Intelligent map construction and evolution system for digital employee skill library

Abstract

The application discloses a digital employee skill library intelligent graph construction and evolution system, and particularly relates to the field of graph data management, and is used for solving the problem that homologous copy skills are wrongly entered into a graph. Through acquiring employee skill related multi-source business data generated in an enterprise, the data is subjected to standardized processing and relationship extraction processing, skill evidence units are formed, and candidate skill relationships and a candidate relationship evidence chain set are obtained based on relationship attribution information. Then, the skill evidence units in the candidate relationship evidence chain set are subjected to evidence blood relationship analysis, evidence blood relationship clusters, evidence blood relationship coincidence matrices and an independent evidence cluster set are formed. Subsequently, blood relationship copy discrimination and relationship risk determination are performed on the candidate skill relationships, relationship risk types and graph access states are obtained. Finally, graph writing processing, expert verification task package assembly processing or relationship suppression processing are performed according to the relationship risk types and the graph access states, and skill graph evolution results are obtained.

Description

Technical Field

[0001] This invention relates to the field of graph data governance, and more specifically, to a system for constructing and evolving an intelligent graph of a digital employee skills database. Background Technology

[0002] In enterprise-wide digital human resource management scenarios, employee skill bases are typically not maintained by a single system, but rather supported by multiple types of business data, including project management systems, training management systems, job requirement systems, performance record systems, talent file systems, and process approval materials. This data may include employee project participation records, course completion records, job skill requirements, performance evaluation snippets, project summary materials, and fields automatically synchronized from business processes. To create a skill graph that can be used for talent matching, competency assessment, training recommendations, and job requirement analysis, current processing typically first standardizes objects such as employee names, employee IDs, skill names, course names, and project names. Then, it uses relation extraction to obtain candidate relationships between employee and skill objects, using identical or similar skill tags appearing in multiple places as the basis for relationship reliability.

[0003] However, during the generation and flow of internal business data, there are often issues such as field synchronization, template reuse, process derivation, and document duplication. For example, skill fields in project work orders may be synchronized to talent profiles, and skill descriptions in talent profiles may be referenced in job application materials or performance summaries; the same training record may also be repeatedly referenced in multiple reports, profile materials, or job matching materials. In this case, although a skill tag may appear in multiple systems or document fragments, this evidence may not necessarily come from independent business facts, but may simply be multiple presentations of the same original evidence after it has spread along the business process.

[0004] In existing skill graph construction processes, relation extraction results often focus more on whether skill tags appear, how frequently they appear, whether the textual descriptions are consistent, and whether the objects at both ends of the relation match. For candidate skill relationships that appear to have multiple sources of support, without reconstructing the source of evidence, the process of evidence generation, and the path of evidence flow, it is easy to misjudge duplicate evidence generated from the same source as independent verification from multiple sources. Such candidate skill relationships may not show low confidence under ordinary confidence evaluation; on the contrary, they may be given a high degree of credibility due to their repeated occurrence.

[0005] Therefore, as the employee skill library continues to evolve, it is necessary to perform evidence lineage analysis on the skill evidence units that support candidate skill relationships, identify whether they are homologous copying states formed by the diffusion of the same evidence lineage cluster, and control whether candidate skill relationships are written into the skill graph, enter the expert verification task package, or enter the relationship suppression process according to the relationship risk type and graph access status, so as to reduce the possibility of high-confidence but substantially homologous erroneous skill relationships entering the skill graph.

[0006] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent graph construction and evolution system for a digital employee skill library. This system acquires multi-source business data related to employee skills generated within the enterprise, performs standardization and relationship extraction processing to form skill evidence units, and obtains candidate skill relationships and candidate relationship evidence chain sets based on relationship attribution information. Then, it performs evidence lineage analysis on the skill evidence units in the candidate relationship evidence chain set to form evidence lineage clusters, evidence lineage overlap matrices, and independent evidence cluster sets. Subsequently, it performs lineage copying discrimination and relationship risk determination on the candidate skill relationships to obtain relationship risk types and graph access status. Finally, based on the relationship risk type and graph access status, it performs graph writing processing, expert verification task package assembly processing, or relationship suppression processing to obtain the skill graph evolution result, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: The evidence generation module acquires multi-source business data related to employee skills generated within the enterprise, performs standardization and relationship extraction processing on the multi-source business data related to employee skills, obtains skill evidence units, and forms a set of candidate skill relationships and candidate relationship evidence chains based on relationship attribution information. The bloodline analysis module performs evidence bloodline analysis on the skill evidence units in the candidate relationship evidence chain set. Based on the evidence description information and the upstream and downstream flow relationship, it determines the bloodline attribution between skill evidence units and forms evidence bloodline clusters, evidence bloodline overlap matrices and independent evidence cluster sets. The risk assessment module performs lineage copying and relationship risk assessment on candidate skill relationships based on the candidate relationship evidence chain set, evidence lineage cluster, evidence lineage overlap matrix and independent evidence cluster set, and obtains the relationship risk type and graph admission status. The graph evolution module controls the graph evolution process of candidate skill relationships based on the relationship risk type and graph admission status. For candidate skill relationships that are automatically admitted, graph writing is performed; for candidate skill relationships that are awaiting expert verification, expert verification task package assembly is performed; and for candidate skill relationships that are temporarily deferred, relationship suppression is performed, thus obtaining the skill graph evolution result.

[0009] Furthermore, multi-source business data related to employee skills is obtained through the enterprise's internal database interface, log interface, file parsing interface, and message synchronization interface. The data processing server organizes the original business records according to the access batch, business occurrence time, and source location, and performs normalization processing on employee objects, skill objects, job titles, project names, course names, dates, text encoding, and field formats to generate a sequence of original evidence records, which serves as input for relation extraction processing.

[0010] Furthermore, the skill evidence unit is generated from structured fields, log entries, document paragraphs, and skill-related fact fragments extracted from the text. The skill evidence unit carries evidence description information, relationship attribution information, and evidence index. The data processing server constructs relationship attribution information according to the objects at both ends of the normalized relationship, the relationship category, the business context identifier, and the extraction result identifier, and organizes the skill evidence units that support the same candidate skill relationship into a candidate relationship evidence chain set.

[0011] Furthermore, evidence lineage analysis includes: extracting the evidence source, evidence generation process, and evidence flow path from the evidence description information of the skill evidence unit; constructing evidence lineage keys according to a fixed field order; and identifying upstream and downstream flow relationships based on the transmission relationship between output nodes and input nodes. The data processing server establishes a lineage merging graph with skill evidence units as nodes, and merges evidence lineage clusters according to the evidence lineage keys and upstream and downstream flow relationships.

[0012] Furthermore, the data processing server determines the bloodline overlap value based on the content inheritance relationship, structural reuse relationship, and business derivation relationship between different evidence bloodline clusters, and writes the bloodline overlap value into the evidence bloodline overlap matrix; when determining the set of independent evidence clusters, it identifies related evidence cluster pairs according to the evidence bloodline overlap matrix and the preset bloodline overlap boundary, and includes evidence bloodline clusters that do not form related evidence cluster pairs with the selected evidence bloodline clusters into the set of independent evidence clusters.

[0013] Furthermore, the lineage copying determination includes: mapping the skill evidence units corresponding to the candidate skill relationships to the evidence lineage clusters to form a distribution state, and determining the homologous copying state based on the maximum cluster proportion and the set of independent evidence clusters; the data processing server also forms an action coverage state based on the business action attributes of the skill evidence units, and forms a content difference state based on the evidence content of the skill evidence units, the evidence lineage clusters, and the evidence lineage overlap matrix.

[0014] Furthermore, the data processing server calculates the number of landing points of statistical skill evidence units in each evidence lineage cluster based on the distribution status, obtains the cluster proportion vector after frequency normalization, and obtains the source overlap rate by combining it with the evidence lineage overlap matrix; at the same time, it constructs the action transfer counting matrix and action transfer matrix based on the business action attributes, evidence flow path and upstream and downstream flow correlation, and obtains the flow rate by the concentration of subsequent flow of business action attributes.

[0015] Furthermore, the data processing server combines the source duplication rate and the flow rate into a two-dimensional feature vector and inputs it into the Bayesian Gaussian mixture discriminant model. The Bayesian Gaussian mixture discriminant model consists of homologous copy risk components and non-homologous copy components, and obtains prior probabilities, mean vectors, and covariance matrices based on historical expert verification samples. The data processing server outputs the copy risk value and combines the homologous copy status, action coverage status, and content difference status to obtain the relationship risk type and graph admission status.

