Human resource data intelligent analysis method and system

Abstract

This invention discloses an intelligent analysis method and system for human resource data, relating to the field of data analysis technology. The method includes acquiring human resource data and constructing modular data analysis objects; standardizing these objects to generate a standardized feature vector space; calculating the parameter sensitivity of hyperparameters using sliding window sampling and dividing the hyperparameters of the prediction model; calculating the probability density distribution difference and topological distance difference between adjacent batch windows to obtain the batch data distribution drift; optimizing the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set using a dynamic inertial weighted particle swarm optimization algorithm and a neighborhood adaptive search algorithm, respectively, to train the optimal prediction model; predicting the employee time window feature matrix; generating a prediction result table; and grouping and aggregating the prediction results according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method and system for intelligent analysis of human resource data. Background Technology

[0002] With the advancement of enterprise digital transformation, human resource data is exhibiting high-dimensional and dynamic characteristics. Traditional analysis methods rely on static models, making it difficult to effectively handle employee characteristics across time windows. Meanwhile, large enterprises have complex organizational structures, with frequent departmental adjustments and management level changes, leading to continuous evolution of underlying data and upper-level organizational states. How to accurately capture data distribution drift caused by organizational changes and achieve rapid adaptive optimization of hyperparameters in complex predictive models to provide accurate employee behavior predictions and organizational decision support has become a critical bottleneck problem that urgently needs to be solved in the field of intelligent human resource data analysis.

[0003] Currently, Chinese invention patent application CN112489657A discloses a data analysis system and method, including a data acquisition terminal and a server terminal. The data acquisition terminal includes a voice acquisition module, a voice production module, a fuzzy setting module, a first transmission module, a display module, an operation module, a first storage module, and a first memory module. This invention, through the data acquisition terminal including the voice acquisition module, voice production module, fuzzy setting module, display module, and operation module, can simultaneously record voice and retrieve keywords from fuzzy accents. It can also generate multiple keywords on the display module for users to select, thereby enabling more accurate identification of the desired query information. Simultaneously, the server terminal can memorize the text generated from fuzzy accents, so that when the same fuzzy accent appears again, the corresponding keywords can be directly displayed through the memory function, saving users time and making it more convenient to use. However, related technologies struggle to effectively address the distributional drift of human resource data and lack adaptive capabilities and accuracy. Summary of the Invention

[0004] The technical problem solved by this invention is that related technologies are difficult to effectively cope with the distribution drift of human resources data caused by frequent organizational restructuring and management level changes, and lack the ability and accuracy to adapt to complex organizational topologies.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for intelligent analysis of human resource data includes the following steps: Step S1: Obtain human resources data and construct a modular data analysis object. Standardize the modular data analysis object to generate a standardized feature vector space. The modular data analysis objects include an employee time window feature matrix and an organizational status information table; Step S2: Calculate the parameter sensitivity of hyperparameters using sliding window sampling, and divide the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters. Step S3: Calculate the probability density distribution difference and topological distance difference between adjacent batch windows to obtain the batch data distribution drift. Step S4: Based on the batch data distribution drift, the dynamic inertial weighted particle swarm algorithm and the neighborhood adaptive search algorithm are used to optimize the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set respectively, and the optimal prediction model is trained. Step S5: Use the optimal prediction model to predict the employee time window feature matrix, generate a prediction result table, and perform grouping and aggregation operations on the prediction results according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table.

[0006] Preferably, step S1 specifically includes: Step S11: Obtain human resources data, which includes employee master data, organizational structure data, attendance data, salary data, and performance data; Human resources data is batch-organized according to time batches, and the sample primary key is determined by the employee identifier and time window identifier in the employee master data. The human resources data is aggregated and calculated according to the sample primary key to generate the employee time window feature matrix. Extract the horizontal departmental affiliation relationships and vertical management hierarchy chains from the organizational structure data, construct an organizational classification topology tree consisting of the group, business unit, first-level department and second-level department, and generate an organizational status information table based on the organizational classification topology tree. The organizational status information table includes organizational level code, on-the-job status field, organizational structure version identifier and time batch identifier. The employee time window feature matrix and the organizational status information table are encapsulated into a modular data analysis object according to the same sample primary key; Step S12: Standardize the modular data analysis object to obtain a standardized feature vector space; The standardization process includes sample alignment, feature alignment, and feature standardization.

