A method and system for labor flow prediction of a spatio-temporal graph network

Abstract

The application relates to the technical field of data prediction, and particularly discloses a labor flow prediction method of a space-time graph network, which comprises the following steps: fusing and analyzing historical labor flow data and regional attribute data of a target region to obtain labor flow fusion features; encoding the spatial adjacency relationship of the target region based on the labor flow fusion features to obtain a labor flow graph; performing spatial self-attention mapping on the labor flow graph to obtain labor flow spatial interaction features of nodes; performing neighborhood feature aggregation on the labor flow spatial interaction features to obtain labor flow features; performing time sequence dependence coding on the labor flow features to obtain labor flow time sequence features; and performing global feature weighted fusion on the labor flow time sequence features to obtain labor flow prediction values; and the application can improve the labor flow prediction efficiency of the space-time graph network.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and in particular to a method and system for predicting labor mobility using spatiotemporal graph networks. Background Technology

[0002] Traditional labor mobility forecasting methods currently available rely primarily on single statistical models or simplified econometric analyses, lacking an effective fusion of spatial dependencies and regional characteristics in labor mobility. These methods typically treat regions as independent entities, failing to capture the complex spatial adjacency relationships between regions, as well as the nonlinear dynamic characteristics of labor mobility, such as push-pull effects, distance decay mechanisms, and multi-order historical inertia. This results in an insufficient characterization of the spatial interactions within labor mobility. Furthermore, existing techniques employ a crude approach to coupling regional attribute data with historical mobility data, often using simple linear superposition or fixed weight combinations. This fails to adapt to the dynamic changes in labor mobility behavior over time and space, causing forecasts to deviate from actual mobility patterns and limiting the model's generalization ability.

[0003] Existing labor mobility prediction methods based on recurrent neural networks or traditional time series analysis often neglect effective temporal dependency encoding and global weighted fusion of mobility features. This makes them ill-suited for handling long-term state transitions and cumulative effects, resulting in slow responses to periodic, trend-based, and sudden changes in labor mobility. Furthermore, existing methods often employ fixed graph structures or predefined adjacency matrices for neighborhood feature aggregation, lacking dynamic feature reorganization capabilities driven by spatial self-attention mechanisms. This hinders adaptive quantification of inter-node correlation strength, leading to insufficient representation of labor mobility features and inadequate prediction accuracy. Therefore, there is an urgent need to develop a labor mobility prediction method capable of jointly modeling spatiotemporal dependencies, dynamically fusing regional attributes and mobility data, and possessing adaptive spatial attention and global weighted fusion mechanisms. This would address the shortcomings of existing technologies, such as insufficient spatial interaction characterization, weak temporal dependency modeling, and inadequate feature representation, thereby improving the accuracy and reliability of labor mobility prediction. Summary of the Invention

[0004] This invention provides a method and system for predicting labor mobility using spatiotemporal graph networks to address the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides a method and system for predicting labor mobility using spatiotemporal graph networks, comprising: S1. The historical labor mobility data and regional attribute data of the target area are fused and analyzed to obtain the fused characteristics of the labor mobility of the target area. S2. Based on the labor mobility fusion characteristics, spatial adjacency relationship encoding is performed on the target area to obtain the labor mobility map of the target area; S3. Perform spatial self-attention mapping on the labor flow map to obtain the spatial interaction characteristics of labor flow of nodes in the target region; S4. Perform neighborhood feature aggregation on the spatial interaction features of the labor mobility to obtain the labor mobility features of the target region; S5. Perform time-dependent encoding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target region; S6. Perform global feature weighted fusion on the time-series characteristics of labor mobility to obtain the predicted value of labor mobility in the target area.

[0006] In a preferred embodiment, the step of fusing and analyzing historical labor mobility data and regional attribute data of the target area to obtain the fused labor mobility characteristics of the target area includes: By performing time-series alignment on the historical labor mobility data of the target area, a numerical sequence of labor mobility data for the target area is obtained. The regional attributes of the target region are parsed to obtain the attribute feature vector of the target region; Autocorrelation measurement is performed on the numerical series of labor mobility in the target region to obtain the mobility dependence strength value of the target region. The labor mobility fusion characteristics are obtained by performing intensity coupling analysis on the flow dependence strength value and the attribute feature vector.

[0007] In a preferred embodiment, the step of performing intensity coupling analysis on the flow dependence strength value and the attribute feature vector to obtain the labor flow fusion characteristics includes: Each time point in the labor mobility numerical sequence is mapped to an embedding vector, which consists of the value at the time point and the values ​​at the time points before the time point. The embedding vectors are stacked as matrices to obtain the delay embedding matrix of the labor mobility numerical sequence. The delayed embedding matrix is ​​decomposed into a low-rank form to obtain a sequence of singular values ​​of the delayed embedding matrix. The long-term dependency coefficient is inverted based on the decay rate of the singular value sequence, and the long-term dependency coefficient is nonlinearly combined with the trace norm of the delay embedding matrix to obtain the flow dependency strength value.

[0008] In a preferred embodiment, the step of encoding the spatial adjacency relationships of the target region based on the labor mobility fusion characteristics to obtain a labor mobility map of the target region includes: Each time point in the labor mobility numerical sequence is mapped to an embedding vector, which consists of the value at the time point and the values ​​at the time points before the time point. The embedding vectors are stacked as matrices to obtain the delay embedding matrix of the labor mobility numerical sequence. The delayed embedding matrix is ​​decomposed into a low-rank form to obtain a sequence of singular values ​​of the delayed embedding matrix. The long-term dependency coefficient is inverted based on the decay rate of the singular value sequence, and the long-term dependency coefficient is nonlinearly combined with the trace norm of the delay embedding matrix to obtain the flow dependency strength value.

[0009] In a preferred embodiment, the step of performing spatial self-attention mapping on the labor flow map to obtain the spatial interaction features of labor flow of nodes in the target region includes: Spatial dependency strength quantification is performed on any two nodes in the labor mobility graph to obtain the association strength value between the two nodes; The attention distribution of the association strength value is reshaped to obtain the attention weights corresponding to the two nodes. Based on the attention weights, the labor flow fusion features of the nodes in the labor flow graph are reorganized to obtain the spatial interaction features of labor flow in the target region.

[0010] In a preferred embodiment, the step of quantifying the spatial dependency strength of two nodes in the labor mobility graph to obtain the association strength value between the two nodes includes: ; in, These are regions with a large outflow of labor. For the target area for labor inflow. For the region and region The strength of the labor mobility correlation value, This is the basic attraction coefficient between the labor outflow source region and the labor inflow target region. and These are the learnable push factor vector of the labor outflow source region and the learnable pull factor vector of the labor inflow target region, respectively. The economic or geographical distance between the source region of the labor outflow and the target region of the labor inflow. Let be the rate of change in labor mobility from the labor outflow source region to the labor inflow target region at the k-th time step. The preset history window length, For push-pull factor strength parameters, For distance attenuation intensity parameters, For historical inertia scaling parameters, For distance scale parameters, Here, t represents the weight of the k-th order historical inertia, and t is the discrete time step index. For the Sigmoid function, It is the hyperbolic tangent function. It is a natural exponential function.

