GNN-based talent relationship evaluation method and system

Abstract

The invention relates to the technical field of talent relationship evaluation, in particular to a GNN-based talent relationship evaluation method and system, and the method comprises the following steps: based on resume event data, classifying and accumulating influence factors, analyzing task and feedback response intervals, calculating multi-source tag differences, comparing the consistency, preferentially selecting accurate tags, and writing the tags into nodes; and adjusting and covering node label contents according to consistency, and generating a label updating result. According to the method, real-time comparison and optimization of multi-source label contents are realized through classification and aggregation of resume data, promotion of dynamic evolution of node information and combination of behavior characteristics of interaction response among nodes, and node attributes and label expressions are adaptively adjusted based on label content distribution and consistency level, so that the real-time comparison and optimization of the multi-source label contents are realized. Cooperative reflection of structure change, behavior transaction and label content in a complex network is supported, and continuous and dynamic discrimination and quantification of data can be kept in a time sequence propulsion and multi-dimensional interaction scene.

Description

Technical Field

[0001] This application relates to the field of talent relationship evaluation technology, and in particular to a talent relationship evaluation method and system based on GNN. Background Technology

[0002] Talent relationship evaluation is a methodology for identifying and assessing interactions between talents within or across organizations. It includes building social networks among talents, analyzing talent influence, identifying and evaluating collaborative relationships, identifying key nodes, and quantifying talent mobility and relationship structures. It systematically utilizes information processing and network analysis techniques to achieve multi-dimensional insights and quantitative research into talent relationships. Traditional talent relationship evaluation relies on methods such as manual surveys, expert scoring, and social network analysis. It involves collecting and organizing qualitative or semi-quantitative data on work interactions, communication frequency, and project collaborations among personnel, and then using statistical analysis or simple network indicator methods to evaluate talent relationships.

[0003] Existing technologies rely on qualitative evaluation or limited network analysis in practical applications. Data is mostly collected statically and entered intermittently. When faced with changes in the structure of relationship networks or the complexity of personnel interactions, it is difficult to achieve dynamic updates of node attributes, and it is difficult to compare tag content from multiple sources. Relationship assessment is prone to missing temporal features and structural changes. The depth of data fusion is limited, resulting in evaluation results that cannot dynamically reflect the multidimensional relationship status within the organization. In real-world scenarios, it is difficult to meet the timeliness and quantification requirements, and the output is often limited to surface descriptions due to static information and simple content. Summary of the Invention

[0004] To address the aforementioned technical issues, this application proposes a talent relationship evaluation method and system based on GNN.

[0005] The technical solution adopted in this application is: a talent relationship evaluation method based on GNN, which constructs a talent relationship network through GNN, where the nodes of the network are people, that is, each evaluated object in the talent relationship network, and the interactions between people are the edges of the network. The implementation steps of the talent relationship evaluation method include:

[0006] S1: Acquire and analyze the resume event data for each talent node. The resume event data includes job promotion, project assignment and cross-departmental transfer resumes. Each resume event is classified into the corresponding influencing factor category. The events are arranged in chronological order. The sum of each influencing factor after classification is calculated. Nodes whose sum meets the judgment criteria are identified. Key event node information is established.

[0007] Talent nodes refer to network nodes in the talent relationship network that represent each evaluated object, usually corresponding to a specific person (such as employee A).

[0008] Resume event data: refers to significant historical events related to the talent, including job promotions, project assignments, and interdepartmental transfers (e.g., promotion, job transfer, interdepartmental involvement, project participation).

[0009] Acquisition and analysis: This involves extracting event records for each individual from various data sources such as personnel files, job change records, project allocation tables, and workflow platforms, and then classifying and organizing each event by type and time.

[0010] Its function is to establish a "resume timeline" for each talent node, and subsequent operations such as impact assessment, key node identification, and GNN modeling can be performed based on these events.

[0011] S2: Based on key event node information, compare the task acceptance time and information feedback time between nodes, divide talent nodes into multiple groups according to identity type, calculate the task response interval and information feedback interval for each group of nodes, analyze the response performance of each node in the group, determine whether the task response interval and information feedback interval are different from the general distribution, classify and organize the behavioral influence factors of the edges, and obtain the node response feature set.

[0012] The identity type refers to the classification information such as job category, job level code, and department affiliation (e.g., "management position", "execution position", "technical position").

[0013] Divide into multiple groups: This refers to grouping all talent nodes according to their identity type, with members in each group having the same type of identity label.

[0014] Analyze each group of nodes: mainly during task response and information feedback analysis, statistically analyze and compare the behavioral characteristics of different identity groups;

[0015] Its function is to analyze the differences in behavioral characteristics of different identity types in terms of task response, collaboration efficiency, etc., which is conducive to accurate profiling and hierarchical management.

[0016] S3: Obtain multi-dimensional tag data of the same talent node from different data sources. Based on the node response feature set, calculate the content differences of multi-source tags in collaboration ability, relationship strength and scope of influence. For the different data source tags of the same node in the same time period, compare the dimension content one by one and compare it with the consistency benchmark to identify the different data source combinations and tag dimension combinations and obtain the tag consistency distribution parameters.

[0017] The different data sources include human resources systems, collaboration platforms, and third-party evaluation systems.

[0018] Same talent node: refers to the same person.

[0019] Multi-dimensional tag data: refers to tags, scores or comments on multiple evaluation dimensions such as a person's collaborative ability, relationship strength, and scope of influence.

[0020] Acquisition: Aggregate all tags from different platforms, systems, or evaluation subjects, and compare and merge multi-source tag data of the same person within the same time period.

[0021] Function: To ensure the diversity and comprehensiveness of tag sources, break the bias brought about by a single information source, and make the final node tags more authentic and representative.

[0022] Differences in data source combinations: Under the same evaluation dimension, if significant differences are found in the tag content of the same person from different data sources (such as personnel, collaboration, external, etc.), these data sources are listed as "differences in data source combinations".

[0023] Tag dimension combination: This refers to the specific dimensions in which differences occur among multi-dimensional tags (such as collaboration ability, relationship strength, scope of influence, etc.).

[0024] Identification process: This refers to comparing the label content of the same node, the same time period, and the same dimension item by item with the consistency benchmark. If the difference exceeds the set threshold, it can be judged as a "difference combination".

[0025] Its function is to provide a basis for grouping in subsequent tag selection, weight adjustment and multi-source fusion, and to selectively filter and merge the most credible tag content.

[0026] S4: Based on the label consistency distribution parameters, determine the historical accuracy parameters of the associated label data source. For the different combinations of label dimensions, prioritize them according to the accuracy parameters, select the label content with the highest accuracy parameters for each dimension, and execute the content writing of the node labels to obtain the preferred label content library.

[0027] Furthermore, the influence factor category in step S1 refers to classifying the influence relationship strength for each type of resume event based on different types such as job promotion, project assignment, and cross-departmental transfer, which facilitates subsequent aggregation analysis.

[0028] The specific steps for obtaining key event node information are as follows:

[0029] S111: Based on the resume event data of talent nodes, analyze the job promotion, project assignment and cross-departmental transfer resume events of talent nodes, assign resume events to corresponding categories according to the resume events, and obtain the resume event ranking set according to the order of occurrence of the events;

[0030] S112: Based on the resume event ranking set, calculate the cumulative impact of various resume events on talent nodes, compare the distribution of different event types under the same node, adjust the ranking and classification criteria of event categories, and obtain the impact factor distribution data.

[0031] S113: Based on the distribution data of impact factors, determine the performance of talent nodes under the screening conditions, analyze the node's identification attributes, number of events participated in and event aggregation tags, optimize the node information content, and establish key event node information.

[0032] Furthermore, the specific steps for obtaining the node response feature set are as follows:

[0033] S211: Based on key event node information, compare the task acceptance time and information feedback time of each node, calculate the task response interval and information feedback interval for each group of nodes, filter out nodes with different response performance, and obtain node interaction time delay data.

[0034] S212: Based on the node interaction time delay data, analyze the impact of node identity type on task response and feedback interval, determine the distribution of nodes in terms of interaction delay, identify nodes with abnormal response behavior, and obtain node response offset comparison data.