[0016] Furthermore, when the graph admission status is automatic admission and the candidate skill relationship meets the automatic admission conditions, the data processing server writes the candidate skill relationship into the skill graph and writes the relationship source, evidence index and admission record; when the graph admission status is pending expert verification and the candidate skill relationship meets the verification trigger conditions, the data processing server reads the candidate skill relationship, skill evidence unit, evidence lineage cluster and evidence lineage overlap matrix, and assembles the expert verification task package.

[0017] Furthermore, the data processing server receives the expert verification results corresponding to the expert verification task package; when the expert verification result is a pass, the server performs graph writing processing on the candidate skill relationship and establishes connections with the skill evidence unit, evidence lineage cluster, and evidence index; when the expert verification result is a fail, the data processing server writes the candidate skill relationship, relationship attribution information, evidence lineage key, risk reason, and evidence index into the relationship suppression table; when the candidate skill relationship is determined to be temporarily deferred, the data processing server writes the candidate skill relationship, relationship attribution information, evidence lineage key, risk reason, and evidence index into the relationship suppression table.

[0018] The technical effects and advantages of the present invention, the intelligent graph construction and evolution system for digital employee skill base: This invention focuses on the construction and evolution of an intelligent graph of a digital employee skill database. It transforms multi-source business data related to employee skills into skill evidence units, and uses a candidate relationship evidence chain set to carry the evidence source, generation process, and flow path of candidate skill relationships. This ensures that the relationship between employee objects and skill objects no longer relies solely on the frequency of skill tag appearances or text consistency. By restoring the lineage of skill evidence units through evidence lineage clusters and evidence lineage overlap matrices, it can identify whether the repeated appearance of the same skill tag in multiple locations such as project work orders, talent files, and job application materials indicates diffusion from the same source, reducing the probability of identically copied evidence being mistaken for multi-source independent evidence.

[0019] This invention establishes relationship risk types and map access status before writing candidate skill relationships into the skill map, providing a clear access control path for the evolution of the employee skill database map. Candidate skill relationships that meet the automatic access criteria can be written into the skill map; those with risks of duplicate entries, insufficient action coverage, or insufficient content differences can proceed to expert verification task package assembly; and those that fail verification or have their access temporarily suspended can undergo relationship suppression. Thus, employee skill relationships in the skill map can maintain correspondence with their evidentiary lineage and access records, reducing the contamination of employee skill profiles by erroneous skill relationships.

[0020] This invention stores candidate skill relationships, relationship attribution information, evidence lineage keys, risk reasons, and evidence indexes through a relationship inhibition table. This enables candidate skill relationships that are identified as risky to be identified and blocked in subsequent graph evolution batches, reducing the repeated entry of erroneous skill relationships with the same or similar lineages into the skill graph. Attached Figure Description

[0021] Figure 1 This is a flowchart of the skill evidence unit generation and candidate skill relationship construction process in this invention.

[0022] Figure 2 This is a flowchart of the candidate skill relationship risk determination and graph injection control process in this invention. Detailed Implementation

[0023] 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 only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 - Figure 2 This invention provides a system for constructing and evolving an intelligent graph of a digital employee skills database, comprising: In the following specific implementation, each step is executed by the data processing server. The data processing server connects to the enterprise's internal database interface, log interface, file parsing interface, and message synchronization interface to standardize, extract evidence, analyze evidence lineage, assess copying risks, and control access to the skill graph for multi-source business data related to employee skills generated within the enterprise, ultimately obtaining the skill graph evolution results. The multi-source business data related to employee skills consists of data records generated during the enterprise's internal business operations that can form evidence description information. These data records can extract necessary information from the evidence source, evidence generation process, evidence flow path, business action attributes, and evidence content. Data that cannot form the aforementioned evidence description information is not used as input for constructing candidate skill relationships in this batch.

[0025] The evidence generation module acquires multi-source business data related to employee skills generated within the enterprise, performs standardization and relationship extraction processing on the multi-source business data related to employee skills to obtain skill evidence units, and forms a set of candidate skill relationships and candidate relationship evidence chains based on relationship attribution information; the following sub-steps S101 to S104 are executed by the evidence generation module.

[0026] S101 Acquisition and Standardization of Multi-Source Business Data Related to Employee Skills: The data processing server uses multi-source business data related to employee skills output from internal enterprise database interfaces, log interfaces, file parsing interfaces, or message synchronization interfaces as the starting point for processing. It receives raw business records according to access batch, business occurrence time, and source location, and includes records that reflect employee participation in business activities, form skill evidence, generate skill tags, or are associated with job requirements into the same processing batch. A processing batch refers to the set of raw business records that are jointly read and enter into relation extraction during a single skill graph construction or evolution process.

[0027] The data processing server performs normalization processing on the original business records in the processing batch, converting employee names, employee numbers, job titles, project names, course names, skill names, and competency levels into a unified representation, and converting dates, text encoding, field delimiters, and document paragraph positions into a comparable format. For cases where the same employee appears in project work orders, training records, and job requirement tables under different names (name, employee number, or account), the data processing server aligns the objects based on the unique identifier of the employee within the enterprise. For cases where the same skill appears under names such as Java development, Java backend, or Java programming, it is grouped into the same skill object based on skill name aliases and skill unification rules.

[0028] After standardization, the data processing server generates a sequence of original evidence records. This sequence is formed by arranging the standardized original business records according to the access batch, business occurrence time, and source location. It preserves the source location and order of each record before relation extraction. For example, employee submissions of interface development records in project work orders, employee course completion records in training platforms, and job requirement skill records in job demand tables are all read and first enter the same processing batch. While employee and skill objects maintain a unified representation after standardization, the original records still retain differences in source location and business occurrence time, serving as the basis for subsequent evidence generation and evidence flow path values.

[0029] S102 Skill-Related Fact Identification and Skill Evidence Unit Generation: The data processing server references the original evidence record sequence and divides each normalized record in the sequence into identifiable segments. It then categorizes structured fields, log entries, document paragraphs, and extracted text containing content that expresses factual connections between employees and skills into skill-related fact fragments. A skill-related fact fragment is the smallest segment of content extracted from a normalized record that can be recognized by relation extraction rules or extraction models; it is used in subsequent processing to generate skill evidence units.

[0030] After skill-related fact fragments are formed, the data processing server performs relation extraction. For structured fields, employee objects, skill objects, and relation categories are determined according to field mapping relationships. For text fragments, an extraction method combining entity recognition and relation trigger word matching is used. First, employee objects, skill objects, and business objects are identified. Then, relation categories are determined based on trigger words such as "complete," "participate," "use," "pass," "responsible for," and "require," as well as the syntactic positions of the objects on both sides of the trigger words. The extraction process outputs an extraction result identifier, which is used to distinguish different relation extraction results that may exist in the same skill-related fact fragment and participates in the determination of subsequent relation attribution information. Field mapping relationships, relation trigger word tables, and business action attribute mapping tables are all stored in the configuration table of the data processing server. When there are multiple relation trigger words in the same skill-related fact fragment, the data processing server first determines candidate trigger words according to the syntactic distance between the trigger word and the employee object or skill object, with trigger words with shorter syntactic distances taking priority. If the syntactic distances are the same, the relation category and business action attribute are determined according to the preset priority in the relation trigger word table.

[0031] The data processing server encapsulates each extraction result into a skill evidence unit. Each skill evidence unit corresponds to an identifiable skill-related fact, and its evidence description information consists of standardized evidence source, evidence generation process, evidence flow path, business action attributes, and evidence content. Specifically, the evidence source indicates the business domain or location from which the skill evidence unit originates; the evidence generation process indicates the process by which the evidence is formed through input, synchronization, extraction, merging, or derivation; the evidence flow path indicates the trajectory of the evidence between business systems, fields, or documents; the business action attributes indicate the type of business behavior corresponding to the evidence; and the evidence content stores the text, field values, or fragments of content that can be extracted or compared. The data processing server also generates an evidence index for each skill evidence unit. This index is used to point to the source location of the corresponding skill evidence unit during graph writing, expert verification task package display, and audit review. The formation of candidate skill relationships and candidate relationship evidence chain sets does not rely solely on the evidence index. When expert verification task package display or access record tracing is required, the data processing server calls the evidence index to read the corresponding evidence content.

[0032] For example, if an employee A completes payment interface development in a project work order, and the work order field shows the technology stack as Java, the data processing server will classify employee A as the corresponding employee object, Java as the corresponding skill object, divide the work order description of completing payment interface development into skill-related fact fragments, and extract the application class relationship between the employee object and the skill object based on the completion and technology stack fields to form a skill evidence unit. If the same employee passes an advanced Java course in the course record, the course record will be processed in the same way to form another skill evidence unit. The two skill evidence units will retain their respective evidence sources, evidence generation processes, evidence flow paths, business action attributes, and evidence content.