[0007] Preferably, the calculation of parameter sensitivity includes: Construct a continuous batch window set based on a preset sliding window sampling method; Based on the joint sampling strategy, cooperative perturbation sampling is performed within the range of values ​​of the hyperparameters to be evaluated to construct the combination space of the hyperparameters to be evaluated; In each batch window of the continuous batch window set, the prediction model is cross-validated using the hyperparameter combination space to be evaluated, and the performance evaluation index value of each hyperparameter combination to be evaluated is obtained in each batch window. The variance contribution of each hyperparameter to be evaluated is calculated based on the performance evaluation index value of each combination of hyperparameters to be evaluated under each batch window, and the average value of the variance contribution under all batch windows is used as the parameter sensitivity of the corresponding hyperparameter to be evaluated. If the parameter sensitivity of the hyperparameter to be evaluated does not exceed the sensitivity threshold, it is added to the static sensitive hyperparameter set; if the parameter sensitivity of the hyperparameter to be evaluated exceeds the sensitivity threshold, it is added to the dynamic sensitive hyperparameter set. The set of statically sensitive hyperparameters, the set of dynamically sensitive hyperparameters, and the parameter sensitivity of each hyperparameter to be evaluated are written into a modular data analysis object.

[0008] Preferably, the expression for calculating the variance contribution of the hyperparameter to be evaluated is as follows: ; in, For the first Within each batch window, the hyperparameters to be evaluated Parameter sensitivity, Index of hyperparameters to be evaluated. In the first Within each batch window, the random variable consists of the performance evaluation index values ​​of each hyperparameter combination to be evaluated under each batch window. For the first The representation of each hyperparameter to be evaluated as a random variable in the hyperparameter combination space. To remove Other combinations of hyperparameters, For the first Within each batch window, the performance evaluation index value Total variance For batch window index, For conditional variance, This is the expected operation.

[0009] Preferably, step S3 specifically includes: Step S31: In the standardized feature vector space, for any two adjacent batch windows, extract the corresponding first batch window sample set and second batch window sample set. Step S32: Divide the first batch of window sample sets and the second batch of window sample sets into corresponding organizational node sample subsets, calculate the difference in probability density distribution of organizational nodes in the first batch of window sample sets and the difference in topological distance under the organizational classification topology tree, and use the difference in topological distance as the difference weight to perform a weighted average of the difference in probability density distribution to obtain the batch data distribution drift. The batch data distribution drift is associated with the identifiers of two adjacent batch windows and stored to form a batch drift record table, which is then written into a modular data analysis object.

[0010] Preferably, step S4 specifically includes: Step S41: Using the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set as optimization objects, perform a global search of the static sensitive hyperparameter set using the dynamic inertia weighted particle swarm algorithm to obtain the optimal solution for the static sensitive hyperparameters. Step S42: Using the optimal solution of statically sensitive hyperparameters as a fixed input, and combining the batch data distribution drift and parameter sensitivity, perform neighborhood adaptive search on the set of dynamically sensitive hyperparameters to obtain the optimal solution of dynamically sensitive hyperparameters. Step S43: Merge the optimal solutions of statically sensitive hyperparameters and dynamically sensitive hyperparameters to obtain the final hyperparameter set, and use the final hyperparameter set to train the prediction model to obtain the optimal prediction model.

