Occupational skill training personalized recommendation system fused with large language model

Abstract

The application relates to the technical field of large language models, and discloses a professional skill training personalized recommendation system fusing a large language model, which comprises a data acquisition module configured to acquire personal portrait data, historical professional state information and current professional state information of a user; and an LLM data enhancement module configured to receive data collected by the data acquisition module, call a pre-deployed large language model, and generate enhanced training data related to the professional development of the user through a prompt word engineering. The application introduces a vision narrowing inhibition module, calculates the concentration degree of a recommendation list on the distribution of professional categories in real time, and automatically compares the concentration degree with a preset threshold value, so that a reordering or filtering mechanism can be triggered in time when the recommendation result shows a homogenization tendency, thereby effectively blocking the "information cocoon" effect caused by the data enhancement deviation of the large language model, and guaranteeing that the user can obtain a diversified professional development vision in the long term.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, and in particular to a personalized recommendation system for vocational skills training that integrates large language models. Background Technology

[0002] With the acceleration of industrial upgrading and digital transformation, the demand for personalized and intelligent management systems in the vocational skills training field is becoming increasingly urgent. Currently, various system solutions have emerged for talent training management and human resource allocation optimization.

[0003] For example, patent application CN115860997A discloses a talent training management method based on vocational skills; patent application CN119831793A discloses an intelligent human resource allocation optimization system; and patent application CN121582034A discloses an intelligent management system and method for vocational skills training. However, the above-mentioned prior art mainly relies on users' historical behavioral data and preset career path rules when recommending personalized courses or training plans. In recent years, large language models (LLMs), due to their powerful text generation and semantic understanding capabilities, have been attempted to be applied to data augmentation to improve the cold start and sparsity problems of recommendation systems. Specifically, the system can call LLMs to generate simulated learning behavior data or cross-domain skill descriptions based on limited user information, thereby enriching the training samples.

[0004] However, the research revealed a unique technical risk when applying Large Language Models (LLMs) to vocational skills training recommendation: the "narrowing of vision" problem caused by data augmentation bias. Because the training corpus of LLMs contains far more information on popular professions (such as AI engineers, big data developers, and full-stack developers) than long-tail professions (such as ancient book restorers and precision mold fitters), LLMs, when generating augmented training data, may over-generate fake data related to a few popular professions. This skewed data distribution is learned and amplified by the recommendation model, causing the system to repeatedly recommend courses related to popular professions to all users (including those who might actually be suited for long-tail professions). Over time, the recommendation information received by users becomes homogenized, artificially narrowing their career development horizons, violating the original intention of personalized recommendations, and hindering users' long-term career development and the diversified balance of the talent market. Summary of the Invention

[0005] This invention provides a personalized recommendation system for vocational skills training that integrates a large language model to solve existing technical problems, thereby addressing the "narrowing of vision" problem caused by data augmentation bias in existing solutions.

[0006] To address the aforementioned technical problems, according to one aspect of the present invention, more specifically, a personalized recommendation system for vocational skills training integrating a large language model, applied to a pre-defined talent training management platform, comprising: The data acquisition module is configured to collect users' personal profile data, historical career status information, and current career status information; The LLM data augmentation module is configured to receive data collected by the data acquisition module and call a pre-deployed large language model to generate augmented training data related to the user's career development through prompt word engineering. The personalized recommendation module is configured to train a vocational skills recommendation model based on raw collected data and enhanced training data to generate an initial list of training courses recommended for users. The field-of-view narrowing suppression module specifically includes: Step 1: Obtain the initial training course recommendation list generated by the personalized recommendation module; Step 2: Calculate the recommendation diversity feedback index of the initial training course recommendation list based on the terminal algorithm; Step 3: Compare the recommended diversity feedback index with the preset diversity threshold; Step 4: When the recommendation diversity feedback index exceeds the diversity threshold, the recommendation re-evaluation mechanism is triggered to reorder or filter the initial training course recommendation list and generate a target training course recommendation list to replace the initial training course recommendation list and output it to the user. Among them, the recommendation diversity feedback index is used to quantitatively characterize the degree of concentration of the initial training course recommendation list in the distribution of occupational categories; The career path planning module is configured to divide the user's career skill association groups into multiple career clusters based on the user's historical career status information using a graph neural network, and determine the principal components of each career cluster to generate multiple candidate career development paths.