[0033] S103 Relationship Attribution Information Construction and Candidate Skill Relationship Aggregation: For each skill evidence unit, relationship attribution information is constructed. The relationship attribution information is jointly determined by the objects at both ends of the normalized relationship pointed to by the skill evidence unit, the relationship category, the business context identifier, and the extraction result identifier. It is used to determine whether multiple skill evidence units support the same candidate skill relationship. Among them, the objects at both ends of the normalized relationship are used to limit the employee object and the skill object, the relationship category is used to limit the relationship semantics such as mastery, possession, participation, application, or improvement, the business context identifier is used to limit the business scenario such as project, course, job requirement, or process node, and the extraction result identifier is used to distinguish different extraction results in the same segment.

[0034] The data processing server aggregates skill evidence units according to relationship attribution information. During aggregation, it first compares whether the objects at both ends of the normalized relationship of the skill evidence unit are consistent, and then compares whether the relationship categories are the same. When the business context identifiers belong to the same business activity or the same job requirement scope, the corresponding skill evidence units are grouped into the same candidate skill relationship. When the objects at both ends of the relationship are consistent but the relationship categories are different, or when the business context identifiers indicate that they point to different business activities, different candidate skill relationships are formed respectively. After aggregation, the candidate skill relationships are still unverified relationships, which only indicate that there may be a corresponding association between a certain employee object and a certain skill object, and have not yet been entered into the skill graph writing process.

[0035] For each candidate skill relationship, the data processing server organizes all skill evidence units supporting that candidate skill relationship into a candidate relationship evidence chain set. The organization process arranges the skill evidence units sequentially according to the order of evidence generation, the order of evidence flow, and the source attribution, retaining the evidence description information, relationship attribution information, and evidence index for each skill evidence unit. For example, employee A has a Java application record in a project work order, a Java course pass record in a course record, and a Java project experience record in a job resume; if, after normalization, all three types of records point to an application or relationship between employee A and the Java skill object, and the business context identifier allows them to be categorized within the same business scope, then the skill evidence units corresponding to the three types of records are aggregated under the same candidate skill relationship, forming a candidate relationship evidence chain set for that candidate skill relationship.

[0036] The output of the S104 candidate relationship evidence chain set and the determination of subsequent call boundaries: Integrity verification is performed on each candidate relationship evidence chain set. The integrity verification is based on whether the evidence lineage analysis module can perform evidence lineage analysis. It checks whether each skill evidence unit retains evidence description information, relationship attribution information, and evidence index, and checks whether the generation order, flow order, and source attribution of skill evidence units in the candidate relationship evidence chain set can be read. When the same original business record forms duplicate fragments due to multiple synchronizations, the data processing server retains the differences in its evidence flow path and marks its source position through the evidence index, so that subsequent evidence lineage analysis can identify whether it belongs to the same source diffusion.

[0037] After integrity verification is completed, the data processing server outputs a set of candidate skill relationships and a set of candidate relationship evidence chains corresponding to each candidate skill relationship. The set of candidate skill relationships represents the range of relationships between employee objects and skill objects to be included in risk assessment in this batch; the set of candidate relationship evidence chains is used to transmit skill evidence units and their evidence description information to the lineage analysis module, enabling the lineage analysis module to construct evidence lineage keys based on evidence source, evidence generation process, evidence flow path, business action attributes, and evidence content, and further generate evidence lineage clusters and evidence lineage overlap matrices. This output does not change the existing relationship state in the skill graph, nor does it trigger graph writing processing.

[0038] Through the evidence generation module, the data processing server completes the construction of skill evidence units, candidate skill relationships, and candidate relationship evidence chains from multi-source business data related to employee skills. Each candidate skill relationship carries a traceable set of skill evidence units, and each skill evidence unit retains evidence description information, relationship attribution information, and evidence index. This forms the evidence lineage analysis input that can be directly called by the lineage analysis module, and provides the object basis for subsequent risk assessment and graph access control.

[0039] The lineage analysis module performs lineage analysis on the skill evidence units in the candidate relationship evidence chain set, determines the lineage attribution between skill evidence units based on evidence description information and upstream and downstream flow relationships, and forms evidence lineage clusters, evidence lineage overlap matrices and independent evidence cluster sets; the following sub-steps S201 to S204 are executed by the lineage analysis module.

[0040] S201 Evidence: Identification of Bloodline Bond Structure and Upstream / Downstream Transfer Relationships: The data processing server takes the candidate relationship evidence chain set output by the evidence generation module as the starting point for processing, reads the corresponding skill evidence unit one by one according to the candidate skill relationship, and extracts the normalized fields corresponding to the evidence source, evidence generation process and evidence flow path from the evidence description information of each skill evidence unit. This reading process takes the candidate relationship evidence chain set as the boundary and does not mix processing across candidate skill relationships, so that the restoration of evidence source under the same candidate skill relationship remains independent.

[0041] The data processing server combines the evidence source, evidence generation process, and evidence flow path in the order of source location, generation action, and flow nodes to obtain the evidence lineage key. The evidence lineage key is generated according to a fixed field order: evidence source, evidence generation process, and evidence flow path. When the evidence flow path contains multiple flow nodes, the data processing server arranges the flow nodes in ascending order according to the business occurrence time or flow record time before writing the evidence lineage key. If a field is missing, a preset null value is used as a placeholder, ensuring that the same evidence lineage key is generated when the same input record is executed repeatedly. The evidence lineage key is only used for evidence lineage parsing and evidence lineage cluster merging; it does not replace relationship attribution information.

[0042] For skill evidence units formed through field synchronization, the evidence lineage key retains the upstream and downstream field positions; for skill evidence units formed through file template reuse, the evidence lineage key retains the template source and generation process; for skill evidence units formed through process derivation, the evidence lineage key retains the business process nodes and derived actions. After the evidence lineage key is formed, the data processing server writes it back to the corresponding skill evidence unit, enabling subsequent merging processing to directly match based on the evidence lineage key.

[0043] After the evidence lineage key is constructed, the data processing server identifies the upstream and downstream flow relationships between skill evidence units based on the evidence flow path. If the output node recorded in the evidence flow path of one skill evidence unit is the same as the input node recorded in the evidence flow path of another skill evidence unit, and the evidence generation processes of the two show synchronization, copying, merging, or derivation actions, then an upstream and downstream flow relationship is determined to exist between them. If there is only the same skill name or the same employee object, but the evidence flow path cannot form a sequential connection from upstream output to downstream input, then an upstream and downstream flow relationship is not formed based on this. Upstream and downstream flow relationships are only established within the candidate relationship evidence chain set corresponding to the same candidate skill relationship. The data processing server does not establish upstream and downstream flow relationships because different candidate skill relationships have the same evidence source, the same business process node, or the same skill object.

[0044] For example, when employee A's Java skills record in the project management platform is synchronized into the talent profile table, and then derived from the talent profile table to the job matching materials, the data processing server identifies the upstream and downstream flow relationships in the order of project management platform records, talent profile table records, and job matching material records, and retains the same or traceable evidence lineage key for the corresponding skill evidence unit.

[0045] S202 skill evidence units are merged and evidence lineage clusters are generated: The data processing server uses the evidence lineage keys and upstream / downstream flow relationships output by S201 as the starting point for processing, and establishes a node in the lineage merging graph for each skill evidence unit in the same candidate relationship evidence chain set. The lineage merging graph is an intermediate structure with skill evidence units as nodes and homologous connections as edges, used to perform evidence lineage cluster merging; homologous connections include connections formed by identical evidence lineage keys, as well as connections formed by established upstream / downstream flow relationships.

[0046] The data processing server performs disjoint-set data collection (D&A) processing on the lineage merge graph. Initially, each skill evidence unit is a separate merge set. When two skill evidence units have the same evidence lineage key, their merge sets are merged. When two skill evidence units have upstream and downstream flow relationships, their merge sets are merged. When multiple skill evidence units form transitive connections through continuous synchronization, continuous derivation, or template reuse, they are merged sequentially according to the connection order until skill evidence units within the same connected range are included in the same merge set. After the D&A process is completed, each connected component in the lineage merge graph is output as an evidence lineage cluster.

[0047] After the evidence lineage clusters are generated, the data processing server records for each evidence lineage cluster the skill evidence units within the cluster, the corresponding candidate skill relationships, the set of evidence lineage keys within the cluster, and the upstream and downstream flow relationships within the cluster. If employee A's Java skill tag first appears in the project work order, is subsequently synchronized to the talent file, and is then referenced in the job matching materials, and all three skill evidence units support the same candidate skill relationship, then the disjoint-set data structure merges the three skill evidence units into the same evidence lineage cluster. If employee A has another independent training record showing that they have passed a Java course, and the evidence lineage key and evidence flow path of this training record are not connected to the project work order chain, then the skill evidence unit corresponding to this training record forms another evidence lineage cluster. Thus, the seemingly scattered evidence in the candidate relationship evidence chain set is restored into evidence lineage clusters organized by source and flow.

