Artificial intelligence-based employee post matching and deploying method and system

Abstract

The invention provides an employee post matching and deploying method and system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the method comprises the steps: generating an optimized employee capability feature vector based on employee historical work data, extracting an employee capability development trend, and combining team cooperation data to obtain an optimized employee capability feature vector; a cross-modal contrast learning network is utilized to obtain an employee comprehensive ability vector, a development potential index is calculated, a post skill knowledge graph is constructed based on post instruction data and historical recruitment data, and a post skill feature matrix and a team structure feature vector are generated through a dynamic heterogeneous graph neural network and an adaptive graph attention network. Calculating an employee-post matching degree score matrix in combination with an employee comprehensive ability vector, inputting the matching degree score matrix and a development potential index into a distributed multi-agent reinforcement learning model, constructing a reward function, generating and evaluating a deployment scheme through hierarchical task decomposition and Monte Carlo tree search, and finally outputting a deployment execution scheme.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an employee position matching and deployment method and system based on artificial intelligence. Background Art

[0002] With the rapid development of artificial intelligence technology, enterprises have an increasing demand for efficient and accurate employee job matching and deployment. Traditional employee job matching and deployment methods mainly rely on manual experience and subjective judgment, which are inefficient and prone to matching errors, making it difficult to adapt to the rapidly changing needs of enterprises;

[0003] At present, existing technologies mainly use employee resume information and job descriptions for text matching, or match based on employee skill tags and job requirements. However, there are still problems such as only considering employees' static skills and experience, difficulty in capturing employees' ability development trends and potential, ignoring the impact of teamwork factors on employee performance, and being unable to make real-time adjustments based on corporate development changes and employee ability improvement.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide an employee position matching and deployment method and system based on artificial intelligence, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides an employee position matching and deployment method based on artificial intelligence, comprising:

[0007] Collecting employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, performing multi-scale feature extraction on the employee historical work data, obtaining an initial employee capability feature vector and adding it to a pre-set diffusion probability model, generating an optimized employee capability feature vector in combination with a Markov chain sampling process, adding the work performance data and the skill evaluation data to a pre-set spatiotemporal hybrid network model, extracting employee capability development trends through convolution operations, adding the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtaining an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and adding it to a feature enhancement network, and obtaining a development potential index in combination with a multi-scale feature pyramid and an adaptive feature aggregation module;

[0008] Collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. Combined with an adaptive graph attention network, hierarchical sampling and dynamic aggregation are performed to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network, and an employee-job matching score matrix is ​​obtained through parameter learning, super network optimization, and knowledge transfer.

[0009] The employee-job matching score matrix and the development potential index are input into a distributed multi-agent reinforcement learning model to construct a state space including the current personnel distribution status and team structure status, and an action space based on the deployment plan space is constructed. A reward function is constructed based on the job matching score, the development potential index and the team collaboration data. An iterative training is performed based on an exploration strategy based on intrinsic incentives, and the deployment strategy is optimized through a hierarchical task decomposition mechanism to generate an initial deployment plan. The initial deployment plan is evaluated through the value network prediction module and the path evaluation module in the Monte Carlo tree search model. A deployment execution plan is output based on the evaluation results, and the performance data of the deployed employees and the team collaboration data are analyzed in combination with a causal reasoning network. The parameters of the multi-agent reinforcement learning model are adjusted in combination with a meta-strategy optimization algorithm.

[0010] In an optional embodiment,

[0011] Collect employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, perform multi-scale feature extraction on the employee historical work data, obtain an initial employee capability feature vector and add it to a preset diffusion probability model, combine the Markov chain sampling process to generate an optimized employee capability feature vector, add the work performance data and the skill evaluation data to a preset spatiotemporal hybrid network model, extract employee capability development trends through convolution operations, add the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtain an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and add it to a feature enhancement network, and obtain a development potential index by combining a multi-scale feature pyramid and an adaptive feature aggregation module, including:

[0012] Collecting employee historical work data, wherein the employee historical work data includes work performance data, skill assessment data and team collaboration data, dividing the employee historical work data into a training set, a validation set and a test set, building a convolutional neural network model, inputting the data in the training set into the convolutional neural network model, calculating an output feature vector through forward propagation, calculating the difference between the output feature vector and the employee's actual ability level based on a loss function, updating network parameters through back propagation, using the data in the validation set to evaluate the model generalization performance and adjust hyperparameters, inputting the employee historical work data into the trained convolutional neural network model, and outputting an initial employee ability feature vector;

[0013] Constructing a graph structure of a diffusion probability model, taking the initial employee capability feature vector as a node feature, constructing edges between nodes based on organizational structure and project cooperation information, calculating the similarity between feature vectors of different nodes, defining a transition probability matrix based on the similarity, randomly selecting a node as a starting node, performing random walk sampling according to the transition probability matrix, recording a sequence of visited nodes to form a sampling path, performing average aggregation on node feature vectors in the sampling path, repeating multiple random walk samplings, and obtaining a distribution of employee capability feature vectors after diffusion;

[0014] The diffused employee capability feature vector distribution is used as the initial distribution of Markov chain sampling, the conditional probability distribution between feature vectors is calculated, a Markov chain transfer matrix is ​​defined based on the conditional probability distribution, an initial state is sampled from the initial distribution, state transfer is iteratively performed according to the Markov chain transfer matrix to form a Markov chain, Markov chain sampling is repeated multiple times to obtain a state sequence, feature vectors in the state sequence are averaged and aggregated, and an optimized employee capability feature vector is output;

[0015] Constructing a cross-modal contrastive learning network including three feature encoders, wherein the three feature encoders are used to process the optimized employee capability feature vector, employee capability development trend vector and team collaboration data respectively, mapping the three modal data to a common feature space through the three feature encoders, introducing a dynamic temperature parameter, calculating the contrast loss between similar sample pairs and dissimilar sample pairs based on the dynamic temperature parameter, optimizing the feature encoder and the dynamic temperature parameter, aligning different modal features in the common space, and concatenating the three feature vectors to obtain an employee comprehensive capability vector;

[0016] Construct a feature enhancement network including a multi-scale feature pyramid and an adaptive feature aggregation module, input the employee's comprehensive ability vector into the bottom layer of the feature pyramid, generate feature maps of different scales through convolution and pooling operations, calculate the feature aggregation weight through the attention mechanism, use the adaptive feature aggregation module to perform weighted fusion on feature maps of different scales, map the fused multi-scale features to a predetermined range through a fully connected layer, and output the employee development potential index.

[0017] In an optional embodiment,

[0018] The contrast loss between similar sample pairs and dissimilar sample pairs is calculated based on the dynamic temperature parameter as shown in the following formula:

[0019] ;

[0020] Among them, L represents the contrast loss, f(x i ) represents the sample x i The eigenvector of f(x j ) represents the sample x j , r represents the dynamic temperature parameter, sim() represents the similarity metric function, f(x k ) represents the sample x k The feature vector of , k represents the sample index.

[0021] In an optional embodiment,

[0022] Collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. Combined with an adaptive graph attention network, hierarchical sampling and dynamic aggregation are performed to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network. Through parameter learning, super network optimization, and knowledge transfer, the employee-job matching score matrix is ​​obtained, including:

[0023] Collect job description data in the human resources system and historical recruitment data in the recruitment system, perform data cleaning operations to remove redundant information and outliers in the job description data and the historical recruitment data, perform data standardization processing to construct a standardized job description corpus and a historical recruitment corpus, input the job description corpus and the historical recruitment corpus into a pre-constructed generative pre-trained language model, call a named entity recognition module to identify skill entities and knowledge entities in the job description corpus and the historical recruitment corpus, call a relationship extraction module to extract the association relationship between the skill entity and the knowledge entity and the job, and establish a job skill knowledge graph with the job as the central node and the skill entity and the knowledge entity as the attribute node;

[0024] Construct a dynamic heterogeneous graph neural network model, input the job skill knowledge graph into the dynamic heterogeneous graph neural network model, calculate the connection weights between nodes based on a time-varying topology learning strategy, dynamically update the graph structure according to the connection weights, construct a heterogeneous subgraph sampler, input the node type information and edge type information of the job skill knowledge graph, generate multiple local subgraphs based on node semantic similarity sampling, perform graph structure distillation operations on the local subgraphs, map job node features and skill node features to a low-dimensional feature space, and fuse the job node features and the skill node features to generate a job skill feature matrix;

[0025] Construct an adaptive graph attention network model, input employee rank information, department information and reporting relationship information, perform hierarchical sampling operations to generate primary team subgraphs, intermediate team subgraphs and advanced team subgraphs; calculate the attention weights of nodes in the primary team subgraphs, the intermediate team subgraphs and the advanced team subgraphs, fuse the feature information of team members at different levels based on the attention weights, and output a team structure feature vector;

[0026] A neural architecture search network model is constructed, the job skill feature matrix, the employee comprehensive ability vector and the team structure feature vector are input into the neural architecture search network model, a parameter sharing module is called to perform shared optimization on network parameters of different candidate architectures, a super network optimization module is called to search for the optimal matching model structure, a knowledge transfer module is constructed, historical job matching data and team collaboration data are input into the knowledge transfer module, migration features across jobs and teams are extracted, the migration features are input into the matching model for model training, and an employee-job matching score matrix is ​​output.

