Training resource recommendation method and system based on knowledge graph and electronic equipment

Abstract

The invention discloses a training resource recommendation method and system based on a knowledge graph and electronic equipment, and relates to the field of enterprise data processing, and the method comprises the steps: constructing a target talent knowledge graph of an enterprise; obtaining an existing talent knowledge graph of the enterprise; matching the target talent knowledge graph with the existing talent knowledge graph by using a graph matching algorithm, and calculating a matching degree; according to a matching result of the graph matching algorithm, training resource recommendation is carried out on enterprise employees so as to improve the matching degree of the existing talent knowledge graph and the target talent knowledge graph; and iteration is repeated until the matching degree is higher than a preset threshold value. According to the invention, enterprises can be helped to systematically evaluate and improve the ability of employees, and market demands and enterprise development objectives can be better adapted.

Description

Technical Field

[0001] The present invention relates to the field of enterprise data processing, and more specifically, to a training resource recommendation method, system and electronic device based on knowledge graph. Background Art

[0002] In today's rapidly changing business environment, companies are faced with the challenge of constantly adjusting and optimizing their talent structure to adapt to market demand. Especially in the process of corporate development, how to accurately evaluate whether the existing talent framework matches the company's strategic goals and future development direction has become a key issue in corporate human resource management.

[0003] However, companies often fail to accurately assess whether the skills and knowledge of existing employees are in line with their long-term development strategies or market needs. Traditional assessment methods rely on static job requirements and self-reporting, which are difficult to fully reflect the actual capabilities of employees. In particular, in industries with rapid technological and business development, the inadequacy of existing talent frameworks is particularly evident.

[0004] Even if companies are aware of the lack of skills and knowledge of existing employees, traditional training methods are often not systematic and cannot provide training from the perspective of the entire company. An enterprise is composed of many employees, and the capabilities of employees systematically constitute the capabilities of an enterprise. Therefore, the recommendation of training resources for employees throughout the enterprise should be more systematic, so that the human resource structure of the entire enterprise can move closer to the target human resource structure. However, there is currently no more effective way to achieve this goal. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a training resource recommendation method, system and electronic device based on knowledge graph, so as to provide training resource recommendations to employees in a more systematic and accurate manner, so as to make the company's existing talent knowledge graph closer to the target talent knowledge graph.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: The first aspect of the present invention discloses a method for recommending training resources based on a knowledge graph, comprising the following steps: Build the enterprise's target talent knowledge map; Obtain the company's existing talent knowledge graph; Matching the target talent knowledge graph with the existing talent knowledge graph using a graph matching algorithm, and calculating the matching degree; Recommending training resources to enterprise employees based on the matching results of the graph matching algorithm to improve the matching degree between the existing talent knowledge graph and the target talent knowledge graph; The existing talent knowledge graph is updated and the matching degree is recalculated until the matching degree is higher than a preset threshold.

[0007] Optionally, the method for constructing the target talent knowledge graph of the enterprise includes: Obtain the company's strategic planning documents, job descriptions, and industry competency standard library, and extract the core competency dimension set of the target position; Construct a capability framework based on ontology modeling, define capability nodes and their hierarchical relationships, and annotate the pre-dependencies and synergy relationships between capabilities; Combined with the enterprise organizational structure information, establish a mapping relationship between departments and target job capabilities, and impose departmental ownership constraints on capability nodes; Link job competencies with the skill certification system in the industry knowledge base, and set the skill maturity level as the dynamic weight of the competency node; Based on the knowledge graph optimization algorithm, the rationality of the ability path is detected, and the graph topology structure is updated to finally generate the target talent knowledge graph containing a timestamp version.

[0008] Optionally, the method for constructing the enterprise's existing talent knowledge graph includes: Obtain employee position information, skill certificate data and 360-degree evaluation results from the enterprise HR system, and build a basic attribute set with employee ID as the core; Analyze employees’ historical project documents and code repositories, extract the frequency of use of technology stacks and task role labels, and use them as dynamic capability indicators; Build the relationship between employees, skills and projects, and calculate skill mastery based on skill certificate level, project application times and peer review scores; The performance evaluation data is mapped to skill nodes based on the graph embedding algorithm, and the implicit skill association is generated using the collaborative filtering algorithm; The ability decay coefficient of skill nodes is calculated using the dynamic time-series graph modeling method, and the impact of cross-departmental collaboration on skill level is analyzed through the graph attention mechanism; Integrate multi-source heterogeneous data, perform conflict detection, and generate a versioned knowledge graph of existing talents based on preset weight priority rules.

