A knowledge graph-based human resource management method and system
By calculating the dispersion coefficient and information distinguishability of business attribute entities in the knowledge graph to generate a differentiated weight matrix, the problem of weakened semantic information in graph neural networks is solved, thereby improving the accuracy and rationality of human resource management decisions.
Patent Information
- Application Number
- CN202611305188.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
AI Technical Summary
In the process of aggregation of graph neural network representations in existing knowledge graphs, the equal weight propagation of connection relationships with different semantic values weakens key and effective semantic information, affecting the rationality of human resource management decisions.
By calculating the dispersion coefficient and information distinguishability of business attribute entities in the human resources knowledge graph, a differentiated weight matrix is generated to quantify the distribution dispersion of entity connection relationships and distinguish semantic importance, suppress the influence of redundant relationships, and improve the ability to preserve semantic signals.
It enables differentiated weighted aggregation of messages in the graph neighborhood, improves the reliability of job matching output results, and ensures that human resource management decisions are aligned with actual business needs.
Smart Images

Figure CN123118048A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of knowledge graph technology, specifically relating to a knowledge graph-based human resource management method and system. Background Technology
[0002] As enterprises continue to advance their digital transformation, knowledge graph-based human resource management technologies have been widely applied in scenarios such as talent profiling, job matching, competency assessment, and human resource risk assessment, becoming an important foundational support for intelligent human resource decision-making.
[0003] Existing knowledge graph-driven human resource management solutions typically employ graph neural networks or embedding propagation methods to aggregate multi-hop neighbor information from the knowledge graph into the representations of each entity. The aggregated entity representations then undergo feature decoding, feature parsing, and business dimension conversion to output corresponding human resource analysis and decision-making results. Furthermore, some solutions incorporate attention weighting, noise filtering, or auxiliary correction mechanisms to further enhance the reliability of human resource entity representations and the accuracy of output results. This approach offers advantages such as strong knowledge association mining capabilities, convenient multi-dimensional information fusion, and ease of integration and deployment within enterprise human resource data platforms, thus finding widespread application in existing knowledge graph-based human resource management systems.
[0004] However, in the actual application of the above knowledge graph processing methods, various connection relationships between different types of nodes in the knowledge graph are used to represent aggregation to the same extent. Low-information-value redundant connections and high-information-value semantic connections are transmitted with the same weight, causing high-information-value semantic signals to be covered by signals from a large number of redundant connections. Key semantic information is attenuated during the propagation process, resulting in a deviation between human resource management decision-making results and actual business needs, thereby affecting the rationality of the enterprise's human resource management strategy. Summary of the Invention
[0005] The purpose of this invention is to solve the problem that the weighted propagation of connection relationships with different semantic values during the aggregation of knowledge graph representations in graph neural networks weakens and interferes with key and effective semantic information, resulting in low rationality of management strategies. Therefore, this invention proposes a human resource management method and system based on knowledge graphs.
[0006] In a first aspect of this invention, a knowledge graph-based human resource management method is proposed, the method comprising: Acquire a human resources knowledge graph and historical interaction data between employees and positions; the human resources knowledge graph contains multiple employee entities, multiple business attribute entities, and various relationship types between employee entities and business attribute entities. The dispersion coefficient of each business attribute entity is calculated based on the human resources knowledge graph; the dispersion coefficient is used to measure the degree of concentration of the distribution of various relationship types connected to the business attribute entity in all connection relationships; Based on the dispersion coefficients of all business attribute entities and the historical interaction data, calculate the information distinguishability of each business attribute entity relative to each employee entity. A differentiated weight matrix is generated based on the information distinguishability of each business attribute entity relative to each employee entity; each element in the differentiated weight matrix represents the degree of importance of the association between the business attribute entity and the corresponding employee entity. Based on the differentiated weight matrix, job matching analysis is performed on all employee entities to obtain job matching results, and human resource management is carried out based on the job matching results.
[0007] Optionally, calculating the dispersion coefficient of each business attribute entity based on the human resources knowledge graph includes: The number of various relationship types connected to each business attribute entity is counted based on the aforementioned human resources knowledge graph. Calculate the proportion of each business attribute entity connected to each type of relation in the total number of relation types based on the number of each type of relation. pass Calculate the dispersion coefficient for each business attribute entity; where, For business attribute entities The dispersion coefficient, Let r be the set of relation types connected to the business attribute entity e. Any relation type in the language, For relation type In business attribute entities The proportion of all connections.
[0008] This scheme can quantify the dispersion of entity connection relationships, distinguish the semantic importance of different connection relationships, avoid treating redundant relationships and effective semantic relationships equally during multi-hop propagation of the graph, suppress representation interference caused by invalid associations, and enable entity representations to retain high-value semantic signals within the graph.
[0009] Optionally, based on the dispersion coefficients of all business attribute entities and the historical interaction data, the information distinguishability of each business attribute entity relative to each employee entity is calculated, including: The set of business attribute entities that the target employee entity has interacted with is determined based on the historical interaction data; the target employee entity is any one of all employee entities. Calculate the conditional entropy of the target employee entity for each business attribute entity in the set of business attribute entities; pass Calculate the information distinguishability of each business attribute entity relative to the target employee entity; wherein... For business attribute entities Relative to the target employee entity Information differentiation For business attribute entities The dispersion coefficient, Indicated in the target employee entity Business attribute entity set Business Attribute Entities The conditional entropy.
