Talent training recommendation method based on federated learning and natural language processing

Abstract

This invention provides a talent cultivation recommendation method based on federated learning and natural language processing, comprising the following steps: processing multi-source heterogeneous data to obtain standard data; constructing a talent knowledge graph capable of incremental updates; anonymizing and desensitizing sensitive data, and, without sharing the original data, performing joint modeling using a federated learning framework to obtain a local model; generating talent profiles and job profiles, and combining multiple target factors and dynamic weights with a preset matching algorithm to obtain a talent matching model; continuously monitoring the similarity gap between the talent profiles and job profiles, and dynamically generating course optimization schemes and job optimization schemes; and providing visual explanations or keyword descriptions of the recommendation process. This invention not only improves matching accuracy and efficiency, breaks down data silos, and protects privacy and security, but also dynamically adapts to changes in market demand, meets the personalized needs of various parties, and promotes precise matching between education and employment.

Description

Technical Field

[0001] This invention relates to the field of job matching technology, and in particular to a talent training recommendation method based on federated learning and natural language processing. Background Technology

[0002] In the modern education and job market, accurately matching curriculum content with job requirements is a critical issue that urgently needs to be addressed in vocational and higher education systems. With the rapid iteration of industrial technologies and the diversification of market demands, traditional education models struggle to quickly respond to companies' skill requirements, leading to a waste of educational resources and a decline in graduates' employment competitiveness.

[0003] Currently, existing job matching methods include, but are not limited to, traditional methods based on keyword matching, matching methods based on scoring and weighting, matching methods based on recommendation systems, and intelligent matching methods based on artificial intelligence. However, existing job matching methods suffer from the following drawbacks: 1) Insufficient matching accuracy: Traditional methods, such as keyword matching, lack a deep understanding of job requirements and candidate capabilities, easily missing potential matches or producing false matches. 2) Insufficient personalization: Most existing systems use general matching rules, ignoring the personalized needs of companies and candidates, such as corporate culture and candidates' career development aspirations. 3) Data silos and privacy issues: Data sharing between educational institutions, companies, and talents is not effective, resulting in incomplete matching information, and the security of sensitive personal and corporate information is difficult to guarantee when using data. 4) Insufficient dynamism: Most current matching systems are statically designed, meaning that fixed recommendation rules are generated after model training, making it impossible to dynamically adjust curriculum design or job recommendations according to rapidly changing market demands. 5) Lack of interpretability: Although intelligent matching methods (such as deep learning) have good results, the matching results are difficult to interpret, making them unconvincing to companies and candidates. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a talent training recommendation method based on federated learning and natural language processing. This method can not only improve matching accuracy and efficiency, break down data silos, and protect privacy and security, but also dynamically adapt to changes in market demand, meet the personalized needs of various parties, and promote the precise matching of education and employment.

[0005] To achieve the above objectives, this invention provides the following solution: a talent training recommendation method based on federated learning and natural language processing, comprising the following steps:

[0006] Collect relevant data from educational institutions, enterprises, and third-party platforms to obtain multi-source heterogeneous data. Store the multi-source heterogeneous data in a data lake, then clean and integrate the multi-source heterogeneous data to obtain standard data, and store the standard data in a data warehouse.

[0007] Construct a talent knowledge graph related to talent development and job skills, and perform real-time incremental updates on the talent knowledge graph;

[0008] Sensitive data in the standard data is anonymized and desensitized, and local models are obtained by joint modeling using a federated learning framework, without sharing the original data between educational institutions and enterprises.

[0009] Based on the standard data, students and positions are characterized in depth to generate dynamically updated talent profiles and position profiles. The matching algorithm is then combined with multiple target factors and dynamic weights. The talent profiles, position profiles and matching algorithms are then deployed into the local model to obtain the talent matching model.

[0010] Continuously monitor the similarity gap between the talent profile and the job profile, and dynamically generate course optimization plans and job optimization plans based on the similarity gap and the output of the talent matching model;

[0011] The output of the talent matching model, the course optimization scheme, and the job optimization scheme are explained visually or described using keywords using explainable AI technology.

[0012] Optionally, relevant data is collected from educational institutions, enterprises, and third-party platforms to obtain multi-source heterogeneous data. This multi-source heterogeneous data is then stored in a data lake. Next, the multi-source heterogeneous data undergoes data cleaning and integration to obtain standard data, which is then stored in a data warehouse. This includes:

[0013] Based on a unified data format and field definition, relevant data from educational institutions, enterprises, and third-party platforms are obtained using API interfaces or web crawling technology to obtain multi-source heterogeneous data.

[0014] A data lake is constructed using a distributed storage system, and the multi-source heterogeneous data is partitioned and stored in the data lake according to the data source and type;

[0015] The multi-source heterogeneous data is processed for missing values, outliers, duplicate data, and data integration to obtain standard data;

[0016] A data warehouse is built using a relational database. Based on the data theme, the standard data is partitioned and stored in the data warehouse, and a query index is created in the data warehouse.

