A talent recommendation system based on matching local productivity with vocational education

Abstract

The application discloses a talent recommendation system based on matching local productivity by vocational education, relates to the technical field of talent recommendation, and comprises a dynamic industry portrait construction module, an expert audit correction module, a personalized path planning module, a talent growth prediction module, a path adjustment module, a talent recommendation module, a data storage module, an interactive interface module and a system management module; the dynamic industry portrait construction module predicts industry trend data through a time sequence hidden Markov model with an attention mechanism, and the expert audit correction module supplements unannounced project information based on a knowledge graph completion algorithm and Bayesian inference, and generates a dynamic industry knowledge graph based on a time decay factor through a graph convolution network feature fusion technology. The dynamic industry portrait construction module and the expert audit correction module can solve the problem that industry information lags behind in the existing talent recommendation system and cannot reflect local industrial dynamic changes in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of talent recommendation, and particularly relates to a talent recommendation system based on matching local productivity by vocational education. BACKGROUND

[0002] The talent recommendation system based on matching local productivity by vocational education is an intelligent platform that realizes accurate matching of talent skills and industrial demand by integrating vocational education talent training data and local industrial post demand by using big data and artificial intelligence technology. The talent recommendation system based on matching local productivity by vocational education helps students to efficiently find jobs, helps enterprises to reduce recruitment costs, promotes optimization of professional settings of vocational education, and promotes upgrading of regional productivity layout. The talent recommendation system based on matching local productivity by vocational education breaks the information barrier between talent supply and demand, builds a virtuous cycle of talent, industry and economy, and enables regional economic sustainable development by vocational education resources.

[0003] At present, in the aspect of industry information acquisition and update, the existing talent recommendation system uses a relatively simple model when predicting industry trends, which cannot fully consider the dynamics and complexity of industry development, resulting in that the prediction of industry trend data lags behind the actual changes, and it is difficult to adapt to the rapidly changing industrial environment. In addition, the existing talent recommendation system lacks the ability to acquire and integrate non-public project information in the industry, and cannot comprehensively and deeply build an industry portrait, so that the talent recommendation lacks sufficient information support and cannot accurately match the actual needs of local industry. On the other hand, in the aspect of career development path planning, only fixed industry standards or limited historical data are used, and the speed of industry technology update and individual differences are ignored, so that the planned career path lacks flexibility and pertinence and cannot meet the development needs of talents at different stages. In the matching link of talents and posts, the matching algorithm used is relatively single, and it is difficult to deeply understand the internal relationship between post skill demand and talent growth potential, resulting in low matching accuracy, a large number of suitable talents and posts missing each other, and waste of human resources and increase of enterprise recruitment costs.

[0004] Therefore, the talent recommendation system based on matching local productivity by vocational education is proposed to solve the above problems. SUMMARY

[0005] The main purpose of the present application is to provide a talent recommendation system based on matching local productivity by vocational education to solve the problems raised in the background.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is as follows: a talent recommendation system based on matching local productivity by vocational education, which comprises a dynamic industry portrait construction module, an expert audit correction module, an individualized path planning module, a talent growth prediction module, a path adjustment module, a talent recommendation module, a data storage module, an interactive interface module and a system management module.

[0007] The dynamic industry portrait construction module predicts industry trend data through a time sequence hidden Markov model with an attention mechanism, and supplements the undisclosed project information based on a knowledge graph completion algorithm and Bayesian inference, generates a dynamic industry knowledge graph based on a time decay factor through a graph convolution network feature fusion technology, and the node update frequency is positively correlated with the feature importance index;

[0008] The personalized path planning module generates a multi-dimensional career development path based on an industry knowledge graph and a multi-agent reinforcement learning hierarchical planning algorithm;

[0009] The talent growth prediction module dynamically predicts talent growth probability using an adaptive weight Transformer model combined with an industry technology iteration index;

[0010] The path adjustment module dynamically adjusts the career development path through a hierarchical multi-objective optimization algorithm combined with a multi-objective optimization game theory conflict resolution mechanism based on Nash equilibrium;

[0011] The talent recommendation module realizes dynamic semantic matching of talents and posts based on a bidirectional graph neural network and a multi-round two-sided matching game model based on a graph attention mechanism;

[0012] The data storage module stores multi-modal data during system operation using a distributed graph database combined with time series data sharding technology;

[0013] The interactive interface module provides an AR visual career path sand table and a multi-modal intelligent question and answer assistant;

[0014] The system management module integrates a regional industry talent supply and demand imbalance risk early warning function based on Kalman filtering and a dynamic permission management system based on attribute-based encryption.

[0015] Preferably, the dynamic industry portrait construction module includes a data acquisition sub-module, a data analysis sub-module, and a portrait update sub-module;

[0016] The data acquisition sub-module uses an intelligent crawler technology based on reinforcement learning combined with a multi-agent collaboration mechanism to real-time capture multi-source data of local industry news, policy documents, and enterprise recruitment information, and filters duplicate data through an adaptive threshold;

[0017] The data analysis sub-module uses a T5-large text generation model combined with a LightGBM ensemble learning algorithm to extract industry key skills, job demand trends, enterprise expansion and contraction dynamic information;

[0018] The profile update submodule uses a temporal hidden Markov model with an attention mechanism to predict industry trends and dynamically adjusts the update strategy by introducing a feature importance formula with a time decay factor.

[0019] The data feature set is F = {f1, f2, ..., f...} n}, where n is the feature dimension, and the formula is as follows:

[0020]

[0021] Among them, f i Let I(f) be the i-th feature in the data feature set F, and t be the current time. i ,t) is a feature f i The importance of the decision tree at the current time t, where m is the number of decision trees, and T is the value of the decision tree at the current time t. j For the j-th decision tree, Gain(f) i ,T j ) represents the feature f in the j-th decision tree. i The information gain brought about, α is the attenuation coefficient (0 < α < 1), t j Let be the construction time of the j-th decision tree.

[0022] Preferably, the expert review and correction module includes an expert team unit and an information supplementation unit;

[0023] The expert team consists of enterprise experts with senior technical titles or more than 10 years of industry experience in the local pillar industries. They use the federated learning collaborative review platform and blockchain evidence storage technology to achieve distributed review of industry profiles.

[0024] The information supplementation unit uses a knowledge graph completion algorithm integrated with Bayesian reasoning to structurally supplement the industry profile with undisclosed project talent demand information provided by experts, and quantifies the credibility of the information through an evidence weight evaluation mechanism.

[0025] Preferably, the personalized route planning module includes a personal information analysis submodule and a route planning submodule;

[0026] The personal information analysis submodule integrates occupational assessment questionnaires, learning behavior logs, and online skills test data through multimodal behavior analysis technology. It combines a principal component analysis model optimized by transfer learning with the results of regional industrial policy text mining to achieve dynamic matching analysis between personal skills and industry trends.