[0027] In an optional embodiment,

[0028] Constructing a heterogeneous subgraph sampler, inputting the node type information and edge type information of the job skill knowledge graph, generating multiple local subgraphs based on node semantic similarity sampling, performing graph structure distillation operations on the local subgraphs, mapping job node features and skill node features to a low-dimensional feature space, and fusing the job node features and the skill node features to generate a job skill feature matrix, including:

[0029] Construct a heterogeneous subgraph sampler, the heterogeneous subgraph sampler includes a node semantic similarity calculation module, the node semantic similarity calculation module includes a multi-head self-attention layer and a feedforward neural network, the node type information and edge type information of the job skill knowledge graph are input into the heterogeneous subgraph sampler, the dot product similarity between node feature vectors is calculated through the multi-head self-attention layer, and the semantic similarity matrix between nodes is calculated based on the dot product similarity and the node type information;

[0030] Constructing an adaptive sampling module, inputting the semantic similarity matrix between nodes and the node degree distribution information into the adaptive sampling module, increasing the sampling weight of highly central nodes based on the semantic similarity matrix between nodes, reducing the sampling weight of edge nodes, and performing sampling operations according to preset job node ratios and skill node ratios to generate multiple local subgraphs;

[0031] Constructing a graph structure distillation module, the graph structure distillation module includes a dual-path encoder, the dual-path encoder includes a feature encoding path and a structure encoding path, wherein the feature encoding path includes a graph convolutional network, and the structure encoding path includes a graph isomorphic network, inputting the local subgraph into the dual-path encoder, extracting node local features through the graph convolutional network, extracting global topological features through the graph isomorphic network, and using a gating mechanism to fuse the node local features and the global topological features to obtain a node feature representation;

[0032] Constructing a contrastive learning module, inputting the node feature representation into the contrastive learning module, constructing the node feature representation in the original feature space and the node feature representation in the low-dimensional embedding space as a positive sample pair, constructing the shuffled node feature representation as a negative sample, performing a feature mapping operation based on the positive sample pair and the negative sample, and mapping the position node feature and the skill node feature to a low-dimensional feature space respectively;

[0033] A feature fusion module is constructed, wherein the feature fusion module includes a node-level fusion layer and a graph-level fusion layer, wherein the node-level fusion layer includes a cross-attention mechanism, and the graph-level fusion layer includes a graph pooling module. The mapped job node features and the skill node features are input into the feature fusion module, and the feature importance weights between the job node features and the skill node features are calculated through the cross-attention mechanism. The node features are aggregated through the graph pooling module to obtain graph-level features. The feature representations at different levels are connected using a residual connection mechanism, and the graph-level features and the feature representations are fused through an adaptive feature aggregation module to generate a job skill feature matrix.

[0034] In an optional embodiment,

[0035] The employee-position matching score matrix and the development potential index are input into a distributed multi-agent reinforcement learning model, a state space including the current personnel distribution status and the team structure status is constructed, an action space based on the deployment plan space is constructed, a reward function is constructed based on the position matching score, the development potential index and the team collaboration data, iterative training is performed based on an exploration strategy based on intrinsic incentives, and the deployment strategy is optimized through a hierarchical task decomposition mechanism to generate an initial deployment plan, the initial deployment plan is evaluated through a value network prediction module and a path evaluation module in a Monte Carlo tree search model, a deployment execution plan is output according to the evaluation result, and the performance data and team collaboration data of the deployed employees are analyzed in combination with a causal reasoning network, and the parameters of the multi-agent reinforcement learning model are adjusted in combination with a meta-strategy optimization algorithm, including:

[0036] Inputting the employee job matching score matrix and the employee development potential index into a distributed multi-agent reinforcement learning model, using a neural network to extract features from the employee job matching score matrix to obtain an employee job matching feature vector, and using a deep learning network to normalize the employee development potential index to obtain a standardized potential index;

[0037] Generate a personnel distribution state vector in the form of a high-dimensional sparse vector, wherein the dimension of the personnel distribution state vector is the product of the number of employees and the number of positions, set the vector element corresponding to the employee's current position to 1, and set the remaining elements to 0, establish a team structure state matrix, the team structure state matrix uses an adjacency matrix to describe the reporting relationship between team members, and combine the personnel distribution state vector and the team structure state matrix to form a state space;

[0038] Generate an employee deployment plan matrix in the form of a two-dimensional matrix, wherein the number of rows of the employee deployment plan matrix is ​​equal to the total number of employees, the number of columns is equal to the total number of positions, the matrix elements of employees assigned to corresponding positions are set to 1, and the matrix elements that are not assigned are set to 0, and the employee deployment plan matrix is ​​used as an action space;

[0039] Calculating the similarity between the employee job matching feature vector and the preset job requirement vector to obtain a matching score, multiplying the standardized potential index by the preset job development coefficient to obtain a potential score, calculating the collaboration density between team members based on the team structure state matrix to obtain a collaboration score, constructing a reward function for the matching score, the potential score and the collaboration score and solving it to obtain a reward value;

[0040] Decompose the overall deployment task into local team deployment subtasks and cross-department deployment subtasks. For the local team deployment subtasks, use a deep reinforcement learning algorithm to train an agent to deploy personnel within the team. For the cross-department deployment subtasks, use a hierarchical reinforcement learning algorithm to train an agent to complete cross-department personnel deployment. Combine the training results of the local team deployment subtasks and the training results of the cross-department deployment subtasks to generate an initial deployment plan.

[0041] The value network prediction module is used to predict the long-term benefits of the initial deployment plan. The improvement value of different deployment decisions on future team performance is calculated based on the graph neural network. The path evaluation module is used to evaluate the deployment path. The Monte Carlo tree search algorithm is used to calculate the cumulative benefits of different deployment plans, and the plan with the highest evaluation value is selected as the optimal deployment plan.

[0042] The employee performance data and team collaboration data after executing the optimal deployment plan are collected, and the employee performance data and the team collaboration data are input into the causal reasoning network. The causal relationship between the deployment decision and the performance and collaboration is analyzed through the structural equation model, a causal relationship diagram is generated, and the key causal path is extracted to obtain the causal analysis results. According to the causal analysis results, the gradient descent algorithm is used to optimize the network parameters in the distributed multi-agent reinforcement learning model, update the feature representation method of the state space, adjust the constraints of the action space, modify the calculation weights of the reward value, and realize the adaptive optimization of the model parameters.

[0043] In an optional embodiment,

[0044] Based on the graph neural network, the improvement value of different deployment decisions on future team performance is calculated. The path evaluation module is used to evaluate the deployment path. The Monte Carlo tree search algorithm is used to calculate the cumulative benefits of different deployment plans. The plan with the highest evaluation value is selected as the optimal deployment plan, including:

[0045] Inputting an employee node feature vector and a position node feature vector into a pre-set graph neural network model, wherein the employee node feature vector includes the employee's skill features and performance data, and the position node feature vector includes the position requirement features and task description features, and constructing a collaboration relationship edge feature and a position matching relationship edge feature based on the employee node feature vector and the position node feature vector;

[0046] A message passing neural network is used to update the node hidden state, a feature aggregation operation is performed on the employee node feature vector and the position node feature vector to generate a node representation vector, and an edge feature aggregation operation is performed on the collaboration relationship edge feature and the position matching relationship edge feature to generate an edge representation vector;

[0047] Constructing a training data set based on pre-acquired historical deployment data, wherein the historical deployment data includes historical deployment plans and team performance indicators, inputting the training data set into the graph neural network model, and optimizing the network parameters of the graph neural network model using a back propagation algorithm;

[0048] Inputting the initial deployment plan into the graph neural network model, calculating the team performance improvement value of each deployment decision based on the node representation vector and the edge representation vector, and generating a decision evaluation vector;

[0049] Constructing a deployment decision tree, taking the initial deployment plan as the root node, generating child nodes based on employee deployment constraints, each child node representing a deployment state, and calculating a performance evaluation score of each child node based on the decision evaluation vector;

[0050] The upper confidence bound algorithm is used to select child nodes. The upper confidence bound value of each child node is calculated according to the average performance evaluation score of the child node, the number of visits to the parent node, the number of visits to the child node, and the preset exploration coefficient. The child node with the largest upper confidence bound value is selected for expansion. The upper confidence bound value adds exploration items based on the average performance evaluation score of the child node. The exploration items increase with the increase of the number of visits to the parent node and decrease with the increase of the number of visits to the child node.

[0051] Performing an expansion operation on the selected child node, generating a new deployment decision based on the employee deployment constraint, and adding the new deployment decision as a new child node to the deployment decision tree;

[0052] The graph neural network model is used to evaluate the performance of the newly added sub-nodes, calculate the team performance improvement value of the newly added sub-nodes, transmit the team performance improvement value back to the root node along the decision path, and update the number of visits and average performance evaluation score of each node on the path;

[0053] The node selection, node expansion, performance evaluation and result return operations are repeatedly performed until a preset search depth is reached, and a decision path with the highest average performance evaluation score is selected in the deployment decision tree as the optimal deployment solution.

[0054] A second aspect of an embodiment of the present invention provides an employee position matching and deployment system based on artificial intelligence, including:

[0055] The first unit is used to collect employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, perform multi-scale feature extraction on the employee historical work data, obtain an initial employee capability feature vector and add it to a preset diffusion probability model, generate an optimized employee capability feature vector in combination with a Markov chain sampling process, add the work performance data and the skill evaluation data to a preset spatiotemporal hybrid network model, extract employee capability development trends through convolution operations, add the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtain an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and add it to a feature enhancement network, and obtain a development potential index in combination with a multi-scale feature pyramid and an adaptive feature aggregation module;

[0056] The second unit is used to collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. The dynamic heterogeneous graph neural network performs hierarchical sampling and dynamic aggregation in combination with an adaptive graph attention network to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network, and an employee-job matching score matrix is ​​obtained through parameter learning, super network optimization, and knowledge transfer.

