Regional talent map construction method based on big data and group algorithm

Abstract

The invention discloses a regional talent map construction method based on big data and a population algorithm. The method comprises the following steps: S1, constructing a standardized talent data set; s2, constructing a dynamic talent network by adopting a depth dynamic graph embedding method; s3, performing path search by combining variation ant colony optimization with near-end strategy optimization, and dynamically adjusting a search range through an adaptive differential variation operator; s4, performing talent flow trend modeling by adopting a time-space diagram neural network based on Transform; s5, dynamically adjusting the matching strategy by adopting a self-adaptive coevolution game; s6, performing multi-objective optimization on the talent supply and demand matching scheme by adopting an MOEA / D algorithm; and S7, constructing a dynamic visual talent map, establishing a spatial topological structure of talent flow by adopting dynamic graph convolution, performing talent group automatic clustering in combination with potential Dirichlet distribution, and outputting a visual talent flow map. According to the method, big data, a group algorithm and a dynamic graph technology are combined, the regional talent flow trend is accurately predicted, and the regional talent map is constructed.

Description

Technical Field

[0001] The present invention relates to the technical field of big data and group algorithm, and in particular to a method for constructing a regional talent map based on big data and group algorithm. Background Art

[0002] With the acceleration of globalization and the rapid development of information technology, the flow of talent has become one of the key factors for countries and regions to promote economic growth, enhance innovation capabilities and competitiveness. Talent is an important production factor to promote social and economic development. The law and characteristics of talent flow between regions directly affect the supply and demand of talent in various places, regional economic development and the allocation of innovation resources. Effectively analyzing and predicting the trend of talent flow within and outside the region and optimizing the matching of talent supply and demand between regions have become important bases for all sectors of society, especially governments, enterprises and universities, in formulating strategies for talent introduction, training and use.

[0003] However, existing regional talent flow analysis and prediction methods are still insufficient in many aspects. Traditional talent flow analysis based on statistical methods usually focuses on retrospective analysis of historical data and lacks dynamic and comprehensive prediction models. Most of these methods use static data analysis models, which makes it difficult to reflect changes in talent flow in real time and cannot accurately capture the complex correlations between different factors. For example, many existing methods rely solely on a single data source (such as recruitment market data or government statistics), ignoring the diversified factors involved in the talent flow process, such as the combined effects of social network interactions, policy changes, industry demand and other factors. At the same time, these methods also fail to fully consider the time dimension, making it difficult to accurately predict and analyze future trends.

[0004] With the continuous development of big data technology and deep learning algorithms, comprehensive analysis and modeling methods based on multi-source heterogeneous data have gradually attracted attention. Existing network analysis methods based on deep learning can solve the limitations of traditional methods to a certain extent, but most studies are still stuck in the construction of theoretical models and the design of algorithmic frameworks, lacking large-scale, multi-dimensional, and multi-level actual data support and application practices. For example, although some studies based on graph neural networks (GNNs) have made certain progress in the fields of social network analysis and talent recommendation, they have problems such as low efficiency and high computational complexity when dealing with dynamic, time-series, and complex data structure regional talent flow problems. Especially in the prediction and optimization of cross-regional and multi-dimensional talent flow, existing technologies cannot provide a global optimal solution and lack the ability to dynamically and real-time optimize talent flow patterns.

[0005] In addition, most of the talent supply and demand matching optimization methods in the existing technology are based on static and single objectives, and fail to fully consider the multi-objective and multi-level characteristics of talent flow. Talent flow not only involves the choice of each individual, but also includes comprehensive optimization problems between regions, such as how to balance the supply and demand of talents between regions, how to improve the efficiency of talent allocation, etc. Although the existing multi-objective optimization algorithms can solve some simple optimization problems, they usually cannot achieve the optimal solution when faced with diverse talent flow patterns. In addition, the existing methods also lack dynamic visualization capabilities based on dynamic graph convolution, and cannot achieve a map that reflects the changes in talent flow in various regions in real time over time.

[0006] In terms of talent flow trend modeling, although the spatiotemporal graph neural network model based on deep learning has strong expressive power in theory and can capture the complex dependencies between space and time dimensions, the existing model faces problems such as high data dimension, information overload and low model training efficiency in practical applications. For large-scale, multi-source heterogeneous data, how to efficiently fuse and extract valuable features, and how to adjust model parameters in real time in a dynamically changing environment are still challenges facing current technology.

[0007] Therefore, how to provide a method for constructing a regional talent map based on big data and group algorithms is an urgent problem that technicians in this field need to solve. Summary of the invention

[0008] One purpose of the present invention is to propose a method for constructing a regional talent map based on big data and group algorithms. The present invention combines deep learning, group algorithms and dynamic graph technology to accurately predict regional talent flow trends, optimize talent supply and demand matching solutions, and generate dynamic visualization talent maps by constructing an optimized talent flow prediction model and spatiotemporal graph neural network. This method not only improves the efficiency of talent allocation and prediction accuracy, but also reflects the dynamic changes of inter-regional talent flow in real time. It has the advantages of high precision, high real-time performance and adaptive optimization, and can effectively guide regional development planning, corporate talent strategy and policy formulation.

[0009] According to an embodiment of the present invention, a method for constructing a regional talent map based on big data and group algorithm includes the following steps:

[0010] S1. Obtain multi-source heterogeneous talent data, including government statistics, recruitment market, social networks and college graduate data, and pre-process them to build a standardized talent data set;

[0011] S2. Based on the standardized talent dataset, a deep dynamic graph embedding method is used to build a dynamic talent network. Talent individuals are used as network nodes, and causal-based structural learning is used to remove redundant information to generate an optimized dynamic talent network.

[0012] S3. Based on the optimized dynamic talent network, talent flow prediction is performed. The path search is performed by using mutation ant colony optimization combined with proximal strategy optimization, and the search range is dynamically adjusted through the adaptive differential mutation operator to generate a talent flow prediction model.

