Professional recommendation post group system and method based on large model and two-way recall

Abstract

The invention provides a professional recommendation post group system and method based on a large model and two-way recall, and relates to the technical field of educational informatization. The post group capability processing module, the professional capability processing module, the capability matching module, the employment matching module, the two-way recall and score merging module and the report generation module are connected with the calculation processing unit, and the calculation processing unit interacts with the large model service interface, the knowledge graph database and the capability vector database. According to the method, deep semantic understanding is carried out on the unstructured professional cultivation scheme through the large language model, the professional ability corresponding to multiple dimensions of course setting, practice links and graduation requirements can be dynamically recognized and quantified, the defects that traditional manual extraction is low in efficiency, quantification is rough, and updating is lagged are overcome, and the method is suitable for large-scale popularization and application. Therefore, the professional ability portrait can reflect the professional training reality more truly, and a high-quality data basis is provided for subsequent matching.

Description

Technical Field

[0001] This invention relates to the field of educational informatization technology, and in particular to a professional recommendation job group system and method based on a large model and dual-path recall. Background Technology

[0002] Currently, higher education faces significant challenges in cultivating talent to meet the dynamically changing job demands of society and industry. With accelerated industrial upgrading and job iteration, the mismatch between professional training programs and actual job requirements is becoming increasingly prominent, leading to a structural contradiction between talent supply and demand. Some graduates face employment difficulties, while others face talent shortages. The core issue lies in the limitations of existing technologies in the precision, efficiency, and depth of matching majors with job requirements, failing to meet dynamic and refined matching needs.

[0003] Existing technologies have several shortcomings in addressing the aforementioned problems: First, they struggle to accurately and dynamically extract and quantify core competencies in professional training programs. Existing methods often rely on manual experience or coarse keyword matching, failing to effectively parse unstructured training program texts. Furthermore, competency quantification lacks scientific basis and fails to reflect changes in competency focus resulting from dynamic revisions of the training program. Second, competency matching mechanisms lack semantic flexibility. Most systems rely on pre-set standard competency vocabularies, unable to identify synonymous competency descriptions, leading to insufficient matching accuracy and adaptability. Third, recommendation dimensions are singular, focusing only on surface-level competency comparisons between majors and positions, without integrating market data such as graduate employment destinations, resulting in recommendations that are out of sync with actual employment trends. Fourth, recommendation results lack interpretability, only outputting recommendation lists or scores, failing to provide users with matching criteria and targeted development suggestions, thus limiting their guiding value. Therefore, this invention proposes a professional job recommendation system and method based on a large model and dual-path recall to address the problems existing in the prior art. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a professional job recommendation system and method based on a large model and dual-path recall. By using a large language model to perform deep semantic understanding of unstructured professional training programs, it can dynamically identify and quantify professional abilities corresponding to multiple dimensions such as curriculum settings, practical components, and graduation requirements. This overcomes the shortcomings of traditional manual extraction, such as low efficiency, coarse quantification, and delayed updates, making the professional ability profile more realistically reflect the actual professional training and providing a high-quality data foundation for subsequent matching.

[0005] To achieve the objectives of this invention, the following technical solution is provided: a professional job recommendation system based on a large model and dual-path recall, comprising a computational processing unit, and job group capability processing module, professional capability processing module, capability matching module, employment matching module, dual-path recall and score merging module, and report generation module connected to the computational processing unit. The computational processing unit interacts with a large model service interface, a knowledge graph database, and a capability vector database, respectively. The job group capability processing module extracts job group capabilities and corresponding requirement scores from the knowledge graph database, encodes them using a pre-trained language model, stores them in the capability vector database, and caches the job group capability data. The professional capability processing module receives professional training program text, calls a large language model through the large model service interface to extract professional capabilities and quantify scores, encodes them, stores them in the capability vector database, and caches the professional capability data.

[0006] The capability matching module calculates the semantic similarity between professional capabilities and job group capabilities based on the capability vector database, and obtains a capability matching score by combining the capability score; the employment matching module extracts graduate employment destination data from the knowledge graph database, constructs a collaborative filtering matrix, and obtains job group employment destination scores through a collaborative filtering algorithm; the dual-path recall and score merging module weighted and merges the capability matching score and employment destination score to obtain a comprehensive recommendation score, and filters a preset number of recommended job group lists according to the score; the report generation module filters the intermediate calculation data corresponding to the recommended job groups, and generates a recommendation report by calling the large language model through the large model service interface.