[0057] The third unit is used to input the employee-job matching score matrix and the development potential index into a distributed multi-agent reinforcement learning model, construct a state space including the current personnel distribution status and team structure status, construct an action space based on the deployment plan space, construct a reward function based on the job matching score, the development potential index and the team collaboration data, perform iterative training based on the exploration strategy of intrinsic incentives, optimize the deployment strategy through a hierarchical task decomposition mechanism, generate an initial deployment plan, evaluate the initial deployment plan through the value network prediction module and the path evaluation module in the Monte Carlo tree search model, output the deployment execution plan according to the evaluation results, analyze the performance data of the deployed employees and the team collaboration data in combination with the causal reasoning network, and adjust the parameters of the multi-agent reinforcement learning model in combination with the meta-strategy optimization algorithm.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] A fourth aspect of the embodiments of the present invention is:

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In the present invention, through multi-scale feature extraction, spatiotemporal hybrid network model and cross-modal comparative learning and other technologies, the ability of employees is evaluated more comprehensively, and combined with generative pre-trained language model and dynamic heterogeneous graph neural network, the job requirements are understood more deeply, thereby improving the matching accuracy, using employees' historical work data, extracting the development trend of employee capabilities through convolution operation, and combining multi-scale feature pyramid and adaptive feature aggregation module to generate development potential index, more accurately predicting the future development potential of employees, using distributed multi-agent reinforcement learning model, combining job matching, development potential and team collaboration data for iterative training, and evaluating and adjusting through Monte Carlo tree search and causal reasoning network to generate better deployment plan, thereby improving the overall efficiency of the team. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of an employee position matching and deployment method based on artificial intelligence according to an embodiment of the present invention;

[0067] Figure 2 Schematic diagram of the structure of an employee position matching and deployment system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of an employee position matching and deployment method based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] S1. Collecting employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, performing multi-scale feature extraction on the employee historical work data, obtaining an initial employee capability feature vector and adding it to a pre-set diffusion probability model, generating an optimized employee capability feature vector in combination with a Markov chain sampling process, adding the work performance data and the skill evaluation data to a pre-set spatiotemporal hybrid network model, extracting employee capability development trends through convolution operations, adding the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtaining an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and adding it to a feature enhancement network, and obtaining a development potential index in combination with a multi-scale feature pyramid and an adaptive feature aggregation module;

[0072] The diffusion probability model is a model based on random processes, used to describe the propagation or diffusion process of information or matter in time and space. The Markov chain sampling process is a random sampling method based on Markov chains, used to generate samples from complex probability distributions. The spatiotemporal hybrid network model is a deep learning model that combines spatial information and temporal information, and is often used to process spatiotemporal sequence data, such as video analysis and traffic flow prediction. The cross-modal contrastive learning network is a deep learning method for learning the association between different modal data. By comparing the representations of different modal data, a cross-modal shared feature space is learned, which is widely used in multimodal learning tasks such as vision-language and image-audio. The contrast loss function is a loss function used to train a cross-modal contrastive learning network. The multi-scale feature pyramid is a network structure for extracting multi-scale features, which is usually used in tasks such as image segmentation and target detection.

[0073] In an optional embodiment,

[0074] Collect employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, perform multi-scale feature extraction on the employee historical work data, obtain an initial employee capability feature vector and add it to a preset diffusion probability model, combine the Markov chain sampling process to generate an optimized employee capability feature vector, add the work performance data and the skill evaluation data to a preset spatiotemporal hybrid network model, extract employee capability development trends through convolution operations, add the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtain an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and add it to a feature enhancement network, and obtain a development potential index by combining a multi-scale feature pyramid and an adaptive feature aggregation module, including:

[0075] Collecting employee historical work data, wherein the employee historical work data includes work performance data, skill assessment data and team collaboration data, dividing the employee historical work data into a training set, a validation set and a test set, building a convolutional neural network model, inputting the data in the training set into the convolutional neural network model, calculating an output feature vector through forward propagation, calculating the difference between the output feature vector and the employee's actual ability level based on a loss function, updating network parameters through back propagation, using the data in the validation set to evaluate the model generalization performance and adjust hyperparameters, inputting the employee historical work data into the trained convolutional neural network model, and outputting an initial employee ability feature vector;

[0076] Constructing a graph structure of a diffusion probability model, taking the initial employee capability feature vector as a node feature, constructing edges between nodes based on organizational structure and project cooperation information, calculating the similarity between feature vectors of different nodes, defining a transition probability matrix based on the similarity, randomly selecting a node as a starting node, performing random walk sampling according to the transition probability matrix, recording a sequence of visited nodes to form a sampling path, performing average aggregation on node feature vectors in the sampling path, repeating multiple random walk samplings, and obtaining a distribution of employee capability feature vectors after diffusion;

[0077] The diffused employee capability feature vector distribution is used as the initial distribution of Markov chain sampling, the conditional probability distribution between feature vectors is calculated, a Markov chain transfer matrix is ​​defined based on the conditional probability distribution, an initial state is sampled from the initial distribution, state transfer is iteratively performed according to the Markov chain transfer matrix to form a Markov chain, Markov chain sampling is repeated multiple times to obtain a state sequence, feature vectors in the state sequence are averaged and aggregated, and an optimized employee capability feature vector is output;

[0078] Constructing a cross-modal contrastive learning network including three feature encoders, wherein the three feature encoders are used to process the optimized employee capability feature vector, employee capability development trend vector and team collaboration data respectively, mapping the three modal data to a common feature space through the three feature encoders, introducing a dynamic temperature parameter, calculating the contrast loss between similar sample pairs and dissimilar sample pairs based on the dynamic temperature parameter, optimizing the feature encoder and the dynamic temperature parameter, aligning different modal features in the common space, and concatenating the three feature vectors to obtain an employee comprehensive capability vector;

[0079] Construct a feature enhancement network including a multi-scale feature pyramid and an adaptive feature aggregation module, input the employee's comprehensive ability vector into the bottom layer of the feature pyramid, generate feature maps of different scales through convolution and pooling operations, calculate the feature aggregation weight through the attention mechanism, use the adaptive feature aggregation module to perform weighted fusion on feature maps of different scales, map the fused multi-scale features to a predetermined range through a fully connected layer, and output the employee development potential index.

[0080] The transition probability matrix is ​​a core concept in the Markov chain, which is used to describe the probability of transitioning from one state to another. Each element of the transition probability matrix represents the probability of the current state transitioning to the next state.

[0081] Collect employee historical work data, including work performance data, such as the number of completed projects, project quality ratings, customer satisfaction, etc.; skill assessment data, such as professional skill level ratings, general skill test results, etc.; and team collaboration data, such as team contribution, communication and collaboration evaluation, etc. For example, collect an employee's work performance data for the past three years, including the number of completed projects, project quality ratings, customer satisfaction ratings, etc. each year; skill assessment data includes professional skill level rating results; and team collaboration data includes team members' evaluations of their contribution and communication and collaboration. Divide the collected data into training set, validation set, and test set, with a ratio of 70%, 15%, and 15%, respectively.

[0082] Construct a convolutional neural network model with the employee's historical work data as input and the initial employee capability feature vector as output. Input the training set data into the convolutional neural network model, calculate the output feature vector through forward propagation, and calculate the difference between the output feature vector and the employee's actual capability level based on the loss function. For example, use mean square error as the loss function. Update the network parameters through the back propagation algorithm to minimize the loss function. Use the validation set data to evaluate the generalization performance of the model and adjust hyperparameters such as learning rate, convolution kernel size, etc. Input all employee historical work data into the trained convolutional neural network model to obtain the initial employee capability feature vector of each employee.

[0083] Construct a diffusion probability model. Use the initial employee capability feature vector as the node feature, and construct edges between nodes based on the organizational structure and project cooperation information. For example, if two employees are in the same department or have worked together on a project, establish an edge between them. Calculate the similarity between feature vectors of different nodes, such as using cosine similarity. Define a transition probability matrix based on the similarity, where each element in the matrix represents the probability of transferring from one node to another. Randomly select a node as the starting node, perform random walk sampling according to the transition probability matrix, and record the sequence of visited nodes to form a sampling path. Aggregate the average value of the node feature vectors in the sampling path. Repeat the random walk sampling multiple times to obtain the distribution of employee capability feature vectors after diffusion.

[0084] Perform Markov chain sampling. Use the diffused employee capability feature vector distribution as the initial distribution of Markov chain sampling. Calculate the conditional probability distribution between feature vectors and define the Markov chain transfer matrix based on the conditional probability distribution. Sample the initial state from the initial distribution and iterate the state transfer according to the Markov chain transfer matrix to form a Markov chain. Repeat Markov chain sampling multiple times to obtain a state sequence. Aggregate the average value of the feature vectors in the state sequence and output the optimized employee capability feature vector.

[0085] A cross-modal contrastive learning network is constructed, which contains three feature encoders, which are used to process the optimized employee capability feature vector, employee capability development trend vector and team collaboration data respectively. The employee capability development trend vector is extracted by inputting work performance data and skill assessment data into the spatiotemporal hybrid network model through convolution operation. The three modal data are mapped to the common feature space through their respective feature encoders. The dynamic temperature parameter is introduced, and the contrast loss between similar sample pairs and dissimilar sample pairs is calculated based on the dynamic temperature parameter. The feature encoder and dynamic temperature parameter are optimized to align the features of different modalities in the common space. The three feature vectors are spliced ​​to obtain the comprehensive employee capability vector.

[0086] Construct a feature enhancement network, including a multi-scale feature pyramid and an adaptive feature aggregation module. Input the employee's comprehensive ability vector to the bottom layer of the feature pyramid, and generate feature maps of different scales through convolution and pooling operations. Calculate the feature aggregation weight through the attention mechanism, and use the adaptive feature aggregation module to perform weighted fusion of feature maps of different scales. Map the fused multi-scale features to the predetermined range through the fully connected layer, and output the employee's development potential index.

[0087] In this embodiment, the employee's historical work data, ability development trends and team collaboration information are comprehensively considered, which can more comprehensively and accurately evaluate the employee's development potential. Through convolutional neural networks, diffusion probability models, Markov chain sampling and other technologies, the interference of human factors is reduced, the objectivity of the evaluation is improved, and an automated evaluation process is realized, which greatly improves the evaluation efficiency and reduces the evaluation cost.

[0088] In an optional embodiment,

[0089] The contrast loss between similar sample pairs and dissimilar sample pairs is calculated based on the dynamic temperature parameter as shown in the following formula:

[0090] ;

[0091] Among them, L represents the contrast loss, f(x i ) represents the sample x i The eigenvector of f(x j ) represents the sample x j , r represents the dynamic temperature parameter, sim() represents the similarity metric function, f(x k ) represents the sample x k The feature vector of , k represents the sample index.

[0092] Among them, the sample x i and sample x jrepresents a positive sample, i and j both represent the positive sample index and the values ​​of i and j are different. The minimum value of i and j is 1 and the maximum value is the size of the current batch. Sample x k represents a negative sample, k represents the index of the negative sample, so the values ​​of k, i, and j are all different. The minimum value of k is 1, and the maximum value is the size of the current batch.