[0035] S213: Based on the node response offset comparison data, calculate the offset magnitude between the task response interval and the information feedback interval, analyze the frequency of node event participation and abnormal response behavior, and use the following formula to filter the node response characteristic interval. The node response feature set is obtained;

[0036] ;

[0037] in, Indicates the first Task response interval for each node pair This represents the average response interval of all node network tasks. Indicates the first The cumulative amount of event participation frequency for each node pair. This indicates the total number of node pairs under the current identity type. This indicates the number of events involved in a single node pair under the current identity type. This indicates the number of abnormal response behaviors under the current identity type. This indicates the proportion of the current identity type.

[0038] Furthermore, the specific steps for obtaining the label consistency distribution parameters are as follows:

[0039] S311: Based on the node response feature set, analyze the node interaction frequency, response category label and relationship strength classification, obtain the tag content of collaboration ability, relationship strength and influence scope in each data source tag, identify the description category and tag characteristics of the tag content in the same dimension, and obtain the tag content classification results;

[0040] S312: Based on the tag content classification results, analyze the tag content in each data source tag in the dimensions of collaboration ability, relationship strength and scope of influence, compare their similarity and deviation, refine the content differences of tags, and obtain tag difference measurement results;

[0041] S313: Based on the label difference measurement results, filter related label combinations, analyze the differences between multi-source labels to determine the distribution characteristics of the label data, and calculate the label consistency difference degree using the following formula. Based on this degree of difference and the distribution characteristics of the label data, a multidimensional comparison is performed to obtain the label consistency distribution parameters;

[0042] ;

[0043] in, This represents the degree of similarity between tag content, indicating the level of similarity between tag content generated from different data sources within the same dimension. This represents the magnitude of the label's numerical deviation, used to indicate the degree of numerical discrepancy between label content and other data within the same dimension. This represents the frequency of evaluation distribution under a specific rating level, indicating the frequency of evaluation distribution for that label within that dimension, and reflecting the concentration of label evaluations. Represents the density of label data distribution, indicating the label The data distribution density along this dimension reflects the concentration of the label data in this dimension. Represents the range span, indicating the label. The span of the data distribution range is used to measure the breadth of the label data within this dimension. This refers to the number of multi-source tag data sources.

[0044] Furthermore, the specific steps for obtaining the preferred tag content library are as follows:

[0045] S411: Based on the label consistency distribution parameters, analyze the key differences in the combination of label dimensions, classify the collaboration ability, relationship strength and scope of influence, filter the corresponding data sources, map the correspondence between each dimension and the data source, and obtain the label source mapping structure.

[0046] S412: Based on the tag source mapping structure, compare the tag content of each data source under each dimension combination, calculate the frequency, content offset and feedback interval of the content provided by each data source, optimize the distribution among participating items, and sort the tag content using the following formula to obtain the optimal tag sorting structure.

[0047] ;

[0048] in, Indicates the first Tag content ranking metrics under a combination of tag dimensions Indicates the first In the combination of the _th label dimension, the _th Each data source provides the frequency of occurrence of the tagged content. Indicates the first In the combination of the _th label dimension, the _th Content offset of each data source Indicates the first The mean of content offset across all data sources under a combination of tag dimensions. Indicates the first In the combination of the _th label dimension, the _th Feedback interval of each data source Indicates the first Feedback expectations under a combination of label dimensions Indicates the total number of data sources;

[0049] S413: Based on the tag optimization and sorting structure, filter the top-ranked tag content, determine the optimal content under each dimension, integrate the description and source information, write the node tag content, and obtain the optimized tag content library.

[0050] Furthermore, it also includes S5: Based on the preferred tag content library, adjust the content of node tags in terms of collaboration ability, relationship strength and influence scope, analyze the consistency between the various contents of the node tag temporary data and the preferred tag content, identify the dimensions that need to be updated, and cover the tag content item by item to obtain the node tag update results.

[0051] Furthermore, the specific steps for obtaining the node label update results are as follows:

[0052] S511: Based on the preferred tag content library, analyze the tags of each dimension in the temporary data of node tags, compare the content descriptions of collaboration ability, relationship strength and influence scope, determine whether there are differences in the content of each dimension tag, filter out the tag dimensions with differences, and obtain a list of dimension deviation indicators.

[0053] S512: Based on the dimensional deviation identifier list, analyze the source of the tags and the historical content of the temporary data of the node tags, determine the difference between the source content and the historical content, identify the tag parameters that need to be covered, and cover them item by item to obtain the node tag update results.

[0054] A talent relationship evaluation system based on Generative Neural Networks (GNNs), the system being used to implement the GNN-based talent relationship evaluation method, the system comprising:

[0055] Resume Attribution Module: Based on resume event data of talent nodes, analyze each job promotion, project assignment and cross-departmental transfer resume, classify each event type into the corresponding influencing factor category, complete the event arrangement according to the time sequence of resume events, calculate the cumulative sum of each influencing factor after classification, identify nodes whose cumulative sum meets the judgment criteria, and establish key event node information.

[0056] Response Feature Module: Based on key event node information, compare the task acceptance time and information feedback time between nodes. For each group of nodes, calculate the task response interval and information feedback interval, analyze the response performance under each node identity type, determine whether the task response and information feedback interval are different from the general distribution, classify and organize the behavioral influence factors of the edges, and obtain the node response feature set.

[0057] Tag Difference Module: Based on the node response feature set, calculate the content differences of multi-source tags in terms of collaboration ability, relationship strength and scope of influence. For the difference data source tags of the same node in the same time period, compare the dimensional content one by one, compare it with the consistency benchmark, identify the key data source combination of differences, and obtain the tag consistency distribution parameters.

[0058] Tag selection module: Based on tag consistency distribution parameters, it determines the historical accuracy parameters of associated tag data sources, prioritizes the tag dimension combinations with key differences according to accuracy parameters, selects the tag content with the highest accuracy parameters for each dimension, and executes the writing of node tag content to obtain the selected tag content library;

[0059] Tag update module: Based on the preferred tag content library, adjust the content of node tags in terms of collaboration ability, relationship strength and influence scope, analyze the consistency between the content of each item of the node tag temporary data and the preferred tag content, identify the dimensions that need to be updated, and cover the tag content item by item to obtain the node tag update results.

[0060] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0061] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0062] The beneficial effects of this application compared to the prior art are as follows: This application promotes the dynamic evolution of node information by classifying and aggregating resume data, and realizes real-time comparison and optimization of multi-source tag content by combining the behavioral characteristics of interaction and response between nodes. Based on the distribution and consistency level of tag content, it adaptively adjusts node attributes and tag expression, supports the coordinated reflection of structural changes, behavioral changes and tag content in complex networks, and enables continuous dynamic discrimination and quantification of data in time-series progression and multi-dimensional interaction scenarios. The talent relationship network presents a brand-new evaluation characteristic of accurate tag iteration, timely response to attribute changes and rich data fusion levels. Attached Figure Description

[0063] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0064] Figure 1 This is a flowchart of the main steps of the method in this application;

[0065] Figure 2 This is a flowchart illustrating the process of obtaining key event node information in this application;

[0066] Figure 3 This is a flowchart illustrating the process of obtaining the node response feature set in this application;

[0067] Figure 4 This is a flowchart illustrating the process of obtaining the label consistency distribution parameters in this application.

[0068] Figure 5 This is a flowchart illustrating the process of obtaining the preferred tag content library in this application;

[0069] Figure 6 This is a flowchart illustrating the process of obtaining the node label update results in this application.

[0070] Figure 7 This application provides a schematic diagram of the internal structure of a computer device. Detailed Implementation

[0071] like Figures 1 to 7As shown, this application provides a talent relationship evaluation method based on GNN. It treats each person as a "node" in the network, and the interactions between people (such as working on projects together or cooperating on tasks) as "connections" (also called "edges"). The GNN (Graph Neural Network) inputs each person's resume, performance, various tags, interaction frequency, and other data as "features" of the nodes / edges. After analysis by the GNN, it can be seen who is the most critical in the relationship network, who has the strongest collaborative ability and influence, who is a marginal person, and which evaluation tags best represent the true ability. Because the tags and relationships are dynamically updated, the output of the GNN is also continuously updated to achieve intelligent evaluation of talent relationships.