[0007] A further improvement is that the job group capability processing module adopts an offline batch processing method, specifically: iterates through each capability of each extracted job group and queries whether the capability exists in the metadata of the capability vector database; if not, it generates capability vectors by encoding through a pre-trained language model and stores the capability vectors and capability-vector ID mapping relationships in the vector storage area and metadata area of ​​the capability vector database, respectively; if yes, it does not update the capability vector database.

[0008] Further improvements are made in the following aspects: In the professional competence processing module, the large language model quantifies the professional competence score based on the curriculum setting, practical links, and graduation requirements of the professional training program, and normalizes the competence score to the [0,10] interval by truncation; the pre-trained language model is the MiniLMv2 model, and the competence vector database is the Faiss vector database.

[0009] A further improvement is made in the following calculation process of the capability matching module: traverse each capability of each job group, query the capability vector database for the K professional capabilities with the highest semantic similarity to the capabilities of that job group, and filter the professional capabilities with similarity reaching a preset threshold; take the professional capability with the highest similarity as the main contributing capability, and the rest that meet the threshold as auxiliary contributing capabilities; the capability matching score is the average of the main contributing capability score and the auxiliary contributing capability score, where the main contributing capability score = similarity × professional capability score × job group capability requirement score, and the auxiliary contributing capability score = similarity² × professional capability score × job group capability requirement score.

[0010] Further improvements are made in the following ways: In the job matching module, the rows of the collaborative filtering matrix represent the same major from different schools, the columns represent different job groups, and the matrix value is the number of graduates of that major who enter the corresponding job group; the collaborative filtering algorithm identifies other schools with the same major that have similar job destinations to the current major and recommends job groups that similar majors often go to in order to obtain an employment destination score.

[0011] A professional job recommendation method based on a large model and dual-path recall includes the following steps:

[0012] S1: Job group capability preprocessing: Extract job group capabilities and corresponding requirement scores from the knowledge graph database, encode them using a pre-trained language model, and store them in the capability vector database to cache job group capability data;

[0013] S2: Professional competency processing: Receives the professional training program text, calls the large language model to extract professional competencies and quantifies their scores, encodes them and stores them in the competency vector database, and caches the professional competency data.

[0014] S3: Ability matching calculation, which calculates the semantic similarity between professional ability and job group ability based on the ability vector database, and obtains the ability matching score by combining the ability score;

[0015] S4: Job matching calculation: Extract graduate employment destination data from knowledge graph database, construct collaborative filtering matrix, and obtain job destination score of job group through collaborative filtering algorithm;

[0016] S5: Dual-path recall and score merging, weighted merging of ability matching score and employment destination score to obtain comprehensive recommendation score, and filtering a preset number of recommended job groups according to the score;

[0017] S6: Generate a recommendation report, filter the intermediate calculation data corresponding to the recommended job groups, and call the large language model to process the data and generate the recommendation report.

[0018] The further improvement is that: in S1, the pre-trained language model is the MiniLMv2 model, and the capability vector database is the Faiss vector database; S1 also includes: querying whether the job group capabilities exist in the metadata of the Faiss vector database, and if not, performing encoding and storage operations.

[0019] Further improvements are made in S2, where the large language model quantifies professional competence scores based on the curriculum, practical components, and graduation requirements of the professional training program, and normalizes the competence scores to the [0,10] interval using a truncation method.

[0020] A further improvement is made in S3, where semantic similarity is obtained through the similarity search function of the capability vector database, and K professional capabilities with similarity reaching a preset threshold are selected; the capability matching score is calculated as follows: main contributing capability score = similarity × professional capability score × job group capability requirement score, auxiliary contributing capability score = similarity² × professional capability score × job group capability requirement score, and the capability matching score is the sum and average of the main contributing capability score and the auxiliary contributing capability score.

[0021] Further improvements are made in S6, where the intermediate calculation data includes a list of professional abilities and scores, details of ability matching between majors and job groups, employment destination data for job groups, collaborative filtering scores, ability matching scores, employment destination scores, and comprehensive recommendation scores; the recommendation report includes an overview of professional abilities, ability matching analysis, employment matching analysis, and personalized development suggestions.

[0022] The beneficial effects of this invention are as follows:

[0023] 1. This invention uses a large language model to perform deep semantic understanding of unstructured professional training programs. It can dynamically identify and quantify the professional abilities corresponding to multiple dimensions such as curriculum settings, practical links, and graduation requirements. It overcomes the shortcomings of traditional manual extraction, such as low efficiency, rough quantification, and lagging updates, so that the professional ability profile more realistically reflects the actual professional training and provides a high-quality data foundation for subsequent matching.