[0010] This solution combines historical interaction data to determine the set of business attribute entities corresponding to the target employee entity. By solving the conditional entropy and using the difference between the dispersion coefficient and the conditional entropy, it obtains the information distinguishability of the business attribute entities for the target employee entity. This solution combines the inherent semantic features of the knowledge graph with the employee's historical business interaction behavior, quantifies the information contribution of different business attribute entities to each employee entity, and makes the assessment of the importance of association take into account both the semantics of the knowledge graph ontology and the historical business behavior, thereby improving the business adaptability of weight allocation.
[0011] Optionally, generating a differentiated weight matrix based on the information distinguishability of each business attribute entity relative to each employee entity includes: The popularity of each business attribute entity is obtained by counting the number of relationships between each business attribute entity and the employee entity. pass Calculate the differential weights between each employee entity and each business attribute entity; where, For employee entities With business attribute entities Differential weights between them For business attribute entities Relative to employee entity Information differentiation For business attribute entities popularity These are preset parameters. It is the set of all business attribute entities in the aforementioned human resources knowledge graph. for Business attribute entities in For business attribute entities The popularity, where c is a preset constant; The differential weight matrix is obtained by arranging the differential weights between all employee entities and all business attribute entities according to the employee entity dimension.
[0012] This solution combines the topological connection characteristics between employee entities and business attribute entities in the human resources knowledge graph. Based on entity popularity and information distinguishability, it calculates the differentiated weights between entities and generates a differentiated weight matrix. This matrix can objectively reflect the actual business value of different entity associations within the graph, weaken the impact of high-frequency but low-distinguishability entities, and achieve refined weighting of graph edge relationships.
[0013] Optionally, the job matching results obtained by performing job matching analysis on all employee entities based on the differentiated weight matrix include: The adjacency matrix and the differential weight matrix of the human resources knowledge graph are input into a preset graph structure optimization model to obtain a human resources knowledge subgraph; The node features of each business attribute entity in the human resources knowledge subgraph are input into a preset graph embedding model to obtain the representation vector corresponding to each business attribute entity. The knowledge graph edge set is obtained by extracting the semantic association structure based on the representation vectors corresponding to all business attribute entities. Based on the edge set of the knowledge graph, semantic aggregation is performed on each business attribute entity to obtain the semantic representation vector of each business attribute entity. Based on the identifier representation vector of each employee entity and the semantic representation vector of each business attribute entity, the matching score between each employee entity and each business attribute entity is calculated to obtain the job matching result.
[0014] This solution integrates differentiated weights into the graph diffusion and embedding process of knowledge graphs, selects effective subgraphs based on the graph topology, and completes targeted semantic aggregation. It suppresses semantic interference caused by redundant associations within the graph, retains high-value association semantic information in the graph, and makes entity representation vectors more aligned with actual business needs. This effectively improves the accuracy of the matching score calculation between employees and business attribute entities.
[0015] Optionally, the knowledge graph edge set is obtained by extracting the semantic association structure based on the representation vectors corresponding to all business attribute entities, including: Construct a vector space based on the representation vectors corresponding to each business attribute entity; The semantic distance set is obtained by calculating the semantic distance between any two business attribute entities in the vector space. Based on the semantic distance set, semantic nearest neighbor aggregation is performed on all business attribute entities to obtain the knowledge graph edge set.
[0016] This scheme can overcome the inherent constraints of the original topological edges of the knowledge graph, mine the hidden semantic associations in the entity vector space of the graph, complete the semantic connection relationships not explicitly recorded in the original graph, filter out invalid noise edges in the original graph topology, and make the generated edge set take into account both the original entity semantics and implicit association features of the graph.
[0017] Optionally, the knowledge graph edge set obtained by performing semantic nearest neighbor aggregation on all business attribute entities based on the semantic distance set includes: Based on the semantic distance set, a preset number of business attribute entities with the closest semantic distance to the target business attribute entity are selected as neighboring entities to obtain a neighboring entity set; the target business attribute entity is any one of the business attribute entities among all business attribute entities. Connect the target business attribute entity with each neighboring entity in the neighboring entity set to form a local semantic edge set; All local semantic edge sets corresponding to business attribute entities are used as semantic edge sets; The semantic edge set is symmetricized to obtain the knowledge graph edge set.
[0018] This scheme controls the generation scale of graph edges by selecting neighboring entities, and then uses symmetry processing to ensure the consistency of the graph topology, avoiding structural deviations caused by unidirectional connections. This results in a reconstructed graph edge set that possesses both semantic relevance and graph structural completeness. Optionally, the semantic representation vector of each business attribute entity is obtained by semantic aggregation based on the edge set of the knowledge graph, including: The target entity is identified by having business attribute entities connected by edges in the knowledge graph edge set as its neighbor entities, thus forming a neighbor entity set. The target entity can be any entity in the knowledge graph edge set. Each edge in the knowledge graph edge set has a relation type identifier. The projection matrix is determined based on the relationship type identifier corresponding to the edge between the target neighbor entity and the target entity; the target neighbor entity is any one of the neighbor entities in the set of neighbor entities. The projection feature vector is obtained by projecting the representation vector of the target neighbor entity onto the projection matrix. Based on the projected feature vector and the representation vector of the target entity, calculate the attention weight of the target neighbor entity relative to the target entity; The semantic representation vector of the target entity is obtained by weighting and aggregating the representation vectors of each neighboring entity based on the attention weights of all neighboring entities.