[0017] Optionally, a talent knowledge graph related to talent development and job skills is constructed, and the talent knowledge graph is updated incrementally in real time, including:

[0018] Collect job-related data, industry-related data, and course-related data to obtain initial text; preprocess the initial text to obtain knowledge text.

[0019] The BERT model is used to perform contextual semantic understanding of the knowledge text, and NLP technology is used to extract entities from the knowledge text. Then, the entities are mapped to a standardized lexicon by combining an external knowledge base to complete the entity standardization operation.

[0020] Dependency parsing is used to extract relationships between entities, and these relationships are then classified.

[0021] Select a graph database, import entities and relations into the selected graph database to obtain a graph structure, use graph embedding technology to map the nodes and relations in the graph structure to a vector space, generate node embedding vectors, complete semantic matching operations, and obtain a talent knowledge graph;

[0022] The system automatically parses new data from the initial text and updates the talent knowledge graph based on the parsing results.

[0023] Optionally, sensitive data in the standard data is anonymized and de-identified, and a local model is obtained by using a federated learning framework for joint modeling, without sharing the original data between educational institutions and enterprises, including:

[0024] The standard data is categorized into personal privacy information, sensitive corporate data, and general data.

[0025] The personal privacy information is encrypted using an irreversible hash algorithm, and a random salt value is added before the hash to prevent rainbow table attacks, thus completing the anonymization process of the personal privacy information.

[0026] For the enterprise's sensitive data, regular expressions are used to identify sensitive fields, and hashing, masking, or obfuscating operations are performed on the sensitive fields to complete the desensitization process of the enterprise's sensitive data.

[0027] Educational institutions and enterprises are set as participants. Each participant uses its own data to train the model locally, obtains the local model and local parameters, and completes the deployment of the local node.

[0028] The local parameters are encrypted using homomorphic encryption to obtain encrypted parameters. The encrypted parameters are received by the central server, and the encrypted parameters are aggregated to obtain global parameters, thus completing the deployment of the cloud node.

[0029] The central server is used to send the global parameters to each of the participating parties to perform a global update of the local model.

[0030] Optionally, based on the aforementioned standard data, students are subjected to in-depth semantic characterization to generate dynamically updated talent profiles, including:

[0031] Collect student multi-source data and store the student multi-source data in the data warehouse; the student multi-source data includes student course grades, skill mastery, internship experience, career preferences and other student data;

[0032] Feature extraction is performed on the student multi-source data to obtain student numerical features, student classification features and student text features. Then, the student numerical features are standardized and normalized, the student classification features are encoded, and the student text features are vectorized to complete the feature transformation.

[0033] Using a fully connected layer, the student numerical features and the student classification features are converted into fixed-dimensional structured data vectors. A pre-trained language model is used to obtain the text data vectors of the student text features. Then, the structured data vectors and the text data vectors are fused to obtain a multi-dimensional student vector representation.

[0034] The student's skills are matched with the skill nodes in the talent knowledge graph to obtain the student's skill semantic vector. A high-dimensional vector of the student's skill semantic vector is generated using a graph neural network. The high-dimensional vector of the student's skill semantic vector is then fused with the multi-dimensional student vector representation to complete the deep semantic characterization of the student and generate a talent profile.

[0035] Optionally, based on the standard data, a deep semantic characterization of the job position is performed to generate a dynamically updated job profile, including:

[0036] Collect multi-source job data and store the multi-source job data in the data warehouse; the multi-source job data includes job requirements, skill requirements, industry attributes, corporate culture, and other job data;

[0037] Feature extraction is performed on the multi-source data of the job positions to obtain job numerical features, job classification features, and job text features. Then, the job numerical features are standardized and normalized, the job classification features are encoded, and the job text features are vectorized to complete the feature transformation.

[0038] Using a fully connected layer, the numerical features and classification features of the job positions are converted into fixed-dimensional structured data vectors. A pre-trained language model is used to obtain text data vectors of the text features of the job positions. The structured data vectors and text data vectors are then fused to obtain a multi-dimensional job vector representation.

[0039] The skills required for the job are matched with the skill nodes in the talent knowledge graph to obtain the job skill semantic vector. A high-dimensional vector of the job skill semantic vector is generated using a graph neural network. The high-dimensional vector of the job skill semantic vector is then fused with the multi-dimensional job vector representation to complete the deep semantic characterization of the job and generate a job profile.

[0040] Optionally, by combining multiple target factors and a dynamic weight-based preset matching algorithm, the talent profile, the job profile, and the matching algorithm are then deployed into the local model to obtain a talent matching model, including:

[0041] The talent profile and the job profile are mapped to a unified vector space to generate talent profile vector and job profile vector.

[0042] Using cosine similarity, the cosine value of the angle between the talent profile vector and the job profile vector is calculated, and using Euclidean distance, the straight-line distance between the talent profile vector and the job profile vector is calculated to obtain the semantic similarity between the talent profile vector and the job profile vector.

