CN119903242A

CN119903242A - Training resource recommendation method, system and electronic device based on knowledge graph

Info

Publication number: CN119903242A
Application number: CN202510387935.8A
Authority: CN
Inventors: 戴炯
Original assignee: Jiangsu Renjia Information Technology Co ltd
Current assignee: Jiangsu Renjia Information Technology Co ltd
Priority date: 2025-03-31
Filing date: 2025-03-31
Publication date: 2025-04-29

Abstract

The present invention discloses a training resource recommendation method, system and electronic device based on knowledge graph, which relates to the field of enterprise data processing, including constructing a target talent knowledge graph of an enterprise; obtaining the enterprise'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; repeating the iteration until the matching degree is higher than a preset threshold. The present invention can help enterprises systematically evaluate and improve the capabilities of employees and better adapt to market demand and enterprise development goals.

Description

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

Technical Field

The invention relates to the field of enterprise data processing, in particular to a training resource recommendation method and system based on a knowledge graph and electronic equipment.

Background

In today's rapidly changing business environment, enterprises face the challenge of continually adjusting and optimizing talent structures to accommodate market demands. Especially in the development process of enterprises, how to accurately evaluate whether the existing talent framework is matched with the strategic targets and the future development direction of the enterprises becomes a key problem in the management of human resources of the enterprises.

However, enterprises often cannot accurately assess whether existing staff's skills and knowledge are consistent with their long-term development strategy or market needs. Traditional assessment methods rely on static job requirements and self-reporting, and it is difficult to comprehensively reflect the actual abilities of staff, and especially in the industry of rapid development of technology and business, the defects of the existing talent frameworks are obvious.

Even if the enterprise is aware of the deficiency of the skills and knowledge of the existing staff, the traditional training method often has no systematicness and cannot be used for training from the whole enterprise. An enterprise is composed of a plurality of employees, and the abilities of the employees systematically form the abilities of an enterprise, so that the recommendation of training resources of the employees of the whole enterprise should be more systematic, so that the human resource structure of the whole enterprise approaches to the target human resource structure direction, however, no effective method is available at present to achieve the goal.

Disclosure of Invention

The invention aims to provide a training resource recommending method, a training resource recommending system and electronic equipment based on a knowledge graph, so that recommendation of training resources is provided for staff more systematically and accurately, and the existing talent knowledge graph of an enterprise is closer to a target talent knowledge graph.

In order to achieve the above purpose, the invention adopts the following technical scheme:

the invention discloses a training resource recommendation method based on a knowledge graph, which comprises the following steps of:

Constructing a target talent knowledge graph of an enterprise;

acquiring the existing talent knowledge graph of an enterprise;

Matching the target talent knowledge graph with the existing talent knowledge graph by using a graph matching algorithm, and calculating the matching degree;

Training resource recommendation is carried out on enterprise staff according to the matching result of the graph matching algorithm so as to improve the matching degree of the existing talent knowledge graph and the target talent knowledge graph;

Updating the existing talent knowledge graph and recalculating the matching degree until the matching degree is higher than a preset threshold value.

Optionally, the method for constructing the target talent knowledge graph of the enterprise includes:

Acquiring a strategic planning document, a post instruction and an industry capability standard library of an enterprise, and extracting a core capability dimension set of a target post;

constructing a capability framework based on an ontology modeling method, defining capability nodes and hierarchical relationships thereof, and labeling preposed dependency relationships and synergistic relationship among capabilities;

establishing a mapping relation between departments and target post capabilities by combining enterprise organization architecture information, and applying department attribution constraint to capability nodes;

Associating the post capability with a skill authentication system in an industry knowledge base, and setting a skill maturity level as a dynamic weight of a capability node;

And detecting the rationality of the capacity path based on a knowledge graph optimization algorithm, updating the graph topological structure, and finally generating the target talent knowledge graph containing the time stamp version.

