Performance management method and device based on knowledge graph and medium

Abstract

The invention provides a performance management method based on a knowledge graph, and the method comprises the steps: carrying out the preliminary integration of multi-source heterogeneous data in advance, obtaining a target attraction event message, converting the discrete target attraction event message into structured graph data through a talent attraction knowledge graph, and carrying out the semantic completion and context association, thereby achieving the target attraction event message. The implicit relation which is not clearly expressed in the message is mined, and omission of the introduction workload is avoided. And then, based on a graph neural network model, the graph-based birth inducing event data is converted into birth inducing event vectors, so that complex birth inducing work of teachers is quantified to be suitable for business logic of different business scenes, and the flexibility of the performance management method is improved. And finally, matching the induction event vector with the preset induction target rule base, thereby realizing intelligent accounting of the completion degree of the induction target, not only improving the accounting precision and efficiency of the performance, but also supporting adjustment and expansion of flexible adjustment induction rules, and improving the suitability of the performance management method.

Description

Technical Field

[0001] This application relates to the field of data processing, and more particularly to a performance management method, device, and medium based on knowledge graphs. Background Technology

[0002] Traditional talent recruitment and development performance management in universities relies on manual data entry and periodic assessments. This method excessively depends on manual verification, resulting in rigid and inflexible performance calculation logic that struggles to adapt to dynamic rule adjustments and real-time decision-making scenarios. Furthermore, manually entered data is susceptible to subjective biases, leading to low efficiency and accuracy. The execution of performance rules also depends on human experience and judgment, resulting in low rule parsing accuracy and limited scalability, failing to meet the real-time processing demands of large-scale, high-concurrency business scenarios. Therefore, improving the efficiency and accuracy of performance calculation has become a pressing technical challenge. Summary of the Invention

[0003] The main purpose of this application is to provide a knowledge graph-based performance management method, computer equipment, and computer-readable storage medium, which aims to improve the efficiency and accuracy of performance calculation.

[0004] To achieve the above objectives, this application provides a knowledge graph-based performance management method, which includes the following steps: acquiring target talent recruitment and development event messages to characterize the workload of teachers to be assessed in talent recruitment and development; performing contextual association and semantic completion on the target talent recruitment and development event messages based on a preset talent recruitment and development knowledge graph to generate graph-based talent recruitment and development event data; generating talent recruitment and development event vectors for teachers to be assessed based on a graph neural network model and the graph-based talent recruitment and development event data; and calculating the completion degree of talent recruitment and development goals for teachers to be assessed based on a preset talent recruitment and development goal rule base and the talent recruitment and development event vectors, so as to complete the performance assessment of teachers' talent recruitment and development based on the completion degree of talent recruitment and development goals.

[0005] In addition, to achieve the above objectives, this application also provides a computer device, the computer device including a processor, a memory, and a knowledge graph-based performance management program stored in the memory and executable by the processor, wherein when the knowledge graph-based performance management program is executed by the processor, it implements the steps of the knowledge graph-based performance management method as described above.

[0006] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a knowledge graph-based performance management program, wherein when the knowledge graph-based performance management program is executed by a processor, it implements the steps of the knowledge graph-based performance management method described above.

[0007] This application provides a knowledge graph-based performance management method. The method first integrates multi-source heterogeneous data to obtain target talent attraction and development event messages. Then, it uses a talent attraction and development knowledge graph to transform these discrete event messages into structured graph data. Through semantic completion and contextual association, it uncovers implicit relationships not explicitly expressed in the messages, avoiding omissions in talent attraction and development workload. Next, based on a graph neural network model, the graph-based event data is converted into attraction and development event vectors, thereby quantifying teachers' complex talent attraction and development work to suit different business scenarios and improve the flexibility of the performance management method. Finally, the attraction and development event vectors are matched with a pre-set base of attraction and development target rules to achieve intelligent calculation of the completion rate of attraction and development targets. This not only improves the accuracy and efficiency of performance calculation but also supports flexible adjustment and expansion of attraction and development rules, enhancing the adaptability of the performance management method. Attached Figure Description

[0008] Figure 1 A flowchart illustrating a knowledge graph-based performance management method provided in this application; Figure 2 A flowchart illustrating another knowledge graph-based performance management method provided in this application.

[0009] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0010] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it require execution in the described order. Some operations / steps may be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0012] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0013] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0014] The knowledge graph-based performance management method involved in the embodiments of this application is mainly applied to computer devices, such as PCs, laptops, and mobile terminals, which have display and processing functions.

[0015] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features described herein are combined with each other.

[0016] Reference Figure 1 , Figure 1 This application provides a flowchart illustrating a knowledge graph-based performance management method.

[0017] like Figure 1 As shown in the figure, this application provides a knowledge graph-based performance management method. In this embodiment, the knowledge graph-based performance management method includes the following steps: Step S101: Obtain the target recruitment event message used to characterize the recruitment workload of teachers to be calculated; Step S102: Based on the preset talent attraction and development knowledge graph, perform context association and semantic completion on the target talent attraction and development event message to generate graph-based talent attraction and development event data; Step S103: Based on the graph neural network model and the graph-based recruitment and training event data, generate the recruitment and training event vector of the teacher to be accounted for; Step S104: Based on the preset talent attraction and development target rule base and the talent attraction and development event vector, calculate the talent attraction and development target completion rate of the teacher to be accounted for, and complete the teacher talent attraction and development performance accounting based on the talent attraction and development target completion rate.

[0018] In this embodiment, the talent attraction and development event message refers to the original data record obtained from various business systems of universities, used to describe teacher talent attraction and development activities. The message may contain information such as teachers' teaching activities, research achievements, student training, and talent introduction, and its format and content may vary from system to system. The talent attraction and development knowledge graph is a knowledge graph specifically constructed for the field of talent attraction and development in universities. This graph integrates entities such as teachers, students, projects, papers, and awards and their interrelationships, and adds rich attribute information, providing a unified and comprehensive knowledge foundation for talent attraction and development performance management. The graph-based talent attraction and development event data refers to the talent attraction and development event data after being processed by the knowledge graph. By performing entity recognition, relation extraction, semantic completion, and context association on the original talent attraction and development event messages, the unstructured or semi-structured message data is transformed into graph data with rich semantic information that conforms to the knowledge graph structure. The talent attraction and development event vector is a numerical vector representation generated by the graph neural network model for the teachers to be accounted for and their related talent attraction and development events. This vector encodes the teacher's personal characteristics, the type, level, and timing of their talent recruitment activities, as well as their relationships with other entities (such as students and projects), providing quantitative input for subsequent performance evaluation. The talent recruitment goal rule base is a predefined set of rules used to assess the teacher's talent recruitment performance. These rules may include specific requirements and weights for aspects such as the number of students mentored, the level of research projects, the number of published papers, and the number of talents recruited, used to calculate the teacher's talent recruitment goal completion rate. The talent recruitment goal completion rate is an indicator calculated based on the talent recruitment goal rule base and the teacher's talent recruitment event vector, measuring the extent to which the teacher has achieved the preset goals in talent recruitment. This indicator is usually expressed as a percentage or other quantitative form, directly reflecting the teacher's performance level.

[0019] Specifically, this involves obtaining target recruitment and development event messages to characterize the recruitment and development workload of the teachers to be assessed. This step can be implemented in two ways: either manually entering raw event data from various business systems into a unified message format; or using periodic data export tools to batch export data from different business systems (e.g., graduate student management system, personnel management system, research management system, financial management system, etc.), followed by format conversion and integration by data processing personnel to form unified event messages. These messages may contain information on teachers' teaching, research, and administrative activities, but their formats and content may differ, requiring further processing.

[0020] Based on a pre-defined talent attraction and development knowledge graph, contextual association and semantic completion are performed on the target talent attraction and development event message to generate graph-based talent attraction and development event data. This step is achieved as follows: First, through text matching or keyword extraction techniques, entities (such as teacher names, student names, project names, paper titles, etc.) and relationships (such as guidance, participation, publication, etc.) related to talent attraction and development are identified from the message. Then, these identified entities and relationships are directly mapped to existing nodes and edges in the talent attraction and development knowledge graph, or new nodes and edges are created. For relationships not explicitly expressed in the message but implicitly present, inference is made using pre-defined simple logical rules. For example, if the message shows "Teacher A is the person in charge of Project B" and "Student C participates in Project B," then it is inferred that "Teacher A indirectly guides Student C." In this way, discrete message information is transformed into structured graph data.

[0021] Based on a graph neural network model and the graph-based recruitment event data, a recruitment event vector for the teacher to be evaluated is generated. This step involves: First, converting the generated graph-based recruitment event data into a graph structure that can be processed by the graph neural network model, where each entity (e.g., teacher, student, project) is considered a node, and the relationships between entities are considered edges. Then, a feature vector is initialized for each node; for example, teacher nodes are initialized based on their title, age, and subject; student nodes are initialized based on their major and year of enrollment. Subsequently, using a graph neural network model (e.g., graph convolution operations), the feature vector of each node is iteratively updated, aggregating the feature information of its neighboring nodes. After multiple aggregations, each node obtains a rich vector representation containing its own features and local neighborhood information. Finally, from these node vectorization representations, the node vector of the teacher to be evaluated and the vectors of the recruitment event nodes directly related to it are extracted and combined to form the teacher's recruitment event vector.

[0022] Based on a pre-defined rule base for talent attraction and development goals and the talent attraction and development event vector, the completion rate of the talent attraction and development goals for the teacher to be evaluated is calculated, and the teacher's talent attraction and development performance is evaluated based on this completion rate. This step is implemented as follows: First, each rule in the talent attraction and development goal rule base (e.g., "supervise at least 2 doctoral students" or "publish at least 1 SCI paper") is manually encoded as a specific rule feature. Then, a simple vector similarity calculation (e.g., cosine similarity) is performed between the teacher's talent attraction and development event vector and these rule features to determine whether the teacher meets each rule. By counting the number of rules met and combining this with the pre-defined total number of goals, the teacher's completion rate of talent attraction and development goals is calculated. For example, if there are 10 talent attraction and development goals in the rule base, and the teacher meets 8 of them, then their completion rate is 80%. Finally, based on this completion rate and combined with the pre-defined performance level standards, the teacher's performance is evaluated.

[0023] The following example will provide a more detailed explanation of the above technical solution: Suppose a university needs to calculate the performance of teacher A in recruitment and development over the past year. Traditionally, teacher A would have to manually fill out various forms, and the department and functional units would then manually verify the data in the graduate student system, personnel system, research system, and financial system—a time-consuming and error-prone process.

[0024] The performance management method provided in this embodiment first obtains target recruitment and training event messages to characterize the recruitment and training workload of Teacher A. Specifically, the system obtains records from the graduate student system of Teacher A's guidance of students B and C, records from the research system of Teacher A's participation in project D and publication of paper E, records from the personnel system of Teacher A's introduction of talent F, and records from the financial system of the funds received for project D. These raw data may exist in different formats, such as CSV files exported from the graduate student system and XML interface data provided by the research system. At this stage, these heterogeneous data are initially integrated into a series of event messages, such as "Teacher A guides student B, time: September 2023" and "Teacher A participates in project D, level: national level, funding: 1 million, time: January 2023".