[0019] This scheme fully utilizes the topological attribute of the relation types inherent in the edge set of the knowledge graph, performs differentiated projection transformations for different graph relations, distinguishes the contribution degree of each neighboring entity within the graph through attention weights, weakens the information interference brought by irrelevant neighbors, and enables the final generated semantic representation vector to fully absorb the relational semantic information of the graph, thereby improving the ability of entity representation to express the relational features of the graph.
[0020] In a second aspect of this invention, a knowledge graph-based human resource management system is proposed, comprising: The data module is used to acquire human resources knowledge graphs and historical interaction data between employees and positions; the human resources knowledge graph includes multiple employee entities, multiple business attribute entities, and various relationship types between employee entities and business attribute entities. The dispersion coefficient calculation module is used to calculate the dispersion coefficient of each business attribute entity based on the human resources knowledge graph; the dispersion coefficient is used to measure the degree of distribution concentration of various relationship types connected to the business attribute entity in all connection relationships; The information distinguishability calculation module is used to calculate the information distinguishability of each business attribute entity relative to each employee entity based on the dispersion coefficient of all business attribute entities and the historical interaction data. The matrix generation module is used to generate a differentiated weight matrix based on the information distinguishability of each business attribute entity relative to each employee entity; each element in the differentiated weight matrix represents the degree of importance of the association between the business attribute entity and the corresponding employee entity. The human resources management module is used to perform job matching analysis on all employee entities based on the differentiated weight matrix to obtain job matching results, and to perform human resources management based on the job matching results.
[0021] The beneficial effects of this invention are as follows: This invention proposes a human resource management method based on a knowledge graph. By acquiring a human resource knowledge graph containing employee entities, business attribute entities, and various relationships, along with historical interaction data between employees and positions, the method first calculates the dispersion coefficient of each business attribute entity. Then, by combining the dispersion coefficient with historical interaction data, it obtains the information distinguishability of each business attribute entity for each employee entity. Based on this, a differentiated weight matrix representing the importance of relationships between entities is constructed. This matrix is then used to perform job matching analysis for employee entities to conduct human resource management. This solution can distinguish the importance of different relationships based on the semantic features of the entities themselves and historical interaction information, achieving differentiated weighted aggregation of graph neighborhood messages, improving the reliability of job matching output results, and ensuring that human resource management decisions are aligned with actual business needs. Attached Figure Description
[0022] The present invention will now be further described with reference to the accompanying drawings.
[0023] Figure 1 A flowchart illustrating a knowledge graph-based human resource management method provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the generation of job matching results provided in an embodiment of the present invention. Detailed Implementation
[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0025] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] This invention provides a knowledge graph-based human resource management method. See also... Figure 1 The method includes the following steps: S101, Obtain the human resources knowledge graph and historical interaction data between employees and positions; S102, Calculate the dispersion coefficient of each business attribute entity based on the human resources knowledge graph; S103, Based on the dispersion coefficients and historical interaction data of all business attribute entities, calculate the information distinguishability of each business attribute entity relative to each employee entity. S104, Generate a differentiated weight matrix based on the information distinguishability of each business attribute entity relative to each employee entity; S105. Based on the differentiated weight matrix, perform job matching analysis on all employee entities to obtain job matching results, and conduct human resource management based on the job matching results. The human resources knowledge graph includes multiple employee entities, multiple business attribute entities, and various relationship types between employee entities and business attribute entities. The dispersion coefficient is used to measure the degree of concentration of the distribution of various relationship types connected to a business attribute entity across all relationships; Each element in the differential weight matrix represents the importance of the association between the business attribute entity and the corresponding employee entity.
[0027] In one implementation, the human resources knowledge graph records the static association facts between employee entities and business attribute entities, that is, nodes represent various entities and edges represent the objectively existing relationship types between employee entities and business attribute entities. The historical interaction data between employees and positions records the dynamic behavioral sequence generated between employee entities and business attribute entities in the time dimension, that is, the factual record of each employee interacting with each business attribute entity with a specific interaction type at a specific point in time. Each record includes employee identifier, business attribute entity identifier, interaction type, interaction timestamp, and interaction frequency, which describes the occurrence pattern of the relationship on the time axis, including the time of occurrence, duration, number of repetitions, and sequence.
[0028] Knowledge graphs provide the semantic distribution features of business attribute entities globally, while historical interaction data provides the behavioral distribution features of individual employees over time. These two types of data characterize the relationships between the same employee and the same business attribute entity from different dimensions. Storing and processing them separately allows for independent extraction and differentiated measurement of global semantic features and individual behavioral features. This enables the calculation of the deviation between the global semantic distribution and the individual behavioral distribution in the semantic space. This deviation directly reflects the contribution of the business attribute entity to the employee's personalized information. This process preserves the temporal information in historical interaction data, facilitating the distinction between high-frequency, regular interactions and low-frequency, random interactions. Simultaneously, it avoids homologous bias between the global and individual distributions, ensuring that information distinguishability maintains its mathematical attribute of the difference between global and individual information during the calculation process. Finally, by aligning and fusing the two independent data sources in the same semantic space, each element of the differentiated weight matrix accurately reflects the contribution of the business attribute entity to the employee's personalized information.
