (c) Iterative Reasoning Based on the Pregel Model—Modeling the Latent Relationship Mining Problem as a Graph Label Propagation and Reasoning Problem. First, a family attribute vector structure is defined. Each set stores a list of node IDs corresponding to that role. Explicitly declared family relationships in the profile are directly written into the family attribute vector of the corresponding node. In each Pregel iteration: Nodes Along the family relationship to the neighbors Send message ;node Merge after receiving all neighbor messages ,in for The set of neighboring nodes. Based on the propagated family attribute vector, four inference rules are applied—sibling inference, uncle-nephew / aunt inference, cousin / cousin inference, and brother-in-law / sister-in-law inference—to discover collateral blood relations and potential relations within three generations that are not directly recorded.

[0044] A real-time risk calculation channel integrating graph computation and stream computation is constructed. Based on the Apache Flink stream computation engine, the risk calculation channel differs from existing technologies in its key design feature: a collaborative mechanism between graph computation results and the stream computation channel. High-confidence implicit relationship results from Pregel iterative inference are serialized into incremental relationship edge data and written to Flink's Broadcast State via a Kafka message queue, ensuring all TaskManager nodes hold the latest implicit relationship edge set in their local memory. Simultaneously, the full knowledge graph relationship data (including original explicit relationship edges and output implicit relationship edges) is loaded into the RocksDB state backend using the personnel ID as the key and the personnel's multi-hop neighbor subgraph (including node attributes, relationship type, and edge weight) as the value. When a personnel appointment or removal event triggers risk calculation: rule condition matching is first performed in the local memory Broadcast State (time less than 1 millisecond). After a match is found, the personnel's relationship subgraph is queried through RocksDB (time less than 5 milliseconds), and a deepest 3-hop graph traversal is performed to retrieve associated personnel with appointment avoidance risks. The system supports dual-mode operation: near real-time batch processing mode—performs a full rule scan and implicit relationship mining on all files from 23:00 to 6:00 the next day, completing a full check of 150,000 files within 8 hours; real-time event-driven mode—continuously monitors Kafka event streams 24 hours a day, completing a full rule scan and graph association query within 340 milliseconds (median) after the personnel appointment and dismissal approval form is submitted.

[0045] Risk scores are used to generate tiered alerts and notifications. The scoring is based on a weighted average of three dimensions: rule matching weight. (Serious violation 1.0 / General violation 0.7 / Minor risk 0.4), Relationship mining confidence level and historical risk records , (Number of risk records in the past 12 months). Alerts are categorized into three levels: red, yellow, and blue, based on preset thresholds. Message push notifications utilize a redundant three-channel design: SMS, government WeChat, email, and in-app messaging. A visual dashboard (based on Vue + ECharts + WebSocket real-time push) displays a risk heatmap, a social relationship network topology (highlighting implicit relationship paths that trigger alerts), alert trend time series analysis, and an alert handling closed-loop dashboard.

[0046] The present invention will now be described in detail below with reference to the accompanying drawings, taking employee personnel file management as an example. Specific implementation methods are as follows: S1: Data Acquisition and Standardization S11: Data Source Access and Acquisition. Once data is available and legally authorized, configure a data access adapter for each accessible data source: For personnel management information systems (typically DM8 or openGauss relational databases), extract full and incremental data via a JDBC connector, with incremental synchronization using the domestic open-source component Canal for real-time Binlog parsing; for personnel file management systems, pull data via REST API or SFTP file transfer; for compliance supervision information systems, obtain relevant case data via a WebService interface; for population basic information systems, synchronize data via a front-end server and message queue; for academic qualification verification systems, perform online academic qualification verification via API and digital certificate authentication. Incremental synchronization is set up after initial full extraction. For large-scale data tables (such as work experience tables, typically with millions of records), set up paginated extraction (5000 rows per page) and a breakpoint resume mechanism (Redis records the extraction offset). Data anonymization: retain the first 6 and last 4 digits of the ID number, retain only the district / county level for detailed addresses, and completely anonymize contact information.