[0007] Furthermore, the terminal algorithm for calculating the recommendation diversity feedback index in the field-of-view narrowing suppression module is as follows: ; In the above formula, This indicates a narrowing of the index; This indicates the total number of occupational categories included in the initial training course recommendation list; Indicates the first The percentage of recommended courses under each occupational category out of the total number of courses in the recommended list.

[0008] Furthermore, the closer the narrowing index is to 1, the more concentrated the occupational distribution of the recommendation list is, and the higher the risk of narrowing the field of vision. When the narrowing index is closer to 0, it indicates that the occupational distribution in the recommendation list is more even.

[0009] Furthermore, the terminal algorithm for calculating the recommendation diversity feedback index in the field-of-view narrowing suppression module is as follows: ; In the above formula, Indicates the diversity of recommendation feedback metrics; This indicates the total number of courses in the initial training course recommendation list; Indicates the first The occupational category to which this course belongs; This represents a predefined set of popular job categories; Indicates the first The ranking weight or confidence score of each course in the recommendation list; The recommendation diversity feedback index is used to quantify the dominance of popular professional courses in the recommendation list.

[0010] Furthermore, the field-of-view narrowing suppression module sends a rescheduling request to the personalized recommendation module, the rescheduling request containing a diversity enhancement parameter.

[0011] Furthermore, in response to the rescheduling request, the personalized recommendation module increases the sampling weight of data from non-popular or long-tail occupational categories in the enhanced training data and decreases the sampling weight of data from popular occupational categories when retraining or inferring the occupational skill recommendation model.

[0012] Furthermore, the LLM data enhancement module specifically comprises: Step 1: Obtain the user's historical career path sequence; Step 2: Construct contextual cue words that include the user's current skill level, career development goals, and popular skill trends in the market; Step 3: Input the contextual prompts into the large language model, instructing the large language model to generate cross-domain professional skill descriptions that are potentially related to the user's current career path as augmented training data.

[0013] Furthermore, the initial list of recommended training courses is matched with multiple candidate career development paths, and general or foundational courses that can cover multiple different candidate paths are given priority.

[0014] Furthermore, after triggering the recommendation re-evaluation mechanism, the field-of-view narrowing suppression module specifically performs the following: Step 1: Record the fingerprint information of this triggered event and the initial training course recommendation list; Step 2: After a preset time period, if the diversity feedback metric for the recommended list generated for the same user still exceeds the threshold, a warning signal will be sent to the platform administrator regarding potential data generation bias in the large language model.

[0015] The personalized recommendation system for vocational skills training that integrates a large language model provided by this invention achieves the following advantages compared to existing technologies: 1. This invention introduces a vision narrowing suppression module to calculate the concentration of the recommended list in terms of occupational category distribution in real time and automatically compare it with a preset threshold. When the recommendation results show a tendency to be homogeneous, it can promptly trigger a re-sorting or filtering mechanism, thereby effectively blocking the "information cocoon" effect caused by the data augmentation bias of the large language model and ensuring that users can obtain a diversified career development vision in the long term.

[0016] 2. This invention designs a narrowing index based on normalized entropy complement, which can accurately measure the central tendency of occupational distribution in the recommendation list in a standardized and quantifiable way. This index is not affected by the length of the recommendation list or the number of occupational categories, making it easy for the system to compare narrowing risks horizontally under different users and application scenarios, and providing a robust mathematical basis for automated suppression decisions.

[0017] 3. By constructing diverse feedback indicators for the dominance of popular career courses and combining them with course ranking weights or confidence scores, this invention can accurately locate situations where popular career courses dominate the recommendation list. This enables the system to quickly identify recommendation biases caused by excessive generation of popular career samples by large language models and then take targeted corrective measures.

[0018] 4. This invention sends a rescheduling request containing diversity enhancement parameters to the personalized recommendation module through the field-of-view narrowing suppression module, and guides it to actively increase the sampling weight of non-popular or long-tail occupational category data during retraining or inference. This achieves closed-loop adaptive control from bias detection to model correction, and can dynamically alleviate the occupational distribution skew caused by data augmentation without manual intervention.