[0093] In this embodiment, by dynamically adjusting the temperature parameters, the contributions of similar sample pairs and dissimilar sample pairs in the loss function can be better balanced, thereby improving the accuracy of model learning. The dynamic temperature parameters can make the model more robust to noise and outliers. By using dynamic temperature parameters, the convergence speed of the model can be accelerated, thereby reducing training time.

[0094] S2. Collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. Combined with an adaptive graph attention network, hierarchical sampling and dynamic aggregation are performed to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network, and an employee-job matching score matrix is ​​obtained through parameter learning, super network optimization, and knowledge transfer.

[0095] The generative pre-trained language model is a natural language processing model based on unsupervised learning. It learns the potential structure and grammatical knowledge of a language through a large-scale corpus pre-training model. The dynamic heterogeneous graph neural network is a neural network that models heterogeneous graph data (containing different types of nodes and edges) and can dynamically learn the spatiotemporal variation characteristics of each node and edge in the graph. The time-varying topology learning is a technology for learning topological structure changes in time series data, especially for those network structures with time changes, such as social networks, communication networks, etc. The graph structure distillation is a model compression method that extracts and learns more concise and efficient graph structure representations from complex graph neural networks through distillation technology. The neural architecture search network is a technology for automatically designing neural network architectures. Through search space and search strategy, the neural architecture search network can automatically explore the optimal network structure without human intervention.

[0096] In an optional embodiment,

[0097] Collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. Combined with an adaptive graph attention network, hierarchical sampling and dynamic aggregation are performed to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network. Through parameter learning, super network optimization, and knowledge transfer, the employee-job matching score matrix is ​​obtained, including:

[0098] Collect job description data in the human resources system and historical recruitment data in the recruitment system, perform data cleaning operations to remove redundant information and outliers in the job description data and the historical recruitment data, perform data standardization processing to construct a standardized job description corpus and a historical recruitment corpus, input the job description corpus and the historical recruitment corpus into a pre-constructed generative pre-trained language model, call a named entity recognition module to identify skill entities and knowledge entities in the job description corpus and the historical recruitment corpus, call a relationship extraction module to extract the association relationship between the skill entity and the knowledge entity and the job, and establish a job skill knowledge graph with the job as the central node and the skill entity and the knowledge entity as the attribute node;

[0099] Construct a dynamic heterogeneous graph neural network model, input the job skill knowledge graph into the dynamic heterogeneous graph neural network model, calculate the connection weights between nodes based on a time-varying topology learning strategy, dynamically update the graph structure according to the connection weights, construct a heterogeneous subgraph sampler, input the node type information and edge type information of the job skill knowledge graph, generate multiple local subgraphs based on node semantic similarity sampling, perform graph structure distillation operations on the local subgraphs, map job node features and skill node features to a low-dimensional feature space, and fuse the job node features and the skill node features to generate a job skill feature matrix;

[0100] Construct an adaptive graph attention network model, input employee rank information, department information and reporting relationship information, perform hierarchical sampling operations to generate primary team subgraphs, intermediate team subgraphs and advanced team subgraphs; calculate the attention weights of nodes in the primary team subgraphs, the intermediate team subgraphs and the advanced team subgraphs, fuse the feature information of team members at different levels based on the attention weights, and output a team structure feature vector;

[0101] A neural architecture search network model is constructed, the job skill feature matrix, the employee comprehensive ability vector and the team structure feature vector are input into the neural architecture search network model, a parameter sharing module is called to perform shared optimization on network parameters of different candidate architectures, a super network optimization module is called to search for the optimal matching model structure, a knowledge transfer module is constructed, historical job matching data and team collaboration data are input into the knowledge transfer module, migration features across jobs and teams are extracted, the migration features are input into the matching model for model training, and an employee-job matching score matrix is ​​output.

[0102] The named entity recognition module is a natural language processing technology used to identify entities with specific meanings (such as names of people, places, dates, etc.) from text. The local subgraph refers to a subset extracted from the overall graph, which is usually used to represent certain specific nodes in the graph and their direct adjacency relationships.

[0103] Collect job description data from the human resources system and historical recruitment data from the recruitment system. For example, job description data includes information such as job title, job description, and job requirements, and historical recruitment data includes information such as applicant resumes, interview evaluations, and hiring results. Clean the collected data to remove redundant information and outliers, such as deleting duplicate job descriptions and applicant resumes that are missing key information. Then, standardize the data, such as unifying the formats of dates in different formats and unifying the salary ranges of different units. Finally, build a standardized job description corpus and historical recruitment corpus.

[0104] Input the processed job description corpus and historical recruitment corpus into the pre-built generative pre-trained language model. For example, use the BERT model as a pre-trained language model. Use the named entity recognition module in the model to identify the skill entities and knowledge entities in the corpus. For example, identify the "Java programming" skill entity from "Proficient in Java programming". At the same time, use the relationship extraction module to extract the association between skill entities and knowledge entities and positions. For example, extract the association between the "Java programming" skill and the "software engineer" position. Finally, establish a job skill knowledge graph with the position as the central node and the skill entity and knowledge entity as the attribute node. For example, the "software engineer" position connects skill nodes such as "Java programming" and "software development".

[0105] Construct a dynamic heterogeneous graph neural network model and input the constructed job skill knowledge graph into the model. Calculate the connection weights between nodes based on the time-varying topology learning strategy. For example, dynamically adjust the connection weights between skill nodes and job nodes according to the co-occurrence frequency and importance of skills. Dynamically update the graph structure according to the calculated connection weights. Construct a heterogeneous subgraph sampler and input the node type information and edge type information of the job skill knowledge graph. For example, input job nodes, skill nodes and the connection relationship between them. Generate multiple local subgraphs based on node semantic similarity sampling. For example, combine job nodes with similar skill requirements and their related skill nodes into a subgraph. Perform graph structure distillation operations on the generated local subgraphs to map job node features and skill node features to a low-dimensional feature space. For example, use a graph convolutional network to extract node features and perform dimensionality reduction. Fusion of job node features and skill node features generates a job skill feature matrix.

[0106] Construct an adaptive graph attention network model. Input employee rank information, department information, and reporting relationship information. For example, input information such as the employee's position level, department, and direct superior. Perform hierarchical sampling operations to generate primary team subgraphs, intermediate team subgraphs, and advanced team subgraphs. For example, grassroots employees in the same department form a primary team subgraph, department managers and grassroots employees form an intermediate team subgraph, and higher-level managers and subordinate department managers form an advanced team subgraph. Calculate the attention weights of nodes in the primary team subgraph, intermediate team subgraph, and advanced team subgraph. For example, calculate the attention weight of an employee based on his or her contribution and influence in the team. Based on the calculated attention weights, fuse the feature information of team members at different levels and output the team structure feature vector.

[0107] Construct a neural architecture search network model. Input the job skill feature matrix, employee comprehensive ability vector, and team structure feature vector into the model. Call the parameter sharing module to optimize the network parameters of different candidate architectures. Call the super network optimization module to search for the optimal matching model structure. Construct a knowledge transfer module and input historical job matching data and team collaboration data into the knowledge transfer module. For example, input successful employee-job matching cases and efficient team collaboration cases in the past. Extract cross-job and cross-team migration features. For example, extract common skill requirements between different jobs and collaboration patterns between different teams. Input the extracted migration features into the matching model for model training. Finally, output the employee-job matching score matrix.

[0108] In this embodiment, by constructing a job skill knowledge graph and a dynamic heterogeneous graph neural network, it is possible to more accurately understand job skill requirements and employee capabilities, thereby improving matching accuracy. Through the adaptive graph attention network model, it is possible to better capture the interactive relationship between team members and the team structure characteristics, thereby optimizing the team structure and improving team effectiveness. Through the neural architecture search network model, it is possible to automatically search for the optimal matching model structure, and combined with knowledge transfer learning, it is possible to achieve intelligent employee-job matching and reduce labor costs.

[0109] In an optional embodiment,

[0110] Constructing a heterogeneous subgraph sampler, inputting the node type information and edge type information of the job skill knowledge graph, generating multiple local subgraphs based on node semantic similarity sampling, performing graph structure distillation operations on the local subgraphs, mapping job node features and skill node features to a low-dimensional feature space, and fusing the job node features and the skill node features to generate a job skill feature matrix, including:

[0111] Construct a heterogeneous subgraph sampler, the heterogeneous subgraph sampler includes a node semantic similarity calculation module, the node semantic similarity calculation module includes a multi-head self-attention layer and a feedforward neural network, the node type information and edge type information of the job skill knowledge graph are input into the heterogeneous subgraph sampler, the dot product similarity between node feature vectors is calculated through the multi-head self-attention layer, and the semantic similarity matrix between nodes is calculated based on the dot product similarity and the node type information;

[0112] Constructing an adaptive sampling module, inputting the semantic similarity matrix between nodes and the node degree distribution information into the adaptive sampling module, increasing the sampling weight of highly central nodes based on the semantic similarity matrix between nodes, reducing the sampling weight of edge nodes, and performing sampling operations according to preset job node ratios and skill node ratios to generate multiple local subgraphs;

[0113] Constructing a graph structure distillation module, the graph structure distillation module includes a dual-path encoder, the dual-path encoder includes a feature encoding path and a structure encoding path, wherein the feature encoding path includes a graph convolutional network, and the structure encoding path includes a graph isomorphic network, inputting the local subgraph into the dual-path encoder, extracting node local features through the graph convolutional network, extracting global topological features through the graph isomorphic network, and using a gating mechanism to fuse the node local features and the global topological features to obtain a node feature representation;

[0114] Constructing a contrastive learning module, inputting the node feature representation into the contrastive learning module, constructing the node feature representation in the original feature space and the node feature representation in the low-dimensional embedding space as a positive sample pair, constructing the shuffled node feature representation as a negative sample, performing a feature mapping operation based on the positive sample pair and the negative sample, and mapping the position node feature and the skill node feature to a low-dimensional feature space respectively;

[0115] A feature fusion module is constructed, wherein the feature fusion module includes a node-level fusion layer and a graph-level fusion layer, wherein the node-level fusion layer includes a cross-attention mechanism, and the graph-level fusion layer includes a graph pooling module. The mapped job node features and the skill node features are input into the feature fusion module, and the feature importance weights between the job node features and the skill node features are calculated through the cross-attention mechanism. The node features are aggregated through the graph pooling module to obtain graph-level features. The feature representations at different levels are connected using a residual connection mechanism, and the graph-level features and the feature representations are fused through an adaptive feature aggregation module to generate a job skill feature matrix.