[0027] The path planning submodule utilizes a multi-agent reinforcement learning-based hierarchical planning algorithm to plan short-term, medium-term and long-term career development paths based on the results of the personal information analysis submodule. The time intervals for each time interval are dynamically adjusted according to the half-life of the industry technical knowledge update cycle, which is defined as the time required for 50% of existing technical knowledge in a certain field to be replaced by new technology.

[0028] Preferably, the time interval of the short-term career development path is 0.5-2 years, the time interval of the medium-term career development path is 2-4 years, and the time interval of the long-term career development path is more than 4 years.

[0029] The path planning submodule dynamically adjusts the time interval according to the half-life of the industry technical knowledge update cycle, which is defined as the time required for 50% of existing technical knowledge in a certain field to be replaced by new technology.

[0030] Short-term path: 0.5≤t s ≤min(2,τ);

[0031] Medium-term path: max(2,τ)≤t m ≤min(4,2τ);

[0032] Long-term path: max(4,2τ)≤t l ;

[0033] Where τ is the half-life of the industry technical knowledge update cycle, defined as the time required for 50% of existing technical knowledge in a certain field to be replaced by new technology, t s , t m and t l are the upper limits of the time intervals of the short-term, medium-term and long-term paths, respectively.

[0034] Preferably, the talent growth prediction module includes a data processing unit and a prediction analysis unit.

[0035] The data processing unit collects historical talent development data and industry dynamic change data, and uses generative adversarial networks combined with adversarial training for data augmentation.

[0036] The prediction analysis unit adapts the Transformer model with adaptive weights, dynamically adjusts the weights of each submodel through reinforcement learning, introduces the industry technology iteration index T(t) as an external feature, and constructs a dynamic time series prediction model:

[0037]

[0038] Where P(t) is the final prediction result at time t, k is the number of submodels, i.e. the total number of submodels integrated in the model, ω i (t) is the weight of the i-th submodel at time t, ε is the industry fluctuation coefficient, P i (t) is the prediction result of the i-th sub-model at time t.

[0039] Preferably, the path adjustment module includes a cause analysis submodule and a scheme adjustment submodule;

[0040] The cause analysis submodule uses the causal forest algorithm combined with bias analysis to introduce the counterfactual reasoning mechanism to identify the reasons for slow skill learning progress and sudden changes in industry demand when the talent fails to achieve the growth target on time.

[0041] The scheme adjustment submodule uses a hierarchical multi-objective optimization algorithm to set dynamic weights γ i (t) for short-term target skill mastery O1, medium-term target job matching degree O2, and long-term target industry contribution value O3, respectively, and constructs a comprehensive evaluation function based on a priority decision model of game theory:

[0042]

[0043] Where γ is the conflict adjustment factor, γ i (t) is the dynamic weight of the i-th target at time t, C j is the j-th target conflict coefficient, and the weight γ i (t) is dynamically updated by the following formula:

[0044]

[0045] Where η is the learning rate, V(π t ) is the value function of strategy π t , represents the partial derivative of O i , is the gradient of the value function V(π t ) of strategy π t with respect to target O i , γ i (t+1) is the weight of the i-th target at time t+1, γ i (t) is the weight of the i-th target at time t.

[0046] Preferably, the talent recommendation module includes a matching submodule and a recommendation execution submodule;

[0047] The matching submodule uses a bidirectional graph neural network based on a graph attention mechanism to model the semantic association between job demand and talent growth path, introduces a dynamic matching degree update mechanism driven by reinforcement learning, and calculates the bidirectional matching probability of the job and the talent.

[0048] The recommendation execution sub-module optimizes the recommendation strategy based on a multi-round bilateral matching game model and combined enterprise feedback data to form a recommendation, feedback and iteration closed loop.

[0049] Preferably, the data storage module adopts a distributed graph database combined with time series data sharding technology to partition and store multi-modal data, and realizes cross-institution data sharing and privacy protection through a federated learning data security aggregation mechanism based on homomorphic encryption.

[0050] Preferably, the interaction interface module provides an AR visual career path sand table, supports multi-modal interaction input methods such as voice, gestures and text, and integrates an intelligent question and answer assistant based on dialogue-enhanced Transformer to answer path planning questions in real time.

[0051] The system management module integrates a regional industry talent supply and demand imbalance risk early warning function based on Kalman filtering, identifies industry talent supply and demand imbalance risks based on an anomaly detection algorithm, and simultaneously adopts an attribute-based encryption-based dynamic permission management system to realize data access control for different roles.

[0052] The present application has the following advantages:

[0053] 1. The present application introduces a time series hidden Markov model with attention mechanism through a dynamic industry portrait construction module to predict industry trend data, and combines an expert review correction module based on a knowledge graph completion algorithm and Bayesian inference to supplement undisclosed project information, generates a dynamic industry knowledge graph based on a time decay factor through graph convolution network feature fusion technology, the node update frequency is positively correlated with the feature importance index, realizes the dynamic update and accurate construction of the industry knowledge graph. Compared with the prior art, the timeliness and accuracy of industry trend prediction can be improved, so that the problem of industry information lag in existing talent recommendation systems, which cannot reflect the dynamic changes of local industry in real time, can be solved.

[0054] 2. The present application adopts a hierarchical planning algorithm based on multi-agent reinforcement learning through a personalized path planning module, generates a multi-dimensional career development path combined with an industry knowledge graph, and dynamically adjusts the time interval of short-term, medium-term and long-term paths according to the half-life period of industry technical knowledge update cycle, realizes the dynamic adaptive planning of career development path. Compared with the prior art, the adaptability of the career path planning to the industry technical iteration can be improved, so that the problem that the traditional path planning model cannot be flexibly adjusted according to the technical update speed, resulting in the disconnection between talent skills and industry demand, can be solved.

[0055] 3.The talent recommendation system based on vocational education matching local productivity according to the present application, through the talent recommendation module, the semantic association of the post demand and the talent growth path is modeled based on the bidirectional graph neural network of the graph attention mechanism, and the dynamic semantic matching is realized by combining the multi-round bilateral matching game model, and the recommendation strategy is optimized through the enterprise feedback data to form a closed loop, so that the precise dynamic matching of talents and posts is realized. Compared with the prior art, the semantic understanding precision and matching efficiency of talent recommendation can be improved, so as to solve the problem that the traditional recommendation algorithm can only be based on keyword matching and cannot deeply understand the post skill demand and the talent growth potential. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 It is the overall system architecture schematic diagram of the present application;

[0057] Figure 2 It is the architecture schematic diagram of the dynamic industry portrait construction module of the present application;

[0058] Figure 3 It is the architecture schematic diagram of the expert audit correction module of the present application;

[0059] Figure 4 It is the architecture schematic diagram of the individualized path planning module of the present application;

[0060] Figure 5 It is the architecture schematic diagram of the talent growth prediction module of the present application;

[0061] Figure 6 It is the architecture schematic diagram of the path adjustment module of the present application;

[0062] Figure 7 It is the architecture schematic diagram of the talent recommendation module of the present application. DETAILED DESCRIPTION

[0063] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application will be further described below in combination with specific embodiments.