[0043] Select multiple objective factors, assign dynamic weights to different objective factors, and then combine optimization algorithms and reinforcement learning to build a multi-objective optimization model;

[0044] By combining the semantic similarity and the multi-objective optimization model, a matching algorithm for calculating the comprehensive matching score is obtained. The matching algorithm is then deployed to the local model to obtain the talent matching model.

[0045] Optionally, the expression for the matching algorithm is:

[0046]

[0047] Where T is the overall matching score, w i For dynamic weights, f i denoted as the score for each objective factor, where m is the total number of objective factors.

[0048] Optionally, the similarity gap between the talent profile and the job profile is continuously monitored, and based on the similarity gap and the output of the talent matching model, course optimization plans and job optimization plans are dynamically generated, including:

[0049] The talent matching model is used to obtain the target position and the skill requirements of the target position. The skill requirements of the target position are then compared, analyzed and updated with the relevant data of the corresponding position in the talent knowledge graph.

[0050] The assessment quantifies students' current skills, compares these skills with the updated skill requirements of the target positions to identify skill gaps, and then uses machine learning algorithms to analyze these skill gaps to recommend relevant courses, obtain course optimization plans, and predict and plan courses based on industry trends and student needs.

[0051] Real-time monitoring of changes in job requirements and curriculum; optimization of target positions based on monitoring results and various target factors to obtain job optimization solutions.

[0052] Optionally, explainable AI technology can be used to provide visual explanations or keyword descriptions of the output of the talent matching model, the course optimization scheme, and the job optimization scheme, including:

[0053] The feature contribution map is used to show the contribution of each feature to the matching and recommendation results; the decision path map is used to show the decision path of the talent matching model; and the keyword cloud is used to show the driving factors of job recommendation.

[0054] Provide dashboards for viewing data usage and the recommendation process, and design privacy policies to achieve transparent management.

[0055] This invention discloses the following technical effects by providing a talent training recommendation method based on federated learning and natural language processing:

[0056] 1. High Matching Accuracy: 1) This invention, through a talent knowledge graph, can construct four core entities—"skills, courses, positions, and industries" and their relationships, supporting more accurate semantic matching and recommendations. Simultaneously, the semantic structure and graph embedding technology of the knowledge graph significantly improve the efficiency and accuracy of matching positions with talents, giving this method stronger semantic understanding and reasoning capabilities. 2) By combining multi-source data and knowledge graphs to construct multi-dimensional talent and position profiles, it can achieve semantic expression of high-dimensional vectors, enhancing the depth and accuracy of the profiles. This makes the similarity calculation between talent and position profiles more accurate. Combined with a multi-objective optimization process, this invention achieves accurate position matching and course optimization, improving recruitment and training efficiency.

[0057] 2. High Security: This invention ensures data security and enhances data protection by anonymizing and desensitizing personal privacy information (such as name, contact information, and ID number) and sensitive corporate data (such as financial information and core technology keywords). By utilizing a federated learning framework, models can be jointly trained between educational institutions and enterprises without sharing data, solving the data silo problem while protecting data privacy.

[0058] 3. Personalized Recommendation and Dynamic Adjustment Scheme: This invention can compare students' current skills with the changing skill requirements of target positions, generate dynamic course optimization schemes and job optimization schemes, provide each student with the most suitable job and course, realize personalized dynamic recommendations, promote the coordinated development of education and industry, and enhance the practicality and competitiveness of talent training.

[0059] 4. High level of trust: This invention makes the data transparent and traceable by visually displaying and explaining the recommendation process and data usage process, which enhances users' understanding and trust and promotes collaborative cooperation between educational institutions and enterprises.

[0060] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention;

[0063] Figure 2 A schematic diagram illustrating the talent and job profile generation process provided in this embodiment of the invention;

[0064] Figure 3 This is a schematic diagram of the job recommendation process provided in an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0067] like Figure 1 As shown, this invention provides a talent training recommendation method based on federated learning and natural language processing, including the following steps:

[0068] 1. Collect relevant data from educational institutions, enterprises, and third-party platforms to obtain multi-source heterogeneous data. Store this multi-source heterogeneous data in a data lake, then perform data cleaning and integration on the multi-source heterogeneous data to obtain standard data, and finally store this standard data in a data warehouse. This includes:

[0069] 1.1 Based on a unified data format (such as JSON format) and field definitions, use API interfaces or web crawling technology to obtain relevant data from educational institutions, enterprises and third-party platforms, resulting in multi-source heterogeneous data.

[0070] Educational institutions: Student information (name, student ID, course grades, internship experience), course information (course name, course content, skills points).

[0071] For companies: Job information (job title, job description, skills required, salary range), recruitment data (interview records, hiring status).

[0072] Third-party platforms: Industry trend data (popular jobs, salary levels), recruitment website data (job requirements, company reviews).

[0073] API Interface: Obtain data in real time through RESTful API or GraphQL interface.

[0074] Batch import: Supports batch data upload in formats such as CSV, Excel, and JSON.

[0075] Web scraping technology: For third-party platforms whose APIs cannot be directly accessed, web scraping technology is used to collect publicly available data.