[0029] In one implementation, the dispersion coefficient is used to quantify the diversity of the types of relationships connected to a business attribute entity. The higher the dispersion coefficient, the more the entity is associated with multiple different types of relationships, and the more dispersed and unfocused the semantics are globally. The lower the dispersion coefficient, the more concentrated the types of relationships connected to the entity are, and the clearer the semantics are.
[0030] In one implementation, information distinguishability is used to characterize how much effective information a business attribute entity can provide for judging the employee's job matching tendency. The higher the information distinguishability, the more effectively the business attribute entity can distinguish the differences between the employee and other employees. That is, the relationship between the employee and the entity has clear regularity and stability, and the entity can serve as an effective indicator signal for judging the employee's job matching tendency. The lower the information distinguishability, the more random or generalized the relationship between the entity and the employee is, and it is difficult to extract effective indicator information for the employee's job matching tendency.
[0031] In one implementation, the following further illustrates the solution using a real-world human resource management scenario: The system acquires a unified human resource knowledge graph across the company. This graph includes 100 employee entities, such as Zhang Ming, Li Si, and Wang Wu; 200 business attribute entities, including entities for various positions, skills, departments, and projects; and 15 relationship types between employees and these business attributes, such as mastering, holding a position, participating in, completing, and reporting to. Simultaneously, the system acquires historical interaction data between employees like Zhang Ming and various positions, including Zhang Ming's employment record (current position: Senior Algorithm Engineer), project participation record (led an intelligent dialogue platform project and participated in a distributed training framework optimization project), training completion record (completed advanced technical management training), and job application record (applied for a technical leadership position).
[0032] The distribution of various relationship types connected to each entity in the human resources knowledge graph was analyzed by traversing each business attribute entity. Taking the skill entity distributed system as an example, the statistics show that this skill involves a total of 200 connections in the global knowledge graph. Among them, 160 connections are associated with the "mastery" relationship type, accounting for 80%; 25 connections are associated with the "job requirement" relationship type, accounting for 12.5%; and 15 connections are associated with the "project application" relationship type, accounting for 7.5%. The distribution is relatively concentrated, and the dispersion coefficient of this skill is calculated to be 0.52. Taking the skill entity communication ability as an example, the system statistics show that this skill involves a total of 600 connections. Among them, 150 connections are related to the mastery relationship type, accounting for 25%; 140 connections are related to the job requirement relationship type, accounting for 23.3%; 125 connections are related to the training objective relationship type, accounting for 20.8%; 110 connections are related to the project collaboration relationship type, accounting for 18.3%; and 75 connections are related to the job promotion standard relationship type, accounting for 12.5%. The distribution of the five relationship types is relatively even, and the dispersion coefficient of this skill is calculated to be 1.58. The system repeats the above process for each business attribute entity in the knowledge graph to obtain the dispersion coefficient values of all business attribute entities, which range from 0.12 to 1.75.
[0033] For employee Zhang Ming, his historical interaction data is retrieved, and the information distinguishability of each business attribute entity relative to Zhang Ming is calculated one by one, based on the dispersion coefficient values of each business attribute entity. Taking the skill entity distributed system as an example, its global dispersion coefficient is 0.52. In Zhang Ming's historical interaction records, this skill appears in 6 out of the 8 core projects he led, 1 out of the 2 positions he held, and 2 out of the 6 training sessions he completed. This historical interaction data shows a highly concentrated distribution pattern. The system calculates its conditional entropy under Zhang Ming's conditions to be 0.18, and the information distinguishability = 0.52 - 0.18 = 0.34. Taking the skill entity communication ability as an example, its global dispersion coefficient is 1.58. In Zhang Ming's historical interaction records, this skill only appears in his completed technical management advanced training records, and only once, showing a low-frequency and random distribution characteristic. The conditional entropy under Zhang Ming's conditions is calculated to be 1.45, and the information distinguishability = 1.58 - 1.45 = 0.13. The calculations for all other employees are performed using the same method.
[0034] All calculated information discrimination values are arranged according to the employee entity dimension, with employees corresponding to one row and business attribute entities as columns, forming a differential weight matrix. In this matrix, in the row corresponding to Zhang Ming, the element value corresponding to the business attribute "distributed systems" is 0.34, the element value corresponding to the business attribute "communication ability" is 0.13, the element value corresponding to the business attribute "Python programming" is 0.07, and the element value corresponding to the business attribute "team management" is 0.31. This differential weight matrix has a dimension of 100×200, with a total of 20,000 elements. Each element directly quantifies the importance of the association between a certain business attribute entity and a certain employee entity.
[0035] Finally, guided by a differentiated weight matrix, conditional denoising processing is performed on the human resources knowledge graph, outputting a knowledge subgraph with higher connection quality. Subsequently, the system extracts the semantic representation vectors of each business attribute entity from this knowledge subgraph and, combined with the identifier representation vectors of each employee entity, calculates the matching score between each employee entity and each job entity. For Zhang Ming, the matching results show a matching score of 0.87 with the Technical Leader position, 0.72 with the Senior Researcher position, 0.65 with the Project Manager position, 0.58 with his current Senior Algorithm Engineer position, and 0.31 with the Product Manager position. Following this method, the matching scores for all employees are determined, ultimately generating job matching results. The Human Resources Department then uses these job matching results for human resource management.