[0025] Based on a pre-defined talent attraction and development knowledge graph, the system performs contextual association and semantic completion on these target talent attraction and development event messages, generating graph-based talent attraction and development event data. For example, the system first identifies entities such as "Teacher A," "Student B," "Project D," "Paper E," and "Talent F" in the messages, as well as relationships such as "Guidance," "Participation," "Publishing," and "Introduction." These entities and relationships are mapped to the talent attraction and development knowledge graph. If a node for "Teacher A" already exists in the knowledge graph, the newly identified information is appended to that node; otherwise, a new node is created. Simultaneously, the system performs semantic completion. For instance, if the message only mentions "Teacher A guided Student B," but the knowledge graph contains information such as "Student B published Paper G in 2023," the system, through reasoning, establishes an indirect "guided student publication" relationship between Teacher A and Paper G, thus completing Teacher A's contribution to talent attraction and development. Thus, all discrete event messages are transformed into a graph-structured data centered on teacher A, containing rich related information such as their guidance of students, participation in projects, publication of papers, and recruitment of talent.

[0026] Based on the graph neural network model and the graph-based mentoring event data, a mentoring event vector for teacher A is generated. The system inputs the aforementioned graph-based mentoring event data into a pre-trained graph neural network model. This model treats entities such as teacher A, student B, project D, and paper E as nodes in the graph, and relationships such as "guidance," "participation," and "publication" as edges. Each node is assigned initial features, such as teacher A's title and age, student B's major and year of enrollment, and project D's level and funding. The graph neural network model learns the deep representation of each node, i.e., the node vector, through multi-layer information aggregation. For example, teacher A's node vector not only contains its own features but also incorporates feature information from neighboring nodes such as the students it has mentored, the projects it has participated in, and the papers it has published. Finally, a comprehensive vector representing teacher A and its mentoring workload is extracted from these node vectors as teacher A's mentoring event vector. This vector is a high-dimensional numerical representation that comprehensively and quantitatively reflects teacher A's mentoring work.

[0027] Based on a pre-defined rule base for talent attraction and development goals and the corresponding event vector, the system calculates the completion rate of teacher A's talent attraction and development goals. This completion rate is then used to perform performance evaluation for teacher talent attraction and development. The rule base may include rules such as "supervising ≥2 doctoral students," "hosting ≥1 national-level project," and "publishing ≥1 SCI Q1 paper." The system matches teacher A's event vector with these rules. For example, through vector similarity calculation, the system determines whether teacher A has supervised more than 2 doctoral students (based on the number and type of student nodes in the event vector), whether they have hosted a national-level project (based on the level information of the project node), and whether they have published an SCI Q1 paper (based on the journal information of the paper node). Based on the matching results, the system counts how many talent attraction and development goals teacher A has completed. Assuming there are 10 goals in the rule base and teacher A has completed 8, their completion rate is 80%. Finally, this completion rate is used to automatically generate teacher A's performance evaluation report, which can be further used for performance rating and reward allocation.

[0028] Compared to traditional manual reporting and periodic assessment methods, this embodiment achieves preliminary integration of multi-source heterogeneous data by using target recruitment event messages to characterize the recruitment workload of teachers to be calculated. By transforming this heterogeneous data into unified event messages, it lays the foundation for subsequent automated processing and effectively alleviates the problem of insufficient multi-source heterogeneous data processing capabilities.

[0029] This embodiment addresses the shortcomings of traditional data processing workflows and the lack of data credibility assurance mechanisms by generating graph-based mentoring event data. Traditional systems only perform simple ETL processing, making it difficult to capture deep semantic relationships and implicit information between data. This embodiment utilizes knowledge graphs to not only transform discrete event messages into structured graph data but also uncover implicit relationships not explicitly expressed in the messages through semantic completion and contextual association, such as the indirect contribution of teachers to student papers. This in-depth data processing approach makes performance data more comprehensive and accurate, providing richer and more reliable evidence for subsequent accounting and effectively avoiding the risk of data falsification that may arise from manual reporting. Furthermore, generating mentoring event vectors for the teacher to be accounted for effectively avoids the difficulty of comprehensively and dynamically assessing the complex mentoring workload of teachers, which often relies on preset fixed indicators and simple statistical methods in traditional performance management. By learning deep feature representations of teachers and their mentoring events from complex graph structure data through a graph neural network model, high-dimensional mentoring event vectors are generated. This vector not only incorporates teachers' direct contributions but also integrates their local neighborhood information within the knowledge graph, thus quantifying teachers' recruitment and training workload more comprehensively and precisely, overcoming the limitations of traditional rules in handling complex business logic. Finally, by matching the recruitment and training event vector with the rule base and calculating the teacher's recruitment and training goal completion rate, automated and intelligent performance evaluation is achieved. This reduces the subjectivity and inefficiency of manual judgment, making the execution of performance rules more efficient and accurate, and supporting more flexible rule adjustments and expansions.

[0030] To further address the aforementioned issues, in this embodiment, step S102 includes: performing entity identification and relation extraction on the target talent attraction and development event message based on predefined entity types and predefined relation types in the talent attraction and development knowledge graph to obtain target talent attraction and development event entities and target talent attraction and development event relations; fusing the target talent attraction and development event entities into corresponding nodes in the talent attraction and development knowledge graph; attaching metadata describing the inherent characteristics of the target talent attraction and development event entities in the target talent attraction and development event message as node attributes to the corresponding nodes in the talent attraction and development knowledge graph; attaching metadata describing the relational characteristics of the target talent attraction and development event in the target talent attraction and development event message as edge attributes to the corresponding edges in the talent attraction and development knowledge graph; performing implicit relation reasoning on the target talent attraction and development event message based on the explicit relations of the fused talent attraction and development knowledge graph; constructing the context association of the target talent attraction and development event message based on the fused implicit relations of the talent attraction and development knowledge graph, and obtaining the graph-based talent attraction and development event data based on the context association.

[0031] In this embodiment, based on predefined entity types and predefined relationship types in the talent attraction and development knowledge graph, entity recognition and relationship extraction are performed on the target talent attraction and development event messages to obtain target talent attraction and development event entities and relationships. The aim is to identify key information units (entities) related to talent attraction and development and their interrelationships (relationships) from unstructured or semi-structured target talent attraction and development event messages. Entity recognition employs rule-based methods, such as predefined keyword lists and regular expression matching, or machine learning-based methods, such as conditional random fields, support vector machines, or deep learning models for training and recognition. Relationship extraction employs pattern-matching methods, such as defining specific syntactic patterns to identify relationships in subject-verb-object structures, or supervised learning methods, such as transforming the relationship extraction task into a classification problem and using neural network models to classify the relationships between entity pairs. The target talent attraction and development event entities are then integrated into corresponding nodes in the talent attraction and development knowledge graph. The purpose is to match and integrate the specific entity instances identified from the messages with abstract concepts or existing entities already present in the talent attraction and development knowledge graph. The fusion process employs entity alignment techniques, such as string similarity-based matching or graph embedding, mapping entities to a low-dimensional vector space and determining whether they belong to the same entity by calculating the distance or similarity between vectors. If the entity in the message is a new instance of an existing entity in the knowledge graph, it is associated with an existing node; if it is a completely new entity, a new node is created in the knowledge graph. Metadata describing the inherent characteristics of the target talent attraction event entity in the message is appended as node attributes to the corresponding node in the talent attraction knowledge graph, aiming to enrich the detailed information of the nodes in the knowledge graph. Metadata is specific descriptive information about the entity, such as a teacher's title, a student's enrollment year, or the project's funding amount. Attaching node attributes is achieved by directly mapping metadata fields to the attribute fields of knowledge graph nodes, or by defining specific data models and ontology to standardize the storage and representation of attributes. Metadata describing the relationship characteristics of the target talent attraction event in the target talent attraction event message is appended as edge attributes to the corresponding edges in the talent attraction knowledge graph. The purpose is to provide a more granular description of the relationships (edges) in the knowledge graph. Relationship characteristic metadata includes information such as relationship strength, timestamp, and source. Appending edge attributes is achieved by defining attribute fields for edges in the graph database; for example, for a "guidance" relationship, attributes such as "guidance start time" and "guidance end time" are appended. Based on the explicit relationships in the fused talent attraction knowledge graph, implicit relationship reasoning is performed on the target talent attraction event message to discover potential relationships that are not explicitly mentioned in the message but logically exist.Implicit relationship reasoning employs rule-based reasoning, such as defining a reasoning rule like "If A guides B, and B participates in C, then A indirectly participates in C"; it also uses graph-based reasoning, such as pathfinding algorithms to discover potential connections between entities; or machine learning-based reasoning, such as knowledge graph embedding models to predict missing relation triples. Based on the talent attraction and development knowledge graph that integrates the implicit relationships, the contextual associations of the target talent attraction and development event messages are constructed, and based on these contextual associations, the graph-based talent attraction and development event data is obtained. The purpose is to organize the knowledge graph information, after entity, relation, and attribute completion and implicit relationship reasoning, into a structured data that comprehensively reflects the target talent attraction and development event. The construction of contextual associations involves extracting entities, relations, and attributes related to the target talent attraction and development event from the knowledge graph in the form of subgraphs. The graph-based talent attraction and development event data is a representation of this subgraph, such as an adjacency matrix, adjacency list, or RDF triple set, which contains the original information of the event message and rich contextual information obtained through knowledge graph completion and reasoning.

[0032] This application's solution overcomes the problems of incomplete information or difficulty in discovering implicit relationships by deeply processing the original target talent attraction event messages. First, using predefined entity types and predefined relationship types in the talent attraction knowledge graph, the message undergoes refined entity identification and relationship extraction, transforming key information into structured target talent attraction event entities and relationships. Subsequently, these identified entities are intelligently integrated into the corresponding nodes of the talent attraction knowledge graph. Simultaneously, metadata describing entity and relationship characteristics in the message is attached to the knowledge graph as node and edge attributes, greatly enriching the detail and expressive power of the knowledge graph. Building on this, the solution further utilizes existing explicit relationships in the integrated talent attraction knowledge graph to infer implicit relationships in the target talent attraction event messages. This means that even if certain associations are not directly mentioned in the message, the system can discover and complete these potential, unexpressed relationships through the logical structure and reasoning capabilities of the knowledge graph. For example, based on the known relationships between teachers and students, and students and projects, the indirect relationship between teachers and projects can be inferred. Finally, these relationships, completed through explicit and implicit methods, are integrated into a knowledge graph. Based on this, a complete contextual association of target breeding event messages is constructed. This contextual association is a graph structure rich in semantic information. It not only contains the original information of the message but also integrates scattered information through the powerful semantic representation capabilities of the knowledge graph, forming comprehensive and accurate graph-based breeding event data. This data format more effectively supports subsequent graph neural network models in generating breeding event vectors, thereby improving the accuracy and comprehensiveness of performance accounting.

[0033] Through the aforementioned technical solution, this application effectively addresses the problems of incomplete information, unclear entity relationships, and unmined implicit information in the original talent recruitment and development event messages. By performing refined entity identification, relationship extraction, and attribute appending on the messages, the knowledge graph more comprehensively and accurately reflects the details of the talent recruitment and development events. Crucially, by reasoning about implicit relationships based on the explicit relationships in the knowledge graph, this solution discovers and completes potential associations not directly expressed in the messages, thereby constructing a richer and more complete contextual association. This graph-based talent recruitment and development event data, after deep semantic completion and association, provides high-quality, high-dimensional input for subsequent talent recruitment and development event vector generation, improving the accuracy and comprehensiveness of teacher talent recruitment and development performance evaluation, and avoiding unfair or inaccurate performance evaluations due to missing information or misunderstandings.