[0011] In a preferred embodiment, the step of aggregating neighborhood features of the spatial interaction characteristics of labor mobility to obtain the labor mobility characteristics of the target region includes: The proximity relationship quantification of the labor flow graph yields the distribution of neighborhood influence intensity between neighboring nodes in the labor flow graph. Based on the neighborhood influence intensity distribution, the labor mobility spatial interaction features of neighboring nodes in the labor mobility graph are weighted and fused to obtain the weighted aggregated neighborhood features of the neighboring nodes. The spatial interaction features of labor mobility of the node are coupled with the corresponding weighted aggregated neighborhood features under gating adaptive coupling to obtain the fusion features of the node. The fusion features of the node are then spliced ​​together in node order to form the labor mobility features of the target region.

[0012] In a preferred embodiment, the step of performing time-dependent encoding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target region includes: The labor mobility characteristics are segmented into time-series windows to obtain a time-series mobility characteristic sequence for the target region. Perform temporal dependency mapping on the features of two adjacent time steps in the temporal flow feature sequence to obtain the state transition intensity value of the temporal flow feature sequence; Based on the state transition intensity value, feature progressive propagation is performed on the temporal flow feature sequence to obtain the cumulative temporal features corresponding to each time step; The accumulated time-series features are aggregated along the time-series dimension to obtain the time-series features of labor mobility in the target region.

[0013] In a preferred embodiment, the step of performing global feature weighted fusion on the temporal characteristics of labor mobility to obtain the predicted labor mobility values ​​for the target region includes: Global contribution decoupling is performed on each feature dimension in the time series characteristics of labor mobility to obtain the adaptive importance weight coefficient corresponding to each feature dimension; Based on the adaptive importance weight coefficients, the feature dimensions in the temporal features of labor mobility are scaled and gated dimension by dimension to obtain the weighted calibrated feature representation of the target region; The weighted calibration feature representation is subjected to nonlinear response synthesis from the feature space to the prediction numerical space to obtain the predicted labor flow values ​​for the target region.

[0014] To address the above problems, the present invention also provides a labor mobility prediction system based on spatiotemporal graph networks, the system comprising: The data fusion and parsing module is used to collaboratively interpret historical labor flow data and regional attribute data of the target area to obtain the labor flow fusion characteristics of the target area. The spatial coding module is used to encode the spatial adjacency relationship of the target area based on the labor mobility fusion characteristics to obtain the labor mobility map of the target area; The spatial self-attention module is used to perform spatial self-attention mapping on the nodes of the labor flow graph to obtain the spatial interaction features of labor flow of the nodes in the target region. The feature aggregation module is used to perform neighborhood feature fusion on the spatial interaction features of the labor mobility to obtain the labor mobility features of the target area. A time-series coding module is used to perform time-series dependency coding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target region. The global weighted fusion module is used to perform global weighted fusion on the temporal features of labor mobility to obtain the predicted value of labor mobility in the target area.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. This invention constructs a labor flow map that reflects the true inter-regional dependencies by fusing and analyzing historical labor flow data with regional attribute data and encoding spatial adjacency relationships based on the fused characteristics of labor flow. On this basis, a spatial self-attention mapping mechanism is employed, utilizing a nonlinear association strength quantification formula that includes learnable parameters such as push-pull factors, distance decay parameters, and multi-order historical inertia weights. This enables dynamic adaptive calculation of the spatial dependency strength between any two nodes, effectively overcoming the shortcomings of existing technologies where fixed adjacency matrices or simple distance metrics are insufficient to characterize the complex push-pull effects, distance decay, and historical inertia influences in labor flow. Simultaneously, through neighborhood feature aggregation and gated adaptive coupling, the aggregation weights can be dynamically adjusted according to the influence strength distribution of neighboring nodes, significantly improving the ability to express the spatial interaction characteristics of labor flow. This allows the model to accurately capture the asymmetric, nonlinear, and time-varying spatial dependencies between different regions, resulting in significantly enhanced regional matching and spatial consistency of the prediction results.

[0017] 2. This invention employs time-dependent encoding of labor mobility characteristics. Through time-series window segmentation, state transition intensity mapping, and feature progressive propagation, it effectively models the cumulative effects and dynamic evolution of labor mobility behavior over long periods, solving the problem of slow response to periodic, trend, and sudden changes in traditional time series analysis methods. Furthermore, by adopting a global weighted fusion mechanism, it decouples the contribution of each feature dimension in the labor mobility time series characteristics and adaptively weights them based on importance, achieving a high-quality nonlinear mapping from the feature space to the prediction numerical space. This avoids information loss and bias caused by simple linear superposition or fixed-weight fusion. Therefore, this invention can integrate spatial interaction features and time-dependent features to output high-precision labor mobility prediction values. It significantly outperforms existing methods in terms of prediction accuracy, robustness, and adaptability to dynamic changes, providing reliable technical support for labor resource allocation and regional economic policy formulation. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for predicting labor mobility using a spatiotemporal graph network, provided in an embodiment of the present invention.

[0019] Figure 2 A functional block diagram of a labor mobility prediction system based on a spatiotemporal graph network provided in an embodiment of the present invention;

[0020] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] This application provides a method for predicting labor mobility using a spatiotemporal graph network. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for predicting labor mobility using a spatiotemporal graph network can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0023] Reference Figure 1 The diagram shown is a flowchart illustrating a labor mobility prediction method using a spatiotemporal graph network according to the present invention. In this embodiment, the labor mobility prediction method using a spatiotemporal graph network includes: S1. The historical labor mobility data and regional attribute data of the target area are fused and analyzed to obtain the fused characteristics of the labor mobility of the target area. In this embodiment of the invention, the step of fusing and analyzing historical labor mobility data and regional attribute data of the target area to obtain the fused labor mobility characteristics of the target area includes: By performing time-series alignment on the historical labor mobility data of the target area, a numerical sequence of labor mobility data for the target area is obtained. The regional attributes of the target region are parsed to obtain the attribute feature vector of the target region; Autocorrelation measurement is performed on the numerical series of labor mobility in the target region to obtain the mobility dependence strength value of the target region. The labor mobility fusion characteristics are obtained by performing intensity coupling analysis on the flow dependence strength value and the attribute feature vector.

[0024] The process of performing intensity coupling analysis on the flow dependence strength value and the attribute feature vector to obtain the labor flow fusion characteristics includes: Each time point in the labor mobility numerical sequence is mapped to an embedding vector, which consists of the value at the time point and the values ​​at the time points before the time point. The embedding vectors are stacked as matrices to obtain the delay embedding matrix of the labor mobility numerical sequence. The delayed embedding matrix is ​​decomposed into a low-rank form to obtain a sequence of singular values ​​of the delayed embedding matrix. The long-term dependency coefficient is inverted based on the decay rate of the singular value sequence, and the long-term dependency coefficient is nonlinearly combined with the trace norm of the delay embedding matrix to obtain the flow dependency strength value.

[0025] Historical labor mobility data for the target region is time-series aligned to obtain a numerical sequence of labor mobility data for the target region. Raw records of labor inflows and outflows are collected for the target region at multiple consecutive discrete time steps, each time step corresponding to a fixed time interval, such as one month. The number of labor flows from all source regions to all target regions within each time step is summarized to form the total flow value for that time step. These values ​​are arranged chronologically to form a one-dimensional numerical list, which is the labor mobility numerical sequence. Each value in the sequence corresponds to a specific time step, and the magnitude of the value reflects the total amount of labor mobility within that time step.