[0036] In one embodiment, calculating the dispersion coefficient of each business attribute entity based on the human resources knowledge graph includes: The number of various relationship types connected to each business attribute entity is counted based on the human resources knowledge graph. Calculate the proportion of each business attribute entity connected to each type of relation in the total number of relation types based on the number of each type of relation. pass Calculate the dispersion coefficient for each business attribute entity; where, For business attribute entities The dispersion coefficient, Let r be the set of relation types connected to the business attribute entity e. Any relation type in the language, For relation type In business attribute entities The proportion of all connections.
[0037] In one implementation, business attribute entities can be skill entities, job entities, or department entities. For each business attribute entity, the frequency of each type of relationship between it and employee entities is counted. Relationship types include mastery relationships, appointment relationships, participation relationships, completion relationships, and application relationships between employee entities and business attribute entities. The statistical results are used to reflect the distribution of relationship types connected to each business attribute entity in the human resource knowledge graph. When the number of various relationship types connected to a business attribute entity is more evenly distributed, the dispersion coefficient of the business attribute entity is higher, indicating that the semantics of the entity in the knowledge graph are more dispersed. When the number of various relationship types connected to a business attribute entity is concentrated in a certain relationship type, the dispersion coefficient of the business attribute entity is lower, indicating that the semantics of the entity in the knowledge graph are more focused.
[0038] In one embodiment, calculating the information distinguishability of each business attribute entity relative to each employee entity, based on the dispersion coefficients and historical interaction data of all business attribute entities, includes: The set of business attribute entities that the target employee entity has interacted with is determined based on historical interaction data; the target employee entity is any one of all employee entities. Calculate the conditional entropy of the target employee entity for each business attribute entity in the set of business attribute entities; pass Calculate the information distinguishability of each business attribute entity relative to the target employee entity; where, For business attribute entities Relative to the target employee entity Information differentiation For business attribute entities The dispersion coefficient, Indicates the target employee entity Business attribute entity set Business Attribute Entities The conditional entropy.
[0039] In one implementation, appointment records, application records, participation records, and completion records constitute the historical interaction conditions of the target employee entity. Conditional entropy refers to calculating the degree of uncertainty remaining for each business attribute entity relative to the target employee entity, given all historical interaction data of the target employee entity. For any business attribute entity, if it appears repeatedly in the historical interaction records of the target employee entity and exhibits a clear regularity, then the uncertainty of the entity under the known employee history is low, and its conditional entropy value is small; conversely, if the entity appears only in a few or random interactions, then the uncertainty of the entity under the known employee history is high, and its conditional entropy value is large. The formula for calculating conditional entropy is... ;in, A collection of entities with business attributes. for Any business attribute entity in the data, In order to be in Down The probability of occurrence.
[0040] In one embodiment, generating a differential weight matrix based on the information distinguishability of each business attribute entity relative to each employee entity includes: The popularity of each business attribute entity is obtained by counting the number of relationships between each business attribute entity and the employee entity. pass Calculate the differential weights between each employee entity and each business attribute entity; where, For employee entities With business attribute entities Differential weights between them For business attribute entities Relative to employee entity Information differentiation For business attribute entities popularity These are preset parameters. It is a collection of all business attribute entities in the human resources knowledge graph. for Business attribute entities in For business attribute entities The popularity, where c is a preset constant; The differential weight matrix is obtained by arranging the differential weights between all employee entities and all business attribute entities according to the employee entity dimension.
[0041] In one implementation, the preset parameters and preset constants are set by technicians, and the preset parameters can be 0.2 and the preset constants can be 0.2.
[0042] In one implementation, global statistical information is introduced by calculating the popularity of each business attribute entity to correct the information discrimination, so as to avoid high popularity entities being given too high weight in the differential weight matrix. In the human resources knowledge graph, some business attribute entities are associated with a large number of employee entities. If the weight is determined solely based on the information discrimination, although the information discrimination of these high popularity entities may be low for individual employees, they are likely to occupy a large proportion in the weight matrix because they are common in the interaction records of most employees, thus affecting the overall balance of the matrix.
[0043] In one embodiment, reference Figure 2 The job matching results obtained by performing job matching analysis on all employee entities based on the differential weight matrix include: S201, Input the adjacency matrix and differential weight matrix of the human resources knowledge graph into the preset graph structure optimization model to obtain the human resources knowledge subgraph; S202, input the node features of each business attribute entity in the human resources knowledge subgraph into the preset graph embedding model to obtain the representation vector corresponding to each business attribute entity; S203, extract the knowledge graph edge set by performing semantic association structure extraction based on the representation vectors corresponding to all business attribute entities; S204, based on the edge set of the knowledge graph, perform semantic aggregation on each business attribute entity to obtain the semantic representation vector of each business attribute entity; S205. Based on the identifier representation vector of each employee entity and the semantic representation vector of each business attribute entity, calculate the matching score between each employee entity and each business attribute entity to obtain the job matching result.