[0072] The main idea is to utilize all of each person's experiences, records of interactions with others, and various evaluation tags, filtering, comparing, and updating them step by step, and finally handing them over to a graph neural network for 'connected analysis', so that everyone in the talent network can be accurately evaluated from multiple angles, in a comprehensive and dynamic manner, to find the most valuable talent and the best cooperative relationship.

[0073] Based on this, the specific implementation steps of this application are as follows:

[0074] S1: Based on the resume event data of talent nodes, analyze each job promotion, project assignment and cross-departmental transfer resume, classify each event type into the corresponding impact factor category, arrange the events according to the time sequence of the resume events, calculate the sum of each impact factor after classification, identify the nodes whose sum meets the judgment criteria, and establish key event node information.

[0075] S2: Based on key event node information, compare the task acceptance time and information feedback time between nodes. For each group of nodes, calculate the task response interval and information feedback interval, analyze the response performance under each node identity type, determine whether the task response interval and information feedback interval are different from the general distribution, classify and organize the behavior influence factors of the edges, and obtain the node response feature set.

[0076] S3: Based on the node response feature set, calculate the content differences of multi-source tags in terms of collaboration ability, relationship strength and scope of influence. For the different data source tags of the same node in the same time period, compare the dimensional content one by one, compare it with the consistency benchmark, identify the key data source combination of differences, and obtain the tag consistency distribution parameters.

[0077] S4: Based on the label consistency distribution parameters, determine the historical accuracy parameters of the associated label data source. For the key difference in the label dimension combination, prioritize according to the accuracy parameters, select the label content with the highest accuracy parameters for each dimension, and execute the content writing of the node labels to obtain the preferred label content library.

[0078] S5: Based on the preferred tag content library, adjust the content of node tags in terms of collaboration ability, relationship strength and influence scope, analyze the consistency between the content of each item of the node tag temporary data and the preferred tag content, identify the dimensions that need to be updated, and cover the tag content item by item to obtain the node tag update results.

[0079] The key event node information includes node identification attributes, number of times the event was participated, and event aggregation tags; the node response feature set includes node interaction frequency, response category label, and relationship strength grading; the tag consistency distribution parameters include tag data distribution curve, tag similarity interval, and tag conflict indicator; the preferred tag content library includes tag preference index, preferred tag source, and preferred content list; and the node tag update results include tag change log, content revision record, and update synchronization mark.

[0080] The data in step S1 comes from personnel files, job change records (promotions, transfers, departmental personnel changes, etc.), project allocation tables, and workflow platforms. The data serves to show "who has experienced what major events," "how many times they have been promoted," "which teams they have led," and "how many times they have worked across departments." For example, employee A has been promoted three times, transferred to two departments, and been assigned to five projects; these are A's "major events on their resume." The system will organize and record all of A's events chronologically, directly identifying who has had more promotions and who has been more frequently promoted, and labeling them as "key talent."

[0081] This step meticulously categorizes, accumulates, and sorts each talent's various resume events (such as promotions, project assignments, and interdepartmental transfers) by category and time sequence, selecting "key nodes" with significant influence. Its essential function is to create a dynamic and time-continuous "resume profile" for each individual, providing a foundation for the dynamic changes in subsequent node characteristics.

[0082] In step S1, a talent node refers to a network node in the talent relationship network that represents each evaluated individual. Node attributes record information such as their personal experience and work records. The influence factor category refers to classifying the influence strength of each type of resume event based on different types of resume events such as job promotion, project assignment, and cross-departmental transfer, which facilitates subsequent aggregation analysis. The completed event arrangement refers to sorting all events of each talent according to the chronological order of their resume events, so that the influence analysis has a temporal causal relationship. The cumulative sum refers to adding the values ​​of the influence factors of each type of event under the same talent node after classification and sorting, to obtain the total influence of that node on that type of event. Nodes that meet the judgment criteria are those whose cumulative sum exceeds the set reference standard or threshold. These nodes are considered to have key influence on that type of event and are selected for the next step of analysis.

[0083] The data in step S2 comes from the task assignment and completion time of the OA system and the message response records of the project collaboration tool. The purpose of the data is to determine how long it takes for a person to respond to a task, how long it takes to provide feedback after completing a task, how smoothly they interact with others, and whether anyone is particularly slow or fast. For example, some people respond immediately when a task comes in, while others take their time to respond. The system will collect and record statistics to calculate whether each person can "cooperate" with others, that is, to see who is an "active participant" and who is a "procrastinator".

[0084] This step performs a detailed analysis of the timing and frequency of task acceptance and feedback between nodes, distinguishing between response speed, cooperation tightness, and abnormal interaction patterns, and generating features that reflect the dynamic strength of the relationship. Its essential function is to make the "interaction relationship" between nodes quantifiable and automatically adjust as the behavior evolves, avoiding the rigidity of static evaluation.

[0085] In step S2, task acceptance time refers to the time or duration from task posting to task acceptance when nodes interact in the talent network; information feedback time refers to the time interval consumed by a node to provide feedback to the task initiator after completing the task and accepting the instruction; response performance refers to the characteristics such as reaction speed and cooperation shown by nodes during the interaction process (such as task acceptance and feedback); whether it differs from the general distribution refers to comparing the statistical distribution of the actual task response and feedback time between nodes with the overall response and feedback time distribution in the network to determine whether certain nodes or relationships have special response behaviors; edge behavior influence factor refers to the quantitative influence parameter assigned to the relationship (i.e., edge) connecting two nodes based on characteristics such as response behavior, used to express the behavioral characteristics of node interaction.

[0086] The data sources for step S3 include scores from the HR system (such as performance evaluation scores and competency tags), evaluations from project collaboration platforms, and external or third-party evaluations. The purpose of this data is to compare these various "evaluation tags": for the same person, scores and tags from different systems may differ, allowing us to identify areas of disagreement. For example, regarding "strong teamwork," collaboration platforms might say you are "quick-witted," while external evaluations might say you have "strong independent ability"—the same person receiving different assessments from three different systems. The system will compare these tags, analyzing them one by one, to determine if the evaluations are accurate and if there are any discrepancies.

[0087] This step collects tags from multiple sources, including personnel, collaboration, and external systems, and compares the tag content of the same node at the same time period dimension by dimension to calculate consistency and differences. Its essential function is to break the limitations of a single information source and achieve the richness and accuracy of tag evaluation through comparison and aggregation, ensuring that node attributes are real, comprehensive, and dynamic.

[0088] In step S3, multi-source tags refer to multi-dimensional tag data on the same talent node's abilities, relationships, and influence from different data sources (such as HR systems, collaboration platforms, third-party evaluations, etc.); content differences refer to the differences in content descriptions or evaluation results of tags on the same node and the same dimension (such as collaboration ability, relationship strength, and scope of influence) from different data sources; same node and same time period refers to multi-source tags collected within a specific time range for the same talent node, ensuring timeliness and comparability; comparison dimension content refers to comparing the content of each evaluation dimension (collaboration ability, relationship strength, scope of influence, etc.) in the tag data item by item; consistency benchmark refers to the reference standard used to measure whether the tag content is consistent, such as content similarity, matching degree, etc., used to determine tag consistency; the key difference data source combination refers to the group of data sources where, after tag content comparison, the content difference between certain data sources in a certain dimension is greater than the consistency benchmark, and these are the objects of subsequent screening and correction.

[0089] The data source for step S4 is all the tags and their sources captured in the previous step, along with historical statistical analysis, such as which system previously provided the most accurate evaluations and which evaluations were frequently biased. The data's purpose is to score and rank tags from different sources, prioritizing historically accurate tags and selecting the more authoritative ones when there are disagreements. If multiple evaluations are all reliable, they can be merged. For example, if the scores for collaboration skills have consistently been accurate, the most accurate result should be used this time. If HR and a project manager have similar opinions, their statements can be combined into a single sentence for a more comprehensive evaluation.

[0090] This step is based on the consistency of multi-source tags and the historical performance of the data source. It optimizes and sorts the tags of each node in each dimension, selects the most reliable and representative content, and writes it in real time. Its essential function is to keep the node tags optimal at all times, so as to sensitively reflect changes in the data source and the evolution of relationships, and support multi-tag fusion and adaptive adjustment in complex scenarios.