[0034] In one embodiment, the implicit relationship reasoning of the target talent attraction and development event message based on the explicit relationships of the fused talent attraction and development knowledge graph includes: Based on the display relationship path template corresponding to the display relationship, at least one candidate relationship chain starting from the teacher node to be verified is constructed, wherein the teacher node to be verified is the teacher node in the target introduction and cultivation event message; Based on a multi-source confidence fusion strategy, candidate relation chains with confidence scores below a preset confidence threshold are filtered to obtain filtered candidate relation chains. Based on the graph embedding model and the candidate relation chain, the completion probability of each missing link in the candidate relation chain is calculated, and the implicit relation is generated based on the missing links whose completion probability exceeds a preset probability threshold.

[0035] In this embodiment, the explicit relationship path template refers to a predefined path pattern in the talent attraction and development knowledge graph used to describe specific types of associations between entities. Its function is to provide a structured search space and reasoning rules for implicit relationship reasoning. These templates are manually defined by domain experts and automatically generated through pattern mining of explicit relationships in the existing knowledge graph. The teacher node to be accounted for refers to the entity node in the talent attraction and development knowledge graph corresponding to the teacher's identity. It is used to focus on the teacher's attraction and development work, ensuring that subsequent implicit relationship reasoning is directly related to the teacher's workload accounting. The candidate relationship chain refers to a series of entities and relationship sequences that may have implicit relationships, constructed from the teacher node to be accounted for, along the explicit relationship path templates in the knowledge graph. Constructing candidate relationship chains is the foundation for discovering implicit relationships.

[0036] Multi-source confidence fusion strategies are methods used to evaluate and integrate confidence scores from different information sources or reasoning paths to improve the accuracy of implicit relation reasoning. These strategies include weighted averaging, Bayesian fusion, or DS evidence theory. A pre-set confidence threshold is a predetermined value used to judge the reliability of candidate relation chains, serving as a screening mechanism to ensure that only sufficiently reliable candidate relation chains proceed to the subsequent completion probability calculation stage. Graph embedding models are techniques that map entities and relations in a knowledge graph to a low-dimensional continuous vector space, providing a quantifiable and computable representation for implicit relation reasoning. Examples include TransE, TransR, DistMult, and ComplEx models. The completion probability of a missing link refers to the likelihood of the existence of an unexpressed relation or entity in a candidate relation chain, quantifying the probability of an implicit relation's existence and providing a decision-making basis for the final generation of implicit relations. The preset probability threshold is a pre-defined value used to determine whether the probability of completing the missing link is high enough to generate implicit relationships. This threshold serves as the final decision criterion, ensuring that only implicit relationships with high confidence are generated. Implicit relationships refer to potential connections between entities that are not explicitly expressed in the original target event message but are discovered through knowledge graph reasoning. Generating implicit relationships is a crucial step in semantic completion.

[0037] For example, implicit relation reasoning based on multi-strategy fusion includes: I. Based on the knowledge graph data with completed node linking and attribute fusion, and using predefined explicit relation path templates, construct candidate relation chains starting from teacher nodes, specifically including: 1. During the system initialization phase, construct a predefined explicit relational path template library using at least one of the following methods: The path template is manually defined based on domain expert knowledge, and frequently occurring patterns are mined from historical data. It is also based on vectorized representation for similar path discovery. The path template adopts a structured representation and includes at least path length, node type sequence, relation type sequence, and constraint field. 2. From the knowledge graph data, locate the corresponding teacher node as the query anchor point based on the globally unique teacher code that identifies a specific teacher; 3. Starting from the anchor teacher node, perform a graph traversal query in the knowledge graph to instantiate the predefined path template into specific candidate relationship chains; the graph traversal query adopts at least one of the following methods: Pattern matching based on graph query language, pattern matching based on graph traversal algorithm, and index-based optimized matching; 4. Perform quality assessment on the candidate relationship chains generated by matching, and filter out low-quality candidate chains; the quality assessment includes: Path integrity check, time window constraint check, and data source confidence check; 5. Summarize the candidate relationship chain instances that have undergone quality assessment and filtering into a candidate relationship chain set.

[0038] Second, a multi-source confidence fusion mechanism is used to screen the candidate relationship chain set, specifically including: Data source authority confidence score, which is the authority score of each fact in the relationship chain calculated based on the authority weight of the business system from which the data is sourced. Time consistency confidence score is a measure of the logical consistency of events occurring within a time window, which examines the continuity of timestamps in a relational chain. Historical accuracy confidence is a priori confidence level assigned based on the verification accuracy of this type of relationship chain in historical data. The above multi-source confidence scores are weighted and fused to filter out candidate relationship chains with confidence scores below a preset threshold, and a set of high-confidence candidate chains is output. Third, based on a high-confidence candidate chain set, the nodes and relations in the set are input into a pre-trained graph embedding model to learn the representations of nodes and relations in a low-dimensional vector space; based on the vector representation, the probability of completing missing links in the relation chain is calculated to predict potential implicit relations; specifically including: For the candidate chain "Teacher A → Instructor → Student B → Publication → Paper C", if the relationship "Student B → Publication → Paper C" is missing, but the graph embedding model predicts that the probability of completing this relationship is higher than the threshold, then the implicit relationship "Teacher A has made an indirect contribution to Paper C" is derived. For the candidate chain "Teacher A → Introduction → Talent D → Participation → Project E", if the relationship "Talent D → Participation → Project E" is missing, but the graph embedding model predicts that the probability of completing this relationship is higher than a threshold, then the implicit relationship "Teacher A has an indirect contribution to Project E" is derived; specifically including: 1. Graph Embedding Model Initialization: Using TransE, RotatE, or ComplEx graph embedding algorithms, a graph embedding model is trained based on historical knowledge graph data to learn the representation of nodes and relations in a low-dimensional vector space. The graph embedding model maps each node and relation type to a fixed-dimensional vector representation, such that in the vector space, if a triple (head entity, relation, tail entity) holds, the head entity vector plus the relation vector is approximately equal to the tail entity vector. 2. Identifying Missing Links: For each high-confidence candidate chain, the missing links are identified. The missing links refer to triple relations in the candidate chain that conform to a predefined path template but do not actually exist in the knowledge graph. Specifically, this includes: For the candidate chain "Teacher A→Guide→Student B→Publish→Paper C", if the relationship "Student B→Publish→Paper C" does not exist in the knowledge graph, then the relationship is identified as a missing link. For the candidate chain "Teacher A→Introduction→Talent D→Participation→Project E", if the relationship "Talent D→Participation→Project E" does not exist in the knowledge graph, then the relationship is identified as a missing link. 3. Calculate the completion probability: Based on the vector representation of the graph embedding model, calculate the completion probability of the missing link; specifically, this includes: for the missing relation (head entity, relation type, tail entity), calculate the sum vector of the head entity vector and the relation type vector; calculate the distance between the sum vector and the tail entity vector, using at least one of Euclidean distance, Manhattan distance, or cosine distance; convert the distance into a completion probability, the smaller the distance, the higher the completion probability. 4. Predicting Implicit Relationships: Based on the completion probability calculation results, predict potential implicit relationships; specifically including: Set a completion probability threshold. When the completion probability is higher than the threshold, the missing link is determined to be a potential implicit relationship. For the missing link "Student B → Publication → Paper C", if the probability of completion is higher than the threshold, the relationship is predicted to exist, and the implicit relationship "Teacher A has made an indirect contribution to Paper C" is deduced. For the missing link "Talent D → Participation → Project E", if the probability of completion is higher than the threshold, the relationship is predicted to exist, and the implicit relationship "Teacher A has made an indirect contribution to Project E" is derived; 5. All predicted potential implicit relationships are summarized into an implicit relationship set.

[0039] IV. Based on the predicted implicit relationships, an attention mechanism is used to automatically learn the weight allocation of different relationship paths to the final performance contribution, including: A multi-hop relation subgraph centered on teacher nodes is constructed, and feature representations of each relation path are extracted. A multi-head self-attention mechanism is employed to calculate the attention weights of different relation paths. Based on these attention weights, a quantified contribution weight is assigned to each implicit relation. Specifically, this includes: 1. Relationship Path Extraction: Construct a multi-hop relationship subgraph centered on teacher nodes, and extract the feature representations of each relationship path; the feature representations include: path length feature: the number of hops in the relationship path; node type feature: the type sequence of nodes in the path; relationship type feature: the type sequence of relationships in the path; attribute feature: the attribute values ​​of nodes and edges in the path, including paper impact factor, project level, funding amount, and participating role; 2. Attention Weight Calculation: A multi-head self-attention mechanism is adopted to calculate the attention weights of different relational paths. Specifically, this includes: mapping the feature representations of each relational path to query vectors, key vectors, and value vectors; calculating the similarity between query vectors and key vectors, and normalizing it to attention weights using the softmax function; and learning attention weights from different representation subspaces using a multi-head mechanism to enhance the model's expressive power. A larger attention weight indicates that the relational path is more important to the performance. 3. Contribution Weight Quantification: Based on attention weight, each implicit relation is assigned a quantified contribution weight; specifically, this includes: combining attention weight with a preset contribution decay or gain model; adjusting the final contribution weight by considering factors such as the length of the relation path, node type, and relation type; and obtaining a set of implicit relations containing quantified contribution weights. V. Based on the verification result signals and their differences from the actual performance labels, a reinforcement learning mechanism is used to dynamically optimize the inference rules and weight allocation strategy; specifically including: The inference rules and weight allocation strategy are considered as the action space of the agent; the accuracy, recall, and F1 score of the verification results are used as reward signals; the optimal combination of inference rules and weight allocation strategy is learned through policy gradient or Q-learning algorithms; specifically including: 1. State Space Definition: The current knowledge graph state, inference rule configuration, and weight allocation strategy are used as the state space for reinforcement learning; the state space includes: the graph structure features of the knowledge graph; the configuration parameters of the inference rule template; and the parameter settings of the weight allocation strategy. 2. Action Space Definition: The adjustment of inference rules and the modification of weight allocation strategies are defined as the action space of reinforcement learning; specifically, this includes: adding, deleting, and modifying inference rule templates; adjusting weight allocation strategy parameters; and optimizing the parameters of contribution decay or gain models. 3. Reward Signal Design: The accuracy, recall, and F1 score of the verification results are used as reward signals; specifically, these include: Accuracy reward: the proportion of verification results that match the true labels; Recall reward: the proportion of correctly identified positive examples out of all positive examples; F1 score reward: the harmonic mean of accuracy and recall; Penalty mechanism: applying negative rewards for overfitting or rule conflicts. 4. Policy Learning and Optimization: Through policy gradient or Q-learning algorithms, learn the optimal combination of inference rules and weight allocation strategies; specifically including: Policy gradient algorithm: directly optimize policy network parameters to maximize expected reward; Q-learning algorithm: learn the action value function and select the action with the highest value; Experience replay mechanism: store historical experience to improve learning efficiency; Target network mechanism: stabilize the training process and prevent overestimation of Q value; 5. Rule and strategy update: Update the learned optimal inference rules and weight allocation strategies to the system.