[0013] S4. Based on the talent flow prediction model, the Transformer-based spatiotemporal graph neural network is used to model the talent flow trend and generate the talent flow trend prediction results;

[0014] S5. Optimize the matching of talent supply and demand based on the prediction results of talent flow trend, dynamically adjust the matching strategy using adaptive co-evolutionary game, and output the talent supply and demand matching plan;

[0015] S6. Use MOEA / D algorithm to perform multi-objective optimization on the talent supply and demand matching scheme to generate the global optimal talent allocation result;

[0016] S7. Based on the globally optimal talent allocation results, a dynamic visual talent map is constructed. Dynamic graph convolution is used to establish the spatial topological structure of talent flow. The talent groups are automatically clustered in combination with the latent Dirichlet distribution to output a visual talent flow map.

[0017] Optionally, the S2 specifically includes:

[0018] S21. Construct an initial talent network G = (V, E, X) based on the standardized talent data set, where V represents the talent node set, E represents the relationship edge set between talents, and X represents the talent feature matrix, and initialize the adjacency matrix;

[0019] S22. Use the deep dynamic graph embedding method to perform temporal modeling on the initial talent network G and construct a dynamic graph G that changes over time. t =(V,E t ,X t ), where E t represents the talent relationship set at time step t, X t Represent the talent characteristics at time step t and generate a dynamic talent network;

[0020] S23. Use the Granger causal analysis method based on causal inference to calculate the causal effect of each influencing factor on talent mobility, select the main influencing factors by setting the contribution threshold τ, and generate an optimized set of influencing factors;

[0021] S24. Based on the optimized set of influencing factors, the causal structure of the dynamic talent network is learned to construct a causal relationship network G. * =(V,E * ,X), where E *represents the edge set after removing redundant relationships based on causal analysis, and uses a structural optimization algorithm to ensure |E * |≤|E| and network connectivity is guaranteed;

[0022] S25. Use the embedding method based on graph neural network to learn the features of talent nodes, generate node embedding, and use the objective function L to obtain the optimized talent network embedding matrix:

[0023]

[0024] Among them, (i,j) represents the connection edge between the i-th talent node and the j-th talent node, σ represents the activation function, and Z i represents the representation of the i-th talent node in the embedding space, Z j represents the representation of the j-th talent node in the embedding space, Z k Represents the representation of the k-th talent node in the embedding space;

[0025] S26. Based on the optimized talent network embedding matrix, calculate the similarity between talents and perform talent clustering, set the number of talent categories to K, use the K-means algorithm to optimize the talent distribution, and finally generate an optimized dynamic talent network.

[0026] Optionally, the S3 specifically includes:

[0027] S31. Based on the optimized dynamic talent network, the talent flow prediction problem is transformed into an optimal path search problem, and the talent flow path is defined as the path from the starting node to the target node;

[0028] S32. In the process of path search, a variant ant colony optimization algorithm is used to simulate multiple ant individuals exploring in the network. The search for high-quality paths is continuously enhanced through the positive feedback mechanism of pheromones. Each ant individual selects a path according to the following fitness function:

[0029]

[0030] Among them, Fit(p) represents the fitness of path p, w i represents the weight of the i-th talent node, d i (p) represents the flow cost of path p at the i-th talent node, exp represents the exponential function, σ i represents the standard deviation of liquidity cost, f j (p) represents the influencing factor of path p on the jth feature, β represents the importance weight of the feature, n represents the total number of talent nodes in the path, and m represents the total number of influencing factors;

[0031] S33, using the proximal strategy optimization algorithm, based on the experience gained in the ant search process, dynamically adjust the search strategy, and update the parameters of the strategy network through back propagation, so that the search process gradually tends to the global optimum:

[0032]

[0033] Among them, θ t+1 represents the parameters of the updated policy network, θ t represents the parameters of the policy network before updating, α represents the learning rate, represents the policy gradient, λ represents the dynamic adjustment factor, Δθ represents the difference between the current strategy and the previous strategy, γ represents the adjustable factor, and C t represents the cumulative return of the current strategy, and E[C] represents the average return of the strategy;

[0034] S34, introduce the adaptive differential mutation operator, dynamically adjust the search range based on individual historical experience and current path search situation, and the mutation operator dynamically changes the step length:

[0035] X new =X old +ΔX adaptive ·(α 1 ·||X r -X best || γ +β 1 ·I(X r ∈P));

[0036] Among them, X new represents the solution after mutation, X old represents the current solution, X r represents a randomly selected individual, X best represents the optimal solution in the current population, ΔX adaptive represents the adaptively adjusted variable step length, γ represents the weight factor of the adjustment distance, and α 1 and β 1 represents the adjustment factor, I(X r ∈P) represents the indicator function, P represents the known high-quality path set, when X r The value is 1 when it belongs to the known high-quality path set P, otherwise it is 0;

[0037] S35. In the process of combining variant ant colony optimization with proximal strategy optimization, the fitness information of the optimal path is used to model the talent flow trend, generate several flow paths, each flow path represents the talent flow pattern between different regions, and select the optimal path output according to the path fitness function;

[0038] S36. Generate a talent flow prediction model according to the selected optimal path, wherein the talent flow prediction model is based on the structure:

[0039]

[0040] Among them, M(t) represents the prediction result at time t, X p (t) represents the influence characteristics of the pth path at time t, Y p (t) represents the historical flow characteristics of the pth path at time t, δ p and p represents the learning parameter, and N represents the number of paths.

[0041] Optionally, the S4 specifically includes:

[0042] S41. According to the talent flow prediction model, a Transformer-based spatiotemporal graph neural network is constructed, wherein the spatiotemporal graph neural network includes an input layer, a spatiotemporal graph convolution layer, a multi-head attention layer, and an output layer, wherein the input layer processes and fuses the spatiotemporal features of different data sources to generate a spatiotemporal feature matrix;

[0043] S42, input the spatiotemporal feature matrix into the spatiotemporal graph convolution layer, perform graph convolution operation, perform convolution processing through the graph convolution kernel, capture the spatiotemporal dependency of talent flow between regions, extract the spatial and temporal patterns of talent flow, and generate a feature vector for each regional node;

[0044] S43, inputting the feature vector of each regional node into the multi-head attention layer, aggregating the feature information of the adjacent nodes of each regional node by weighted summation, and generating a weighted feature representation;

[0045] S44, in the spatiotemporal graph neural network, encoding the spatial position and time information of the regional node into the regional node feature through position encoding, and generating a regional node feature vector with spatiotemporal information;

[0046] S45. In the output layer, the weighted feature representation and the regional node feature vector with spatiotemporal information are fused through the fully connected layer to generate the talent flow trend prediction results for each region in the future time period.