[0047] It should be noted that the method described in this invention can operate under the condition of a single data source. When only personnel file management system data is available, knowledge graph construction is based on explicit kinship, work experience, and educational background information in this single data source; implicit relationship mining can still be carried out through iterative reasoning of family members (such as inferring sibling relationships based on family member information filled in in an individual's personnel file); compliance verification can be completed based on information within the scope of this data source. The more data sources connected, the higher the coverage dimensions and confidence of the knowledge graph; however, even if only the personnel file management system is connected, this method can still complete implicit relationship mining and compliance verification based on explicit relationships.

[0048] S12: Data Cleaning and Quality Governance. Missing Value Handling: Records with missing values ​​in the 9 required core fields are directly marked as anomalies and fed back to the source system for correction; non-required fields are filled using the median (numerical) or mode (categorical) of employees of the same age group within the same unit. Outlier Detection employs a multi-method approach: box plots detect abnormal distributions in numerical fields; Z-scores detect extreme values ​​deviating from the mean; logical rules detect inconsistencies between fields (e.g., records with birth dates earlier than 1900, or work start dates earlier than birth dates, etc., which are illogically invalid). Format Standardization: Dates are standardized to ISO 8601 format (YYYY-MM-DD); place names are uniformly coded according to GB / T 2260 standard; education levels are uniformly mapped to the GB / T 4658 national standard level; ethnicity is uniformly mapped to the GB / T 3304 code. Duplicate Record Detection uses the ID number as the unique primary key for exact matching; for records with missing ID numbers, Levenshtein edit distance fuzzy matching is performed using the name, date of birth, and gender combination key.

[0049] S13: Entity Alignment and Data Dictionary Construction. Establish a semantically standardized mapping table covering all fields of personnel files, including date representation mapping, place name standardization mapping (including historical place names to current standard place names), educational level coding table (including correspondence between Chinese and foreign educational qualifications), and institution name standardization mapping table (full name-abbreviation-former name correspondence). Cross-system entity alignment is performed in two rounds: the first round uses the ID number as the exact matching key (expected matching rate 85%-90%); the second round uses a combination of name, date of birth, and gender for fuzzy matching. Entities with an alignment confidence score higher than 0.95 are automatically merged; those with a confidence score between 0.80 and 0.95 require manual review; and those with a confidence score lower than 0.80 are retained as independent entities.

[0050] S2: Constructing a multi-dimensional knowledge graph oriented towards people's social relationships: See Knowledge Graph Construction Figure 3 As shown: Entity extraction employs a hybrid strategy combining rule templates and NLP named entity recognition (structured fields are directly extracted using field mapping templates, while semi-structured and unstructured text uses a pre-trained BERT-BiLSTM-CRF named entity recognition model). Entity types include personnel entities, unit entities, family member entities, and social relationship entities. Relationship extraction covers four types of relationship edges: kinship edges, educational ties edges, geographical ties edges, and professional ties edges. Relationship extraction incorporates multi-source cross-validation: the confidence level is 0.95 if the same relationship is recorded in both the organizational department and the household registration system, and 0.70 if it exists only in a single data source. The fused knowledge graph is stored using the domestic open-source graph database TuGraph (attribute graph model), establishing node ID indexes, ID card number attribute indexes, name attribute indexes, and edge type indexes.

[0051] S3: Building a dedicated structured rule base for HR compliance: S31: Rule Classification and Definition. For example... Figure 4As shown, the four categories of rules are: Rules for verifying age, work experience, and identity (e.g., age restrictions of 16-65 years old at the time of employment, minimum age of 18 years old, minimum age of 21-28 years old for undergraduate graduation, minimum age of 24-32 years old for master's graduation, minimum age of 27-38 years old for doctoral graduation, minimum matching of education level and job level, etc.); Rules for recusal (e.g., recusal for spouses, direct blood relatives, collateral blood relatives within three generations, and close relatives by marriage, with clear definitions of recusal scope for positions within the same agency or with direct superior-subordinate relationships or supervisory and restraining relationships, etc.); Rules for job rotation (e.g., a maximum of 5 consecutive years for high-risk positions (personnel, finance, procurement, approval), and a maximum of 8 years for general positions, etc.); Rules for compliance review (e.g., marking at least 3 risk clues per year as "attention," restricting promotion within 3 years after disciplinary action, and marking incomplete rectification of compliance issues as "risk," etc.).