[0076] The field definitions and mappings are shown in Table 1 below:

[0077] Table 1 Field Mapping Table

[0078] Data source Original field name Standard field name Educational institutions Student ID student_id Educational institutions Name name enterprise Job Title job_title Third-party platform Skill Requirements skill_requirements

[0079] 1.2 Construct a data lake using a distributed storage system (such as Hadoop HDFS or Amazon S3), and partition and store the multi-source heterogeneous data into the data lake according to the data source and type.

[0080] 1.3 The multi-source heterogeneous data is processed for missing values, outliers, duplicate data, and data integration to obtain standard data.

[0081] 1.31 Missing value handling includes:

[0082] Deletion method: For non-critical fields (such as remarks), directly delete records with missing values.

[0083] Fill method: For key fields (such as student grades, job salaries), use the following fill method:

[0084] Mean filling: Filling with the mean of similar data (such as average score).

[0085] Interpolation method: interpolation is performed based on time series data.

[0086] Machine learning prediction: Using regression models to predict missing values.

[0087] 1.32 Outlier handling includes:

[0088] Rule detection: Set reasonable ranges (e.g., score range 0-100, salary range 3000-50000), and mark data outside the range as abnormal.

[0089] Statistical detection: Use box plot method (IQR) or Z-score method to detect outliers.

[0090] Handling method: Outliers can be either deleted or replaced with the median.

[0091] 1.33 Duplicate data processing includes:

[0092] Use hash algorithms or unique identifiers (such as student IDs or job IDs) to detect duplicate data. Merge duplicate records, keeping only the latest or most complete data.

[0093] 1.34 Data integration includes:

[0094] Uniform ID mapping for entities: Generate a unique identifier (UUID) for each entity (such as a student or a job position).

[0095] Data Linking and Merging: Use primary keys and foreign keys to link data from different data sources. Example: Merge student course grades and internship experience.

[0096] 1.4 Build a data warehouse using relational databases (such as PostgreSQL and Amazon Redshift). Based on the data theme (to support fast retrieval and analysis), store the standard data partitions in the data warehouse and create query indexes (such as student_id and job_id) in the data warehouse to improve query efficiency.

[0097] 2. Construct a talent knowledge graph related to talent development and job skills, and perform real-time incremental updates to the talent knowledge graph. This includes:

[0098] 2.1 Collect job-related data (job requirement documents provided by enterprises or job information from recruitment websites), industry-related data (industry trend analysis reports provided by third-party platforms), and course-related data (course outlines, teaching objectives, skill points, etc. provided by educational institutions) to obtain initial text. Preprocess the initial text to obtain knowledge text.

[0099] Text preprocessing: This involves processing the text such as word segmentation, stop word removal, and part-of-speech tagging. Example:

[0100] Input text: `"Requires proficiency in Python, machine learning, and data visualization."`

[0101] Preprocessing result: `["Requires","proficiency","in","Python",","machinelearning",","datavisualization"]`.

[0102] 2.2 Combine the BERT model to perform contextual semantic understanding on the knowledge text, and use NLP technology to extract entities (such as skills, positions, and industries) from the knowledge text. Then, combine external knowledge bases (such as LinkedIn Skills, ONET) to map the entities to a standardized vocabulary (such as mapping "Python" to "PythonProgramming" in the skills library) to complete the entity standardization operation.

[0103] 2.3 Use dependency parsing to extract relationships between entities and categorize these relationships (e.g., "need," "contain," "depend on"). Example:

[0104] Input text: `"Python is required for this position."`

[0105] Dependency relationship: `[("Python","required","Skill-Job")]`.

[0106] 2.4 Select a graph database (such as Neo4j), import entities and relations into the selected graph database to obtain a graph structure, and use graph embedding technology (such as Node2Vec, GraphSAGE) to map the nodes and relations in the graph structure to a vector space to generate node embedding vectors, complete semantic matching operations, and obtain a talent knowledge graph.

[0107] Spectral structure:

[0108] Node types: Skill, Job, Course, Industry.

[0109] Edge types: require, contain, depend on, belong to.

[0110] 2.5 The system automatically parses new data from the initial text. Specifically, when a new job posting or course is launched, the system automatically parses the text and extracts entities and relationships. Based on the parsing results, the talent knowledge graph is updated.

[0111] 3. Sensitive data in the standard data is anonymized and de-identified. Without sharing the original data between educational institutions and enterprises, joint modeling is performed using a federated learning framework (open-source federated learning frameworks such as TensorFlow Federated and PySyft) to obtain a local model. This includes:

[0112] 3.1 The standard data is categorized into personal privacy information (such as name, contact information, ID number, etc.), sensitive corporate data (such as financial information, core technology keywords, etc.), and general data (such as job descriptions, course content, industry trends, etc.). Different protection strategies are adopted for different types of data to ensure privacy protection without affecting data usability.