[0048] S203 Evidence of Bloodline Overlap Matrix Generation: The data processing server uses the evidence lineage clusters output by S202 as the starting point for processing. It traverses the corresponding evidence lineage clusters according to the candidate skill relationships, and sequentially determines the content inheritance relationship, structural reuse relationship, and business derivation relationship between different evidence lineage clusters. The content inheritance relationship is determined by the fragment inclusion, keyword sequence preservation, and semantic reuse between evidence content within the cluster; the structural reuse relationship is determined by the template consistency of the form, document paragraph, or field structure where the evidence content within the cluster is located; and the business derivation relationship is determined by whether there is a derivation link between the business object or process node corresponding to the skill evidence unit within the cluster.

[0049] For any two evidence lineage clusters, the data processing server generates content inheritance, structure reuse, and business derivation tags, respectively. The content inheritance tag indicates whether there is fragment inclusion, keyword sequence retention, or semantic reuse in the evidence content of the two evidence lineage clusters. The structure reuse tag indicates whether the two evidence lineage clusters originate from the same form template, document template, or field structure sequence. The business derivation tag indicates whether the two evidence lineage clusters are derived from the same business object or the same process node. The content inheritance tag is determined based on the keyword sequence retention ratio. When the keyword sequence retention ratio reaches the content inheritance limit, the content inheritance tag is valid. The content inheritance limit is obtained by taking the quartile of the keyword sequence retention ratios of evidence cluster pairs confirmed as content duplication in the historical expert verification sample, sorted in ascending order, and written into the configuration table. The calculation process for the keyword sequence retention ratio is as follows: The data processing server first performs word segmentation and stop word removal on the evidence content in the two evidence lineage clusters, retaining words related to employee objects, skill objects, business objects, project names, course names, job names, and business actions to form the first keyword sequence and the second keyword sequence; then, the number of identical keywords in the two keyword sequences is counted, and the number of keywords in the shorter keyword sequence is used as the denominator to obtain the keyword sequence retention ratio; if either keyword sequence is empty, the keyword sequence retention ratio is 0.

[0050] Suppose evidence of kinship clusters and evidence of bloodline clusters The bloodline overlap value between them is The data processing server generates content inheritance tags respectively. Structural reuse mark and business-derived tags When the evidence content in two evidence clusters contains overlapping segments, or when the retention rate of keyword sequences after word segmentation reaches the content inheritance limit, ,otherwise When the form template identifier, document template identifier, or field structure sequence corresponding to two evidence lineage clusters are consistent, ,otherwise When two evidence lineage clusters correspond to the same business object, and the evidence flow path shows a derivation relationship from the same process node, ,otherwise .when hour, ;when and hour, Take the highest overlap level; when and hour, Take the content overlap level; when and hour, Take the template overlap level; when , and hour, The independent level is determined. High overlap level, content overlap level, template overlap level, and independent level are all stored in the configuration table. When the historical expert verification sample is insufficient to statistically determine the level values, the high overlap level is initialized to 1, the content overlap level to 0.75, the template overlap level to 0.50, and the independent level to 0. After the historical expert verification sample is formed, the data processing server updates the above levels based on the evidence clusters confirmed by experts as homologous copying, content copying, template reuse, and independent source. The updated levels only apply to subsequent processing batches and meet the following requirements. High overlap level Content overlap level Template overlap level Independent level .

[0051] in, Indicates the first The evidence of bloodline and the first The degree of overlap in blood relations between individual evidence bloodline clusters is used to form an evidence bloodline overlap matrix; Whether a content inheritance relationship is established is determined by comparing the content of the evidence. Indicates whether a structural reuse relationship is valid, derived from a form template, document template, or field structure sequence; Whether a business derivation relationship is established depends on the business object, process node, and evidence flow path. and This represents the sorting number of evidence lineage clusters under the same candidate skill relationship. The rows and columns of the evidence lineage overlap matrix correspond to evidence lineage clusters under the same candidate skill relationship, and the matrix elements represent the degree of overlap between the evidence lineage clusters corresponding to the rows and columns.

[0052] In the context of an enterprise employee skills database, if one evidence lineage cluster originates from a project work order and another from job application materials, but both contain the same project description fragments and skill tags, and the job application materials' workflow points to project work order derivation, the data processing server determines that the two have a content inheritance relationship and a business derivation relationship, and writes a high overlap level in the evidence lineage overlap matrix. If two evidence lineage clusters originate from different employee training templates, with the same field structure but different course names, business contexts, and evidence content, then a template overlap level or an independent level is written. This evidence lineage overlap matrix is ​​then used as the matrix input when the risk assessment module calculates the source duplication rate.

[0053] S204 Determination and Output of Independent Evidence Cluster Sets: The data processing server uses the evidence lineage clusters output by S202 and the evidence lineage overlap matrix output by S203 as the starting point for processing, and reads the preset lineage overlap boundary. The preset lineage overlap boundary is used to determine whether two evidence lineage clusters should be processed as related sources of evidence. When the number of evidence clusters confirmed as homologous copies in the historical expert verification samples is sufficient to calculate the lower quartile, the data processing server sorts the lineage overlap values ​​corresponding to the sample set in ascending order and takes the lower quartile as the preset lineage overlap boundary; when the number of historical expert verification samples is insufficient, the data processing server reads the initial preset lineage overlap boundary from the configuration table. The initial preset lineage overlap boundary is the overlap level of the content saved in the configuration table in S203; the preset lineage overlap boundary updated according to the expert verification samples only applies to the new processing batch and does not reverse the already output evidence lineage clusters, evidence lineage overlap matrices, and map admission status.

[0054] The data processing server compares the overlap values ​​of evidence clusters one by one according to the evidence lineage overlap matrix. If the overlap value between two evidence clusters reaches the preset overlap limit, or if there is an upstream and downstream flow relationship between the two evidence clusters as identified by S201, then the two are marked as an associated evidence cluster pair; if the overlap value is lower than the preset overlap limit and there is no upstream and downstream flow relationship, then the two are marked as an independent evidence cluster pair.

[0055] Before filtering the set of independent evidence clusters, the data processing server first performs a fixed sorting of the evidence lineage clusters corresponding to the same candidate skill relationship. The sorting rules are as follows: first, they are sorted in ascending order according to the earliest business occurrence time of the skill evidence units within the evidence lineage cluster; if the earliest business occurrence times are the same, they are sorted in descending order according to the number of skill evidence units within the evidence lineage cluster; if they are still the same, they are sorted in lexicographical order according to the evidence lineage key. When any skill evidence unit within an evidence lineage cluster lacks a business occurrence time, the data processing server uses the access time of that skill evidence unit when it enters the processing batch as the replacement time; if the access time is still missing, the generation time of the evidence index record is used as the replacement time. The data processing server selects evidence lineage clusters sequentially according to this fixed order and determines whether the current evidence lineage cluster forms a related evidence cluster pair with the selected evidence lineage clusters. Evidence lineage clusters that do not form a related evidence cluster pair with the selected evidence lineage clusters are included in the independent evidence cluster set. Evidence lineage clusters that form a related evidence cluster pair with the selected evidence lineage clusters are retained in the evidence lineage clusters, but are not counted as new independent sources in the independent evidence cluster set. This ensures that the same candidate relationship evidence chain set yields a consistent independent evidence cluster set when executed repeatedly.

[0056] After screening, the data processing server outputs the evidence lineage clusters, evidence lineage overlap matrix, and independent evidence cluster set corresponding to the candidate skill relationships, while retaining the correspondence between the evidence lineage clusters and the skill evidence units and candidate relationship evidence chain set. If the lineage overlap values ​​between the project work order cluster, talent file cluster, and job application material cluster all reach the preset lineage overlap threshold, they are only processed as mutually related evidence sources. If the lineage overlap value between the training record cluster and the above evidence lineage clusters is lower than the preset lineage overlap threshold, and there is no upstream or downstream flow relationship, the training record cluster is included in the independent evidence cluster set. Based on this, the risk assessment module can count the number of landing points of the skill evidence units corresponding to the candidate skill relationships in each evidence lineage cluster, and use the independent evidence cluster set for same-source copying status identification.