[0024] 2. This invention uses pre-trained language model vector embedding and capability vector database to achieve semantic matching for capability ontology without standard. It can intelligently identify capability descriptions with the same name but different names, solving the problem of rigid matching in traditional methods. At the same time, by integrating capability matching score and employment destination score through a dual-path recall strategy, the recommendation results take into account both professional hard skills and market soft skills, greatly improving practicality and accuracy.

[0025] 3. This invention utilizes a large language model to transform intermediate computational data into a recommendation report that includes professional ability analysis, matching details, employment trends, and personalized suggestions. This completely changes the traditional black-box recommendation system, not only increasing users' trust in the recommendation results but also providing specific data support and decision-making basis for students' career planning and the optimization of university professional training programs, significantly enhancing the application value of the system. Attached Figure Description

[0026] Figure 1 This is a diagram illustrating the overall architecture of the present invention;

[0027] Figure 2 This is a schematic diagram of the job group capability processing module of the present invention;

[0028] Figure 3 This is a schematic diagram of the professional capability processing module of the present invention;

[0029] Figure 4 This is a schematic diagram of the capability matching module of the present invention;

[0030] Figure 5 This is a schematic diagram of the employment matching module of the present invention;

[0031] Figure 6 This is a schematic diagram of the dual-path recall and score merging module of the present invention;

[0032] Figure 7 This is a schematic diagram of the report generation module of the present invention. Detailed Implementation

[0033] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0034] Example 1

[0035] according to Figure 1 , 2 As shown in Figures 3, 4, 5, 6, and 7, this embodiment proposes a professional job recommendation system and method based on a large model and dual-path recall. The system includes a job cluster capability processing module, a professional capability processing module, a capability matching module, an employment matching module, a dual-path recall and score merging module, and a report generation module. The core of the system runs on a single computational processing unit and interacts with a large model (LLM) service interface, a knowledge graph database, and a capability vector database (Faiss).

[0036] The system comprises the following components: User Input Layer: Receives the text of the professional training program to be analyzed and inputs it into the Large Model (LLM) service interface for processing. Knowledge Graph Database: Stores the competency names, demand scores, and graduate employment destination data for job clusters. Job Cluster Competency Processing Module: Extracts job cluster competencies and scores in batches from the knowledge graph and encodes the competency names into vectors using a pre-trained model, storing them in the Faiss vector database. Faiss Vector Database: Faiss (FacebookAISimilaritySearch) is a vector database for efficient similarity search and dense vector clustering. Professional Competency Processing Module: Analyzes the professional training program text using the Large Model (LLM), extracts professional competencies, and quantifies their scores. Encodes the competency names into vectors using a pre-trained model and stores them in the Faiss vector database. Competency Matching Module: Calculates competency matching scores using appropriate algorithms based on professional competencies, job cluster competencies, and the Faiss vector database. Employment Matching Module: Establishes a collaborative filtering matrix based on graduate employment destination data from the knowledge graph and calculates employment destination scores by combining this data with employment data from other schools for the same major. The dual-path recall and score merging module: This module weights and merges the ability matching score and employment destination score to obtain a comprehensive recommendation score, and generates a recommendation list based on this score. The report generation module: Based on the recommendation list, this module filters intermediate calculation data and provides it to the large-scale model (LLM) to generate a detailed report.

[0037] Output: Recommended Job Group List: The final output to the user is a list of the Top 10 recommended job groups. Recommendation Report: The final output to the user is a detailed analysis report.

[0038] The job cluster capability processing module is primarily responsible for building the basic data required for system operation, especially preprocessing job cluster capabilities, including extracting capabilities, scores, and coding abilities from the knowledge graph and storing them in the database. This process is performed offline in batches to ensure efficiency during online matching.