[0113] 3.2 The aforementioned personal privacy information is encrypted using an irreversible hash algorithm (such as SHA-256), ensuring that even if the data is leaked, the original information cannot be recovered. A random salt value is added before the hash to prevent rainbow table attacks, thus completing the anonymization process of the personal privacy information. Example:

[0114] Enter: `Name: Zhang San`

[0115] Output: `Hash value: 3a7bd3e2360a3d5a6d8b6f4f1a1e5f6c`

[0116] Example of salting treatment:

[0117] Input: `Name: Zhang San + Salt Value: random123`

[0118] Output: `Hash value: 9f8b3e2a6d7c4f1b2a3e5f6d8b6f4c1a`

[0119] 3.3 For the enterprise's sensitive data, use regular expressions to identify sensitive fields, and perform hashing, masking, or obfuscation operations on the sensitive fields to complete the desensitization process of the enterprise's sensitive data.

[0120] Example of masking:

[0121] Input: `Financial data: 2024 revenue was 50 million yuan`

[0122] Output: `Financial data: 2024 revenue was 10,000 yuan`

[0123] Example of blurring:

[0124] Input: Core technology: Deep learning model optimization

[0125] Output: Core technology: Deep learning.

[0126] 3.4 Educational institutions and enterprises are set as participants. Each participant uses its own data to train the model locally, obtains the local model and local parameters (such as the weight matrix), and completes the deployment of the local node.

[0127] 3.5 Encrypt the local parameters using homomorphic encryption to obtain encrypted parameters. Example:

[0128] Input: `Model parameters: [0.1, 0.2, 0.3]`

[0129] Output: `encryption parameters:[Enc(0.1),Enc(0.2),Enc(0.3)]`.

[0130] The encrypted parameters are received by a central server, and then aggregated to obtain global parameters, thus completing the deployment of the cloud node; Example:

[0131] Input: `Enc(0.1)+Enc(0.2)+Enc(0.3)`

[0132] Output: `Enc(0.6)`.

[0133] 3.6 The global parameters are sent to each participant using a central server (deployed in the cloud) to perform a global update of the local model. Example:

[0134] Input: `Global parameters:[Enc(0.6)]`

[0135] Output: `Update local model after decryption`.

[0136] 4. For example Figure 2As shown, based on the standard data, students and positions are characterized in a deep semantic way, generating dynamically updated talent profiles and position profiles. A matching algorithm is then pre-set with multiple target factors and dynamic weights, and the talent profiles, position profiles, and matching algorithms are deployed into the local model to obtain a talent matching model.

[0137] 4.1 Generation of Talent Profiles:

[0138] 4.11 Collect student multi-source data and store the student multi-source data in the data warehouse; the student multi-source data includes:

[0139] Student course grades, including student scores and rankings in each course;

[0140] Skill mastery is assessed based on data such as course grades, assessment results, and skills certifications.

[0141] Internship experience, including the company, position, duration, and performance.

[0142] Career preferences, including students' career interests, target positions, and desired industries;

[0143] Other student data includes interests, soft skills assessments, and project experience.

[0144] 4.12 Feature extraction is performed on the student multi-source data to obtain:

[0145] Student numerical characteristics, such as course grades, internship duration, and skills mastery scores;

[0146] Student classification characteristics, including the type of courses taken, skill categories, types of internship companies, and career preference categories;

[0147] Student text characteristics, including course descriptions, internship descriptions, and career goal descriptions.

[0148] The student numerical features are then standardized and normalized, the student classification features are encoded (one-hot encoding or embedding encoding), and the student text features are vectorized to complete the feature transformation.

[0149] 4.13 Using a fully connected layer, the student numerical features and the student classification features are converted into fixed-dimensional structured data vectors. A pre-trained language model is used to obtain the text data vectors of the student text features. The structured data vectors and the text data vectors are then fused to obtain a multi-dimensional student vector representation.

[0150] 4.14 Match the student's skills with the skill nodes in the talent knowledge graph to obtain the student's skill semantic vector. Use a graph neural network to generate a high-dimensional vector of the student's skill semantic vector. Then fuse the high-dimensional vector of the student's skill semantic vector with the multi-dimensional student vector representation to complete the deep semantic characterization of the student and generate a talent profile.

[0151] 4.2 Generation of Job Profiles:

[0152] 4.21 Collect multi-source job data and store the multi-source job data in the data warehouse; the multi-source job data includes:

[0153] Job requirements, job description, responsibilities, and essential skills;

[0154] Skills requirements, including technical skills, soft skills, and certification requirements;

[0155] Industry attributes, including the industry it belongs to, industry development trends, and industry standards;

[0156] Corporate culture, including corporate values, work environment, team structure, etc.

[0157] Other job information includes salary range, work location, and career advancement path.

[0158] 4.22 Feature extraction is performed on the multi-source data of the aforementioned positions to obtain:

[0159] Job numerical characteristics, salary range, years of work experience requirements, etc.;

[0160] Job classification characteristics, including job category, skill category, and industry category;

[0161] Job description features include job description, job responsibilities description, and company culture description.

[0162] The numerical features of the job positions are then standardized and normalized, the classification features of the job positions are encoded, and the textual features of the job positions are vectorized to complete the feature transformation.