[0057] Through the lineage analysis module, the data processing server completes the lineage analysis of the candidate relationship evidence chain set, forming evidence lineage keys, upstream and downstream flow relationships, evidence lineage clusters, evidence lineage overlap matrices, and independent evidence cluster sets corresponding to the candidate skill relationships. These objects maintain a correspondence with the skill evidence units and the candidate relationship evidence chain set, and can be directly used by the risk assessment module for cluster proportion vectors, source duplication rates, homologous copy status, and subsequent relationship risk assessment.

[0058] The risk assessment module performs lineage copying and relationship risk assessment on candidate skill relationships based on the candidate relationship evidence chain set, evidence lineage cluster, evidence lineage overlap matrix, and independent evidence cluster set, and obtains the relationship risk type and graph admission status; the following sub-steps S301 to S304 are executed by the risk assessment module.

[0059] S301 Candidate Skill Relationship Distribution Status Analysis and Homologous Copying Status Identification: The data processing server takes the candidate skill relationships and candidate relationship evidence chain sets output by the evidence generation module, as well as the evidence lineage clusters and independent evidence cluster sets output by the lineage analysis module, as its starting point. It reads the corresponding candidate relationship evidence chain sets one by one according to the candidate skill relationships and maps each skill evidence unit to the evidence lineage cluster formed by the lineage analysis module. This mapping is completed based on the attribution result of the skill evidence unit in the evidence lineage cluster, outputting the distribution status of candidate skill relationships in different evidence lineage clusters. The distribution status refers to the number of skill evidence units supporting the same candidate skill relationship falling into each evidence lineage cluster and their corresponding cluster attribution relationship.

[0060] The data processing server counts the number of landing points of the skill evidence units corresponding to the candidate skill relationship in each evidence lineage cluster and determines the evidence lineage cluster corresponding to the maximum number of landing points. When the proportion of the maximum number of landing points to the total number of skill evidence units corresponding to the candidate skill relationship reaches the homogeneity set limit, the homogeneity replication state is established. When the skill evidence units supporting the candidate skill relationship are distributed across multiple evidence lineage clusters, but the number of evidence lineage clusters that can enter an independent evidence cluster set is less than 2, the homogeneity replication state is established. When the skill evidence units are distributed across multiple evidence lineage clusters, and at least 2 mutually independent evidence lineage clusters can support the candidate skill relationship, the homogeneity replication state is not established. The homogeneity set limit is obtained by taking the quartile of the largest cluster proportion of candidate skill relationships confirmed as homogeneous replication in the historical expert verification samples, sorted in ascending order. When historical samples are insufficient, the data processing server reads the initial homogeneity set limit from the configuration table, which is 0.60. After subsequent historical expert verification samples meet the statistical conditions, the data processing server updates the homogeneity set limit, and the updated homogeneity set limit only applies to subsequent processing batches. The maximum cluster proportion is derived from the cluster proportion vector corresponding to the candidate skill relationship, and the set of independent evidence clusters is derived from the lineage analysis module.

[0061] For example, the candidate skill relationship between employee A and a Java skill object is supported by project work orders, talent files, and job application materials. The lineage analysis module has grouped the skill evidence units corresponding to the project work orders, talent files, and job application materials into the same evidence lineage cluster. Based on this, the data processing server determines that the distribution of the candidate skill relationship is concentrated on the same source diffusion path, and the same source duplication status is established. If another training record forms an independent evidence cluster and also supports the candidate skill relationship between employee A and a Java skill object, the data processing server retains this independent source in the distribution status, providing a basis for differentiation for subsequent source duplication rate calculation and relationship risk determination.

[0062] S302 Source Repetition Rate Calculation and Business Action Flow Rate Calculation:

[0063] The data processing server uses the distribution state formed by S301, the candidate relationship evidence chain set output by the evidence generation module, and the evidence lineage clusters and evidence lineage overlap matrix output by the lineage analysis module as the starting point for processing. For any candidate skill relationship, the data processing server counts the number of landing points of the skill evidence units corresponding to the candidate skill relationship in each evidence lineage cluster, and performs frequency normalization processing based on the total number of skill evidence units corresponding to the candidate skill relationship to obtain the cluster proportion vector. The dimension order of the cluster proportion vector is consistent with the row and column order of the evidence lineage overlap matrix, both arranged according to the fixed sorting of evidence lineage clusters under the same candidate skill relationship. Subsequently, the cluster proportion vector and the evidence lineage overlap matrix are subjected to a quadratic form operation to obtain the source overlap rate: , in, For candidate skill relationships The source overlap rate is used to indicate the degree of overlap in ancestral origin among the evidence supporting the candidate skill relationship; For candidate skill relationships The cluster proportion vector, each dimension of which is obtained by dividing the number of skill evidence unit landing points in the corresponding evidence lineage cluster by the total number of skill evidence units corresponding to the candidate skill relationship; For candidate skill relationships The corresponding evidence lineage overlap matrix is ​​generated by the lineage analysis module based on content inheritance relationships, structural reuse relationships, and business derivation relationships. This is the transpose of the cluster proportion vector. The source weight rate is a dimensionless statistical scale used to enter the Bayesian-Gaussian mixture discriminant model.

[0064] After the source weight ratio is calculated, the data processing server constructs an action transition counting matrix based on the business action attributes, evidence flow paths, and upstream and downstream flow relationships within the same candidate relationship evidence chain set. The rows of the action transition counting matrix correspond to the current business action attribute, and the columns correspond to the subsequent business action attributes. When there is an upstream and downstream flow relationship between two skill evidence units, the data processing server reads the business action attribute of the upstream skill evidence unit as the current business action attribute and reads the business action attribute of the downstream skill evidence unit as the subsequent business action attribute, incrementing the corresponding matrix position by 1. After traversing all upstream and downstream flow relationships, row normalization is performed on the rows with subsequent actions to obtain the action transition matrix.

[0065] Establish candidate skill relationships The corresponding action transition counting matrix is , of which elements Indicates business action attributes The transfer of upstream and downstream related relationships to subsequent business action attributes The number of times. For business action attributes with subsequent actions. The data processing server calculates the total number of transfers. and in accordance with Obtain the action transition matrix elements in .in, For candidate skill relationships The action transition counting matrix; For business action attributes Transfer to business action attributes The number of times; For business action attributes The total number of subsequent transfers; This is the action transition matrix; For business action attributes Transfer to business action attributes The proportion.

[0066] When candidate skills relationship When there are no upstream or downstream action transfer records in the corresponding action transfer counting matrix, the total number of subsequent transfers for all business action attributes is 0, and the data processing server will then consider this candidate skill relationship. flow rate A value of 0 is used to represent an unobserved, standardized transfer path. For candidate skill relationships... Business action attributes The concentration of subsequent flows is the largest proportion of subsequent transfers in the row corresponding to this business action attribute, i.e. Business action attributes Its proportion in action transfer is: Candidate Skill Relationships flow rate It is obtained according to the following formula: , in, For candidate skill relationships The flow rate is used to indicate the degree to which the evidence supporting the candidate skill relationship is solidified in the flow of business actions; This refers to the attributes of the current business action. For attributes of subsequent business actions; For business action attributes The total number of subsequent transfers; For business action attributes Transfer to business action attributes The proportion; Indicates business action attributes The proportion of the most concentrated subsequent flows. The closer the flow rate is to 1, the more concentrated the business actions are to a few fixed subsequent paths.

[0067] S303 Action Coverage Status and Content Difference Status Formation: The data processing server takes the candidate relationship evidence chain set output by the evidence generation module, the action transition counting matrix formed by S302, and the business action attributes and evidence content in the skill evidence unit as the starting point for processing. It then statistically analyzes the types of business action attributes covered by the skill evidence units supporting the candidate skill relationship. The data processing server deduplicates the business action attributes supporting the same candidate skill relationship to form an action coverage set. This action coverage set, which is the deduplicated set of business action attributes of the skill evidence units corresponding to the candidate skill relationship, is used to determine whether the evidence covers different business stages.

[0068] When the action coverage set contains only one or more of the following: input, synchronization, and template derivation, and does not contain any of the following business action attributes: formation, execution, verification, and feedback, the action coverage status output is insufficient. When the action coverage set contains at least two different types of business action attributes: formation, execution, verification, and feedback, the action coverage status output is sufficient. When the action coverage set contains only one type of business action attribute: formation, execution, verification, and feedback, if the skill evidence unit corresponding to that business action attribute comes from the evidence lineage cluster in the independent evidence cluster set, the action coverage status output is sufficient; otherwise, the output is insufficient.