[0011] Preferably, in step S41, the global search of the statically sensitive hyperparameter set using the dynamic inertia-weighted particle swarm optimization algorithm specifically includes: Initialize the particle positions and velocities of the particle swarm in a statically sensitive hyperparameter set; In each iteration, the dynamic inertia weight is gradually decayed and updated. Based on the updated dynamic inertia weight, the particle velocity and particle position are updated. The static sensitivity hyperparameters corresponding to each updated particle are cross-validated and evaluated on a continuous batch window set to obtain the performance evaluation index value of the updated particle in each batch window. The particle fitness is calculated, and the individual optimal particle and the global optimal particle are updated according to the particle fitness. In continuous iterations, if the improvement in the particle fitness of the globally optimal particle does not exceed the early termination threshold, the iteration is stopped. Output the optimal solution of static sensitive hyperparameters corresponding to the globally optimal particle, and write the optimal solution of static sensitive hyperparameters into a modular data analysis object.

[0012] Preferably, in step S42, the neighborhood adaptive search of the dynamically sensitive hyperparameter set specifically includes: Historical optimization trajectories are read from modular data analysis objects to construct an autocovariance matrix, and the mean of historical optimization trajectories is calculated as the benchmark center of the dynamic sensitive hyperparameter set; A Gaussian distribution model is constructed by multiplying the scaling factor by the autocovariance matrix, where the scaling factor is calculated as follows: ; in, Scaling factor For expansion coefficient, This is the drift adjustment coefficient. This refers to the drift degree of the batch data distribution. During the iteration process, Monte Carlo sampling is performed on the Gaussian distribution model to generate a candidate parameter set; For each candidate parameter in the candidate parameter set, cross-validation is performed in a continuous batch window set to obtain the performance evaluation index value corresponding to the candidate parameter. The surrogate model is updated according to the candidate parameter and the corresponding performance evaluation index value. The expected improvement criterion value of the candidate parameter set is calculated based on the surrogate model. The candidate parameter with the largest expected improvement criterion value is taken as the candidate optimal solution of the dynamic sensitive hyperparameter in the current iteration until the preset iteration stopping condition is met. The optimal solution of the dynamic sensitive hyperparameter is then output and written into the modular data analysis object.

[0013] Preferably, step S5 specifically includes: Select the employee time window feature matrix corresponding to the target batch window as the inference input, load the optimal prediction model and perform prediction calculation on the inference input to obtain the prediction output corresponding to each sample primary key, associate the prediction output with the organization status information table according to the sample primary key, and generate the prediction result table. The prediction result table is grouped and aggregated according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table. The prediction result table and the organizational summary result table are then written into a modular data analysis object.

[0014] A human resources data intelligent analysis system includes a data acquisition module, a hierarchical module, a difference calculation module, an optimization module, and a prediction output module; The data acquisition module is used to acquire human resources data, construct modular data analysis objects, standardize the modular data analysis objects, and generate a standardized feature vector space. The modular data analysis objects include an employee time window feature matrix and an organizational status information table; The hierarchical module is used to calculate the parameter sensitivity of hyperparameters using sliding window sampling, thereby dividing the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters. The difference calculation module is used to calculate the probability density distribution difference and topological distance difference between adjacent batch windows to obtain the batch data distribution drift degree. The optimization module is used to optimize the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set respectively using the dynamic inertial weighted particle swarm algorithm and the neighborhood adaptive search algorithm, and train to obtain the optimal prediction model. The prediction output module is used to predict the employee time window feature matrix using the optimal prediction model, generate a prediction result table, and perform grouping and aggregation operations on the prediction results according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table.