[0047] Optionally, the S5 specifically includes:

[0048] S51. Based on the prediction results of talent flow trend, an adaptive co-evolutionary game model is constructed. The adaptive co-evolutionary game model includes three game subjects, including talents, enterprises and policy makers. Each game subject has a different strategy space, and a supply and demand matching strategy is generated through game interaction.

[0049] S52. Set the fitness function of each game subject. The talent fitness function is set according to the selected flow area and matching results, the enterprise fitness function is set according to the recruitment target and talent matching, and the policy maker fitness function is set according to the set regional development strategy and talent introduction policy;

[0050] S53. Through evolutionary game theory, simulate the game process between game players, where each game player updates its own strategy according to the fitness function, dynamically adjust the talent supply and demand matching strategy through the evolutionary process, and optimize the strategy selection of the game players;

[0051] S54. During the game, according to the interaction between the game players, the step size and update frequency of the strategy selection are adjusted, and the speed of strategy adjustment is controlled through the adaptive learning mechanism;

[0052] S55. Through repeated iterations of the adaptive co-evolutionary game model, a talent supply and demand matching solution is output.

[0053] Optionally, the S6 specifically includes:

[0054] S61, initializing the population of the MOEA / D algorithm by randomly generating multiple initial solutions, each of which represents a talent allocation plan, wherein each talent allocation plan includes the optimal talent allocation ratio, flow path and supply and demand matching information of different regions;

[0055] S62. Define multiple optimization goals, including minimizing the cost of talent mobility, maximizing talent matching efficiency, and maximizing the balance between talent supply and demand among regions, and assign initial weights to each goal;

[0056] S63, using a weighted linear combination method to adjust the weight of each objective, and through an adaptive adjustment strategy, during the optimization process, the weight of each objective is dynamically adjusted according to changes in the current optimization stage:

[0057]

[0058] in, represents the weight of the i-th target in the t+1th generation, represents the weight of the i-th target in the t-th generation, κ represents the weight adjustment factor, and f i (x best ) represents the objective function value of the current optimal solution, f i (x current ) represents the objective function value of the current solution;

[0059] S64. In each iteration, individuals with fitness higher than the set threshold are selected from the current population through selection operations, new solutions are generated through crossover operations, and diversity is introduced through mutation operations. When any of the following conditions is met, the iteration process is terminated and the global optimal talent allocation result is output:

[0060] The first condition is that the number of iterations reaches the preset maximum number of iterations;

[0061] The second condition is that the change between the optimal solution of each generation and the optimal solution of the previous generation is less than a preset threshold.

[0062] Optionally, the S7 specifically includes:

[0063] S71. Based on the global optimal talent allocation result, a dynamic visual talent map is constructed, wherein the dynamic visual talent map includes a plurality of nodes and edges connecting the nodes, wherein the nodes represent various regions, the weights of the edges represent the intensity of talent flow between regions, the size of the nodes represents the number of talents in the region, and the colors represent the talent supply and demand status of the region;

[0064] S72. Establishing a spatial topological structure of talent flow through a dynamic graph convolutional network, wherein the dynamic graph convolutional network is used to model the dynamic changes of talent flow between regions, and capturing the spatiotemporal characteristics of talent flow through convolution operations, generating a spatiotemporal feature vector of each node, wherein the spatiotemporal feature vector contains real-time talent flow information of the region;

[0065] S73. In the dynamic graph convolution network, a time window sliding strategy is used to dynamically update the graph convolution kernel to adapt to the impact of time changes on the talent flow pattern;

[0066] S74, automatically clustering the talent groups by latent Dirichlet distribution, wherein the latent Dirichlet distribution clusters the talent types, flow paths and demand intensity characteristics in the region, and the clustering results reflect the potential trends and group characteristics of talent flow in the region;

[0067] S75. Based on the clustering results, the flow of talents is displayed through visualization methods. The chord diagram technology is used to display the talent flow paths between regions. The chord width in the chord diagram represents the intensity of talent flow, and the color of the chord represents the flow direction.

[0068] S76. Combining the clustering results of the latent Dirichlet distribution and the dynamic visualization of the chord diagram, a visualized talent flow map is generated. The visualized talent flow map is dynamically updated according to time changes to provide real-time information on the trend, direction and intensity of talent flow between regions.

[0069] The beneficial effects of the present invention are:

[0070] First, based on the deep dynamic graph embedding method adopted by the present invention, we can extract valuable information from multiple data sources and build an optimized dynamic talent network. This method can not only effectively capture the complex spatiotemporal dependencies of talent flow, but also eliminate redundant information and improve the accuracy and efficiency of the network model. Compared with traditional static analysis methods, the present invention can dynamically adjust and update the talent network structure according to changes in time, making the prediction results of talent flow more accurate and real-time.

[0071] Secondly, the present invention innovatively solves the path search problem by combining variant ant colony optimization with proximal strategy optimization. By simulating the behavior of multiple ant individuals and combining adaptive differential mutation operators, the present invention can quickly and efficiently search for the optimal talent flow path in a complex dynamic environment. In addition, the ability to dynamically adjust the search strategy and change the step length enables the search process to be continuously optimized as the environment changes, thereby improving the accuracy and adaptability of the prediction model.

[0072] In terms of talent flow trend modeling, the present invention combines a Transformer-based spatiotemporal graph neural network. Compared with traditional methods, the spatiotemporal graph neural network integrates spatial and temporal information and considers the spatiotemporal dependency of talent flow between regions when modeling, thereby better reflecting the dynamics of talent flow between regions. Through the multi-head attention mechanism, the network can automatically learn and focus on the key factors affecting talent flow, making the prediction of talent flow trend more accurate and scientific, and able to reflect the changing trend in real time.