[0116] The dot product similarity is a method for calculating the similarity of two vectors. By calculating the dot product of two vectors, the degree of similarity in the same direction can be measured. The dual-path encoder is a deep learning model architecture that uses two parallel paths (usually two different neural networks or modules) to process different features or information of input data respectively. The positive sample refers to a sample that matches the target category or query condition in a machine learning or information retrieval task, and the negative sample refers to a sample that does not match the target category or query condition in a machine learning or information retrieval task.

[0117] Construct a heterogeneous subgraph sampler, which includes a node semantic similarity calculation module consisting of a multi-head self-attention layer and a feedforward neural network. Input the node type information (e.g., job node, skill node) and edge type information (e.g., have, need) of the job skill knowledge graph into the sampler. The multi-head self-attention layer calculates the dot product similarity between node feature vectors, and combines the node type information to calculate the semantic similarity matrix between nodes. For example, if two nodes are both skill nodes and their feature vectors are very similar, the semantic similarity between them will also be high.

[0118] Construct an adaptive sampling module. Input the semantic similarity matrix between nodes and the node degree distribution information into the module. According to the semantic similarity matrix between nodes, increase the sampling weight of highly central nodes (for example, job nodes with many skills or skill nodes required by many jobs), and reduce the sampling weight of edge nodes. According to the preset job node ratio and skill node ratio (for example, job node: skill node = 1:3), perform sampling operations to generate multiple local subgraphs. For example, around a specific job node, you can sample skill nodes with high semantic similarity and other job nodes related to these skill nodes to form a local subgraph.

[0119] A graph structure distillation module is constructed, which contains a dual-path encoder consisting of a feature encoding path and a structure encoding path. The feature encoding path uses a graph convolutional network to extract node local features. The structure encoding path uses a graph isomorphism network to extract global topological features. The local subgraph is input into the dual-path encoder to extract node local features and global topological features respectively. A gating mechanism is used to fuse these two features to obtain a node feature representation. For example, the gating mechanism can dynamically adjust the weights of the local features and global topological features of the node according to their importance.

[0120] Construct a contrastive learning module. Input the node feature representation into the module. Construct the node feature representation in the original feature space and the node feature representation in the low-dimensional embedding space as positive sample pairs. Construct the shuffled node feature representation as negative samples. Perform feature mapping operations based on the positive sample pairs and negative samples to map the position node features and skill node features to the low-dimensional feature space respectively. For example, a fully connected layer can be used to map high-dimensional features to low-dimensional space.

[0121] Construct a feature fusion module, including a node-level fusion layer and a graph-level fusion layer. The node-level fusion layer uses a cross-attention mechanism to calculate the feature importance weights between job node features and skill node features. The graph-level fusion layer uses a graph pooling module to aggregate node features to obtain graph-level features. A residual connection mechanism is used to connect feature representations at different levels. The adaptive feature aggregation module fuses graph-level features and feature representations to generate a job-skill feature matrix. For example, the adaptive feature aggregation module can dynamically adjust the weights of features at different levels according to their importance. Assuming that a local subgraph contains a job node and three skill nodes, after feature extraction and fusion, a vector containing job features and skill features can be obtained. This vector is the job-skill feature representation of the local subgraph. The job-skill feature representations of all local subgraphs are combined to obtain the final job-skill feature matrix.

[0122] In this embodiment, through heterogeneous subgraph sampling, rich local structural information in the job skill knowledge graph can be captured, and through graph structure distillation, complex graph structure information can be compressed into a low-dimensional feature space, thereby improving the expressiveness of the model. The comparative learning module can learn more discriminative node feature representations, thereby enhancing the generalization ability of the model. Mapping high-dimensional features to low-dimensional space can reduce the computational complexity of the model and improve the training efficiency of the model.

[0123] S3. Input the employee-job matching score matrix and the development potential index into the distributed multi-agent reinforcement learning model, construct a state space including the current personnel distribution status and team structure status, construct an action space based on the deployment plan space, construct a reward function based on the job matching score, development potential index and team collaboration data, perform iterative training based on the exploration strategy of intrinsic incentives, optimize the deployment strategy through the hierarchical task decomposition mechanism, generate an initial deployment plan, evaluate the initial deployment plan through the value network prediction module and path evaluation module in the Monte Carlo tree search model, output the deployment execution plan according to the evaluation results, and analyze the employee performance data and team collaboration data after deployment in combination with the causal reasoning network, and adjust the parameters of the multi-agent reinforcement learning model in combination with the meta-strategy optimization algorithm.

[0124] The distributed multi-agent reinforcement learning model is a reinforcement learning method in a multi-agent system, in which multiple agents collaborate or compete in a shared environment, and each agent maximizes its individual or global rewards by learning interaction strategies with other agents and the environment. The intrinsically motivated exploration strategy is a reinforcement learning strategy that promotes exploration by providing intrinsic incentives for agents. The Monte Carlo tree search model is a search algorithm for decision-making problems, commonly used in chess, games, robotics and other fields. It evaluates different nodes in a decision tree by simulating random game processes and selects the optimal action. The path evaluation module is a module for evaluating the performance and effect of an agent or system on a given path. The causal reasoning network is a network based on causal relationship modeling, which aims to infer the behavior or impact of a system by observing and analyzing the causal relationship between variables.

[0125] In an optional embodiment,

[0126] The employee-position matching score matrix and the development potential index are input into a distributed multi-agent reinforcement learning model, a state space including the current personnel distribution status and the team structure status is constructed, an action space based on the deployment plan space is constructed, a reward function is constructed based on the position matching score, the development potential index and the team collaboration data, iterative training is performed based on an exploration strategy based on intrinsic incentives, and the deployment strategy is optimized through a hierarchical task decomposition mechanism to generate an initial deployment plan, the initial deployment plan is evaluated through a value network prediction module and a path evaluation module in a Monte Carlo tree search model, a deployment execution plan is output according to the evaluation result, and the performance data and team collaboration data of the deployed employees are analyzed in combination with a causal reasoning network, and the parameters of the multi-agent reinforcement learning model are adjusted in combination with a meta-strategy optimization algorithm, including:

[0127] Inputting the employee job matching score matrix and the employee development potential index into a distributed multi-agent reinforcement learning model, using a neural network to extract features from the employee job matching score matrix to obtain an employee job matching feature vector, and using a deep learning network to normalize the employee development potential index to obtain a standardized potential index;

[0128] Generate a personnel distribution state vector in the form of a high-dimensional sparse vector, wherein the dimension of the personnel distribution state vector is the product of the number of employees and the number of positions, set the vector element corresponding to the employee's current position to 1, and set the remaining elements to 0, establish a team structure state matrix, the team structure state matrix uses an adjacency matrix to describe the reporting relationship between team members, and combine the personnel distribution state vector and the team structure state matrix to form a state space;

[0129] Generate an employee deployment plan matrix in the form of a two-dimensional matrix, wherein the number of rows of the employee deployment plan matrix is ​​equal to the total number of employees, the number of columns is equal to the total number of positions, the matrix elements of employees assigned to corresponding positions are set to 1, and the matrix elements that are not assigned are set to 0, and the employee deployment plan matrix is ​​used as an action space;

[0130] Calculating the similarity between the employee job matching feature vector and the preset job requirement vector to obtain a matching score, multiplying the standardized potential index by the preset job development coefficient to obtain a potential score, calculating the collaboration density between team members based on the team structure state matrix to obtain a collaboration score, constructing a reward function for the matching score, the potential score and the collaboration score and solving it to obtain a reward value;

[0131] Decompose the overall deployment task into local team deployment subtasks and cross-department deployment subtasks. For the local team deployment subtasks, use a deep reinforcement learning algorithm to train an agent to deploy personnel within the team. For the cross-department deployment subtasks, use a hierarchical reinforcement learning algorithm to train an agent to complete cross-department personnel deployment. Combine the training results of the local team deployment subtasks and the training results of the cross-department deployment subtasks to generate an initial deployment plan.

[0132] The value network prediction module is used to predict the long-term benefits of the initial deployment plan. The improvement value of different deployment decisions on future team performance is calculated based on the graph neural network. The path evaluation module is used to evaluate the deployment path. The Monte Carlo tree search algorithm is used to calculate the cumulative benefits of different deployment plans, and the plan with the highest evaluation value is selected as the optimal deployment plan.

[0133] The employee performance data and team collaboration data after executing the optimal deployment plan are collected, and the employee performance data and the team collaboration data are input into the causal reasoning network. The causal relationship between the deployment decision and the performance and collaboration is analyzed through the structural equation model, a causal relationship diagram is generated, and the key causal path is extracted to obtain the causal analysis results. According to the causal analysis results, the gradient descent algorithm is used to optimize the network parameters in the distributed multi-agent reinforcement learning model, update the feature representation method of the state space, adjust the constraints of the action space, modify the calculation weights of the reward value, and realize the adaptive optimization of the model parameters.

[0134] The high-dimensional sparse vector form is a vector representation method in which data is represented with a high dimension but most of its elements are zero (sparse).

[0135] Collect employee job matching scores and development potential index data. For example, obtain the matching scores of each employee for each job through the evaluation system to form a matrix. For example, employee A's matching score for job 1 is 0.8, and the matching score for job 2 is 0.6, and so on. At the same time, obtain the development potential index of each employee. For example, employee A's potential index is 90 points, employee B's potential index is 85 points, and so on. Input these data into the distributed multi-agent reinforcement learning model.

[0136] Preprocess the input data. Use a neural network to extract the features of the employee job matching score matrix and obtain the job matching feature vector for each employee. For example, the feature vector of employee A may be [0.75, 0.2, 0.05], which means that he is more suitable for a technical position. At the same time, use a deep learning network to normalize the employee development potential index to obtain a standardized potential index. For example, 90 points and 85 points are standardized to 0.9 and 0.85 respectively.

