An employee training demand intelligent prediction method, system, device and medium

Abstract

The application relates to an employee training demand intelligent prediction method, system, device and medium. The method comprises the following steps: collecting enterprise employee micro and organizational macro time sequence data, constructing a dynamic heterogeneous graph sequence containing employee, skill, organizational target node and corresponding dynamic edge; outputting employee skill state and organizational future skill demand intensity through a double-channel time sequence encoder; constructing an enhanced static graph and identifying a skill gap through heterogeneous graph convolution to generate a skill gap priority list; based on the skill prerequisite relationship, a deep Q network is used to plan a personalized learning path, and a knapsack optimization model is used to complete path combination and resource allocation. Through the dynamic heterogeneous graph modeling of individual skill state and organizational strategic demand, heterogeneous graph convolution feature enhancement, reinforcement learning and path planning, etc., the accuracy of employee training demand prediction, the personalized adaptability of training path and the utilization efficiency of enterprise training resources are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, system, device and medium for intelligent prediction of employee training needs. Background Technology

[0002] With the continuous advancement of enterprise digital transformation and talent system construction, the accurate matching of employee skills training with organizational strategic needs has become crucial for enhancing core competitiveness. Current methods for predicting employee training needs and planning learning paths largely rely on manual experience, static data statistics, and single-dimensional analysis, making it difficult to adapt to complex and ever-changing business scenarios and dynamically evolving skill requirements. Furthermore, existing employee training needs prediction technologies typically process employee skill data in isolation, failing to construct a dynamic heterogeneous graph structure encompassing employee nodes, skill nodes, and organizational goal nodes. This makes it impossible to effectively model the temporal correlation features of multiple types of dynamic edges, such as skill acquisition, employee collaboration, and organizational needs. Moreover, the lack of a dual-channel temporal encoding mechanism makes it difficult to simultaneously analyze the individual employee's skill mastery status and the organization's macro-strategic evolution trend, resulting in insufficient accuracy in predicting future organizational skill needs. In addition, existing employee analysis solutions, when identifying skill gaps, do not utilize heterogeneous graph convolution to deeply enhance node features, failing to comprehensively assess employees' real-time status, historical mastery levels, and future demand intensity for quantitative evaluation, leading to a lack of objectivity and scientific rigor in skill prioritization. Summary of the Invention

[0003] Therefore, it is necessary to provide an intelligent prediction method, system, device, and medium for employee training needs to address the aforementioned technical issues, aiming to improve the accuracy and foresight of employee training needs prediction and enhance the objectivity and quantitative scientific nature of skills gap identification.

[0004] Firstly, this application provides an intelligent prediction method for employee training needs, including:

[0005] We collect micro-level employee data and macro-level organizational data from multiple time steps within a historical time window of the enterprise. We then perform feature preprocessing on the micro-level employee data and macro-level organizational data to generate corresponding time-series feature sequences. Based on these time-series feature sequences, we construct a dynamic heterogeneous graph sequence that includes employee nodes, skill nodes, and organizational goal nodes. Each snapshot of the dynamic heterogeneous graph sequence includes nodes, initial node features, and dynamic edges. The dynamic edges include skill acquisition edges, employee collaboration edges, and organizational demand edges. The organizational demand edges are directed edges between organizational goal nodes and skill nodes.

[0006] The dynamic heterogeneous graph sequence is input into a dual-channel time encoder that includes employee and strategic channels. The employee channel analyzes the employee skill mastery status and outputs the real-time employee skill status vector and the final employee skill mastery matrix. The strategic channel analyzes the skill evolution trend and outputs the future organizational skill demand intensity vector.

[0007] An enhanced static graph is constructed based on the real-time skill status vector of employees, the employee skill mastery matrix, and the intensity vector of future organizational skill demand. Multi-layer heterogeneous graph convolution is performed on the enhanced static graph to obtain an enhanced employee node representation matrix and an enhanced skill node representation matrix. Based on these matrices, the predicted correlation strength between employees and skills is calculated. This predicted correlation strength is compared with the current actual mastery level of employees in the employee skill mastery matrix to identify the skill gaps of each employee. A comprehensive demand score is then calculated, generating a priority list of employee skill gaps.

[0008] Based on a pre-constructed directed graph of skill prerequisites, target skills are selected according to the priority list of employee skill gaps. A deep Q-network is used for reinforcement learning path planning to generate personalized learning paths for employees. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation plan for each employee.

[0009] In one embodiment, the dynamic heterogeneous graph sequence is input into a dual-channel time encoder including an employee channel and a strategy channel. The employee channel is used to analyze the employee's skill mastery status, and the output includes a real-time employee skill state vector and a final employee skill mastery matrix, including:

[0010] S101. Extract the initial features of employee nodes, initial features of skill nodes, and temporal weights of skill acquisition edges at each time step in the dynamic heterogeneous graph sequence. Retrieve the temporal tensor of employee skill acquisition from the temporal feature sequence. The temporal tensor of employee skill acquisition includes the time step of each employee's last verification of each skill and the cumulative number of verifications.

[0011] S102. For each employee node and each skill node, for each time step, combining the temporal weights of the skill acquisition edge and the temporal tensor of the employee skill acquisition, an improved exponential decay function is used to calculate the skill mastery level of each employee at the corresponding time step, and to construct the employee skill mastery matrix.

[0012] S103. For each time step, the initial features of the skill nodes corresponding to each employee before the time step are used as the value vector, and the initial features of the employee nodes corresponding to each employee at the time step are used as the query vector. The time difference encoding between the last verification time step and the current time step for each employee for each skill and the degree of skill mastery are fused together. The attention weight of each skill to the current employee state is calculated through a multi-head attention mechanism.

[0013] S104. Based on the attention weight, the value vector is weighted and aggregated to obtain the real-time skill state vector of the employee at the current time step.

[0014] S105. Repeat S102 to S104 step by step to obtain the real-time skill state vector and employee skill mastery matrix of each employee at each time step. Extract the real-time skill state vector of the last time step as the employee's real-time skill state vector, and extract the employee's skill mastery matrix of the last time step as the final employee skill mastery matrix.

[0015] In one embodiment, skill evolution trends are analyzed through a strategic channel to output a vector of future organizational skill demand intensity, including:

[0016] Extract the feature vectors of organizational target nodes, the edge weights of organizational demand edges, and the feature vectors of skill nodes associated with organizational demand edges at each time step in the dynamic heterogeneous graph sequence, and arrange them in chronological order to form a strategic semantic vector sequence.

[0017] The strategic semantic vector sequence is input into the LSTM encoder for time step encoding to obtain the encoded hidden state at each time step. The encoded hidden state of the last time step is extracted as the global historical evolution feature.

[0018] The decoder hidden state is initialized with global historical evolution features. An attention mechanism is introduced in each decoding step to calculate the attention weight of each decoding step for each encoded hidden state. The encoded hidden states are weighted and summed based on the attention weights to generate a context vector. The context vector and the current decoder hidden state are fused and calculated through the gating mechanism of the LSTM decoder to obtain the updated decoder hidden state.

[0019] In the first decoding step, the context vector is transformed and concatenated with the global historical evolution features to obtain the first concatenated feature. The first concatenated feature is input into the LSTM decoder and combined with the updated decoder hidden state to generate the strategic semantic vector prediction value for the first future time step. In the next decoding step, the context vector is concatenated with the strategic semantic vector prediction value generated in the previous decoding step and input into the LSTM decoder. Combined with the updated decoder hidden state corresponding to the next decoding step, recursive decoding is performed to generate the strategic semantic vector prediction values ​​for each future time step.

[0020] The predicted strategic semantic vector values ​​for each future time step are input into a pre-defined mapping network and converted into the demand probability of each skill through a softmax activation function, resulting in a sequence of organizational skill demand intensity vectors for the future preset time steps. The vector of the first future time step in the organizational skill demand intensity vector sequence is extracted as the future organizational skill demand intensity vector.

[0021] In one embodiment, an enhanced static graph is constructed based on the employee's real-time skill state vector, the employee's skill mastery matrix, and the future organizational skill demand intensity vector. Multi-layer heterogeneous graph convolution is performed on the enhanced static graph to obtain an enhanced employee node representation matrix and an enhanced skill node representation matrix. Based on these matrices, the predicted correlation strength between employees and skills is calculated. This predicted correlation strength is compared with the employee's current actual skill mastery level in the employee's skill mastery matrix to identify each employee's skill gap. A comprehensive demand score is then calculated, generating a priority list of employee skill gaps, including:

[0022] Using the real-time skill status vector of employees as the employee node feature, and the initial features of skill nodes in the dynamic heterogeneous graph sequence as the skill node feature, the future organizational skill demand intensity vector is mapped to the same dimension as the skill node feature through a preset projection matrix to obtain the future organizational target node feature.

[0023] The values ​​in the employee skill mastery matrix are used as the weights of the skill acquisition edges between employee nodes and skill nodes. The temporal weights of employee collaboration edges are extracted from the graph snapshot of the last time step of the dynamic heterogeneous graph sequence and used as the weights of the collaboration edges between employee nodes. The components of the future organizational skill demand intensity vector are used as the weights of the virtual demand edges between the future organizational target nodes and the corresponding skill nodes. An enhanced static graph is constructed by combining the characteristics of the future organizational target nodes.

[0024] The enhanced static graph is subjected to L-layer heterogeneous graph convolution. In each convolution layer, for each node, the features of neighboring nodes connected by different types of edges are aggregated. During the aggregation process, an independent relational transformation matrix is ​​configured for each type of edge, and the contribution weight of neighboring nodes of the same type is calculated through an attention mechanism.