[0026] The regional attributes of the target region are analyzed to obtain the attribute feature vector. Multidimensional static attributes of the target region are extracted, including total area, resident population, GDP, average wage level, unemployment rate, industrial structure coefficient, and transportation convenience score. The raw values ​​of these attributes are normalized, mapping each attribute value to the interval between zero and one, eliminating dimensional differences between different attributes. All normalized attribute values ​​are arranged in a fixed dimensional order to form a multidimensional numerical vector, which is the attribute feature vector. Each dimension of the vector represents a specific regional attribute, and the value in each dimension indicates the relative strength of that attribute after normalization.

[0027] Autocorrelation is measured on the numerical series of labor mobility in the target region to obtain the mobility dependence strength value. A copy of the labor mobility numerical series is made and shifted forward by one time step to obtain a lagged series. The sum of the products of the original and lagged series at each corresponding position is calculated, and then divided by the series length to obtain the autocovariance of the first lag. The variance of the original series is used as the denominator, and the first-lag autocovariance is divided by this variance to obtain the first-lag autocorrelation coefficient. Similarly, the autocorrelation coefficients for the second, third, and subsequent lags up to multiple time steps are calculated. The average of the absolute values ​​of all calculated autocorrelation coefficients is taken as the mobility dependence strength value. The larger this value, the stronger the temporal dependence of the labor mobility numerical series, and the more significant the influence of historical values ​​on current values.

[0028] A strength coupling analysis is performed on the flow dependency strength value and attribute feature vector to obtain the labor flow fusion feature. The flow dependency strength value is used as a scalar factor, multiplied by the value of each dimension in the attribute feature vector to obtain a weighted attribute vector. Simultaneously, the value of the most recent time step in the labor flow numerical sequence is taken as the current flow state value, and this current flow state value is copied into a vector with the same number of dimensions as the attribute feature vector to form a flow state vector. The weighted attribute vector and the flow state vector are added dimension by dimension to obtain the fusion vector. A nonlinear activation function is applied to this fusion vector to compress the values ​​of each dimension to a suitable dynamic range. The transformed fusion vector is the labor flow fusion feature. This fusion feature simultaneously includes the region's static attribute information, the current flow state information, and the autocorrelation dependency strength information of the historical flow sequence.

[0029] Each time point in the labor mobility numerical sequence is mapped to an embedding vector, which consists of the value at that time point and the values ​​of the time points preceding it. These embedding vectors are then stacked to obtain the delayed embedding matrix of the labor mobility numerical sequence. An embedding dimension is determined, which specifies the number of time points included in each embedding vector. For each time point in the labor mobility numerical sequence, the value at that time point and the values ​​of several consecutive time points preceding it are extracted; the number of these values ​​equals the embedding dimension. These values ​​are arranged in chronological order from earliest to latest into a one-dimensional vector, which is the embedding vector corresponding to that time point. The first few time points in the sequence are padded with zero values ​​because there are insufficient preceding time points. All the embedding vectors corresponding to all time points are arranged in chronological order from top to bottom, forming a two-dimensional matrix. The number of rows in this matrix represents the number of time points, and the number of columns represents the embedding dimension. This two-dimensional matrix is ​​the delayed embedding matrix.

[0030] A low-rank decomposition of the delayed embedding matrix yields its singular value sequence. The delayed embedding matrix is ​​treated as a rectangular numerical table, and decomposed into a product of three matrices using an iterative numerical computation method. The column vectors of the first matrix are mutually orthogonal and have normalized lengths. The second matrix is ​​a diagonal matrix with a series of non-negative values ​​arranged from top left to bottom right along its diagonal in descending order. The row vectors of the third matrix are mutually orthogonal and have normalized lengths. The non-negative values ​​on the diagonal of the second matrix are the singular values ​​of the delayed embedding matrix. These singular values ​​are then extracted in descending order to form a one-dimensional numerical list, which is the singular value sequence. Each value in the singular value sequence reflects the energy of the delayed embedding matrix in different directions.

[0031] The long-term dependency coefficient is derived from the decay rate inversion of singular value sequences. The first singular value in the sequence is calculated, followed by the arithmetic mean of the second to last singular values. The first singular value is divided by this arithmetic mean to obtain a ratio. The ratios of all adjacent singular values ​​in the sequence are then calculated (second divided by first, third by second, and so on), resulting in a series of ratios. The arithmetic mean of these ratios is then calculated as the average decay rate. The first singular value is multiplied by the average decay rate, and then divided by the sum of the first singular value and the arithmetic mean. The result is the long-term dependency coefficient. A larger coefficient indicates a stronger long-term dependency in the labor mobility data sequence, and a more persistent impact of the sequence's historical information on the present and future.

[0032] The mobility dependency strength value is obtained by nonlinearly combining the long-term dependency coefficient with the trace norm of the delayed embedding matrix. The trace norm of the delayed embedding matrix is ​​calculated, which is equal to the sum of all singular values ​​in the singular value sequence. The long-term dependency coefficient is used as a weighting factor, and the product of this weighting factor and the trace norm is calculated. Simultaneously, the difference between this weighting factor and the result is calculated and multiplied by the square root of the trace norm. These products are then added together to obtain the combined value. A nonlinear compression function is applied to this combined value, mapping any non-negative input value to the interval between zero and one. The resulting value is the mobility dependency strength value. This strength value comprehensively reflects the autocorrelation strength and overall energy scale of the labor mobility numerical sequence; a larger value indicates a more significant autocorrelation of the sequence.

[0033] The beneficial effects are as follows: by mapping the numerical sequence of labor mobility to a delayed embedding matrix and performing low-rank decomposition, the long-term dependency coefficient is derived based on the singular value decay rate. This coefficient is then nonlinearly combined with the trace norm to obtain the mobility dependency strength value. Finally, this strength value is coupled with the attribute feature vector for strength analysis, generating a fused feature of labor mobility that integrates historical autocorrelation dependency information and regional static attributes. This process overcomes the shortcomings of traditional methods that simply superimpose or ignore the influence of temporal autocorrelation on attribute features. It achieves accurate quantification and effective fusion of the inherent long-term dependencies in labor mobility data, enabling the fused feature to simultaneously characterize historical evolution patterns and represent inherent regional attributes, providing a more complete and accurate information foundation for subsequent spatial modeling.

[0034] S2. Based on the labor mobility fusion characteristics, spatial adjacency relationship encoding is performed on the target area to obtain the labor mobility map of the target area; In this embodiment of the invention, the labor mobility fusion feature encodes the spatial adjacency relationship of the target region to obtain a labor mobility map of the target region, including: The labor mobility integration characteristics are analyzed to obtain the identifier list and association strength matrix of the target region; Using the identifier list as the node list and the association strength matrix as the edge list, a labor flow graph of the target region is constructed.

[0035] The labor mobility integration characteristics are analyzed to obtain a list of identifiers and a correlation strength matrix for the target area. The labor mobility integration characteristics are a multi-dimensional data structure containing identifier information for all sub-regions within the target area and labor mobility correlation information between these sub-regions. Unique codes for all sub-regions are extracted from the labor mobility integration characteristics; each code represents an independent region node. These codes are arranged in any fixed order to form a list, which is the identifier list. A two-dimensional numerical table is extracted from the labor mobility integration characteristics, where the rows and columns correspond to the sub-region codes in the identifier list. The values ​​at the rightmost position in the table represent the sub-region codes. Line 1 The numerical value of the column position indicates the first column in the identifier list. The region and the first The strength of labor mobility between regions is a two-dimensional numerical table that shows the frequency or closeness of labor mobility between the two regions.