[0009] Optionally, the method for calculating the matching degree specifically includes: Using graph neural networks, the target talent knowledge graph and the existing talent knowledge graph are encoded into vector representations, while retaining the structure and semantic information of the graphs. Based on the vector representation, the global similarity of the two graphs is calculated as an initial matching degree; Using graph matching algorithms, identify key subgraph structures in the target talent knowledge graph and the existing talent knowledge graph, including local patterns of entities and relationships; and calculate local matching degrees through entity alignment and relationship matching; The initial matching degree and the local matching degree are fused to generate a final matching degree, wherein the fusion method includes weighted averaging or machine learning model optimization.

[0010] Optionally, the method for calculating the matching degree specifically includes: Initialize the graph matching model and use the graph neural network to calculate the matching degree between the target talent’s knowledge graph and the existing employee’s knowledge graph as the preliminary matching result; Execute actions and update matching results, using the reinforcement learning model to select actions to optimize graph matching, including adjusting the weights of the matching algorithm or updating employee skill calibration, and recalculating the matching degree through the graph matching algorithm; Calculate rewards and feed them back to the reinforcement learning model. Depending on the degree of improvement in the matching results, positive or negative rewards are given according to the reward function to optimize the matching strategy. Update strategy, based on reward feedback, use Q-learning or deep Q network or Actor-Critic method to update the strategy of reinforcement learning model, with the goal of maximizing future cumulative rewards; Repeated training and updating to improve the accuracy and adaptability of graph matching.

[0011] Optionally, the Actor-Critic method includes: Use Actor network computing strategy; Use the Critic network evaluation strategy to calculate the state-action value function; The Actor network is optimized and adjusted based on the feedback from the Critic network.

[0012] Optionally, the recommending of training resources to enterprise employees according to the matching result of the graph matching algorithm to improve the matching degree between the existing talent knowledge graph and the target talent knowledge graph specifically includes: Based on the matching degree, identify the differences between the existing talent knowledge map and the target talent knowledge map, and determine the skills gaps of employees; Build a knowledge graph of training resources and define the relationship between training resources and skills; Use recommendation algorithms to generate training resource recommendations for employees based on skill gaps.

[0013] Optionally, the recommendation algorithm includes the following steps: The global influence of the target skill node in the strategic map is calculated by improving the PageRank algorithm, and a priority ranking is generated; Use graph attention network to model employees’ historical learning paths and extract implicit learning pattern features; The meta-path reasoning method is used to mine multi-hop association paths between gap skill nodes and training resources, and the path confidence scores are calculated; Introducing the multi-armed bandit algorithm, dynamically adjusting the weight distribution of collaborative filtering recommendations and knowledge-driven recommendations, and updating the reward function based on real-time employee feedback data; To address cross-departmental skill gaps, a transfer learning model is established to project successful training programs of high-matching departments into the feature space of low-matching departments. The recommendation results are encoded into a training plan graph with time and space constraints. The nodes represent training tasks with time windows, and the edges represent the pre- and post-logical relationships between tasks and the skill gain transfer function, ultimately generating a personalized course path planning sequence.

[0014] The second aspect of the present invention discloses a training resource recommendation system based on knowledge graph, comprising: The target talent knowledge graph construction module is used to construct the target talent knowledge graph of the enterprise; The existing talent knowledge graph construction module is used to obtain the enterprise's existing talent knowledge graph; A graph matching module, used to match the target talent knowledge graph with the existing talent knowledge graph using a graph matching algorithm, and calculate a matching degree; An iterative update module is used to recommend training resources to enterprise employees based on the matching results of the graph matching algorithm, update the existing talent knowledge graph, and repeatedly execute the graph matching module until the matching degree is higher than a preset threshold.

[0015] As a third aspect of the present invention, an electronic device is disclosed, comprising: A memory and a processor, wherein the memory is used to store a computer program; The processor is used to execute any one of the above methods when calling the computer program.

[0016] The advantage of the present invention over the prior art is that, by utilizing knowledge graphs and graph matching algorithms, the accuracy and flexibility of enterprises in talent management can be effectively improved. In traditional human resource management, enterprises often face difficulties in evaluating the capabilities of existing employees, especially in a constantly changing market and technical environment. Using the present invention, a target talent knowledge graph that is highly consistent with the development goals of the enterprise can be accurately constructed, and the matching degree between the target and the target can be ensured to be continuously improved by continuously updating the existing talent knowledge graph. In this way, enterprises can promptly discover the gaps in employee skills and recommend training resources in a targeted manner, thereby helping employees to improve their capabilities that are more consistent with the future development direction of the enterprise. In addition, through iterative optimization, the training needs of employees are continuously adjusted dynamically, which can effectively avoid the inefficient use of a single training resource and improve the overall training efficiency. Using advanced technologies such as graph neural networks, the potential correlation in the talent knowledge graph can be mined from a deeper level, making the recommended training resources more accurate and targeted. Employees can receive customized training based on their individual skill gaps, thereby not only improving their personal abilities, but also providing a steady stream of support for the long-term development of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is a flow chart of constructing the target talent knowledge graph of the enterprise according to the method of the present invention; Figure 3 It is a flow chart of constructing the existing talent knowledge graph of the enterprise according to the method of the present invention; Figure 4 is a flow chart of the first graph matching method of the present invention; Figure 5 is a flow chart of a second graph matching method of the present invention; Figure 6 is a flow chart of a recommended algorithm of the method of the present invention; Figure 7 It is a schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0018] The specific implementation of the present invention will be described below in conjunction with the accompanying drawings.