[0019] 5. This invention utilizes a graph neural network through a career path planning module to divide user skills into multiple career clusters and generate multiple candidate development paths. At the same time, it matches the initial recommendation list with these paths, prioritizing the recommendation of general or basic courses that can cover multiple different candidate paths simultaneously. This significantly improves the compatibility of career paths and the flexibility of users' long-term development while maintaining recommendation relevance. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention; Figure 2 This is a diagram showing the relationship between the narrowing index in this invention; Figure 3 This is a flowchart of Embodiment 3 of the present invention. Detailed Implementation

[0021] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1

[0023] like Figure 1 As shown, according to one aspect of the present invention, a personalized recommendation system for vocational skills training integrating a large language model is provided, applied to a pre-set talent training management platform, comprising: The data acquisition module is configured to collect users' personal profile data, historical career status information, and current career status information; The LLM data augmentation module is configured to receive data collected by the data acquisition module and call a pre-deployed large language model to generate augmented training data related to the user's career development through prompt word engineering. The personalized recommendation module is configured to train a vocational skills recommendation model based on raw collected data and enhanced training data to generate an initial list of training courses recommended for users. The field-of-view narrowing suppression module specifically includes: Step 1: Obtain the initial training course recommendation list generated by the personalized recommendation module; Step 2: Calculate the recommendation diversity feedback index of the initial training course recommendation list based on the terminal algorithm; Step 3: Compare the recommended diversity feedback index with the preset diversity threshold; Step 4: When the recommendation diversity feedback index exceeds the diversity threshold, the recommendation re-evaluation mechanism is triggered to reorder or filter the initial training course recommendation list and generate a target training course recommendation list to replace the initial training course recommendation list and output it to the user. Among them, the recommendation diversity feedback index is used to quantitatively characterize the degree of concentration of the initial training course recommendation list in the distribution of occupational categories; The career path planning module is configured to divide the user's career skill association groups into multiple career clusters based on the user's historical career status information using a graph neural network, and determine the principal components of each career cluster to generate multiple candidate career development paths.

[0024] This embodiment proposes a complete architecture for a personalized recommendation system for vocational skills training that integrates Large Language Model (LLM). The system acquires user profiles and career status information through a data acquisition module. An LLM data augmentation module uses cue word engineering to generate cross-domain, forward-looking augmented training data. A personalized recommendation module trains a recommendation model based on the original and augmented data and generates an initial course list. A vision narrowing suppression module calculates the concentration of the recommendation list in the distribution of career categories in real time. When a threshold is exceeded, a re-ranking or filtering mechanism is triggered to generate a diversified target list. Simultaneously, a career path planning module uses a graph neural network to divide user skills into multiple career clusters, generating multiple candidate development paths and matching them with the recommendation results. This system forms a closed-loop collaborative workflow of "data acquisition - LLM augmentation - recommendation generation - diversity assessment - path planning".

[0025] Compared to traditional recommendation systems that rely solely on users' historical behavior, this embodiment is the first to deeply integrate LLM generation capabilities with diversity suppression mechanisms and career path planning, alleviating the problems of "information cocoons" and "narrowed perspectives" from the source. By introducing career clustering and multi-path matching strategies, the system not only recommends courses that users are interested in, but also proactively guides them to explore cross-domain or potential development paths, improving the long-term career adaptability and structured growth value of the recommendations, and significantly enhancing the platform's ability to assist users in career planning.

[0026] Example 2

[0027] like Figure 2 As shown, according to one aspect of the present invention, a personalized recommendation system for vocational skills training integrating a large language model is provided, applied to a pre-set talent training management platform, wherein the terminal algorithm for calculating the recommendation diversity feedback index in the vision narrowing suppression module is as follows: ; In the above formula, This indicates a narrowing of the index; This indicates the total number of occupational categories included in the initial training course recommendation list; Indicates the first The percentage of recommended courses under each occupational category out of the total number of courses in the recommended list.

[0028] In this embodiment, the closer the narrowing index is to 1, the more concentrated the occupational distribution of the recommendation list is, and the higher the risk of narrowing the field of vision; the closer the narrowing index is to 0, the more even the occupational distribution of the recommendation list is.

[0029] The formula above is used to calculate the recommendation diversity feedback index of the recommendation list. Essentially, it is the complement of normalized entropy, used to quantify the concentration of occupational categories in the recommendation list. The detailed derivation is given below: I. Assume the initial training course recommendation list contains a total of Different job categories (e.g., "software development", "marketing", "data analysis", etc.). For the first Each occupational category is defined as having a probability of appearing in the recommendation list as follows: ; In the above formula, Indicates belonging to the first The number of courses in each occupational category; of which: ; In the above formula, This represents the total number of courses in the recommendation list. Clearly, the probability distribution... satisfy ,and .