[0091] In step S4, the historical accuracy parameter refers to the statistical indicators such as accuracy, authority, and reliability of each data source tag in historical evaluations, which are used for subsequent priority ranking; the tag dimension combination refers to the set composed of multiple tag dimensions (collaboration ability, relationship strength, scope of influence) and their corresponding data sources, which are used to filter the optimal tag source; the tag content refers to the specific evaluation description or score on each dimension such as collaboration ability, relationship strength, and scope of influence, which is the substantive content of the node tag parameter; the content writing refers to inputting or overwriting the priority-selected tag content into the node tag database according to the standard format, in preparation for subsequent evaluation and analysis.

[0092] The data source in step S5 is the newly selected tags, the original old tags of the nodes (talents), and update records. The purpose of this data is to replace old tags with the latest and most reliable ones, ensuring that each talent's tags are the freshest and most accurate; and to guarantee that the tags "keep up with the current talent market," preventing information lag from affecting decision-making. For example, if A's performance has improved significantly in the past two years, and people's evaluation of A has changed, then A's tags will be automatically corrected and updated in a timely manner, and the system will record each change.

[0093] This step involves continuously monitoring the differences between node tag content and historical temporary tags, dynamically overriding each dimension identified as changing or conflicting, and recording change logs. Its essential function is to ensure that the evaluation tags of nodes continuously evolve with changes in network structure and interaction status, realizing "live data" of talent relationship networks within the organization, rather than "dead tags".

[0094] In step S5, the temporary node label data refers to the historical labels, pending labels, or conflicting label data temporarily stored under the node object during the multi-round label judgment and screening process, awaiting final consistency judgment and update; the preferred label content refers to the label content given by the most authoritative and reliable data source in the current node and current dimension, which is determined after the aforementioned sorting and comparison; the dimension for parameter update refers to those dimensions that need to be covered or replaced by preferred labels in the various evaluation dimensions of node labels (collaboration ability, relationship strength, scope of influence, etc.) through comparison.

[0095] In step S1, by summarizing and aggregating the event features of talent resumes, rich basic features are provided for the GNN to evaluate talent relationships. In step S2, the interaction features between talents are used as edge attributes of the GNN by utilizing the task response and information feedback behaviors between nodes to characterize the strength and dynamics of talent relationships. In step S3, node features are dynamically improved based on the differences in multi-source label content, providing the GNN with multi-dimensional and realistic input data that reflects talent relationships. In step S4, node label expressions are optimized with high-accuracy labels, providing a reliable basis for the GNN's evaluation tasks such as talent classification and ability recognition. In step S5, by dynamically correcting node labels, a data foundation is provided for the GNN to continuously output timely and consistent talent relationship evaluation results.

[0096] Please see Figure 2 The specific steps for obtaining key event node information are as follows:

[0097] S111: Based on the resume event data of talent nodes, analyze the job promotion, project assignment and cross-departmental transfer resume events of talent nodes, assign them to the corresponding categories according to the event type, and obtain the resume event ranking set by combining the order of the events.

[0098] First, resume data containing timestamps, job titles, projects, and departments needs to be extracted from the talent database. For each record, its event type needs to be analyzed to determine if it involves changes in job title, additions to project participation records, or changes to the department field. These are then categorized into three event sets: job promotion, project assignment, and interdepartmental transfer. After categorization, for the same talent node, the events are sorted in ascending order according to their occurrence time in the resume, forming a resume event sorting set. This sorting must retain the event type label and occurrence time. For example, if a talent... I first participated in project code A1 in March 2021, was promoted from junior to intermediate position in January 2022, and transferred from department D1 to department D2 in June 2022. The sequence of events is project assignment, job promotion, and interdepartmental transfer. This sequencing process must ensure the complete chronological order of events. If multiple similar events occur within a similar time period, they need to be re-compared and sorted according to the specific timestamp precision. After sequencing, the sequence set corresponding to each node is numbered and identified to form a complete talent resume timeline structure, thereby providing a temporal logical basis for subsequent event aggregation and behavior identification.

[0099] S112: Based on the resume event ranking set, calculate the cumulative impact of various resume events on talent nodes, compare the distribution of different event types under the same node, adjust the ranking and classification criteria of event categories, and obtain the impact factor distribution data.

[0100] The frequency of each type of event in a node is statistically analyzed. Job promotion, project assignment, and interdepartmental transfer are initialized as three statistical dimensions, and each type of event is counted separately. For example, if a talent node has 3 job promotions, 4 project assignments, and 1 interdepartmental transfer, then a preliminary total impact value is assigned to these three types of events. For instance, the cumulative value for job promotion is 3, for project assignment is 4, and for interdepartmental transfer is 1, resulting in a preliminary impact structure. This structure is then normalized. By comparing whether the proportion of job promotions exceeds the sum of project assignments and interdepartmental transfers, the dominant event type can be identified. If the proportion of a type is significantly high, it can be considered as the presence of dominant event characteristics. Then, taking nodes as units, the overall average distribution and range of variation of the three types of events in all nodes are statistically analyzed. If cross-departmental events occur with significantly low frequency in the overall nodes but frequently in some nodes, the event classification weight settings need to be manually reviewed and the corresponding score reference for a certain type of event needs to be adjusted so that the three types of events form a reasonable range distribution in the influence calculation. After the settings are completed, the influence values ​​of all nodes need to be updated to form new influence factor distribution data. This data is used to determine whether the node characteristic performance meets the key standards.

[0101] S113: Based on the distribution data of impact factors, determine the performance of talent nodes under the screening conditions, analyze the node identification attributes, number of events participated in and event aggregation tags, optimize the node information content, and establish key event node information;

[0102] For each talent node, the three impact values ​​of job promotion, project assignment, and inter-departmental transfer are read one by one. It is determined whether any value exceeds the general performance threshold for that dimension across all nodes. For example, a value above 10 for the average job promotion impact value across all nodes is considered significant. If a node's job promotion value exceeds this value, that node is identified as the object to be analyzed. The corresponding talent identification information for that node is extracted, including basic attributes such as talent number, department, job level, and years of service. Simultaneously, the frequency of occurrence of these attributes in the three dimensions is counted, recording the number of job promotions, project assignments, and inter-departmental transfers. The event frequency vector is generated, and keyword frequency analysis is performed on the event tags corresponding to the node in the event ranking set. For example, if keywords such as "supervisor", "senior" and "nomination" frequently appear in job promotion, the aggregate tag under this dimension is set as "promotion to supervisor". The tag content is based on the top items with actual frequency. Combining node identifier, number of events participated in and aggregate tag, a unified structured key event node information is output. This structured key event node information includes a node ID field, three event frequencies, three impact values ​​and three sets of tag descriptions, which serve as an important data source for subsequent graph neural network construction and tag initialization.

[0103] Please see Figure 3 The specific steps for obtaining the node response feature set are as follows:

[0104] S211: Based on key event node information, compare the task acceptance time and information feedback time of each node, calculate the task response interval and information feedback interval for each group of nodes, filter out nodes with different response performance, and obtain node interaction time delay data.

[0105] The "groups" here refer to talent nodes grouped according to their "identity type." This means that instead of analyzing all nodes at once, we first group them based on each person's identity type (such as job category, job level code, department affiliation, etc.), and then analyze the task response and feedback characteristics of nodes within each group. Each group of nodes refers to a group of talent nodes with the same identity type attribute; for example, "all R&D employees" is one group, and "all middle-level managers" is another.

[0106] Extract the task reception time and information feedback time fields involved in the actual task execution process of each node. For each task record, extract the timestamp from the task issuance time field and record the time when the task receiving node first confirms receipt as the task reception time. Then extract the timestamp of the task completion node reporting the task execution status or result to the initiating node as the information feedback time. For each pair of nodes, construct a task interaction record set indexed by the node pair. Each record contains three data items: task issuance time, reception time, and feedback time. Then perform difference processing on each record to calculate the task response interval as the time difference between the reception time and the issuance time, and calculate the information feedback interval as the time difference between the feedback time and the reception time.

[0107] "Group" refers to the results of grouping by identity type, and "node pair" is the combination of all pairs of nodes that have actually interacted with each other within each group. The analysis is progressively advanced from "within group - node pair - task interaction". Statistical results can be regressed to individual nodes, or the response differences between groups and between node pairs can be compared. The process involves first grouping (each group contains nodes), then enumerating all node pairs within each group, and finally analyzing the interaction behavior data.