[0052] S32: Formal Expression of Rules. Design a custom domain description language layer to encapsulate the query statements and indicator calculation logic. A typical job avoidance rule is formally expressed as follows: Condition part—There are personnel A and personnel B, whose relationship type is one of spouse, direct blood relative, collateral blood relative within three generations, or close in-law, and they belong to the same unit and have a direct superior-subordinate or supervisory relationship; Action part—Generate a job avoidance violation warning, with a severity level of high.

[0053] S33: Rule Management and Deployment. Rule management utilizes a web-based visual interface, supporting online editing, syntax checking, sandbox testing, and approval for release. Version management employs a Git-flow strategy. Rule changes undergo a three-tiered approval process: business experts review the rule's logical correctness, technical experts review its execution efficiency and security, and administrators approve its release. Rule deployment supports hot deployment: new rules are compiled into a Rete network and dynamically loaded into the running rule engine without requiring a service restart. The rule repository is equipped with an independent sandbox testing environment, using anonymized historical data to replay and verify the effectiveness and false positive rate of new rules. Only rules with a false positive rate below 5% can be released to the production environment.

[0054] S4: Execute the graph computation algorithm to mine implicit relationships: Traditional relational databases face two fundamental difficulties in analyzing social networks: combinatorial explosion of JOIN operations and the inability to discover only explicit relationships. This invention fundamentally solves these problems through a graph computing paradigm, and the three types of graph computing operations (Louvain community discovery, two-way BFS, and Pregel iterative inference) work together to achieve implicit relationship discovery results that cannot be achieved by a single algorithm.

[0055] S41: Implementation of the Louvain community detection algorithm: The algorithm input is a knowledge graph. ,in For personnel entity node set, For family member node set, For a set of unit nodes, It is a relational edge set.

[0056] The algorithm consists of two phases of alternating iterations until the modularity no longer increases: Phase 1 (Modularity Optimization): For each node i, calculate the modularity increment after removing it from the current community and adding it to the community of its neighboring nodes. .choose The largest neighboring community moves, and this process is repeated until a local optimum is reached.

[0057] Phase 2 (Graph Compression): Each community obtained in Phase 1 is shrunk into a super node, and the edge weights between communities are aggregated into the edge weights between super nodes in the new graph. Phase 1 is then repeated for the new graph.

[0058] In the employee relationship graph, since the edge weights are assigned based on the quality score of multi-source cross-validation (high-confidence edges have higher weights), Louvain's modularity optimization naturally favors aggregating closely linked groups with high-confidence relationships, excluding nodes connected only by low-confidence relationships from the core community. For each detected community, when the internal edge density... With global average edge density When the ratio is greater than 3 and the number of community nodes is not less than 5, the community is marked as a suspicious group with dense relationships.

[0059] S42: Two-ended BFS shortest path analysis: Since kinship edges (approximately 65%) and colleague edges (approximately 25%) dominate in the employee relationship graph, the search prioritizes along these two types of high-weighted edges. Bidirectional BFS performs a breadth-first search simultaneously from both the source and target nodes, alternating layers at each step until the two search fronts meet. Utilizing the native indexes of the TuGraph graph database (node ​​ID B+ tree index and edge type bitmap index), the response time for a single shortest path query is controlled within 100 milliseconds. When the path length does not exceed 3 and the path contains a specific kinship edge combination pattern, the path is marked as a suspected implicit nepotism path and used as a candidate seed for Pregel iterative inference.

[0060] S43: Iterative inference algorithm based on Pregel model: Table 1 below shows the complete processing flow of the Pregel iterative inference algorithm: Table 1. Processing flow of the Pregel iterative inference algorithm

[0061] In this embodiment, the algorithm's time complexity is O(n log n). ;in For iteration rounds; The size of the knowledge graph node set is the union of three types of nodes: personnel entity nodes, family member nodes, and unit nodes (after deduplication). The size of the knowledge graph edge set is the union of edges representing four types of relationships: kinship, academic ties, geographical ties, and professional ties. On a knowledge graph containing 1 million nodes and 5 million edges, the computation time for full implicit relationship mining is 10-30 minutes. The convergence guarantee of the three iterations is based on the monotonically increasing property of family attribute vectors.

[0062] in, Represents a node In the The family attribute vector at the end of each iteration has an ordered quadruple structure: {father node set, mother node set, spouse node set, sibling node set}. During iteration, each subset only adds members and does not remove existing members. This holds true consistently; furthermore, since all the information required for the inference rules can be obtained within a propagation range of no more than 3 hops, the nodes... Each subset of the family attribute vector saturates after at most 3 rounds of iteration, and the algorithm will inevitably converge to a fixed point within at most 3 rounds.