Optionally, the method for constructing the existing talent knowledge graph of the enterprise comprises the following steps:

staff post information, skill certificate data and 360-degree evaluation results in an enterprise HR system are obtained, and a basic attribute set taking staff ID as a core is constructed;

Analyzing employee history project documents and code warehouses, extracting the use frequency and task role labels of a technical stack, and taking the use frequency and task role labels as dynamic capability indexes;

Constructing the association relation between staff and skills and projects, and calculating the skill mastery degree based on the skill certificate level, the project application times and the peer review score;

mapping performance evaluation data to skill nodes based on a graph embedding algorithm, and generating implicit skill association by adopting a collaborative filtering algorithm;

calculating the capacity attenuation coefficient of the skill node by using a dynamic time sequence diagram modeling method, and analyzing the influence of cross-department collaboration on the skill level through a diagram attention mechanism;

and integrating the multi-source heterogeneous data, performing conflict detection, and generating a versioned existing talent knowledge graph according to a preset weight priority rule.

Optionally, the method for calculating the matching degree specifically includes:

coding a target talent knowledge graph and an existing talent knowledge graph into vector representation by using a graph neural network, and reserving structure and semantic information of the graph;

based on the vector representation, calculating the global similarity of the two maps to be used as an initial matching degree;

Identifying key sub-graph structures in the target talent knowledge graph and the existing talent knowledge graph by utilizing a graph matching algorithm, wherein the key sub-graph structures comprise local modes of entities and relations;

and fusing the initial matching degree and the local matching degree to generate a final matching degree, wherein the fusing method comprises weighted average or machine learning model optimization.

initializing a map matching model, and calculating the matching degree of the target talent knowledge map and the existing employee knowledge map by using a graph neural network to serve as a primary matching result;

performing actions and updating a matching result, and selecting actions for optimizing map matching by using the reinforcement learning model, wherein the actions comprise adjusting the weight of a matching algorithm or updating staff skill calibration, and recalculating the matching degree through the map matching algorithm;

Calculating rewards and feeding back the rewards to the reinforcement learning model, and giving positive or negative rewards according to the rewards function according to the improvement degree of the matching result so as to optimize the matching strategy;

Updating a strategy, based on reward feedback, of updating the reinforcement learning model by using a Q-learning or deep Q network or an Actor-Critic method, with the aim of maximizing future accumulated rewards;

and training and updating repeatedly to improve the accuracy and adaptability of map matching.

Optionally, the Actor-Critic method includes:

Calculating a strategy by using an Actor network;

calculating a state-action value function by using a Critic network evaluation strategy;

and optimizing and adjusting the Actor network according to feedback of the Critic network.

Optionally, training resource recommendation is performed on enterprise staff according to the matching result of the graph matching algorithm, so as to improve the matching degree of the existing talent knowledge graph and the target talent knowledge graph, and specifically includes:

According to the matching degree, identifying the difference between the existing talent knowledge graph and the target talent knowledge graph, and determining the skill gap of the staff;

constructing a training resource knowledge graph, and defining the association relationship between training resources and skills;

and generating a training resource recommendation scheme of the staff based on the skill gap by using a recommendation algorithm.

Optionally, the recommendation algorithm includes the following steps:

Calculating the global influence of the target skill node in the strategic map by improving the PageRank algorithm, and generating priority ranking;

Modeling an employee history learning path by using a drawing and meaning network, and extracting implicit learning mode features;

adopting a meta-path reasoning method to mine multi-hop associated paths between the gap skill nodes and training resources, and calculating path confidence scores;

introducing a multi-arm slot machine algorithm, dynamically adjusting weight distribution of collaborative filtering recommendation and knowledge driving recommendation, and updating a reward function according to employee real-time feedback data;

Aiming at the cross-department skill gap, a transfer learning model is established, and a successful training scheme of a high-matching-degree department is projected to a feature space of a low-matching-degree department;

And encoding the recommended result into a training plan graph with space-time constraint, wherein nodes represent training tasks with time windows, edges represent front-back logic relations and skill gain transfer functions among the tasks, and finally, a personalized course path planning sequence is generated.