[0025] The features of neighboring nodes are linearly transformed based on the relational transformation matrix to obtain the transformed features. The transformed features and contribution weights are then combined and weighted to sum them up, and the representation vector of the current node is updated.

[0026] After performing L-layer heterogeneous graph convolution, an enhanced employee node representation matrix is ​​constructed based on the final representation vector of each employee node, and an enhanced skill node representation matrix is ​​constructed based on the final representation vector of each skill node.

[0027] Calculate the dot product of each employee node representation in the enhanced employee node representation matrix and each skill node representation in the enhanced skill node representation matrix, and normalize using the sigmoid function to obtain the predicted association strength between employees and skills;

[0028] The predicted correlation strength is subtracted from the employee's current actual mastery level in the employee skill mastery matrix to obtain the skill gap value. Skills with skill gap values ​​greater than the preset gap threshold are selected as potential demand skills for employees.

[0029] For each potential skill in demand, a comprehensive demand score is calculated by combining the skill gap value, the corresponding demand intensity in the future organizational skill demand intensity vector, and the accessibility index from the employee's current skills to the potential skill in demand, calculated based on the directed graph of skill prerequisite relationships. The potential skills of each employee are then sorted in descending order of comprehensive demand score to generate a priority list of employee skill gaps.

[0030] In one embodiment, the nodes of the directed graph of skill prerequisites are the skill entity nodes corresponding to the skill nodes in the dynamic heterogeneous graph, and the edges represent the prerequisite dependencies between skill entity nodes.

[0031] Based on a pre-constructed directed graph of skill prerequisites, target skills are selected according to a priority list of employee skill gaps. A deep Q-network is used for reinforcement learning path planning to generate personalized learning paths for each employee. These personalized learning paths are scored, and based on the comprehensive score and the pre-set total amount of corporate training resources, a knapsack optimization model is used to select the optimal path combination, generating the final recommended learning path and resource allocation plan for each employee, including:

[0032] S201. For each skill entity node in the directed graph of skill prerequisite relationships, associate the corresponding skill node representation in the enhanced skill node representation matrix; for each employee, select the top K skills with the highest comprehensive demand from the employee skill gap priority list as target skills, construct a target skill set, and extract the employee's skill mastery vector from the employee skill mastery matrix as the employee's current skill mastery vector;

[0033] S202. Configure a deep Q-network for each target skill. The state of the deep Q-network is defined as a concatenated vector of the employee's current skill mastery vector, the employee's current corresponding skill node representation, and the target skill node representation. The action is defined as the set of the next learnable skills that satisfy the prior dependency relationship starting from the current skill entity node, and construct the corresponding reward function.

[0034] S203. Initialize the current network parameters of the deep Q network and the parameters of the target network, set the training hyperparameters, and keep the network structure of the target network consistent with the network structure of the deep Q network. The training hyperparameters include the batch sampling size and the target network parameter synchronization period.

[0035] S204. Based on the current skill mastery vector and the directed graph of skill prerequisites, generate the initial state in the training scenario, update the initial state by executing actions in the action set, obtain the next state, and obtain the corresponding reward based on the reward function. Combine the initial state, actions, rewards and the next state into a sample and store the sample in the experience replay pool.

[0036] S205. Randomly sample samples of batch size from the experience replay pool. Calculate the temporal difference error based on the next state Q value output by the target network and the current state Q value output by the current depth Q network. Construct a loss function using the temporal difference error and calculate the loss function value.

[0037] S206. Update the current network parameters of the deep Q network using gradient descent based on the loss function value. Every time the target network parameter synchronization period is reached, copy the current network parameters of the deep Q network to the target network and update the target network parameters.

[0038] S207. Repeat steps S204 to S206 until the loss function value is less than the preset convergence threshold or the number of training iterations reaches the preset maximum number of iterations, to obtain a trained deep Q-network. For each target skill, starting from the initial skill entity node corresponding to the employee's current skill mastery vector, select the action with the largest Q value based on the Q value of each action output by the trained deep Q-network, and gradually advance until the target skill entity node is reached, generating an ordered sequence of skill entity nodes. The sequence of skill entity nodes is used as the personalized learning path for the target skill.

[0039] S208. Construct a scoring function to calculate the comprehensive score of each personalized learning path; take each employee's personalized learning path as an alternative project, calculate the training course cost corresponding to each personalized learning path, and accumulate the resource consumption; construct a knapsack optimization model with the comprehensive score as the benefit, resource consumption as the cost, and the preset total amount of corporate training resources as the constraint; wherein, the scoring function is used to quantify the comprehensive demand for target skills corresponding to the personalized learning path, the path length of the personalized learning path, and the learning speed feasibility factor calculated based on the employee's historical learning data.

[0040] S209. Solve the knapsack optimization model using a preset optimization algorithm and select the optimal path combination that meets the preset conditions. The optimal path combination includes the optimal personalized learning path for each employee. Determine the final recommended learning path for each employee based on the optimal path combination. Combine the resource consumption of each final recommended learning path with the preset total amount of corporate training resources, allocate the resource allocation quota corresponding to each final recommended learning path, and generate a resource allocation plan.

[0041] In one embodiment, the mathematical expression for the accessibility index of an employee's current skills to potential required skills is:

[0042]

[0043]

[0044]

[0045] in, For the first Each employee's current skills to the next Availability indicators for potential skills in demand; For the first The first employee to the The path entropy of a potential skill requirement; This is a numerical stability correction term; For the first From the collection of skills already acquired by each employee to the first A set of effective prerequisite paths for potential required skills; for Any of the prerequisite paths in the list; For the first path Length; For the first path Any skill node in the game; For the first path Remove target skills The remaining set of skill nodes; For the first Each employee took a step in the last historical moment. skills The actual level of mastery; For the first path The probability weights are used to characterize the prior paths. The relative importance of each effective prerequisite path; for Path Any of the preceding paths.

[0046] Secondly, this application also provides an intelligent prediction system for employee training needs, including:

[0047] The data fusion and heterogeneous graph construction module is used to collect micro-level employee data and macro-level organizational data at multiple time steps within the enterprise's historical time window. It performs feature preprocessing on the micro-level employee data and macro-level organizational data to generate corresponding time-series feature sequences. Based on these time-series feature sequences, it constructs a dynamic heterogeneous graph sequence including employee nodes, skill nodes, and organizational target nodes. Each snapshot of the dynamic heterogeneous graph sequence includes nodes, initial node features, and dynamic edges. Dynamic edges include skill acquisition edges, employee collaboration edges, and organizational demand edges. Organizational demand edges are directed edges between organizational target nodes and skill nodes.

[0048] The dual-channel temporal evolution perception module is used to input the dynamic heterogeneous graph sequence into the dual-channel temporal encoder, which includes the employee channel and the strategy channel. The employee channel analyzes the employee skill mastery status and outputs the real-time skill status vector of the employee and the final employee skill mastery matrix. The strategy channel analyzes the skill evolution trend and outputs the future organizational skill demand intensity vector.

[0049] The intelligent skills gap identification module is used to construct an enhanced static graph based on the real-time skill status vector of employees, the employee skill mastery matrix, and the intensity vector of future organizational skill demand. Multi-layer heterogeneous graph convolution is performed on the enhanced static graph to obtain an enhanced employee node representation matrix and an enhanced skill node representation matrix. Based on these matrices, the predicted correlation strength between employees and skills is calculated. This predicted correlation strength is compared with the employee's current actual skill mastery level in the employee skill mastery matrix to identify each employee's skills gap, calculate the comprehensive demand degree, and generate a priority list of employee skills gaps.

[0050] The learning path planning and resource optimization module is used to select target skills based on a pre-built directed graph of skill prerequisites and a priority list of employee skill gaps. It then uses a deep Q-network to plan reinforcement learning paths and generate personalized learning paths for employees. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation plan for each employee.

[0051] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0052] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0053] The aforementioned intelligent prediction method, system, equipment, and medium for employee training needs first integrates micro-level employee data with macro-level organizational data to construct a dynamic heterogeneous graph sequence containing employee-skill-organizational goal nodes, laying a data foundation for subsequent intelligent prediction of training needs. Second, a dual-channel temporal encoder synchronously analyzes employee skill mastery status and organizational strategic evolution trends, outputting real-time employee status vectors and future demand intensity vectors, overcoming the shortcomings of traditional single-dimensional static analysis. Furthermore, multi-layer heterogeneous graph convolution is performed based on the enhanced static graph to comprehensively calculate employee skill gaps and overall demand, improving the objectivity and quantitative scientific rigor of skill gap identification. Finally, a deep Q-network is used to plan personalized learning paths, combined with a knapsack optimization model for resource allocation, achieving intelligent full-process prediction of training needs and learning path planning, ensuring accurate matching between enterprise talent supply and strategic needs. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart of an intelligent prediction method for employee training needs is provided as an exemplary embodiment of the present invention;

[0056] Figure 2 A flowchart illustrating a method for generating a priority list of employee skills gaps, as provided in an exemplary embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of an intelligent prediction system for employee training needs, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1 As shown, an intelligent prediction method for employee training needs is provided. This embodiment illustrates the method's application to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0060] S101: Collect micro-level employee data and macro-level organizational data at multiple time steps within the enterprise's historical time window, perform feature preprocessing on the micro-level employee data and macro-level organizational data to generate corresponding time-series feature sequences; construct a dynamic heterogeneous graph sequence including employee nodes, skill nodes, and organizational target nodes based on the time-series feature sequences. Each snapshot of the dynamic heterogeneous graph sequence includes nodes, initial node features, and dynamic edges. Dynamic edges include skill acquisition edges, employee collaboration edges, and organizational demand edges; organizational demand edges are directed edges between organizational target nodes and skill nodes.