[0036] A labor mobility graph for the target region is constructed using an identifier list as the node list and an association strength matrix as the edge list. An empty directed graph is created, consisting of nodes and directed edges connecting them. Each sub-region in the identifier list is encoded as a node in the directed graph, with the number of nodes equal to the length of the identifier list. For each non-zero value in the association strength matrix, its row and column indices are determined. The node corresponding to the row index becomes the starting node of the directed edge, and the node corresponding to the column index becomes the ending node. The non-zero value is used as the weight of this directed edge. All directed edges corresponding to non-zero values ​​and their weights are added to the directed graph; this set of directed edges constitutes the edge list. After adding all nodes and edges, the resulting directed graph is the labor mobility graph. This labor mobility graph comprehensively depicts the spatial adjacency relationships and intensity distribution of labor mobility between sub-regions within the target region.

[0037] The beneficial effect is that by analyzing the fusion characteristics of labor mobility, the identifier list of each sub-region within the target area and the correlation strength matrix between them are directly extracted. A labor mobility graph is then constructed using the identifier list as nodes and the correlation strength matrix as edge list. This labor mobility graph transforms the originally scattered fusion characteristics into a structured graph data form, clearly and completely depicting the spatial adjacency relationships between nodes in each region and the distribution of labor mobility intensity. This provides a standardized graph structure input for subsequent spatial self-attention mapping and neighborhood feature aggregation, avoiding prediction bias caused by ambiguity or missing representations of inter-regional relationships.

[0038] S3. Perform spatial self-attention mapping on the labor flow map to obtain the spatial interaction characteristics of labor flow of nodes in the target region; In this embodiment of the invention, the step of performing spatial self-attention mapping on the labor flow map to obtain the spatial interaction features of labor flow of nodes in the target region includes: Spatial dependency strength quantification is performed on any two nodes in the labor mobility graph to obtain the association strength value between the two nodes; The attention distribution of the association strength value is reshaped to obtain the attention weights corresponding to the two nodes. Based on the attention weights, the labor flow fusion features of the nodes in the labor flow graph are reorganized to obtain the spatial interaction features of labor flow in the target region.

[0039] The step of quantifying the spatial dependency strength of two nodes in the labor mobility graph to obtain the association strength value between the two nodes includes: The formula for calculating the correlation strength value is as follows: ; in, These are regions with a large outflow of labor. Target areas for labor inflow The value representing the correlation strength between the labor outflow source region and the labor inflow region is the labor mobility correlation strength value. This is the basic attraction coefficient between the labor outflow source region and the labor inflow target region. and These are the learnable push factor vector of the labor outflow source region and the learnable pull factor vector of the labor inflow target region, respectively. The economic or geographical distance between the source region of the labor outflow and the target region of the labor inflow. Let be the rate of change in labor mobility from the labor outflow source region to the labor inflow target region at the k-th time step. The preset history window length, The preset push-pull factor strength parameters, The preset distance attenuation intensity parameter, The preset historical inertia scaling parameters, The preset distance scale parameters, Here, t represents the weight of the k-th order historical inertia, and t is the discrete time step index. For the Sigmoid function, It is the hyperbolic tangent function. It is a natural exponential function.

[0040] Spatial dependency strength is quantified for any two nodes in the labor mobility graph to obtain the correlation strength value between the two nodes. For the source node of labor outflow and the target node of labor inflow, the pre-stored basic attraction coefficient between the two nodes is first obtained. Then, a learnable push factor value is extracted from the labor mobility fusion characteristics of the source node and a learnable pull factor value is extracted from the labor mobility fusion characteristics of the target node. The push factor value and the pull factor value are multiplied and then divided by the product of the length of the push factor vector and the length of the pull factor vector to obtain the push-pull similarity. After multiplying the push-pull similarity by a preset push-pull factor strength parameter, the natural exponential function is applied for transformation to obtain the push-pull influence factor. At the same time, the economic distance or geographical distance value between the source node and the target node is obtained. The distance value is divided by a preset distance scale parameter, minus one, and then multiplied by a distance decay strength parameter. The result is used as the input of the natural exponential function to obtain the distance decay factor. One plus the distance decay factor is used as the denominator, and the basic attraction coefficient multiplied by the push-pull influence factor is used as the numerator to calculate the spatial basic dependency value. Furthermore, the change rate series of labor mobility from the source node region to the target node region is obtained at multiple historical time steps. For each change rate, a sigmoid function is applied for nonlinear compression. The compressed values ​​for each time step are multiplied by the corresponding order of historical inertia weights. All products are summed and then divided by the sum of the historical inertia weights of each order to obtain the weighted average historical impact value. This weighted average historical impact value is multiplied by a historical inertia scaling parameter and then scaled using a hyperbolic tangent function to obtain the historical inertia impact factor. The spatial basis dependency value is multiplied by the historical inertia impact factor; the resulting product is the correlation strength value between two nodes. This process is repeated for each pair of nodes in the labor mobility graph to obtain the correlation strength values ​​between all node pairs.

[0041] Attention distribution reshaping is performed on the association strength values ​​to obtain the attention weights corresponding to the two nodes. For each target node in the labor mobility graph, the association strength values ​​from all source nodes to that target node are collected, forming a list of association strength values. The maximum value in this list is found, and each association strength value in the list is subtracted from the maximum value to obtain a set of offset values. The natural exponential function is applied to each offset value to calculate a set of exponential values. The sum of these exponential values ​​is calculated, and each exponential value is divided by the sum to obtain the attention weight from the corresponding source node to the target node. The sum of all attention weights from source nodes to the target node equals one. The above process is repeated for each target node in the labor mobility graph to obtain the attention weights for each pair of source nodes to the target node.

[0042] Based on attention weights, the labor mobility fusion features of nodes in the labor mobility graph are recombined to obtain the spatial interaction features of labor mobility in the target region. For each target node in the labor mobility graph, the labor mobility fusion feature vectors of all source nodes are obtained. Each dimension value in the labor mobility fusion feature vector of each source node is multiplied by the attention weight from that source node to the target node to obtain the weighted feature vector of that source node. The weighted feature vectors of all source nodes are summed element-wise according to their corresponding dimensions to obtain the recombined feature vector of the target node. This recombined feature vector aggregates the influence of all source nodes on the target node, and its dimension is the same as that of the original labor mobility fusion feature vector. The recombined feature vectors of all target nodes in the labor mobility graph are arranged according to the order of the nodes in the identifier list to form a two-dimensional array, which is the spatial interaction feature of labor mobility. Each row in this feature corresponds to a node, and each column corresponds to a feature dimension. The values ​​reflect the interaction patterns between nodes under the spatial self-attention mechanism.

[0043] The basic attraction coefficient is directly calculated from the historical total labor mobility statistics between the source and target regions. Specifically, it is obtained by collecting the total number of labor mobility trips from the source region to the target region over multiple past time steps, then dividing this sum by the total number of trips between all region pairs to obtain a normalized basic attraction coefficient. This coefficient reflects the inherent labor mobility preference between the two regions without considering other factors. When the basic attraction coefficient increases, the association strength value increases proportionally, indicating a stronger natural attraction between the two regions; when the basic attraction coefficient decreases, the association strength value decreases linearly.