In a second aspect of the present invention, a training resource recommendation system based on a knowledge graph is disclosed, comprising:

the target talent knowledge graph construction module is used for constructing a target talent knowledge graph of an enterprise;

The existing talent knowledge graph construction module is used for acquiring the existing talent knowledge graph of the enterprise;

the image matching module is used for matching the target talent knowledge graph with the existing talent knowledge graph by utilizing an image matching algorithm and calculating the matching degree;

and the iteration updating module is used for recommending training resources for enterprise staff according to the matching result of the graph matching algorithm, updating the existing talent knowledge graph and repeatedly executing the graph matching module until the matching degree is higher than a preset threshold.

As a third aspect of the present invention, there is disclosed an electronic apparatus comprising:

a memory and a processor, the memory for storing a computer program;

The processor is configured to perform any of the methods described above when the computer program is invoked.

Compared with the prior art, the method has the advantage that the accuracy and the flexibility of enterprises in talent management can be effectively improved by utilizing the knowledge graph and the graph matching algorithm. In traditional human resource management, enterprises often face difficulties in assessing the capabilities of existing employees, especially in changing market and technical environments. By using the method, the target talent knowledge graph which is matched with the enterprise development target in height can be accurately constructed, and the matching degree between the target talent knowledge graph and the target is ensured to be continuously improved by continuously updating the existing talent knowledge graph. By the method, the enterprise can timely find the gap in the skills of the staff and pertinently recommend training resources, so that the staff is helped to promote the ability of being more matched with the future development direction of the enterprise. In addition, through iterative optimization, the training requirements of staff are dynamically adjusted continuously, so that the inefficient use of single training resources can be effectively avoided, and the overall training efficiency is improved. By utilizing advanced technologies such as a graph neural network, potential relevance in a talent knowledge graph can be mined from a deeper level, so that recommended training resources are more accurate and targeted. Staff can receive customized training according to individual skill gaps, so that personal capacity is improved, and continuous support is provided for long-term development of enterprises.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a flow chart of the construction of a target talent knowledge graph of an enterprise according to the method of the invention;

FIG. 3 is a flow chart of the construction of a prior talent knowledge graph of an enterprise in accordance with the method of the present invention;

FIG. 4 is a flow chart of a first graph matching method of the present invention;

FIG. 5 is a flow chart of a second graph matching method of the present invention;

FIG. 6 is a flowchart of a recommended algorithm for the method of the present invention;

Fig. 7 is a schematic diagram of the system of the present invention.

Detailed Description

The following describes specific embodiments of the present invention with reference to the drawings.

The first aspect of the present invention is as shown in fig. 1, which is a training resource recommendation method based on a knowledge graph, comprising the following steps:

Constructing a target talent knowledge graph of an enterprise;

acquiring the existing talent knowledge graph of an enterprise;

Fig. 2 shows a method for constructing a target talent knowledge graph of an enterprise. The build process begins with the acquisition of data, specifically, the need to extract information from strategic planning documents of the enterprise (such as a three year development blueprint), post specifications (such as job responsibilities description of the data scientist), and industry capability criteria libraries (such as skill criteria of the IT industry).

The core capabilities required for the target post can be identified through natural language processing techniques such as named entity recognition and keyword extraction.

For example, assuming an enterprise planning develops artificial intelligence business, strategic planning mentions "enhanced AI technology capability," the job of "machine learning model development" is listed in the job specification, and "deep learning" skills are defined in the industry standard library. After analyzing the data, a core set of capabilities, such as { machine learning, data analysis, programming capabilities }, can be extracted.

If the goal of the enterprise is to become an "intelligent manufacturing leader," it is possible to extract "automation control" capability from strategic documents, "PLC programming" skills from post specifications, and "industry Internet of things" knowledge from industry standards, forming a capability set { automation control, PLC programming, industry Internet of things }.

Next, a framework can be built for these core capabilities based on the ontology modeling approach.