[0039] See Figure 2 Specifically, the process includes the following steps:

[0040] Start: The job group capability processing flow is initiated. Extracting job group capabilities and scores from the knowledge graph: The system first accesses the knowledge graph database to batch retrieve data for all defined job groups, including the required capability list and corresponding requirement scores for each job group. Traversing each capability within each job group: The system iteratively processes each capability within each job group. Checking if the capability is already stored in Faiss: The system checks if the capability name has an encoded vector. The Faiss database contains two data items: an ID and vector data, and metadata storing the mapping relationship between job group capability names and IDs. Checking if the capability name exists in the metadata: Existing: If the query result shows that the encoding for the capability name already exists, the vector database does not need to be updated. Not Existing: If the capability name appears for the first time, the system enters the encoding process. Pre-trained model encoding: The system calls the pre-trained MiniLMv2 model to encode the capability name, generating a high-dimensional numerical vector representing the semantic information of the capability name. MiniLMv2 (MinimalLanguageModelv2) is a lightweight yet efficient pre-trained language model capable of converting sentences or phrases into semantically rich vector embeddings. Update the Faiss vector database: Add the newly generated vectors to the Faiss vector database; Faiss assigns a unique internal ID to each added vector. Then, add a mapping between capability names and vector IDs in the metadata. Cache job group capability data: Cache job group capabilities and scores. End: Data preprocessing complete.

[0041] The professional competence processing module is responsible for processing the professional training program input by the user, dynamically extracting and quantifying the competences within it, and providing a "profile" of professional competences for subsequent matching.

[0042] See Figure 3The specific process includes the following steps: Start: The professional competency processing flow is initiated. Receiving the professional training program: The system receives the raw text data of the professional training program input by the user. Calling the Large Model for Analysis, Extraction, and Quantification: The system inputs the professional training program text into the Large Model (LLM) service interface. LLM utilizes its powerful text understanding and information extraction capabilities to intelligently identify and extract the core competencies cultivated by the major from the unstructured description of the training program, such as data analysis, project management, communication and coordination, and proficiency in specific programming languages. LLM not only extracts competency names, but further, it can comprehensively evaluate and calculate a quantitative score for the extracted competencies based on multiple dimensions of information in the training program, including curriculum design, practical components, graduation requirements, credit allocation, and teaching objectives. For example, if a major offers multiple data analysis-related courses, has data analysis practical projects, and emphasizes data analysis skills in graduation requirements, LLM will assign a higher score to the data analysis skills. Normalized Score: The raw ability scores calculated by the large model are normalized to the range [0,10] using a truncation method to ensure consistency with the unit of measurement of the job group ability scores. Traversing Each Professional Ability: The system iteratively processes each professional ability extracted by the large model. Checking if an Ability is Already Stored in Faiss: The system checks if the ability name has an encoded vector. The Faiss database contains two data items: an ID and vector data, and metadata storing the mapping relationship between job group ability names and IDs. Checking if the ability name exists in the metadata: Existing: If the query results show that the encoding of the ability name already exists, the vector database does not need to be updated. Not Existing: If the ability name appears for the first time, the system enters the encoding process. Pre-trained Model Encoding: The system calls the pre-trained MiniLMv2 model to encode the ability name, generating a high-dimensional numerical vector that represents the semantic information of the ability name. MiniLMv2 (MinimalLanguageModelv2) is a lightweight yet efficient pre-trained language model capable of converting sentences or phrases into semantically rich vector embeddings. The process involves updating the Faiss vector database: adding newly generated vectors to the Faiss vector database, where Faiss assigns a unique internal ID to each added vector. Then, a mapping between ability names and vector IDs is added to the metadata. Professional ability data is cached: professional abilities and scores are cached. The professional ability processing is complete.

[0043] The competency matching module is one of the core recommendation paths, responsible for calculating the matching degree between professional training capabilities and the competencies required by job groups.

[0044] See Figure 4The specific process includes the following steps: Start: The capability matching process is initiated. Traverse each capability of each job group: The system iterates through each job group sequentially and matches each capability within each group. Query the K most similar capabilities in Faiss: Based on the currently calculated job group capabilities, first query the vector ID in the metadata based on the capability name, then use the obtained vector ID to query the vector in Faiss. Use the Faiss database method to find the K most similar capabilities, and filter out similar capabilities with a similarity threshold, obtaining a list of capabilities and similarities. Check if the K capabilities are in the professional capabilities: Match each capability of the professional capability to the list of similar capabilities using keywords that are exactly equal. Professional capability matches a similar capability: Highest similarity: This professional capability is considered a primary contributing capability. Not the highest similarity: This professional capability is considered a secondary contributing capability. Professional capability does not match a similar capability: This professional capability is a failed match. Calculate the skill matching score:

[0045]

[0046] For the overall score of a single job group, the score for the primary contributing ability is the product of the similarity score, the professional ability score, and the job group ability score. For the secondary contributing ability, the similarity score is squared, which is a designed decay rule. Finally, the scores are summed and averaged to obtain the overall score. Calculating the job group matching score: Each job group is traversed sequentially to obtain the overall ability score for each job group. End: The ability matching process is complete.