[0040] VI. Based on the candidate relation chain set, graph pruning techniques are used to reduce unnecessary computational overhead, including: Pruning based on degree centrality; pruning based on time windows; and pruning based on importance sampling; specifically including: 1. Degree Centrality Pruning: Pruning is performed based on the degree centrality of nodes, removing nodes with excessively low degree centrality. Specifically, this includes: calculating the degree centrality of each node in the knowledge graph; setting a degree centrality threshold and pruning nodes with degree centrality below the threshold; these nodes contribute little to relation propagation, and pruning them has little impact on the final result. 2. Time window pruning: Pruning is performed based on the time cycle requirements of performance accounting; specifically, this includes: setting a time range according to the effective time window of performance accounting; pruning nodes and relationships outside the time window; and retaining only events and relationships that occur within the time window. 3. Importance Sampling Pruning: Random walk or PageRank algorithm is used to sample important nodes and relationships for reasoning; specifically, it includes: sampling important nodes starting from teacher nodes based on the random walk algorithm; calculating the importance score of nodes based on the PageRank algorithm and sampling important nodes; setting the sampling ratio to control the size of the pruned graph; importance sampling pruning reduces computational complexity while ensuring reasoning accuracy; 7. Caching frequently accessed inference paths and intermediate results allows for direct return of results from the cache when the same or similar inference requests are encountered again, avoiding duplicate calculations. The caching mechanism employs an LRU or LFU strategy. Specifically, this includes: 1. Cache Key Design: Design a cache key to uniquely identify inference requests; the cache key includes: a globally unique code for the teacher node; a hash value for the inference rule template; a time window parameter; and a pruning strategy parameter; 2. Caching: Store the inference results in the cache; specifically, this includes: storing the complete inference results, including the set of implicit relations and contribution weights; storing intermediate calculation results, such as embedding vectors and attention weights; using memory caching or distributed caching for storage. 3. Cache Query: When a new inference request is received, the cache is queried first; specifically, this includes: checking if a matching result exists in the cache based on the cache key; if it exists and has not expired, the cached result is returned directly; if it does not exist or has expired, the complete inference process is executed. 4. Cache Update and Eviction: Cache management employs LRU (Least Recently Used) or LFU (Least Frequently Used) strategies; specifically including: LRU strategy: eviction of the least recently used cache item; LFU strategy: eviction of the least frequently used cache item; setting cache capacity and expiration time to prevent cache from consuming excessive resources; 5. Cache Hit Rate Monitoring: Monitor cache hit rate and evaluate caching effectiveness; specifically including: counting cache hits and total requests; calculating cache hit rate and optimizing caching strategies; adjusting cache capacity and eviction policies based on hit rate.

[0041] Through the above technical solution, this embodiment represents various potential association paths that the teacher to be assessed may have through candidate relationship chains, such as establishing connections with broader introduction and training events through intermediate entities such as students and projects. To improve the accuracy of the inference results, a multi-source confidence fusion strategy is further introduced. This strategy evaluates the confidence of each constructed candidate relationship chain, comprehensively considering multiple factors such as path length, entity type, and relationship strength, and fusing evidence from different information sources. By filtering candidate relationship chains with confidence scores below a preset confidence threshold, potential associations with insufficient evidence or low reliability are effectively eliminated, resulting in a more refined and reliable set of filtered candidate relationship chains. Subsequently, this application uses a graph embedding model to conduct in-depth analysis of the filtered candidate relationship chains. The graph embedding model maps entities and relationships in the knowledge graph to a low-dimensional vector space, thereby capturing the semantic and structural information between them. Based on these vector representations, the model calculates the probability of completing each missing link in the candidate relationship chain, that is, assessing the possibility of the existence of a relationship or entity that is not explicitly expressed. Finally, the system generates the corresponding implicit relationship only when the probability of completing the missing link exceeds a preset probability threshold. These implicit relationships were subsequently added to the talent attraction and development knowledge graph, thereby enriching the connectivity of the knowledge graph and making the semantic completion of the target talent attraction and development event messages more comprehensive. In this way, this application discovers and utilizes indirect contributions and associations that do not exist directly in the original messages, solving the problem that relying solely on explicit relationships may lead to incomplete semantic completion, and improving the accuracy and comprehensiveness of talent attraction and development workload accounting.

[0042] In one embodiment, in order to fully capture the deep contributions of teachers through talent attraction and development event vectors, the construction of contextual associations for the target talent attraction and development event messages based on a talent attraction and development knowledge graph that integrates the implicit relationships includes: The implicit relationship is added as a new triple to the talent attraction and development knowledge graph; Taking the teacher node to be accounted for as the center, extract the first-degree related subgraph corresponding to the first-degree neighbor node from the added talent introduction and cultivation knowledge graph. The first-degree neighbor node has a direct relationship with the teacher node to be accounted for. Based on the first-degree related subgraph, the second-degree related subgraph corresponding to the second-degree neighbor node is extracted from the added talent attraction and cultivation knowledge graph. The second-degree neighbor node has an indirect relationship with the teacher node to be calculated. The first-degree association subgraph and the second-degree association subgraph are integrated to obtain the context association.

[0043] In this embodiment, adding implicit relationships as new triples to the talent attraction and development knowledge graph means formally incorporating logically existing but inferred connections into the knowledge graph in the form of entity-relation-entity (triples). This aims to enrich the structure and content of the knowledge graph, making it contain more comprehensive information and providing a broader data foundation for subsequent contextual association construction. This is achieved through graph database insertion operations, such as using the CREATE statement in Neo4j to add new nodes and relationships, or adding new triple declarations in RDF storage. Another approach is to build an incremental update mechanism to import the inferred implicit relationships into the knowledge graph in batches and update the index.

[0044] Centered on the teacher node to be evaluated, extracting the first-degree neighbor subgraph from the added talent attraction and development knowledge graph involves identifying and extracting all entity nodes directly connected to the teacher node and their relationships. A first-degree neighbor node is a node directly connected to the teacher node, and the first-degree subgraph contains the teacher node, its direct neighbors, and their relationships. The purpose of extracting the first-degree subgraph is to capture the most direct and core talent attraction and development activities and related entities, forming its core talent attraction and development circle. This is done by using graph traversal algorithms (such as Breadth-First Search (BFS) or Depth-First Search (DFS)) starting from the teacher node and searching for all nodes with a distance of 1 and their connecting edges. For example, in a graph database, a query like `MATCH (t:Teacher)-[r]-(n)WHERE t.id = 'Teacher ID' RETURN t, r, n` can be executed. Another approach is to use graph computing frameworks (such as Apache Giraph or GraphX) for single-hop neighbor queries.

[0045] Based on the first-degree-of-factory subgraph, extracting the second-degree-of-factory subgraph corresponding to the second-degree-of-factory neighbor nodes from the added talent attraction and development knowledge graph refers to further identifying entity nodes and their relationships that are directly connected to these first-degree-of-factory neighbor nodes but indirectly connected to the teacher node to be assessed, based on the already identified first-degree-of-factory neighbor nodes. Second-degree-of-factory neighbor nodes are nodes indirectly connected to the teacher node to be assessed (connected through first-degree-of-factory neighbor nodes). The second-degree-of-factory subgraph further expands the context scope, capturing the indirect influence and contribution of the teacher to be assessed. This helps to comprehensively evaluate the teacher attraction and development performance, as the impact of many attraction and development efforts is indirect. This involves starting from each node in the set of first-degree-of-factory neighbor nodes, performing a graph traversal again, and finding all nodes with a distance of 1 (relative to the first-degree-of-factory neighbor nodes), but these nodes cannot be the teacher node to be assessed itself. For example, executing a query like MATCH (t:Teacher)-[r1]-(n1)-[r2]-(n2) WHERE t.id = 'Teacher ID to be assessed' AND n2<>t RETURN t, r1, n1, r2, n2 in the graph database. Another approach is to proceed in two steps: first, obtain the first-degree neighbors, and then, starting from these first-degree neighbors, obtain their first-degree neighbors again (excluding the teachers to be accounted for).

[0046] Integrating a first-degree-of-connection subgraph with a second-degree-of-connection subgraph to obtain a contextual connection involves performing a union operation on the node and edge sets of the two subgraphs to form a more comprehensive and richer view of the teacher recruitment context. This integrated graph structure includes both direct and indirect teacher recruitment activities, providing comprehensive input for subsequent recruitment event vector generation. Since the second-degree-of-connection subgraph is an extension of the first-degree-of-connection subgraph, there will be overlapping nodes and edges between them (i.e., first-degree neighbor nodes and edges connecting them). Therefore, the integration process mainly involves adding second-degree neighbor nodes and edges connecting first-degree neighbors to the first-degree-of-connection subgraph, forming a local subgraph containing the teacher to be evaluated, their direct neighbors, and indirect neighbors. In a graph database, this is achieved by combining query results or constructing a new graph object in memory.

[0047] This application's approach first formally incorporates the implicit relationships derived through reasoning into the talent attraction and development knowledge graph, thereby expanding the information dimensions of the knowledge graph. Subsequently, focusing on the teacher to be assessed, it extracts their directly related first-degree neighbor nodes and indirectly related second-degree neighbor nodes hierarchically, constructing first-degree and second-degree relationship subgraphs respectively. The first-degree relationship subgraph focuses on the teacher's core attraction and development activities and direct impact, while the second-degree relationship subgraph further reveals the teacher's indirect contributions and broader influence through their directly related entities. Finally, by integrating these two subgraphs, this approach constructs a comprehensive, structured, and teacher-centric contextual relationship. This hierarchical and progressive construction method ensures that the formed contextual relationships not only contain explicit information but also fully utilize implicit information, thus providing a solid foundation for subsequently generating more accurate and representative attraction and development event vectors.

[0048] Through the aforementioned technical solution, this application effectively integrates the implicit relationships derived from reasoning into the context of teacher recruitment and development performance evaluation, and systematically constructs direct and indirect recruitment and development event relationships centered on the teacher to be evaluated. This not only compensates for the potential limitations of the original messages and explicit relationships, but more importantly, it captures the deep-seated, non-explicit contributions of teachers in talent recruitment and development work, such as the indirect influence generated through guiding students to participate in projects and students publishing papers. This comprehensive and structured contextual relationship provides a solid foundation for subsequently generating more accurate and representative recruitment and development event vectors, thereby making the evaluation of teachers' talent recruitment and development performance more precise and fair, and avoiding evaluation bias caused by incomplete information.

[0049] In one embodiment, to accurately identify and construct a first-degree-of-first-order relational subgraph directly related to the teacher node to be evaluated, and to ensure the accuracy and completeness of the contextual association, the step of extracting the first-degree-of-first-order relational subgraph corresponding to the first-degree-of-first-order neighbor nodes from the added talent attraction and development knowledge graph, centered on the teacher node to be evaluated, includes: Starting from the teacher node to be evaluated, traverse all outgoing and incoming edges of the teacher node to be evaluated, and determine the first-degree neighbor node directly connected to the teacher node to be evaluated; Extract the edges between the teacher node to be evaluated and all its first-degree neighbors to obtain the set of direct relations; direct relations include mentorship, participation, membership, and referral relations. The set of teacher nodes to be accounted for, the set of first-degree neighbor nodes, and the set of direct relationships are combined to construct the first-degree association subgraph; In this embodiment, the teacher node to be evaluated refers to the entity representing the teacher to be evaluated in the talent introduction and development knowledge graph. This node is the center for constructing the first-degree-of-fact subgraph and the starting point for all subsequent relation traversals and subgraph extraction. It is a node with a unique identifier, such as the teacher's employee ID or a globally unique teacher code, as its unique ID in the graph. Traversing all outgoing and incoming edges of the teacher node to be evaluated means that the system explores all direct connections of the teacher node to be evaluated in the knowledge graph. This is achieved through the query language of the graph database (such as using the Cypher query statement to find the direct neighbors of a node in Neo4j) or through a programmatic graph traversal algorithm (such as a one-level traversal of breadth-first search or depth-first search). Determining the first-degree-of-fact neighbor nodes directly connected to the teacher node to be evaluated means identifying other entity nodes directly connected to the teacher node to be evaluated through an edge during the traversal process. These nodes represent entities with the most direct association with the teacher to be evaluated and are the core components for constructing the first-degree-of-fact subgraph. Extracting the edges between the teacher node to be evaluated and all its first-degree neighbors to obtain the set of direct relations means collecting all edges and their types connecting the teacher node to be evaluated and these first-degree neighbors. These edges clarify the specific relationship types between the teacher to be evaluated and its first-degree neighbors. Direct relations include mentorship, participation, membership, and introduction; these are predefined relationship types used to describe the direct connections between teachers and entities such as students, projects, and teams, thus defining the scope and semantics of the first-degree association subgraph. Combining the teacher node to be evaluated, the set of first-degree neighbors, and the set of direct relations to construct the first-degree association subgraph means integrating the central node, its direct neighbors, and the edges connecting them to form a local graph structure. This is represented by constructing a subgraph data structure (such as an adjacency list or adjacency matrix containing specific nodes and edges).