[0044] The learnable push and learnable pull factor vectors are parameters automatically updated through backpropagation during model training. The learnable push factor vector represents the source region's ability to push labor outwards, while the learnable pull factor vector represents the target region's ability to attract labor. During model initialization, the learnable push and learnable pull factor vectors for each region are set to random, small numerical vectors. Then, as the training data iterates, the model continuously adjusts these vector values ​​based on the prediction error. The learnable push and learnable pull factor vectors have the same length, typically set to a fixed dimension, such as 32. When the overall value in the learnable push factor vector increases, it indicates an enhanced willingness to push labor from the source region, which makes the labor attracted from the target region more attractive. The association strength values ​​from the source region to all target regions increase. When the overall value in the learnable pull factor vector increases, it indicates that the attraction of the target region is enhanced, which will increase the association strength values ​​from all source regions to the target region. Changes in the vector direction of the learnable push factor vector and the learnable pull factor vector also affect the matching degree between a specific source region and a specific target region. The more consistent the direction, the higher the push-pull similarity and the greater the association strength value.

[0045] The push-pull factor strength parameter is a preset hyperparameter, manually set before model training and kept constant during training. This parameter scales the influence of push-pull similarity on the association strength value. When the push-pull factor strength parameter is large, small changes in push-pull similarity are amplified by the exponential function, making the association strength value extremely sensitive to the matching degree between the learnable push factor vector and the learnable pull factor vector. When the push-pull factor strength parameter is small, even if the push-pull similarity is high, the exponentially transformed value will not be too large, thus reducing the dependence of the association strength value on the push-pull factor.

[0046] Distance refers to the economic or geographical distance between the source and target regions, a value directly derived from regional attribute data. Economic distance is calculated by the absolute difference in per capita GDP between the two regions, while geographical distance is calculated by the actual distance in kilometers between the center points of the two regions. In practical applications, either one distance can be chosen, or a weighted combination of the two distances can be used. As the distance increases, the fractional value in the distance decay factor increases, leading to a larger exponential term in the denominator and a larger overall denominator. This results in a decrease in the spatial basis dependency value; that is, the greater the distance between the two regions, the weaker the correlation strength.

[0047] The distance scale parameter is a preset hyperparameter used to normalize distance values. Physically, this parameter represents the half-life of the distance. When the value is large, the value obtained by dividing the same distance by the distance scale parameter becomes smaller, resulting in a gentler distance decay effect, maintaining a high correlation strength even between distant regions. Conversely, when the distance scale parameter is small, the value obtained by dividing the same distance by the distance scale parameter becomes larger, leading to a more severe distance decay effect, and only very close regions can exhibit a significant correlation strength.

[0048] The distance attenuation strength parameter is a preset hyperparameter used to control the nonlinearity of the influence of distance on the correlation strength value. When the distance attenuation strength parameter is large, the exponent is very sensitive to changes in distance; even a slightly greater distance will cause the denominator to increase sharply, thus significantly reducing the correlation strength value. When the distance attenuation strength parameter is zero, the denominator is always equal to two, and distance has no effect on the correlation strength value. When the distance attenuation strength parameter is positive, the larger the distance, the smaller the correlation strength value; the larger the distance attenuation strength parameter, the faster the attenuation.

[0049] The historical window length is a preset hyperparameter that indicates how many past time steps of labor mobility change rate are considered when calculating historical inertia effects. This value is set manually based on the time granularity of the data; for example, monthly data could be set to twelve, indicating that the historical effects of the past year are considered. Increasing the historical window length incorporates information from more historical time steps, allowing the correlation strength value to reflect longer-term trends, but also increases the computational load. Conversely, decreasing the historical window length focuses only on recent changes, resulting in a faster response to sudden changes, but may miss long-term cyclical patterns.

[0050] No. The historical inertia weights are parameters learned through backpropagation during model training. The offset at each time step... Each historical inertia weight corresponds to an independent weight value, representing the degree to which the historical rate of change at that time step influences the current association strength value. During model initialization, all historical inertia weights are set to the same positive value. As training progresses, the model automatically adjusts the magnitude of each weight. Typically, historical inertia weights corresponding to more recent time steps learn larger values, while those corresponding to more distant time steps learn smaller values. When a certain historical inertia weight increases, the influence of the rate of change in labor mobility at that time step on the current association strength value strengthens; conversely, when the historical inertia weight decreases, the historical information of that time step is weakened.

[0051] The rate of change in labor mobility is calculated directly from historical data. Specifically, it is obtained in the [missing information - likely a date or time]. At the first time step, calculate the labor mobility from the source region to the target region, and then calculate the second time step. The rate of change of labor mobility at one time step is calculated by subtracting the next value from the previous value and then dividing by the next value. This rate of change reflects the dynamic fluctuations in labor mobility. When the rate of change is positive, it indicates that mobility is increasing, and after compression by the Sigmoid function, it will yield a value greater than 0.5, thus strengthening the historical inertia factor. When the rate of change is negative, it indicates that mobility is decreasing, and the Sigmoid output is less than 0.5, weakening the historical inertia factor.

[0052] The historical inertia scaling parameter is a preset hyperparameter used to control the output range of the historical inertia influence factor after passing through the hyperbolic tangent function. The output range of the hyperbolic tangent function is from -1 to +1. The historical inertia scaling parameter is multiplied by the weighted average historical influence value before being input into the hyperbolic tangent function. When the historical inertia scaling parameter is large, even if the historical influence value is small, it will approach the extreme value of +1 or -1 after passing through the hyperbolic tangent, causing the output of the historical inertia influence factor to tend to saturate. When the historical inertia scaling parameter is small, the output of the historical inertia influence factor is approximately linear, and the change is relatively gentle. The role of the historical inertia scaling parameter is to adjust the upper limit of the contribution of historical information to the correlation strength value.

[0053] The beneficial effects are as follows: By quantifying the spatial dependency strength of any two nodes in the labor mobility map and introducing the joint calculation of multiple nonlinear factors such as push-pull factors, distance decay, and historical inertia, a correlation strength value that can truly reflect the complex flow relationships between regions is generated. Then, the correlation strength value undergoes attention distribution reshaping to normalize and make the attention weights of each node pair comparable. Finally, based on the attention weights, the labor mobility fusion characteristics of the nodes are weighted and recombined to obtain spatial interaction characteristics of labor mobility rich in spatial interaction information. This process overcomes the shortcomings of fixed adjacency matrices or simple distance metrics in dynamically characterizing asymmetric, time-varying spatial dependencies, achieving adaptive quantification and feature reconstruction of interactions between nodes, and providing an accurate and realistic spatial feature expression for subsequent neighborhood feature aggregation.

[0054] S4. Perform neighborhood feature aggregation on the spatial interaction features of the labor mobility to obtain the labor mobility features of the target region; In this embodiment of the invention, the step of aggregating neighborhood features of the spatial interaction features of labor mobility to obtain the labor mobility features of the target region includes: The proximity relationship quantification of the labor flow graph yields the distribution of neighborhood influence intensity between neighboring nodes in the labor flow graph. Based on the neighborhood influence intensity distribution, the labor mobility spatial interaction features of neighboring nodes in the labor mobility graph are weighted and fused to obtain the weighted aggregated neighborhood features of the neighboring nodes. The spatial interaction features of labor mobility of the node are coupled with the corresponding weighted aggregated neighborhood features under gating adaptive coupling to obtain the fusion features of the node. The fusion features of the node are then spliced ​​together in node order to form the labor mobility features of the target region.