[0108] For example, if a task is issued at 9:00 AM on July 1, 2023, received at 11:00 AM on July 1, and responded at 10:00 AM on July 2, then the task response interval is 2 hours and the information feedback interval is 23 hours. After repeating this process to process all the interaction data between nodes, two time delay indicators are obtained for each group of nodes. Then, using the nodes as indexes, a dataset of average response interval and average feedback interval for each node in all interactions is constructed. The two indicators of each node are compared with the overall average level of all nodes to determine whether there is a delay characteristic deviation. If the task response interval of a node is higher than the mean of all nodes plus twice the standard deviation, or the information feedback interval exceeds the upper limit of the distribution range, then the node is recorded as a node with a difference in response performance, and its corresponding delay interval type (task reception delay or feedback delay) is labeled. The delay time data of all nodes that meet the conditions are output as node interaction time delay data.

[0109] S212: Based on the node interaction time delay data, analyze the impact of node identity type on task response and feedback interval, determine the distribution of nodes in terms of interaction delay, identify nodes with abnormal response behavior, and obtain node response offset comparison data.

[0110] For each node, extract the job category, job level code, department affiliation, and other identity type information fields. Divide the nodes into multiple groups according to identity type, such as management, execution, and R&D groups. Within each identity type group, summarize the average task response interval and average feedback interval values ​​of its member nodes. Then, compare the internal average of each group with the overall average response interval of all nodes to determine whether its response performance shows a systematic deviation from other groups. For example, if the average response interval of the execution group is 1 hour and the overall average is 3 hours, then this group is designated as a fast-response node group. At the same time, calculate the maximum and minimum response interval ranges in each group and extract the difference to construct the response. Based on the interval fluctuation index, compare whether the fluctuation level within the group is within the preset standard range. If the fluctuation difference of a group exceeds 10 hours, record the inconsistent behavior characteristics within the group. Further examine the feedback interval data of individual nodes within each group. If the average feedback time of a single node is more than twice the median of all nodes and the frequency of occurrence exceeds 50% of the total number of tasks of the node, it is determined that the node has abnormal response behavior. Record its node identification information, corresponding delay type, and abnormal response behavior label, construct a node response offset record table, and select the set of nodes with obvious delay performance deviations from all nodes, and record the distribution results of their identity types as node response offset comparison data.

[0111] S213: Based on the node response offset comparison data, calculate the offset magnitude between the task response interval and the information feedback interval, analyze the frequency of node event participation and abnormal response behavior, and use formula (1) to filter the node response characteristic interval. The node response feature set is obtained;

[0112] (1);

[0113] in, Indicates the first The task response interval for a single node pair (two nodes / two nodes) This represents the average response interval of all node network tasks. Indicates the first The cumulative amount of event participation frequency for each node pair. This represents the total number of node pairs under the current identity type. This indicates the number of events involved in a single node pair under the current identity type. This indicates the number of abnormal response behaviors under the current identity type. This indicates the proportion of the current identity type.

[0114] The node response feature interval quantity refers to a comprehensive characterization of the degree of deviation between the task response interval of all node pairs under the same identity type and the average task response interval of the network. Its function is to reflect the complex dynamics of node pair responses through a single indicator, which facilitates grouping, filtering and classification, and serves as a core component of the node response feature set.

[0115] Extract the set of node pairs under each identity type, and count their task response intervals in turn. Average task response rate across the entire network Frequency of event participation Total number of node pairs Number of abnormal response behaviors and the current proportion of nodes of different identity types. Furthermore, multiple heterogeneous data are normalized. For example, the normalization range is set to [0, 1]. Each parameter is linearly mapped according to its maximum and minimum values ​​before being included in the calculation. Taking a technical position as an example, the response intervals of its five nodes to tasks are 305 seconds, 240 seconds, 150 seconds, 180 seconds, and 200 seconds, respectively, with corresponding participation frequencies of 3, 5, 8, 4, and 6. After normalization, they are as follows:

[0116] Normalized to 0.92; After normalization, it becomes 0.89;

[0117] Normalized to 0.22; Normalized to 0.26;

[0118] Normalized to 0.64; Normalized to 0.51;

[0119] Normalized to 0.44; Normalized to 0.39;

[0120] Normalized to 0.12; Normalized to 0.21;

[0121] Normalized to 0.78; Normalized to 0.51;

[0122] Normalized to 0.24; Normalized to 0.00;

[0123] Normalized to 0.33; Normalized to 0.34;

[0124] Normalized to 0.40; Normalized to 0.14;

[0125] Normalized to 0.56; Normalized to 0.43;

[0126] Substituting the normalized data into the offset term product in sequence, we get:

[0127] Group 1 offset product: ;

[0128] Group 2 offset product: ;

[0129] Group 3 offset product: ;

[0130] Group 4 offset product: ;

[0131] Group 5 offset product: ;

[0132] Sum the five sets of results:

[0133] ;

[0134] in, Set the number of abnormal response behaviors for technical positions. Normalized to 0.60, the node count ratio Normalized to 0.32, the denominator is calculated as follows:

[0135] ;

[0136] Continue substituting into the overall formula:

[0137] ;

[0138] The results indicate that, under the current technical job title type, the comprehensive response characteristic interval is... The value is 0.0145; this value will be used in conjunction with other identity types in the future. Comparative analysis is performed to construct a complete set of node response features, because It reflects the weighted result of the offset after multi-factor normalization. The closer it is to 0, the closer the response behavior is to the average performance. The higher the value, the greater the deviation in the response behavior. This feature set will be used as one of the input data in the subsequent multi-source label consistency judgment and label credibility ranking process to realize the continuous transfer between data structures.

[0139] Please see Figure 4 The specific steps for obtaining the label consistency distribution parameters are as follows:

[0140] S311: Based on the node response feature set, analyze the node interaction frequency, response category label and relationship strength classification, obtain the tag content of collaboration ability, relationship strength and influence scope in each data source tag, identify the description category and tag characteristics of the tag content in the same dimension, and obtain the tag content classification results;

[0141] First, the task participation count of each node within different time periods is extracted from the task records and divided by the time span to obtain interaction frequency data. For example, if node A participates in 15 tasks within 30 days, the frequency is 0.5 times / day. Then, the task response category field is retrieved, and each participation of the node in the task flow is marked as "active reception," "passive response," "delayed feedback," or "instant feedback," etc. The frequency of each type is summarized to form the response category label. Next, the task relationship edge data between this node and other nodes is read. The relationship strength level is set based on indicators such as the number of tasks between the two parties and the number of mutual response records. If two nodes have more than 10 round-trip tasks and more than 7 two-way responses within 30 days, it is defined as a strong relationship; less than 5 times is a weak relationship. Each edge is assigned a strength label according to this standard. Then, the label content of the corresponding node in each data source is obtained. Collaboration labels such as "strong coordination ability" and "good execution" are extracted from the human resources system. Relationship labels such as "frequent communication" and "rapid feedback" are extracted from the collaboration platform. In the third-party evaluation, the scope of influence labels such as "covering multiple projects" and "strong resource linkage ability" are identified. The labels of each dimension are semantically classified. For example, "collaboration", "execution" and "cooperation" are classified as collaboration ability, "close", "frequent" and "weak connection" are classified as relationship strength, and "cross-department", "link center" and "widespread influence" are classified as scope of influence. Frequency statistics are performed on the same semantic labels and they are divided according to the occurrence structure to form a classification number and label characteristic identification field. The label content classification results are output.

[0142] S312: Based on the tag content classification results, analyze the tag content in each data source tag in the dimensions of collaboration ability, relationship strength and scope of influence, compare their similarity and deviation, refine the content differences of tags, and obtain tag difference measurement results;

[0143] First, extract the tag text for the same node from the HR system, collaboration platform, and third-party platform. These are then categorized into comparison pools according to their dimensions. Word decomposition and structural analysis are performed. For example, the HR tag is "strong execution," the collaboration platform tag is "timely feedback," and the external tag is "strong independent work ability." If the semantic focus is scattered, it is judged as a semantically inconsistent tag and recorded as a difference item in that dimension. Next, the proportion of the same keywords used in the same dimension across the three data sources is calculated. For example, "smooth communication" has a coverage rate of 80% in the HR system and 30% in the collaboration platform, a difference exceeding 30 percentage points, and is categorized as a distribution offset item. Furthermore, the length of each tag content is analyzed. For example, the average length of HR tags is 6 characters, platform tags are 12 characters, and external tags are 9 characters. The average difference is calculated and used as a basis for expression style differences. Finally, the number of semantically inconsistent items, distribution offset items, and expression difference items is summarized. A scoring standard is set: if two of the three indicators are offset, the dimension is marked as "moderate difference"; if all three are significant, it is "significant difference"; and if only one is slightly offset, it is "low difference." The three-dimensional tag difference levels for each node are labeled and output to form a complete tag difference measurement result.