[0137] Construct the state space. Represent the personnel distribution state as a high-dimensional sparse vector whose dimension is the number of employees multiplied by the number of positions. If an employee is in a certain position, the corresponding vector element value is 1, otherwise it is 0. For example, if there are 3 employees and 3 positions, and employee A is in position 1, the vector is [1, 0, 0, 0, 0, 0, 0, 0]. At the same time, construct the team structure state matrix and use the adjacency matrix to describe the reporting relationship between team members. For example, if employee A reports to employee B, the element value corresponding to A and B in the matrix is ​​1. Combine the personnel distribution state vector and the team structure state matrix to form the state space.

[0138] Construct an action space. Use a two-dimensional matrix to represent the employee deployment plan. The number of rows in the matrix is ​​equal to the total number of employees, and the number of columns is equal to the total number of positions. If an employee is assigned to a certain position, the corresponding matrix element value is 1, otherwise it is 0.

[0139] Construct a reward function. Calculate the similarity between the employee's job matching feature vector and the preset job requirement vector to obtain a matching score. Multiply the standardized potential index by the preset job development coefficient to obtain a potential score. Calculate the collaboration density between team members based on the team structure state matrix to obtain a collaboration score. Perform a weighted sum of the matching score, potential score, and collaboration score to obtain a reward value.

[0140] Conduct model training. Decompose the overall deployment task into local team deployment subtasks and cross-department deployment subtasks. For the local team deployment subtask, use the deep reinforcement learning algorithm to train the agent to deploy personnel within the team. For the cross-department deployment subtask, use the hierarchical reinforcement learning algorithm to train the agent to complete cross-department personnel deployment. Combine the training results of the local team deployment subtask and the cross-department deployment subtask to generate an initial deployment plan.

[0141] Evaluate the initial deployment plan. Use the value network prediction module to predict the long-term benefits of the initial deployment plan, for example, predict the team performance improvement value in the next three months. Calculate the impact of different deployment decisions on future team performance based on the graph neural network. Use the path evaluation module to evaluate the deployment path and calculate the cumulative benefits of different deployment plans. Use the Monte Carlo tree search algorithm to select the plan with the highest evaluation value as the optimal deployment plan.

[0142] Execute the optimal deployment plan and optimize the model. Collect employee performance data and team collaboration data after the implementation of the plan. Input these data into the causal reasoning network, analyze the causal relationship between deployment decisions and performance and collaboration, and generate a causal relationship graph. Based on the results of causal analysis, use the gradient descent algorithm to optimize the network parameters in the distributed multi-agent reinforcement learning model, update the feature representation of the state space, adjust the constraints of the action space, modify the calculation weight of the reward value, and realize the adaptive optimization of the model parameters.

[0143] In this embodiment, deployment plans are automatically generated through intelligent algorithms, reducing manual intervention, which helps to improve the efficiency of personnel deployment. Deployment based on multi-dimensional factors such as job matching, development potential, and team collaboration helps to achieve the best configuration of personnel, thereby improving the overall performance of the team. Taking the development potential of employees into consideration helps to provide employees with more suitable positions and development opportunities, thereby promoting their personal growth.

[0144] In an optional embodiment,

[0145] Based on the graph neural network, the improvement value of different deployment decisions on future team performance is calculated. The path evaluation module is used to evaluate the deployment path. The Monte Carlo tree search algorithm is used to calculate the cumulative benefits of different deployment plans. The plan with the highest evaluation value is selected as the optimal deployment plan, including:

[0146] Inputting an employee node feature vector and a position node feature vector into a pre-set graph neural network model, wherein the employee node feature vector includes the employee's skill features and performance data, and the position node feature vector includes the position requirement features and task description features, and constructing a collaboration relationship edge feature and a position matching relationship edge feature based on the employee node feature vector and the position node feature vector;

[0147] A message passing neural network is used to update the node hidden state, a feature aggregation operation is performed on the employee node feature vector and the position node feature vector to generate a node representation vector, and an edge feature aggregation operation is performed on the collaboration relationship edge feature and the position matching relationship edge feature to generate an edge representation vector;

[0148] Constructing a training data set based on pre-acquired historical deployment data, wherein the historical deployment data includes historical deployment plans and team performance indicators, inputting the training data set into the graph neural network model, and optimizing the network parameters of the graph neural network model using a back propagation algorithm;

[0149] Inputting the initial deployment plan into the graph neural network model, calculating the team performance improvement value of each deployment decision based on the node representation vector and the edge representation vector, and generating a decision evaluation vector;

[0150] Constructing a deployment decision tree, taking the initial deployment plan as the root node, generating child nodes based on employee deployment constraints, each child node representing a deployment state, and calculating a performance evaluation score of each child node based on the decision evaluation vector;

[0151] The upper confidence bound algorithm is used to select child nodes. The upper confidence bound value of each child node is calculated according to the average performance evaluation score of the child node, the number of visits to the parent node, the number of visits to the child node, and the preset exploration coefficient. The child node with the largest upper confidence bound value is selected for expansion. The upper confidence bound value adds exploration items based on the average performance evaluation score of the child node. The exploration items increase with the increase of the number of visits to the parent node and decrease with the increase of the number of visits to the child node.

[0152] Performing an expansion operation on the selected child node, generating a new deployment decision based on the employee deployment constraint, and adding the new deployment decision as a new child node to the deployment decision tree;

[0153] The graph neural network model is used to evaluate the performance of the newly added sub-nodes, calculate the team performance improvement value of the newly added sub-nodes, transmit the team performance improvement value back to the root node along the decision path, and update the number of visits and average performance evaluation score of each node on the path;

[0154] The node selection, node expansion, performance evaluation and result return operations are repeatedly performed until a preset search depth is reached, and a decision path with the highest average performance evaluation score is selected in the deployment decision tree as the optimal deployment solution.

[0155] The message passing neural network is a graph neural network model used to process graph structured data. The upper confidence bound algorithm is a strategy for balancing exploration and exploitation, and is often used in reinforcement learning, optimal decision-making, and multi-armed bandit problems. The decision path refers to a series of steps or the order of actions in the decision-making process, and is usually used to describe the specific process that an intelligent agent or algorithm goes through when making a decision.

[0156] Collect employee information, including skills, past performance, etc., to form an employee feature vector. At the same time, collect job information, such as job responsibilities, required skills, etc., to form a job feature vector. In addition, it is necessary to collect collaborative relationship data between team members and matching data between employees and positions to construct collaborative relationship edge features and job matching relationship edge features. For example, Xiao Zhang is good at programming and communication, and has excellent performance. This information constitutes his employee feature vector. The programming engineer position requires proficiency in Java and Python and strong communication skills. This information constitutes the feature vector of the position. Xiao Zhang and Xiao Li often cooperate and have a high degree of tacit cooperation, which constitutes the collaborative relationship edge feature between them. Xiao Zhang's skills are highly matched with the requirements of the programming engineer position, which constitutes the job matching relationship edge feature between them.

[0157] The collected employee feature vectors, job feature vectors, collaboration relationship edge features, and job matching relationship edge features are input into the graph neural network. The graph neural network aggregates these features to generate node representation vectors and edge representation vectors. For example, through the message passing mechanism, Xiao Zhang's node representation vector will incorporate the information of his neighbor nodes (colleagues with whom he collaborates) and the information of the edges connected to him (matching relationships with positions).

[0158] The graph neural network model is trained using historical deployment data, including past deployment plans and corresponding team performance indicators. The model parameters are continuously adjusted through the back-propagation algorithm so that the model can accurately predict the impact of different deployment plans on team performance. For example, after Xiao Zhang was transferred to the programming engineer position in the past, the team performance increased by 10%. This data is input into the model for training, so that the model can learn the positive impact of this deployment relationship on performance.

[0159] After the model training is completed, it is used to evaluate new deployment plans. The initial deployment plan is input into the trained graph neural network model. The model will calculate the improvement value of each deployment decision on team performance based on the node representation vector and edge representation vector, and generate a decision evaluation vector. For example, if Xiao Wang is transferred from the testing position to the development position, the model predicts that the team performance will improve by 5%, and this value is recorded in the decision evaluation vector.

[0160] The initial deployment plan is used as the root node of the decision tree. Different child nodes are generated according to the employee deployment constraints, such as budget constraints, staffing, etc. Each child node represents a possible deployment state. The performance evaluation score of each child node is calculated using the decision evaluation vector. For example, the root node represents the current team staffing, one child node represents transferring Xiao Wang to the development position, and another child node represents transferring Xiao Li to the product manager position.

[0161] The Monte Carlo tree search algorithm is used to select child nodes for expansion using the upper confidence bound algorithm. The algorithm comprehensively considers the average performance evaluation score of the child nodes, the number of visits to the parent node and the child node, and the preset exploration coefficient, and selects the child node with the largest upper confidence bound value for expansion. For example, although a child node currently has a low average performance evaluation score, it has a small number of visits and still has great potential. The upper confidence bound algorithm will give priority to exploring such nodes.

[0162] Expand the selected subnodes to generate new deployment decisions and add these decisions as new subnodes to the decision tree. For example, select the subnode of transferring Xiao Wang to the development position to expand and generate new subnodes, such as transferring Xiao Li to the product manager position or keeping Xiao Li's position unchanged.

[0163] Use the graph neural network model to evaluate the performance of the newly added child nodes, calculate their team performance improvement value, and transmit the results back to the root node to update the number of visits and average performance evaluation scores of each node on the path.

[0164] The node selection, node expansion, performance evaluation and result return operations are repeated until the preset search depth is reached. Finally, the decision path with the highest average performance evaluation score is selected in the decision tree as the optimal deployment solution.

[0165] In this embodiment, the graph neural network model is used to more accurately predict the impact of different deployment plans on team performance, avoiding the limitations of traditional methods that rely on subjective judgment. The Monte Carlo tree search algorithm can search for the optimal deployment plan on a global scale, avoiding local optimal solutions, thereby finding the plan that greatly improves team performance. By taking employee characteristics, job characteristics, and the relationships between them as input, the model can more comprehensively consider the impact of personnel relationships and job matching on team performance, thereby formulating a more reasonable deployment plan.