[0050] This application's scheme systematically traverses all direct outgoing and incoming edges from the teacher node to be evaluated, thereby comprehensively identifying all directly connected first-degree neighbor nodes. Based on this, all edges between the teacher node to be evaluated and these first-degree neighbor nodes are explicitly extracted and categorized into a set of direct relationships, with specific types clearly defined such as mentorship, participation, membership, and referral relationships. Finally, the teacher node to be evaluated, the set of first-degree neighbor nodes, and the set of direct relationships are combined to construct a complete first-degree relational subgraph. This explicit and structured extraction process ensures that referral and mentoring event information directly related to the teacher to be evaluated is comprehensively and accurately captured, avoiding information omissions or inaccuracies. In this way, the constructed contextual relationships more accurately reflect the teacher's direct contributions and connections, providing a more solid and reliable foundation for the subsequent generation of referral and mentoring event vectors.

[0051] Through the above technical solutions, this application ensures that the construction process of the first-degree-of-connection subgraph is more systematic and accurate, avoiding potential omissions or biases when extracting direct teacher-related information. This helps improve the completeness and reliability of the graph-based recruitment and training event data, enabling the recruitment and training event vectors subsequently generated based on the graph neural network model to more accurately represent the recruitment and training workload of the teachers to be accounted for. By clarifying the types of direct relationships such as guidance relationships, participation relationships, affiliation relationships, and introduction relationships, the identification of teachers' direct contributions becomes more accurate, providing a more solid and reliable data foundation for the subsequent generation of recruitment and training event vectors and performance accounting, thereby improving the accuracy and fairness of the performance management method.

[0052] In one embodiment, to avoid relying solely on the subgraph formed by first-degree neighbor nodes directly associated with the teacher to be accounted for, which might fail to fully capture the indirect contributions of teachers in the talent attraction and development process, leading to an incomplete and inaccurate assessment of teachers' actual workload and thus low accounting accuracy, the step of extracting the second-degree neighbor node corresponding to the second-degree neighbor node from the added talent attraction and development knowledge graph based on the first-degree neighbor subgraph includes: Starting from each neighbor node in the set of first-degree neighbor nodes, traverse the outgoing and incoming edges of each neighbor node to determine the second-degree neighbor nodes indirectly connected to the teacher node to be calculated. Obtain the edges connecting the first-degree neighbor node set and the second-degree neighbor node set to obtain the indirect relationship set; the indirect relationship is used to represent the indirect contribution of the teacher to be accounted for. The indirect relationships mentioned above include student-published papers, student-participated projects, projects led by introduced talents, and project funding received. The second-degree related subgraph is constructed by combining the first-degree related subgraph, the set of second-degree neighbor nodes, and the set of indirect relationships.

[0053] In this embodiment, to identify nodes that are not directly connected to the teacher node to be evaluated but are connected through first-degree neighbors (i.e., second-degree neighbors), the system starts from each neighbor node in the set of first-degree neighbors and traverses the outgoing and incoming edges of each neighbor node to determine the second-degree neighbors indirectly connected to the teacher node to be evaluated. These second-degree neighbors represent the indirect influence range of the teacher to be evaluated. For example, a breadth-first search (BFS) or depth-first search (DFS) can be performed on each node in the set of first-degree neighbors, starting from that first-degree neighbor node, exploring its directly connected nodes, and marking these nodes as second-degree neighbors, while ensuring that these second-degree neighbors do not include the teacher node to be evaluated itself or any already identified first-degree neighbors. Alternatively, this can be done efficiently using the query language of a graph database, by specifying the node reached from a first-degree neighbor node via an edge, thereby quickly identifying second-degree neighbors.

[0054] After identifying second-degree neighbors, it is necessary to clarify the specific relationships connecting first-degree and second-degree neighbors and categorize these relationships into an indirect relationship set. These indirect relationships are key elements for measuring the indirect contributions of the faculty to be accounted for. One approach is to record the specific edges (i.e., relationship type and direction) connecting first-degree and second-degree neighbors while identifying them, and collect these edges and their attributes to form an indirect relationship set. Another approach is to construct the indirect relationship set by traversing all edges connecting nodes in these two sets in the knowledge graph and filtering out edges that meet the criteria after obtaining the sets of first-degree and second-degree neighbors. These indirect relationships specifically include student-published papers, student-participated projects, projects led by introduced talents, and project funding received. "Student-published papers" refers to academic papers published by students supervised by the faculty to be accounted for, whose output indirectly reflects the faculty's training achievements. "Student-participated projects" refers to research projects participated in by students supervised by the faculty to be accounted for, whose progress and results indirectly reflect the faculty's contributions in project guidance and team building. "Projects led by introduced talents" refers to research projects led by talents introduced or assisted by the faculty member to be accounted for. The successful implementation of these projects indirectly reflects the faculty member's role in talent introduction and team expansion. "Funding received for projects" refers to research funding obtained for projects indirectly related to the faculty member to be accounted for, such as funding obtained for projects led by their students or introduced talents. This indirectly reflects the faculty member's contribution in resource acquisition and academic influence. Finally, the identified first-degree related subgraph, the set of second-degree neighbor nodes, and the set of indirect relationships connecting them are integrated to form a more comprehensive and richer second-degree related subgraph. This is based on the existing first-degree related subgraph, adding newly identified second-degree neighbor nodes as new nodes to the subgraph, and adding edges from the set of indirect relationships as new edges to the subgraph, thereby expanding to form a second-degree related subgraph. Alternatively, all nodes and edges in the first-degree related subgraph, the set of second-degree neighbor nodes, and the set of indirect relationships can be collected first, and then a new graph structure, i.e., a second-degree related subgraph, can be reconstructed from these elements using graph construction tools or libraries.

[0055] In constructing the contextual association of target talent attraction and development event messages, this application first uses a talent attraction and development knowledge graph that incorporates implicit relationships. Centered on the node corresponding to the teacher to be assessed, it extracts the first-degree neighbor subgraphs directly associated with that teacher. Building upon this, to more comprehensively capture the teacher's workload in talent attraction and development, particularly their indirect contributions, this application further traverses the outgoing and incoming edges of each neighbor node in the first-degree neighbor node set to identify the second-degree neighbor nodes indirectly connected to the teacher node to be assessed. These second-degree neighbor nodes represent entities that establish connections with the teacher to be assessed through first-degree neighbor nodes, such as students mentored by the teacher's students, or projects participated in by the teacher's collaborators. Subsequently, the system obtains the edges connecting the first-degree neighbor node set and the second-degree neighbor node set, forming a set of indirect relationships. These relationships explicitly represent the indirect contributions of the teacher to be assessed, such as papers published by students, projects participated in by students, projects led by the introduced talent, and project funding received. Finally, the constructed first-degree-of-fact subgraph, the newly identified set of second-degree-of-fact neighbor nodes, and the set of indirect relationships are combined to construct a more complete second-degree-of-fact subgraph. In this way, this application not only considers teachers' direct talent development activities but also delves into their indirect influence through students, collaborators, and introduced talent. This makes the constructed contextual relationships more comprehensive and accurate in reflecting teachers' actual workload and contributions in talent development, thus solving the problem of incomplete performance evaluation caused by relying solely on direct relationship information.

[0056] Through the aforementioned technical solution, this application extracts a second-degree-of-grained subgraph from the talent attraction and development knowledge graph, containing the indirect contributions of teachers to be assessed. This allows for the construction of contextual relationships in target talent attraction and development event messages, considering not only teachers' direct attraction and development activities but also delving into their indirect influences through student mentoring, team collaboration, and resource introduction. For example, the achievements of students mentored by teachers and the progress of projects led by talents introduced by teachers are all included in the performance evaluation. This effectively solves the problem of incomplete assessment of teachers' talent attraction and development workload and inaccurate performance accounting results due to relying solely on direct relational information. Therefore, the accounting of teachers' talent attraction and development performance is more comprehensive, objective, and accurate, more realistically reflecting teachers' actual contributions to talent attraction and development.

[0057] In one embodiment, to transform complex graph-based recruitment event data into a vector representation that comprehensively characterizes teachers' recruitment workload, the step of generating the recruitment event vector of the teacher to be accounted for based on the graph neural network model and the graph-based recruitment event data includes: The graph-based mentoring event data is converted into a graph structure containing preset mentoring entity nodes and preset mentoring event edges. The mentoring entity nodes include teacher nodes, event nodes, student nodes, and project nodes. The mentoring event edges include guidance relationships, participation relationships, and membership relationships. Each node's features are initialized. The features of the teacher node include age, professional title, subject area, and historical performance score; the features of the event node include event type, event level, occurrence of the event, and ongoing event; the features of the student node include year of enrollment, major, and training level; and the features of the project node include project level, funding amount, and research period. Based on graph convolutional networks or graph attention networks, the initialized nodes are subjected to neighbor node feature aggregation and feature update to obtain a node vectorization representation of each node. The node vectorization representation is used to vectorize the self-features and local neighborhood information of each node. From the node vectorization representation of each node, extract the teacher feature vector of the teacher to be accounted for and the event feature vector used to characterize the workload of the teacher to be accounted for in recruitment and training, as the recruitment and training event vector.

[0058] In this embodiment, converting the graph-based talent attraction and development event data into a graph structure containing preset talent attraction and development entity nodes and preset talent attraction and development event edges means constructing a standardized graph data structure from the graph-based data generated in the preceding steps, which has undergone contextual association and semantic completion, according to pre-defined entity types and relation types. Talent attraction and development entity nodes are the basic units in the graph, representing key participants or objects in talent attraction and development activities. For example, teacher nodes represent individual teachers participating in talent attraction and development work, event nodes represent specific talent attraction and development activities or achievements, student nodes represent students being trained, and project nodes represent research or teaching projects. Talent attraction and development event edges represent the interactions or connections between these entity nodes. For example, a mentoring relationship represents a teacher's guidance to a student, a participation relationship represents a teacher or student participating in a project, and a membership relationship represents a student or project belonging to a particular teacher or team. This conversion is achieved by directly mapping the triples (entity-relationship-entity) in the graph-based data to the nodes and edges of the graph, or by constructing data structures such as adjacency matrices and adjacency lists to represent the graph.