[0044] In one implementation, the training process of the preset graph structure optimization model is as follows: First, a training dataset is acquired, which includes the adjacency matrix of the sample knowledge graph and the sample differential weight matrix. Each element in the sample differential weight matrix represents the importance of the association between the corresponding employee entity and the corresponding business attribute entity in the sample knowledge graph. After training begins, the adjacency matrix of the sample knowledge graph is used as the initial training graph structure, and the sample differential weight matrix is used as the training condition. A forward noise addition process is executed according to a preset total number of steps. In each step, random noise is added to the current graph structure according to the noise scheduling coefficient corresponding to that step. This process is iterated until the preset total number of steps is reached, resulting in the noise-added graph structure for each noise-added step. Subsequently, the noise-added graph structure for each noise-added step, the time step information corresponding to each noise-added step, and the sample differential weight matrix are input together into the denoising network to be trained. The denoising network then predicts the noise added in each noise-added step. Finally, the loss function value is calculated based on the difference between the noise predicted by the denoising network and the actual noise added, and the parameters of the denoising network are updated based on the loss function value. Then, the forward denoising process is returned and iterated until the loss function value meets the preset convergence condition. The training stops when the loss function value meets the preset convergence condition. The denoising network that meets the preset convergence condition at this time is used as the preset graph structure optimization model.
[0045] The knowledge graph structure is purified through a conditional denoising mechanism using a pre-defined graph structure optimization model. This model takes a differential weight matrix as input and determines the retention strength of each connection based on its corresponding differential weight value during denoising. This ensures that connections with high differential weights are retained in the denoised graph structure, while connections with low differential weights are removed, transforming the original human resources knowledge graph into a knowledge subgraph with higher connection quality. This process is automatically streamlined by the pre-defined graph structure optimization model based on the differential weight matrix, preserving semantic connections with high informational value for job matching while filtering out redundant connections without high informational value for job matching.
[0046] In one implementation, node features are used to describe the basic attribute information of the business attribute entity in the knowledge graph. However, the original node features between different business attribute entities are not comparable and cannot directly reflect the semantic relationship between entities. Therefore, it is necessary to input the node features of each business attribute entity into a preset graph embedding model for feature transformation.
[0047] The preset graph embedding model is a pre-trained neural network model that can map high-dimensional original node features to a low-dimensional dense vector space, outputting a representation vector corresponding to each business attribute entity. The representation vector is a fixed-length numerical vector, which not only retains the original attribute information of the business attribute entity in the knowledge graph, but also incorporates the local neighborhood structure information of the entity in the denoised graph structure.
[0048] In one implementation, the identifier representation vector of each employee entity originates from the employee entity's own identifier embedding vector. This vector represents the individual identity information of the employee entity. The job matching process involves using the identifier representation vector of each employee entity as the initial representation vector on the employee side and the semantic representation vector of each business attribute entity as the initial representation vector on the business attribute side. Multi-layer graph convolutional propagation is performed on the employee-business attribute bipartite graph. During each layer of propagation, each employee entity aggregates the current representation vectors of its neighboring business attribute entities to update its own representation vector, and each business attribute entity aggregates the current representation vectors of its neighboring employee entities to update its own representation vector. After multiple layers of iterative propagation, the initial identifier representation vector of the employee entity incorporates the semantic information of its neighboring business attribute entities, and the initial semantic representation vector of the business attribute entity also incorporates the identity information of its neighboring employee entities. After completing the collaborative propagation, for each employee entity, the inner product operation is performed on its final output employee representation vector and the final output business attribute representation vectors of each business attribute entity to obtain the matching degree score between the employee entity and each business attribute entity. The job matching result is determined based on the matching degree score.
[0049] In one embodiment, the knowledge graph edge set obtained by extracting the semantic association structure based on the representation vectors corresponding to all business attribute entities includes: Construct a vector space based on the representation vectors corresponding to each business attribute entity; The semantic distance set is obtained by calculating the semantic distance between any two business attribute entities in the vector space. The knowledge graph edge set is obtained by performing semantic nearest neighbor aggregation on all business attribute entities based on the semantic distance set.
[0050] In one implementation, the process of constructing the vector space involves using the representation vectors corresponding to each business attribute entity as the dimensions of the vector space, and determining the number of dimensions of the representation vectors as the number of dimensions of the vector space, so that the representation vector of each business attribute entity and the position of the entity in the vector space form a one-to-one mapping relationship. This process transforms each business attribute entity from a discrete and incomparable symbolic form into a continuous coordinate point in the same mathematical space, so that each business attribute entity has a unified reference benchmark and a measurable relative positional relationship in the space.
[0051] In one implementation, Euclidean distance is used to calculate semantic distance.
[0052] In one embodiment, the knowledge graph edge set obtained by performing semantic nearest neighbor aggregation on all business attribute entities based on the semantic distance set includes: Based on the semantic distance set, a preset number of business attribute entities with the closest semantic distance to the target business attribute entity are selected as neighboring entities to obtain the neighboring entity set; the target business attribute entity is any one of the business attribute entities among all business attribute entities. Connect the target business attribute entity with each neighboring entity in the neighboring entity set to form a local semantic edge set; All local semantic edge sets corresponding to business attribute entities are used as semantic edge sets; The semantic edge set is symmetricized to obtain the knowledge graph edge set.