[0047] The job matching module is responsible for using graduate employment destination data and calculating the job matching degree between majors and job groups through collaborative filtering algorithms.

[0048] See Figure 5 The specific process includes the following steps: Start: The job matching process is initiated. Extracting employment data from the knowledge graph: The system extracts large-scale graduate employment destination data from the knowledge graph database. This data typically includes information such as: school, major, job cluster, and the number of graduates entering that job cluster. Constructing a collaborative filtering matrix for the same major across different schools: The system uses the extracted data to construct a sparse matrix. Rows represent different schools with the same major (which can be considered "users" in collaborative filtering), columns represent different job clusters (which can be considered "items" in collaborative filtering), and values ​​in the matrix represent the number of graduates from that major entering a particular job cluster. Collaborative filtering to calculate school similarity: The system applies a collaborative filtering algorithm to this matrix. It identifies other schools with the same major whose employment destinations are similar to the current school's, and then recommends job clusters frequently visited by graduates from these similar schools with the same major. Calculating job cluster matching score: The output of the collaborative filtering algorithm is the employment destination score for each job cluster relative to the current major. This score reflects the historical tendency or probability of graduates from that major entering a particular job cluster. End: The job matching process is completed.

[0049] The dual-path recall and score merging module is responsible for weighted merging of the scores from ability matching and job matching to generate the final comprehensive recommendation score.

[0050] See Figure 6 The specific process includes the following steps: Start: The dual-path recall and score merging process is initiated. Ability Matching Score: Obtain the ability matching score between each job group and the current major from the ability matching module. Employment Matching Score: Obtain the employment destination score between each job group and the current major from the employment matching module. Weighted Merging Score: The system presets or dynamically defines two core weights: the weight of ability and the weight of employment. Typically, the sum of the two is 1 to ensure the rationality of the total score. These weights can be adjusted according to business needs; for example, increasing the ability weight if more emphasis is placed on "hard skills," and increasing the employment weight if more emphasis is placed on "market trends." Iterate through each job group, processing each job group to be recommended, and calculate its final comprehensive recommendation score according to the following formula:

[0051]

[0052] Top 10 Job Groups: The system sorts all job groups from highest to lowest based on their overall recommendation scores. Based on a preset recommendation quantity N (e.g., 10), the system selects the N highest-scoring job groups from the sorted list to form the final recommendation list for the user. End: The dual-path recall and score merging process is complete.

[0053] The report generation module is responsible for filtering intermediate calculation data for job clusters based on the final recommendation list and using a large model to generate detailed and interpretable recommendation reports.

[0054] See Figure 7The specific process includes the following steps: Start: The report generation process is initiated. Data is filtered based on the top 10 job groups: To generate a detailed report, the system collects all key data generated during the matching process, including but not limited to: the professional competency list and scores extracted by the large model; details of competency matching between the major and each job group; historical employment destination data and collaborative filtering scores for the job groups; weighting of competency matching and employment destination scores; and the final comprehensive recommendation score for each job group. The system retains the data for the top 10 job groups based on the list output by the dual-path recall and score merging module. The filtered data is input into the large model: The system inputs the filtered computational data related to the top 10 job groups into the large model through the large model (LLM) service interface and designs specific prompts. The large model generates a detailed report: The large model analyzes the data based on the prompts and generates a detailed recommendation report. LLM utilizes its powerful text understanding and generation capabilities to conduct in-depth analysis of this data and write a detailed recommendation report containing the following: Professional Competency Overview: A summary and explanation of the various competencies extracted from the professional training program. Competency Matching Analysis: This section details the alignment between the skills trained in the major and the skills required for the recommended job clusters, highlighting key competency points for successful matching. Employment Matching Analysis: Based on the collaborative filtering results, this section analyzes the main employment flows and trends of graduates from this major. Development Recommendations: Based on the strengths and weaknesses of competency matching and market trends, this section provides students with personalized suggestions for skill enhancement and career development guidance. Conclusion: The entire recommendation system process is complete.

[0055] Large LLM (LLM) selection: DeepseekR1, a mainstream technology that excels in general semantic understanding and text generation, is chosen. Its capabilities in processing unstructured text, performing semantic understanding, and complex reasoning far surpass traditional NLP methods, providing the foundation for dynamic, refined capability extraction and intelligent report generation.