[0064] Embodiment one, please refer to Figures 1 to 2 The talent recommendation system based on vocational education matching local productivity includes a dynamic industry portrait construction module, an expert audit correction module, an individualized path planning module, a talent growth prediction module, a path adjustment module, a talent recommendation module, a data storage module, an interactive interface module and a system management module.

[0065] The dynamic industry portrait construction module predicts industry trend data through a time sequence hidden Markov model with an attention mechanism, and supplements the undisclosed project information based on a knowledge graph completion algorithm and Bayesian inference, generates a dynamic industry knowledge graph based on a time decay factor through graph convolution network feature fusion technology, and the node update frequency is positively correlated with the feature importance index;

[0066] The personalized path planning module generates a multi-dimensional career development path based on the industry knowledge graph through a multi-agent reinforcement learning hierarchical planning algorithm;

[0067] The talent growth prediction module dynamically predicts talent growth probability by using an adaptive weight Transformer model combined with an industry technology iteration index;

[0068] The path adjustment module dynamically adjusts the career development path through a hierarchical multi-objective optimization algorithm combined with a multi-objective optimization game theory conflict resolution mechanism based on Nash equilibrium;

[0069] The talent recommendation module realizes dynamic semantic matching of talents and posts based on a bidirectional graph neural network and a multi-round two-sided matching game model based on a graph attention mechanism;

[0070] The data storage module stores multi-modal data in the system operation process by using a distributed graph database combined with time sequence data sharding technology;

[0071] The interactive interface module provides an AR visual career path sand table and a multi-modal intelligent question and answer assistant;

[0072] The system management module integrates a regional industry talent supply and demand imbalance risk early warning function based on Kalman filtering and a dynamic permission management system based on attribute-based encryption.

[0073] The dynamic industry portrait construction module includes a data acquisition submodule, a data analysis submodule, and a portrait update submodule;

[0074] The data acquisition submodule uses intelligent crawler technology based on reinforcement learning combined with a multi-agent collaboration mechanism to real-time capture multi-source data of local industry news, policy documents, and enterprise recruitment information, and filters duplicate data through adaptive threshold;

[0075] The data analysis submodule uses a T5-large text generation model combined with a LightGBM ensemble learning algorithm to extract industry key skills, job demand trends, enterprise expansion and contraction dynamic information;

[0076] The portrait update submodule predicts industry trend changes by using a time sequence hidden Markov model with an attention mechanism, and dynamically adjusts the update strategy through a feature importance formula with a time decay factor;

[0077] The data feature set is F = {f1, f2, …, f n}, where n is the feature dimension, and the formula is as follows:

[0078]

[0079] where f i is the i-th feature in the data feature set F, t is the current time, I(f i , t) is the importance of the feature f i at the current time t, m is the number of decision trees, T j is the j-th decision tree, Gain(f i , T j ) represents the information gain brought by the feature f i in the j-th decision tree, a is a decay coefficient (0 < a < 1), and t j is the construction time of the j-th decision tree.

[0080] Further, the talent recommendation system based on matching the vocational education to the local productivity covers nine core modules, including a dynamic industry portrait construction module, an expert review and correction module, a personalized path planning module, a talent growth prediction module, a path adjustment module, a talent recommendation module, a data storage module, an interactive interface module, and a system management module. The modules work in cooperation, realize the whole-process function from industry trend analysis to talent development planning and accurate job recommendation through the front-end algorithm and technology, and adapt to the demand of local productivity for vocational education talents.

[0081] The data acquisition sub-module adopts an intelligent crawler technology based on reinforcement learning. The technology constructs a deep Q network as a core decision model. In the data acquisition process, the intelligent crawler takes the state of each data capture as input, including the type of collected data, website access state, etc. The DQN network outputs actions to select the next data source or page to be captured. The intelligent crawler will receive reward feedback based on the collection results. Positive rewards are given for successfully capturing valid data, and negative rewards are given for capturing repeated or invalid data. Through the continuous trial, feedback, and learning process, the data acquisition strategy is optimized.

[0082] In combination with the multi-agent collaboration mechanism, multiple data acquisition agents are deployed at different network nodes. Each agent has an independent DQN model, and they work in parallel to capture local industry news, policy documents, and enterprise recruitment information in real time. The collected data are filtered by dynamically calculating the cosine similarity and other indicators to adapt to the threshold. Data with a similarity higher than the threshold are identified as repeated data and removed to ensure the uniqueness and effectiveness of the collected data.

[0083] The data analysis submodule innovatively combines the T5-large text generation model with the LightGBM ensemble learning algorithm. The T5-large text generation model is based on the Transformer architecture and has strong semantic understanding capabilities. When processing the collected text data, the T5-large model first performs word segmentation and encoding on the text, captures semantic information in the text through multiple attention mechanisms, and extracts information such as industry key skills or job requirements contained in the text.

[0084] The LightGBM ensemble learning algorithm uses a gradient boosting framework to build a feature selection model by iteratively training multiple weak learners. It uses optimization techniques such as histogram algorithms to filter and strengthen the features extracted by the T5-large model, improving the accuracy and stability of feature extraction. Through the collaborative work of these two technologies, the system can accurately extract industry key skill requirements, job demand trends, and dynamic information about enterprise expansion or contraction, providing key data support for subsequent industry portrait construction.

[0085] The portrait update submodule uses a time-series hidden Markov model with an attention mechanism to predict industry trends. Traditional time-series hidden Markov models have difficulty distinguishing the influence of different data on industry trends when processing large amounts of data. The attention mechanism introduced in this system calculates the attention weight of each time step data, focusing on key data and time nodes that have a greater impact on industry trends. Specifically, during model training, the system calculates the correlation between data and industry trend changes, assigning different attention weights to each data point. The higher the weight, the greater the proportion of the data in the model prediction, thereby improving the accuracy of the prediction.

[0086] Combining the feature importance formula with a time decay factor The system dynamically adjusts the update strategy of the industry portrait. Over time, the importance of features changes according to factors such as the decay coefficient α and the information gain of the decision tree. When the importance index of a certain feature reaches the system's preset threshold or the manually set threshold, the portrait update submodule will prioritize updating the industry portrait content related to that feature, ensuring that the industry portrait can reflect the latest industry dynamics in a timely manner.

[0087] Through the above technical solutions, in actual application tests, the system's prediction delay for industry trends has been reduced from 72 hours to 20 hours, and it can identify more than 85% of emerging job requirements in advance. The completeness of industry portrait information has been improved from 65% to 92%, providing more comprehensive and accurate industry information support for talent recommendation.