[0144] S313: Based on the label difference measurement results, filter the associated label combinations, analyze the differences between multi-source labels, determine the distribution characteristics of the label data, and calculate the label consistency difference degree using formula (2). Based on this degree of difference and the distribution characteristics of the label data, a multidimensional comparison is performed to obtain the label consistency distribution parameters;

[0145] (2);

[0146] in, This represents the degree of similarity between tag content, indicating the level of similarity between tag content generated from different data sources within the same dimension. This represents the magnitude of the label's numerical deviation, used to indicate the degree of numerical discrepancy between label content and other data within the same dimension. This represents the frequency of evaluation distribution under a specific rating level, indicating the frequency of evaluation distribution for that label within that dimension, and reflecting the concentration of label evaluations. Represents the density of label data distribution, indicating the label The data distribution density along this dimension reflects the concentration of the label data in this dimension. Represents the range span, indicating the label. The span of the data distribution range is used to measure the breadth of the label data within this dimension. This refers to the number of multi-source tag data sources.

[0147] Tag consistency variance is a metric used to measure the consistency difference of tags generated from different data sources on the same dimension. It reflects the degree of difference between the tag content given by different data sources or evaluation systems under the same node and the same dimension. It is used to quantify the degree of difference in tag content across multiple data sources or evaluation systems and helps identify which data sources provide more consistent or more accurate tags under the same dimension.

[0148] The data sources for the three dimensions of collaboration ability, relationship strength, and scope of influence were sequentially combined and filtered. Data source combinations with significant differences in characteristics across each dimension were selected. Let the data sources involved in the collaboration ability dimension be Source 1, Source 2, and Source 3, where the degree of similarity... The value is 0.82, representing the magnitude of the numerical deviation. The evaluation grading frequency is 0.11. The label data distribution density is 3. , , The values ​​are 0.64, 0.52, and 0.46 respectively, with a range of [missing information]. , , The values ​​are 0.35, 0.41, and 0.39 respectively, representing the number of data sources involved. The expression is 3. First, calculate the expression on the left side of the formula:

[0149] , ;

[0150] Multiplying together Divide by The left term is:

[0151] ;

[0152] Then perform a summation operation on the right-hand side, sequentially... and Multiplying the corresponding products, we get:

[0153] ;

[0154] ;

[0155] ;

[0156] Summing the three together and then dividing by 3, the right-hand term is:

[0157] ;

[0158] Subtracting the two terms from each side yields the following result:

[0159] ;

[0160] The negative result indicates that there is a serious inconsistency in the distribution of the current tag group in terms of collaboration capability, and it is necessary to rebuild the tag comparison combination or change the reference data source group.

[0161] Then, the same processing is applied to the relation strength dimension, let:

[0162] , , ;

[0163] After normalization, we get They are 0.72, 0.68, and 0.65 respectively. The values ​​are 0.40, 0.38, and 0.36. Calculate the left-hand side:

[0164] ;

[0165] The item on the right is:

[0166] ;

[0167] ;

[0168] ;

[0169] The average value is:

[0170] ;

[0171] The difference is:

[0172] ;

[0173] Finally, perform the same operation on the scope of influence dimension, and set:

[0174] , , ;

[0175] After normalization The values ​​are 0.59, 0.55, and 0.48. The values ​​are 0.38, 0.35, and 0.32.

[0176] The left term is calculated as follows:

[0177] ;

[0178] The item on the right is:

[0179] ;

[0180] ;

[0181] ;

[0182] The average value is:

[0183] ;

[0184] Subtraction yields:

[0185] ;

[0186] The three sets of difference results were -0.183, -0.231, and -0.151, all of which were negative, indicating a high degree of inconsistency in the distribution across all dimensions. Ultimately, these three sets of data sources were identified as data source combinations with discrepancies, and their label consistency distribution parameters were output. The key difference points were mainly manifested in the distribution breadth exceeding the boundary due to the high distribution density and large span interval. This characteristic can be used for subsequent strategy adjustments for node label selection and content sorting.

[0187] Please see Figure 5 The specific steps for obtaining the preferred tag content library are as follows:

[0188] S411: Based on the label consistency distribution parameters, analyze the key differences in the combination of label dimensions, classify the collaboration ability, relationship strength and scope of influence, filter the corresponding data sources, map the correspondence between each dimension and the data source, and obtain the label source mapping structure.

[0189] Extract data source combinations that differ in label dimensions for each node identified in the preceding steps. Calculate the similarity score between data source labels for each node across three dimensions: collaboration ability, relationship strength, and scope of influence. Select the dimension label group with the largest difference value and inconsistent semantic classification as the key difference dimension combination. For example, in the collaboration ability dimension, node B has the HR label "strong cooperation awareness" and the business platform label "slow execution speed." These two labels are semantically positive and negative, respectively, and their matching score is less than 0.2, thus recording them as key difference labels. Then, classify the label content involved in this dimension combination separately. If the label contains "..." Word roots such as "collaboration," "promotion," "cooperation," and "execution" are categorized as collaboration capability dimension tags. If word roots such as "frequency," "density," "response," and "interaction" exist, they are categorized as relationship strength dimension tags. If words such as "influence," "radiation," "cross-level," and "multi-team" are included, they are classified as influence scope tags. After categorization, the number of times each data source uses the tags in the corresponding dimensions is counted, and the corresponding source data source for each type of tag is recorded. For example, if the collaboration platform frequently uses the tag "cross-team collaboration," a correspondence between the tag and the platform's data source is established. Finally, a combined table of node dimension, tag semantic category, and source information is output, forming a tag source mapping structure.

[0190] S412: Based on the tag source mapping structure, compare the tag content of each data source under each dimension combination, calculate the frequency, content offset and feedback interval of the content provided by each data source, optimize the distribution among the participating items, and sort the tag content using formula (3) to obtain the optimal tag sorting structure.

[0191] (3);

[0192] in, Indicates the first Tag content ranking metrics under a combination of tag dimensions Indicates the first In the combination of the _th label dimension, the _th Each data source provides the frequency of occurrence of the tagged content. Indicates the first In the combination of the _th label dimension, the _th Content offset of each data source Indicates the first The mean of content offset across all data sources under a combination of tag dimensions. Indicates the first In the combination of the _th label dimension, the _th Feedback interval of each data source Indicates the first Feedback expectations under a combination of label dimensions This indicates the total number of data sources.

[0193] Based on the tag source mapping structure, compare the tag content of each data source under each dimension combination, analyze the quantity of tag content provided by each data source under the corresponding dimension, and determine the frequency of occurrence by counting the number of times each tag is repeatedly labeled on the node. For example, data sources 101, 102, and 103 provide tag content frequencies of 12, 15, and 9 respectively under the "Collaboration Ability-1" dimension. After normalization, these frequencies are recorded as follows: , , Calculate the difference between the content of each tag group and the average content under the dimensional combination, and use the numerical distance between the coded tag content and the reference content as the offset. The original offsets are respectively , , After normalization, they are respectively , , Simultaneously calculate the mean of all data source offsets in this dimension combination. , set as Further, retrieve the feedback time interval of the corresponding tag content in the feedback records of each data source. The original intervals were 6, 5, and 7 days, respectively, and after normalization, we obtained... , , Corresponding feedback expectations The average feedback interval for this dimension combination is set to after normalization. Substitute the above parameters into formula (3).