Capability is defined as nodes in a graph and their hierarchical relationships are noted. For example, "programming capability" can be divided into a base layer (like Python grammar) and a higher layer (like algorithm optimization), and it is also necessary to identify pre-dependencies between capabilities (like learning "machine learning basis" before learning "deep learning") and synergistic relationships (like "data analysis" and "business understanding" in combination to improve decision support capability).

Taking the post of data scientist as an example, the capability framework may comprise a basic node of Python programming pointing to an advanced node of machine learning, and the pre-dependence is that the Python programming needs to be mastered first, and a synergistic effect exists between data visualization and statistical analysis, so that the capability of data insight can be improved together.

Based on the framework, the constraint of department attribution is added to the capability node by combining with the organization architecture (such as research and development department and market department) of enterprises. The research and development department may require an "algorithm development" capability and the market department may require a "market research" capability. In a science and technology company, the "AI engineer" post of the research and development department may be mapped to { machine learning, deep learning }, while the "customer manager" of the sales department is mapped to { communication ability, customer relationship management }.

In addition, the capability definitions within the enterprise are aligned with industry skill certification authorities (e.g., compTIA certification, AWS certification) and skill maturity levels (e.g., primary, intermediate, high) are set for the capability nodes as dynamic weights. For example, for "cloud computing" capability, if an employee holds a "AWS CERTIFIED Solutions Architect" certificate, it may be labeled "advanced" with a weight of 0.9, and if only basic training is completed, it is labeled "primary" with a weight of 0.3.

And finally, optimizing the map and generating a versioned target knowledge map. Knowledge-graph optimization algorithms (such as path consistency detection) are used to check the rationality of capability paths, ensuring that pre-nodes such as "deep-learned" paths do not miss "linear algebra". The map is time stamped after optimization (such as 2023-10-01_v1.0), so that the change of the strategic adjustment of the enterprise can be conveniently tracked. If the 'big data processing' path is found to lack the prepositive capacity of a 'distributed system', nodes are automatically supplemented and the topological structure is updated, and finally, a complete target talent knowledge graph is generated.

FIG. 3 shows a construction method of the prior talent knowledge graph of the enterprise of the present invention. In particular embodiments, the first step in constructing this graph is to extract employee base data from the enterprise HR system, including post information (e.g., job name), skill certificate data (e.g., certificate name and level), and 360 degree assessment results (e.g., ratings of leadership and colleagues).

And integrating the data to form a basic attribute set by taking employee ID as a core. For example, employee A's set of attributes may be { post: data analyst, certificate: tensorFlow authentication, evaluation score: 4.5/5}, and employee B may be { post: software engineer, certificate: java SE 8, evaluation score: 4.0/5}.

Dynamic capability indicators can be extracted by parsing the project experience of the employee. By utilizing natural language processing and code analysis technology, historical project documents and code warehouses of staff are analyzed, and the use frequency (such as Python accounting for 70%) of a technical stack and task role labels (such as 'front-end development') are obtained.

For example, employee A has participated in three projects over the past year, and the code repository shows that he has used Python and Pandas up to 80% frequently, and the document marks his role as "data engineer" and deduces therefrom that he has "data processing" capabilities.

Relationships between employees, skills, and items are stored in a graph database, such as "employee a-participation-item B-use-skill C". The skill level (weight 40%), the project application times (weight 40%) and the peer review score (weight 20%) were combined and the skill mastery was calculated by weighted average.

For example, employee A's "machine learning" mastery may be calculated as:

(certificate middle level 0.5×40%) + (item 5 times 0.8×40%) + (review 4.5/5×20%) =0.67. The "Java Programming" mastery of employee B was based on certificate high (0.9), project 10 times (1.0) and review 4.0/5, with a calculation of 0.92.

Additionally, performance data can be mapped to skill nodes using graph embedding algorithms (e.g., deepWalk), generating a low-dimensional vector representation, and inferring implicit associations based on employee skill similarities. For example, if both employees A and B are good at "Python programming," it can be speculated that B may also have "data analysis" skills of A. Employee C is highly performance and good at "front-end development," and may speculate through collaborative filtering that he is "UI design" capable.