[0030] II. Shannon entropy is a classic indicator for measuring the uncertainty of a probability distribution, defined as: ; The logarithm (log) is usually taken as the natural logarithm (base e) or a common logarithm, but the base is canceled out during subsequent normalization, which does not affect the result. Entropy The range of values ​​for is: Minimum value The probability of a class is 1 if and only if the probability of the other classes is 0 (completely concentrated distribution); the maximum value is... If and only if all categories are equally distributed, i.e. When it reaches (uniform distribution).

[0031] Third, to map the entropy value to a fixed interval [0, 1] for easier comparison, the entropy is normalized. The maximum possible entropy is used. Remove, and you get the normalized entropy: ; At this point, when the distribution is completely uniform, Therefore When the distribution is completely concentrated, Therefore .

[0032] Normalized entropy It is a diversity indicator: the closer its value is to 1, the more even the distribution and the higher the diversity; the closer its value is to 0, the more concentrated the distribution and the lower the diversity (the higher the risk of narrowing).

[0033] IV. When The closer the value is to 1, the more concentrated the occupational distribution in the recommendation list, and the higher the risk of narrowing the scope; when... "The closer the value is to 0, the more evenly the occupational distribution in the recommendation list is." This indicates that it is actually a diversity indicator in the same direction as the normalized entropy (i.e., close to 0 when evenly distributed, and close to 1 when concentrated). The given formula is: ; This formula calculates the complement of the normalized entropy. Its properties are completely uniformly distributed. ,but Completely centralized distribution ,but .

[0034] Therefore, the formula above actually calculates a concentration index (or narrowing index), and the larger the value, the more concentrated the distribution.

[0035] This indicator has a clear mathematical interpretation and a standardized value range [0,1], facilitating the system to set a uniform diversity threshold for automated judgment. Compared to simply counting the number of popular courses, this indicator reflects the overall uniformity of the distribution and is unaffected by list length and the number of categories, making it suitable for horizontal comparison of recommendation results across different users and time windows. This calculation method provides a robust and interpretable quantitative basis for the narrowing of the field of view suppression module, enabling the recommendation system to automatically identify narrowing risks without relying on manual intervention.

[0036] Example 3

[0037] like Figure 3 As shown, according to one aspect of the present invention, a personalized recommendation system for vocational skills training integrating a large language model is provided, applied to a pre-set talent training management platform, wherein the terminal algorithm for calculating the recommendation diversity feedback index in the vision narrowing suppression module is as follows: ; In the above formula, Indicates the diversity of recommendation feedback metrics; This indicates the total number of courses in the initial training course recommendation list; Indicates the first The occupational category to which this course belongs; This represents a predefined set of popular job categories; Indicates the first The ranking weight or confidence score of each course in the recommendation list; The recommendation diversity feedback index is used to quantify the dominance of popular professional courses in the recommendation list.

[0038] In personalized recommendation systems enhanced by Large Language Models (LLM), if the data augmentation module excessively generates false training samples related to a few popular professions, the recommendation model will continuously output such courses to all users, resulting in "information cocoons" or "narrowing of vision." To automatically identify and quantify this narrowing risk, a calculable recommendation diversity feedback index needs to be constructed, which directly reflects the dominance of popular profession courses in the recommendation list. Therefore, the derivation of the above formula is as follows: I. Suppose the initial training course recommendation list generated by the system for a user is as follows: ; in, This represents the total number of courses in the recommendation list (M>1); for the j-th course in the list... Define its occupational category as (For example, "administration and organization management", "artistic creation", "medical care" etc.).

[0039] Predefine a set of popular job categories This set can be obtained by statistically analyzing the most frequently occurring data in historical global recommendation data. Each job category can be obtained, or manually configured by the platform administrator based on market popularity.

[0040] II. To determine whether a single course belongs to a popular career category, an indicator function is introduced: ; This function only contributes non-zero values ​​to courses in popular career categories. Considering that courses ranked higher in the recommendation list have a greater impact on the user's view (users typically pay more attention to the first few recommendations), each course needs to be assigned a weight. This is to reflect its importance in the recommendation process. Weighting It is usually defined as: Inverse sorting weight: ; in, This indicates the position of the course in the list (starting from 1); The model confidence score is obtained by directly using the confidence score (e.g., softmax probability value) output by the recommendation model for the course. This formula adopts a general form. The specific calculation method is not limited, but it is required that... Furthermore, it is positively correlated with the importance of the course (the higher the ranking or the higher the confidence level, the better). The larger (the larger).

[0041] 3. Perform a weighted sum of all courses, but only include courses belonging to popular career categories, to obtain the numerator: ; This value represents the total weight of all popular career courses in the recommended list. The numerator is larger when the list includes more popular courses and those courses are ranked higher.