[0073] In addition, the present invention adopts an adaptive co-evolutionary game model in the matching of talent supply and demand, and optimizes the strategic interaction between various game subjects. By simulating the game process between talents, enterprises and policy makers, we can dynamically adjust the matching strategy to ensure the balance of talent supply and demand. Compared with the traditional single-objective optimization method, the present invention can balance multiple objectives and generate the global optimal talent allocation result through a multi-objective optimization algorithm, thereby achieving more efficient talent resource allocation.

[0074] In terms of visualization, the present invention uses a technology that combines dynamic graph convolutional networks with latent Dirichlet distribution, breaking through the limitations of traditional static visualization. By constructing a dynamic visualization talent map and combining the spatial topological structure of talent flow, the present invention can display talent flow trends and flow information in real time, helping decision makers to intuitively understand the pattern, intensity and direction of talent flow between regions. Such a dynamic visualization map not only improves the accuracy of decision-making, but also provides strong support for various talent management policies. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0076] Figure 1 This is an overall flow chart of the method for constructing a regional talent map based on big data and group algorithm proposed in the present invention. DETAILED DESCRIPTION

[0077] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0078] refer to Figure 1 ,The regional talent map construction method based on big data and group algorithm includes the following steps:

[0079] S1. Obtain multi-source heterogeneous talent data, including government statistics, recruitment market, social networks and college graduate data, and pre-process them to build a standardized talent data set;

[0080] S2. Based on the standardized talent dataset, a deep dynamic graph embedding method is used to build a dynamic talent network. Talent individuals are used as network nodes, and causal-based structural learning is used to remove redundant information to generate an optimized dynamic talent network.

[0081] S3. Based on the optimized dynamic talent network, talent flow prediction is performed. The path search is performed by using mutation ant colony optimization combined with proximal strategy optimization, and the search range is dynamically adjusted through the adaptive differential mutation operator to generate a talent flow prediction model.

[0082] S4. Based on the talent flow prediction model, the Transformer-based spatiotemporal graph neural network is used to model the talent flow trend and generate the talent flow trend prediction results;

[0083] S5. Optimize the matching of talent supply and demand based on the prediction results of talent flow trend, dynamically adjust the matching strategy using adaptive co-evolutionary game, and output the talent supply and demand matching plan;

[0084] S6. Use MOEA / D algorithm to perform multi-objective optimization on the talent supply and demand matching scheme to generate the global optimal talent allocation result;

[0085] S7. Based on the globally optimal talent allocation results, a dynamic visual talent map is constructed. Dynamic graph convolution is used to establish the spatial topological structure of talent flow. The talent groups are automatically clustered in combination with the latent Dirichlet distribution to output a visual talent flow map.

[0086] In this implementation, S2 specifically includes:

[0087] S21. Construct an initial talent network G = (V, E, X) based on the standardized talent data set, where V represents the talent node set, E represents the relationship edge set between talents, and X represents the talent feature matrix, and initialize the adjacency matrix;

[0088] S22. Use the deep dynamic graph embedding method to perform temporal modeling on the initial talent network G and construct a dynamic graph G that changes over time. t =(V,E t ,X t ), where E t represents the talent relationship set at time step t, X t Represent the talent characteristics at time step t and generate a dynamic talent network;

[0089] S23. Use the Granger causal analysis method based on causal inference to calculate the causal effect of each influencing factor on talent mobility, select the main influencing factors by setting the contribution threshold τ, and generate an optimized set of influencing factors;

[0090] S24. Based on the optimized set of influencing factors, the causal structure of the dynamic talent network is learned to construct a causal relationship network G. * =(V,E * ,X), where E * represents the edge set after removing redundant relationships based on causal analysis, and uses a structural optimization algorithm to ensure |E * |≤|E| and network connectivity is guaranteed;

[0091] S25. Use the embedding method based on graph neural network to learn the features of talent nodes, generate node embedding, and use the objective function L to obtain the optimized talent network embedding matrix:

[0092]

[0093] Among them, (i,j) represents the connection edge between the i-th talent node and the j-th talent node, σ represents the activation function, and Z i represents the representation of the i-th talent node in the embedding space, Z j represents the representation of the j-th talent node in the embedding space, Z k Represents the representation of the k-th talent node in the embedding space;

[0094] S26. Based on the optimized talent network embedding matrix, calculate the similarity between talents and perform talent clustering, set the number of talent categories to K, use the K-means algorithm to optimize the talent distribution, and finally generate an optimized dynamic talent network.

[0095] In this implementation, S3 specifically includes:

[0096] S31. Based on the optimized dynamic talent network, the talent flow prediction problem is transformed into an optimal path search problem, and the talent flow path is defined as the path from the starting node to the target node;

[0097] S32. In the process of path search, a variant ant colony optimization algorithm is used to simulate multiple ant individuals exploring in the network. The search for high-quality paths is continuously enhanced through the positive feedback mechanism of pheromones. Each ant individual selects a path according to the following fitness function:

[0098]

[0099] Among them, Fit(p) represents the fitness of path p, w i represents the weight of the i-th talent node, d i (p) represents the flow cost of path p at the i-th talent node, exp represents the exponential function, σ i represents the standard deviation of liquidity cost, f j (p) represents the influencing factor of path p on the jth feature, β represents the importance weight of the feature, n represents the total number of talent nodes in the path, and m represents the total number of influencing factors;

[0100] S33, using the proximal strategy optimization algorithm, based on the experience gained in the ant search process, dynamically adjust the search strategy, and update the parameters of the strategy network through back propagation, so that the search process gradually tends to the global optimum:

[0101]

[0102] Among them, θ t+1 represents the parameters of the updated policy network, θ t represents the parameters of the policy network before updating, α represents the learning rate, represents the policy gradient, λ represents the dynamic adjustment factor, Δθ represents the difference between the current strategy and the previous strategy, γ represents the adjustable factor, and C t represents the cumulative return of the current strategy, and E[C] represents the average return of the strategy;

[0103] S34, introduce the adaptive differential mutation operator, dynamically adjust the search range based on individual historical experience and current path search situation, and the mutation operator dynamically changes the step length:

[0104] X new =X old +ΔX adaptive ·(α 1 ·||X r -X best ||γ +β 1 ·I(X r ∈P));