[0088] Embodiment Two, please refer to Figure 3 Based on the basis of Embodiment One, the expert review and correction module includes an expert team unit and an information supplement unit.

[0089] The expert team unit is composed of enterprise experts with senior technical titles or more than 10 years of industry experience in local pillar industries. Through the federated learning collaborative review platform, the distributed review of industry portraits is realized based on blockchain storage technology;

[0090] The information supplement unit uses knowledge graph completion algorithm and Bayesian inference to supplement the unpublicized project talent demand information provided by experts to the industry portrait in a structured manner, and quantifies the information credibility through the evidence weight evaluation mechanism.

[0091] Further, the expert team works through the federated learning collaborative review platform. The core advantage of the federated learning mechanism is that it can realize cross-enterprise and cross-institutional collaborative review of industry portraits without revealing the original data.

[0092] Specifically, each expert retains the original data locally and only uploads encrypted model parameter update information. The platform aggregates these parameters to iteratively optimize the industry portrait model, avoiding data leakage risks while protecting enterprise trade secrets. In addition, based on blockchain storage technology, every operation in the review process, including data upload, model parameter update, and review opinion recording, is recorded in the form of a hash value on the blockchain. This storage method makes the review process tamper-proof, and once recorded, it cannot be modified. At the same time, it can be traced back, and the full picture of the review process can be queried at any time, greatly improving the credibility and authority of the industry portrait.

[0093] The information supplement unit uses knowledge graph completion algorithm and Bayesian inference. The knowledge graph completion algorithm uses the relationships between nodes in the constructed industry knowledge graph to learn vector representations of nodes and relationships through deep learning models, and then infers missing information. Taking the talent demand prediction of a new industry as an example, the algorithm can predict the existing but not yet clear job skill combination based on the existing relationships between enterprises, jobs, and skills. Bayesian inference then dynamically updates the confidence in the information based on prior knowledge and newly acquired evidence.

[0094] When experts provide undisclosed project talent demand information, Bayesian inference combines information from similar past projects as prior knowledge to evaluate new information and update its credibility. To further ensure the reliability of the supplementary information, an evidence weight evaluation mechanism is introduced, which quantitatively evaluates the expert authority, the consistency of information and knowledge graph, and the credibility of information sources. Expert authority includes title, experience, and field relevance, and the credibility of information sources includes enterprise level and data timeliness. Each dimension is given different weights, and the final credibility score is calculated by weighted calculation. Only information that meets a certain credibility threshold will be structured and supplemented to the industry portrait, thereby achieving precise improvement of the content of the industry portrait. The credibility threshold can be automatically calculated by the system based on historical data or dynamically adjusted by the expert committee to ensure the overall consistency of the supplementary information and the industry portrait.

[0095] Embodiment three, please refer to Figure 4 Based on the basis of embodiment one and embodiment two, the personalized path planning module includes a personal information analysis submodule and a path planning submodule.

[0096] The personal information analysis submodule integrates career assessment questionnaires, learning behavior logs, and online skill test data through multi-modal behavior analysis technology, combines a principal component analysis model optimized by transfer learning, and introduces regional industry policy text mining results to achieve dynamic matching analysis of personal skills and industry trends.

[0097] The path planning submodule uses a hierarchical planning algorithm based on multi-agent reinforcement learning to plan short-term, medium-term, and long-term career development paths based on the results of the personal information analysis submodule. The time intervals for each time interval are dynamically adjusted based on the half-life of the industry technical knowledge update cycle, which is defined as the time required for 50% of existing technical knowledge in a certain field to be replaced by new technology.

[0098] The time interval for short-term career development paths is 0.5-2 years, the time interval for medium-term career development paths is 2-4 years, and the time interval for long-term career development paths is more than 4 years.

[0099] The path planning submodule dynamically adjusts the time interval based on the half-life of the industry technical knowledge update cycle, which is defined as the time required for 50% of existing technical knowledge in a certain field to be replaced by new technology.

[0100] Short-term path: 0.5≤t s ≤min(2,τ);

[0101] Medium-term path: max(2,τ)≤t m ≤min(4,2τ);

[0102] Long-term path: max(4,2τ)≤tl ;

[0103] where τ is the half-life of the industry technology knowledge update cycle, defined as the time required for 50% of existing technology knowledge in a certain field to be replaced by new technology, t s , t m , and t l are the upper limits of the time intervals for short-term, medium-term, and long-term paths, respectively.

[0104] Further, the personalized path planning module generates multi-dimensional career development paths based on a multi-agent reinforcement learning hierarchical planning algorithm combined with an industry knowledge graph. The multi-agent reinforcement learning algorithm uses a centralized training-distributed execution architecture, with three agents responsible for short-term, medium-term, and long-term path planning. Each agent consists of a policy network with three fully connected layers and a value network with two fully connected layers, with full connection between nodes. The hierarchical planning algorithm decomposes the career development path planning problem into strategic, tactical, and operational layers, and optimizes the path through inter-layer information exchange.

[0105] The personal information analysis sub-module uses multi-modal behavior analysis technology to break down data barriers and deeply integrate career assessment questionnaires, learning behavior logs, and online skill test data. These data sources are diverse and cover a wide range of information, including subjective self-evaluation, objective learning behavior records, and actual skill level tests, providing a rich data foundation for a comprehensive understanding of individual capabilities.

[0106] At the data processing level, a principal component analysis (PCA) model optimized by transfer learning is introduced. Traditional PCA models are prone to increased computational complexity due to excessive data features when processing high-dimensional data, and it is difficult to effectively extract key features. However, by using transfer learning, the system can borrow model parameters from related fields that have been trained, and transfer them to the current data processing task, significantly improving the model's adaptability to new data. On this basis, dimensionality reduction processing can quickly and accurately extract features that are crucial to career development path planning from massive amounts of data. For example, in the career assessment questionnaire, personal career interest preference features are extracted, learning efficiency, knowledge mastery progress, and other features are mined from learning behavior logs, and professional skill level features are obtained from online skill test data.

[0107] Moreover, the sub-module also innovatively introduces regional industrial policy text mining results. Through natural language processing (NLP) technology, local industrial policy documents are deeply analyzed to mine key information such as industrial development direction, key support areas, and skill demand trends. Dynamic matching analysis of these industrial policy information and personal skill data can reflect the degree of fit between personal skills and industry trends in real time. For example, when the local industrial policy strongly supports the development of the new energy vehicle industry, the system will analyze the individual's skill mastery in the new energy vehicle-related field and determine whether it meets the needs of industry development, providing accurate and real-time personal information basis for subsequent career path planning.