[0053] In one implementation, the preset quantity is determined by technical personnel, and can specifically be 10.
[0054] In one implementation, the semantic edge set is constructed by selecting a preset number of semantically closest business attribute entities as neighboring entities for each business attribute entity. This construction method results in each edge in the semantic edge set starting from the business attribute entity that made the selection and ending at the selected neighboring business attribute entity. That is, the semantic edge set is essentially a directed edge set. When entity A selects entity B as its neighboring entity, a directed edge is formed from A to B. However, entity B may not select entity A as its neighboring entity, so there may not be a corresponding directed edge in the direction from B to A. If this directed semantic edge set is directly used as the graph structure basis for subsequent graph propagation and semantic aggregation, information can only propagate in a single direction, that is, from the entity that made the selection to the selected neighboring entity. It cannot propagate in the opposite direction from the selected neighboring entity back to the entity that made the selection. This results in some entities only being able to receive neighbor information but not send information to their neighbors, or only being able to send information but not receive neighbor information, causing a one-way bias in the flow of information and structural incompleteness during graph propagation. By symmetrization, each directed edge in the semantic edge set is supplemented with its reverse edge, transforming the original unidirectional directed edge into a bidirectional undirected edge. This allows any two business attribute entities to form a bidirectional reachable connection as long as there is a proximity relationship in any direction, thus transforming the asymmetric directed graph into a symmetric undirected graph.
[0055] In one embodiment, semantic aggregation of each business attribute entity based on the knowledge graph edge set yields the semantic representation vector of each business attribute entity, including: The neighbor entity set is obtained by taking the business attribute entities that the target entity is connected to by edges in the knowledge graph edge set as the neighbor entities of the target entity; the target entity is any entity in the knowledge graph edge set; each edge in the knowledge graph edge set has a relation type identifier; The projection matrix is determined based on the relationship type identifier corresponding to the edge between the target neighbor entity and the target entity; the target neighbor entity is any one of the neighbor entities in the set of neighbor entities. The projection feature vector is obtained by projecting the representation vector of the target neighbor entity onto the projection matrix. Based on the projected feature vector and the representation vector of the target entity, calculate the attention weight of the target's neighboring entities relative to the target entity. The semantic representation vector of the target entity is obtained by weighting and aggregating the representation vectors of each neighboring entity based on the attention weights of all neighboring entities.
[0056] In one implementation, different edges in the knowledge graph edge set correspond to different relation types. Since the semantic information carried by different relation types differs fundamentally, directly aggregating the representation vectors of neighboring entities under different relation types within the same semantic space would cause semantic features of different relation types to become confuse, resulting in the aggregated representation vector failing to accurately reflect the semantic features of the target entity under a specific relation dimension. By introducing a relation type-specific projection matrix, the representation vectors of neighboring entities are mapped from the common semantic space to the dedicated semantic subspace corresponding to that relation type. This ensures that neighboring entities belonging to the same relation type maintain a consistent semantic projection direction within this subspace, while neighboring entities of different relation types are projected separately into their respective subspaces. This enables differentiated semantic fusion based on relation type perception during subsequent attention weight calculation and weighted aggregation, avoiding semantic interference between different relation types and improving the accuracy of the semantic aggregation result in representing the semantic features of the target entity under different relation dimensions.
[0057] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.
Claims
1. A knowledge graph-based human resource management method, characterized in that, The method includes: Acquire a human resources knowledge graph and historical interaction data between employees and positions; the human resources knowledge graph contains multiple employee entities, multiple business attribute entities, and various relationship types between employee entities and business attribute entities. The dispersion coefficient of each business attribute entity is calculated based on the human resources knowledge graph; the dispersion coefficient is used to measure the degree of concentration of the distribution of various relationship types connected to the business attribute entity in all connection relationships; Based on the dispersion coefficients of all business attribute entities and the historical interaction data, calculate the information distinguishability of each business attribute entity relative to each employee entity. A differentiated weight matrix is generated based on the information distinguishability of each business attribute entity relative to each employee entity; each element in the differentiated weight matrix represents the degree of importance of the association between the business attribute entity and the corresponding employee entity. Based on the differentiated weight matrix, job matching analysis is performed on all employee entities to obtain job matching results, and human resource management is carried out based on the job matching results.
2. The knowledge graph-based human resource management method according to claim 1, characterized in that, The dispersion coefficient for each business attribute entity is calculated based on the aforementioned human resources knowledge graph, including: The number of various relationship types connected to each business attribute entity is counted based on the aforementioned human resources knowledge graph. Calculate the proportion of each business attribute entity connected to each type of relation in the total number of relation types based on the number of each type of relation. pass Calculate the dispersion coefficient for each business attribute entity; where, For business attribute entities The dispersion coefficient, Let r be the set of relation types connected to the business attribute entity e. Any relation type in the language, For relation type In business attribute entities The proportion of all connections.