[0194] The calculation process is as follows:

[0195] Item 1: , , , ;

[0196] Item 2: , , , ;

[0197] Item 3: , , , ;

[0198] The molecular part is:

[0199] ;

[0200] Denominator part:

[0201] Item 1: ;

[0202] Item 2: ;

[0203] Item 3: ;

[0204] The denominator is:

[0205] ;

[0206] The final calculation yields:

[0207] ;

[0208] The results indicate that, within the "Collaboration Capability-1" dimension combination, the ranking index obtained from the joint analysis of frequency, offset, and feedback interval among the data sources involved in the content ranking calculation for this tag is: It can be used to identify the optimal tag items and generate the tag optimization ranking structure in subsequent screening and sorting. The formula establishes a mathematical reference structure for tag content ranking by calculating the product deviation of content quality, frequency of provision and response behavior and comparing the overall distribution of feedback behavior, so that the evaluation process has statistical rationality and computational consistency.

[0209] S413: Based on the tag optimization and sorting structure, filter the top-ranked tag content, determine the optimal content under each dimension, integrate the description and source information, write the node tag content, and obtain the optimized tag content library.

[0210] Tag content ranking index refers to a comprehensive numerical reference for measuring and ranking the quality of tag content provided by different data sources under the same tag dimension combination. It is calculated by comprehensively analyzing the performance, offset and feedback of tag content from each data source. It is the main numerical parameter used to judge the priority when comprehensively ranking and filtering candidate tag content.

[0211] First, for each of the three dimensions—collaboration ability, relationship strength, and scope of influence—the corresponding data source labels for each node are extracted. Then, the accuracy parameter ranking results constructed previously are used to sort the label sources according to the priority of each dimension from high to low. For example, if node C's labels for the collaboration ability dimension come from the HR system, collaboration platform, and external evaluations, with accuracy scores of 0.87, 0.81, and 0.68 respectively, then the HR system label is selected as the primary content. The collaboration label "stable execution willingness" for the current node is extracted from this data source. Then, the collaboration content "proactive cooperation" and "timely task follow-up" are extracted from lower-ranked data sources. The primary content is then selected as the secondary content. The descriptions are placed in the core tag field, and the descriptions of the remaining tags are filled in the auxiliary fields in a concise order. The original data source and the timestamp of acquisition are recorded. When processing the tags of each node, if the similarity score of the top two source contents exceeds 0.8 and there are no key contradictory terms, multi-tag integration can be performed to merge them into a complete description. For example, "stable execution intention" and "proactive cooperation" can be integrated into "stable execution intention, proactive cooperation". After the tag description is constructed, it is uniformly formatted as a "dimension-tag content-data source-time" structure record and written into the node tag content field to generate a complete preferred tag content library.

[0212] Please see Figure 6 The specific steps for obtaining the node label update results are as follows:

[0213] S511: Based on the preferred tag content library, analyze the tags of each dimension in the temporary data of node tags, compare the content descriptions of collaboration ability, relationship strength and influence scope, determine whether there are differences in the content of each dimension tag, filter out the tag dimensions with differences, and obtain a list of dimension deviation indicators.

[0214] For each node, extract the three tags: collaboration ability, relationship strength, and scope of influence under the current dimension. Then, read the historical tag descriptions for the corresponding dimension from the node's temporary tag data, establish a field mapping for the three tags, and perform sentence splitting and word grouping operations on the mapped content. Extract keyword sets for comparison. For example, if node D's collaboration ability tag in the temporary data is "possesses team leadership potential," while the preferred tag is "skilled at coordinating cross-departmental teams," then extract the keywords "leadership," "coordination," "team," and "cross-departmental" for comparison. If the overlap is less than 40%, it is considered a content difference. After performing the above steps for each tag dimension... The number of dimensions identified as discrepancies across the three dimensions is counted, and these discrepancies are recorded in a deviation list. When performing these steps, a benchmark value for keyword overlap is set, uniformly set to 50%. Dimensions below this value are marked as discrepancies. Additionally, tag length differences are calculated; if the preferred tag length differs from the temporary tag length by more than 10 characters, it is also marked as a structural deviation. Combining keyword comparison results and structural difference judgments, a dimension deviation flag list is constructed for each node. This list, indexed by node number, lists tag content difference flags for the three dimensions of collaboration capability, relationship strength, and influence scope, indicating whether subsequent tag updates are required.

[0215] S512: Based on the dimensional deviation identifier list, analyze the source of the tags and the historical content of the temporary data of the node tags, determine the difference between the source content and the historical content, determine the tag parameters that need to be covered, and cover them item by item to obtain the node tag update results;

[0216] For each node, the tag items marked as differences in dimensions are further extracted for their tag source information and historical write timestamps. The data source record of the currently preferred tag content is compared side-by-side with the record of the historically stored tags. First, it is determined whether the two contents come from the same system. If they come from different data sources, it is further determined whether the accuracy ranking value of the current preferred tag source is higher than that of the historical content source. For example, in the collaboration dimension, the current tag comes from the organizational performance platform, and the historical tag comes from the project log platform. The performance platform ranks first in accuracy in this dimension, while the project log platform ranks third. Then it is determined that the tag content needs to be overwritten. If the current tag and the historical tag have the same source... However, if the content description changes significantly, for example, from "stable execution" to "outstanding promotion capability", then the difference in the keyword set before and after the change is compared. If the number of different keywords exceeds two, then the tag is also considered to need to be updated, and this dimension is added to the tag coverage item. Then, the tags to be covered are written into the main structure field of the node tag one by one. The coverage operation adopts the full field coverage method, that is, the original value is directly replaced by the complete tag description, source information and update timestamp in the preferred tag. After the replacement, a tag change record structure is generated, which records the tag content before and after the change, the data source and the operation time. Finally, the tag change records of all nodes are summarized and output to form the node tag update result.

[0217] This application also proposes a talent relationship evaluation system based on GNN, including:

[0218] Resume Attribution Module: Based on resume event data of talent nodes, analyze each job promotion, project assignment and cross-departmental transfer resume, classify each event type into the corresponding influencing factor category, complete the event arrangement according to the time sequence of resume events, calculate the cumulative sum of each influencing factor after classification, identify nodes whose cumulative sum meets the judgment criteria, and establish key event node information.

[0219] Response Feature Module: Based on key event node information, compare the task acceptance time and information feedback time between nodes. For each group of nodes, calculate the task response interval and information feedback interval, analyze the response performance under each node identity type, determine whether the task response and information feedback interval are different from the general distribution, classify and organize the behavioral influence factors of the edges, and obtain the node response feature set.

[0220] Tag Difference Module: Based on the node response feature set, calculate the content differences of multi-source tags in terms of collaboration ability, relationship strength and scope of influence. For the difference data source tags of the same node in the same time period, compare the dimensional content one by one, compare it with the consistency benchmark, identify the key data source combination of differences, and obtain the tag consistency distribution parameters.

[0221] Tag selection module: Based on tag consistency distribution parameters, it determines the historical accuracy parameters of associated tag data sources, prioritizes the tag dimension combinations with key differences according to accuracy parameters, selects the tag content with the highest accuracy parameters for each dimension, and executes the writing of node tag content to obtain the selected tag content library;

[0222] Tag update module: Based on the preferred tag content library, adjust the content of node tags in terms of collaboration ability, relationship strength and influence scope, analyze the consistency between the content of each item of the node tag temporary data and the preferred tag content, identify the dimensions that need to be updated, and cover the tag content item by item to obtain the node tag update results.

[0223] Each module in the above system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0224] This application also provides a computer device, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data related to the talent evaluation method of this application embodiment. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a talent evaluation method.

[0225] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0226] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A talent relationship evaluation method based on GNN, characterized in that: A talent relationship network is constructed using Generative Neural Networks (GNNs). The nodes of the network are individuals, i.e., each person being evaluated within the network. Interactions between individuals serve as the edges of the network. The implementation steps of the talent relationship evaluation method include: S1: Acquire and analyze the resume event data for each talent node. The resume event data includes job promotion, project assignment and cross-departmental transfer resumes. Each resume event is classified into the corresponding influencing factor category. The events are arranged in chronological order. The sum of each influencing factor after classification is calculated. Nodes whose sum meets the judgment criteria are identified. Key event node information is established. S2: Based on key event node information, compare the task acceptance time and information feedback time between nodes, divide talent nodes into multiple groups according to identity type, calculate the task response interval and information feedback interval for each group of nodes, analyze the response performance of each node in the group, determine whether the task response interval and information feedback interval are different from the general distribution, classify and organize the behavioral influence factors of the edges, and obtain the node response feature set. S3: Obtain multi-dimensional tag data of the same talent node from different data sources. Based on the node response feature set, calculate the content differences of multi-source tags in collaboration ability, relationship strength and scope of influence. For the different data source tags of the same node in the same time period, compare the dimension content one by one and compare it with the consistency benchmark to identify the different data source combinations and tag dimension combinations and obtain the tag consistency distribution parameters. S4: Based on the label consistency distribution parameters, determine the historical accuracy parameters of the associated label data source. For the different combinations of label dimensions, prioritize them according to the accuracy parameters, select the label content with the highest accuracy parameters for each dimension, and execute the content writing of the node labels to obtain the preferred label content library.