[0108] The path planning sub-module uses a hierarchical planning algorithm based on multi-agent reinforcement learning to plan short-term, medium-term, and long-term career development paths based on the results of the personal information analysis sub-module. The strategy network of each agent in the multi-agent reinforcement learning part contains 3 fully connected layers with 256, 128, and 64 neurons, respectively, and the activation function uses ReLU. The value network contains 2 fully connected layers with 128 and 64 neurons, respectively, and the activation function uses a linear function. Agents communicate asynchronously through a message queue and collaboratively generate career development paths based on node information in the industry knowledge graph. The hierarchical planning algorithm divides path planning into strategic, tactical, and operational layers. The strategic layer determines the direction of career development, the tactical layer sets stage goals, and the operational layer plans specific learning and practice tasks. Information is transmitted between layers through state transition matrices to achieve dynamic optimization of the path.

[0109] The specific dynamic adjustment logic is as follows:

[0110] Short-term career development path: time interval set to 0.5 ≤ t s ≤ min(2, τ), when the industry technology updates slowly, the upper limit of the short-term path is 2 years, ensuring that talents can master basic skills and adapt to the current needs of the industry in the short term; if τ is less than 2 years, the upper limit is taken as τ, so that the short-term path closely follows the rhythm of rapid changes in the industry and adjusts the direction of talent development in a timely manner.

[0111] Medium-term career development path: time interval max(2, τ) ≤ t m ≤ min(4, 2τ), such setting can give talents enough time to deepen skills and advance in their careers when the industry technology develops steadily; when the industry technology updates faster, the lower and upper limits of the medium-term path will be dynamically adjusted according to τ to ensure that talents always keep pace with industry technological progress in the medium-term development stage.

[0112] Long-term career development path: time interval max(4, 2τ) ≤ t lThe long-term path setting fully considers the long-term development trend of the industry technology. When the industry technology update cycle is long, the talent is provided with a stable long-term development plan. If the industry technology updates rapidly, the lower limit of the long-term path will be increased accordingly to ensure that the talent has foresight in long-term development and constantly adapts to the challenges of new technology.

[0113] Through this dynamic time interval adjustment mechanism based on the industry technology knowledge update cycle half-life, the path planning sub-module can make the career development path planning highly adaptable to the industry technology iteration, effectively avoiding the problem that the talent skills are out of touch with the industry demand in the traditional path planning model, and providing scientific, reasonable and forward-looking career development guidance for the talents of vocational education.

[0114] Through actual application verification, the post matching success rate of users using the system for career path planning is increased by 31 percentage points compared with the traditional method, reaching 89%. In the case of rapid iteration of industry technology, the average adjustment cycle of the user's career development path is shortened from 6 months to 1.5 months, effectively improving the adaptation efficiency of talents and industry demand.

[0115] Embodiment four, please refer to Figure 5 Based on the basis of embodiment one, embodiment two and embodiment three, the talent growth prediction module includes a data processing unit and a prediction analysis unit.

[0116] The data processing unit collects historical talent development data and industry dynamic change data, and uses a generative adversarial network combined with adversarial training for data enhancement.

[0117] The prediction analysis unit adaptively weights the Transformer model, dynamically adjusts the weights of each sub-model through reinforcement learning, introduces the industry technology iteration index T(t) as an external feature, and constructs a dynamic time series prediction model:

[0118]

[0119] Wherein, P(t) is the final prediction result at the current time t, k is the number of sub-models, i.e. the total number of sub-models integrated in the model, ω i (t) is the weight of the i-th sub-model at time t, ε is the industry fluctuation coefficient, P i (t) is the prediction result of the i-th sub-model at time t.

[0120] Further, the data processing unit collects historical talent development and industry dynamic data, covering multi-dimensional information such as education, skills, market demand and technological innovation, to provide a basis for prediction.

[0121] To enhance the diversity and richness of data, the data processing unit introduces a data augmentation technique combining generative adversarial networks with adversarial training. The generative adversarial network consists of a generator and a discriminator. The role of the generator is to generate new data similar to the original data distribution based on the input random noise. These new data are similar in features and structure to the real data but have certain differences, thereby expanding the number of data samples. For example, when processing talent work experience data, the generator can generate work experience data simulating different industries and different positions, enriching the diversity of data.

[0122] The discriminator is responsible for determining whether the input data is real data or fake data generated by the generator. In the adversarial training process, the generator continuously optimizes itself to generate more realistic data to deceive the discriminator, while the discriminator also continuously learns to improve its ability to distinguish between real and fake data. This adversarial game between the generator and the discriminator promotes the continuous improvement of the quality of the data generated by the generator. The augmented data not only increases in quantity but also is closer to real data in quality, providing more sufficient and representative data support for the subsequent prediction analysis unit, effectively alleviating the problem of insufficient model training caused by insufficient data.

[0123] The prediction analysis unit adopts a self-adaptive weight Transformer model based on Transformer architecture optimization, dynamically adjusts the weights of sub-models through reinforcement learning, and realizes accurate prediction of talent growth probability. One of the important innovations of the model is that it dynamically adjusts the weights of each sub-model through reinforcement learning. Traditional model fusion methods usually combine sub-models with fixed weights, which is difficult to adapt to complex and variable data characteristics and prediction needs. The self-adaptive weight Transformer model in this system uses reinforcement learning algorithm to construct a reward mechanism by taking the prediction results of the model as feedback signals. For example, when the prediction results of the model are consistent with the actual talent growth, positive rewards are given; if the prediction deviation is large, negative feedback is given. Through continuous trial and error and learning, the model can automatically optimize the contribution of each sub-model according to data characteristics and prediction needs, so that each sub-model can perform best in different prediction scenarios. This dynamic weight adjustment mechanism greatly improves the flexibility and adaptability of the model, enabling it to more accurately capture the complex rules in the talent growth process.

[0124] Introducing industry technology iteration index T(t) as an external feature to construct a dynamic time series prediction model The industry technology iteration index T(t) considers multiple factors such as the speed of new technology research and development, the popularity of technology application, and the frequency of technology update, and can quantitatively reflect the iteration speed and development trend of industry technology.

[0125] In the model, P(t) is the final prediction result at the current time t, k is the number of sub-models, i.e., the total number of sub-models integrated in the model, ω i (t) is the weight of the i-th sub-model at time t, ε is the industry fluctuation coefficient, used to adjust the influence degree of the industry technology iteration index on the prediction result, P i (t) is the prediction result of the i-th sub-model at time t.

[0126] By incorporating the industry technology iteration index into the model, the prediction result can reflect the impact of industry technology changes on talent growth in real time. For example, when the industry technology iteration index T(t) rises rapidly, it means that the industry technology updates faster, and the model will adjust the prediction result accordingly, and be more inclined to predict that talents need to speed up skill updating to adapt to industry development, thereby providing more forward-looking growth recommendations for talents, and providing more accurate reference for talent training and recruitment decisions for vocational education institutions and enterprises.