[0061] Specifically, employee micro-data directly reflects individual skill bases, learning trajectories, and collaboration patterns, while organizational macro-data records corporate strategic orientation and business development needs. Illustratively, multiple consecutive time steps can be divided according to a preset historical time window. Micro-data such as employee static attributes, skill acquisition records, and collaboration records, as well as macro-data such as strategic planning documents and project initiation data, can be systematically collected at each time step. Subsequently, through data cleaning, standardization, and missing value imputation, feature preprocessing operations are performed to eliminate data noise and format differences, generating a time-series feature sequence with a unified structure that can be directly used for modeling. However, since single static data cannot capture the temporal changes in the relationships between elements, a dynamic heterogeneous graph sequence can be constructed based on this sequence. This involves constructing employee nodes, skill nodes, and organizational goal nodes to represent the core analysis objects, and using skill acquisition edges, employee collaboration edges, and organizational need edges to represent the core relationships between nodes. Graph snapshots at each time step record the node status and relationships at that moment. The resulting dynamic heterogeneous graph sequence can completely preserve the temporal evolution information of the relationships between elements, providing structured input for subsequent dual-channel time-series analysis.

[0062] S102: Input the dynamic heterogeneous graph sequence into a dual-channel time encoder that includes employee channel and strategic channel. Analyze employee skill mastery status through the employee channel and output the real-time employee skill status vector and the final employee skill mastery matrix. Analyze skill evolution trends through the strategic channel and output the future organizational skill demand intensity vector.

[0063] Specifically, the employee channel focuses on the dynamic evolution of individual skill mastery status. Based on the node characteristics and correlation weights of each time step in a dynamic heterogeneous graph sequence, it captures the changing patterns of employee skills over time. The output real-time employee skill status vector accurately represents the individual's current skill level, while the final employee skill mastery matrix quantifies the individual's historical mastery of each skill. The strategic channel, on the other hand, focuses on mining the temporal evolution trend of organizational skill needs. By analyzing the feature vectors of organizational target nodes at each time step, it uncovers the changing patterns of strategic orientation and business needs, predicts the types and intensity of skills urgently needed by the company within a set future time window, and outputs a vector of future organizational skill demand intensity. This vector provides a clear basis for subsequent supply and demand matching. Through this parallel processing, the synchronous acquisition of individual status and organizational needs can be achieved, avoiding the misalignment between training needs and corporate development direction caused by single-dimensional analysis, and laying the foundation for subsequent identification of skills gaps.

[0064] S103: Construct an enhanced static graph based on the employee's real-time skill state vector, the employee's skill mastery matrix, and the future organizational skill demand intensity vector; perform multi-layer heterogeneous graph convolution on the enhanced static graph to obtain an enhanced employee node representation matrix and an enhanced skill node representation matrix; calculate the predicted correlation strength between employees and skills based on the enhanced employee node representation matrix and the enhanced skill node representation matrix, compare the predicted correlation strength with the employee's current actual mastery level in the employee skill mastery matrix, identify the skill gap of each employee, calculate the comprehensive demand degree, and generate a priority list of employee skill gaps.

[0065] Specifically, by integrating three key information categories—individual real-time status, historical mastery level, and future organizational needs—outputted by S102, the direct correlation between employees, skills, and future organizational goals can be supplemented, forming a more complete feature association network and resulting in an enhanced static graph. Subsequently, multi-layer heterogeneous graph convolution is performed on the enhanced static graph. This allows for in-depth analysis of the impact of different types of associations (skill acquisition, employee collaboration, and organizational needs) on node features, enhancing the effectiveness of employee and skill node representations and making node features more reflective of their core attributes in training demand scenarios. In the skills gap identification stage, the predicted correlation strength between enhanced employee node representations and enhanced skill node representations can be calculated to quantify employees' future adaptability to various skills. By comparing this with employees' current actual mastery level, individual skill gaps can be directly identified. The comprehensive demand degree can be calculated by combining multiple dimensions such as the size of the skills gap and the intensity of future organizational needs to avoid the limitations of ranking based on a single indicator. The final generated employee skills gap priority list can objectively reflect the urgency and importance of individual training, providing a precise basis for subsequent personalized learning path planning.

[0066] S104: Based on a pre-built directed graph of skill prerequisites, target skills are selected according to the priority list of employee skill gaps. A deep Q-network is used for reinforcement learning path planning to generate personalized learning paths for employees. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation plan for each employee.

[0067] Specifically, by introducing a directed graph of skill prerequisites, the planned learning path can be ensured to conform to the logic of starting with basic skills and progressing to advanced skills, based on the objective laws of skill learning. This avoids inefficient training caused by infeasible paths. Furthermore, deep Q-networks are used for reinforcement learning path planning. Intelligent algorithms can explore the optimal learning sequence that suits an individual's skill level, dynamically adjusting the learning path based on the employee's current skill status and skill prerequisite constraints to achieve personalized adaptation. Subsequently, personalized learning paths are scored to select high-quality solutions with reasonable path lengths and strong adaptability. By applying a knapsack optimization model, under the constraint of a preset total amount of corporate training resources, the path combination with the highest total benefit is selected through global optimization, using the comprehensive path score as the benefit and path resource consumption as the cost. This not only ensures that each employee receives a suitable training plan but also achieves efficient use of training resources, avoiding waste or uneven distribution. Through this process, the final recommended learning path and resource allocation plan for employees can be output, directly guiding the company to conduct precise training and improve training effectiveness and resource utilization efficiency.

[0068] The aforementioned method effectively addresses the limitations of traditional training needs analysis, such as its single dimension and lack of relevant features, by constructing a dynamic heterogeneous graph sequence containing multiple types of nodes and dynamic edges. Secondly, it employs a dual-channel temporal encoder to simultaneously analyze employee skill status and organizational strategic evolution trends, overcoming the shortcomings of traditional solutions that suffer from disconnected supply and demand forecasting and insufficient accuracy. Furthermore, by using enhanced static graphs and heterogeneous graph convolution to accurately identify and prioritize skill gaps, it further addresses the issues of subjective bias and mismatch in traditional skills assessment. Finally, it generates personalized learning paths based on reinforcement learning and optimizes resource allocation through a knapsack model, achieving dynamic adaptation between training needs and resource allocation, thereby improving the accuracy, personalization, and resource utilization efficiency of employee training.

[0069] In one embodiment, a dynamic heterogeneous graph sequence is input into a dual-channel time encoder, including an employee channel and a strategy channel. The employee channel is used to analyze employee skill mastery status, outputting a real-time employee skill state vector and a final employee skill mastery matrix, including:

[0070] S111. Extract the initial features of employee nodes, initial features of skill nodes, and temporal weights of skill acquisition edges at each time step in the dynamic heterogeneous graph sequence. Retrieve the employee skill acquisition temporal tensor from the temporal feature sequence. The employee skill acquisition temporal tensor includes the last verification time step and the cumulative number of verifications for each employee for each skill.

[0071] Specifically, the initial features of employee nodes encompass both static attributes and dynamic performance-related characteristics, including but not limited to skill certificate acquisition, job suitability scores, historical training participation records, and work task completion quality assessment results. Initial features of skill nodes may include the skill's professional field, skill difficulty level, frequency of skill application scenarios, and skill update cycle. The temporal weights of skill acquisition edges can be calculated based on data such as the employee's actual application frequency of the skill within the corresponding time step, skill assessment pass rate, and the skill's contribution to work tasks, used to quantify the strength of the acquisition correlation between the employee and the skill at that time step. Furthermore, an employee skill acquisition temporal tensor can be retrieved from the preprocessed temporal feature sequence. This tensor has a three-dimensional structure; for example, the first dimension corresponds to the employee's individual ID, the second dimension corresponds to the skill ID, and the third dimension contains two key temporal information: the time step identifier corresponding to each employee's last valid verification use of each skill, and the cumulative number of successful verifications of the skill up to that time step. This tensor allows for precise capture of the temporal patterns of employee skill acquisition and the trajectory of proficiency changes.

[0072] S112. For each employee node and each skill node, for each time step, combining the temporal weights of the skill acquisition edge and the temporal tensor of the employee skill acquisition, an improved exponential decay function is used to calculate the skill mastery level of each employee at the corresponding time step, and an employee skill mastery matrix is ​​constructed.

[0073] Specifically, the improved exponential decay function can simultaneously integrate the effects of time decay and cumulative verification frequency, overcoming the limitations of traditional single-dimensional evaluation. For example, for each employee, each skill, and each time step, we can first obtain the temporal weights of the skill acquisition edges corresponding to that time step, and then extract the employee's last verification time step and cumulative verification count for that skill from the employee's skill acquisition temporal tensor. Substituting this data into the improved exponential decay function:

[0074]

[0075] in, This represents the degree to which the i-th employee has mastered the j-th skill at time step t, and its value ranges from [0,1]. Let be the temporal weight of the skill acquisition edge between employee i and skill j at time step t; As a balancing factor, it is used to adjust the contribution ratio of the time decay term and the cumulative number of verifications term. Its value range is (0,1). The optimal value is determined through experimental verification to balance the influence of time decay and usage frequency. This is the time step corresponding to the current calculation. The time step for employee i's last verification of skill j. The time interval between two verifications represents the degree of skill mastery; the longer the interval, the more significant the decline in skill mastery. This is the time decay constant, used to control the decay rate. Its value is determined based on the business cycle and the forgetting pattern of skills. This represents the cumulative number of times employee i has passed the verification for skill j. This function represents the maximum cumulative number of successful verifications for all skills by all employees. The normalization process converts this cumulative number of verifications into a value within the range [0,1], ensuring consistency with the dimensions of the time decay term. Simultaneously, the function sets two constraints: the cumulative number of verifications for skill j by employee i... This indicates that the employee has never effectively mastered the skill. At the current time step Use time step with the last verification When they are the same, the time decay term takes the value of 1. This accurately reflects the level of skill mastery an employee achieves upon passing skills verification. Based on the above calculation process, by iterating through all employees, skills, and time steps, a dimension can be constructed that represents the total number of employees. The total number of skills is represented by an employee skill mastery matrix, where each element corresponds to a quantitative value of a specific employee's mastery of a specific skill at a specific time step.