[0055] The labor mobility graph is quantified by proximity relations to obtain the distribution of neighborhood influence intensity between neighboring nodes. An edge list of the labor mobility graph is obtained, where each edge records a directed connection from a source node to a target node and its weight value, which represents the original neighborhood influence intensity between the two nodes. For each node in the labor mobility graph, all source nodes pointing to that node and their corresponding edge weights are collected and listed. The maximum value in this list is found, and this maximum value is subtracted from each edge weight in the list to obtain a set of offset values. The natural exponential function is applied to each offset value to obtain a set of exponential values. The sum of these exponential values ​​is calculated, and each exponential value is divided by the sum to obtain the distribution of the neighborhood influence intensity intensity of the corresponding source node to the target node. The sum of the neighborhood influence intensity distribution values ​​from all source nodes to the same target node equals one. The above process is repeated for each node in the labor mobility graph to obtain the neighborhood influence intensity distribution between all nodes and their neighboring nodes.

[0056] Based on the distribution of neighborhood influence intensity, the spatial interaction features of labor mobility among neighboring nodes in the labor mobility graph are weighted and fused to obtain the weighted aggregated neighborhood features of neighboring nodes. For each target node in the labor mobility graph, the spatial interaction feature vectors of labor mobility corresponding to all source nodes pointing to that target node are obtained. The spatial interaction features of labor mobility are a two-dimensional array, with each row corresponding to a node and each column corresponding to a feature dimension. The row vector corresponding to each source node is extracted from this two-dimensional array. Each dimension value in the row vector of each source node is multiplied by the neighborhood influence intensity distribution value of that source node on the target node to obtain the weighted neighborhood feature vector of that source node. The weighted neighborhood feature vectors of all source nodes are summed element-wise according to their corresponding dimensions to obtain a new vector with the same dimensions as the original row vectors. This new vector is the weighted aggregated neighborhood feature of the target node. This feature aggregates the comprehensive influence of all neighboring nodes on the target node at the spatial interaction level.

[0057] The spatial interaction features of labor mobility at a node are adaptively coupled with the corresponding weighted aggregated neighborhood features to obtain the node's fused features. These fused features are then concatenated sequentially to form the labor mobility features of the target region. For each node in the labor mobility graph, the spatial interaction feature vector of that node and the corresponding weighted aggregated neighborhood feature vector are obtained. These two vectors are concatenated along their feature dimensions to form a combined vector of twice the length. An element-wise nonlinear transformation using a sigmoid function is applied to this combined vector to obtain a set of gating coefficient vectors with values ​​between zero and one. Simultaneously, the spatial interaction feature vector of labor mobility and the weighted aggregated neighborhood feature vector are added element-wise to obtain the preliminary fused vector. Each value in the gating coefficient vector is multiplied by the corresponding value in the preliminary fused vector to obtain the gating-adjusted vector. Then, one minus each value in the gating coefficient vector is used as a complementarity coefficient, which is multiplied by the corresponding value in the original spatial interaction feature vector of the node to obtain the complementarity-preserved vector. The gating-adjusted vector and the complementarity-preserved vector are added element-wise to obtain the fused feature vector of that node. This fused feature vector retains the spatial interaction information of each node while selectively incorporating the aggregation effects of neighboring nodes. Following the order of nodes in the identifier list in the labor flow map, the fused feature vectors of all nodes are arranged sequentially to form a two-dimensional array, which represents the labor flow characteristics of the target region. Each row in this feature array corresponds to a node, and each column corresponds to a feature dimension, fully encoding the final feature representation of each node under the influence of its spatial neighborhood.

[0058] The beneficial effects are as follows: by quantifying the proximity relationships in the labor mobility map, an accurate distribution of neighborhood influence intensity is generated. Based on this distribution, the spatial interaction characteristics of labor mobility of neighboring nodes are weighted and fused to obtain weighted aggregated neighborhood features. Then, through gated adaptive coupling, the spatial interaction characteristics of the nodes themselves are dynamically fused with the aggregated neighborhood features, generating labor mobility features that retain individual node information while incorporating the comprehensive influence of the neighborhood. This process overcomes the shortcomings of fixed adjacency weights or simple average aggregation, which cannot adapt to the differences in influence intensity between nodes. It achieves adaptive weighting of neighborhood features and nonlinear gated fusion of node features, enabling labor mobility features to more accurately reflect the modulation effect of local spatial structure on labor mobility behavior.

[0059] S5. Perform time-dependent encoding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target region; In this embodiment of the invention, the step of performing time-dependent encoding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target region includes: The labor mobility characteristics are segmented into time-series windows to obtain a time-series mobility characteristic sequence for the target region. Perform temporal dependency mapping on the features of two adjacent time steps in the temporal flow feature sequence to obtain the state transition intensity value of the temporal flow feature sequence; Based on the state transition intensity value, feature progressive propagation is performed on the temporal flow feature sequence to obtain the cumulative temporal features corresponding to each time step; The accumulated time-series features are aggregated along the time-series dimension to obtain the time-series features of labor mobility in the target region.

[0060] Labor mobility features are segmented into time-series windows to obtain a sequence of time-series mobility features for the target region. Each labor mobility feature is a three-dimensional data block containing feature vectors for each node across multiple consecutive time steps. A fixed window length is set, which determines the number of time steps processed each time step. Starting from the first time step, all node features across the corresponding time steps of the next consecutive window length are extracted to form a window data block. This entire window data block is stored as an element in a list. The window is then slid forward one time step, extracting all node features across the next consecutive window length to form the next window data block, which is also stored in the list. This process of sliding the window forward and extracting data blocks is repeated until the last time step is also included in a window. Each element in the final list is a window data block, arranged chronologically to form the sequence of time-series mobility features. Each window data block contains complete feature information for all nodes across all time steps within the window.

[0061] Temporal dependency mapping is performed on the features of two adjacent time steps in the time-series flow feature sequence to obtain the state transition intensity value of the time-series flow feature sequence. For each window data block in the time-series flow feature sequence, all node features of the last time step in that window are extracted to form the current state feature set. Then, all node features of the first time step in the next adjacent window data block are extracted to form the next state feature set. The feature vector of each node in the current state feature set is subtracted element-wise from the feature vector of the corresponding node in the next state feature set to obtain the difference vector. The sum of squares of all elements in the difference vector is calculated, and the square root is taken to obtain the state difference value of the node between the current window and the next window. The arithmetic mean of the state difference values ​​of all nodes is calculated, and this mean is the state transition intensity value between adjacent windows. The above calculation is repeated for each pair of adjacent windows in the time-series flow feature sequence to obtain a series of state transition intensity values. These values ​​are arranged in chronological order, reflecting the drastic change of labor mobility state over time.

[0062] Based on the state transition intensity value, a feature propagation process is performed on the temporal flow feature sequence to obtain the cumulative temporal features corresponding to each time step. A cumulative feature array is initialized, using all node features from the last time step in the first window data block as the initial cumulative features. For each subsequent state transition intensity value, this intensity value is used as the decay coefficient for the current time step. Each node feature vector in the cumulative features of the previous time step is multiplied by one and subtracted from the decay coefficient to obtain the retained portion. Simultaneously, all node features from the last time step in the current window data block are obtained, and each node feature vector is multiplied by the decay coefficient to obtain the updated portion. The retained portion and the updated portion are added node-by-node and element-by-element to obtain the cumulative feature for the current time step. This recursive calculation is performed sequentially for each time step, generating a cumulative feature for each time step. The cumulative features of all time steps are arranged in chronological order to form a three-dimensional data block. Each time step in this data block corresponds to a cumulative feature, which integrates all historical information from the initial time to the current time.