[0056] The choice of MiniLMv2: MiniLMv2 is a lightweight yet high-performance pre-trained model suitable for generating high-quality embeddings. Compared to large models like BERT, it maintains high semantic representation capabilities while consuming fewer computational resources and achieving faster inference speeds. This is crucial for this system, which requires frequent capability vectorization and similarity calculations, ensuring real-time responsiveness.

[0057] The choice of Faiss vector database: Faiss is an open-source, high-efficiency vector similarity search library launched by Facebook AI. It supports various Approximate Nearest Neighbor (ANN) algorithms, enabling extremely fast searches in large-scale vector datasets. This invention requires rapid matching of massive job competency vectors and dynamically generated professional competency vectors; Faiss's high performance is its core advantage. Its open-source nature also means lower licensing costs and more flexible deployment methods.

[0058] Knowledge graphs as a data source: Knowledge graphs provide a structured and semantically rich way to store and manage job skill requirements, historical employment data, and more. They ensure the authority, consistency, and scalability of this data, avoiding the inefficiency and uncertainty of repeatedly extracting data from unstructured text.

[0059] Semantic similarity threshold: key to connecting "non-standard capability ontologies". This threshold (usually between 0 and 1) determines the degree of similarity between two capability vectors to be considered a match. Too low a threshold introduces noise, while too high a threshold may miss valid matches. Its optimal value is determined through cross-validation with a large amount of real-world data and manual evaluation.

[0060] Dual-path recall weighting: These two weights determine the relative importance of the competency matching score and the employment destination score in the final comprehensive recommendation score. If the goal is to emphasize the hard skills of professional training, a higher competency weight (0.7) can be set. If more emphasis is placed on actual market employment orientation, the employment weight can be appropriately increased (0.5). The weights are not fixed and can be dynamically adjusted and optimized through A / B testing, user feedback, or expert experience.

[0061] Collaborative Filtering Algorithm Selection and Parameters: There are various collaborative filtering algorithms, such as user-based KNN, item-based KNN, and matrix factorization (SVD, ALS). The specific choice affects the algorithm's ability to handle data sparsity and its recommendation accuracy. For new fields, item-based collaborative filtering may be more advantageous; for large datasets, matrix factorization is more efficient.

[0062] Example 2

[0063] according to Figure 1 , 2 As shown in Figures 3, 4, 5, 6, and 7, this embodiment proposes a professional job recommendation system and method based on a large model and dual-path recall.

[0064] System Deployment and Data Preparation

[0065] Deploy computing processing units (such as server clusters), configure large model service interfaces (connecting to GPT series or open source LLM models), knowledge graph databases (store job group capabilities (such as "data analysis" and required score 8), graduate employment destination data (school A - computer science major - internet job group - 50 people)), and Faiss vector databases (used to store capability vectors and metadata).

[0066] The job group capability processing module performs offline batch processing: It extracts the capabilities and requirement scores of all job groups (such as Internet and Fintech) from the knowledge graph, encodes capabilities such as "data analysis" and "data processing skills" into vectors through the MiniLMv2 model, queries Faiss metadata, and if the capability does not exist, stores the vector and "capability-vector ID" mapping in Faiss, thus caching the job group capability data.

[0067] Professional skills processing

[0068] The program receives the curriculum text of a university's "Computer Science and Technology" major, calls LLM to extract the major's capabilities (such as "data analysis" and "Python programming"), and quantifies the "data analysis" capability score as 9 based on "3 data analysis courses + 2 practical projects + graduation requirement including a data analysis thesis", normalizing it to the [0,10] interval; it generates a vector of "data analysis" through MiniLMv2 encoding, stores it in Faiss and caches it.

[0069] Capability matching calculation

[0070] The ability matching module iterates through the "data analysis" ability of the Internet job group, queries the K=5 professional abilities with the highest similarity in Faiss, and filters "data analysis" (similarity 0.98, main contributing ability) and "data processing" (similarity 0.85, auxiliary contributing ability) with a similarity ≥ 0.8; calculates the ability matching score: (0.98×9×8+0.85²×8×8) / 2≈70.2.

[0071] Job matching calculation

[0072] The job matching module extracts computer science employment data from the knowledge graph and constructs a collaborative filtering matrix (rows: computer science majors in schools A and B; columns: internet and fintech job groups; values: 50 and 30). Through the collaborative filtering algorithm, it identifies school B with similar employment prospects to computer science majors in school A, recommends internet job groups frequently visited by students from school B, and calculates an employment destination score of 85.