[0105] Among them, X new represents the solution after mutation, X old represents the current solution, X r represents a randomly selected individual, X best represents the optimal solution in the current population, ΔX adaptive represents the adaptively adjusted variable step length, γ represents the weight factor of the adjustment distance, and α 1 and β 1 represents the adjustment factor, I(X r ∈P) represents the indicator function, P represents the known high-quality path set, when X r The value is 1 when it belongs to the known high-quality path set P, otherwise it is 0;

[0106] S35. In the process of combining variant ant colony optimization with proximal strategy optimization, the fitness information of the optimal path is used to model the talent flow trend, generate several flow paths, each flow path represents the talent flow pattern between different regions, and select the optimal path output according to the path fitness function;

[0107] S36. Generate a talent flow prediction model according to the selected optimal path, wherein the talent flow prediction model is based on the structure:

[0108]

[0109] Among them, M(t) represents the prediction result at time t, X p (t) represents the influence characteristics of the pth path at time t, Y p (t) represents the historical flow characteristics of the pth path at time t, δ p and p represents the learning parameter, and N represents the number of paths.

[0110] In this implementation manner, the S4 specifically includes:

[0111] S41. According to the talent flow prediction model, a Transformer-based spatiotemporal graph neural network is constructed, wherein the spatiotemporal graph neural network includes an input layer, a spatiotemporal graph convolution layer, a multi-head attention layer, and an output layer, wherein the input layer processes and fuses the spatiotemporal features of different data sources to generate a spatiotemporal feature matrix;

[0112] S42, input the spatiotemporal feature matrix into the spatiotemporal graph convolution layer, perform graph convolution operation, perform convolution processing through the graph convolution kernel, capture the spatiotemporal dependency of talent flow between regions, extract the spatial and temporal patterns of talent flow, and generate a feature vector for each regional node;

[0113] S43, inputting the feature vector of each regional node into the multi-head attention layer, aggregating the feature information of the adjacent nodes of each regional node by weighted summation, and generating a weighted feature representation;

[0114] S44, in the spatiotemporal graph neural network, encoding the spatial position and time information of the regional node into the regional node feature through position encoding, and generating a regional node feature vector with spatiotemporal information;

[0115] S45. In the output layer, the weighted feature representation and the regional node feature vector with spatiotemporal information are fused through the fully connected layer to generate the talent flow trend prediction results for each region in the future time period.

[0116] In this implementation manner, S5 specifically includes:

[0117] S51. Based on the prediction results of talent flow trend, an adaptive co-evolutionary game model is constructed. The adaptive co-evolutionary game model includes three game subjects, including talents, enterprises and policy makers. Each game subject has a different strategy space, and a supply and demand matching strategy is generated through game interaction.

[0118] S52. Set the fitness function of each game subject. The talent fitness function is set according to the selected flow area and matching results, the enterprise fitness function is set according to the recruitment target and talent matching, and the policy maker fitness function is set according to the set regional development strategy and talent introduction policy;

[0119] S53. Through evolutionary game theory, simulate the game process between game players, where each game player updates its own strategy according to the fitness function, dynamically adjust the talent supply and demand matching strategy through the evolutionary process, and optimize the strategy selection of the game players;

[0120] S54. During the game, according to the interaction between the game players, the step size and update frequency of the strategy selection are adjusted, and the speed of strategy adjustment is controlled through the adaptive learning mechanism;

[0121] S55. Through repeated iterations of the adaptive co-evolutionary game model, a talent supply and demand matching solution is output.

[0122] In this implementation manner, S6 specifically includes:

[0123] S61, initializing the population of the MOEA / D algorithm by randomly generating multiple initial solutions, each of which represents a talent allocation plan, wherein each talent allocation plan includes the optimal talent allocation ratio, flow path and supply and demand matching information of different regions;

[0124] S62. Define multiple optimization goals, including minimizing the cost of talent mobility, maximizing talent matching efficiency, and maximizing the balance between talent supply and demand among regions, and assign initial weights to each goal;

[0125] S63, using a weighted linear combination method to adjust the weight of each objective, and through an adaptive adjustment strategy, during the optimization process, the weight of each objective is dynamically adjusted according to changes in the current optimization stage:

[0126]

[0127] in, represents the weight of the i-th target in the t+1th generation, represents the weight of the i-th target in the t-th generation, κ represents the weight adjustment factor, and f i (x best ) represents the objective function value of the current optimal solution, f i (x current ) represents the objective function value of the current solution;

[0128] S64. In each iteration, individuals with fitness higher than the set threshold are selected from the current population through selection operations, new solutions are generated through crossover operations, and diversity is introduced through mutation operations. When any of the following conditions is met, the iteration process is terminated and the global optimal talent allocation result is output:

[0129] The first condition is that the number of iterations reaches the preset maximum number of iterations;

[0130] The second condition is that the change between the optimal solution of each generation and the optimal solution of the previous generation is less than a preset threshold.

[0131] In this implementation manner, the S7 specifically includes:

[0132] S71. Based on the global optimal talent allocation result, a dynamic visual talent map is constructed, wherein the dynamic visual talent map includes a plurality of nodes and edges connecting the nodes, wherein the nodes represent various regions, the weights of the edges represent the intensity of talent flow between regions, the size of the nodes represents the number of talents in the region, and the colors represent the talent supply and demand status of the region;

[0133] S72. Establishing a spatial topological structure of talent flow through a dynamic graph convolutional network, wherein the dynamic graph convolutional network is used to model the dynamic changes of talent flow between regions, and capturing the spatiotemporal characteristics of talent flow through convolution operations, generating a spatiotemporal feature vector of each node, wherein the spatiotemporal feature vector contains real-time talent flow information of the region;

[0134] S73. In the dynamic graph convolution network, a time window sliding strategy is used to dynamically update the graph convolution kernel to adapt to the impact of time changes on the talent flow pattern;

[0135] S74, automatically clustering the talent groups by latent Dirichlet distribution, wherein the latent Dirichlet distribution clusters the talent types, flow paths and demand intensity characteristics in the region, and the clustering results reflect the potential trends and group characteristics of talent flow in the region;

[0136] S75. Based on the clustering results, the flow of talents is displayed through visualization methods. The chord diagram technology is used to display the talent flow paths between regions. The chord width in the chord diagram represents the intensity of talent flow, and the color of the chord represents the flow direction.