Dynamic timing diagram modeling can also calculate the decay factor of skills, such as "SQL" skills if not useful for a year, the mastery decays by 10%. Analysis of the impact of cross-department collaboration on skills by graph attention mechanisms, such as 20% improvement in "communication ability" after employee A has collaborated with market segments.

After integrating these data, if the HR data and the item data conflict with each other (e.g., HR flag "primary", item display "advanced"), the HR data and the item data are processed according to a priority rule of a preset weight (HR 40%, item 60%), and finally a current talent knowledge map with a timestamp, such as "2023-10-01_v1.0", is generated.

The next step is to compare the target and the existing talents knowledge graph by using graph matching algorithm, and calculate their matching degree, so as to quantify the difference between the existing talents and the target talents.

As shown in fig. 4, one implementation is to encode two atlases into vectors using the graph neural network GNN, preserving structural information (such as competency paths) and semantic information (such as skill weights).

GNNs may aggregate information of each node and its neighbors through a multi-layer convolution operation. For example, in a target talent knowledge graph, a "data mining" node may be connected to nodes such as "statistics", "programmability", etc., and these connections may be encoded as part of a vector, reflecting the location and importance of the node in the graph.

The attributes of the node, such as skill level (e.g., a "data mining" level of proficiency of 0.8), the department or post requirements, may also be incorporated into the encoding process. For example, in an existing employee graph, the "data mining" is 0.5, and the target graph is 0.8, and this difference is represented in the vector values.

In a specific example, it is assumed that the coding vector of the "data mining" node in the target map is [0.8,0.4,0.6] and the coding vector in the existing map is [0.5,0.3,0.4], and the two vectors are generated by GNN, and include dual information of structure and semantics, so as to lay a foundation for subsequent similarity calculation.

After obtaining the vector representations, global similarity of the two atlases is calculated as initial matching based on these vectors. The global similarity reflects the degree of fit of the overall map, and generally cosine similarity is used as a measurement standard, so that the consistency of vector directions can be effectively measured in a high-dimensional space. The calculation mode can adopt a cosine similarity formula:

;

Wherein the method comprises the steps of AndThe encoding vectors of the target and existing atlases, respectively.

Global similarity, while providing an overall view, may mask differences in local structure. Therefore, the invention introduces a graph matching algorithm, identifies the key sub-graph structure in the target and the existing graph, and calculates the local matching degree.

Important local patterns in the map can be extracted through subgraph isomorphism or pattern mining technology. For example, a skill path of "programmability→data mining→machine learning" may exist in a target atlas, while "programmability→data analysis→visualization" may exist in an existing atlas. These paths reflect the dependency of skill.

After identifying the subgraph, the corresponding nodes are aligned by similarity calculation based on the attributes. For example, "data mining" and "data analysis" may be aligned due to functional proximity, but proficiency differences may result in reduced matching. Relationship matching focuses on edge consistency, such as whether the dependence of "programmability" to subsequent skills in both maps is the same. The local match may be calculated by a similarity weighted sum of the alignment entities and the relationships.

For example, assume that the subgraph of the target map is "programmability→data mining (0.8)", and the existing map is "programmability→data analysis (0.6)".

The "programmability" nodes are fully aligned, with a similarity of 1;

The functions of the data mining and the data analysis are similar but the proficiency is different, the similarity is assumed to be 0.7, the relationship of the edges is consistent, and the similarity is 1. The local match may be:

;

This value indicates that the local structures are relatively similar, but with some differences.

To comprehensively reflect the matching degree of the atlas, the global similarity (initial matching degree) and the local matching degree can be fused to generate final matching degree. The fusion method can be either a simple weighted average or can be optimized by means of a machine learning model.

The weighted average method is as follows:

;

Where α is an adjustable weight, e.g., set to 0.6 to more emphasize global similarity.