[0063] S44: Calibration of key parameters for confidence modeling: Confidence decay factor =0.85 was determined through a parameter scanning experiment. Experimental method: 1000 full-field records were randomly selected from 150,000 records. Experts independently annotated implicit kinship relationships within these records, and unanimous confirmation of the relationships by the experts was used as the gold standard. Run the complete Pregel inference process with values ​​of 0.70, 0.75, 0.80, 0.85, 0.90, and 0.95 respectively, and calculate each... Precision, Recall, and Value. The experimental results are: =0.70 =74.3%, =0.75 =77.6%, =0.80 =80.2%, =0.85 =83.0% (optimal) =0.90 =82.6%, =0.95 =80.7%. (Selection) =0.85 is used as the default decay factor. Relationship type weights Determined by averaging the scores from independent participants.

[0064] S5: Construct a real-time risk calculation channel that integrates graph computation and stream computation: S51: Setting up a stream computing environment. For example... Figure 5 As shown, a stream computing cluster is built using Apache Flink 1.17 or later, deploying 3 to 5 TaskManager nodes (each with 16 CPU cores and 64GB of memory) and 1 JobManager node. Resource management uses YARN or Kubernetes. Checkpoints are set at 60-second intervals, with incremental checkpointing enabled. The state backend uses RocksDB with SSD storage.

[0065] S52: Graph Computation-Stream Computation Linkage Mechanism (a) Implicit Relation Knowledge Injection – After Pregel iterative inference is completed, the high-confidence implicit relationship results are serialized into incremental relation edge messages and written to the message queue via Kafka Topic. Flink jobs consume this Topic in Broadcast State mode, and each TaskManager node holds the latest implicit relation edge set in its local memory. The update strategy adopts a full replacement mode – after each implicit relationship mining task is completed (once daily in batch mode, incremental update in real-time mode), a complete high-confidence edge snapshot is generated and broadcast for update.

[0066] (b) Stateful Relationship Subgraphs – The entire knowledge graph relationship data (including original explicit and implicit relationship edges) is loaded into the RocksDB state backend with the employee ID as the key and the employee's 3-hop neighbor subgraph (including neighbor node IDs, relationship types, edge weights, and credibility scores) as the value. The size of each subgraph is controlled within 100KB (the average neighbor degree of the employee relationship subgraph is about 15-25).

[0067] (c) Event-Driven Compliance Validation – When a personnel appointment / removal event reaches the Flink operator: 1. Perform rule LHS condition matching in BroadcastState (memory lookup, less than 1 millisecond); 2. Query the 3-hop neighbor subgraph from RocksDB using the employee ID (SSD random read, less than 5 milliseconds); 3. Perform the deepest 3-hop DFS / BFS traversal on the subgraph to retrieve the violation relationship path; 4. Generate a risk event record (including event ID, triggering employee ID, associated employee ID, violation rule type, relationship path, confidence level, and timestamp). In real-time event-driven mode, the end-to-end median latency is 340 milliseconds, and the 99th percentile latency is 1.2 seconds.

[0068] S53: Out-of-order processing and fault tolerance. Out-of-order arrival of multi-source heterogeneous data is handled through a watermark mechanism (maximum out-of-order tolerance of 30 seconds). Events with delays exceeding the watermark are output to the delay processing channel for separate compensation calculation via a side output mechanism. A checkpoint mechanism guarantees exactly-once semantics. Back pressure monitoring and adaptive rate limiting mechanisms are established.

[0069] S6: Tiered Early Warning and Push Notifications S61: Risk score calculation. For example... Figure 6 As shown, each risk event is scored using a weighted average of three dimensions: rule matching weight. (Serious violation 1.0 / General violation 0.7 / Minor risk 0.4), Relationship mining confidence level (Directly using S4 output confidence values) and historical risk records , (Number of risk records in the past 12 months). Total risk score. The weighting coefficients are determined using the Analytic Hierarchy Process (AHP), a judgment matrix is ​​constructed, eigenvectors are calculated, and consistency checks are performed. The weighting factor is determined after the value is less than 0.1. The weighting factor can be adjusted by region in the configuration file.