[0163] 4.23 Using a fully connected layer, the numerical features and classification features of the job positions are converted into fixed-dimensional structured data vectors. A pre-trained language model is used to obtain the text data vectors of the text features of the job positions. The structured data vectors and text data vectors are then fused to obtain a multi-dimensional job vector representation.

[0164] 4.24 Match the skills required for the job with the skill nodes in the talent knowledge graph to obtain the job skill semantic vector. Use a graph neural network to generate a high-dimensional vector of the job skill semantic vector. Then fuse the high-dimensional vector of the job skill semantic vector with the multi-dimensional job vector representation to complete the deep semantic characterization of the job and generate a job profile.

[0165] 4.3 such as Figure 3 As shown, the generation of the talent matching model includes:

[0166] 4.31 Map the talent profile and the job profile to a unified vector space to generate talent profile vector and job profile vector.

[0167] 4.32 Using cosine similarity, calculate the cosine value of the angle between the talent profile vector and the job profile vector, and using Euclidean distance, calculate the straight-line distance between the talent profile vector and the job profile vector to obtain the semantic similarity between the talent profile vector and the job profile vector.

[0168] 4.33 Select multiple objective factors and assign dynamic weights to different objective factors, and then combine optimization algorithms and reinforcement learning to construct a multi-objective optimization model.

[0169] Multiple target factors are considered, such as: 1) Student personal preferences: including career preferences, expected salary, and work location preferences. These preferences are quantified through data sources such as questionnaires and course selection records. 2) Company recruitment preferences: including skill requirements, experience requirements, and cultural fit. Preference parameters are extracted based on job descriptions and historical recruitment data. 3) Potential development space: assessing the role of the position in promoting students' future career development, such as promotion opportunities and skills enhancement. Scores are derived through industry development trends and job growth path analysis. 4) Other factors: such as market demand and job scarcity.

[0170] Dynamic weighting: Identified based on market changes, company urgency, seasonality, and other dynamic factors. Market and company data are collected and processed in real-time using stream processing frameworks (such as Apache Kafka and Apache Flink), triggering weight adjustments. Example: Current market demand for "data science" is high, so related skills are weighted more, improving the matching score.

[0171] Optimization algorithms, such as linear weighted models (suitable for simple weighted scoring). Multi-objective optimization algorithms, such as Pareto optimization (ensuring balance among multiple objectives).

[0172] Reinforcement learning: Train the agent to dynamically adjust weights based on feedback to optimize matching results.

[0173] Example: Use deep reinforcement learning (such as DQN) to adjust weights based on historical matching results to improve the satisfaction of future matches.

[0174] 4.34 Combining the semantic similarity and the multi-objective optimization model, a matching algorithm for calculating the comprehensive matching score is obtained. The matching algorithm is then deployed into the local model to obtain the talent matching model.

[0175] The expression for the matching algorithm is:

[0176]

[0177] Where T is the overall matching score, w i For dynamic weights, f i denoted as the score for each objective factor, where m is the total number of objective factors.

[0178] 5. For example Figure 3 As shown, the similarity gap between the talent profile and the job profile is continuously monitored. Based on the similarity gap and the output of the talent matching model, course optimization plans and job optimization plans are dynamically generated. This includes:

[0179] 5.1 Use the talent matching model to obtain the target position and the skill requirements of the target position, and compare, analyze and update the skill requirements of the target position with the relevant data of the corresponding position in the talent knowledge graph.

[0180] 5.2 Assess and quantify students' current skills, compare them with the updated skill requirements of the target positions to identify skill gaps, and then use machine learning algorithms (such as cluster analysis and classification models) to analyze these skill gaps in order to recommend relevant courses, obtain course optimization plans, and predict and plan courses based on industry trends and student needs.

[0181] Assessing and quantifying students' current skills: Based on students' course grades, skill certifications, internship experience, and other information, assess their current skill levels. Use skill nodes in a knowledge graph and graph embedding technology to quantify students' skill levels.

[0182] 5.3 Monitor changes in job requirements and courses in real time. Based on the monitoring results and various target factors, optimize the target positions to obtain a job optimization plan.

[0183] 6. Utilize explainable AI technology to provide visual explanations or keyword descriptions of the output of the talent matching model, the course optimization plan, and the job optimization plan. This includes:

[0184] 6.1 Use feature contribution plots (bar charts, pie charts, etc.) to show the contribution of each feature to the matching and recommendation results. For example, a bar chart can show "Skill matching degree: 40%, Internship experience: 30%, Cultural fit: 20%".

[0185] 6.2 Utilize decision path diagrams, using decision trees or flowcharts to illustrate the decision path of the talent matching model. For example, a decision path diagram might show "Skill matching degree > 80% → Recommend position A".

[0186] 6.3 Utilize keyword clouds to demonstrate the driving factors for job recommendations. For example, the keyword cloud can display "skills matching, job requirements, and cultural fit" as the basis for recommendations.