[0069] After the action coverage state is established, the data processing server continues to read the evidence content from the same candidate relationship evidence chain set and compares it in groups according to the evidence lineage cluster attribution results in the lineage analysis module. The determination of the content difference state is based on the evidence content, evidence lineage cluster attribution, and evidence lineage overlap matrix. The data processing server first performs a standardized text comparison on the evidence content under the same candidate skill relationship. If the evidence content of different skill evidence units only has changes in field position, format, encoding, or fixed template statements, the content difference state output is insufficient. If at least two skill evidence units belong to different evidence lineage clusters, and the lineage overlap value between the corresponding evidence lineage clusters is lower than the preset lineage overlap limit, and the evidence content of the two skill evidence units respectively contains descriptions of employee objects, skill objects, and business objects under different business contexts, the content difference state output is sufficient.

[0070] For example, if the candidate skill relationship between employee A and a Java skill object appears in the same project description in project work orders, talent files, and job application materials, and the business action attributes are all synchronous or derived, the data processing server outputs an action coverage status of insufficient and a content difference status of insufficient. If the same candidate skill relationship is also supported by project execution records and course pass records, and these correspond to execution actions and verification actions respectively, with the evidence content describing interface development and course assessment results respectively, then the data processing server outputs an action coverage status of sufficient and marks the content difference status as sufficient. These two statuses, along with the same-origin copy status, source duplication rate, and flow stability rate, are included in the relationship risk assessment.

[0071] S304 Copy Risk Value Output and Relationship Risk Assessment: The data processing server uses the source duplication rate and flow stability obtained from S302, the same-source copy status obtained from S301, and the action coverage status and content difference status obtained from S303 as the starting point for processing. (Regarding candidate skill relationships...) The data processing server will reduce the source redundancy rate. and flow rate Composition of two-dimensional feature vectors The Bayesian-Gaussian mixture discriminant model is a non-neural network statistical discriminant model, which does not contain neural network layers, neurons, activation functions, optimizers, or backpropagation training processes; the model consists of homology copying risk components and non-homology copying components, among which... This indicates the risk of homologous copying. This represents non-homologous copy components, and the model input is candidate skill relationships. Two-dimensional feature vectors The model output is the risk value of replication. The training dataset is derived from historical expert-verified samples. Each sample includes at least the source duplication rate, flow rate, relationship risk type, and expert verification result. The model training objective is to maximize the probability density of labeled samples under their corresponding components, which is equivalent to minimizing the negative log-likelihood of labeled samples. Since the sample components are determined by the expert verification results and relationship risk type, the data processing server uses supervised parameter estimation to obtain the prior probability of each component. Mean vector Covariance Matrix Initial model parameters are stored in the model configuration; in the absence of historical samples, prior probabilities are... and All values ​​are initialized to 0.5, and the mean vector of the homologous copy risk components is used. Initialize to The mean vector of non-homologous copying components Initialize to The covariance matrices of both types of components are initialized as two-dimensional matrices with diagonal elements of 0.05 and off-diagonal elements of 0. After the historical expert verification samples are formed, the data processing server updates the above model parameters with the historical expert verification samples. The updated model parameters only apply to subsequent processing batches.

[0072] Historical expert-verified samples were divided into two components based on expert verification results and relationship risk type: samples that failed expert verification and had a relationship risk type of homology copying risk were classified as the homology copying risk component, denoted as... Samples that pass expert verification and have a low-risk relationship are classified as non-homologous copy components, denoted as... For each component class, the data processing server calculates the prior probability by dividing the number of samples for that component by the total number of samples for both components. The mean vector is the average of the two-dimensional feature vectors of the component sample. The covariance matrix is ​​the covariance of the two-dimensional eigenvectors of the component sample. When the number of samples for any component is less than 3, the data processing server does not update the parameters of the Bayesian-Gaussian mixture discriminant model, but uses the previous version of the model parameters from the configuration table. If there are no previous version of the model parameters, candidate skill relationships that trigger the homology replication state, insufficient action coverage state, or insufficient content difference state are identified as awaiting expert verification. To prevent the covariance matrix from becoming non-invertible, the data processing server adds a preset regularization term to the diagonal of the covariance matrix, with the regularization term taking the smallest positive number from the configuration table. After completing the above parameter acquisition, the data processing server calculates the candidate skill relationships according to the following posterior probabilities. Copy risk value: , in, For candidate skill relationships The risk value of copying; The two-dimensional Gaussian probability density function is defined according to... Calculate, where, Covariance matrix The determinant, Covariance matrix The inverse matrix, It is a natural exponential function. Indicates candidate skill relationships The two-dimensional feature vector relative to the components The offset of the mean vector; It is a two-dimensional feature vector composed of source weight ratio and flow stability. and These are the prior probabilities of homologous copy risk components and non-homologous copy components, respectively. and These are the mean vectors of the two types of components, respectively; and These are the covariance matrices of the two types of components, respectively. Derived from cluster proportion vector and evidence lineage overlap matrix, Derived from the action transition matrix, , , The data is derived from historical expert-verified samples or initial samples confirmed by experts. When historical expert-verified samples are insufficient to estimate the parameters of the two types of components, the data processing server does not force the model parameter update for the current batch. Instead, it identifies candidate skill relationships that trigger insufficient homology, insufficient action coverage, or insufficient content difference states as pending expert verification and updates the model configuration used in subsequent batches after the expert verification results are formed.

[0073] The copy risk threshold is determined by historical expert-verified samples. The data processing server first calculates the copy risk value for each candidate skill relationship in the historical expert-verified samples, and identifies samples that failed expert verification and whose relationship risk type is homogeneous copy risk as risk samples. Then, the copy risk values ​​of the risk samples are sorted in ascending order, and the 25th percentile is taken as the copy risk threshold and written into the model configuration. The 25th percentile is determined as follows: Assume the number of risk samples is... The data processing server arranges the copy risk values ​​of the risk samples in ascending order and selects the sorting position as... The copy risk value is used as the copy risk threshold, where This indicates rounding up. If the number of risk samples is insufficient, the initial copy risk threshold is read from the configuration table; after subsequent risk samples are formed, the data processing server updates the model configuration, and the updated copy risk threshold only applies to subsequent processing batches. The copy risk threshold is used to determine whether candidate skill relationships need to enter expert verification, and its input sources are the copy risk values ​​of historical expert verification samples and the expert verification results.

[0074] After the copy risk value is generated, the data processing server reads the copy risk threshold from the model configuration. When the copy risk value is not less than the copy risk threshold, the relationship risk type is determined to be same-origin copy risk, and the graph admission status is determined to be pending expert verification. When the copy risk value is lower than the copy risk threshold, the same-origin copy status, action coverage status, and content difference status are read and combined for judgment. If the same-origin copy status is valid, the relationship risk type is determined to be same-origin copy risk, and the graph admission status is determined to be pending expert verification. If the action coverage status is insufficient, the relationship risk type is determined to be insufficient action coverage risk, and the graph admission status is determined to be pending expert verification. If the content difference status is insufficient, the relationship risk type is determined to be insufficient content difference risk, and the graph admission status is determined to be pending expert verification. If the copy risk value is lower than the copy risk threshold, the same-origin copy status is not valid, and both the action coverage status and the content difference status are sufficient, the relationship risk type is determined to be regular low risk, and the graph admission status is determined to be automatic admission. When multiple risk conditions are met simultaneously, the relationship risk type is recorded as the primary risk cause in the order of source copying risk, insufficient action coverage risk, and insufficient content difference risk, and the remaining risk conditions are written as secondary risk causes into the access record or expert verification task package.

[0075] For candidate skill relationships with recorded relationship suppression information, the data processing server reads the relationship attribution information and evidence lineage key from the relationship suppression table after the above determination. The final determination order of the graph admission status is as follows: first, determine whether the relationship suppression table is hit; if it is, it is determined to be temporarily suspended; if the relationship suppression table is not hit, then determine whether it is pending expert verification based on the copy risk value, relationship risk type, same-source copy status, action coverage status, and content difference status; if the conditions for temporary admission and pending expert verification are not met, but the conditions for automatic admission are met, it is determined to be automatically admitted. If the relationship attribution information of the current candidate skill relationship is the same as the record in the relationship suppression table, and the evidence lineage key corresponding to the current candidate skill relationship is the same as the evidence lineage key in the suppressed record, or the lineage overlap value between the evidence lineage cluster corresponding to the current candidate skill relationship and the evidence lineage cluster corresponding to the suppressed record reaches the preset lineage overlap limit, then the data processing server determines the graph admission status of the candidate skill relationship as temporarily suspended and retains the original relationship risk type as the risk reason.

[0076] Through the risk assessment module, the data processing server completes the lineage copying and relationship risk assessment of candidate skill relationships, generating information on source copying status, action coverage status, content difference status, source duplication rate, flow determination rate, copying risk value, relationship risk type, and graph admission status. These results maintain a correspondence with candidate skill relationships, candidate relationship evidence chain sets, evidence lineage clusters, and evidence lineage overlap matrices, and can be directly used by the graph evolution module for automatic admission, expert verification task package assembly, or deferred admission processing.