[0070] S62: Three-level early warning classification and notification. Red alert ( The message is immediately pushed to the compliance supervision department and compliance management department, generating a verification work order with a time limit of 48 hours; a yellow alert ( The alert is sent to the organization and personnel department, generating a verification work order within 7 working days; Blue Alert ( The alerts are pushed to the archives management department and included in the monthly centralized processing batches. The message push adopts a three-channel redundancy design—the first channel is SMS, the second channel is WeChat for government affairs, and the third channel is email and internal messaging. Each alert message carries a risk event summary, risk level tag, and a URL leading directly to the risk details page. The system maintains a closed-loop record of the entire process: generation, push, receipt, verification, handling, and archiving, recording the operator, operation time, and operation content at each stage.

[0071] S63: Visualization. The visualization interface is based on the Vue front-end framework and ECharts rendering engine, with real-time data push via WebSocket. It features four functional areas: a risk heatmap (supporting drill-down at the provincial, municipal, and county levels), a relationship network topology map (force-directed graph or radial layout, highlighting implicit relationship paths, supporting multi-level node expansion and backtracking queries), early warning trend analysis (line charts displaying 7-day, 30-day, 90-day, and 365-day trends and year-on-year / month-on-month comparisons), and an early warning handling dashboard (handling progress, closure rate, and automatic escalation upon timeout).

[0072] II. System Architecture and Hardware Implementation Environment: like Figure 2 As shown, this invention is implemented through a system architecture consisting of the following core modules: Data governance module – Deploys the domestic open-source ETL tool DataX for data governance; The knowledge graph module is built on the domestic open-source graph database TuGraph. It stores a knowledge graph of social relationships of people covering four dimensions: blood ties, academic ties, geographical ties, and professional ties, and provides graph query and graph computing analysis interfaces. The rules engine module deploys the domestic open-source rules engine URule, configures and manages a structured rule base for the field of human resources compliance, and provides online editing, compilation and deployment, hot reloading and version management functions for rules; Graph computation module – executes Louvain community detection, two-end BFS and Pregel iterative inference algorithms; Risk computation module – deploys an Apache Flink stream computing cluster to perform real-time rule matching and graph association queries through a graph computing-stream computing linkage mechanism; The early warning push module—is based on a comprehensive weighted risk score to issue graded early warnings, and distributes early warning information to the corresponding responsible entities through a multi-channel message push mechanism; The visualization module provides visualizations of risk heatmaps, relationship network topology, and early warning trend analysis.

[0073] The modules communicate asynchronously and decoupledly through a Kafka message queue.

[0074] Hardware Environment: 3-5 compute servers (Phytium FT-2000+ or ​​Kunpeng 920 processors, with at least 128GB of memory), optionally equipped with Huawei Ascend 910 NPU accelerator cards for NLP model inference; storage is an SSD solid-state drive array (RAID5, total usable capacity at least 10TB) for storing standardized archival data warehouse, knowledge graph data, and risk record data; network is Gigabit Ethernet, with internal network latency between modules less than 1 millisecond, and network bandwidth between message queue clusters at least 1Gbps. Software Dependencies: Kylin V10 operating system, Bisheng JDK 11 or later, Apache Flink 1.17 or later, Apache Kafka 3.0 or later, TuGraph, URule. The system is fully adapted to the domestic IT innovation environment.

[0075] The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment, and will not be repeated here.

[0076] III. Verification of Technical Effects: The verification results in the simulated test environment (deployed based on approximately 150,000 simulated personnel files) are shown in Table 2 below: Table 2 Validation Results

[0077] The pilot results above verify that the present invention has achieved significant improvements over traditional solutions in three key indicators: audit coverage, accuracy of implicit relationship discovery, and real-time compliance verification delay; the measured data for each indicator are shown in Table 2 above.