[0137] S76. Combining the clustering results of the latent Dirichlet distribution and the dynamic visualization of the chord diagram, a visualized talent flow map is generated. The visualized talent flow map is dynamically updated according to time changes to provide real-time information on the trend, direction and intensity of talent flow between regions.

[0138] Embodiment 1:

[0139] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the prediction of talent flow and optimization of supply and demand matching in a certain region. The region is a medium-sized city in a certain province in China, facing a series of problems such as accelerated talent flow, uneven regional development, and unreasonable talent distribution. In this case, the matching degree between the talent demand and supply of enterprises and universities in the region is low, especially in terms of technical talents and high-end talents, which seriously affects the high-quality development of the economy and the improvement of regional competitiveness. According to government statistics, the city experienced a more obvious brain drain in the first half of 2023, especially in information technology, artificial intelligence, financial services and other industries. The loss of high-end talents is more serious.

[0140] In this context, we applied the method of the present invention to the region and adopted a regional talent map construction method based on big data and group algorithms, attempting to improve the problem of uneven talent distribution by optimizing talent flow prediction and supply and demand matching strategies, thereby promoting high-quality development of the regional economy.

[0141] During the implementation process, we first obtained multi-source heterogeneous data from government statistics, recruitment markets, social networks, and college graduates, performed data preprocessing, and constructed a standardized talent data set. Through data fusion, we formed dynamic data of various types of talents in the region and analyzed them in combination with historical data. Through the deep dynamic graph embedding method, we successfully constructed a dynamic talent network in the region, and performed causal structure learning on it, extracting key influencing factors related to talent mobility.

[0142] Next, we used the mutation ant colony optimization algorithm combined with proximal strategy optimization to search for paths, and dynamically adjusted the search range through the adaptive differential mutation operator. By predicting the talent flow path, we identified the flow trend from high-end technical talent-intensive areas to talent-shortage areas, and then generated a talent flow prediction model, providing data support for subsequent supply and demand matching optimization.

[0143] On this basis, we used the Transformer-based spatiotemporal graph neural network to model the talent flow trend. Through model training and optimization, we successfully predicted the main direction and trend of talent flow in the region in the next quarter, providing a scientific basis for policy makers, corporate recruitment and employment of college graduates.

[0144] In order to further improve the matching degree of talent supply and demand, we adopted an adaptive co-evolutionary game model and optimized the supply and demand matching strategy of talent flow by dynamically adjusting the strategies between game players. In this process, we also considered the cost of talent flow, the talent demand situation in the region, and policy factors, and gradually optimized a set of dynamic adjustment plans that meet the needs of regional development. Through continuous game interaction, enterprises, talents and policymakers have reached a more reasonable balance in talent allocation.

[0145] Finally, we used the MOEA / D algorithm to perform multi-objective optimization on the talent supply and demand matching scheme and generated the global optimal talent allocation result. By fine-tuning the supply and demand matching of different regions and different talent categories, we finally generated an optimized regional talent distribution map, which can dynamically display the supply and demand status and talent flow trends of each region in real time, providing strong support for subsequent talent policies and corporate decisions.

[0146] During the implementation process, we also updated the data in real time and visualized the method. Through the dynamic graph convolutional network, we were able to capture the spatiotemporal characteristics of talent flow and establish a dynamic talent flow model with spatial topology. Combining the latent Dirichlet distribution to cluster the talent groups, we successfully clustered the various types of talents in the region according to their flow trends and demand characteristics, and visualized them on the map. In this way, policymakers and companies can intuitively see the flow direction, flow intensity and flow trend of talents in each region, helping them make more accurate decisions.

[0147] Specific data proves that after applying the technology of the present invention, the flow of high-end talents in the region has been significantly improved.

[0148] Table 1 Comparison of changes in key talent flow and supply-demand matching indicators before and after implementation

[0149] index Before implementation After implementation Rate of change High-end technical talent loss rate (%) 18.43 12.67 -5.76 Matching degree of enterprise talent demand (%) 63.21 71.45 8.24 Employment matching degree of college graduates (%) 57.39 64.58 7.19 Prediction accuracy of talent flow (%) 69.25 83.16 13.91 Balance of talent supply and demand in the region (%) 56.81 69.12 12.31

[0150] According to the data provided in Table 1 above, it can be seen that the present invention has achieved remarkable results in improving the flow of talents and matching supply and demand. First, the significant reduction in the loss rate of high-end technical talents shows that after the implementation of the method of the present invention, the loss of high-end talents can be effectively slowed down. The loss rate of high-end technical talents dropped from 18.43% to 12.67%, a decrease of 5.76%. This change reflects that the present invention has helped the region retain more technical talents by optimizing the allocation and flow paths of talents, providing stronger human resource guarantees for the long-term development of the regional economy.

[0151] Secondly, the improvement in the matching degree of enterprise talent needs further verifies the effectiveness of the present invention in optimizing the supply and demand of the talent market. The matching degree of enterprise talent needs increased from 63.21% to 71.45%, an increase of 8.24%. This shows that after implementing dynamic optimization matching based on deep learning, enterprises can more accurately find talents that meet their needs, thereby improving recruitment efficiency and reducing the waste of human resources.

[0152] In addition, the improvement of the employment matching degree of college graduates is also one of the important achievements of the present invention. The employment matching degree of college graduates increased from 57.39% to 64.58%, an increase of 7.19%. This change shows that the supply and demand matching optimization scheme provided by the present invention can effectively promote the connection between college graduates and market demand, improve the accuracy of graduates' employment, and provide better employment guidance for colleges and universities.

[0153] The improvement in the accuracy of talent flow prediction further highlights the technical advantages of the present invention. The accuracy of talent flow prediction has increased from 69.25% to 83.16%, an increase of 13.91%. This improvement shows that the present invention can more accurately predict the flow trend of talents in a region through deep graph neural networks and spatiotemporal modeling technology. This accurate prediction provides strong data support for governments and enterprises to formulate talent introduction and flow policies, and further optimizes the allocation of talent resources.