[0059] Initializing the features of each node refers to assigning initial numerical attribute representations to each entity node in the graph structure. These features are a quantitative representation of the inherent attributes of the node itself, providing basic information for subsequent graph neural network processing. For example, the features of a teacher node include its age, professional title, academic discipline, and past performance scores, reflecting the teacher's qualifications and professional background; the features of an event node include the type of event (e.g., paper publication, project approval), level (e.g., national, provincial), occurrence time, and duration, describing the nature and scale of the event; the features of a student node include the year of enrollment, major, and training level (e.g., undergraduate, master's, doctoral), reflecting the student's training stage and professional background; the features of a project node include the project level, funding amount, and research period, describing the scale and importance of the project. These features are converted into initial vector representations through one-hot encoding, word embedding, or numerical normalization.

[0060] Based on graph convolutional networks (GCNs) or graph attention networks (GATs), the aggregation and updating of neighbor node features of initialized nodes refers to iteratively processing the node features in a graph structure using the powerful capabilities of GNNs. GCNs update node representations by aggregating the features of each node's neighbors and performing nonlinear transformations based on the node's own features. GATs further introduce an attention mechanism, allowing the model to assign different weights to different neighbor nodes, thus more flexibly capturing the importance of neighbor information. Through multiple layers of such aggregation and update operations, the vectorized representation of each node not only includes its own initial features but also incorporates information from its local neighborhood (i.e., directly or indirectly connected entities), thus comprehensively reflecting the contextual semantics of the node in the graph.

[0061] From the node vectorization representations of each node, the teacher feature vector and event feature vector representing the teacher's recruitment workload are extracted as recruitment event vectors. This means that after the graph neural network processing, specific vectors related to the teacher are identified and extracted from the final vectorization representations of all nodes. The teacher feature vector directly extracts the vectorization representation corresponding to the teacher node, containing the teacher's attributes and its local structural information in the graph. The event feature vector represents various events related to the teacher's recruitment workload. It is obtained by aggregating (e.g., summing, averaging, or max pooling) the vectors of all event nodes (e.g., project nodes, student nodes, paper nodes, etc.) directly or indirectly related to the teacher, or by selecting only the event node vectors directly related to the teacher. Finally, these extracted teacher feature vectors and event feature vectors are combined (e.g., concatenated or weighted summation) to form a comprehensive recruitment event vector, which comprehensively and accurately represents the teacher's recruitment workload.

[0062] This application's approach transforms graph-based talent recruitment and development event data into a standardized graph structure and initializes rich features for each node, laying the foundation for graph neural network processing. Subsequently, graph convolutional networks or graph attention networks are used to aggregate and update node features with their neighbors, enabling each node's vectorized representation to integrate its own attributes and contextual information within the complex relational network. This mechanism allows teacher node vectors to not only reflect their own characteristics but also indirectly reflect the features of related entities such as the students they guide and the projects they participate in, thus comprehensively capturing the teacher's recruitment and development workload. Finally, by extracting the teacher's own feature vector and the aggregated event feature vector, a comprehensive recruitment and development event vector is formed. This vector, in a high-dimensional, dense numerical form, accurately quantifies the teacher's multi-dimensional contribution to talent recruitment and development.

[0063] Through the aforementioned technical solution, this application transforms complex, multi-relational talent attraction and development knowledge graph data into a structured, computable graph representation. Graph neural networks are then used to deeply aggregate and update node features, generating a comprehensive, accurate, and context-rich talent attraction and development event vector. This vector not only includes the teacher's own attributes but also incorporates multi-dimensional information from their student guidance, project participation, and output activities, effectively solving the problem of traditional methods' difficulty in comprehensively quantifying teachers' talent attraction and development workload. This refined vector representation provides high-quality input for subsequent calculations of talent attraction and development goal completion, improving the accuracy of talent attraction and development performance evaluation in universities.

[0064] Reference Figure 2 , Figure 2 A flowchart illustrating another knowledge graph-based performance management method provided in this application.

[0065] In this embodiment, step S104 specifically includes: Step S1041: Encode each rule in the breeding target rule base into a rule feature vector, and calculate the matching degree between the breeding event vector and each rule feature vector; Step S1042: Based on the preset matching degree threshold and each matching degree, determine the number of recruitment and training events completed by the teacher to be accounted for; Step S1043: Calculate the completion rate of the breeding target based on the total number of breeding events and the number of breeding events completed.

[0066] In this embodiment, encoding each rule in the talent attraction and development target rule base into a rule feature vector means converting a predefined series of standards or requirements used to measure the performance of teacher talent attraction and development into a machine-processable numerical vector form. These rules are qualitative, such as "supervising a doctoral student," and quantitative, such as "publishing an SCI paper with an impact factor greater than 3." In terms of implementation, Natural Language Processing (NLP) technology is used to parse the rule text, extracting key information such as keywords, entities, and numerical values, and mapping them to predefined feature dimensions to form a vector representation of the rule. For example, the rule "supervising a doctoral student" is encoded as a vector containing features such as "supervising," "doctoral student," and "quantity." Another approach is to use pre-trained language models, such as BERT or Word2Vec, to directly embed the rule text into a high-dimensional vector space, generating semantically rich rule feature vectors to capture the deeper meaning of the rules. The matching degree between the talent attraction and development event vector and each rule feature vector is calculated to quantify the degree of conformity between the teacher's actual talent attraction and development activities (represented by the talent attraction and development event vector) and each talent attraction and development target rule (represented by the rule feature vector). This is achieved through various similarity calculation methods, such as using cosine similarity to measure the directional similarity between two vectors; the closer the value is to 1, the higher the matching degree. Alternatively, Euclidean distance is calculated and converted into a matching degree score; the smaller the Euclidean distance, the higher the matching degree. Another approach is to design a small neural network layer that takes the introduction / recruitment event vector and rule feature vector as input and outputs a matching degree score. Based on a preset matching degree threshold and each matching degree, the number of introduction / recruitment events completed by the teacher to be evaluated is determined. Its purpose is to objectively quantify how many introduction / recruitment goals the teacher has actually completed based on the matching degree results. One implementation is that for each rule in the introduction / recruitment goal rule base, if its matching degree with the introduction / recruitment event vector exceeds a preset matching degree threshold, then the introduction / recruitment event corresponding to that rule is considered completed, and the completion count is accumulated. Another approach is to assign different weights to different rules; when the matching degree exceeds the threshold, the weight of the corresponding rule is added to the completion count to reflect the importance of different rules. Based on the total number of mentoring events and the number of completed mentoring events, the completion rate of the mentoring goal is calculated, aiming to provide a quantitative indicator of the overall completion status of the teacher's mentoring goal. This is usually achieved through a simple ratio calculation, that is, dividing the determined number of completed mentoring events by the total number of all rules in the mentoring goal rule base. If the rules have weights, a weighted average method is used, that is, the completion rate is equal to the sum of the weights of all completed rules divided by the sum of the weights of all rules.

[0067] This application's solution encodes each rule in the talent attraction target rule base into a rule feature vector, enabling a quantitative comparison between these rules and the talent attraction event vectors of the teachers to be evaluated. By calculating the matching degree between the talent attraction event vectors and the rule feature vectors, the system accurately measures the degree to which the teachers' actual talent attraction work conforms to the preset targets. Subsequently, based on a preset matching degree threshold, the system automatically identifies and counts the number of talent attraction events completed by the teachers. Finally, the ratio of the number of completed events to the total number of events is used to objectively calculate the talent attraction target completion rate. This series of steps transforms the teachers' complex talent attraction activity data into quantifiable performance indicators, achieving automation, standardization, and refinement from raw event data to performance evaluation. It effectively solves the problems of strong subjectivity and low efficiency in traditional performance evaluation and fully utilizes the rich information contained in the talent attraction event vectors.

[0068] Through the above technical solution, this application efficiently and accurately matches and quantifies complex talent attraction and development goal rules with teachers' talent attraction and development event vectors, thereby achieving automated calculation of the completion rate of talent attraction and development goals. This not only improves the efficiency and objectivity of performance accounting and avoids subjective biases that may be caused by manual evaluation, but also makes full use of the rich and multi-dimensional information contained in the talent attraction and development event vectors, thereby improving the accuracy of performance accounting results.

[0069] In one embodiment, a construction process is further provided, the construction process including: The knowledge graph ontology definition is a formal description of the entity types, relation types, and attributes of a knowledge graph. Entity types include teachers, students, teams, colleges, disciplines, projects, papers, patents, conferences, journals, awards, etc. These types are categorized based on the characteristics of talent recruitment and development performance management in universities, aiming to comprehensively cover all subjects and objects related to teacher recruitment and development. In addition to the types listed above, further refinement is made according to actual needs. For example, "teacher" is subdivided into "professor," "associate professor," "lecturer," etc., or entity types such as "course" and "textbook" are added to more precisely describe recruitment and development activities. Relationship types include mentoring relationships (mentor, mentored), affiliation relationships (belong to, contain), participation relationships (participate, organize), publication relationships (publish, published), award relationships (receive, awarded), and citation relationships (cite, cited). These relationship types quantify various interaction patterns between teachers and students, teachers and projects, teachers and achievements, etc. In addition to these explicit relationships, implicit relationships are defined, such as "collaboration relationships" (inferred through joint project participation or published papers) or "influence relationships" (inferred through citation chains) to enrich the expressive power of the knowledge graph. Attributes include teachers (name, title, degree, research direction), students (name, student ID, major, year of enrollment), projects (name, level, funding, start and end dates), and papers (title, journal, publication date, impact factor). These attributes provide detailed descriptive information about entities, helping to distinguish different entity instances and supporting more complex queries and analyses. Beyond these basic attributes, more attributes are added based on practical application scenarios, such as teachers' "year of employment" and "contact information," students' "educational level" and "graduation destination," project "project leader" and "participants," and papers' "author list" and "keywords." The role of the knowledge graph ontology definition is to provide a unified and structured framework for the construction of the knowledge graph, ensuring the consistency and standardization of knowledge representation. In practical applications, ontology languages ​​(such as OWL and RDF Schema) are used for definition. By clarifying the hierarchical structure, attribute constraints, and relational semantics between entities, the foundation for subsequent knowledge extraction, fusion, and reasoning is laid. Another approach is to manually construct the ontology model through interviews with domain experts and analysis of existing data patterns, and then use ontology editing tools for visual editing and management.

[0070] The construction process also includes entity extraction, which involves extracting entity information from the identity-aligned event messages. Named Entity Recognition (NER) technology is used to identify entities such as teacher names, student names, team names, college names, subject names, and project names. The role of entity extraction is to transform the textual information in the original event messages into structured entity data, providing basic elements for the construction of the knowledge graph. Entity extraction is implemented using various techniques. In addition to methods based on rule matching, dictionary matching, and machine learning (such as BiLSTM-CRF or BERT models), pre-trained language models based on deep learning (such as RoBERTa and XLNet) are also used for entity recognition. These models demonstrate stronger capabilities in handling complex contexts and long texts.

[0071] The construction process also includes relation extraction, which involves extracting relationships between entities from identity-aligned event messages. The purpose of relation extraction is to connect discrete entities to form a meaningful knowledge structure. Relation extraction employs various techniques, such as rule-based relation extraction, dependency parsing-based relation extraction, and relation classification based on deep learning (e.g., BERT or Transformer models). It also utilizes distant supervision methods, automatically labeling training data using existing knowledge bases to train the relation extraction model on large-scale corpora.