[0078] As described in the above embodiments, this invention provides a method and system for mining implicit relationships in personnel files and providing early warning of compliance risks based on graph-flow linkage. The system systematically constructs a multi-dimensional social relationship knowledge graph for personnel files, achieving unified modeling and integration of four-dimensional relationships: kinship, academic ties, geographical ties, and professional ties. It systematizes and formalizes personnel management compliance requirements into a machine-executable structured rule base, supporting online editing, hot deployment, and version management. It discovers implicit relationships not directly recorded in the files through graph computing algorithms and effectively controls the false alarm rate of inference through confidence modeling. It coordinates the implicit relationship knowledge generated by offline graph computing with online flow computing pipelines to complete compliance verification based on the latest relationship graph as soon as personnel appointment and removal events occur. The entire process is systematically integrated to construct a complete technical solution that can be practically deployed in a domestic information technology innovation environment. Compared with existing technologies, the technical advantages of this invention are as follows: (1) Significantly improved audit coverage: This method completes the automated physical examination of all 150,000 files within 8 hours, with a coverage rate of 100%; compared with the traditional manual sampling mode (coverage rate of about 2.4%), it is about 41.7 times higher.

[0079] (2) Implicit Relationship Detection Capability: During the three-month pilot period, 237 pairs of implicit kinship relationships with high confidence were identified, and 198 pairs were confirmed by manual verification, achieving an accuracy rate of 83.5%. Among them, 162 pairs (81.8%) involved collateral blood relatives within three generations, and 45 pairs (22.7%) also involved violations of recusal rules, all of which were issues that had not been detected by manual review. This effect relies on the synergy of three types of graph computation operations: Louvain identifies suspicious groups to narrow the search scope, BFS quickly calculates the length of the relationship chain, and Pregel accurately infers the specific relationship type and confidence level.

[0080] (3) Real-time compliance verification capability: In the real-time event-driven mode, the end-to-end median latency is 340 milliseconds and the 99th percentile latency is 1.2 seconds, which meets the immediate response requirements of the personnel appointment and dismissal approval process. This effect directly benefits from the graph computing-stream computing linkage mechanism—implicit relationship knowledge is available in the local memory of the computing node through Broadcast State, compressing the verification latency from the second level to the hundred millisecond level.

[0081] (4) Effectiveness of early warning: During the pilot period, the red warning verification rate was 90.5% (38 / 42), the yellow warning verification rate was 70.0% (362 / 517), and the number of violations actively discovered by the system was 3.2 times that of the traditional risk clue recording channels.

[0082] (5) Compatibility with domestic IT innovation: The system has been fully adapted to the domestic IT innovation environment (Kylin V10 + Kunpeng 920 CPU + Ascend 910 NPU + DM8 database), and all core components (TuGraph / URule / Apache Flink / Apache Kafka / DataX) are open source or domestic independent products.

[0083] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, electronic devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.

[0085] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0086] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for mining implicit relationships in personnel files and providing early warning of compliance risks based on graph-flow linkage, characterized in that, The method includes the following steps: Using multi-source heterogeneous personnel file data as the data source, after cleaning and standardization, a knowledge graph containing personnel information, work experience, family members and social relationship entities is constructed based on a graph database. Transform human resources management compliance requirements into machine-executable structured rules; Graph computation operations are performed on the knowledge graph to mine and infer implicit relationships that are not directly recorded, and the inferred relationships are classified based on confidence level; A real-time risk calculation channel linking graph computing and stream computing is constructed. Implicit relationship results that meet the confidence requirements are written to the broadcast state of the stream computing engine through a message queue. At the same time, the full knowledge graph relationship data is loaded to the state backend. When personnel appointment and removal events trigger risk calculation, rule condition matching is first completed in local memory, and then the relationship subgraph is queried through the state backend to retrieve related personnel with appointment avoidance risks. The system uses a weighted scoring system that combines rule matching weights, relationship mining confidence levels, and historical risk records. It then classifies warning levels according to preset thresholds and distributes them to the corresponding responsible entities and regulatory nodes.

2. The method according to claim 1, characterized in that, The data sources of multi-source heterogeneous personnel file data include, but are not limited to, one or more of the following: personnel management information systems, personnel file management systems, compliance supervision information systems, basic population information systems, or academic qualification verification systems; The knowledge graph's relationship types cover four dimensions: kinship, academic relationship, geographical relationship, and professional relationship; it also performs dereference resolution and entity merging for the same entity from different data sources, conducts multi-source cross-validation for the same relationship, and assigns differentiated credibility weights based on the number of systems in which the relationship appears.

3. The method according to claim 1, characterized in that, The rule engine uses a DSL syntax to define a structured rule base, which includes rules for age, work experience and identity verification, avoidance of conflicts of interest, tenure rotation rules and compliance review rules. Each rule consists of a condition part and an action part.