[0154] Finally, the improvement in the balance between supply and demand of talents in the region shows the actual effect of the present invention in optimizing the distribution of talents in the region. The balance between supply and demand increased from 56.81% to 69.12%, an increase of 12.31%. This change reflects that the present invention can dynamically adjust the supply and demand relationship of different regions and different types of talents through the application of multi-objective optimization algorithms, realize the reasonable flow and allocation of talents between regions, and alleviate the long-standing imbalance between supply and demand of talents.

[0155] In summary, the present invention significantly improves the accuracy of talent flow prediction, the efficiency of talent supply and demand matching, and effectively slows down the loss of high-end technical talents through the regional talent map construction method based on big data and group algorithm. These results show that the present invention has great potential in practical application, and can provide powerful talent flow optimization tools for local governments, enterprises and universities, and help the rational allocation of talent resources and the development of regional economy.

[0156] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a regional talent map based on big data and group algorithms, characterized in that: The steps include: S1. Obtain multi-source heterogeneous talent data, including government statistics, recruitment market, social networks and college graduate data, and pre-process them to build a standardized talent data set; S2. Based on the standardized talent dataset, a deep dynamic graph embedding method is used to build a dynamic talent network. Talent individuals are used as network nodes, and causal-based structural learning is used to remove redundant information to generate an optimized dynamic talent network. S3. Based on the optimized dynamic talent network, talent flow prediction is performed. The path search is performed by using mutation ant colony optimization combined with proximal strategy optimization, and the search range is dynamically adjusted through the adaptive differential mutation operator to generate a talent flow prediction model. S4. Based on the talent flow prediction model, the Transformer-based spatiotemporal graph neural network is used to model the talent flow trend and generate the talent flow trend prediction results; S5. Optimize the matching of talent supply and demand based on the prediction results of talent flow trend, dynamically adjust the matching strategy using adaptive co-evolutionary game, and output the talent supply and demand matching plan; S6. Use MOEA / D algorithm to perform multi-objective optimization on the talent supply and demand matching scheme to generate the global optimal talent allocation result; S7. Based on the globally optimal talent allocation results, a dynamic visual talent map is constructed. Dynamic graph convolution is used to establish the spatial topological structure of talent flow. The talent groups are automatically clustered in combination with the latent Dirichlet distribution to output a visual talent flow map.

2. The method for constructing a regional talent map based on big data and group algorithm according to claim 1 is characterized in that: The S2 specifically includes: S21. Construct an initial talent network G = (V, E, X) based on the standardized talent data set, where V represents the talent node set, E represents the relationship edge set between talents, and X represents the talent feature matrix, and initialize the adjacency matrix; S22. Use the deep dynamic graph embedding method to perform temporal modeling on the initial talent network G and construct a dynamic graph G that changes over time. t =(V,E t ,X t ), where E t represents the talent relationship set at time step t, X t Represent the talent characteristics at time step t and generate a dynamic talent network; S23. Use the Granger causal analysis method based on causal inference to calculate the causal effect of each influencing factor on talent mobility, select the main influencing factors by setting the contribution threshold τ, and generate an optimized set of influencing factors; S24. Based on the optimized set of influencing factors, the causal structure of the dynamic talent network is learned to construct a causal relationship network G. * =(V,E * ,X), where E * represents the edge set after removing redundant relationships based on causal analysis, and uses a structural optimization algorithm to ensure |E * |≤|E| and network connectivity is guaranteed; S25. Use the embedding method based on graph neural network to learn the features of talent nodes, generate node embedding, and use the objective function L to obtain the optimized talent network embedding matrix: Among them, (i,j) represents the connection edge between the i-th talent node and the j-th talent node, σ represents the activation function, and Z i represents the representation of the i-th talent node in the embedding space, Z j represents the representation of the j-th talent node in the embedding space, Z k Represents the representation of the k-th talent node in the embedding space; S26. Based on the optimized talent network embedding matrix, calculate the similarity between talents and perform talent clustering, set the number of talent categories to K, use the K-means algorithm to optimize the talent distribution, and finally generate an optimized dynamic talent network.

3. The method for constructing a regional talent map based on big data and group algorithm according to claim 1 is characterized in that: The S3 specifically includes: S31. Based on the optimized dynamic talent network, the talent flow prediction problem is transformed into an optimal path search problem, and the talent flow path is defined as the path from the starting node to the target node; S32. In the process of path search, a variant ant colony optimization algorithm is used to simulate multiple ant individuals exploring in the network. The search for high-quality paths is continuously enhanced through the positive feedback mechanism of pheromones. Each ant individual selects a path according to the following fitness function: Among them, Fit(p) represents the fitness of path p, w i represents the weight of the i-th talent node, d i (p) represents the flow cost of path p at the i-th talent node, exp represents the exponential function, σ i represents the standard deviation of liquidity cost, f j (p) represents the influencing factor of path p on the jth feature, β represents the importance weight of the feature, n represents the total number of talent nodes in the path, and m represents the total number of influencing factors; S33, using the proximal strategy optimization algorithm, based on the experience gained in the ant search process, dynamically adjust the search strategy, and update the parameters of the strategy network through back propagation, so that the search process gradually tends to the global optimum: Among them, θ t+1 represents the parameters of the updated policy network, θ t represents the parameters of the policy network before updating, α represents the learning rate, represents the policy gradient, λ represents the dynamic adjustment factor, Δθ represents the difference between the current strategy and the previous strategy, γ represents the adjustable factor, and C t represents the cumulative return of the current strategy, and E[C] represents the average return of the strategy; S34, introduce the adaptive differential mutation operator, dynamically adjust the search range based on individual historical experience and current path search situation, and the mutation operator dynamically changes the step length: X new =X old +ΔX adaptive ·(α1·||X r -X best || γ +β1·I(X r ∈P)); Among them, X new represents the solution after mutation, X old represents the current solution, X r represents a randomly selected individual, X best represents the optimal solution in the current population, ΔX adaptive represents the adaptively adjusted variable step length, γ represents the weight factor of the adjustment distance, α1 and β1 represent the adjustment factors, I(X r ∈P) represents the indicator function, P represents the known high-quality path set, when X r The value is 1 when it belongs to the known high-quality path set P, otherwise it is 0; S35. In the process of combining variant ant colony optimization with proximal strategy optimization, the fitness information of the optimal path is used to model the talent flow trend, generate several flow paths, each flow path represents the talent flow pattern between different regions, and select the optimal path output according to the path fitness function; S36. Generate a talent flow prediction model according to the selected optimal path, wherein the talent flow prediction model is based on the structure: Among them, M(t) represents the prediction result at time t, X p (t) represents the influence characteristics of the pth path at time t, Y p (t) represents the historical flow characteristics of the pth path at time t, δ p and p represents the learning parameter, and N represents the number of paths.