[0076] S113. For each time step, the initial features of the skill nodes corresponding to each employee before the time step are used as the value vector, and the initial features of the employee nodes corresponding to each employee at the time step are used as the query vector. The time difference encoding between the last verification time step and the current time step for each employee for each skill and the degree of skill mastery are integrated. The attention weight of each skill to the current employee state is calculated through a multi-head attention mechanism.

[0077] To illustrate, for each time step, the initial features of each employee's node at that time step can be used as a query vector. The query vector dimension is the same as the initial feature dimension of the employee node, representing the employee's current basic state. The initial features of the skill nodes corresponding to all previous time steps are used as a value vector, with the value vector dimension matching the initial feature dimension of the skill nodes, providing historical feature support related to skills. Subsequently, feature fusion processing is performed. First, the time difference between the last verification time step for each employee and the current time step for each skill is calculated. Then, a sine / cosine coding method is used to convert this time difference into a fixed-dimensional time difference encoded vector. This coding method effectively captures the periodicity and relative position information of the time series.

[0078] Subsequently, the time difference encoded vector and the employee's quantified mastery of the skill can be concatenated. A linear transformation is then applied to map the concatenated vector to the same dimension as the query vector and value vector, resulting in a fused feature vector. The query vector, value vector, and fused feature vector are then input into a multi-head attention mechanism module. This module can contain multiple parallel attention heads, each independently calculating its attention weight. Specifically, three different linear transformation matrices are used to map the query vector, value vector, and fused feature vector to obtain query projection vector, value projection vector, and fused projection vector, respectively. The dot product similarity between the query projection vector and the fused projection vector is calculated and normalized using a scaling factor to obtain an initial similarity score. Finally, a softmax function is used to normalize the initial similarity score, yielding the attention weight of each skill for each attention head on the current employee's state. This weight quantifies the influence of different skills on the employee's current state.

[0079] S114. Based on the attention weight, the value vector is weighted and aggregated to obtain the real-time skill state vector of the employee at the current time step.

[0080] Specifically, for each attention head, the aggregated feature vector of that attention head is obtained by element-wise multiplying the attention weight corresponding to each skill with the corresponding value projection vector, and then summing the weighted vectors of all skills. Subsequently, the aggregated feature vectors of all attention heads are concatenated, and a linear transformation is applied to the concatenated vector for dimensional integration and feature optimization, resulting in the real-time skill state vector of the employee at the current time step. This vector integrates multi-dimensional information such as the employee's basic state, historical skill characteristics, time decay effect, and skill influence weights, accurately representing the employee's current comprehensive skill level.

[0081] S115. Repeat S112 to S114 step by step to obtain the real-time skill state vector and employee skill mastery matrix of each employee at each time step. Extract the real-time skill state vector of the last time step as the employee's real-time skill state vector, and extract the employee's skill mastery matrix of the last time step as the final employee skill mastery matrix.

[0082] Through the aforementioned iterative calculations and feature updates at each time step, the dynamic evolution of employees' skill mastery status is accurately captured, ensuring the timeliness and accuracy of the output results. Furthermore, since the last time step of the dynamic heterogeneous graph sequence corresponds to the latest historical data, reflecting employees' latest skill status and mastery level, the real-time skill state vectors of all employees at that time step can be extracted as the real-time skill state vectors required for subsequent analysis. Simultaneously, the employee skill mastery matrix corresponding to that time step can be extracted as the final employee skill mastery matrix, providing a quantitative basis for subsequent skills gap identification.

[0083] In one embodiment, by analyzing skill evolution trends through a strategic channel, a vector of future organizational skill demand intensity is output, including:

[0084] Extract the feature vectors of organizational target nodes, the edge weights of organizational demand edges, and the feature vectors of skill nodes associated with organizational demand edges at each time step in the dynamic heterogeneous graph sequence, and arrange them in chronological order to form a strategic semantic vector sequence.

[0085] The strategic semantic vector sequence is input into the LSTM encoder for time step encoding to obtain the encoded hidden state at each time step. The encoded hidden state of the last time step is extracted as the global historical evolution feature.

[0086] The decoder hidden state is initialized with global historical evolution features. An attention mechanism is introduced in each decoding step to calculate the attention weight of each decoding step for each encoded hidden state. The encoded hidden states are weighted and summed based on the attention weights to generate a context vector. The context vector and the current decoder hidden state are fused and calculated through the gating mechanism of the LSTM decoder to obtain the updated decoder hidden state.

[0087] In the first decoding step, the context vector is transformed and concatenated with the global historical evolution features to obtain the first concatenated feature. The first concatenated feature is input into the LSTM decoder and combined with the updated decoder hidden state to generate the strategic semantic vector prediction value for the first future time step. In the next decoding step, the context vector is concatenated with the strategic semantic vector prediction value generated in the previous decoding step and input into the LSTM decoder. Combined with the updated decoder hidden state corresponding to the next decoding step, recursive decoding is performed to generate the strategic semantic vector prediction values ​​for each future time step.

[0088] The predicted strategic semantic vector values ​​for each future time step are input into a pre-defined mapping network and converted into the demand probability of each skill through a softmax activation function, resulting in a sequence of organizational skill demand intensity vectors for the future preset time steps. The vector of the first future time step in the organizational skill demand intensity vector sequence is extracted as the future organizational skill demand intensity vector.

[0089] Specifically, we can first extract temporal features from the dynamic heterogeneous graph sequence. This involves sequentially reading the graph structure data corresponding to each historical time step and extracting the feature vectors of organizational target nodes, the edge weights of organizational demand edges, and the features of skill nodes associated with those edges at each time step. The organizational target node feature vectors characterize the enterprise's macro-level attributes at the corresponding time step, such as strategic planning, business layout, capacity targets, and development direction. The edge weights of organizational demand edges characterize the urgency of the organization's demand for related skills at the corresponding time step. The skill node features characterize attributes such as skill type, application area, capability level, and iteration cycle. By concatenating and integrating these multi-dimensional features according to the chronological order of historical time steps, a continuous strategic semantic vector sequence can be formed. This sequence can fully preserve the temporal correlation features of organizational skill requirements evolving over time.

[0090] The constructed strategic semantic vector sequence is then input into an LSTM encoder. Leveraging the LSTM encoder's ability to capture long-term dependencies in time-series data, the strategic semantic vector at each time step is encoded sequentially. Through the synergistic action of forget gates, input gates, and output gates, historical time-series information can be filtered, stored, and updated, yielding the encoded hidden state for each time step. The sum of the encoded hidden states from all time steps constitutes the encoded feature set representing the historical evolution of organizational skill requirements. Furthermore, the encoded hidden state corresponding to the last time step can be extracted as the global historical evolution feature. This feature integrates the strategic semantic information from all historical time steps, comprehensively representing the overall evolutionary pattern of organizational skill requirements.

[0091] Specifically, by assigning values ​​to the initial hidden states of the LSTM decoder using global historical evolution features, an attention mechanism can be introduced into each decoding step of the decoder. This dynamically weights the encoded hidden states at each historical time step, calculating the attention score for each encoded hidden state in the current decoding step and processing it with a normalization function to obtain the attention weight. A weighted summation operation is then performed on all encoded hidden states based on this attention weight to generate a context vector guiding the current decoding. Furthermore, the gating structure of the LSTM decoder fuses the context vector with the current decoder hidden state, enabling iterative updates to the decoder's hidden states. This allows the decoder to dynamically focus on historical temporal information that has a crucial impact on predicting future skill demands during the decoding process. Illustratively, the attention weight can be calculated as follows:

[0092]

[0093] in, For the first Attention weights corresponding to each decoding step , , These are the learnable parameter matrix and parameter vector in the attention mechanism. The set of encoded hidden states for all time steps of the encoder. For the first The decoder hidden state of each decoding step. The hyperbolic tangent activation function is used, and softmax is a normalized exponential function used to convert attention scores into attention weights with values ​​between 0 and 1 that sum to 1.

[0094] Furthermore, in the first decoding step, the generated context vector and global historical evolution features can be linearly transformed, concatenated, and then input into the LSTM decoder. Combined with the updated decoder hidden state, the algorithm outputs the strategic semantic vector prediction value for the first future time step. In each subsequent decoding step, the context vector generated in the current decoding step is concatenated with the strategic semantic vector prediction value output from the previous decoding step. The concatenated result is input into the LSTM decoder, and recursive decoding is performed in conjunction with the updated decoder hidden state for the corresponding decoding step, sequentially generating strategic semantic vector prediction values ​​for all future preset time steps. By inputting the strategic semantic vector prediction values ​​for all future time steps into a preset mapping network, which can employ a fully connected layer structure, the high-dimensional strategic semantic vectors are mapped to a dimensional space consistent with the number of skill nodes. The softmax activation function then converts the mapping result into the demand probability corresponding to each skill. This demand probability represents the organization's demand intensity for the corresponding skill, thus yielding a sequence of organizational skill demand intensity vectors for the future preset time steps. The organizational skill demand intensity vector corresponding to the first future time step in the sequence is selected as the final output future organizational skill demand intensity vector, which can be used to characterize the organization's demand for various skills in the near future.