4. The method according to claim 1, characterized in that, Perform various graph computation operations on the knowledge graph, including: Community detection and dense subgraph detection: The Louvain community detection algorithm is executed to divide the graph into several communities, calculate the edge density within the community, and mark a community as a densely connected group when the edge density of a community is higher than the average level of the entire graph and the size exceeds a preset threshold. Shortest path analysis: Perform bidirectional BFS search on personnel entity pairs to calculate the length of the relationship chain. When the number of hops in the shortest path does not exceed a preset threshold and the path contains a specific relationship edge combination pattern, it is marked as a suspected implicit relationship. Iterative reasoning based on the Pregel model: Define a family attribute vector. In each iteration, the node propagates the family attribute vector to its neighbors along the family member relationship edge. The receiver merges all neighbor messages and updates its own vector. Based on the propagated family attribute vector, multiple reasoning rules are applied to discover collateral blood relations and potential relationships that are not directly recorded.

5. The method according to claim 4, characterized in that, The Louvain community detection algorithm employs a two-stage iteration, wherein: Phase 1 is modularity optimization: For each node, calculate the modularity increment after removing it from the current community and adding it to the community of a neighboring node, select the neighboring community with the largest increment to move to, and repeat until a local optimum is reached; Phase 2 is graph compression: each community obtained in Phase 1 is shrunk into a super node, and the edge weights between communities are aggregated into the edge weights between super nodes in the new graph. Phase 1 is repeated for the new graph. The two phases are executed alternately until the modularity no longer increases; for each detected community, when the ratio of internal edge density to external edge density exceeds a preset threshold and the community size exceeds a preset minimum size, the community is marked as a suspected group with dense relationships.

6. The method according to claim 4, characterized in that, For each inferred implicit relation, a chain-decay confidence model is performed, and the confidence scores are graded based on the calculated scores. The confidence score calculation formula is as follows: in, To infer the implicit relationship The overall confidence level, and its range. This indicates the degree of credibility of the invention regarding the implicit relationship; implicit relationship The implicit relation triples obtained through Pregel iterative reasoning consist of two personnel nodes and the relationship type between them; To infer the relationship The set of ordered edges contained in the propagation path, that is, the ordered sequence formed along the edges of the knowledge graph from the starting point personnel to the ending point personnel; Number the edge sequentially within the path. = [1,2, …, L]; For the first on the path Data quality score for each edge; For the first on the path Relationship type weights of edges; The number of edges contained in the propagation path; The confidence decay factor is defined as follows: along the inference path, the confidence decays by α for each edge encountered, satisfying the condition... .

7. The method according to claim 1, characterized in that, The linkage mechanism between graph computing and stream computing is as follows: The implicit relationship results that meet the confidence requirements are serialized into incremental relationship edge data and written to the broadcast state of the stream computing engine through a message queue, so that all computing nodes hold the latest implicit relationship edge set in their local memory. At the same time, the full knowledge graph relationship data is loaded into the state backend with the personnel identifier as the key and the relationship subgraph of that personnel as the value. When the personnel appointment and removal event triggers risk calculation, the rule condition matching is first completed in the broadcast state and the time taken is less than the first preset time. After the match is hit, the relationship subgraph of that personnel is queried through the state backend and the time taken is less than the second preset time. Graph traversal is performed on the relationship subgraph to retrieve the related personnel with appointment avoidance risks.

8. The method according to claim 1, characterized in that, The risk calculation channel supports dual-mode operation, specifically: Near real-time batch processing mode: Utilizes the batch processing capabilities of the stream computing engine to perform a complete rule scan and graph implicit relationship mining on the full archive within a preset time period; Real-time event-driven mode: continuously monitors the data change event stream, and completes a full rule scan and graph association query of the person and its related entities in real time when the personnel appointment and dismissal approval form is submitted.

9. A system for mining implicit relationships in personnel files and providing early warning of compliance risks based on graph-flow linkage, characterized in that, The system includes: a data governance module, a knowledge graph module, a rules engine module, a graph computing module, a risk calculation module, an early warning push module, and a visualization display module. It uses the method described in any one of claims 1-8 to realize the mining of implicit relationships in personnel files and early warning of compliance risks.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to perform the method as described in any one of claims 1-8.

Images