[0073] Dual-path recall and report generation

[0074] The dual-path recall and score merging module has a capability weight of 0.6 and an employment weight of 0.4. The comprehensive recommendation score is 70.2 × 0.6 + 85 × 0.4 = 76.12. The top 10 recommended job groups are selected by sorting the scores (the Internet job group ranks first).

[0075] The report generation module filters intermediate data (skill matching details, employment data, etc.) for this job group, inputs LLM, and generates a report: "Computer science major's 'data analysis' skill score is 9, with a similarity of 0.98 to the Internet job group's requirements (score 8), indicating a high match; 50 people have entered this job group in the past, showing a good employment trend; it is recommended to improve 'cloud computing' skills to enhance competitiveness."

[0076] This invention utilizes Large LLM (LLM) for deep semantic understanding of unstructured professional training programs, enabling dynamic and precise identification and quantification of various training capabilities. This overcomes the shortcomings of traditional methods, such as low efficiency, coarse quantification, and delayed updates, allowing professional capability profiles to more realistically and precisely reflect the actual situation of professional training. Through comprehensive consideration of multi-dimensional information, the scientific and objective nature of capability scores is ensured, providing a high-quality, dynamically updated data foundation for subsequent matching. This invention employs pre-trained models such as MiniLMv2 for vector embedding of non-standard capability ontology and uses the Faiss vector database for efficient semantic similarity calculation. This enables the system to intelligently identify and match semantically similar but differently expressed capabilities, effectively solving the matching omissions and rigidity problems caused by relying on fixed standard tables or precise name correspondences in existing technologies. This highly flexible semantic matching mechanism greatly improves the recall and accuracy of matching between professional capabilities and job group capabilities, more realistically reflecting their deep compatibility. This invention innovatively introduces a dual-path recall strategy, weighting and merging capability matching scores with collaborative filtering scores based on employment destination data. This overcomes the limitations of single-dimensional recommendations, ensuring that the recommendations consider not only the hard skills of professional training but also the soft skills relevant to actual job market employment. This multi-dimensional integration provides more comprehensive, market-relevant, and valuable recommendations, offering students balanced guidance that considers both academics and career prospects. This invention leverages the powerful text generation capabilities of Large Module Models (LLM), using all intermediate computational data and matching details as input to automatically generate detailed, interpretable, and insightful recommendation reports. This completely transforms the black-box output of traditional recommendation systems, providing users with clear recommendation criteria, in-depth professional skills analysis, historical employment trend reviews, and even personalized development suggestions. This highly intelligent report generation significantly enhances user trust in the recommendation results and truly guides students in personalized career development planning. It also provides universities with concrete data support and decision-making basis for optimizing their professional training programs.

[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A professional job recommendation system based on a large model and dual-path recall, comprising a computational processing unit, and job group capability processing module, professional capability processing module, capability matching module, employment matching module, dual-path recall and score merging module, and report generation module connected to the computational processing unit, characterized in that: The computing and processing unit interacts with the large model service interface, the knowledge graph database, and the capability vector database, respectively. The job group capability processing module is used to extract job group capabilities and corresponding requirement scores from the knowledge graph database, encode them using a pre-trained language model, store them in the capability vector database, and cache the job group capability data. The professional capability processing module is used to receive professional training program text, call the large language model through the large model service interface to extract professional capabilities and quantify scores, encode them, store them in the capability vector database, and cache the professional capability data. The capability matching module is used to calculate the semantic similarity between professional capabilities and job group capabilities based on the capability vector database, and to obtain a capability matching score by combining the capability score. The employment matching module is used to extract graduate employment destination data from the knowledge graph database, construct a collaborative filtering matrix, and obtain employment destination scores for job groups through a collaborative filtering algorithm; the dual-path recall and score merging module is used to weight and merge the ability matching score and employment destination score to obtain a comprehensive recommendation score, and filter a preset number of recommended job groups according to the score; the report generation module is used to filter the intermediate calculation data corresponding to the recommended job groups, and generate a recommendation report by calling the large language model through the large model service interface.

2. The professional job recommendation system based on a large model and dual-path recall according to claim 1, characterized in that: The job group capability processing module adopts an offline batch processing method, which specifically involves: traversing each capability of each extracted job group and querying whether the capability exists in the metadata of the capability vector database; If no, then the capability vector is generated by encoding through a pre-trained language model, and the capability vector and the capability-vector ID mapping relationship are stored in the vector storage area and metadata area of ​​the capability vector database, respectively; if yes, then the capability vector database is not updated.