[0019] The first aspect of the present invention is as follows Figure 1 The figure shows a training resource recommendation method based on knowledge graph, which includes the following steps: Build the enterprise's target talent knowledge map; Obtain the company's existing talent knowledge graph; Matching the target talent knowledge graph with the existing talent knowledge graph using a graph matching algorithm, and calculating the matching degree; Recommending training resources to enterprise employees based on the matching results of the graph matching algorithm to improve the matching degree between the existing talent knowledge graph and the target talent knowledge graph; The existing talent knowledge graph is updated and the matching degree is recalculated until the matching degree is higher than a preset threshold.

[0020] like Figure 2 The following is a method for constructing a target talent knowledge graph for an enterprise. The construction process starts with data acquisition. Specifically, information needs to be extracted from the enterprise's strategic planning documents (such as a three-year development blueprint), job descriptions (such as job descriptions for data scientists), and industry capability standard libraries (such as skill standards for the IT industry).

[0021] Through natural language processing techniques, such as named entity recognition and keyword extraction, the core capabilities required for the target position can be identified.

[0022] For example, suppose a company plans to develop artificial intelligence business, the strategic plan mentions "strengthening AI technical capabilities", the job description lists the responsibility of "machine learning model development", and the industry standard library defines "deep learning" skills. After analyzing these data, a core set of capabilities can be extracted, such as {machine learning, data analysis, programming capabilities}.

[0023] If the company's goal is to become a "smart manufacturing leader", it may extract "automation control" capabilities from strategic documents, "PLC programming" skills from job descriptions, and "Industrial Internet of Things" knowledge from industry standards to form a capability set {automation control, PLC programming, Industrial Internet of Things}.

[0024] Next, based on the ontology modeling method, a framework can be built for these core capabilities.

[0025] Capabilities are defined as nodes in the graph, and their hierarchical relationships are labeled. For example, "programming ability" can be divided into basic layers (such as Python syntax) and advanced layers (such as algorithm optimization). It is also necessary to indicate the pre-dependencies between capabilities (such as the need to master "machine learning basics" before learning "deep learning") and synergistic relationships (such as the combination of "data analysis" and "business understanding" can improve decision support capabilities).

[0026] Taking the position of "data scientist" as an example, the capability framework may include the basic node "Python programming" pointing to the advanced node "machine learning", and the prerequisite is "must master Python programming first"; and there is a synergy between "data visualization" and "statistical analysis", which can improve data insight capabilities together.

[0027] On the basis of this framework, we also need to combine the organizational structure of the enterprise (such as R&D department, marketing department) to add departmental affiliation constraints to the capability nodes. The R&D department may need "algorithm development" capabilities, and the marketing department may need "market research" capabilities. In a technology company, the "AI engineer" position in the R&D department may be mapped to {machine learning, deep learning}, and the "account manager" in the sales department may be mapped to {communication skills, customer relationship management}.

[0028] In addition, the internal competency definition of the enterprise should be aligned with the industry skill certification system (such as CompTIA certification, AWS certification), and skill maturity levels (such as elementary, intermediate, and advanced) should be set for competency nodes as dynamic weights. For example, for the "cloud computing" capability, if an employee holds the "AWS Certified Solutions Architect" certificate, it can be marked as "advanced" with a weight of 0.9; if only basic training is completed, it can be marked as "elementary" with a weight of 0.3.

[0029] Finally, the graph is optimized and a versioned target knowledge graph is generated. The rationality of the capability path is checked using knowledge graph optimization algorithms (such as path consistency detection) to ensure that, for example, the "deep learning" path does not lack pre-nodes such as "linear algebra". After optimization, a timestamp is added to the graph (such as "2023-10-01_v1.0") to facilitate tracking of changes after corporate strategy adjustments. If it is found that the "big data processing" path lacks pre-capabilities such as "distributed systems", the nodes are automatically supplemented and the topology is updated to eventually generate a complete target talent knowledge graph.

[0030] like Figure 3 The figure shows the construction method of the existing talent knowledge map of the enterprise of the present invention. In a specific embodiment, the first step of constructing this map is to extract the basic data of employees from the enterprise HR system, including job information (such as job title), skill certificate data (such as certificate name and level) and 360-degree evaluation results (such as evaluations of leaders and colleagues).