4. The method for constructing a regional talent map based on big data and group algorithm according to claim 1 is characterized in that: The S4 specifically includes: S41. According to the talent flow prediction model, a Transformer-based spatiotemporal graph neural network is constructed, wherein the spatiotemporal graph neural network includes an input layer, a spatiotemporal graph convolution layer, a multi-head attention layer, and an output layer, wherein the input layer processes and fuses the spatiotemporal features of different data sources to generate a spatiotemporal feature matrix; S42, input the spatiotemporal feature matrix into the spatiotemporal graph convolution layer, perform graph convolution operation, perform convolution processing through the graph convolution kernel, capture the spatiotemporal dependency of talent flow between regions, extract the spatial and temporal patterns of talent flow, and generate a feature vector for each regional node; S43, inputting the feature vector of each regional node into the multi-head attention layer, aggregating the feature information of the adjacent nodes of each regional node by weighted summation, and generating a weighted feature representation; S44, in the spatiotemporal graph neural network, encoding the spatial position and time information of the regional node into the regional node feature through position encoding, and generating a regional node feature vector with spatiotemporal information; S45. In the output layer, the weighted feature representation and the regional node feature vector with spatiotemporal information are fused through the fully connected layer to generate the talent flow trend prediction results for each region in the future time period.

5. The method for constructing a regional talent map based on big data and group algorithm according to claim 1 is characterized in that: The S5 specifically includes: S51. Based on the prediction results of talent flow trend, an adaptive co-evolutionary game model is constructed. The adaptive co-evolutionary game model includes three game subjects, including talents, enterprises and policy makers. Each game subject has a different strategy space, and a supply and demand matching strategy is generated through game interaction. S52. Set the fitness function of each game subject. The talent fitness function is set according to the selected flow area and matching results, the enterprise fitness function is set according to the recruitment target and talent matching, and the policy maker fitness function is set according to the set regional development strategy and talent introduction policy; S53. Through evolutionary game theory, simulate the game process between game players, where each game player updates its own strategy according to the fitness function, dynamically adjust the talent supply and demand matching strategy through the evolutionary process, and optimize the strategy selection of the game players; S54. During the game, according to the interaction between the game players, the step size and update frequency of the strategy selection are adjusted, and the speed of strategy adjustment is controlled through the adaptive learning mechanism; S55. Through repeated iterations of the adaptive co-evolutionary game model, a talent supply and demand matching solution is output.

6. The method for constructing a regional talent map based on big data and group algorithm according to claim 1 is characterized in that: The S6 specifically includes: S61, initializing the population of the MOEA / D algorithm by randomly generating multiple initial solutions, each of which represents a talent allocation plan, wherein each talent allocation plan includes the optimal talent allocation ratio, flow path and supply and demand matching information of different regions; S62. Define multiple optimization goals, including minimizing the cost of talent mobility, maximizing talent matching efficiency, and maximizing the balance between talent supply and demand among regions, and assign initial weights to each goal; S63, using a weighted linear combination method to adjust the weight of each objective, and through an adaptive adjustment strategy, during the optimization process, the weight of each objective is dynamically adjusted according to changes in the current optimization stage: in, represents the weight of the i-th target in the t+1th generation, represents the weight of the i-th target in the t-th generation, κ represents the weight adjustment factor, and f i (x best ) represents the objective function value of the current optimal solution, f i (x current ) represents the objective function value of the current solution; S64. In each iteration, individuals with fitness higher than the set threshold are selected from the current population through selection operations, new solutions are generated through crossover operations, and diversity is introduced through mutation operations. When any of the following conditions is met, the iteration process is terminated and the global optimal talent allocation result is output: The first condition is that the number of iterations reaches the preset maximum number of iterations; The second condition is that the change between the optimal solution of each generation and the optimal solution of the previous generation is less than a preset threshold.

7. The method for constructing a regional talent map based on big data and group algorithm according to claim 1 is characterized in that: The S7 specifically includes: S71. Based on the global optimal talent allocation result, a dynamic visual talent map is constructed, wherein the dynamic visual talent map includes a plurality of nodes and edges connecting the nodes, wherein the nodes represent various regions, the weights of the edges represent the intensity of talent flow between regions, the size of the nodes represents the number of talents in the region, and the colors represent the talent supply and demand status of the region; S72. Establishing a spatial topological structure of talent flow through a dynamic graph convolutional network, wherein the dynamic graph convolutional network is used to model the dynamic changes of talent flow between regions, and capturing the spatiotemporal characteristics of talent flow through convolution operations, generating a spatiotemporal feature vector of each node, wherein the spatiotemporal feature vector contains real-time talent flow information of the region; S73. In the dynamic graph convolution network, a time window sliding strategy is used to dynamically update the graph convolution kernel to adapt to the impact of time changes on the talent flow pattern; S74, automatically clustering the talent groups by latent Dirichlet distribution, wherein the latent Dirichlet distribution clusters the talent types, flow paths and demand intensity characteristics in the region, and the clustering results reflect the potential trends and group characteristics of talent flow in the region; S75. Based on the clustering results, the flow of talents is displayed through visualization methods. The chord diagram technology is used to display the talent flow paths between regions. The chord width in the chord diagram represents the intensity of talent flow, and the color of the chord represents the flow direction. S76. Combining the clustering results of the latent Dirichlet distribution and the dynamic visualization of the chord diagram, a visualized talent flow map is generated. The visualized talent flow map is dynamically updated according to time changes to provide real-time information on the trend, direction and intensity of talent flow between regions.

Images