[0187] 6.4 Provide dashboards for viewing data usage and the recommendation process. For example, after clicking "Skill Matching," users can view the matching status of specific skills.

[0188] 6.5 Include privacy policies to ensure transparent management. For example:

[0189] Data Source Explanation: Clearly state the source of the data used (e.g., student-submitted resumes, job postings from companies). For example, "The system used your course grades, internship experience, and career preference information to generate job recommendations."

[0190] Data Usage Description: Provide a detailed explanation of the data's intended use (e.g., for profile generation, matching model training). For example, "Your data will be used to generate personalized job recommendations and will not be used for other commercial purposes."

[0191] Privacy Policy Design: Data access control, clearly defining which users or organizations can access which data. For example, "Only authorized educational institutions and businesses can access your publicly available resume information."

[0192] Transparent approval process: Set up approval processes on the management side and record the approval history for each data access and recommendation operation.

[0193] Therefore, by providing a talent training recommendation method based on federated learning and natural language processing, this invention can not only improve matching accuracy and efficiency, break down data silos, and protect privacy and security, but also dynamically adapt to changes in market demand, meet the personalized needs of various parties, and promote the precise matching of education and employment.

[0194] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0195] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A talent training recommendation method based on federated learning and natural language processing, characterized in that, Includes the following steps: Collect relevant data from educational institutions, enterprises, and third-party platforms to obtain multi-source heterogeneous data. Store the multi-source heterogeneous data in a data lake, then clean and integrate the multi-source heterogeneous data to obtain standard data, and store the standard data in a data warehouse. Construct a talent knowledge graph related to talent development and job skills, and perform real-time incremental updates on the talent knowledge graph; Sensitive data in the standard data is anonymized and desensitized, and local models are obtained by joint modeling using a federated learning framework, without sharing the original data between educational institutions and enterprises. Based on the standard data, students and positions are characterized in depth to generate dynamically updated talent profiles and position profiles. The matching algorithm is then combined with multiple target factors and dynamic weights. The talent profiles, position profiles and matching algorithms are then deployed into the local model to obtain the talent matching model. Continuously monitor the similarity gap between the talent profile and the job profile, and dynamically generate course optimization plans and job optimization plans based on the similarity gap and the output of the talent matching model; The output of the talent matching model, the course optimization scheme, and the job optimization scheme are explained visually or described using keywords using explainable AI technology. Sensitive data in the standard data is anonymized and desensitized. Then, without sharing the original data between educational institutions and enterprises, a federated learning framework is used for joint modeling to obtain a local model, including: The standard data is categorized into personal privacy information, sensitive corporate data, and general data. The personal privacy information is encrypted using an irreversible hash algorithm, and a random salt value is added before the hash to prevent rainbow table attacks, thus completing the anonymization process of the personal privacy information. For the enterprise's sensitive data, regular expressions are used to identify sensitive fields, and hashing, masking, or obfuscating operations are performed on the sensitive fields to complete the desensitization process of the enterprise's sensitive data. Educational institutions and enterprises are set as participants. Each participant uses its own data to train the model locally, obtains the local model and local parameters, and completes the deployment of the local node. The local parameters are encrypted using homomorphic encryption to obtain encrypted parameters. The encrypted parameters are received by the central server, and the encrypted parameters are aggregated to obtain global parameters, thus completing the deployment of the cloud node. The central server is used to send the global parameters to each of the participating parties to perform a global update of the local model.

2. The talent training recommendation method based on federated learning and natural language processing according to claim 1, characterized in that, Data is collected from educational institutions, enterprises, and third-party platforms to obtain multi-source heterogeneous data. This multi-source heterogeneous data is then stored in a data lake. Next, the multi-source heterogeneous data undergoes data cleaning and integration to obtain standard data, which is then stored in a data warehouse. This process includes: Based on a unified data format and field definition, relevant data from educational institutions, enterprises, and third-party platforms are obtained using API interfaces or web crawling technology to obtain multi-source heterogeneous data. A data lake is constructed using a distributed storage system, and the multi-source heterogeneous data is partitioned and stored in the data lake according to the data source and type; The multi-source heterogeneous data is processed for missing values, outliers, duplicate data, and data integration to obtain standard data; A data warehouse is built using a relational database. Based on the data theme, the standard data is partitioned and stored in the data warehouse, and a query index is created in the data warehouse.

3. The talent training recommendation method based on federated learning and natural language processing according to claim 2, characterized in that, Construct a talent knowledge graph related to talent development and job skills, and perform real-time incremental updates to the talent knowledge graph, including: Collect job-related data, industry-related data, and course-related data to obtain initial text; preprocess the initial text to obtain knowledge text. The BERT model is used to perform contextual semantic understanding of the knowledge text, and NLP technology is used to extract entities from the knowledge text. Then, the entities are mapped to a standardized lexicon by combining an external knowledge base to complete the entity standardization operation. Dependency parsing is used to extract relationships between entities, and these relationships are then classified. Select a graph database, import entities and relations into the selected graph database to obtain a graph structure, use graph embedding technology to map the nodes and relations in the graph structure to a vector space, generate node embedding vectors, complete semantic matching operations, and obtain a talent knowledge graph; The system automatically parses new data from the initial text and updates the talent knowledge graph based on the parsing results.