[0031] With employee ID as the core, these data are integrated to form a basic attribute set. For example, the attribute set of employee A may be {Position: Data Analyst, Certificate: TensorFlow Certification, Evaluation Score: 4.5 / 5}; employee B may be {Position: Software Engineer, Certificate: Java SE 8, Evaluation Score: 4.0 / 5}.

[0032] By analyzing employees’ project experience, dynamic capability indicators can be extracted. Using natural language processing and code analysis technology, employees’ historical project documents and code repositories can be analyzed to obtain the frequency of use of technology stacks (for example, Python accounts for 70%) and task role labels (for example, “front-end development”).

[0033] For example, employee A has participated in three projects in the past year. The code repository shows that he used Python and Pandas 80% of the time. The document marked his role as "data engineer", from which it can be inferred that he has "data processing" capabilities.

[0034] The relationship between employees, skills and projects is stored in the graph database, such as "employee A-participation-project B-use-skill C". The skill mastery is calculated by weighted average based on the skill certificate level (weight 40%), project application times (weight 40%) and peer review score (weight 20%).

[0035] For example, employee A’s mastery of “machine learning” can be calculated as follows: (Certificate Intermediate 0.5 × 40%) + (Project 5 0.8 × 40%) + (Review 4.5 / 5 × 20%) = 0.67. Employee B's mastery of "Java Programming" is based on the certificate advanced (0.9), 10 projects (1.0) and review 4.0 / 5, and the result is 0.92.

[0036] In addition, graph embedding algorithms (such as DeepWalk) can be used to map performance data to skill nodes, generate low-dimensional vector representations, and infer implicit associations based on employee skill similarities. For example, if employees A and B are both good at "Python programming", it can be inferred that B may also have A's "data analysis" skills. Employee C has high performance and is good at "front-end development", and collaborative filtering may be used to infer that he has "UI design" capabilities.

[0037] Dynamic time-series graph modeling can also calculate the attenuation coefficient of skills. For example, if the "SQL" skill is not used for a year, the mastery will decay by 10%. The graph attention mechanism can be used to analyze the impact of cross-departmental collaboration on skills. For example, after employee A cooperates with the marketing department, his "communication ability" improves by 20%.

[0038] After integrating these data, if there is a conflict in skill assessment between HR data and project data (for example, HR is marked as "junior" and project is shown as "senior"), it will be processed according to the priority rules with preset weights (HR 40%, project 60%), and finally a knowledge graph of existing talents with a timestamp will be generated, such as "2023-10-01_v1.0".

[0039] The next step is to use a graph matching algorithm to compare the target and existing talent knowledge graphs and calculate their matching degree, with the aim of quantifying the gap between existing talent and target talent.

[0040] like Figure 4As shown, one implementation method is to use graph neural network GNN to encode two graphs into vectors, retaining structural information (such as capability path) and semantic information (such as skill weight).

[0041] GNN can aggregate the information of each node and its neighbors through multi-layer convolution operations. For example, in the target talent knowledge graph, the "data mining" node may be connected to nodes such as "statistics" and "programming ability". These connections will be encoded as part of the vector, reflecting the position and importance of the node in the graph.

[0042] Node attributes, such as skill proficiency (e.g., "data mining" proficiency is 0.8), department or job requirements, are also included in the encoding process. For example, in an existing employee graph, "data mining" is 0.5, and in the target graph it is 0.8, and this difference will be reflected in the value of the vector.

[0043] In a specific example, suppose the encoding vector of the "data mining" node in the target graph is [0.8, 0.4, 0.6], while that in the existing graph is [0.5, 0.3, 0.4]. These two vectors are generated by GNN, containing both structural and semantic information, laying the foundation for subsequent similarity calculations.

[0044] After obtaining the vector representation, the global similarity of the two graphs is calculated based on these vectors as the initial matching degree. Global similarity reflects the overall fit of the graphs. Cosine similarity is usually used as a metric because it can effectively measure the consistency of vector directions in high-dimensional space. The calculation method can use the formula of cosine similarity: ; in and are the encoding vectors of the target and existing graphs, respectively.

[0045] Although global similarity can provide an overall perspective, it may mask the differences in local structures. To this end, the present invention introduces a graph matching algorithm to identify key subgraph structures in the target and existing graphs and calculate local matching.

[0046] The important local patterns in the graph can be extracted through subgraph isomorphism or pattern mining technology. For example, the target graph may have a skill path of "programming ability → data mining → machine learning", while the existing graph may be "programming ability → data analysis → visualization". These paths reflect the dependencies of skills.