[0127] In summary, the talent growth prediction module realizes dynamic and accurate prediction of talent growth trend through innovative design and collaborative operation of data processing unit and prediction analysis unit, further improves the talent recommendation system based on matching local productivity of vocational education, effectively solves the problems of data shortage, poor model adaptability and inability to fully consider industry dynamic changes in traditional talent prediction methods.

[0128] Embodiment five, please refer to Figure 6 Based on the basis of embodiment one, embodiment two, embodiment three and embodiment four, the path adjustment module includes a cause analysis submodule and a scheme adjustment submodule;

[0129] The cause analysis submodule identifies the reasons for slow skill learning progress and sudden changes in industry demand when the talent fails to achieve the growth target on time by using the causal forest algorithm combined with bias analysis and introducing the counterfactual reasoning mechanism;

[0130] The scheme adjustment submodule uses a hierarchical multi-objective optimization algorithm to set dynamic weights γ i (t) for short-term target skill mastery O1, medium-term target job matching degree O2 and long-term target industry contribution value O3, respectively, and constructs a comprehensive evaluation function based on a priority decision model of game theory:

[0131]

[0132] Wherein, γ is the conflict adjustment factor, γ i (t) is the dynamic weight of the i-th target at time t, C j is the j-th target conflict coefficient, and the weight γ i(t) is dynamically updated by the following equation:

[0133]

[0134] where η is the learning rate, V(π t ) is the value function of policy π t , denotes the partial derivative with respect to O i , is the gradient of the value function V(π t ) of policy π t with respect to the objective O i , γ i (t+1) is the weight of the i-th objective at time t+1, and γ i (t) is the weight of the i-th objective at time t.

[0135] Further, when the talent fails to achieve the growth target on time, the cause analysis submodule starts the deep diagnosis mechanism of multi-technology fusion. The cause analysis submodule uses the causal forest algorithm to quickly locate the key factors affecting talent growth in complex data. For example, when analyzing the problem of slow improvement of programming skills of vocational education talents, the causal forest algorithm can quantify the causal effect of each factor on skill improvement from many potential factors such as learning time allocation, course difficulty adaptation, and practical project participation, and locate the real influencing factors.

[0136] Combined with bias analysis technology, the actual development trajectory of talents is systematically compared with the preset growth target. Through time series analysis, not only the absolute deviation value of skill mastery progress, job matching degree and other indicators can be found, but also the trend of deviation can be excavated, such as the turning point of skill growth rate slowing down. For example, if it is found that the job matching degree of a certain talent in the middle stage of career development is continuously lower than expected, the bias analysis can accurately locate the gap in specific skill areas, such as the lack of mastery of emerging data analysis tools.

[0137] The counterfactual reasoning mechanism is introduced innovatively to build a virtual scenario to deduce the causal relationship. This mechanism simulates the growth results of talents under other possible conditions by assuming different conditions and intervention measures, so as to infer the specific reasons for the failure to achieve the target. For example, when the industry demand changes due to the application of new technology, counterfactual reasoning can simulate whether the decline in job matching degree can be avoided if the talent has access to new technology training in advance, and thus determine the causal relationship between the sudden change of industry demand and the lag of talent skills, providing accurate problem diagnosis basis for subsequent scheme adjustment.

[0138] The solution adjustment submodule, guided by the diagnostic results of the cause analysis submodule, employs a hierarchical multi-objective optimization algorithm to construct a dynamic weight adjustment and priority decision-making system. Dynamic weights γ are assigned to the short-term goal of skill mastery O1, the mid-term goal of job matching O2, and the long-term goal of industry contribution value O3. i (t), which breaks through the limitations of traditional fixed-weight programming.

[0139] Weight dynamic update formula This achieves adaptive weight adjustment, where η is the learning rate, controlling the step size of weight adjustment to avoid over- or under-adjustment, and V(π) t ) is the strategy π t The value function quantifies the strengths and weaknesses of a career development strategy by evaluating its overall benefits across various objectives. Indicates O i The partial derivatives, For strategy π t Value function V(π) t For target O i The gradient of γ indicates the direction and magnitude of the improvement in policy value resulting from adjusting the weights of each objective. i (t+1) represents the weight of the i-th target at time t+1, γ i (t) represents the weight of the i-th objective at time t. For example, when a sudden change in industry demand causes a sharp drop in job matching, this formula can quickly increase the weight of the mid-term objective job matching O2, guiding the system to prioritize optimizing job matching solutions.

[0140] Constructing a comprehensive evaluation function based on a priority decision-making model of game theory Balancing conflict and cooperation among multiple objectives, where γ is a conflict moderating factor. i (t) represents the dynamic weight of the i-th target at time t, C j Let be the conflict coefficient of the j-th target.

[0141] The conflict mitigation factor γ can be dynamically adjusted according to the degree of conflict between objectives, avoiding optimization failure caused by contradictions between objectives. For example, when there is a resource allocation conflict between short-term skill learning investment and long-term industry contribution development, γ will reduce the weight of the conflicting objective and reallocate resources to improve overall returns. By continuously iterating and updating the weights, this module can optimize career development paths in real time, such as adjusting the order of course learning, recommending targeted training programs, or planning cross-domain skill expansion programs, ensuring that talent development always aligns with dynamic industry needs and individual growth potential.

[0142] The path adjustment module forms a closed-loop feedback mechanism through the accurate diagnosis of the cause analysis submodule and the dynamic optimization of the scheme adjustment submodule, effectively solving the problem of path rigidity in traditional career planning that is difficult to cope with complex changes, providing development strategies with dynamic adaptability for vocational education talents, and strengthening the matching accuracy of the talent recommendation system and local productivity demand.

[0143] Embodiment six, please refer to Figure 7 Based on the basis of embodiment one, embodiment two, embodiment three, embodiment four and embodiment five, the talent recommendation module includes a matching submodule and a recommendation execution submodule.

[0144] The matching submodule uses a bidirectional graph neural network based on graph attention mechanism to model the semantic association of job requirements and talent growth paths at the same time, and introduces a dynamic matching degree update mechanism driven by reinforcement learning to calculate the bidirectional matching probability of jobs and talents.

[0145] The recommendation execution submodule is based on a multi-round bilateral matching game model, combines enterprise feedback data to optimize the recommendation strategy, and forms a recommendation, feedback and iterative closed loop.

[0146] Further, the talent recommendation module is based on a bidirectional graph neural network based on graph attention mechanism and a multi-round bilateral matching game model to realize dynamic semantic matching of talents and jobs. The matching submodule uses a bidirectional graph neural network based on graph attention mechanism to model the semantic association of job requirements and talent growth paths at the same time. During network training, the Adam optimizer is used, the initial value of the learning rate is set to 0.001, the batch size is 32, and the network parameters are updated through the back propagation algorithm to realize deep modeling of the semantic association of jobs and talents.