[0072] Furthermore, the construction process also includes knowledge fusion, which involves integrating knowledge extracted from different data sources. The purpose of knowledge fusion is to eliminate data heterogeneity and improve the completeness and accuracy of the knowledge graph. Knowledge fusion includes entity alignment, relation alignment, and attribute fusion. Entity alignment, in addition to similarity-based or graph-based methods, also employs clustering methods to group similar entities together for manual or semi-automatic verification. Relation alignment, besides merging relations, also uses ontology mapping or semantic matching techniques to unify relations representing the same semantics but with different names from different data sources. Attribute fusion, in addition to voting, confidence-weighted, or timestamp-based recent update methods, also uses machine learning methods to train models to determine the priority or credibility of different attribute values. Knowledge fusion also includes graph embedding-based entity alignment, using graph embedding models such as TransE and TransR to learn vector representations of entities and relations, and performing entity alignment through vector similarity calculation. The purpose of graph embedding-based entity alignment is to utilize graph structure information for entity alignment, improving the accuracy and efficiency of alignment. In addition to models like TransE and TransR, more advanced graph embedding models such as RotatE and ComplEx are employed. These models excel in handling complex relationship types and multi-hop relationships. Furthermore, knowledge fusion includes entity alignment based on active learning. For entities that are difficult to align, an active learning strategy is used to select the most informative samples for manual annotation, improving alignment accuracy. The purpose of active learning-based entity alignment is to maximize the performance of the entity alignment model with limited manual annotation costs. Besides selecting the most informative samples for manual annotation, strategies such as uncertainty sampling, density sampling, or diversity sampling are used to select samples requiring manual annotation, further improving model performance. Moreover, knowledge fusion also includes entity alignment based on graph neural networks. Graph neural networks are used to learn entity representations, and entity alignment is performed through graph matching algorithms. The purpose of graph neural network-based entity alignment is to leverage the powerful feature learning capabilities of graph neural networks to extract entity features from complex graph structures, thereby achieving more accurate entity alignment. In addition to graph matching algorithms, attention-based graph neural network models, such as GAT (Graph Attention Network), are used to better capture the importance of entity neighbor information, thus improving the accuracy of entity alignment.

[0073] The construction process includes knowledge storage, which involves storing the constructed knowledge graph in a graph database, such as Neo4j, JanusGraph, or TigerGraph, to support efficient graph querying and computation. The role of knowledge storage is to provide efficient knowledge management and retrieval capabilities. Besides graph databases like Neo4j, JanusGraph, or TigerGraph, RDF triple storage (such as Virtuoso or AllegroGraph) or columnar storage-based graph databases (such as HugeGraph) are also used. Specifically, this embodiment first defines the entity types, relation types, and attributes related to talent attraction and development through a knowledge graph ontology definition. This sets a unified semantic standard for subsequent data extraction and integration, ensuring the internal consistency of the knowledge graph. Based on this, the entity extraction and relation extraction steps identify key entities and their relationships from the original event messages, transforming unstructured text information into structured triple forms. Subsequently, the knowledge fusion step integrates knowledge from different data sources that may contain heterogeneity or redundancy. Through techniques such as entity alignment, relation alignment, and attribute fusion, data conflicts are eliminated, improving the completeness and accuracy of the knowledge graph. In particular, by introducing entity alignment methods based on graph embedding, active learning, and graph neural networks, entity recognition and matching problems in complex scenarios are handled more effectively, improving fusion accuracy. Finally, the constructed knowledge graph is stored in a graph database, ensuring efficient knowledge management and retrieval. Through the above construction process, the resulting talent attraction and development knowledge graph possesses higher quality and richer semantic information. When this knowledge graph is applied to the aforementioned performance management methods, performing contextual association and semantic completion on target talent attraction and development event messages, it more accurately identifies entities and relationships within the events and performs reasoning and completion based on the rich background knowledge in the graph, thereby generating more accurate and comprehensive graph-based talent attraction and development event data. This high-quality graph-based data will directly improve the accuracy of subsequent graph neural network models generating talent attraction and development event vectors, thus enhancing the accuracy and reliability of teacher talent attraction and development performance accounting.

[0074] Through the methods described above, this embodiment comprehensively covers all aspects of knowledge graph construction, from ontology definition, entity extraction, relation extraction, knowledge fusion to knowledge storage. It also incorporates various machine learning and deep learning techniques, particularly entity alignment methods based on graph embedding, active learning, and graph neural networks, thereby improving the quality and accuracy of the knowledge graph construction. This provides an accurate knowledge foundation for subsequent contextual association and semantic completion, ensuring that the generated graph-based talent recruitment and development event data more accurately reflects teachers' recruitment and development workload, ultimately leading to more precise results in the performance evaluation of teacher talent recruitment and development.

[0075] In one embodiment, a further step is proposed: before obtaining the target recruitment and development event message representing the workload of teachers to be assessed, firstly, original recruitment and development event data representing the workload of teachers to be assessed is obtained, and the original recruitment and development event data is converted into an initial recruitment and development event message; subsequently, the identity identifiers of teachers to be assessed in the initial recruitment and development event message are mapped across systems to generate a target global teacher code for the teachers to be assessed, and based on the target global teacher code, the initial recruitment and development event message is aligned to obtain the target recruitment and development event message. Specifically, before obtaining the target recruitment and development event message representing the workload of teachers to be assessed, talent recruitment and development event data is collected in real time from multiple heterogeneous business systems, and the collected original event data is converted into standard event messages; the business systems include a graduate system, a human resources system, a research system, and a financial system. Next, the standard event message is received, and the teacher identity identifiers therein are mapped across systems to generate a globally unique teacher code for each teacher, which is then written into the message, and the identity-aligned event message is output as the target recruitment and development event message. Specifically, the cross-system mapping process based on teacher identity identifiers includes: first, identity identifier extraction, i.e., extracting teacher identity identifiers from standard event messages, including identifier fields such as employee ID, ID number, mobile phone number, and email address. Then, a precise matching query is performed, checking a preset teacher identity mapping table to see if a completely matching record exists. This teacher identity mapping table stores the identifier correspondence between teachers in different systems, such as employee ID in the personnel system, ID in the research system, account in the graduate student system, and ID number in the finance system. For records where no completely matching record is found, a fuzzy matching algorithm is used for intelligent association. This fuzzy matching algorithm includes similarity matching based on name pinyin, such as calculating the edit distance or cosine similarity of name pinyin; matching based on the last few digits of the ID number, such as extracting the last 4 or 6 digits of the ID number for matching; matching based on the last 4 digits of the mobile phone number, such as extracting the last 4 digits of the mobile phone number for matching; and matching based on the email username, such as extracting the username before the @ symbol in the email address for matching. After matching is completed, a confidence score is calculated for the matching results. The scoring criteria include the number of matched fields (more matched fields result in higher confidence); the weight of the matched fields (e.g., key fields like ID numbers and employee IDs have higher weight than others); matching similarity (higher similarity in fuzzy matching results in higher confidence); and historical matching records (if the teacher has previous successful matching records, the confidence score is increased accordingly). Matching results with a confidence score below a preset threshold trigger a manual review process, where an administrator confirms or corrects the match. After identity verification, a globally unique code is generated for each teacher. This code uses a UUID or Snowflake algorithm to generate a unique identifier.Next, the mapping relationship is updated by writing the confirmed mapping relationship into the teacher identity mapping table and updating the mapping relationship. Finally, event message enhancement is performed by writing the globally unique teacher code into the standard event message and outputting the identity-aligned event message. The fuzzy matching algorithm also includes similarity calculation based on combined features, which combines multiple features (such as name pinyin, last few digits of ID number, and last few digits of mobile phone number) to calculate a comprehensive similarity; a matching model based on machine learning, such as logistic regression, support vector machine, or neural network models, which trains the matching model based on historical matching data to predict the matching probability; and entity linking based on graph neural networks, which represents teacher entities as nodes in a graph structure, learns entity representations through graph neural networks, and calculates the similarity between entities. The generation of the globally unique teacher code also includes code structure design, such as using a segmented code structure, including fields such as school code, college code, teacher type, and serial number; code mapping relationship management, i.e., establishing a mapping relationship table between the globally unique code and each system identifier, supporting bidirectional queries; and code version control, i.e., supporting code version management and maintaining backward compatibility when coding rules change.

[0076] The above solution first collects raw recruitment event data in real time from multiple heterogeneous business systems and converts it into initial recruitment event messages in a unified format, laying the foundation for subsequent identity alignment. Then, by mapping teacher identity identifiers in the initial recruitment event messages across systems, a target global teacher code is generated. Based on this code, the messages are aligned to obtain the target recruitment event message. By combining precise matching and fuzzy matching strategies, the solution effectively addresses the problem of inconsistent teacher identity identifiers across different systems. Precise matching quickly identifies known correspondences, while fuzzy matching intelligently associates information such as name pinyin, ID number, mobile phone number, and email address, and ensures matching accuracy through a confidence scoring mechanism. Low-confidence results are subject to manual review, further improving data quality. The final globally unique teacher code provides a stable and unified identity identifier for each teacher, ensuring that all recruitment event data can be accurately attributed to the corresponding teacher. In this way, this application effectively integrates and unifies the teacher recruitment and development workload data scattered across different systems, providing a high-quality and complete data foundation for subsequent knowledge graph-based contextual association, semantic completion, and the generation of recruitment and development event vectors, thereby improving the accuracy of talent recruitment and development performance accounting.

[0077] Through the aforementioned technical solution, this application effectively addresses the issues of data fragmentation and inaccuracy caused by heterogeneous data sources and inconsistent identity identifiers in the performance management of talent recruitment and development in universities. By employing a refined cross-system identity mapping mechanism, including precise matching, multi-dimensional fuzzy matching, confidence scoring, and manual review, it ensures that all recruitment and development event data for each teacher across different business systems can be accurately aggregated under their unique identity identifier. This not only significantly improves the completeness and consistency of recruitment and development event data, providing high-quality input for subsequent knowledge graph construction and graph neural network analysis, but also fundamentally guarantees the accuracy of teacher talent recruitment and development workload calculation and performance evaluation, avoiding deviations caused by incorrect data attribution, thereby improving the overall accuracy of performance management calculations.

[0078] In one embodiment, this application further proposes collecting recruitment and training event data related to the teacher to be accounted for. Specifically, this includes collecting the teacher's student guidance records and teaching hours data from the graduate student system; collecting the teacher's recruitment records, on-the-job status, and professional title information from the personnel system; collecting the teacher's project participation records, paper publication records, and patent application records from the research system; and collecting funding receipt and reward distribution records related to the teacher from the financial system. The standard event message adopts a structured data format and includes at least a unique event identifier, event type, occurrence time, associated teacher local identifier, and key event metrics.