[0063] The cumulative time-series features are aggregated along the time-series dimension to obtain the labor mobility time-series features of the target region. The cumulative time-series features are a three-dimensional data block with three dimensions: time step, node, and feature. Along the time step dimension, the feature vectors of all nodes at each time step are considered as a whole. The average value of the same feature dimension for the same node across all time steps is calculated, resulting in a two-dimensional array where rows correspond to nodes and columns correspond to feature dimensions. The variance of the values ​​for the same feature dimension for the same node across all time steps is then calculated, resulting in another two-dimensional array of the same dimension. The average and variance arrays are concatenated along the feature dimension to form a new two-dimensional array with the same number of nodes but doubled feature dimensions. This new two-dimensional array represents the labor mobility time-series features. Each row in this feature corresponds to a node, and each column corresponds to a statistical descriptive dimension, comprehensively depicting the central tendency and dispersion of each node's features along the time dimension.

[0064] The beneficial effects are as follows: by segmenting labor mobility characteristics into a sequence through time-series windows, calculating the state transition intensity value between adjacent windows, and then performing feature progressive propagation based on this intensity value, the cumulative time-series features at each time step integrate all historical information from the initial moment to the current moment. Finally, through feature aggregation along the time-series dimension, a labor mobility time-series feature containing both mean and variance is obtained. This method overcomes the shortcomings of traditional time-series models, such as difficulty in effectively capturing long-term dependencies and state abrupt changes, and slow response to periodic trends. It achieves adaptive decay and recursive fusion of historical information, so that the final output time-series feature reflects both the central trend of labor mobility and its fluctuation dispersion, providing a more stable and information-rich feature foundation for subsequent global weighted fusion prediction.

[0065] S6. Perform global feature weighted fusion on the time-series characteristics of labor mobility to obtain the predicted value of labor mobility in the target area.

[0066] In this embodiment of the invention, the step of performing global feature weighted fusion on the temporal features of labor mobility to obtain the predicted value of labor mobility in the target area includes: Global contribution decoupling is performed on each feature dimension in the time series characteristics of labor mobility to obtain the adaptive importance weight coefficient corresponding to each feature dimension; Based on the adaptive importance weight coefficients, the feature dimensions in the temporal features of labor mobility are scaled and gated dimension by dimension to obtain the weighted calibrated feature representation of the target region; The weighted calibration feature representation is subjected to nonlinear response synthesis from the feature space to the prediction numerical space to obtain the predicted labor flow values ​​for the target region.

[0067] Global contribution decoupling is performed on each feature dimension in the labor mobility time-series features to obtain adaptive importance weight coefficients for each feature dimension. The labor mobility time-series features are a two-dimensional array, where rows correspond to nodes in the target region, and columns correspond to different feature dimensions of each node. Element-wise summation of all rows and columns of this two-dimensional array yields a one-dimensional vector. The length of this vector equals the total number of feature dimensions, and each value represents the cumulative sum of the corresponding feature dimension across all nodes. Each value in this one-dimensional vector is divided by the sum of all values ​​to obtain a set of normalized values, each corresponding to the global original contribution of a feature dimension. A non-linear transformation function is applied to the global original contribution of each feature dimension for recalibration. This transformation function maps the input values ​​to another numerical range, moderately increasing excessively small contributions and compressing excessively large contributions. The resulting set of values ​​is used as the adaptive importance weight coefficient for each feature dimension. The value of this weight coefficient is between zero and one, and the sum of the weight coefficients for all feature dimensions is not necessarily one. Each weight coefficient independently reflects the importance of the corresponding feature dimension to the final prediction result.

[0068] Based on adaptive importance weighting coefficients, a dimension-wise gated scaling is applied to the feature dimensions of the labor mobility time-series features to obtain a weighted calibrated feature representation of the target region. Each column of the two-dimensional array of labor mobility time-series features—that is, the column vector composed of the values ​​of each feature dimension across all nodes—is extracted. For the k-th feature dimension, the adaptive importance weighting coefficient corresponding to that feature dimension is obtained, and this coefficient is multiplied by each value in the column vector of that feature dimension to obtain a scaled column vector. This process is repeated for each feature dimension to obtain a set of scaled column vectors. These scaled column vectors are then concatenated according to the original feature dimension order to form a two-dimensional array with the same number of rows and columns as the labor mobility time-series features; this two-dimensional array is the weighted calibrated feature representation. In this representation, feature dimensions that originally contributed significantly to prediction have their numerical amplitude amplified, while feature dimensions that contributed less have their numerical amplitude suppressed, thus adaptively adjusting the influence of each feature dimension.

[0069] A nonlinear response synthesis from the feature space to the predicted numerical space is performed on the weighted calibration feature representation to obtain the predicted labor mobility value for the target region. Treating the weighted calibration feature representation as a whole input, all values ​​in the two-dimensional array are first flattened to form a one-dimensional vector. This one-dimensional vector is then input into a multi-layer nonlinear mapping structure consisting of multiple cascaded computational layers. In each layer, the input vector is first multiplied by a pre-trained weight matrix to obtain an intermediate vector. Then, a bias value is added to each value in this intermediate vector. Finally, a nonlinear activation function is applied to each value in the biased vector, and the result is output to the next layer. After layer-by-layer nonlinear transformation, a scalar value is finally obtained, which is the predicted labor mobility value for the target region. This predicted value represents the estimated total labor mobility of the target region within a specified future time step, given historical data and regional attributes.

[0070] The beneficial effect is that by decoupling the global contribution of each feature dimension in the time-series characteristics of labor mobility, independent adaptive importance weight coefficients are obtained for each dimension. Based on this, the feature dimensions are then subjected to dimension-wise gating and scaling to obtain a targeted calibrated weighted feature representation. Finally, this representation is mapped to the prediction numerical space through multi-layer nonlinear response synthesis. This method overcomes the shortcomings of traditional methods, such as equal-weighted feature processing or fixed-weight fusion, which are difficult to adapt to the dynamic differences in the contributions of different features. It achieves adaptive weighting and nonlinear recombination of feature dimensions, enabling the prediction model to highlight key features and suppress redundant noise, thereby significantly improving the accuracy and generalization ability of labor mobility prediction values.

[0071] like Figure 2 The diagram shown is a functional block diagram of a spatiotemporal graph network-based labor mobility prediction system provided in an embodiment of the present invention.

[0072] The spatiotemporal graph network-based labor mobility prediction system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the spatiotemporal graph network-based labor mobility prediction system may include a data fusion and parsing module 101, a spatial encoding module 102, a spatial self-attention module 103, a feature aggregation module 104, a temporal encoding module 105, and a global weighted fusion module 106. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0073] In this embodiment, the functions of each module / unit are as follows: The data fusion and parsing module 101 is used to collaboratively interpret the historical labor flow data and regional attribute data of the target area to obtain the labor flow fusion characteristics of the target area. The spatial coding module 102 is used to encode the spatial adjacency relationship of the target area based on the labor flow fusion characteristics to obtain the labor flow map of the target area; The spatial self-attention module 103 is used to perform spatial self-attention mapping on the nodes of the labor flow graph to obtain the spatial interaction features of labor flow of the nodes in the target region. The feature aggregation module 104 is used to perform neighborhood feature fusion on the spatial interaction features of labor mobility to obtain the labor mobility features of the target area. The time-series coding module 105 is used to perform time-series dependency coding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target area. The global weighted fusion module 106 is used to perform global weighted fusion on the temporal features of labor mobility to obtain the predicted value of labor mobility in the target area.