[0042] IV. To eliminate list length To account for the influence of absolute weight values, the numerator needs to be divided by the sum of the weights of the entire recommendation list: ; The denominator represents the sum of the weights of all recommended courses.

[0043] V. Define the recommendation diversity feedback index as follows: ; In the above formula, Indicates the diversity of recommendation feedback metrics; This indicates the total number of courses in the initial training course recommendation list; Indicates the first The occupational category to which this course belongs; This represents a predefined set of popular job categories; Indicates the first The ranking weight or confidence score of a course in the recommendation list.

[0044] This indicates that none of the courses in the recommended list belong to popular career categories (all are long-tail or emerging careers), thus broadening the scope and eliminating the risk of narrowing it. This indicates that all courses in the recommended list belong to popular career categories, and non-popular courses have a weight of 0 (i.e., the list is completely dominated by popular courses), which poses a serious risk of narrowing the scope of the recommendations.

[0045] This metric directly addresses the bias issue that may arise from the LLM data augmentation module overgenerating samples related to popular professions. It can quickly identify whether the recommendation list is dominated by a few popular profession categories. By introducing ranking weights, the metric gives higher influence to top-ranked popular courses, which is more in line with actual user attention behavior. When the metric exceeds a threshold, the system can accurately pinpoint the specific cause of the narrowing, namely "popular profession bias," facilitating subsequent targeted sampling weight adjustments or rescheduling strategies. This effectively suppresses the tendency of the recommendation system to repeatedly push homogeneous popular content to users.

[0046] Example 4

[0047] like Figure 1As shown, according to one aspect of the present invention, a personalized recommendation system for vocational skills training that integrates a large language model is provided and applied to a preset talent training management platform, wherein a vision narrowing suppression module sends a rescheduling request to a personalized recommendation module, the rescheduling request including a diversity enhancement parameter.

[0048] In this embodiment, in response to a rescheduling request, the personalized recommendation module increases the sampling weight of data from non-popular or long-tail occupational categories in the enhanced training data and decreases the sampling weight of data from popular occupational categories when retraining or inferring the occupational skills recommendation model.

[0049] In this embodiment, the LLM data enhancement module specifically comprises: Step 1: Obtain the user's historical career path sequence; Step 2: Construct contextual cue words that include the user's current skill level, career development goals, and popular skill trends in the market; Step 3: Input the contextual prompts into the large language model, instructing the large language model to generate cross-domain professional skill descriptions that are potentially related to the user's current career path as augmented training data.

[0050] In this embodiment, the initial training course recommendation list is matched with multiple candidate career development paths, and general or basic courses that can cover multiple different candidate paths are given priority.

[0051] In this embodiment, the field-of-view narrowing suppression module, after triggering the recommendation re-evaluation mechanism, specifically does the following: Step 1: Record the fingerprint information of this triggered event and the initial training course recommendation list; Step 2: After a preset time period, if the diversity feedback metric for the recommended list generated for the same user still exceeds the threshold, a warning signal will be sent to the platform administrator regarding potential data generation bias in the large language model.

[0052] When the field-of-view suppression module detects that the recommendation diversity index exceeds a threshold, it sends a rescheduling request containing diversity enhancement parameters to the personalized recommendation module. During model retraining or inference, the personalized recommendation module increases the sampling weights of data from less popular or long-tail occupational categories in the enhanced training data and decreases the sampling weights of popular occupational categories. Simultaneously, the LLM data augmentation module guides the generation of cross-domain skill descriptions by constructing prompts containing the user's current skills, development goals, and market trends. The system also matches the recommendation list with multiple candidate paths generated by the career path planning module, prioritizing general or foundational courses. Furthermore, the system records trigger events and continuously monitors them; if thresholds are exceeded multiple times, an LLM data deviation warning is issued to the administrator.

[0053] This embodiment implements a complete closed-loop adaptive mechanism from "diversity detection" to "model retraining" and then to "data augmentation correction," dynamically suppressing the narrowing of the recommendation field without requiring manual redesign of the recommendation strategy. By prioritizing recommendations of general courses covering multiple career paths, the system can significantly improve path compatibility while maintaining recommendation relevance. The long-term early warning mechanism provides system-level monitoring capabilities for LLM generation biases, helping the platform to promptly identify and correct systematic biases generated by the large language model during the data augmentation stage, ensuring the long-term robustness and fairness of the recommendation system.