[0015] The beneficial effects of this invention are as follows: This invention accurately divides the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters by using sliding window sampling and variance contribution calculation. It also combines batch data distribution drift and uses dynamic inertial weighted particle swarm optimization algorithm and neighborhood adaptive search algorithm for hierarchical collaborative optimization. The hyperparameter hierarchical optimization strategy significantly reduces the computational complexity of model parameter tuning, ensuring the global optimality of static hyperparameters while enabling real-time agile responses to data changes through neighborhood adaptive search that incorporates drift degree. By calculating the probability density distribution differences of organizational nodes in adjacent batch windows and the topological distance differences under the organizational classification topology tree, a difference weight is introduced to calculate the batch data distribution drift degree, deeply exploring the impact of dynamic evolution of organizational structure on human resource data distribution, overcoming the difficulty of traditional methods in capturing data drift caused by organizational topology changes. By constructing a modular data analysis object encapsulated by employee time window feature matrix and organizational status information table, the decoupling and efficient association of features and organizational hierarchy encoding are achieved, supporting precise grouping and aggregation operations by organizational hierarchy to generate organizational summary result tables, thus improving the flexibility, adaptability, and business application value of intelligent human resource data analysis under complex enterprise organizational structures. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of an intelligent human resources data analysis method according to an embodiment of the present invention; Figure 2 This is a basic flowchart of a human resources data intelligent analysis system provided in one embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1, referring to Figure 1This paper provides a method for intelligent analysis of human resource data, which includes the following steps: Step S1: Obtain human resources data and construct a modular data analysis object. Standardize the modular data analysis object to generate a standardized feature vector space. Modular data analysis objects include employee time window feature matrices and organizational status information tables; Step S2: Calculate the parameter sensitivity of hyperparameters using sliding window sampling, and divide the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters. Step S3: Calculate the probability density distribution difference and topological distance difference between adjacent batch windows to obtain the batch data distribution drift. Step S4: Based on the batch data distribution drift, the dynamic inertial weighted particle swarm algorithm and the neighborhood adaptive search algorithm are used to optimize the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set respectively, and the optimal prediction model is trained. Step S5: Use the optimal prediction model to predict the employee time window feature matrix, generate a prediction result table, and perform grouping and aggregation operations on the prediction results according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table.

[0019] In a specific embodiment, step S1 specifically includes: Step S11: Obtain human resources data, which includes employee master data, organizational structure data, attendance data, salary data, and performance data. Human resources data is batch-organized according to time batches, and the sample primary key is determined by the employee identifier and time window identifier in the employee master data. The human resources data is aggregated and calculated according to the sample primary key to generate the employee time window feature matrix. Extract the horizontal departmental affiliation relationships and vertical management hierarchy chains from the organizational structure data, construct an organizational classification topology tree consisting of the group, business units, first-level departments, and second-level departments, and generate an organizational status information table based on the organizational classification topology tree. The organizational status information table includes organizational level code, on-the-job status field, organizational structure version identifier, and time batch identifier. Specifically, the processing logic for acquiring human resources data includes reading employee master data, organizational structure data, attendance data, salary data, and performance data from the enterprise's human resources-related information system. Fields used to identify the same employee in the above data are uniformly mapped to employee identifiers. Time fields in each data source are uniformly mapped to the original time fields required for time batch identifiers. When multiple records of the same employee identifier exist in the same time batch, the attendance data, salary data, and performance data are summarized to obtain human resources data.

[0020] The specific processing logic for batching human resources data based on time batches includes mapping the original time fields in each data source to time batch identifiers according to preset time batching rules; determining the sample primary key through the employee identifier and time window identifier in the employee master data, where the time window identifier is obtained by combining continuous time batch identifiers; calculating the mean of attendance data, salary data, and performance data according to the sample primary key to generate feature value records that correspond one-to-one with the sample primary key; and arranging these feature value records according to the sample primary key to form the row data source of the employee time window feature matrix.

[0021] The preset time batching rules determine the granularity and boundaries of time batches, and merge the original time fields according to the granularity and boundaries to generate time batch identifiers. The preferred granularity of time batches is quarterly, and the boundaries of time batches at the quarterly granularity are preferably based on the start and end dates of the natural quarter. For original time fields in attendance data that are in daily units, they are mapped to the corresponding time batch identifiers according to their respective natural quarters. For original time fields in payroll and performance data that are in periodic units, they are mapped to the corresponding time batch identifiers according to the start and end dates covered by their period. When the original time field of a record spans multiple time batches, the record is assigned to the time batch identifier of the end date.

[0022] The specific processing logic for aggregating human resources data by sample primary key includes generating a corresponding sample primary key for each record in each data source and deleting records with missing sample primary keys; grouping records by sample primary key; calculating attendance-related aggregate values ​​by summing within each sample primary key group for attendance data; calculating salary-related aggregate values ​​by summing within each sample primary key group for salary data; and calculating performance-related aggregate values ​​by summing within each sample primary key group for performance data.