4. The talent training recommendation method based on federated learning and natural language processing according to claim 3, characterized in that, Based on the aforementioned standard data, students are subjected to in-depth semantic characterization to generate dynamically updated talent profiles, including: Collect multi-source student data and store it in the data warehouse; the multi-source student data includes student course grades, skill mastery, internship experience, and career preferences; Feature extraction is performed on the student multi-source data to obtain student numerical features, student classification features and student text features. Then, the student numerical features are standardized and normalized, the student classification features are encoded, and the student text features are vectorized to complete the feature transformation. Using a fully connected layer, the student numerical features and the student classification features are converted into fixed-dimensional structured data vectors. A pre-trained language model is used to obtain the text data vectors of the student text features. Then, the structured data vectors and the text data vectors are fused to obtain a multi-dimensional student vector representation. The student's skills are matched with the skill nodes in the talent knowledge graph to obtain the student's skill semantic vector. A high-dimensional vector of the student's skill semantic vector is generated using a graph neural network. The high-dimensional vector of the student's skill semantic vector is then fused with the multi-dimensional student vector representation to complete the deep semantic characterization of the student and generate a talent profile.

5. The talent training recommendation method based on federated learning and natural language processing according to claim 4, characterized in that, Based on the aforementioned standard data, a deep semantic characterization of the job positions is performed to generate dynamically updated job profiles, including: Collect multi-source job data and store the multi-source job data in the data warehouse; the multi-source job data includes job requirements, skill requirements, industry attributes, and corporate culture; Feature extraction is performed on the multi-source data of the job positions to obtain job numerical features, job classification features, and job text features. Then, the job numerical features are standardized and normalized, the job classification features are encoded, and the job text features are vectorized to complete the feature transformation. Using a fully connected layer, the numerical features and classification features of the job positions are converted into fixed-dimensional structured data vectors. A pre-trained language model is used to obtain text data vectors of the text features of the job positions. The structured data vectors and text data vectors are then fused to obtain a multi-dimensional job vector representation. The skills required for the job are matched with the skill nodes in the talent knowledge graph to obtain the job skill semantic vector. A high-dimensional vector of the job skill semantic vector is generated using a graph neural network. The high-dimensional vector of the job skill semantic vector is then fused with the multi-dimensional job vector representation to complete the deep semantic characterization of the job and generate a job profile.

6. The talent training recommendation method based on federated learning and natural language processing according to claim 5, characterized in that, By combining multiple target factors and a dynamically weighted preset matching algorithm, and then deploying the talent profile, the job profile, and the matching algorithm into the local model, a talent matching model is obtained, including: The talent profile and the job profile are mapped to a unified vector space to generate talent profile vector and job profile vector. Using cosine similarity, the cosine value of the angle between the talent profile vector and the job profile vector is calculated, and using Euclidean distance, the straight-line distance between the talent profile vector and the job profile vector is calculated to obtain the semantic similarity between the talent profile vector and the job profile vector. Select multiple objective factors, assign dynamic weights to different objective factors, and then combine optimization algorithms and reinforcement learning to build a multi-objective optimization model; By combining the semantic similarity and the multi-objective optimization model, a matching algorithm for calculating the comprehensive matching score is obtained. The matching algorithm is then deployed to the local model to obtain the talent matching model.

7. The talent training recommendation method based on federated learning and natural language processing according to claim 6, characterized in that, The expression for the matching algorithm is: ； in, To calculate the overall matching score, For dynamic weights, Scores for each objective factor, The total number of target factors.

8. The talent training recommendation method based on federated learning and natural language processing according to claim 7, characterized in that, Continuously monitor the similarity gap between the talent profile and the job profile, and dynamically generate course optimization plans and job optimization plans based on the similarity gap and the output of the talent matching model, including: The talent matching model is used to obtain the target position and the skill requirements of the target position. The skill requirements of the target position are then compared, analyzed and updated with the relevant data of the corresponding position in the talent knowledge graph. The assessment quantifies students' current skills, compares these skills with the updated skill requirements of the target positions to identify skill gaps, and then uses machine learning algorithms to analyze these skill gaps to recommend relevant courses, obtain course optimization plans, and predict and plan courses based on industry trends and student needs. Real-time monitoring of changes in job requirements and curriculum; optimization of target positions based on monitoring results and various target factors to obtain job optimization solutions.

9. The talent training recommendation method based on federated learning and natural language processing according to claim 8, characterized in that, The output of the talent matching model, the course optimization scheme, and the job optimization scheme are explained visually or with keywords using explainable AI technology, including: The feature contribution map is used to show the contribution of each feature to the matching and recommendation results; the decision path map is used to show the decision path of the talent matching model; and the keyword cloud is used to show the driving factors of job recommendation. Provide dashboards for viewing data usage and the recommendation process, and design privacy policies to achieve transparent management.

Images