The machine learning optimization can train a regression model by utilizing the history matching data, input global matching degree and local matching degree, and output final matching degree. This way, a nonlinear relationship between the two can be captured. For example, the trained model may consider local differences to have a greater impact on certain posts, thereby adjusting the weights.

Besides static matching degree calculation, the invention also provides a dynamic matching degree optimization method based on reinforcement learning.

As shown in fig. 5, the method adapts to complexity and dynamic change of the map structure through continuous learning and strategy adjustment, and improves accuracy of matching degree.

In the initial stage, a Graph Neural Network (GNN) can be utilized to calculate the preliminary matching degree of the target talent knowledge graph and the existing employee knowledge graph, and the preliminary matching degree is used as a reference result. The specific process is similar to the above method, vector representation is generated by GNN encoding, and cosine similarity is calculated. For example, the preliminary matching degree may be 0.82, which indicates the similarity of the two maps in the initial state.

Subsequently, a reinforcement learning model is introduced to optimize the matching result by performing a specific action. These actions include adjusting the weight parameters of the matching algorithm or updating employee skill calibrations, with the matching degree recalculated after each action.

For example, the weight of "machine learning" skills in GNN coding is increased, adjusting from 1.0 to 1.3, as it is more important in the target atlas. The "machine learning" proficiency of employee B was updated from 0.5 to 0.7 based on the new assessment data.

In the updating process, the GNN and similarity calculation is restarted after adjustment, and the matching degree is possibly increased from 0.82 to 0.85. This dynamic adjustment can better reflect the actual situation. To guide the reinforcement learning model, a reward function may be designed to give feedback based on the change in the degree of matching. Rewards are generally defined as the difference between a new match and an old match. For example, from 0.82 to 0.85, the prize is 0.03, and if it drops to 0.80, the prize is-0.02. The rewarding function may also incorporate business objectives such as adding rewards if the action improves the matching of key skills.

Next, based on the reward feedback, the strategy may be updated using a reinforcement learning algorithm with the goal of maximizing the 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 gets a positive prize, whose Q value rises. In DQN, a neural network is used to fit the Q function, which is suitable for complex scenes. In the Actor-Critic, policy selection and value assessment are combined, as described in more detail in one embodiment below. After one update, the model may be more prone to adjusting the weights of the key skills.

The matching degree is gradually optimized through multiple rounds of action execution, rewarding feedback and strategy adjustment. Training may continue until the degree of matching stabilizes or reaches a preset threshold. For example, after 15 rounds of iteration, the matching degree is improved from 0.82 to 0.89, the model learns a better matching strategy, and the adaptability is obviously improved.

In the above embodiment, the Actor-Critic method is an efficient policy updating mode, combines the advantages of policy gradient and value function, and is suitable for dynamically adjusting the matching process. The Actor network can select an optimization action according to the current map matching state. The state may be the current degree of matching and profile characteristics and the action includes adjusting weights or skill calibrations. The Actor network may output a probability distribution of actions, e.g., a probability of "tune 'data mining' weight to 1.2" of 0.7. For example, when the current matching degree is 0.85, the Actor suggests to raise the "data mining" weight.

Critic networks may be used to evaluate policies. The Critic network evaluates the action selected by the Actor, calculates a state-action value function (Q value), reflecting the expected benefit of the action. An example evaluation procedure is that for "data mining" weights to 1.2", critic might output a Q value of 0.04, indicating a match of 0.04 is expected to be raised. For example, if the degree of matching increases to 0.87 after the operation, critic verifies the effect.

Based on Critic's evaluation, the Actor adjusts parameters by gradient ramp-up, increasing the probability of high Q-value motion. The optimization logic is that the Q value is positive, the selection trend of the action is enhanced, and if the Q value is negative, the probability is reduced. For example, the actor increases the tendency to adjust the "data mining" weight to optimize the matching policy because the Q value is 0.04. This collaboration mechanism makes the Actor-Critic approach excellent in complex matching scenarios.