[0166] Figure 2 FIG. 1 is a schematic diagram of the structure of an employee position matching and deployment system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0167] The first unit is used to collect employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, perform multi-scale feature extraction on the employee historical work data, obtain an initial employee capability feature vector and add it to a preset diffusion probability model, generate an optimized employee capability feature vector in combination with a Markov chain sampling process, add the work performance data and the skill evaluation data to a preset spatiotemporal hybrid network model, extract employee capability development trends through convolution operations, add the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtain an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and add it to a feature enhancement network, and obtain a development potential index in combination with a multi-scale feature pyramid and an adaptive feature aggregation module;

[0168] The second unit is used to collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. The dynamic heterogeneous graph neural network performs hierarchical sampling and dynamic aggregation in combination with an adaptive graph attention network to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network, and an employee-job matching score matrix is ​​obtained through parameter learning, super network optimization, and knowledge transfer.

[0169] The third unit is used to input the employee-job matching score matrix and the development potential index into a distributed multi-agent reinforcement learning model, construct a state space including the current personnel distribution status and team structure status, construct an action space based on the deployment plan space, construct a reward function based on the job matching score, the development potential index and the team collaboration data, perform iterative training based on the exploration strategy of intrinsic incentives, optimize the deployment strategy through a hierarchical task decomposition mechanism, generate an initial deployment plan, evaluate the initial deployment plan through the value network prediction module and the path evaluation module in the Monte Carlo tree search model, output the deployment execution plan according to the evaluation results, analyze the performance data of the deployed employees and the team collaboration data in combination with the causal reasoning network, and adjust the parameters of the multi-agent reinforcement learning model in combination with the meta-strategy optimization algorithm.

[0170] According to a third aspect of the embodiments of the present invention,

[0171] An electronic device is provided, comprising:

[0172] processor;

[0173] a memory for storing processor-executable instructions;

[0174] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0175] A fourth aspect of the embodiments of the present invention is:

[0176] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0177] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An employee position matching and deployment method based on artificial intelligence, characterized in that: include: Collecting employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, performing multi-scale feature extraction on the employee historical work data, obtaining an initial employee capability feature vector and adding it to a pre-set diffusion probability model, generating an optimized employee capability feature vector in combination with a Markov chain sampling process, adding the work performance data and the skill evaluation data to a pre-set spatiotemporal hybrid network model, extracting employee capability development trends through convolution operations, adding the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtaining an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and adding it to a feature enhancement network, and obtaining a development potential index in combination with a multi-scale feature pyramid and an adaptive feature aggregation module; Collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. Combined with an adaptive graph attention network, hierarchical sampling and dynamic aggregation are performed to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network, and an employee-job matching score matrix is ​​obtained through parameter learning, super network optimization, and knowledge transfer. The employee-job matching score matrix and the development potential index are input into a distributed multi-agent reinforcement learning model to construct a state space including the current personnel distribution status and team structure status, and an action space based on the deployment plan space is constructed. A reward function is constructed based on the job matching score, the development potential index and the team collaboration data. An iterative training is performed based on an exploration strategy based on intrinsic incentives, and the deployment strategy is optimized through a hierarchical task decomposition mechanism to generate an initial deployment plan. The initial deployment plan is evaluated through the value network prediction module and the path evaluation module in the Monte Carlo tree search model. A deployment execution plan is output based on the evaluation results, and the performance data of the deployed employees and the team collaboration data are analyzed in combination with a causal reasoning network. The parameters of the multi-agent reinforcement learning model are adjusted in combination with a meta-strategy optimization algorithm.

2. The method according to claim 1, characterized in that Collect employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, perform multi-scale feature extraction on the employee historical work data, obtain an initial employee capability feature vector and add it to a preset diffusion probability model, combine the Markov chain sampling process to generate an optimized employee capability feature vector, add the work performance data and the skill evaluation data to a preset spatiotemporal hybrid network model, extract employee capability development trends through convolution operations, add the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtain an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and add it to a feature enhancement network, and obtain a development potential index by combining a multi-scale feature pyramid and an adaptive feature aggregation module, including: Collecting employee historical work data, wherein the employee historical work data includes work performance data, skill assessment data and team collaboration data, dividing the employee historical work data into a training set, a validation set and a test set, building a convolutional neural network model, inputting the data in the training set into the convolutional neural network model, calculating an output feature vector through forward propagation, calculating the difference between the output feature vector and the employee's actual ability level based on a loss function, updating network parameters through back propagation, using the data in the validation set to evaluate the model generalization performance and adjust hyperparameters, inputting the employee historical work data into the trained convolutional neural network model, and outputting an initial employee ability feature vector; Constructing a graph structure of a diffusion probability model, taking the initial employee capability feature vector as a node feature, constructing edges between nodes based on organizational structure and project cooperation information, calculating the similarity between feature vectors of different nodes, defining a transition probability matrix based on the similarity, randomly selecting a node as a starting node, performing random walk sampling according to the transition probability matrix, recording a sequence of visited nodes to form a sampling path, performing average aggregation on node feature vectors in the sampling path, repeating multiple random walk samplings, and obtaining a distribution of employee capability feature vectors after diffusion; The diffused employee capability feature vector distribution is used as the initial distribution of Markov chain sampling, the conditional probability distribution between feature vectors is calculated, a Markov chain transfer matrix is ​​defined based on the conditional probability distribution, an initial state is sampled from the initial distribution, state transfer is iteratively performed according to the Markov chain transfer matrix to form a Markov chain, Markov chain sampling is repeated multiple times to obtain a state sequence, feature vectors in the state sequence are averaged and aggregated, and an optimized employee capability feature vector is output; Constructing a cross-modal contrastive learning network including three feature encoders, wherein the three feature encoders are used to process the optimized employee capability feature vector, employee capability development trend vector and team collaboration data respectively, mapping the three modal data to a common feature space through the three feature encoders, introducing a dynamic temperature parameter, calculating the contrast loss between similar sample pairs and dissimilar sample pairs based on the dynamic temperature parameter, optimizing the feature encoder and the dynamic temperature parameter, aligning different modal features in the common space, and concatenating the three feature vectors to obtain an employee comprehensive capability vector; Construct a feature enhancement network including a multi-scale feature pyramid and an adaptive feature aggregation module, input the employee's comprehensive ability vector into the bottom layer of the feature pyramid, generate feature maps of different scales through convolution and pooling operations, calculate the feature aggregation weight through the attention mechanism, use the adaptive feature aggregation module to perform weighted fusion on feature maps of different scales, map the fused multi-scale features to a predetermined range through a fully connected layer, and output the employee development potential index.

3. The method according to claim 2, characterized in that The contrast loss between similar sample pairs and dissimilar sample pairs is calculated based on the dynamic temperature parameter as shown in the following formula: ； Among them, L represents the contrast loss, f(x i ) represents the sample x i The eigenvector of f(x j ) represents the sample x j , r represents the dynamic temperature parameter, sim() represents the similarity metric function, f(x k ) represents the sample x k The feature vector of , k represents the sample index.

4. The method according to claim 1, characterized in that Collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. Combined with an adaptive graph attention network, hierarchical sampling and dynamic aggregation are performed to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network. Through parameter learning, super network optimization, and knowledge transfer, the employee-job matching score matrix is ​​obtained, including: Collect job description data in the human resources system and historical recruitment data in the recruitment system, perform data cleaning operations to remove redundant information and outliers in the job description data and the historical recruitment data, perform data standardization processing to construct a standardized job description corpus and a historical recruitment corpus, input the job description corpus and the historical recruitment corpus into a pre-constructed generative pre-trained language model, call a named entity recognition module to identify skill entities and knowledge entities in the job description corpus and the historical recruitment corpus, call a relationship extraction module to extract the association relationship between the skill entity and the knowledge entity and the job, and establish a job skill knowledge graph with the job as the central node and the skill entity and the knowledge entity as the attribute node; Construct a dynamic heterogeneous graph neural network model, input the job skill knowledge graph into the dynamic heterogeneous graph neural network model, calculate the connection weights between nodes based on a time-varying topology learning strategy, dynamically update the graph structure according to the connection weights, construct a heterogeneous subgraph sampler, input the node type information and edge type information of the job skill knowledge graph, generate multiple local subgraphs based on node semantic similarity sampling, perform graph structure distillation operations on the local subgraphs, map job node features and skill node features to a low-dimensional feature space, and fuse the job node features and the skill node features to generate a job skill feature matrix; Construct an adaptive graph attention network model, input employee rank information, department information and reporting relationship information, perform hierarchical sampling operations to generate primary team subgraphs, intermediate team subgraphs and advanced team subgraphs; calculate the attention weights of nodes in the primary team subgraphs, the intermediate team subgraphs and the advanced team subgraphs, fuse the feature information of team members at different levels based on the attention weights, and output a team structure feature vector; A neural architecture search network model is constructed, the job skill feature matrix, the employee comprehensive ability vector and the team structure feature vector are input into the neural architecture search network model, a parameter sharing module is called to perform shared optimization on network parameters of different candidate architectures, a super network optimization module is called to search for the optimal matching model structure, a knowledge transfer module is constructed, historical job matching data and team collaboration data are input into the knowledge transfer module, migration features across jobs and teams are extracted, the migration features are input into the matching model for model training, and an employee-job matching score matrix is ​​output.