[0095] In one embodiment, such as Figure 2 As shown, a priority list of employee skills gaps can be generated through the following steps:

[0096] S201: Using the real-time skill state vector of employees as the employee node feature, and the initial features of skill nodes in the dynamic heterogeneous graph sequence as the skill node feature, the future organizational skill demand intensity vector is mapped to the same dimension as the skill node feature through a preset projection matrix to obtain the future organizational target node feature.

[0097] S202: Use the values ​​in the employee skill mastery matrix as the weights of the skill acquisition edges between employee nodes and skill nodes. Extract the temporal weights of employee collaboration edges from the graph snapshot of the last time step of the dynamic heterogeneous graph sequence as the weights of collaboration edges between employee nodes. Use each component of the future organizational skill demand intensity vector as the weights of the virtual demand edges between the future organizational target nodes and the corresponding skill nodes. Combine the characteristics of the future organizational target nodes to construct an enhanced static graph.

[0098] S203: Perform L-layer heterogeneous graph convolution on the enhanced static graph. In each convolution, for each node, aggregate the features of neighboring nodes connected by different types of edges. During the aggregation process, configure an independent relational transformation matrix for each type of edge and calculate the contribution weight of neighboring nodes of the same type through an attention mechanism.

[0099] S204: Perform a linear transformation on the features of neighboring nodes based on the relational transformation matrix to obtain the transformed features. Combine the transformed features and contribution weights to perform a weighted summation and update the representation vector of the current node.

[0100] S205: After performing L-layer heterogeneous graph convolution, an enhanced employee node representation matrix is ​​constructed based on the final representation vector of each employee node, and an enhanced skill node representation matrix is ​​constructed based on the final representation vector of each skill node. The dot product of each employee node representation in the enhanced employee node representation matrix and each skill node representation in the enhanced skill node representation matrix is ​​calculated, and the predicted association strength between employees and skills is obtained by normalization using the sigmoid function.

[0101] S206: Subtract the predicted correlation strength from the employee’s current actual mastery level in the employee skill mastery matrix to obtain the skill gap value, and select skills with skill gap values ​​greater than the preset gap threshold as the potential demand skills of employees.

[0102] S207: For each potential skill requirement, combine the skill gap value, the corresponding demand intensity in the future organizational skill demand intensity vector, and the accessibility index from the employee's current skills to the potential skill requirement calculated based on the directed graph of skill prerequisite relationships. Calculate the comprehensive demand degree by weighted summation. Sort each employee's potential skills requirement in descending order of comprehensive demand degree to generate an employee skill gap priority list.

[0103] Specifically, the real-time skill status vector of employees can be used as the initial feature of employee nodes. This vector integrates multi-dimensional information about the employee's current skill status, accurately representing individual skill levels. Furthermore, the initial features of skill nodes in the dynamic heterogeneous graph sequence can be used as the initial features of skill nodes, encompassing attributes such as the skill's professional field, difficulty level, and application scenario. By linearly transforming the vector of future organizational skill demand intensity through a preset projection matrix and mapping it to the same dimensional space as the skill node features, the features of future organizational target nodes can be obtained. These features integrate the organization's future skill demands into the graph structure, providing strategic constraints for subsequent analysis. Regarding edge weight construction, the values ​​in the employee skill mastery matrix can be used as the weights of skill acquisition edges between employee nodes and skill nodes. These weights directly quantify the employee's historical mastery of skills. Additionally, the temporal weights of employee collaboration edges can be extracted from the graph snapshot at the last time step of the dynamic heterogeneous graph sequence and used as the weights of collaboration edges between employee nodes, reflecting the closeness of collaboration relationships among employees. Furthermore, the components of the future organizational skill demand intensity vector can be used as weights for the virtual demand edges between future organizational target nodes and corresponding skill nodes. These weights represent the urgency of the organization's future demand for each skill. Through the construction of the above node features and edge weights, an enhanced static graph containing employee nodes, skill nodes, future organizational target nodes, and three types of edges can be formed. This graph structure simultaneously integrates individual skill status, historical collaboration relationships, and future organizational needs, providing comprehensive input information for subsequent heterogeneous graph convolution.

[0104] Furthermore, an L-layer heterogeneous graph convolution can be performed on the enhanced static graph, where the number of convolutional layers L can be determined experimentally, and L is a positive integer. In each convolutional layer, for each node, the features of neighboring nodes connected by different types of edges are aggregated. Since the semantic meanings of different edge types differ, an independent relational transformation matrix can be configured for each edge type to adapt to the feature transformation requirements of different edge types. Illustratively, an attention mechanism can be introduced during the aggregation process to calculate the contribution weight of neighboring nodes of the same type. This weight is used to quantify the importance of different neighboring nodes to the feature update of the current node, ensuring that the features of important neighbors occupy a higher proportion during the aggregation process. Subsequently, a linear transformation is performed on the neighboring node features based on the relational transformation matrix to obtain the transformed neighboring features. Then, the transformed features are weighted and summed using the attention weights to update the representation vector of the current node. This allows the node representation to integrate feature information from different types of neighbors, improving the expressive power of the features. After completing the L-layer heterogeneous graph convolution, an enhanced employee node representation matrix can be constructed based on the final representation vector of each employee node. Each row of this matrix corresponds to the comprehensive skill status representation of an employee. Similarly, an enhanced skill node representation matrix can be constructed based on the final representation vector of each skill node. Each row of this matrix corresponds to the comprehensive attribute representation of a skill. These two types of matrices provide deep feature support for subsequent supply and demand matching.

[0105] Specifically, a dot product can be performed on the representation vector of each employee node in the enhanced employee node representation matrix and the representation vector of each skill node in the enhanced skill node representation matrix. The dot product result is used to measure the fit potential between employees and skills. Normalizing the dot product result using the sigmoid activation function, mapping it to the [0,1] interval, yields the predicted correlation strength between employees and skills. Subtracting the predicted correlation strength from the employee's current actual skill level in the employee skill mastery matrix yields the skill gap value. This value quantifies the gap between the employee's current skill level and future fit potential. Skills with skill gap values ​​greater than a preset gap threshold are selected as potential skills needed by employees. The preset gap threshold can be dynamically adjusted based on the company's training objectives and skill update frequency.

[0106] Furthermore, when calculating the overall demand for each potential skill, an accessibility indicator from an employee's current skills to the potential skill can be introduced. This indicator is calculated by fusing path entropy and path weight, as shown in the following formula:

[0107]

[0108]

[0109]

[0110] In the above formula, For the first Each employee's current skills to the next The accessibility index of a potential skill is indicated by the value of the index. The higher the value, the more sufficient the basic conditions for employees to learn the skill and the lower the difficulty of learning it. For the first The first employee to the The path entropy of a potential skill requirement is used to quantify the concentration of prerequisite paths. The smaller the entropy value, the clearer the path and the more stable the accessibility. This is a numerical stability correction term used to avoid the computational anomaly of a denominator of 0 when the path entropy is 0; For the first Each employee's existing skill set to the next A set of effective prerequisite paths for potential required skills. An effective prerequisite path refers to an ordered sequence of skill nodes that satisfy the dependency constraints in a directed graph of skill prerequisite relationships. for Any of the prerequisite paths in the list; For the first path The length of the path, i.e., the number of skill nodes contained in the path; For the first path Any skill node in the game; For the first path Remove target skills The remaining set of skill nodes; For the first Each employee took a step in the last historical moment. skills The actual level of mastery, which is taken from the final employee skills mastery matrix; For the first path The probability weights are used to characterize the prior paths. Among all effective prerequisite paths, the shorter paths have a higher probability weight, which aligns with the learning logic of starting with the easier paths and progressing to the more difficult ones. for Path Any other path that is not included in the priority path; This is the sum of the inverses of the lengths of all valid precedent paths, used to normalize path probabilities and ensure that the sum of the probability weights of all paths is 1.

[0111] Furthermore, based on the skills gap value, the corresponding demand intensity and attainability indicators in the future organizational skills demand intensity vector, a comprehensive demand degree can be calculated through weighted summation. The weighting coefficients can be adjusted based on corporate training objectives and business priorities. Finally, by sorting each employee's potential skills in descending order of comprehensive demand degree, an employee skills gap priority list can be generated. This list provides a precise priority basis for subsequent personalized learning path planning, ensuring that training resources are prioritized for skills with higher demand.

[0112] In one embodiment, the nodes of the directed graph of skill prerequisite relationships are skill entity nodes corresponding to skill nodes in the dynamic heterogeneous graph, and the edges represent the prerequisite dependencies between skill entity nodes. Based on the pre-constructed directed graph of skill prerequisite relationships, target skills are selected according to the employee skill gap priority list, and a deep Q-network is used for reinforcement learning path planning to generate personalized learning paths for employees. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation scheme for each employee, which may include:

[0113] S211. For each skill entity node in the directed graph of skill prerequisite relationships, associate the corresponding skill node representation in the enhanced skill node representation matrix. For each employee, select the top K skills with the highest comprehensive demand from the employee skill gap priority list as target skills, construct a target skill set, and extract the employee's skill mastery vector from the employee skill mastery matrix as the employee's current skill mastery vector.