2. The talent relationship evaluation method based on GNN according to claim 1, characterized in that: The impact factor category in step S1 refers to classifying the strength of the impact relationship for each type of resume event, such as job promotion, project assignment, and interdepartmental transfer, to facilitate subsequent aggregation analysis. The specific steps for obtaining key event node information are as follows: S111: Based on the resume event data of talent nodes, analyze the job promotion, project assignment and cross-departmental transfer resume events of talent nodes, assign resume events to corresponding categories according to the resume events, and obtain the resume event ranking set according to the order of occurrence of the events; S112: Based on the resume event ranking set, calculate the cumulative impact of various resume events on talent nodes, compare the distribution of different event types under the same node, adjust the ranking and classification criteria of event categories, and obtain the impact factor distribution data. S113: Based on the distribution data of impact factors, determine the performance of talent nodes under the screening conditions, analyze the node's identification attributes, number of events participated in and event aggregation tags, optimize the node information content, and establish key event node information.

3. The talent relationship evaluation method based on GNN according to claim 1, characterized in that: The specific steps for obtaining the node response feature set are as follows: S211: Based on key event node information, compare the task acceptance time and information feedback time of each node, calculate the task response interval and information feedback interval for each group of nodes, filter out nodes with different response performance, and obtain node interaction time delay data. S212: Based on the node interaction time delay data, analyze the impact of node identity type on task response and feedback interval, determine the distribution of nodes in terms of interaction delay, identify nodes with abnormal response behavior, and obtain node response offset comparison data. S213: Based on the node response offset comparison data, calculate the offset magnitude between the task response interval and the information feedback interval, analyze the frequency of node event participation and abnormal response behavior, and use the following formula to filter the node response characteristic interval. The node response feature set is obtained; ； in, Indicates the first Task response interval for each node pair This represents the average response interval of all node network tasks. Indicates the first The cumulative amount of event participation frequency for each node pair. This represents the total number of node pairs under the current identity type. This indicates the number of events involved in a single node pair under the current identity type. This indicates the number of abnormal response behaviors under the current identity type. This indicates the proportion of the current identity type.

4. The talent relationship evaluation method based on GNN according to claim 1, characterized in that: The specific steps for obtaining the label consistency distribution parameters are as follows: S311: Based on the node response feature set, analyze the node interaction frequency, response category label and relationship strength classification, obtain the tag content of collaboration ability, relationship strength and influence scope in each data source tag, identify the description category and tag characteristics of the tag content in the same dimension, and obtain the tag content classification results; S312: Based on the tag content classification results, analyze the tag content in each data source tag in the dimensions of collaboration ability, relationship strength and scope of influence, compare their similarity and deviation, refine the content differences of tags, and obtain tag difference measurement results; S313: Based on the label difference measurement results, filter related label combinations, analyze the differences between multi-source labels to determine the distribution characteristics of the label data, and calculate the label consistency difference degree using the following formula. Based on this degree of difference and the distribution characteristics of the label data, a multidimensional comparison is performed to obtain the label consistency distribution parameters; ； in, This represents the degree of similarity between tag content, indicating the level of similarity between tag content generated from different data sources within the same dimension. This represents the magnitude of the label's numerical deviation, used to indicate the degree of numerical discrepancy between label content and other data within the same dimension. This represents the frequency of evaluation distribution under a specific rating level, indicating the frequency of evaluation distribution for that label within that dimension, and reflecting the concentration of label evaluations. Represents the density of label data distribution, indicating the label The data distribution density along this dimension reflects the concentration of the label data in this dimension. Represents the range span, indicating the label. The span of the data distribution range is used to measure the breadth of the label data within this dimension. This refers to the number of multi-source tag data sources.

5. The talent relationship evaluation method based on GNN according to claim 1, characterized in that: The specific steps for obtaining the preferred tag content library are as follows: S411: Based on the label consistency distribution parameters, analyze the key differences in the combination of label dimensions, classify the collaboration ability, relationship strength and scope of influence, filter the corresponding data sources, map the correspondence between each dimension and the data source, and obtain the label source mapping structure. S412: Based on the tag source mapping structure, compare the tag content of each data source under each dimension combination, calculate the frequency, content offset and feedback interval of the content provided by each data source, optimize the distribution among participating items, and sort the tag content using the following formula to obtain the optimal tag sorting structure. ； in, Indicates the first Tag content ranking metrics under a combination of tag dimensions Indicates the first In the combination of the _th label dimension, the _th Each data source provides the frequency of occurrence of the tagged content. Indicates the first In the combination of the _th label dimension, the _th Content offset of each data source Indicates the first The mean of content offset across all data sources under a combination of tag dimensions. Indicates the first In the combination of the _th label dimension, the _th Feedback interval of each data source Indicates the first Feedback expectations under a combination of label dimensions Indicates the total number of data sources; S413: Based on the tag optimization and sorting structure, filter the top-ranked tag content, determine the optimal content under each dimension, integrate the description and source information, write the node tag content, and obtain the optimized tag content library.

6. The talent relationship evaluation method based on GNN according to claim 1, characterized in that: It also includes S5: Based on the preferred tag content library, adjust the content of node tags in terms of collaboration ability, relationship strength and influence scope, analyze the consistency between the content of each item of the node tag temporary data and the preferred tag content, identify the dimensions that need to be updated, and cover the tag content item by item to obtain the node tag update results.

7. The talent relationship evaluation method based on GNN according to claim 6, characterized in that: The specific steps for obtaining the node label update results are as follows: S511: Based on the preferred tag content library, analyze the tags of each dimension in the temporary data of node tags, compare the content descriptions of collaboration ability, relationship strength and influence scope, determine whether there are differences in the content of each dimension tag, filter out the tag dimensions with differences, and obtain a list of dimension deviation indicators. S512: Based on the dimensional deviation identifier list, analyze the source of the tags and the historical content of the temporary data of the node tags, determine the difference between the source content and the historical content, identify the tag parameters that need to be covered, and cover them item by item to obtain the node tag update results.

8. A talent relationship evaluation system based on GNN, characterized in that: The system is used to implement the talent relationship evaluation method based on GNN as described in any one of claims 1-7, and the system comprises: Resume Attribution Module: Based on resume event data of talent nodes, analyze each job promotion, project assignment and cross-departmental transfer resume, classify each event type into the corresponding influencing factor category, complete the event arrangement according to the time sequence of resume events, calculate the cumulative sum of each influencing factor after classification, identify nodes whose cumulative sum meets the judgment criteria, and establish key event node information. Response Feature Module: Based on key event node information, compare the task acceptance time and information feedback time between nodes. For each group of nodes, calculate the task response interval and information feedback interval, analyze the response performance under each node identity type, determine whether the task response and information feedback interval are different from the general distribution, classify and organize the behavioral influence factors of the edges, and obtain the node response feature set. Tag Difference Module: Based on the node response feature set, calculate the content differences of multi-source tags in terms of collaboration ability, relationship strength and scope of influence. For the difference data source tags of the same node in the same time period, compare the dimensional content one by one, compare it with the consistency benchmark, identify the key data source combination of differences, and obtain the tag consistency distribution parameters. Tag selection module: Based on tag consistency distribution parameters, it determines the historical accuracy parameters of associated tag data sources, prioritizes the tag dimension combinations with key differences according to accuracy parameters, selects the tag content with the highest accuracy parameters for each dimension, and executes the writing of node tag content to obtain the selected tag content library; Tag update module: Based on the preferred tag content library, adjust the content of node tags in terms of collaboration ability, relationship strength and influence scope, analyze the consistency between the content of each item of the node tag temporary data and the preferred tag content, identify the dimensions that need to be updated, and cover the tag content item by item to obtain the node tag update results.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.

Images