[0047] After identifying the subgraphs, the corresponding nodes are aligned by calculating the attribute-based similarity. For example, "data mining" and "data analysis" may be aligned due to functional similarity, but the difference in proficiency will result in a lower match. Relationship matching focuses on the consistency of edges, such as whether the dependencies from "programming ability" to subsequent skills in the two graphs are the same. Local matching can be calculated by weighted summing the similarities of aligned entities and relationships.

[0048] For example: suppose the subgraph of the target graph is "Programming ability → Data mining (0.8)", and the existing graph is "Programming ability → Data analysis (0.6)".

[0049] The “Programming Ability” nodes are completely aligned, with a similarity of 1; "Data mining" and "data analysis" have similar functions but different proficiency, and the similarity is assumed to be 0.7; the relationship between the edges is consistent, and the similarity is 1. The local matching degree may be: ; This value indicates that the local structures are relatively similar, but there are some differences.

[0050] In order to fully reflect the matching degree of the graph, the global similarity (initial matching degree) and the local matching degree can be fused to generate the final matching degree. The fusion method can be either a simple weighted average or optimized with the help of a machine learning model.

[0051] The weighted average method is: ; Here, α is an adjustable weight, for example, set to 0.6 to give more emphasis to global similarity.

[0052] Machine learning optimization can use historical matching data to train a regression model, input global and local matching, and output the final matching. This method can capture the nonlinear relationship between the two. For example, after training, the model may believe that local differences have a greater impact on certain positions, and thus adjust the weights.

[0053] In addition to static matching degree calculation, the present invention also proposes a dynamic matching degree optimization method based on reinforcement learning.

[0054] like Figure 5 As shown in the figure, this method improves the accuracy of matching by continuously learning and adjusting strategies to adapt to the complexity and dynamic changes of the graph structure.

[0055] In the initial stage, the graph neural network (GNN) can be used to calculate the preliminary matching degree between the target talent knowledge graph and the existing employee knowledge graph as the benchmark result. The specific process is similar to the above method, generating a vector representation through GNN encoding, and then calculating the cosine similarity. For example, the preliminary matching degree may be 0.82, indicating the similarity of the two graphs in the initial state.

[0056] Subsequently, a reinforcement learning model is introduced to optimize the matching results by performing specific actions, such as adjusting the weight parameters of the matching algorithm or updating the employee's skill rating, and recalculating the matching degree after each action.

[0057] For example, the weight of the "machine learning" skill in GNN encoding is increased from 1.0 to 1.3 because it is more important in the target graph. Based on the new evaluation data, the "machine learning" proficiency of employee B is updated from 0.5 to 0.7.

[0058] During the update process, the GNN and similarity calculations are rerun after adjustment, and the match may be improved from 0.82 to 0.85. This dynamic adjustment can better reflect the actual situation. To guide the reinforcement learning model, a reward function can be designed to give feedback based on the change in match. The reward is usually defined as the difference between the new match and the old match. For example, if it increases from 0.82 to 0.85, the reward is 0.03; if it drops to 0.80, the reward is -0.02. The reward function can also be combined with business goals, such as if the action improves the match of key skills, an additional reward is added.

[0059] Next, based on reward feedback, a reinforcement learning algorithm can be used to update the strategy, with the goal of maximizing long-term cumulative rewards. Common algorithms include Q-learning, deep Q network (DQN), and Actor-Critic methods. In Q-learning, the value of each state-action pair is recorded by updating the Q table. For example, increasing the "machine learning" weight to obtain a positive reward will increase its Q value. In DQN, a neural network is used to fit the Q function, which is suitable for complex scenarios. In Actor-Critic, strategy selection and value evaluation are combined, as shown in the following example. After an update, the model may be more inclined to adjust the weights of key skills.

[0060] Through multiple rounds of action execution, reward feedback, and strategy adjustment, the matching degree is gradually optimized. Training can continue until the matching degree stabilizes or reaches a preset threshold. For example, after 15 rounds of iterations, the matching degree increased from 0.82 to 0.89. The model has learned a better matching strategy and its adaptability has been significantly enhanced.

[0061] In the above embodiment, the Actor-Critic method is an efficient strategy update method that combines the advantages of policy gradient and value function and is suitable for dynamically adjusting the matching process. The Actor network can select optimization actions based on the current graph matching state. The state may be the current matching degree and graph features, and the action includes adjusting weights or skill calibration values. The Actor network may output a probability distribution of actions, such as the probability of "adjusting the 'data mining' weight to 1.2" is 0.7. For example, when the current matching degree is 0.85, the Actor recommends increasing the "data mining" weight.

[0062] You can use the Critic network to evaluate the strategy. The Critic network evaluates the action selected by the Actor and calculates the state-action value function (Q value), which reflects the expected benefit of the action. Evaluation process example: For "adjust the 'data mining' weight to 1.2", the Critic may output a Q value of 0.04, indicating an expected increase in matching by 0.04. For example, if the matching degree rises to 0.87 after the action, the Critic verifies its effect.