[0147] The matching submodule uses a bidirectional graph neural network based on graph attention mechanism to model the semantic association of job requirements and talent growth paths at the same time. In actual construction of graph structure, the various requirements of the post and the abilities possessed by the talent are abstracted as nodes in the graph, and the association relationship between the nodes is taken as the edge, for example, the requirements of the post include professional skills, work experience and qualification certificates, the abilities possessed by the talent include learning experience, project experience and skill mastery, and the association relationship between the nodes is the specific requirement of a certain skill for the post.

[0148] The graph attention mechanism plays a key role in this regard, as it can automatically assign different attention weights to nodes and edges based on their importance in the graph structure, focusing on nodes and edges that play a key role in modeling semantic associations. For example, for a software development position, the graph attention mechanism will assign higher attention weights to programming language skill nodes, project development experience nodes, and other key features during the modeling process. At the same time, the Q-learning algorithm dynamically updates the matching strategy, adjusts the matching probability based on changes in job requirements and talent growth in real time, and gives negative penalties in the opposite case. In this way, the system dynamically calculates the two-way matching probability between jobs and talents, achieving accurate matching analysis of talents and jobs.

[0149] The recommendation execution sub-module is based on a multi-round bilateral matching game model. In this model, enterprises and talents are considered as game parties, each with their own interests and decision-making strategies. In the initial stage of the recommendation process, the system generates a preliminary matching plan based on the two-way matching probability calculated by the matching sub-module. Enterprises and talents evaluate and decide on this plan based on their own needs, and if they are not satisfied, they will propose adjustments. The system simulates the process of continuous gaming between the two parties through multiple rounds of matching attempts and strategy adjustments. For example, after the first round of matching, the enterprise believes that the talent lacks a certain skill, and the system will select talents with the corresponding skills in subsequent matching based on this feedback. If the talent believes that the salary of the position does not meet expectations, the system will also adjust the recommended plan.

[0150] At the same time, the system optimizes the recommendation strategy based on enterprise feedback data. After the talent is hired, the enterprise will provide feedback on the talent's actual performance data on the job, such as work efficiency, skill application, and team collaboration ability. The system uses these feedback data as an important basis for training and optimizing the recommendation strategy using supervised learning algorithms such as support vector machines. Through continuous iteration, a closed loop of recommendation, feedback, and iteration is formed, significantly improving the accuracy and effectiveness of talent recommendation and achieving dynamic semantic matching between talents and positions.

[0151] Embodiment seven, please refer to Figure 1 Based on the basis of embodiments one, two, three, four, five, and six, the data storage module uses a distributed graph database combined with time series data sharding technology to store multi-modal data in partitions, and implements cross-institution data sharing and privacy protection through a federated learning data security aggregation mechanism based on homomorphic encryption.

[0152] The interactive interface module provides an AR visual career path sand table, supports multi-modal interactive input methods such as voice, gestures, and text, and integrates an intelligent question and answer assistant based on dialogue-enhanced Transformer to answer path planning questions in real time.

[0153] The system management module integrates a regional industry talent supply and demand imbalance risk early warning function based on Kalman filtering, identifies industry talent supply and demand imbalance risks based on an anomaly detection algorithm, and simultaneously adopts a dynamic permission management system based on attribute-based encryption to realize data access control for different roles.

[0154] Further, the data storage module adopts a distributed graph database combined with time series data sharding technology to partition store multi-modal data generated during system operation, the distributed graph database can efficiently store and manage data with complex relationships, the time series data sharding technology stores data according to the time attribute, improves data storage and query efficiency, and through a federated learning data security aggregation mechanism based on homomorphic encryption, cross-institutional data sharing is realized under the premise of ensuring data privacy and security, rich data support is provided for each module of the system, and the safety and privacy of data in the sharing process are ensured.

[0155] The interactive interface module provides an AR visual career path sand table, users can interact with the sand table through multi-modal interaction input methods such as voice, gestures and text, and the sand table displays the career development path in a three-dimensional visual form, enabling users to more intuitively understand their career development plan, and the module integrates an intelligent question and answer assistant based on dialogue-enhanced Transformer, which can understand the user's question intention, combine system knowledge and data to answer user questions about path planning in real time, provide personalized consulting services, and improve user experience.

[0156] The system management module integrates a regional industry talent supply and demand imbalance risk early warning function based on Kalman filtering and a dynamic permission management system based on attribute-based encryption, the risk early warning function based on Kalman filtering, through real-time monitoring and analysis of regional industry talent supply and demand data, uses Kalman filtering algorithm to predict and correct data, identifies industry talent supply and demand imbalance risks in a timely manner, and sends early warning signals, the dynamic permission management system based on attribute-based encryption dynamically allocates and manages data access permissions according to the roles and attributes of users, ensures that users of different roles can only access data within their permission range, and protects the security and confidentiality of system data.

[0157] In the present application, first, the dynamic industry portrait construction module of the talent recommendation system based on matching local productivity of vocational education collects local industry multi-source data through intelligent crawlers, extracts key information through a data analysis submodule, and then predicts industry trends and updates portraits by combining a time series hidden Markov model and a time decay factor, the expert team in the expert review and correction module reviews the portraits, and the information supplement unit supplements undisclosed information to perfect the industry knowledge graph.

[0158] Next, the personalized path planning module integrates personal multimodal data with industry trends and uses a hierarchical planning algorithm to generate short, medium and long-term career development paths. The talent growth prediction module processes and enhances historical and dynamic data, and then predicts the probability of talent growth through an adaptive weighted Transformer model.

[0159] When talent fails to meet the target, the path adjustment module's cause analysis submodule identifies the reasons, the solution adjustment submodule adjusts the path accordingly, the talent recommendation module's matching submodule calculates the match probability between the job and the talent, and the recommendation execution submodule combines enterprise feedback to optimize the strategy and complete the recommendation.

[0160] Finally, the data storage module uses a distributed graph database and time-series data sharding technology to store multimodal data and ensures the security of cross-institutional data sharing. The interactive interface module provides AR sandbox and intelligent question answering to enhance user experience. The system management module realizes talent supply and demand risk warning and dynamic permission management.

[0161] 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 this invention is defined by the appended claims and their equivalents.