[0023] The horizontal departmental affiliation relationship includes the affiliation mapping relationship between employee identifiers and hierarchical nodes in the group, business unit, first-level department, and second-level department. When the organizational structure data records multiple departmental affiliations corresponding to an employee identifier in the same time batch, the departmental affiliation record after the last change in that time batch is preferred to determine the unique departmental affiliation of that employee identifier in that time batch.

[0024] The vertical management hierarchy chain includes the parent-child relationship between the group, business units, first-level departments, and second-level departments, as well as the identifier of the superior node used to represent the connection between departmental nodes. When there are multiple vertical management hierarchy chains in the organizational structure data, the system reads whether there is an identifier field in the organizational structure data to indicate the main chain. If it exists, the corresponding vertical management hierarchy chain is selected as the main chain according to the identifier field. If there is no identifier field in the organizational structure data that can directly indicate the main chain, the multiple vertical management hierarchy chains are sorted according to a preset sorting criterion in the same batch at the same time, and the vertical management hierarchy chain with the first sort is selected as the main chain. The remaining chains are retained as auxiliary information in the original record of the organizational structure data for traceability.

[0025] The preset sorting criteria include prioritizing the vertical management hierarchy chain corresponding to the parent node identifier after the last change within the time batch. After determining the main chain, the remaining vertical management hierarchy chains are retained as auxiliary information in the original organizational structure data for traceability, but they do not participate in the generation of node connection relationships in the organizational classification topology tree.

[0026] The specific processing logic for generating an organizational status information table based on the organizational classification topology tree includes: constructing a set of organizational classification topology tree nodes and node connections based on the vertical management hierarchy chain, consisting of the group, business units, first-level departments, and second-level departments; mapping employee identifiers to department nodes in the organizational classification topology tree based on the horizontal departmental affiliation relationship, and generating organizational hierarchy codes on this basis, preferably hierarchical path codes from the group to the second-level departments; extracting the on-duty status field from the employee master data and writing it into the organizational status information table according to the sample primary key; within each time batch, extracting the set of organizational nodes and node connections covered by that time batch from the organizational structure data, and extracting the affiliation mapping relationship from employee identifiers to organizational nodes within that time batch; if the organizational structure data provides an organizational structure version... If the current number, publication timestamp, and publication sequence number can be directly used as version identifiers, then these fields will be written into the organization status information table as organization structure version identifiers. If no directly usable version identifier field is provided in the organization structure data, then an organization structure snapshot for each time batch will be constructed based on the set of organization nodes, the set of node connection relationships, and the attribution mapping relationship. The key fields of the organization structure snapshot will be concatenated in a fixed order to generate a unique identifier. When there are multiple organization structure change records in the same time batch, the organization structure version identifier will be generated using the organization structure snapshot after the last change in that time batch. This organization structure version identifier and the time batch identifier will be written into the organization status information table according to the sample primary key, resulting in an organization status information table that is consistent with the employee time window feature matrix according to the sample primary key.

[0027] Obtain the department affiliation relationship from the employee master data, map the organization status information table through organizational hierarchy coding and concatenate it into the employee time window feature matrix, so that the organization status data is transformed into employee-level features, thereby aligning them under the same sample primary key.

[0028] The employee time window feature matrix and the organizational status information table are encapsulated into a modular data analysis object according to the same sample primary key; Step S12: Standardize the modular data analysis object to obtain a standardized feature vector space; Standardization processes include sample alignment, feature alignment, and feature standardization.