[0063] Based on the Critic's evaluation, the Actor adjusts parameters through gradient ascent to increase the probability of high Q-value actions. The optimization logic is that when the Q value is positive, the selection tendency of the action is strengthened; if it is negative, the probability is reduced. For example, because the Q value is 0.04, the Actor increases the tendency to adjust the "data mining" weight and optimizes the matching strategy. This collaborative mechanism enables the Actor-Critic method to perform well in complex matching scenarios.

[0064] After calculating the matching degree, training resources can be further recommended to employees based on the matching results to narrow the gap between the existing talent knowledge map and the target map. This process includes skill gap identification, resource association construction, and recommendation plan generation.

[0065] Specifically, the matching results of the graph matching algorithm can reveal the specific gap between the existing talent knowledge graph and the target talent knowledge graph. This gap is mainly reflected in the coverage and proficiency of skills. By comparing the attribute values ​​of skill nodes in the two graphs, such as proficiency or importance, the skill gaps of employees can be accurately located. For example, assuming that the target graph requires the proficiency of "cloud computing" skills to reach 0.8, and the current proficiency of an employee is only 0.5, then this 0.3 gap becomes the basis for recommending training resources.

[0066] Similarly, in the reinforcement learning strategy of the second method mentioned above, by observing the adjustment strategy of the existing talent knowledge graph, we can know what kind of adjustment strategy can improve the matching degree. For example, if in the existing talent knowledge graph, after improving the proficiency of a certain employee's "cloud computing" skills in the existing talent knowledge graph, the overall matching degree will be improved, then in the subsequent recommended training resources, the "cloud computing" skills can be recommended.

[0067] In order to effectively connect skill gaps with training resources, it is necessary to build a training resource knowledge graph. The purpose of this graph is to associate various types of training resources - such as online courses, practical workshops or professional certification exams - with specific skill nodes. The establishment of associations can rely on the content introduction of the resource, learning objectives, or learning effects reflected in historical data. For example, a "Basics of Cloud Computing Course" may be labeled as being able to improve the proficiency of "Cloud Computing" skills by 0.2. In this way, the training resource knowledge graph provides a structured resource library for the recommendation process, allowing the system to quickly find matching training content based on skill requirements.

[0068] With the foundation of skill gaps and resource maps, the recommendation algorithm can generate personalized training resource recommendation plans. The algorithm can integrate multiple factors, including the degree of skill gaps, the strength of the correlation between training resources and skills, and the individual learning habits of employees. For example, if an employee has a shortcoming in the skill of "data visualization", the algorithm will select highly relevant courses from the resource map, and at the same time consider the employee's past habit of preferring video teaching rather than text reading, and finally recommend an "advanced data visualization video course". Such recommendations can not only effectively fill skill gaps, but also improve employees' learning experience and results.

[0069] In the design of recommendation algorithms, a fusion of multiple technologies can be used.

[0070] like Figure 6 As shown in the figure, the execution process of the entire algorithm is guided by the strategic goals of the enterprise, while taking into account the personalized needs of employees and the knowledge sharing needs across departments. Specifically, the algorithm can evaluate the importance of the target skills in the enterprise strategy map through an improved PageRank method. This method imagines the skill nodes as "pages" in a network, and the dependencies between skills are similar to "links", and then superimposed with the weight factors of the enterprise strategy. For example, in a company focusing on "smart manufacturing", the "Industrial Internet of Things" skill may get a higher influence score because of its strategic value. Based on such analysis, the algorithm can determine which skill gaps have a greater impact on the enterprise as a whole, and thus give priority to recommending relevant training resources.

[0071] At the same time, the algorithm uses the graph attention network to conduct an in-depth analysis of employees' past learning paths and unearth implicit learning pattern characteristics. This network can capture employees' preferences by focusing on key nodes in the learning history, such as preferring hands-on practice rather than theoretical learning. This personalized insight makes the recommendation results more in line with employees' actual needs and improves the participation and completion rate of training. In addition, in order to more accurately connect skill gaps with training resources, the algorithm can also use the meta-path reasoning method. This method can mine multi-level association paths in heterogeneous graphs, such as "employee-lack-skill A-can be improved through-course B", and assign a confidence score to each path based on the effect of historical data, and finally filter out the most reliable resource options.

[0072] In order to make the recommendation strategy more dynamic, a multi-armed bandit mechanism can also be introduced. This mechanism can balance between collaborative filtering recommendations and knowledge-driven recommendations, and continuously adjust according to real-time feedback from employees. For example, collaborative filtering recommends resources based on similarities between employees, while knowledge-driven recommendations rely directly on the relationship between resources and skills. In the initial stage, the two methods may each account for half of the weight, but if feedback shows that employees prefer knowledge-driven recommendations, the algorithm will gradually increase the proportion of this part. This adaptive adjustment allows the recommendation system to be continuously optimized as it is used.