[0114] Specifically, in a directed graph of skill prerequisite relationships, if a directed edge exists from skill entity node A to skill entity node B, it indicates that learning skill B requires prior mastery of skill A. The skill node representation in the enhanced skill node representation matrix associated with each skill entity node in the directed graph of skill prerequisite relationships is denoted by . This representation integrates the semantic attributes and contextual information of the skill, providing feature support for reinforcement learning. For each employee, the top K skills with the highest overall demand are selected from the employee skill gap priority list as target skills, constructing a target skill set. The value of K can be adjusted according to the company's training scale and resource constraints; for example, it is usually set to 3 to 5. The employee's skill mastery vector is extracted from the employee skill mastery matrix as the employee's current skill mastery vector. This vector has the same dimension as the total number of skills, and each element corresponds to the employee's mastery of the corresponding skill, directly reflecting the employee's current skill base.

[0115] S212. Configure a deep Q-network for each target skill. The state of the deep Q-network is defined as a concatenated vector of the employee's current skill mastery vector, the employee's current corresponding skill node representation, and the target skill node representation. The action is defined as the set of the next learnable skills that satisfy the prior dependency relationship starting from the current skill entity node, and construct the corresponding reward function.

[0116] Specifically, the state vector of a deep Q-network integrates employee skill foundation, skill semantic information, and learning objectives, accurately representing the core elements of the learning scenario. An action is defined as the set of next learnable skills that satisfy prerequisite dependencies, starting from the current skill entity node. This means only skills meeting prerequisite relationships are allowed as the next learning target, ensuring path feasibility. The reward function quantifies the merits of each action. For example, a positive reward is given when the selected skill is a direct prerequisite of the target skill and the employee has a high level of mastery of that skill; a negative penalty is given when the path length exceeds a preset threshold; and a large positive reward is given when the target skill is reached, guiding the agent to converge towards the optimal path. The reward function can be... , For state Next action The reward The reward coefficient is based on the level of prior skill mastery. To assess employees' mastery of the skills corresponding to the actions. This is the path length penalty coefficient. This represents the current path length. To achieve the high reward coefficient for the target skill, This is an indicator function; it takes a value of 1 when the skill corresponding to the action is the target skill, and 0 otherwise.

[0117] S213. Initialize the current network parameters of the deep Q network and the parameters of the target network, set the training hyperparameters, and ensure that the network structure of the target network is consistent with that of the deep Q network. The training hyperparameters include the batch sampling size and the target network parameter synchronization period.

[0118] Specifically, the network weights can be initialized using a uniform Xavier distribution, and the network structure of the target network should be consistent with that of the deep Q-network to ensure consistent feature extraction capabilities. Training hyperparameters can include batch size, target network parameter synchronization period, discount factor, and learning rate. The batch size controls the number of samples used in each training iteration, and the target network parameter synchronization period is used to periodically update the target network parameters to prevent oscillations during training.

[0119] S214. Based on the employee's current skill mastery vector and the directed graph of skill prerequisites, generate the initial state in the training scenario, update the initial state by executing actions in the action set, obtain the next state, and obtain the corresponding reward based on the reward function. Combine the initial state, actions, rewards and the next state into a sample and store the sample in the experience replay pool.

[0120] Specifically, the initial state in the training scenario corresponds to the set of skill entity nodes that the employee has already mastered. The initial state is updated by executing actions from the action set to obtain the next state, i.e., transitioning from the current skill node to the next skill node corresponding to the action, and the corresponding reward is obtained based on the reward function. The initial state, actions, rewards, and next state are combined into a sample, which can be stored in an experience replay pool. The experience replay pool is used to store samples generated during training, breaking the temporal correlation between samples and improving training stability.

[0121] S215. Randomly sample samples of batch size from the experience replay pool. Calculate the temporal difference error based on the next state Q-value output by the target network and the current state Q-value output by the current depth Q-network. Construct a loss function using the temporal difference error and calculate the loss function value.

[0122] As an illustration, the formula for calculating timing difference error can be:

[0123]

[0124] in, This refers to timing difference error; This is the current reward; Discount factor; For the target network in state Next action Q value; For the current network state Next action Q value; These are the current network parameters; The target network parameters are defined. A mean squared error loss function is constructed using the temporal difference error, and the loss function value is calculated to guide the network parameter updates.

[0125] S216. Update the current network parameters of the deep Q network using gradient descent based on the loss function value. Every time the target network parameter synchronization period is reached, copy the current network parameters of the deep Q network to the target network and update the target network parameters.

[0126] Specifically, the current network parameters of the deep Q-network can be updated using the Adam gradient descent method based on the loss function value, with the optimization direction being to minimize the loss function value. At each target network parameter synchronization period, the current network parameters of the deep Q-network are copied to the target network, and the target network parameters are updated to ensure that the target network can stably provide a reliable Q-value reference.

[0127] S217. Repeat steps S214 to S216 until the loss function value is less than the preset convergence threshold or the number of training iterations reaches the preset maximum number of iterations, to obtain a trained deep Q-network. For each target skill, starting from the initial skill entity node corresponding to the employee's current skill mastery vector, select the action with the largest Q value based on the Q value of each action output by the trained deep Q-network, and gradually advance until the target skill entity node is reached, generating an ordered sequence of skill entity nodes. The sequence of skill entity nodes is used as the personalized learning path for the target skill.

[0128] Specifically, when the loss function value of the deep Q-network is less than the preset convergence threshold or the number of training iterations reaches the preset maximum number of iterations, the network parameters converge to a stable state, at which point the trained deep Q-network is obtained. For each target skill, an initial skill entity node can be determined based on the employee's current skill mastery vector. This node corresponds to the skill node in the employee's already mastered skill set that has the closest pre-requirement relationship with the target skill, serving as the starting point for path planning. During the path generation phase, a greedy strategy can be used for action selection. That is, for the current state, the action with the largest Q-value among all executable actions output by the deep Q-network is directly selected, and the skill node corresponding to this action is the skill to be learned next. Furthermore, the state transition process follows the constraints of the directed graph of skill pre-requirement relationships, only allowing transitions from the current skill node to its direct successor skill node, ensuring that the learning order conforms to the objective law of learning from basic to advanced skills. Each time an action selection is performed, the corresponding skill node is added to the path sequence, and the current state is updated to the state vector corresponding to that skill node. This process is repeated until the target skill entity node is reached. The resulting ordered sequence of skill entity nodes is the personalized learning path for the target skill. This path not only satisfies the prerequisite dependency constraint, but also achieves comprehensive optimization of multiple objectives such as learning efficiency and difficulty adaptability through the Q-value maximization selection of deep Q-networks, ensuring the adaptability of the path to individual employees.

[0129] S218. Construct a scoring function to calculate the comprehensive score of each personalized learning path; take each employee's personalized learning path as an alternative project, calculate the training course cost corresponding to each personalized learning path, and accumulate the resource consumption. With the comprehensive score as the benefit, resource consumption as the cost, and the preset total amount of corporate training resources as the constraint, construct a knapsack optimization model; wherein, the scoring function is used to quantify the comprehensive demand for target skills corresponding to the personalized learning path, the path length of the personalized learning path, and the learning speed feasibility factor calculated based on the employee's historical learning data.

[0130] Specifically, the scoring function is used to quantify the comprehensive value of personalized learning paths. This function integrates three core dimensions: the comprehensive demand for target skills, path length, and the feasibility factor of employee learning speed. This allows for... , For the first Path for each employee Overall score For path Corresponding target skills The overall demand level is indicated by this value; a higher value indicates that the skill is more important to employees and the organization. For path The length of the path, i.e., the number of skill nodes it contains, is used. A longer path results in a higher learning cost; therefore, negative weights are employed. For the first The feasibility factor for an employee's learning speed is calculated based on the employee's historical learning data, such as the average mastery speed of similar skills learned in the past and the completion rate of training courses. The faster the speed and the higher the completion rate, the better. The larger the value, the more adaptable the employee is to the path. , , For the weighting coefficients, satisfying Its value can be dynamically adjusted according to the company's training objectives, such as increasing it when focusing on organizational needs. When focusing on improving learning efficiency .

[0131] Subsequently, each employee's personalized learning path can be considered as an alternative project. The cost of training courses for each path can be calculated, including the training course duration for the corresponding skill, instructor fees, and material resource consumption. The total resource consumption for the path is obtained by summing the course costs for all skills in the path. Using comprehensive scores as the benefit, resource consumption as the cost, and a preset total amount of corporate training resources as constraints, a knapsack optimization model can be constructed. This model, through global optimization, can maximize training benefits under resource constraints, thus avoiding the problem of overall resource waste due to optimal paths for a single employee, ensuring the fairness and efficiency of training resource allocation.

[0132] S219. Solve the knapsack optimization model using a preset optimization algorithm and select the optimal path combination that meets the preset conditions. The optimal path combination includes the optimal personalized learning path for each employee. Determine the final recommended learning path for each employee based on the optimal path combination. Combine the resource consumption of each final recommended learning path with the preset total amount of corporate training resources, allocate the resource allocation quota corresponding to each final recommended learning path, and generate a resource allocation plan.