Claims

1.A talent recommendation system matching local productivity based on vocational education, characterized by: The talent recommendation system based on matching local productivity with vocational education includes a dynamic industry portrait construction module, an expert review correction module, a personalized path planning module, a talent growth prediction module, a path adjustment module, a talent recommendation module, a data storage module, an interactive interface module, and a system management module; The dynamic industry portrait construction module predicts industry trend data through a time-series hidden Markov model with an attention mechanism, and supplements undisclosed project information based on a knowledge graph completion algorithm and Bayesian inference, generates a dynamic industry knowledge graph based on a time decay factor through graph convolution network feature fusion technology, and the node update frequency is positively correlated with the feature importance index; The personalized path planning module generates a multi-dimensional career development path based on a multi-agent reinforcement learning hierarchical planning algorithm combined with an industry knowledge graph; The talent growth prediction module dynamically predicts talent growth probability using an adaptive weight Transformer model combined with an industry technology iteration index; The path adjustment module dynamically adjusts the career development path through a hierarchical multi-objective optimization algorithm combined with a multi-objective optimization game theory conflict resolution mechanism based on Nash equilibrium; The talent recommendation module realizes dynamic semantic matching of talents and positions based on a bidirectional graph neural network and a multi-round two-sided matching game model based on graph attention mechanism; The data storage module stores multi-modal data during system operation using a distributed graph database combined with time-series data sharding technology; The interactive interface module provides an AR visual career path sand table and a multi-modal intelligent question and answer assistant; The system management module integrates a regional industry talent supply and demand imbalance risk early warning function based on Kalman filtering and a dynamic permission management system based on attribute-based encryption; The dynamic industry portrait construction module includes a data acquisition submodule, a data analysis submodule, and a portrait update submodule; The data acquisition submodule uses intelligent crawler technology based on reinforcement learning combined with a multi-agent collaboration mechanism to real-time capture multi-source data of local industry news, policy documents, and enterprise recruitment information, and filters duplicate data through an adaptive threshold; The data analysis submodule uses a T5-large text generation model combined with a LightGBM ensemble learning algorithm to extract industry key skills, job demand trends, enterprise expansion and contraction dynamic information; The portrait update submodule uses a time-series hidden Markov model with an attention mechanism to predict industry trends, and dynamically adjusts the update strategy through a feature importance formula with a time decay factor; The data feature set is where n is the feature dimension, and the formula is as follows: ； wherein, is the i-th feature in the feature set F, t is the current time, is the feature importance at the current time t, m is the number of decision trees, is the j-th decision tree, denotes the information gain brought by the feature in the j-th decision tree, is the decay coefficient (0 < 1), is the construction time of the j-th decision tree; The expert review correction module includes an expert team unit and an information supplement unit; The expert team unit is composed of enterprise experts with senior technical titles or more than 10 years of industry experience in local pillar industries, and realizes distributed review of industry portraits based on blockchain storage technology through a federated learning collaborative review platform; The information supplement unit uses a knowledge graph completion algorithm to integrate Bayesian inference, structures the undisclosed project talent demand information provided by experts into the industry portrait, and quantifies the information credibility through an evidence weight evaluation mechanism; The personalized path planning module comprises a personal information analysis submodule and a path planning submodule; The personal information analysis submodule integrates professional evaluation questionnaires, learning behavior logs, and online skill test data through multi-modal behavior analysis technology, combines a principal component analysis model optimized by transfer learning, and introduces regional industry policy text mining results to achieve dynamic matching analysis of personal skills and industry trends. The path planning submodule uses a hierarchical planning algorithm based on multi-agent reinforcement learning to plan short-term, medium-term, and long-term career development paths based on the results of the personal information analysis submodule. The talent growth prediction module comprises a data processing unit and a prediction analysis unit. The data processing unit collects historical talent development data and industry dynamic change data, and uses a generative adversarial network combined with adversarial training for data augmentation. The prediction analysis unit self-adapting weight transformer model dynamically adjusts the weights of each sub-model through reinforcement learning, and introduces an industry technology iteration index As an external feature, a dynamic time series prediction model is constructed: ； wherein, is the final prediction result for the current time t, k is the number of sub-models, i.e. the total number of sub-models integrated in the model, is the weight of the i-th sub-model at time t, , is the industry volatility coefficient, is the prediction result of the i-th sub-model at time t; The path adjustment module comprises a cause analysis submodule and a scheme adjustment submodule. The cause analysis submodule uses a causal forest algorithm combined with bias analysis to introduce a counterfactual reasoning mechanism to identify the reasons for slow skill learning progress and sudden changes in industry demand when the talent fails to achieve the growth target on time. The scheme adjustment sub-module adopts a hierarchical multi-objective optimization algorithm to adjust the short-term target skill mastery degree , the medium-term target post matching degree , and the long-term target industry contribution value , respectively sets dynamic weights , and constructs a comprehensive evaluation function based on a priority decision model of game theory: ； wherein, is a conflict regulation factor, is a dynamic weight of the ith objective at time t, is a conflict coefficient of the jth objective, , and the weight is dynamically updated by the following equation: ； in, For learning rate, For strategy The value function, Indicates to The partial derivatives, For strategy value function For the target gradient, For time The weight of the i-th target. Let be the weight of the i-th objective at time t; The talent recommendation module comprises a matching submodule and a recommendation execution submodule. The matching submodule uses a bidirectional graph neural network based on graph attention mechanism to model the semantic association between job requirements and talent growth paths, introduces a reinforcement learning driven dynamic matching degree update mechanism, and calculates the bidirectional matching probability of job and talent. The recommendation execution submodule optimizes the recommendation strategy based on multi-round bilateral matching game model combined with enterprise feedback data to form a recommendation, feedback, and iteration closed loop. 2.The talent recommendation system matching local productivity based on vocational education according to claim 1, characterized in that: The time interval of the short-term career development path is 0.5-2 years, the time interval of the medium-term career development path is 2-4 years, and the time interval of the long-term career development path is more than 4 years. The path planning submodule dynamically adjusts the time interval based on the half-life of the industry technology knowledge update cycle, which is the time required for 50% of existing technical knowledge in a certain field to be replaced by new technology. Short term path: 00.5 < t < 0.5 ≤ ; Mid-term path: ≤ ≤ ; Long term path: ≤ ; wherein, T is the half-life of the industry technology knowledge update cycle, defined as the time required for 50% of the existing knowledge in a field to be replaced by new technology, , and are the upper bounds of the time intervals for short, medium, and long term paths, respectively. 3.The talent recommendation system matching local productivity based on vocational education according to claim 1, characterized in that: The data storage module uses a distributed graph database combined with time series data sharding technology to store multi-modal data in partitions, and realizes cross-institution data sharing and privacy protection through a federated learning data security aggregation mechanism based on homomorphic encryption. 4.The talent recommendation system matching local productivity based on vocational education according to claim 1, characterized in that: The interactive interface module provides an AR visual career path sand table, supports multi-modal interactive input methods such as voice, gesture, and text, and integrates an intelligent question and answer assistant based on dialogue-enhanced Transformer to answer path planning questions in real time. The system management module integrates a regional industry talent supply and demand imbalance risk early warning function based on Kalman filtering, identifies industry talent supply and demand imbalance risks based on an anomaly detection algorithm, and uses a dynamic permission management system based on attribute-based encryption to realize data access control for different roles.

Images