After the matching is calculated, training resources can be recommended to staff further according to the matching result so as to reduce the gap between the existing talent knowledge graph and the target graph. This process includes skill gap identification, resource association construction, and recommendation generation.

Specifically, the matching result of the graph matching algorithm can reveal a 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 the skills. By comparing attribute values, such as proficiency or importance, of skill nodes in the two maps, a skill gap of an employee can be precisely located. For example, assuming that the target atlas requires a proficiency of "cloud computing" skills of 0.8, and the existing proficiency of a staff is only 0.5, this 0.3 gap becomes the basis for recommending training resources.

Similarly, in the reinforcement learning strategy in the second method, by observing the adjustment strategy of the existing talent knowledge graph, the degree of matching can be improved by knowing what adjustment strategy is. For example, if the proficiency of the "cloud computing" skill of a staff in the existing talent knowledge graph is improved, the overall matching degree is improved, and then the "cloud computing" skill can be recommended in the subsequent recommendation training resources.

In order to effectively connect the skill gap with the training resources, a training resource knowledge graph needs to be constructed. The purpose of this atlas is to associate various training resources, such as online courses, practice workshops or professional certification exams, with specific skill nodes. The establishment of the association relationship may depend on the content profile of the resource, the learning objective, or the learning effect reflected in the history data. For example, a "cloud computing base course" may be labeled as being able to promote a "cloud computing" skill of 0.2. In this way, the training resource knowledge graph provides a structured resource library for the recommendation process, so that the system can quickly find matched training contents according to skill requirements.

Based on the skill gap and the resource map, the recommendation algorithm can generate a personalized training resource recommendation scheme. The algorithm may integrate a number of factors, including the degree of skill gaps, the strength of association of training resources with skills, and the learning habits of the staff individuals. For example, an employee may have a short board in the "data visualization" skill, and the algorithm may pick a course highly relevant to the employee from the resource map, and eventually recommend a "data visualization video advanced course" taking into account the habit that the employee is more inclined to video teaching rather than text reading. Such recommendation can not only effectively fill the skill blank, but also promote the learning experience and effect of staff.

Fusion of various techniques may be employed in the design of the recommendation algorithm.

As shown in fig. 6, the whole algorithm execution process is guided by the enterprise strategic target, and meanwhile, the personalized requirements of staff and the knowledge sharing requirements of cross departments are considered. Specifically, the algorithm may evaluate the importance of the target skills in the enterprise strategic profile by a modified PageRank method. The method imagines skill nodes as 'pages' in a network, and the dependency relationship among the skills is similar to 'links', and then the weight factors of enterprise strategy are overlapped. For example, in an enterprise focusing "smart manufacturing," an "industrial internet of things" skill may get a higher impact score due to its strategic value. Based on such analysis, the algorithm can determine which skill gaps have a greater impact on the enterprise as a whole, thereby preferentially recommending relevant training resources.

Meanwhile, the algorithm utilizes the graph attention network to conduct deep analysis on the learning paths of the staff, and hidden learning mode features are mined. Such networks can capture employee preferences, such as a preference to practice with hands over theoretical learning, by focusing on key nodes in the learning history. The personalized insight enables the recommended result to be more in line with the actual demands of staff, and improves the participation degree and the completion rate of training. In addition, in order to more accurately connect the skill gap with the training resource, the algorithm can also adopt a meta-path reasoning method. The method can excavate multi-level associated paths in the heterogeneous map, such as 'staff-lack-skill A-pass-course B-promote', and mark confidence score for each path according to the effect of the historical data, and finally screen the most reliable resource options.

In order to make the recommended strategy more dynamic, a multi-arm slot machine mechanism can also be introduced. This mechanism enables trade-offs between collaborative filtering recommendations and knowledge driven recommendations and is continuously adjusted based on the employee's real-time feedback. For example, collaborative filtering may recommend resources based on similarities between employees, while knowledge driven is directly dependent on the association of resources with skills. The initial two ways may each weigh half, but if the feedback shows that the employee is more inclined to knowledge driven recommendations, the algorithm will step up the weight of this part. This adaptive adjustment allows the recommendation system to be optimized continuously with use.