[0054] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A professional skill training personalized recommendation system fused with a large language model, applied to a preset talent training management platform, characterized in that, include: The data acquisition module is configured to collect users' personal profile data, historical career status information, and current career status information; The LLM data augmentation module is configured to receive data collected by the data acquisition module and call a pre-deployed large language model to generate augmented training data related to the user's career development through prompt word engineering. The personalized recommendation module is configured to train a vocational skills recommendation model based on raw collected data and enhanced training data to generate an initial list of training courses recommended for users. The field-of-view narrowing suppression module specifically includes: Step 1: Obtain the initial training course recommendation list generated by the personalized recommendation module; Step 2: Calculate the recommendation diversity feedback index of the initial training course recommendation list based on the terminal algorithm; Step 3: Compare the recommended diversity feedback index with the preset diversity threshold; Step 4: When the recommendation diversity feedback index exceeds the diversity threshold, the recommendation re-evaluation mechanism is triggered to reorder or filter the initial training course recommendation list and generate a target training course recommendation list to replace the initial training course recommendation list and output it to the user. Among them, the recommendation diversity feedback index is used to quantitatively characterize the degree of concentration of the initial training course recommendation list in the distribution of occupational categories; The career path planning module is configured to divide the user's career skill association groups into multiple career clusters based on the user's historical career status information using a graph neural network, and determine the principal components of each career cluster to generate multiple candidate career development paths.

2. The personalized recommendation system for vocational skills training integrating a large language model as described in claim 1, characterized in that: The terminal algorithm for calculating the recommendation diversity feedback index in the field-of-view narrowing suppression module is as follows: ； In the above formula, This indicates a narrowing of the index; This indicates the total number of occupational categories included in the initial training course recommendation list; Indicates the first The percentage of recommended courses under each occupational category out of the total number of courses in the recommended list.

3. The personalized recommendation system for vocational skills training integrating a large language model as described in claim 2, characterized in that: When the narrowing index is closer to 1, it means that the occupational distribution of the recommendation list is more concentrated and the risk of narrowing the field of vision is higher. When the narrowing index is closer to 0, it indicates that the occupational distribution in the recommendation list is more even.

4. The personalized recommendation system for vocational skills training integrating a large language model according to claim 1, characterized in that: The terminal algorithm for calculating the recommendation diversity feedback index in the field-of-view narrowing suppression module is as follows: ； In the above formula, Indicates the diversity of recommendation feedback metrics; This indicates the total number of courses in the initial training course recommendation list; Indicates the first The occupational category to which this course belongs; This represents a predefined set of popular job categories; Indicates the first The ranking weight or confidence score of each course in the recommendation list; The recommendation diversity feedback index is used to quantify the dominance of popular professional courses in the recommendation list.

5. The personalized recommendation system for vocational skills training integrating a large language model according to claim 1, characterized in that: The field-of-view narrowing suppression module sends a rescheduling request to the personalized recommendation module, and the rescheduling request includes a diversity enhancement parameter.

6. The personalized recommendation system for vocational skills training integrating a large language model according to claim 1, characterized in that: In response to the rescheduling request, the personalized recommendation module increases the sampling weight of data from non-popular or long-tail occupational categories in the enhanced training data and decreases the sampling weight of data from popular occupational categories when retraining or inferring the occupational skill recommendation model.

7. The personalized recommendation system for vocational skills training integrating a large language model according to claim 1, characterized in that: The LLM data enhancement module specifically includes: Step 1: Obtain the user's historical career path sequence; Step 2: Construct contextual cue words that include the user's current skill level, career development goals, and popular skill trends in the market; Step 3: Input the contextual prompts into the large language model, instructing the large language model to generate cross-domain professional skill descriptions that are potentially related to the user's current career path as augmented training data.

8. The personalized recommendation system for vocational skills training integrating a large language model according to claim 1, characterized in that: The initial list of recommended training courses is matched with multiple candidate career development paths, and general or foundational courses that can cover multiple different candidate paths are given priority.

9. The personalized recommendation system for vocational skills training integrating a large language model according to claim 1, characterized in that: The field-of-view narrowing suppression module, after triggering the recommendation re-evaluation mechanism, specifically does the following: Step 1: Record the fingerprint information of this triggered event and the initial training course recommendation list; Step 2: After a preset time period, if the diversity feedback metric for the recommended list generated for the same user still exceeds the threshold, a warning signal will be sent to the platform administrator regarding potential data generation bias in the large language model.

Images