3. The knowledge graph-based human resource management method according to claim 1, characterized in that, Based on the dispersion coefficients of all business attribute entities and the historical interaction data, the information distinguishability of each business attribute entity relative to each employee entity is calculated, including: The set of business attribute entities that the target employee entity has interacted with is determined based on the historical interaction data; the target employee entity is any one of all employee entities. Calculate the conditional entropy of the target employee entity for each business attribute entity in the set of business attribute entities; pass Calculate the information distinguishability of each business attribute entity relative to the target employee entity; wherein... For business attribute entities Relative to the target employee entity Information differentiation For business attribute entities The dispersion coefficient, Indicated in the target employee entity Business attribute entity set Business Attribute Entities The conditional entropy.
4. The knowledge graph-based human resource management method according to claim 1, characterized in that, A differential weight matrix is generated based on the information distinguishability of each business attribute entity relative to each employee entity, including: The popularity of each business attribute entity is obtained by counting the number of relationships between each business attribute entity and the employee entity. pass Calculate the differential weights between each employee entity and each business attribute entity; where, For employee entities With business attribute entities Differential weights between them For business attribute entities Relative to employee entity Information differentiation For business attribute entities popularity These are preset parameters. It is the set of all business attribute entities in the aforementioned human resources knowledge graph. for Business attribute entities in For business attribute entities The popularity, where c is a preset constant; The differential weight matrix is obtained by arranging the differential weights between all employee entities and all business attribute entities according to the employee entity dimension.
5. The knowledge graph-based human resource management method according to claim 1, characterized in that, Based on the aforementioned differential weight matrix, a job matching analysis is performed on all employee entities to obtain the following job matching results: The adjacency matrix and the differential weight matrix of the human resources knowledge graph are input into a preset graph structure optimization model to obtain a human resources knowledge subgraph; The node features of each business attribute entity in the human resources knowledge subgraph are input into a preset graph embedding model to obtain the representation vector corresponding to each business attribute entity. The knowledge graph edge set is obtained by extracting the semantic association structure based on the representation vectors corresponding to all business attribute entities. Based on the edge set of the knowledge graph, semantic aggregation is performed on each business attribute entity to obtain the semantic representation vector of each business attribute entity. Based on the identifier representation vector of each employee entity and the semantic representation vector of each business attribute entity, the matching score between each employee entity and each business attribute entity is calculated to obtain the job matching result.
6. The knowledge graph-based human resource management method according to claim 5, characterized in that, The knowledge graph edge set is obtained by extracting the semantic association structure based on the representation vectors corresponding to all business attribute entities, including: Construct a vector space based on the representation vectors corresponding to each business attribute entity; The semantic distance set is obtained by calculating the semantic distance between any two business attribute entities in the vector space. Based on the semantic distance set, semantic nearest neighbor aggregation is performed on all business attribute entities to obtain the knowledge graph edge set.
7. The knowledge graph-based human resource management method according to claim 6, characterized in that, Based on the semantic distance set, semantic nearest neighbor aggregation is performed on all business attribute entities to obtain the knowledge graph edge set, which includes: Based on the semantic distance set, a preset number of business attribute entities with the closest semantic distance to the target business attribute entity are selected as neighboring entities to obtain a neighboring entity set; the target business attribute entity is any one of the business attribute entities among all business attribute entities. Connect the target business attribute entity with each neighboring entity in the neighboring entity set to form a local semantic edge set; All local semantic edge sets corresponding to business attribute entities are used as semantic edge sets; The semantic edge set is symmetricized to obtain the knowledge graph edge set.
8. The knowledge graph-based human resource management method according to claim 5, characterized in that, Based on the edge set of the knowledge graph, semantic aggregation is performed on each business attribute entity to obtain the semantic representation vector of each business attribute entity, including: The target entity is identified by having business attribute entities connected by edges in the knowledge graph edge set as its neighbor entities, thus forming a neighbor entity set. The target entity can be any entity in the knowledge graph edge set. Each edge in the knowledge graph edge set has a relation type identifier. The projection matrix is determined based on the relationship type identifier corresponding to the edge between the target neighbor entity and the target entity; the target neighbor entity is any one of the neighbor entities in the set of neighbor entities. The projection feature vector is obtained by projecting the representation vector of the target neighbor entity onto the projection matrix. Based on the projected feature vector and the representation vector of the target entity, calculate the attention weight of the target neighbor entity relative to the target entity; The semantic representation vector of the target entity is obtained by weighting and aggregating the representation vectors of each neighboring entity based on the attention weights of all neighboring entities.
9. A knowledge graph-based human resource management system, characterized in that, The system includes: The data module is used to acquire human resources knowledge graphs and historical interaction data between employees and positions; the human resources knowledge graph includes multiple employee entities, multiple business attribute entities, and various relationship types between employee entities and business attribute entities. The dispersion coefficient calculation module is used to calculate the dispersion coefficient of each business attribute entity based on the human resources knowledge graph; the dispersion coefficient is used to measure the degree of distribution concentration of various relationship types connected to the business attribute entity in all connection relationships; The information distinguishability calculation module is used to calculate the information distinguishability of each business attribute entity relative to each employee entity based on the dispersion coefficient of all business attribute entities and the historical interaction data. The matrix generation module is used to generate a differentiated weight matrix based on the information distinguishability of each business attribute entity relative to each employee entity; each element in the differentiated weight matrix represents the degree of importance of the association between the business attribute entity and the corresponding employee entity. The human resources management module is used to perform job matching analysis on all employee entities based on the differentiated weight matrix to obtain job matching results, and to perform human resources management based on the job matching results.