3. The professional job recommendation system based on a large model and dual-path recall according to claim 1, characterized in that: In the professional competence processing module, the large language model quantifies the professional competence score based on the curriculum, practical components, and graduation requirements of the professional training program, and normalizes the competence score to the [0,10] interval using a truncation method; the pre-trained language model is the MiniLMv2 model, and the competence vector database is the Faiss vector database.

4. The professional job recommendation system based on a large model and dual-path recall according to claim 1, characterized in that: The calculation process of the capability matching module is as follows: traverse each capability of each job group, query the capability vector database for the K professional capabilities with the highest semantic similarity to the capabilities of that job group, and filter the professional capabilities with similarity reaching a preset threshold; take the professional capability with the highest similarity as the main contributing capability, and the rest that meet the threshold as auxiliary contributing capabilities; the capability matching score is the average of the main contributing capability score and the auxiliary contributing capability score, where the main contributing capability score = similarity × professional capability score × job group capability requirement score, and the auxiliary contributing capability score = similarity² × professional capability score × job group capability requirement score.

5. The professional job recommendation system based on a large model and dual-path recall according to claim 1, characterized in that: In the job matching module, the rows of the collaborative filtering matrix represent the same major from different schools, the columns represent different job groups, and the matrix value is the number of graduates of that major who enter the corresponding job group. The collaborative filtering algorithm identifies other schools with the same major that have similar job destinations to the current major and recommends job groups that similar majors often go to in order to obtain an employment destination score.

6. A method for professional job recommendation clusters based on a large model and dual-path recall, applying the professional job recommendation cluster system based on a large model and dual-path recall as described in any one of claims 1-5, characterized in that, Includes the following steps: S1: Job group capability preprocessing: Extract job group capabilities and corresponding requirement scores from the knowledge graph database, encode them using a pre-trained language model, and store them in the capability vector database to cache job group capability data; S2: Professional competency processing: Receives the professional training program text, calls the large language model to extract professional competencies and quantifies their scores, encodes them and stores them in the competency vector database, and caches the professional competency data. S3: Ability matching calculation, which calculates the semantic similarity between professional ability and job group ability based on the ability vector database, and obtains the ability matching score by combining the ability score; S4: Job matching calculation: Extract graduate employment destination data from knowledge graph database, construct collaborative filtering matrix, and obtain job destination score of job group through collaborative filtering algorithm; S5: Dual-path recall and score merging, weighted merging of ability matching score and employment destination score to obtain comprehensive recommendation score, and filtering a preset number of recommended job groups according to the score; S6: Generate a recommendation report, filter the intermediate calculation data corresponding to the recommended job groups, and call the large language model to process the data and generate the recommendation report.

7. The professional recommendation job group method based on a large model and dual-path recall according to claim 6, characterized in that: In step S1, the pre-trained language model is the MiniLMv2 model, and the capability vector database is the Faiss vector database. Step S1 also includes: querying whether the job group capabilities exist in the metadata of the Faiss vector database, and if not, performing encoding and storage operations.

8. The professional recommendation job group method based on a large model and dual-path recall according to claim 6, characterized in that: In S2, the large language model quantifies professional competence scores based on the curriculum, practical components, and graduation requirements of the professional training program, and normalizes the competence scores to the [0,10] interval using a truncation method.

9. The professional recommendation job group method based on a large model and dual-path recall according to claim 6, characterized in that: In step S3, semantic similarity is obtained through the similarity search function of the capability vector database, and K professional capabilities with similarity reaching a preset threshold are selected. The capability matching score is calculated as follows: main contributing capability score = similarity × professional capability score × job group capability requirement score, auxiliary contributing capability score = similarity² × professional capability score × job group capability requirement score, and the capability matching score is the sum and average of the main contributing capability score and the auxiliary contributing capability score.

10. The professional recommendation job group method based on a large model and dual-path recall according to claim 6, characterized in that: In S6, the intermediate calculation data includes a list of professional abilities and scores, details of ability matching between majors and job groups, employment destination data of job groups, collaborative filtering scores, ability matching scores, employment destination scores, and comprehensive recommendation scores. The recommended report includes an overview of professional capabilities, capability matching analysis, employment matching analysis, and personalized development suggestions.

Images