[0077] The graph evolution module controls the graph evolution process of candidate skill relationships based on the relationship risk type and graph admission status. It performs graph writing processing on candidate skill relationships that are automatically admitted, performs expert verification task package assembly processing on candidate skill relationships that are pending expert verification, and performs relationship suppression processing on candidate skill relationships that are temporarily deferred to obtain the skill graph evolution result. The following sub-steps S401 to S404 are executed by the graph evolution module.

[0078] S401 Automatic Admission Criteria Determination and Candidate Skill Relationship Writing: The data processing server takes the relationship risk type, graph access status, copy risk value, same-source copy status, action coverage status, and content difference status output by the risk assessment module as its starting point. Simultaneously, it reads the candidate skill relationships and candidate relationship evidence chain sets generated by the evidence generation module, as well as the evidence lineage clusters and evidence lineage overlap matrices generated by the lineage analysis module. The data processing server first reads the graph access status item by item according to the candidate skill relationships. When the graph access status is automatic access, it continues to verify whether the candidate skill relationship meets the automatic access conditions. This verification is based on the judgment results already generated by the risk assessment module and does not recalculate the source duplication rate, flow rate, and copy risk value.

[0079] The execution path for automatic admission is as follows: The data processing server reads the relationship risk type of the candidate skill relationship. If the relationship risk type is low risk, it continues to read the replication risk value and compares it with the replication risk threshold in the model configuration. When the replication risk value is lower than the replication risk threshold, and the same-source replication status is not valid, the action coverage status is sufficient, the content difference status is sufficient, and there is no relationship suppression information matching the candidate skill relationship in the relationship suppression table, the candidate skill relationship is determined to meet the automatic admission conditions. Among them, matching relationship suppression information means that the relationship attribution information of the current candidate skill relationship is the same as the relationship attribution information in the relationship suppression table, and the evidence lineage key corresponding to the current candidate skill relationship is the same as the evidence lineage key in the relationship suppression table, or the lineage overlap value between the evidence lineage cluster corresponding to the current candidate skill relationship and the evidence lineage cluster recorded in the relationship suppression table reaches the preset lineage overlap limit. If any check item is not met, the data processing server transfers the candidate skill relationship to the verification trigger condition judgment of S402 or postpones the admission path, and does not perform the graph writing process.

[0080] For candidate skill relationships that meet the automatic admission criteria, the data processing server performs graph writing. During writing, a relationship ontology is first generated based on the objects at both ends of the normalized relationship and the relationship category in the candidate skill relationship. Then, the evidence index of the skill evidence unit is read from the candidate relationship evidence chain set as the relationship source index, and the corresponding relationship risk type, graph admission status, duplication risk value, source duplication rate, and flow determination rate are written into the admission record. For example, if the candidate skill relationship between employee A and the Java skill object is determined to be of low risk by the risk assessment module, and the training record cluster and project execution record cluster constitute independent evidence sources, the data processing server writes employee A, the Java skill object, and their relationship category into the skill graph, while retaining the evidence indexes of the training records and project execution records, enabling the skill relationship to be traced back to the corresponding skill evidence unit during subsequent audits.

[0081] S402 Verification Trigger Condition Determination and Expert Verification Task Package Assembly: The data processing server uses candidate skill relationships from S401 that have not entered the automatic admission writing path, as well as the relationship risk type and graph admission status output by the risk assessment module, as the starting point for processing. The data processing server reads the graph admission status of each candidate skill relationship. When the graph admission status is pending expert verification, or when the relationship risk type is any of the following: homogeneous copy risk, insufficient action coverage risk, or insufficient content difference risk, the candidate skill relationship is determined to meet the verification trigger condition. Similarly, when the copy risk value is not less than the copy risk threshold, the homogeneous copy status is valid, the action coverage status is insufficient, or the content difference status is insufficient, the candidate skill relationship is also determined to meet the verification trigger condition. Candidate skill relationships that meet the verification trigger condition are not directly written into the skill graph.

[0082] After the verification trigger condition is met, the data processing server reads the candidate skill relationship, the skill evidence unit supporting the candidate skill relationship, and the evidence index from the evidence generation module; it reads the corresponding evidence lineage cluster and evidence lineage overlap matrix from the lineage analysis module; and it reads the source duplication rate, flow determination rate, copy risk value, relationship risk type, and graph access status from the risk judgment module. Finally, it assembles an expert verification task package according to the candidate skill relationship as the main index. In the expert verification task package, skill evidence units are grouped and arranged according to evidence lineage clusters; the lineage overlap information related to the candidate skill relationship in the evidence lineage overlap matrix is ​​converted into a displayable inter-cluster overlap description; and the evidence content is read through the evidence index and presented together with its corresponding evidence lineage cluster, allowing experts to view the evidence source, copy path, and risk judgment basis within the same task package.

[0083] For example, a candidate skill relationship between employee A and a Python skill object appears simultaneously in the employee profile, job application materials, and performance summary. However, the lineage analysis module shows that all three pieces of evidence are derived from the same project work order. The risk assessment module outputs the risk of duplicate origin and requires expert verification. Based on this, the data processing server generates an expert verification task package, presenting the lineage overlap information between the project work order cluster, employee profile cluster, and job application material cluster along with the evidence content, along with the source duplication rate, flow rate, and duplication risk value, for experts to determine whether the candidate skill relationship can be included in the skill graph as a genuine skill relationship.

[0084] S403 Expert Verification Result Reception and Spectrum Writing Processing: The data processing server uses the expert verification task package generated by S402 as its starting point. The expert verification result is a record of expert confirmation or rejection of the authenticity of candidate skill relationships based on the expert verification task package. It points to at least one candidate skill relationship and carries a conclusion of whether the verification passed or failed. The expert verification result, as a structured confirmation record received by the data processing server, does not participate in the recalculation of source duplication rate, flow stability rate, and copy risk value. The data processing server performs graph writing processing or relationship suppression processing only based on the candidate skill relationship identifier, verification conclusion, and task package identifier in the expert verification result.

[0085] After receiving the expert verification results, the data processing server first verifies whether they correspond to the candidate skill relationships, evidence lineage clusters, and evidence indexes in the expert verification task package. If the correspondence is valid, the server proceeds to either graph writing or relationship suppression processing according to the expert verification results. When the expert verification result is passed, the data processing server executes graph writing. This writing process follows the relationship ontology generation method in S401 and writes the expert verification results as part of the admission record into the skill graph. Simultaneously, it establishes an index connection between the skill evidence units, evidence lineage clusters, relevant lineage overlap information, source overlap rate, flow rate, and copy risk value in the candidate relationship evidence chain set and the already written skill relationships. If the same relationship ontology already exists in the skill graph, the data processing server does not generate a new relationship ontology but instead appends the relationship source, evidence index, and admission record for this batch. If the same relationship ontology does not exist in the skill graph, the corresponding skill relationship is newly written.

[0086] In the enterprise employee skills database implementation, after reviewing the expert verification task package for employee A and the Python skill object, the expert confirmed that although the project work order was derived multiple times, the work order content truly reflected employee A's undertaking of Python script development tasks, and the expert verification result was a pass. Based on this, the data processing server wrote the relationship between employee A and the Python skill object into the skills graph, and wrote the project work order cluster, derived material cluster, and corresponding evidence index as traceable sources into the access record. This ensures that although the skill relationship has undergone the expert verification path, it can still serve as the data foundation for subsequent talent matching and competency assessment.

[0087] S404 Relation Suppression Processing and Skill Map Evolution Result Generation: The data processing server uses the expert verification results received by S403 and the map access status output by the risk assessment module as the starting point for processing. If the expert verification result is a failure, or if the candidate skill relationship has been determined to be temporarily suspended in the risk assessment module, the data processing server performs relationship suppression processing. The relationship suppression processing first reads the relationship attribution information of the candidate skill relationship, then reads the evidence lineage key and relationship risk type corresponding to the skill evidence unit supporting the candidate skill relationship, and writes the candidate skill relationship, relationship attribution information, evidence lineage key, risk cause, and corresponding evidence index into the relationship suppression table. When there are already records with the same relationship attribution information and the same evidence lineage key in the relationship suppression table, the data processing server does not add new records, but appends the risk cause, evidence index, and processing batch identifier of this batch to the existing records; when the relationship attribution information is the same but the evidence lineage key is different, and the lineage overlap value between the corresponding evidence lineage clusters reaches the preset lineage overlap limit, the data processing server writes the current evidence lineage key as the associated lineage key into the existing record.