[0079] The collection of recruitment and training event data related to the teachers to be accounted for aims to comprehensively obtain information related to the teachers' recruitment and training workload from multiple dimensions and sources, ensuring the completeness and accuracy of the data. This is achieved through establishing data interfaces, configuring data scraping tools, or using API calls. Data on the teacher's student guidance records and teaching hours are collected from the graduate student system to reflect the teacher's direct input in talent cultivation; data such as student graduation status and course teaching quality evaluation can also be used as supplementary data. Recruitment records, on-the-job status, and professional title information are collected from the personnel system to characterize the teacher's identity attributes and career development; information such as contract duration and job level can also be included. Project participation records, paper publication records, and patent application records are collected from the research system to quantify the teacher's contribution to scientific research innovation and academic output; data such as the transformation of research results and academic conference reports can also be considered. Funding receipts and reward distribution records related to the teacher are collected from the financial system to reflect the teacher's performance in resource acquisition and recognition; data such as the use of research funds and details of various awards can also be used as a reference. The standard event message adopts a structured data format to unify the representation of data from different sources, facilitating subsequent automated processing and analysis. The structured data format is JSON, XML, or a predefined data table format, ensuring clear and consistent data field definitions. A unique event identifier ensures the traceability of each event record, preventing duplication or omissions. The event type categorizes different types of mentoring activities, such as "supervising doctoral students" or "publishing SCI papers." The occurrence time records the specific moment the event occurred, providing a foundation for time series analysis. The associated teacher local identifier is the teacher's identifier in the original business system at the time the event occurred, crucial for subsequent identity alignment. Key event metrics include core data quantifying the event, such as the number of students supervised, paper impact factor, and project funding amount.

[0080] This application's solution comprehensively collects recruitment and development event data related to teachers to be accounted for from multiple heterogeneous business systems, including graduate student systems, personnel systems, research systems, and financial systems, ensuring the comprehensiveness and multi-dimensionality of recruitment and development workload information. This raw data undergoes unified processing, converting it into standard event messages in a structured data format. These messages include at least a unique event identifier, event type, occurrence time, associated teacher local identifier, and key event metrics. This standardized data format effectively solves the problem of inconsistent data formats across different systems, thereby improving the accuracy of subsequent contextual association and semantic completion based on a talent recruitment and development knowledge graph, making the graph-based recruitment and development event data richer and more accurate. Furthermore, the recruitment and development event vectors generated based on a graph neural network model more accurately represent the recruitment and development workload of teachers to be accounted for, thereby improving the calculation results of recruitment and development target completion and effectively solving the problems of incomplete and inaccurate data caused by dispersed data sources and inconsistent formats. Through the above technical solution, this application comprehensively and accurately collects recruitment and development event data related to teachers to be accounted for from multiple heterogeneous systems and unifies it into structured standard event messages. This not only solves the problem of scattered and inconsistent raw data, ensuring the integrity and consistency of data input, but also provides a high-quality and standardized data foundation for subsequent knowledge graph construction and graph neural network analysis, thereby improving the accuracy of the performance evaluation results for teacher talent attraction and development.

[0081] In one embodiment, this application further proposes a training process for a graph neural network model, which includes: using historically labeled teacher workload graph data and their corresponding real talent introduction target completion labels as training samples, with the goal of minimizing the difference between the model prediction results and the real labels, and optimizing the model parameters using supervised learning.

[0082] This approach introduces a graph neural network (GNN) model training process to ensure that the model learns effective feature representations from complex recruitment event data. The training process utilizes historically labeled teacher workload graph data as input, which contains rich entities (such as teachers, students, and projects) and their relationships (such as guidance, participation, and affiliation), presented in a graph structure. Simultaneously, the actual recruitment target completion labels corresponding to these graph data serve as supervisory signals to guide model learning. During training, the GNN model extracts features and aggregates information from the input graph data to generate prediction results. Subsequently, by calculating the difference between the model's predictions and the actual labels, and aiming to minimize this difference, supervised learning is employed, using optimization algorithms (such as backpropagation and gradient descent) to iteratively adjust the model's internal parameters. This iterative optimization allows the model to gradually learn the complex patterns and rules inherent in the data, thereby improving its accuracy in generating recruitment event vectors. The fully trained and optimized graph neural network model more accurately captures subtle changes and deep correlations in teachers' recruitment and development workload. This enables the model to more precisely extract and generate recruitment and development event vectors for teachers to be evaluated from the graph-based recruitment and development event data, effectively overcoming the inaccurate feature extraction problems that may be caused by untrained models. Therefore, this solution improves the reliability of subsequent recruitment and development target completion calculations and the accuracy of performance evaluation, providing a solid data foundation and decision support for university talent recruitment and development performance management.

[0083] The above method is implemented as a computer program that runs on a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and the computer program. The computer program includes program instructions that, when executed, cause the processor to execute any knowledge graph-based performance management method. The processor provides computational and control capabilities to support the operation of the entire computer device. The internal memory provides the environment for the computer program in the non-volatile storage media to run; when executed by the processor, this program causes the processor to execute any knowledge graph-based performance management method. The network interface is used for network communication, such as sending assigned tasks.

[0084] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement any of the knowledge graph-based performance management methods provided in the embodiments of this application.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A performance management method based on knowledge graphs, characterized in that, The performance management method includes the following steps: Obtain target recruitment and training event messages that characterize the recruitment and training workload of teachers to be calculated; Based on a pre-defined talent attraction and development knowledge graph, contextual association and semantic completion are performed on the target talent attraction and development event messages to generate graph-based talent attraction and development event data. Based on the graph neural network model and the graph-based recruitment and training event data, the recruitment and training event vector of the teacher to be accounted for is generated; Based on the preset talent attraction and development target rule base and the talent attraction and development event vector, the completion degree of the talent attraction and development target of the teacher to be accounted for is calculated, so as to complete the performance accounting of teacher talent attraction and development based on the completion degree of the talent attraction and development target.

2. The performance management method as described in claim 1, characterized in that, The process involves performing contextual association and semantic completion on the target talent attraction and development event messages based on a pre-defined talent attraction and development knowledge graph to generate graph-based talent attraction and development event data, including: Based on the predefined entity types and predefined relationship types in the talent attraction and development knowledge graph, entity recognition and relationship extraction are performed on the target attraction and development event message to obtain the target attraction and development event entity and the target attraction and development event relationship. The target talent attraction and development event entities are integrated into the corresponding nodes in the talent attraction and development knowledge graph; The metadata describing the inherent characteristics of the target talent attraction event entity in the target talent attraction event message is attached as a node attribute to the corresponding node in the talent attraction knowledge graph; The metadata describing the relationship characteristics of the target talent attraction and development event in the target talent attraction and development event message is attached as an edge attribute to the corresponding edge in the talent attraction and development knowledge graph; Based on the explicit relationships of the fused talent attraction and development knowledge graph, implicit relationship reasoning is performed on the target talent attraction and development event messages; Based on the talent attraction and development knowledge graph that integrates the implicit relationships, the context association of the target attraction and development event message is constructed, and based on the context association, the graph-based attraction and development event data is obtained.

3. The performance management method as described in claim 2, characterized in that, The implicit relationship reasoning of the target talent attraction and development event message based on the explicit relationship of the fused talent attraction and development knowledge graph includes: Based on the display relationship path template corresponding to the display relationship, at least one candidate relationship chain starting from the teacher node to be verified is constructed, wherein the teacher node to be verified is the teacher node in the target introduction and cultivation event message; Based on a multi-source confidence fusion strategy, candidate relation chains with confidence scores below a preset confidence threshold are filtered to obtain filtered candidate relation chains. Based on the graph embedding model and the candidate relation chain, the completion probability of each missing link in the candidate relation chain is calculated, and the implicit relation is generated based on the missing links whose completion probability exceeds a preset probability threshold.

4. The performance management method as described in claim 2, characterized in that, The construction of contextual associations for the target talent attraction and development event messages based on the talent attraction and development knowledge graph that integrates the implicit relationships includes: The implicit relationship is added as a new triple to the talent attraction and development knowledge graph; Taking the teacher node to be accounted for as the center, extract the first-degree related subgraph corresponding to the first-degree neighbor node from the added talent introduction and cultivation knowledge graph. The first-degree neighbor node has a direct relationship with the teacher node to be accounted for. Based on the first-degree related subgraph, the second-degree related subgraph corresponding to the second-degree neighbor node is extracted from the added talent attraction and cultivation knowledge graph. The second-degree neighbor node has an indirect relationship with the teacher node to be calculated. The first-degree association subgraph and the second-degree association subgraph are integrated to obtain the context association.

5. The performance management method as described in claim 4, characterized in that, The step of extracting a first-degree related subgraph corresponding to a first-degree neighbor node from the added talent attraction and development knowledge graph, centered on the node corresponding to the teacher to be accounted for, includes: Starting from the teacher node to be evaluated, traverse all outgoing and incoming edges of the teacher node to be evaluated, and determine the first-degree neighbor node directly connected to the teacher node to be evaluated; Extract the edges between the teacher node to be evaluated and all its first-degree neighbors to obtain the set of direct relations; direct relations include mentorship, participation, membership, and referral relations. The set of teacher nodes to be calculated, the set of first-degree neighbor nodes, and the set of direct relationships are combined to construct the first-degree association subgraph.

6. The performance management method as described in claim 4, characterized in that, The step of extracting the second-degree neighbor node corresponding to the second-degree related subgraph from the added talent attraction and cultivation knowledge graph based on the first-degree related subgraph includes: Starting from each neighbor node in the set of first-degree neighbor nodes, traverse the outgoing and incoming edges of each neighbor node to determine the second-degree neighbor nodes indirectly connected to the teacher node to be calculated. Obtain the edges connecting the first-degree neighbor node set and the second-degree neighbor node set to obtain the indirect relationship set; the indirect relationship is used to represent the indirect contribution of the teacher to be accounted for. The indirect relationships mentioned above include student-published papers, student-participated projects, projects led by introduced talents, and project funding received. The second-degree related subgraph is constructed by combining the first-degree related subgraph, the set of second-degree neighbor nodes, and the set of indirect relationships.

7. The performance management method as described in claim 1, characterized in that, The generation of the teacher's recruitment event vector based on the graph neural network model and the graph-based recruitment event data includes: The graph-based mentoring event data is converted into a graph structure containing preset mentoring entity nodes and preset mentoring event edges. The mentoring entity nodes include teacher nodes, event nodes, student nodes, and project nodes. The mentoring event edges include guidance relationships, participation relationships, and membership relationships. Each node's features are initialized. The features of the teacher node include age, professional title, subject area, and historical performance score; the features of the event node include event type, event level, occurrence of the event, and ongoing event; the features of the student node include year of enrollment, major, and training level; and the features of the project node include project level, funding amount, and research period. Based on graph convolutional networks or graph attention networks, the initialized nodes are subjected to neighbor node feature aggregation and feature update to obtain a node vectorization representation of each node. The node vectorization representation is used to vectorize the self-features and local neighborhood information of each node. From the node vectorization representation of each node, extract the teacher feature vector of the teacher to be accounted for and the event feature vector used to characterize the workload of the teacher to be accounted for in recruitment and training, as the recruitment and training event vector.

8. The performance management method as described in any one of claims 1-7, characterized in that, The calculation of the teacher's achievement rate in attracting and cultivating talent, based on a preset rule base for talent attraction and cultivation goals and the talent attraction and cultivation event vector, includes: Each rule in the breeding target rule base is encoded into a rule feature vector, and the matching degree between the breeding event vector and each rule feature vector is calculated; Based on the preset matching degree threshold and each matching degree, the number of recruitment and training events completed by the teacher to be accounted for is determined; The completion rate of the breeding target is calculated based on the total number of breeding events and the number of breeding events completed.

9. A computer device, characterized in that, The computer device includes a processor, a memory, and a knowledge graph-based performance management program stored in the memory and executable by the processor, wherein when the knowledge graph-based performance management program is executed by the processor, it implements the steps of the knowledge graph-based performance management method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a knowledge graph-based performance management program, wherein when the knowledge graph-based performance management program is executed by a processor, it implements the steps of the knowledge graph-based performance management method as described in any one of claims 1 to 8.

Images