[0074] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0075] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0077] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0078] The embodiments of this application can acquire and process relevant data based on an artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting labor mobility using spatiotemporal graph networks, characterized in that, The method includes: S1. The historical labor mobility data and regional attribute data of the target area are fused and analyzed to obtain the fused characteristics of the labor mobility of the target area. S2. Based on the labor mobility fusion characteristics, spatial adjacency relationship encoding is performed on the target area to obtain the labor mobility map of the target area; S3. Perform spatial self-attention mapping on the labor flow map to obtain the spatial interaction characteristics of labor flow of nodes in the target region; S4. Perform neighborhood feature aggregation on the spatial interaction features of the labor mobility to obtain the labor mobility features of the target region; S5. Perform time-dependent encoding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target region; S6. Perform global feature weighted fusion on the time-series characteristics of labor mobility to obtain the predicted value of labor mobility in the target area.

2. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 1, characterized in that, The process of fusing and analyzing historical labor mobility data and regional attribute data of the target area to obtain the fused characteristics of labor mobility in the target area includes: By performing time-series alignment on the historical labor mobility data of the target area, a numerical sequence of labor mobility data for the target area is obtained. The regional attributes of the target region are parsed to obtain the attribute feature vector of the target region; Autocorrelation measurement is performed on the numerical series of labor mobility in the target region to obtain the mobility dependence strength value of the target region. The labor mobility fusion characteristics are obtained by performing intensity coupling analysis on the flow dependence strength value and the attribute feature vector.

3. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 2, characterized in that, By performing a strength coupling analysis on the flow dependence intensity value and the attribute feature vector, the labor flow fusion characteristics are obtained, including: Each time point in the labor mobility numerical sequence is mapped to an embedding vector, which consists of the value at the time point and the values ​​at the time points before the time point. The embedding vectors are stacked as matrices to obtain the delay embedding matrix of the labor mobility numerical sequence. The delayed embedding matrix is ​​decomposed into a low-rank form to obtain a sequence of singular values ​​of the delayed embedding matrix. The long-term dependency coefficient is inverted based on the decay rate of the singular value sequence, and the long-term dependency coefficient is nonlinearly combined with the trace norm of the delay embedding matrix to obtain the flow dependency strength value.

4. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 1, characterized in that, Based on the aforementioned labor mobility fusion characteristics, spatial adjacency encoding is performed on the target region to obtain a labor mobility map of the target region, including: The labor mobility integration characteristics are analyzed to obtain the identifier list and association strength matrix of the target region; Using the identifier list as the node list and the association strength matrix as the edge list, a labor flow graph of the target region is constructed.

5. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 1, characterized in that, The step of performing spatial self-attention mapping on the labor mobility map to obtain the spatial interaction features of labor mobility in the target region includes: Spatial dependency strength quantification is performed on any two nodes in the labor mobility graph to obtain the association strength value between the two nodes; The attention distribution of the association strength value is reshaped to obtain the attention weights corresponding to the two nodes. Based on the attention weights, the labor flow fusion features of the nodes in the labor flow graph are reorganized to obtain the spatial interaction features of labor flow in the target region.

6. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 5, characterized in that, The step of quantifying the spatial dependency strength of two nodes in the labor mobility graph to obtain the association strength value between the two nodes includes: in, These are regions with a large outflow of labor. Target areas for labor inflow For the region and region The strength of the labor mobility correlation value, This is the basic attraction coefficient between the labor outflow source region and the labor inflow target region. and These are the learnable push factor vector of the labor outflow source region and the learnable pull factor vector of the labor inflow target region, respectively. The economic or geographical distance between the source region of the labor outflow and the target region of the labor inflow. Let be the rate of change in labor mobility from the labor outflow source region to the labor inflow target region at the k-th time step. The preset history window length, For push-pull factor strength parameters, For distance attenuation intensity parameters, For historical inertia scaling parameters, For distance scale parameters, Here, t represents the weight of the k-th order historical inertia, and t is the discrete time step index. For the Sigmoid function, It is the hyperbolic tangent function. It is a natural exponential function.

7. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 1, characterized in that, The process of aggregating neighborhood features of the spatial interaction characteristics of labor mobility to obtain the labor mobility characteristics of the target region includes: The proximity relationship quantification of the labor flow graph yields the distribution of neighborhood influence intensity between neighboring nodes in the labor flow graph. Based on the neighborhood influence intensity distribution, the labor mobility spatial interaction features of neighboring nodes in the labor mobility graph are weighted and fused to obtain the weighted aggregated neighborhood features of the neighboring nodes. The spatial interaction features of labor mobility of the node are coupled with the corresponding weighted aggregated neighborhood features under gating adaptive coupling to obtain the fusion features of the node. The fusion features of the node are then spliced ​​together in node order to form the labor mobility features of the target region.

8. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 1, characterized in that, The labor mobility characteristics are time-dependently encoded to obtain the time-series characteristics of labor mobility in the target region, including: The labor mobility characteristics are segmented into time-series windows to obtain a time-series mobility characteristic sequence for the target region. Perform temporal dependency mapping on the features of two adjacent time steps in the temporal flow feature sequence to obtain the state transition intensity value of the temporal flow feature sequence; Based on the state transition intensity value, feature progressive propagation is performed on the time-series flow feature sequence to obtain the cumulative time-series features corresponding to each time step; The accumulated time-series features are aggregated along the time-series dimension to obtain the time-series features of labor mobility in the target region.

9. The labor mobility prediction method based on spatiotemporal graph networks as described in claim 1, characterized in that, The step of linearly mapping the temporal characteristics of labor mobility to obtain the predicted labor mobility values ​​for the target region includes: Global contribution decoupling is performed on each feature dimension in the time series characteristics of labor mobility to obtain the adaptive importance weight coefficient corresponding to each feature dimension; Based on the adaptive importance weight coefficients, the feature dimensions in the temporal features of labor mobility are scaled and gated dimension by dimension to obtain the weighted calibrated feature representation of the target region; The weighted calibration feature representation is subjected to nonlinear response synthesis from the feature space to the prediction numerical space to obtain the predicted labor flow values ​​for the target region.

10. A labor mobility prediction system based on spatiotemporal graph networks, characterized in that, The system for implementing the labor mobility prediction method of the spatiotemporal graph network of claim 1 includes: The data fusion and parsing module is used to collaboratively interpret historical labor flow data and regional attribute data of the target area to obtain the labor flow fusion characteristics of the target area. The spatial coding module is used to encode the spatial adjacency relationship of the target area based on the labor mobility fusion characteristics to obtain the labor mobility map of the target area; The spatial self-attention module is used to perform spatial self-attention mapping on the nodes of the labor flow graph to obtain the spatial interaction features of labor flow of the nodes in the target region. The feature aggregation module is used to perform neighborhood feature fusion on the spatial interaction features of the labor mobility to obtain the labor mobility features of the target region. A time-series coding module is used to perform time-series dependency coding on the labor mobility characteristics to obtain the time-series characteristics of labor mobility in the target region. The global weighted fusion module is used to perform global weighted fusion on the temporal features of labor mobility to obtain the predicted value of labor mobility in the target area.

Images