[0073] In view of the differences in skill requirements across departments, a transfer learning model can also be designed. This model can adjust the successful training experience of departments with higher skill matching and apply it to departments with lower matching. For example, the technical department's "artificial intelligence" training program has a significant effect. Through transfer learning, it can be quickly adapted to a version suitable for the sales department to help sales staff master relevant knowledge. This method not only improves training efficiency, but also promotes experience sharing within the company.

[0074] Finally, the algorithm will integrate the recommendation results into a training plan diagram with time and space constraints. This diagram shows specific training tasks in the form of nodes, and each node is marked with the time window of the task, such as "complete the basic course from October 10 to 20"; and the edges connecting the nodes reflect the logical order between tasks and the transfer relationship of skill improvement, such as completing the basic course before entering the advanced stage. In this way, employees can obtain a clear learning path that takes into account the rationality of time arrangement and ensures the continuity of skill improvement.

[0075] like Figure 7 As shown, the second aspect of the present invention discloses a training resource recommendation system based on knowledge graph, which can be used to implement the above method, including: The target talent knowledge graph construction module is used to construct the target talent knowledge graph of the enterprise; The existing talent knowledge graph construction module is used to obtain the enterprise's existing talent knowledge graph; A graph matching module, used to match the target talent knowledge graph with the existing talent knowledge graph using a graph matching algorithm, and calculate a matching degree; An iterative update module is used to recommend training resources to enterprise employees based on the matching results of the graph matching algorithm, update the existing talent knowledge graph, and repeatedly execute the graph matching module until the matching degree is higher than a preset threshold.

[0076] More specifically, the implementation of the system can rely on modular design, dividing the functions into multiple independent parts, and the entire system can be deployed on a cloud computing platform, such as AWS or Alibaba Cloud. In terms of hardware configuration, high-performance servers can be used, equipped with multi-core CPUs, large-capacity RAM, and SSD storage to ensure that the graph database and machine learning models can run efficiently.

[0077] In terms of programming implementation, a variety of languages ​​and technologies can be used. Python can develop algorithms and models with its rich library support (such as TensorFlow, PyTorch, and NLTK). Java is used to build high-concurrency backend services to ensure the stability of the system under multi-user access. The development of the user interface can rely on JavaScript, which is implemented through the React framework to provide a smooth interactive experience. At the data storage level, Neo4j can be used as a graph database to manage knowledge graphs. It is good at processing complex node and relationship data; at the same time, PostgreSQL is used to store metadata and user interaction records to provide reliable management of structured data. During the development process, you can follow the agile development method, use Git for version control, and use Jenkins for continuous integration and deployment.

[0078] Therefore, the third aspect of the present invention discloses an electronic device, comprising: A memory and a processor, wherein the memory is used to store a computer program; The processor is used to execute the above method when calling the computer program.

[0079] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A training resource recommendation method based on knowledge graph, characterized in that: The following steps are involved: Build the enterprise's target talent knowledge map; Obtain the company's existing talent knowledge graph; Matching the target talent knowledge graph with the existing talent knowledge graph using a graph matching algorithm, and calculating the matching degree; Recommending training resources to enterprise employees based on the matching results of the graph matching algorithm to improve the matching degree between the existing talent knowledge graph and the target talent knowledge graph; The existing talent knowledge graph is updated and the matching degree is recalculated until the matching degree is higher than a preset threshold.

2. The method for recommending training resources based on knowledge graph according to claim 1, characterized in that: The method for constructing the target talent knowledge graph of an enterprise includes: Obtain the company's strategic planning documents, job descriptions, and industry competency standard library, and extract the core competency dimension set of the target position; Construct a capability framework based on ontology modeling, define capability nodes and their hierarchical relationships, and annotate the pre-dependencies and synergy relationships between capabilities; Combined with the enterprise organizational structure information, establish a mapping relationship between departments and target job capabilities, and impose departmental ownership constraints on capability nodes; Link job competencies with the skill certification system in the industry knowledge base, and set the skill maturity level as the dynamic weight of the competency node; Based on the knowledge graph optimization algorithm, the rationality of the ability path is detected, and the graph topology structure is updated to finally generate the target talent knowledge graph containing a timestamp version.