For the difference of skill demands of cross departments, a migration learning model can also be designed. The model can be used for adjusting successful training experience of departments with higher skill matching degree and applying the adjusted training experience to departments with lower matching degree. For example, the technical departments have obvious scheme effects on 'artificial intelligence' training, and the technical departments can be quickly adapted to be a version suitable for sales departments through transfer learning, so that sales personnel can be helped to master relevant knowledge. The method not only improves training efficiency, but also promotes experience sharing in enterprises.

Finally, the algorithm integrates the recommendation into a training plan with space-time constraints. The atlas displays specific training tasks in the form of nodes, each node marks a time window of the task, such as 'completing basic courses 10 months 10 days to 20 days', and the edges connecting the nodes embody the transfer relation of logic sequence and skill improvement between the tasks, such as entering the advanced stage after completing the basic courses. In this way, the employee can obtain a clear learning path, both taking into account the rationality of the scheduling and ensuring the consistency of skill improvement.

As shown in fig. 7, in a second aspect of the present invention, a training resource recommendation system based on a knowledge graph is disclosed, which may be used to implement the above method, including:

More specifically, the implementation of the system can rely on a modular design to divide the functionality into multiple independent parts, with the entire system deployed on a cloud computing platform, such as AWS or ali cloud. In hardware configuration, a high performance server may be used, equipped with a multi-core CPU, a high capacity RAM, and SSD storage to ensure efficient operation of the graph database and machine learning model.

In terms of programming implementation, a variety of languages and techniques may be employed. Python can enable the development of algorithms and models with its rich library support (e.g., tensorFlow, pyTorch and NLTK). Java is used for constructing high concurrency back-end service, and guaranteeing the stability of the system under multi-user access. The development of the user interface can depend on JavaScript, and is realized through a compact framework in particular, so as to provide smooth interaction experience. The data storage layer can select Neo4j as a graph database to manage knowledge graphs, which is good at processing complex node and relationship data, and meanwhile, postgreSQL is used for storing metadata and user interaction records to provide reliable management of structured data. In the development process, the agile development method can be followed, version control is carried out through Git, and continuous integration and deployment are realized by means of Jenkins.

Thus, in a third aspect of the present invention, an electronic device is disclosed comprising:

a memory and a processor, the memory for storing a computer program;

The processor is configured to execute the above method when the computer program is called.

The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should be covered by the protection scope of the present invention by making equivalents and modifications to the technical solution and the inventive concept thereof.

Claims

1. The training resource recommendation method based on the knowledge graph is characterized by comprising the following steps of:

Constructing a target talent knowledge graph of an enterprise;

acquiring the existing talent knowledge graph of an enterprise;

2. The training resource recommendation method based on the knowledge graph of claim 1,

The method for constructing the target talent knowledge graph of the enterprise comprises the following steps:

3. The training resource recommendation method based on the knowledge graph according to claim 1, wherein the method for constructing the existing talent knowledge graph of the enterprise comprises the following steps:

4. The training resource recommendation method based on a knowledge graph according to claim 1, wherein the method for calculating the matching degree specifically comprises:

5. The training resource recommendation method based on a knowledge graph according to claim 1, wherein the method for calculating the matching degree specifically comprises:

6. The knowledge-graph-based training resource recommendation method of claim 5, wherein the Actor-Critic method comprises:

Calculating a strategy by using an Actor network;

7. The training resource recommendation method based on the knowledge graph according to claim 1, wherein the training resource recommendation is performed on enterprise staff 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 comprising:

8. The training resource recommendation method based on the knowledge graph of claim 7,

The recommendation algorithm comprises the following steps:

9. A training resource recommendation system based on a knowledge graph, comprising:

10. An electronic device, comprising:

a memory and a processor, the memory for storing a computer program;

The processor is configured to perform the method of any of claims 1-8 when the computer program is invoked.