[0133] Indicatively, the knapsack optimization model belongs to the constrained 0-1 knapsack problem, and its objective function and constraints can be expressed as follows:

[0134]

[0135] in, The total number of employees in the company; For the first The collection of all personalized learning paths for each employee; For the first Path for each employee Overall score; For the first Path for each employee Resource consumption; This represents the preset total amount of corporate training resources; For 0-1 decision variables, a value of 1 indicates that the first option is selected. Employee's Path A value of 0 indicates no selection. The core objective of this model is to maximize the sum of the overall scores of all selected paths within the constraint of total resources, while ensuring that each employee is assigned at most one personalized learning path. For solving this model, when the number of employees and paths is relatively small, a dynamic programming algorithm can be used to obtain the globally optimal solution. The state in the dynamic programming can be defined as... This indicates that resource consumption does not exceed The maximum overall score that can be obtained at the initial state The remaining states are initialized to negative infinity. The state transition equation can be:

[0136]

[0137] Iterate through all paths of all employees, update the state array sequentially, and finally... The maximum comprehensive score is used to determine the selected path combination by backtracking the state update process. When the problem scale is large, such as when the number of employees exceeds 10,000 or the total number of paths exceeds 10,000, the time complexity of dynamic programming will increase significantly. Therefore, a greedy algorithm can be used for approximate solution. This involves first sorting all candidate paths from highest to lowest according to the ratio of "comprehensive score / resource consumption," prioritizing the path with the highest ratio, deducting the corresponding resource consumption after each path is selected, and repeating this process until the remaining resources are insufficient to select any path, or all employees have been assigned paths. While the greedy algorithm cannot guarantee a global optimum, it can obtain an approximate optimal solution in polynomial time, meeting the efficiency requirements of large-scale scenarios. After obtaining the solution results of the model, the decision variables can be used to determine the optimal solution. The value of is used to filter the optimal path combination, for example, if Then the path Determined as the number The final recommended learning path for each employee, if all If the resource consumption of all paths for the employee exceeds the remaining resources, then resource allocation can be adjusted or paths can be replanned according to the company's training strategy.

[0138] As an example, when generating a resource allocation plan, you can first calculate the total resource consumption of all selected paths. ,make sure For each selected path The calculation method for its resource allocation quota can be... ,like If no path is selected, resources can be allocated evenly among employees or reserved. In practice, the allocation can be fine-tuned based on training priority. For example, resources can be increased for the skill paths with the highest overall demand, while ensuring that the sum of all allocated resources does not exceed [a certain threshold]. Ultimately, this can lead to a resource allocation scheme that balances efficiency and fairness, enabling the efficient use of training resources.

[0139] Based on the same inventive concept, this application also provides an intelligent prediction system for employee training needs to implement the aforementioned intelligent prediction method for employee training needs. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent prediction system for employee training needs provided below can be found in the limitations of the intelligent prediction method for employee training needs described above, and will not be repeated here.

[0140] In one exemplary embodiment, such as Figure 3 As shown, an intelligent prediction system 300 for employee training needs is provided, comprising:

[0141] The data fusion and heterogeneous graph construction module 301 is used to collect micro-level employee data and macro-level organizational data at multiple time steps within the enterprise's historical time window. It performs feature preprocessing on the micro-level employee data and macro-level organizational data to generate corresponding time-series feature sequences. Based on these time-series feature sequences, it constructs a dynamic heterogeneous graph sequence including employee nodes, skill nodes, and organizational target nodes. Each snapshot of the dynamic heterogeneous graph sequence includes nodes, initial node features, and dynamic edges. Dynamic edges include skill acquisition edges, employee collaboration edges, and organizational demand edges. Organizational demand edges are directed edges between organizational target nodes and skill nodes.

[0142] The dual-channel temporal evolution perception module 302 is used to input the dynamic heterogeneous graph sequence into a dual-channel temporal encoder including the employee channel and the strategic channel. The employee channel analyzes the employee skill mastery status and outputs the real-time employee skill status vector and the final employee skill mastery matrix. The strategic channel analyzes the skill evolution trend and outputs the future organizational skill demand intensity vector.

[0143] The skills gap intelligent identification module 303 is used to construct an enhanced static graph based on the employee's real-time skills status vector, the employee's skills mastery matrix, and the future organizational skills demand intensity vector; perform multi-layer heterogeneous graph convolution on the enhanced static graph to obtain an enhanced employee node representation matrix and an enhanced skills node representation matrix; calculate the predicted correlation strength between employees and skills based on the enhanced employee node representation matrix and the enhanced skills node representation matrix; compare the predicted correlation strength with the employee's current actual mastery level in the employee skills mastery matrix to identify the skills gap of each employee, calculate the comprehensive demand degree, and generate an employee skills gap priority list;

[0144] The learning path planning and resource optimization module 304 is used to select target skills based on a pre-built directed graph of skill prerequisite relationships and a priority list of employee skill gaps. It then uses a deep Q-network to perform reinforcement learning path planning, generating personalized learning paths for employees. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation plan for each employee.

[0145] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the intelligent prediction method for employee training needs according to this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0146] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the intelligent prediction method for employee training needs according to this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0147] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligently predicting employee training needs, characterized in that, The method includes: The system collects micro-level employee data and macro-level organizational data from multiple time steps within a historical time window of the enterprise. It then performs feature preprocessing on the micro-level employee data and macro-level organizational data to generate corresponding time-series feature sequences. Based on these time-series feature sequences, it constructs a dynamic heterogeneous graph sequence including employee nodes, skill nodes, and organizational target nodes. Each snapshot of the dynamic heterogeneous graph sequence includes nodes, initial node features, and dynamic edges. The dynamic edges include skill acquisition edges, employee collaboration edges, and organizational demand edges. The organizational demand edges are directed edges between the organizational target nodes and the skill nodes. The dynamic heterogeneous graph sequence is input into a dual-channel time encoder that includes an employee channel and a strategic channel. The employee channel is used to analyze the employee's skill mastery status and output the real-time employee skill status vector and the final employee skill mastery matrix. The strategic channel is used to analyze the skill evolution trend and output the future organizational skill demand intensity vector. An enhanced static graph is constructed based on the employee's real-time skill state vector, the employee's skill mastery matrix, and the future organizational skill demand intensity vector. Multi-layer heterogeneous graph convolution is performed on the enhanced static graph to obtain an enhanced employee node representation matrix and an enhanced skill node representation matrix. Based on the enhanced employee node representation matrix and the enhanced skill node representation matrix, the predicted correlation strength between employees and skills is calculated. This predicted correlation strength is compared with the employee's current actual skill mastery level in the employee's skill mastery matrix to identify the skill gap for each employee. A comprehensive demand score is then calculated, generating a priority list of employee skill gaps. Based on a pre-constructed directed graph of skill prerequisites, target skills are selected according to the employee skill gap priority list. A deep Q-network is used for reinforcement learning path planning to generate personalized learning paths for each employee. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation scheme for each employee.

2. The method according to claim 1, characterized in that, The step of inputting the dynamic heterogeneous graph sequence into a dual-channel time encoder including an employee channel and a strategy channel, analyzing the employee skill mastery status through the employee channel, and outputting the employee real-time skill status vector and the final employee skill mastery matrix includes: S101. Extract the initial features of employee nodes, initial features of skill nodes, and temporal weights of skill acquisition edges at each time step in the dynamic heterogeneous graph sequence. Retrieve the temporal tensor of employee skill acquisition from the temporal feature sequence. The temporal tensor of employee skill acquisition includes the last verification time step and the cumulative number of verifications for each employee for each skill. S102. For each employee node and each skill node, for each time step, combining the temporal weights of the skill acquisition edge and the temporal tensor of the employee skill acquisition, an improved exponential decay function is used to calculate the skill mastery degree of each employee at the corresponding time step, and the employee skill mastery matrix is ​​constructed. S103. For each time step, the initial features of the skill nodes corresponding to each employee before the time step are used as the value vector, and the initial features of the employee nodes corresponding to each employee at the time step are used as the query vector. The time difference encoding between the last verification time step and the current time step for each employee for each skill and the mastery level of the skill are fused together. The attention weight of each skill to the current employee state is calculated through a multi-head attention mechanism. S104. The value vector is weighted and aggregated based on the attention weight to obtain the real-time skill state vector of the employee at the current time step; S105. Repeat S102 to S104 step by step to obtain the real-time skill state vector and the employee skill mastery matrix of each employee at each time step. Extract the real-time skill state vector of the last time step as the employee's real-time skill state vector, and extract the employee's skill mastery matrix of the last time step as the final employee skill mastery matrix.

3. The method according to claim 1, characterized in that, The process of analyzing skill evolution trends through the strategic channel and outputting a vector of future organizational skill demand intensity includes: The feature vectors of organizational target nodes, the edge weights of organizational demand edges, and the skill node features associated with the organizational demand edges at each time step in the dynamic heterogeneous graph sequence are extracted and arranged in chronological order to form a strategic semantic vector sequence. The strategic semantic vector sequence is input into an LSTM encoder for time-step encoding to obtain the encoded hidden state of each time step. The encoded hidden state of the last time step is extracted as the global historical evolution feature. The decoder hidden state is initialized with the global historical evolution features. An attention mechanism is introduced in each decoding step. The attention weight of each decoding step to each encoded hidden state is calculated. The encoded hidden states are weighted and summed based on the attention weights to generate a context vector. The context vector and the current decoder hidden state are fused and calculated through the gating mechanism of the LSTM decoder to obtain the updated decoder hidden state. In the first decoding step, the context vector and the global historical evolution feature are transformed and concatenated to obtain a first concatenated feature. The first concatenated feature is input into the LSTM decoder, and combined with the updated decoder hidden state, a strategic semantic vector prediction value for the first future time step is generated. In the next decoding step, the context vector and the strategic semantic vector prediction value generated in the previous decoding step are concatenated and input into the LSTM decoder. Combined with the updated decoder hidden state corresponding to the next decoding step, recursive decoding is performed to generate strategic semantic vector prediction values ​​for each future time step. The predicted strategic semantic vector values ​​for each future time step are input into a preset mapping network and converted into the demand probability of each skill through a softmax activation function to obtain a sequence of organizational skill demand intensity vectors for the future preset time step; the vector of the first future time step in the sequence of organizational skill demand intensity vectors is extracted as the future organizational skill demand intensity vector.