3. The method for recommending training resources based on knowledge graph according to claim 1, characterized in that: The method for constructing the enterprise's existing talent knowledge graph includes: Obtain employee position information, skill certificate data and 360-degree evaluation results from the enterprise HR system, and build a basic attribute set with employee ID as the core; Analyze employees’ historical project documents and code repositories, extract the frequency of use of technology stacks and task role labels, and use them as dynamic capability indicators; Build the relationship between employees, skills and projects, and calculate skill mastery based on skill certificate level, project application times and peer review scores; The performance evaluation data is mapped to skill nodes based on the graph embedding algorithm, and the implicit skill association is generated using the collaborative filtering algorithm; The ability decay coefficient of skill nodes is calculated using the dynamic time-series graph modeling method, and the impact of cross-departmental collaboration on skill level is analyzed through the graph attention mechanism; Integrate multi-source heterogeneous data, perform conflict detection, and generate a versioned knowledge graph of existing talents based on preset weight priority rules.

4. The method for recommending training resources based on knowledge graph according to claim 1, characterized in that: The method for calculating the matching degree specifically includes: Using graph neural networks, the target talent knowledge graph and the existing talent knowledge graph are encoded into vector representations, while retaining the structure and semantic information of the graphs. Based on the vector representation, the global similarity of the two graphs is calculated as an initial matching degree; Using graph matching algorithms, identify key subgraph structures in the target talent knowledge graph and the existing talent knowledge graph, including local patterns of entities and relationships; and calculate local matching degrees through entity alignment and relationship matching; The initial matching degree and the local matching degree are fused to generate a final matching degree, wherein the fusion method includes weighted averaging or machine learning model optimization.

5. The method for recommending training resources based on knowledge graph according to claim 1, characterized in that: The method for calculating the matching degree specifically includes: Initialize the graph matching model and use the graph neural network to calculate the matching degree between the target talent’s knowledge graph and the existing employee’s knowledge graph as the preliminary matching result; Execute actions and update matching results, using the reinforcement learning model to select actions to optimize graph matching, including adjusting the weights of the matching algorithm or updating the employee skill calibration, and recalculating the matching degree through the graph matching algorithm; Calculate rewards and feed them back to the reinforcement learning model. Depending on the degree of improvement in the matching results, positive or negative rewards are given according to the reward function to optimize the matching strategy. Update strategy, based on reward feedback, use Q-learning or deep Q network or Actor-Critic method to update the strategy of reinforcement learning model, with the goal of maximizing future cumulative rewards; Repeated training and updating to improve the accuracy and adaptability of graph matching.

6. The method for recommending training resources based on knowledge graph according to claim 5, characterized in that: The Actor-Critic method includes: Use Actor network computing strategy; Use the Critic network evaluation strategy to calculate the state-action value function; The Actor network is optimized and adjusted based on the feedback from the Critic network.

7. The method for recommending training resources based on knowledge graph according to claim 1, characterized in that: The recommending of training resources to enterprise employees according to the matching results of the graph matching algorithm to improve the matching degree between the existing talent knowledge graph and the target talent knowledge graph specifically includes: Based on the matching degree, identify the differences between the existing talent knowledge map and the target talent knowledge map, and determine the skills gaps of employees; Build a knowledge graph of training resources and define the relationship between training resources and skills; Use recommendation algorithms to generate training resource recommendations for employees based on skill gaps.

8. The method for recommending training resources based on knowledge graph according to claim 7, characterized in that: The recommendation algorithm includes the following steps: The global influence of the target skill node in the strategic map is calculated by improving the PageRank algorithm, and a priority ranking is generated; Use graph attention network to model employees’ historical learning paths and extract implicit learning pattern features; The meta-path reasoning method is used to mine multi-hop association paths between gap skill nodes and training resources, and the path confidence scores are calculated; Introducing the multi-armed bandit algorithm, dynamically adjusting the weight distribution of collaborative filtering recommendations and knowledge-driven recommendations, and updating the reward function based on real-time employee feedback data; To address cross-departmental skill gaps, a transfer learning model is established to project successful training programs of high-matching departments into the feature space of low-matching departments. The recommendation results are encoded into a training plan graph with time and space constraints. The nodes represent training tasks with time windows, and the edges represent the pre- and post-logical relationships between tasks and the skill gain transfer function, ultimately generating a personalized course path planning sequence.

9. A training resource recommendation system based on knowledge graph, characterized in that: include: The target talent knowledge graph construction module is used to construct the target talent knowledge graph of the enterprise; The existing talent knowledge graph construction module is used to obtain the enterprise's existing talent knowledge graph; A graph matching module, used to match the target talent knowledge graph with the existing talent knowledge graph using a graph matching algorithm, and calculate a matching degree; An iterative update module is used to recommend training resources to enterprise employees based on the matching results of the graph matching algorithm, update the existing talent knowledge graph, and repeatedly execute the graph matching module until the matching degree is higher than a preset threshold.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the memory is used to store a computer program; The processor is configured to execute the method according to any one of claims 1 to 8 when calling the computer program.

Images