5. The method according to claim 4, characterized in that Constructing a heterogeneous subgraph sampler, inputting the node type information and edge type information of the job skill knowledge graph, generating multiple local subgraphs based on node semantic similarity sampling, performing graph structure distillation operations on the local subgraphs, mapping job node features and skill node features to a low-dimensional feature space, and fusing the job node features and the skill node features to generate a job skill feature matrix, including: Construct a heterogeneous subgraph sampler, the heterogeneous subgraph sampler includes a node semantic similarity calculation module, the node semantic similarity calculation module includes a multi-head self-attention layer and a feedforward neural network, the node type information and edge type information of the job skill knowledge graph are input into the heterogeneous subgraph sampler, the dot product similarity between node feature vectors is calculated through the multi-head self-attention layer, and the semantic similarity matrix between nodes is calculated based on the dot product similarity and the node type information; Constructing an adaptive sampling module, inputting the semantic similarity matrix between nodes and the node degree distribution information into the adaptive sampling module, increasing the sampling weight of highly central nodes based on the semantic similarity matrix between nodes, reducing the sampling weight of edge nodes, and performing sampling operations according to preset job node ratios and skill node ratios to generate multiple local subgraphs; Constructing a graph structure distillation module, the graph structure distillation module includes a dual-path encoder, the dual-path encoder includes a feature encoding path and a structure encoding path, wherein the feature encoding path includes a graph convolutional network, and the structure encoding path includes a graph isomorphic network, inputting the local subgraph into the dual-path encoder, extracting node local features through the graph convolutional network, extracting global topological features through the graph isomorphic network, and using a gating mechanism to fuse the node local features and the global topological features to obtain a node feature representation; Constructing a contrastive learning module, inputting the node feature representation into the contrastive learning module, constructing the node feature representation in the original feature space and the node feature representation in the low-dimensional embedding space as a positive sample pair, constructing the shuffled node feature representation as a negative sample, performing a feature mapping operation based on the positive sample pair and the negative sample, and mapping the position node feature and the skill node feature to a low-dimensional feature space respectively; A feature fusion module is constructed, wherein the feature fusion module includes a node-level fusion layer and a graph-level fusion layer, wherein the node-level fusion layer includes a cross-attention mechanism, and the graph-level fusion layer includes a graph pooling module. The mapped job node features and the skill node features are input into the feature fusion module, and the feature importance weights between the job node features and the skill node features are calculated through the cross-attention mechanism. The node features are aggregated through the graph pooling module to obtain graph-level features. The feature representations at different levels are connected using a residual connection mechanism, and the graph-level features and the feature representations are fused through an adaptive feature aggregation module to generate a job skill feature matrix.

6. The method according to claim 1, characterized in that The employee-position matching score matrix and the development potential index are input into a distributed multi-agent reinforcement learning model, a state space including the current personnel distribution status and the team structure status is constructed, an action space based on the deployment plan space is constructed, a reward function is constructed based on the position matching score, the development potential index and the team collaboration data, iterative training is performed based on an exploration strategy based on intrinsic incentives, and the deployment strategy is optimized through a hierarchical task decomposition mechanism to generate an initial deployment plan, the initial deployment plan is evaluated through a value network prediction module and a path evaluation module in a Monte Carlo tree search model, a deployment execution plan is output according to the evaluation result, and the performance data and team collaboration data of the deployed employees are analyzed in combination with a causal reasoning network, and the parameters of the multi-agent reinforcement learning model are adjusted in combination with a meta-strategy optimization algorithm, including: Inputting the employee job matching score matrix and the employee development potential index into a distributed multi-agent reinforcement learning model, using a neural network to extract features from the employee job matching score matrix to obtain an employee job matching feature vector, and using a deep learning network to normalize the employee development potential index to obtain a standardized potential index; Generate a personnel distribution state vector in the form of a high-dimensional sparse vector, wherein the dimension of the personnel distribution state vector is the product of the number of employees and the number of positions, set the vector element corresponding to the employee's current position to 1, and set the remaining elements to 0, establish a team structure state matrix, the team structure state matrix uses an adjacency matrix to describe the reporting relationship between team members, and combine the personnel distribution state vector and the team structure state matrix to form a state space; Generate an employee deployment plan matrix in the form of a two-dimensional matrix, wherein the number of rows of the employee deployment plan matrix is ​​equal to the total number of employees, the number of columns is equal to the total number of positions, the matrix elements of employees assigned to corresponding positions are set to 1, and the matrix elements that are not assigned are set to 0, and the employee deployment plan matrix is ​​used as an action space; Calculating the similarity between the employee job matching feature vector and the preset job requirement vector to obtain a matching score, multiplying the standardized potential index by the preset job development coefficient to obtain a potential score, calculating the collaboration density between team members based on the team structure state matrix to obtain a collaboration score, constructing a reward function for the matching score, the potential score and the collaboration score and solving it to obtain a reward value; Decompose the overall deployment task into local team deployment subtasks and cross-department deployment subtasks. For the local team deployment subtasks, use a deep reinforcement learning algorithm to train an agent to deploy personnel within the team. For the cross-department deployment subtasks, use a hierarchical reinforcement learning algorithm to train an agent to complete cross-department personnel deployment. Combine the training results of the local team deployment subtasks and the training results of the cross-department deployment subtasks to generate an initial deployment plan. The value network prediction module is used to predict the long-term benefits of the initial deployment plan. The improvement value of different deployment decisions on future team performance is calculated based on the graph neural network. The path evaluation module is used to evaluate the deployment path. The Monte Carlo tree search algorithm is used to calculate the cumulative benefits of different deployment plans, and the plan with the highest evaluation value is selected as the optimal deployment plan. The employee performance data and team collaboration data after executing the optimal deployment plan are collected, and the employee performance data and the team collaboration data are input into the causal reasoning network. The causal relationship between the deployment decision and the performance and collaboration is analyzed through the structural equation model, a causal relationship diagram is generated, and the key causal path is extracted to obtain the causal analysis results. According to the causal analysis results, the gradient descent algorithm is used to optimize the network parameters in the distributed multi-agent reinforcement learning model, update the feature representation method of the state space, adjust the constraints of the action space, modify the calculation weights of the reward value, and realize the adaptive optimization of the model parameters.

7. The method according to claim 6, characterized in that Based on the graph neural network, the improvement value of different deployment decisions on future team performance is calculated. The path evaluation module is used to evaluate the deployment path. The Monte Carlo tree search algorithm is used to calculate the cumulative benefits of different deployment plans. The plan with the highest evaluation value is selected as the optimal deployment plan, including: Inputting an employee node feature vector and a position node feature vector into a pre-set graph neural network model, wherein the employee node feature vector includes the employee's skill features and performance data, and the position node feature vector includes the position requirement features and task description features, and constructing a collaboration relationship edge feature and a position matching relationship edge feature based on the employee node feature vector and the position node feature vector; A message passing neural network is used to update the node hidden state, a feature aggregation operation is performed on the employee node feature vector and the position node feature vector to generate a node representation vector, and an edge feature aggregation operation is performed on the collaboration relationship edge feature and the position matching relationship edge feature to generate an edge representation vector; Constructing a training data set based on pre-acquired historical deployment data, wherein the historical deployment data includes historical deployment plans and team performance indicators, inputting the training data set into the graph neural network model, and optimizing the network parameters of the graph neural network model using a back propagation algorithm; Inputting the initial deployment plan into the graph neural network model, calculating the team performance improvement value of each deployment decision based on the node representation vector and the edge representation vector, and generating a decision evaluation vector; Constructing a deployment decision tree, taking the initial deployment plan as the root node, generating child nodes based on employee deployment constraints, each child node representing a deployment state, and calculating a performance evaluation score of each child node based on the decision evaluation vector; The upper confidence bound algorithm is used to select child nodes. The upper confidence bound value of each child node is calculated according to the average performance evaluation score of the child node, the number of visits to the parent node, the number of visits to the child node, and the preset exploration coefficient. The child node with the largest upper confidence bound value is selected for expansion. The upper confidence bound value adds exploration items based on the average performance evaluation score of the child node. The exploration items increase with the increase of the number of visits to the parent node and decrease with the increase of the number of visits to the child node. Performing an expansion operation on the selected child node, generating a new deployment decision based on the employee deployment constraint, and adding the new deployment decision as a new child node to the deployment decision tree; The graph neural network model is used to evaluate the performance of the newly added sub-nodes, calculate the team performance improvement value of the newly added sub-nodes, transmit the team performance improvement value back to the root node along the decision path, and update the number of visits and average performance evaluation score of each node on the path; The node selection, node expansion, performance evaluation and result return operations are repeatedly performed until a preset search depth is reached, and a decision path with the highest average performance evaluation score is selected in the deployment decision tree as the optimal deployment solution.

8. An employee position matching and deployment system based on artificial intelligence, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect employee historical work data, wherein the employee historical work data includes work performance data, skill evaluation data and team collaboration data, perform multi-scale feature extraction on the employee historical work data, obtain an initial employee capability feature vector and add it to a preset diffusion probability model, generate an optimized employee capability feature vector in combination with a Markov chain sampling process, add the work performance data and the skill evaluation data to a preset spatiotemporal hybrid network model, extract employee capability development trends through convolution operations, add the employee capability development trends, the optimized employee capability feature vector and the team collaboration data to a cross-modal contrastive learning network, obtain an employee comprehensive capability vector based on a contrast loss function and a dynamic temperature parameter solution and add it to a feature enhancement network, and obtain a development potential index in combination with a multi-scale feature pyramid and an adaptive feature aggregation module; The second unit is used to collect job description data and historical recruitment data and add them to a pre-built generative pre-trained language model for semantic understanding and knowledge extraction, build a job skill knowledge graph, add the job skill knowledge graph to a dynamic heterogeneous graph neural network, and generate a job skill feature matrix through time-varying topology learning, heterogeneous subgraph sampling, and graph structure distillation. The dynamic heterogeneous graph neural network performs hierarchical sampling and dynamic aggregation in combination with an adaptive graph attention network to obtain a team structure feature vector. The job skill feature matrix, the employee comprehensive ability vector, and the team structure feature vector are added to a neural architecture search network, and an employee-job matching score matrix is ​​obtained through parameter learning, super network optimization, and knowledge transfer. The third unit is used to input the employee-job matching score matrix and the development potential index into a distributed multi-agent reinforcement learning model, construct a state space including the current personnel distribution status and team structure status, construct an action space based on the deployment plan space, construct a reward function based on the job matching score, the development potential index and the team collaboration data, perform iterative training based on the exploration strategy of intrinsic incentives, optimize the deployment strategy through a hierarchical task decomposition mechanism, generate an initial deployment plan, evaluate the initial deployment plan through the value network prediction module and the path evaluation module in the Monte Carlo tree search model, output the deployment execution plan according to the evaluation results, analyze the performance data of the deployed employees and the team collaboration data in combination with the causal reasoning network, and adjust the parameters of the multi-agent reinforcement learning model in combination with the meta-strategy optimization algorithm.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images