[0088] The conditions for hitting the relation inhibition table are: the relation attribution information of the current candidate skill relationship is the same as the relation attribution information in the relation inhibition table, and the evidence lineage key corresponding to the current candidate skill relationship is the same as the evidence lineage key in the relation inhibition table; or the lineage overlap value between the evidence lineage cluster corresponding to the current candidate skill relationship and the evidence lineage cluster recorded in the relation inhibition table reaches a preset lineage overlap limit. When the relation inhibition table is hit, the data processing server determines the graph admission status of the current candidate skill relationship as temporarily suspended admission and uses the risk reasons in the relation inhibition table as the admission blocking basis for the candidate skill relationship. For relation ontologies that have already entered the skill graph, relation inhibition processing does not delete other skill relationships that have been confirmed by experts or have been automatically admitted; it only performs write blocking on candidate skill relationships that have not passed the verification or have been temporarily suspended admission, and retains the evidence lineage key for subsequent batch matching.

[0089] After all candidate skill relationships have undergone automatic admission, expert verification, or relationship suppression processing, the data processing server generates skill graph evolution results. These results include newly added skill relationships in this batch, skill relationships confirmed by expert verification, candidate skill relationships written into the relationship suppression table, and corresponding evidence index records. Simultaneously, the final graph admission status, relationship risk type, and admission path for each candidate skill relationship are recorded. Therefore, when viewing employee skill library evolution records, enterprises can distinguish between directly written, expert-confirmed written, and suppressed relationships, and can trace back to the corresponding skill evidence unit based on the evidence index.

[0090] Through the graph evolution module, the data processing server controls the graph evolution of candidate skill relationships based on relationship risk type and graph admission status, generating automatic admission writing results, expert verification task packages, expert-verified graph writing results, relationship suppression table records, and skill graph evolution results. These results preserve the correspondence between candidate skill relationships, evidence lineage clusters, evidence lineage overlap matrices, skill evidence units, and evidence indexes, ensuring traceable admission criteria for this batch of skill graph updates.

[0091] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A system for constructing and evolving an intelligent graph of a digital employee skills database, characterized in that: The data processing server includes an evidence generation module, a kinship analysis module, a risk assessment module, and a graph evolution module. The evidence generation module acquires multi-source business data related to employee skills generated within the enterprise, performs standardization and relationship extraction processing on the multi-source business data related to employee skills, obtains skill evidence units, and forms a set of candidate skill relationships and candidate relationship evidence chains based on relationship attribution information. The bloodline analysis module performs evidence bloodline analysis on the skill evidence units in the candidate relationship evidence chain set. Based on the evidence description information and the upstream and downstream flow relationship, it determines the bloodline attribution between skill evidence units and forms evidence bloodline clusters, evidence bloodline overlap matrices and independent evidence cluster sets. The risk assessment module performs lineage copying and relationship risk assessment on candidate skill relationships based on the candidate relationship evidence chain set, evidence lineage cluster, evidence lineage overlap matrix and independent evidence cluster set, and obtains the relationship risk type and graph admission status. The graph evolution module controls the graph evolution process of candidate skill relationships based on the relationship risk type and graph admission status. For candidate skill relationships that are automatically admitted, graph writing is performed; for candidate skill relationships that are awaiting expert verification, expert verification task package assembly is performed; and for candidate skill relationships that are temporarily deferred, relationship suppression is performed, thus obtaining the skill graph evolution result. Evidence lineage analysis includes: extracting the evidence source, evidence generation process, and evidence flow path from the evidence description information of skill evidence units; constructing evidence lineage keys according to a fixed field order; and identifying upstream and downstream flow relationships based on the transmission relationship between output nodes and input nodes. The data processing server establishes a lineage merging graph with skill evidence units as nodes and merges evidence lineage clusters according to the evidence lineage keys and upstream and downstream flow relationships. The data processing server determines the bloodline overlap value based on the content inheritance relationship, structural reuse relationship and business derivation relationship between different evidence bloodline clusters, and writes the bloodline overlap value into the evidence bloodline overlap matrix; when determining the set of independent evidence clusters, it identifies related evidence cluster pairs according to the evidence bloodline overlap matrix and the preset bloodline overlap boundary, and includes evidence bloodline clusters that do not form related evidence cluster pairs with the selected evidence bloodline clusters into the set of independent evidence clusters. The bloodline copying judgment includes: mapping the skill evidence units corresponding to the candidate skill relationship to the evidence bloodline clusters to form a distribution state, and determining the homologous copying state based on the maximum cluster proportion and the set of independent evidence clusters; the data processing server also forms an action coverage state based on the business action attributes of the skill evidence units, and forms a content difference state based on the evidence content of the skill evidence units, the evidence bloodline clusters and the evidence bloodline overlap matrix. The data processing server counts the number of landing points of the skill evidence units corresponding to the candidate skill relationship in each evidence lineage cluster, and determines the evidence lineage cluster corresponding to the maximum number of landing points. When the proportion of the maximum number of landing points to the total number of skill evidence units corresponding to the candidate skill relationship reaches the homogeneous set boundary, the homogeneous replication state is established. When the skill evidence units supporting the candidate skill relationship are distributed in multiple evidence lineage clusters, but the number of evidence lineage clusters that can enter the independent evidence cluster set is less than 2, the homogeneous replication state is established. When the skill evidence units are distributed in multiple evidence lineage clusters, and there are at least 2 evidence lineage clusters with mutually independent sources that can support the candidate skill relationship, the homogeneous replication state is not established. The data processing server calculates the number of landing points of the statistical skill evidence units in each evidence lineage cluster based on the distribution status, obtains the cluster proportion vector after frequency normalization, and obtains the source overlap rate by combining it with the evidence lineage overlap matrix; at the same time, it constructs the action transfer counting matrix and action transfer matrix based on the business action attributes, evidence flow path and upstream and downstream flow correlation, and obtains the flow rate by the concentration of subsequent flow of business action attributes.

2. The intelligent graph construction and evolution system for digital employee skill base according to claim 1, characterized in that, Employee skill-related multi-source business data is obtained through the enterprise's internal database interface, log interface, file parsing interface, and message synchronization interface. The data processing server organizes the original business records according to the access batch, business occurrence time, and source location, and performs normalization processing on employee objects, skill objects, job titles, project names, course names, dates, text encoding, and field formats to generate a sequence of original evidence records, which serves as the input for relation extraction processing.

3. The intelligent graph construction and evolution system for digital employee skill base according to claim 2, characterized in that, Skill evidence units are generated from structured fields, log entries, document paragraphs, and skill-related fact fragments extracted from text. Each skill evidence unit carries evidence description information, relationship attribution information, and evidence index. The data processing server constructs relationship attribution information based on the objects at both ends of the normalized relationship, the relationship category, the business context identifier, and the extraction result identifier, and organizes the skill evidence units that support the same candidate skill relationship into a candidate relationship evidence chain set.

4. The intelligent graph construction and evolution system for digital employee skill base according to claim 1, characterized in that, The data processing server combines the source duplication rate and the flow rate into a two-dimensional feature vector and inputs it into the Bayesian Gaussian mixture discriminant model. The Bayesian Gaussian mixture discriminant model consists of homologous copy risk components and non-homologous copy components, and obtains prior probabilities, mean vectors, and covariance matrices based on historical expert verification samples. The data processing server outputs the copy risk value and combines homologous copy status, action coverage status, and content difference status to obtain the relationship risk type and graph admission status.

5. The intelligent graph construction and evolution system for digital employee skill base according to claim 4, characterized in that, When the skill graph admission status is automatic admission and the candidate skill relationship meets the automatic admission conditions, the data processing server writes the candidate skill relationship into the skill graph and writes the relationship source, evidence index and admission record. When the map admission status is pending expert verification and the candidate skill relationship meets the verification triggering conditions, the data processing server reads the candidate skill relationship, skill evidence unit, evidence lineage cluster and evidence lineage overlap matrix, and assembles the expert verification task package.

6. The intelligent graph construction and evolution system for digital employee skill base according to claim 5, characterized in that, The data processing server receives the expert verification results corresponding to the expert verification task package; when the expert verification result is a pass, the server performs graph writing processing on the candidate skill relationship and establishes connections with the skill evidence unit, evidence lineage cluster, and evidence index; when the expert verification result is a fail, the data processing server writes the candidate skill relationship, relationship attribution information, evidence lineage key, risk cause, and evidence index into the relationship suppression table. When a candidate skill relationship is determined to be temporarily deferred, the data processing server writes the candidate skill relationship, relationship attribution information, evidence lineage key, risk reason, and evidence index into the relationship suppression table.

Images