4. The method according to claim 1, characterized in that, An enhanced static graph is constructed based on the employee's real-time skill status vector, the employee's skill mastery matrix, and the future organizational skill demand intensity vector. Multi-layer heterogeneous graph convolution is performed on the enhanced static graph to obtain the enhanced employee node representation matrix and the enhanced skill node representation matrix; Based on the enhanced employee node representation matrix and the enhanced skill node representation matrix, the predicted association strength between employees and skills is calculated. This predicted association strength is then compared with the employee's current actual skill mastery level in the employee skill mastery matrix to identify the skill gap for each employee. A comprehensive demand score is calculated, and a priority list of employee skill gaps is generated, including: Using the real-time skill state vector of the employee as the employee node feature, and the initial feature of the skill node in the dynamic heterogeneous graph sequence as the skill node feature, the future organizational skill demand intensity vector is mapped to the same dimension as the skill node feature through a preset projection matrix to obtain the future organizational target node feature. The values ​​in the employee skill mastery matrix are used as the weights of the skill acquisition edges between the employee nodes and the skill nodes. The temporal weights of the employee collaboration edges are extracted from the graph snapshot of the last time step of the dynamic heterogeneous graph sequence and used as the weights of the collaboration edges between the employee nodes. Each component of the future organizational skill demand intensity vector is used as the weight of the virtual demand edges between the future organizational target node and the corresponding skill node. The enhanced static graph is constructed by combining the features of the future organizational target node. The enhanced static graph is subjected to L-layer heterogeneous graph convolution. In each convolution, for each node, the features of neighboring nodes connected by different types of edges are aggregated. During the aggregation process, an independent relational transformation matrix is ​​configured for each type of edge, and the contribution weight of neighboring nodes of the same type is calculated through an attention mechanism. The neighbor node features are linearly transformed based on the relational transformation matrix to obtain transformed features. The transformed features and the contribution weights are then combined and weighted to update the representation vector of the current node. After performing the L-layer heterogeneous graph convolution, the enhanced employee node representation matrix is ​​constructed based on the final representation vector of each employee node, and the enhanced skill node representation matrix is ​​constructed based on the final representation vector of each skill node. Calculate the dot product of each employee node representation in the enhanced employee node representation matrix and each skill node representation in the enhanced skill node representation matrix, and normalize using the sigmoid function to obtain the predicted association strength between the employee and the skill; The predicted correlation strength is subtracted from the current actual mastery level of the employee corresponding to the employee skill mastery matrix to obtain the skill gap value. Skills with skill gap values ​​greater than a preset gap threshold are selected as the potential demand skills of the employee. For each potential skill in demand, the comprehensive demand score is calculated by combining the skill gap value, the corresponding demand intensity in the future organizational skill demand intensity vector, and the accessibility index from the employee's current skills to the potential skill in demand, calculated based on the directed graph of skill prerequisite relationships. The potential skills of each employee are then sorted in descending order of the comprehensive demand score to generate a priority list of employee skill gaps.

5. The method according to claim 1, characterized in that, The nodes of the directed graph of skill prerequisites are the skill entity nodes corresponding to the skill nodes in the dynamic heterogeneous graph, and the edges represent the prerequisite dependencies between the skill entity nodes. The pre-built directed graph of skill prerequisite relationships selects target skills according to the employee skill gap priority list, and uses a deep Q-network for reinforcement learning path planning to generate the employee's personalized learning path. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation plan for each employee, including: S201. Associate each skill entity node in the directed graph of skill prerequisites with the corresponding skill node representation in the enhanced skill node representation matrix; for each employee, select the top K skills with the highest overall demand from the employee skill gap priority list as the target skills, construct a target skill set, and extract the employee's skill mastery vector from the employee skill mastery matrix as the employee's current skill mastery vector. S202. Configure a deep Q-network for each target skill. The state of the deep Q-network is defined as a concatenated vector of the employee's current skill mastery vector, the employee's current corresponding skill node representation, and the target skill node representation. The action is defined as the next set of learnable skills that satisfy the prior dependency relationship starting from the current skill entity node, and construct the corresponding reward function. S203. Initialize the current network parameters of the deep Q network and the parameters of the target network, and set the training hyperparameters. The network structure of the target network is consistent with the network structure of the deep Q network. The training hyperparameters include the batch sampling size and the target network parameter synchronization period. S204. Based on the employee's current skill mastery vector and the directed graph of the skill prerequisite relationship, generate an initial state in the training scenario, update the initial state by executing actions in the action set, obtain the next state, obtain the corresponding reward based on the reward function, and form a sample by combining the initial state, the action, the reward and the next state, and store the sample in the experience replay pool. S205. Randomly sample samples of the batch sampling size from the experience replay pool, calculate the temporal difference error based on the next state Q value output by the target network and the current state Q value output by the current depth Q network, construct a loss function based on the temporal difference error and calculate the loss function value. S206. Update the current network parameters of the deep Q network using gradient descent based on the loss function value. Every time the target network parameter synchronization period is reached, copy the current network parameters of the deep Q network to the target network and update the target network parameters. S207. Repeat the operations from S204 to S206 until the loss function value is less than the preset convergence threshold or the number of training iterations reaches the preset maximum number of iterations, to obtain a trained deep Q network; for each target skill, starting from the initial skill entity node corresponding to the employee's current skill mastery vector, select the action with the largest Q value according to the Q value of each action output by the trained deep Q network, and gradually advance until the target skill entity node is reached, generating an ordered sequence of skill entity nodes, and using the sequence of skill entity nodes as the personalized learning path for the target skill; S208. Construct a scoring function to calculate the comprehensive score of each personalized learning path; take each employee's personalized learning path as a candidate project, calculate the training course cost corresponding to each personalized learning path, accumulate the resource consumption, and construct the knapsack optimization model with the comprehensive score as the benefit, the resource consumption as the cost, and the preset total amount of corporate training resources as constraints; wherein, the scoring function is used to quantify the comprehensive demand for the target skills corresponding to the personalized learning path, the path length of the personalized learning path, and the learning speed feasibility factor calculated based on the employee's historical learning data; S209. Solve the knapsack optimization model using a preset optimization algorithm to select the optimal path combination that meets the preset conditions; the optimal path combination includes the optimal personalized learning path corresponding to each employee; determine the final recommended learning path for each employee based on the optimal path combination, and allocate the resource allocation quota corresponding to each final recommended learning path by combining the resource consumption of each final recommended learning path with the preset total amount of corporate training resources, and generate the resource allocation scheme.

6. The method according to claim 4, characterized in that, The mathematical expression for the attainability index of the employee's current skills to the potential required skills is: in, For the first Each employee's current skills to the next The availability index of each potential skill requirement; For the first The first employee to the The path entropy of a potential skill requirement; This is a numerical stability correction term; For the first From the collection of skills already acquired by each employee to the first A set of effective prerequisite paths for potential required skills; for Any of the prerequisite paths in the list; For the first path Length; For the first path Any skill node in the game; For the first path Remove target skills The remaining set of skill nodes; For the first Each employee took a step in the last historical moment. skills The actual level of mastery; For the first path The probability weights are used to characterize the prior paths. The relative importance of each effective prerequisite path; for Path Any of the preceding paths.

7. An intelligent prediction system for employee training needs, characterized in that, The system includes: The data fusion and heterogeneous graph construction module is used to collect micro-level employee data and macro-level organizational data at multiple time steps within a historical time window of the enterprise. It performs feature preprocessing on the micro-level employee data and macro-level organizational data to generate corresponding temporal feature sequences. Based on these temporal feature sequences, it constructs a dynamic heterogeneous graph sequence including employee nodes, skill nodes, and organizational target nodes. Each snapshot of the dynamic heterogeneous graph sequence includes nodes, initial node features, and dynamic edges. The dynamic edges include skill acquisition edges, employee collaboration edges, and organizational demand edges. The organizational demand edges are directed edges between the organizational target nodes and the skill nodes. The dual-channel temporal evolution perception module is used to input the dynamic heterogeneous graph sequence into a dual-channel temporal encoder including an employee channel and a strategic channel. The employee channel analyzes the employee skill mastery status and outputs the real-time employee skill status vector and the final employee skill mastery matrix. The strategic channel analyzes the skill evolution trend and outputs the future organizational skill demand intensity vector. The skills gap intelligent identification module is used to construct an enhanced static graph based on the employee's real-time skills status vector, the employee's skills mastery matrix, and the future organizational skills demand intensity vector; perform multi-layer heterogeneous graph convolution on the enhanced static graph to obtain an enhanced employee node representation matrix and an enhanced skills node representation matrix; calculate the predicted correlation strength between employees and skills based on the enhanced employee node representation matrix and the enhanced skills node representation matrix; compare the predicted correlation strength with the employee's current actual mastery level in the employee skills mastery matrix to identify the skills gap of each employee, calculate the comprehensive demand degree, and generate an employee skills gap priority list; The learning path planning and resource optimization module is used to select target skills based on a pre-built directed graph of skill prerequisite relationships and the employee skill gap priority list, and to perform reinforcement learning path planning using a deep Q-network to generate personalized learning paths for the employees. The personalized learning paths are scored, and based on the comprehensive score and the preset total amount of corporate training resources, the optimal path combination is selected through a knapsack optimization model to generate the final recommended learning path and resource allocation scheme for each employee.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images