[0029] Specifically, the human resources business cycle identifier corresponding to each time batch is recorded in the organization status information table, and the organizational structure version identifier and policy interpretation version identifier corresponding to the time batch are also recorded. Perform sample alignment operation on the employee time window feature matrix and the organization status information table. The sample alignment operation includes taking the intersection of the sample primary keys of the employee time window feature matrix and the organization status information table as the aligned sample set, and deleting sample records that do not belong to the aligned sample set. The feature alignment operation is performed on the employee time window feature matrix. The feature alignment operation includes deleting feature variables that are always zero in the aligned sample set, deleting feature variables with zero variance in the aligned sample set, and deleting feature variables with missing proportions exceeding the preset rules, and forming the aligned employee time window feature matrix. The observation intensity normalization operation is performed on the aligned employee time window feature matrix. The observation intensity normalization operation includes normalizing the event count features according to a unified observable duration caliber, which includes the number of days on duty, the number of effective working days, the scheduled working hours, and the contractually agreed working hours. The aligned employee time window feature matrix is ​​subjected to feature standardization to obtain the standardization result. The feature standardization operation includes robust standardization of numerical features and centered log-ratio transformation (CLR) for proportional features. The standardized results are written into a modular data analysis object to form a standardized feature vector space.

[0030] In a specific embodiment, calculating parameter sensitivity includes: Construct a continuous batch window set based on a preset sliding window sampling method; Each batch window in a consecutive batch window set contains multiple consecutive time batches; The preset sliding window is used to construct a continuous batch of windows, and the preferred setting range is: When the time batch is monthly, the sliding window length is preferably 3 to 6 consecutive time batches, and the set of consecutive batch windows preferably covers the window sequence formed by the most recent 6 to 12 time batches. When the time batch is quarterly, the sliding window length is preferably 2 to 4 consecutive time batches, and the set of consecutive batch windows preferably covers the window sequence formed by the most recent 4 to 8 time batches. This method requires cross-validation of the prediction model using the hyperparameter combination space to be evaluated in each batch window of the consecutive batch window set, and generates performance evaluation index records that can be used to calculate variance contribution. Therefore, the sliding window length and the coverage of the consecutive batch window set must simultaneously satisfy the constraints of the number of samples available for cross-validation within the batch window and the number of adjacent windows available for drift measurement between batch windows.

[0031] Based on the joint sampling strategy, cooperative perturbation sampling is performed within the range of values ​​of the hyperparameters to be evaluated to construct the combination space of the hyperparameters to be evaluated; In each batch window of the continuous batch window set, the prediction model is cross-validated using the hyperparameter combination space to be evaluated, and the performance evaluation index value of each hyperparameter combination to be evaluated is obtained in each batch window. The variance contribution of each hyperparameter to be evaluated is calculated based on the performance evaluation index value of each combination of hyperparameters to be evaluated under each batch window, and the average value of the variance contribution under all batch windows is used as the parameter sensitivity of the corresponding hyperparameter to be evaluated. If the parameter sensitivity of the hyperparameter to be evaluated does not exceed the sensitivity threshold, it is added to the static sensitive hyperparameter set; if the parameter sensitivity of the hyperparameter to be evaluated exceeds the sensitivity threshold, it is added to the dynamic sensitive hyperparameter set. The set of statically sensitive hyperparameters, the set of dynamically sensitive hyperparameters, and the parameter sensitivity of each hyperparameter to be evaluated are written into a modular data analysis object.

[0032] Specifically, based on the joint sampling strategy, collaborative perturbation sampling is performed within the value range of the hyperparameters to be evaluated. The specific processing logic for constructing the hyperparameter combination space to be evaluated includes determining the set of hyperparameters to be evaluated for the prediction model and setting a value range for each hyperparameter; simultaneously generating values ​​for multiple hyperparameters to be evaluated within the value range based on the joint sampling strategy to obtain a hyperparameter combination to be evaluated, and repeating the above synchronous generation process to obtain multiple hyperparameter combinations to be evaluated, forming the hyperparameter combination space to be evaluated; during the generation process, validity checks and deduplication are performed on the hyperparameter combinations to be evaluated to ensure that each hyperparameter combination to be evaluated in the hyperparameter combination space can be used for subsequent cross-validation evaluation and that there are no duplicate combination records.

[0033] When the prediction model is a gradient boosting tree model, the range of values ​​for the hyperparameters to be evaluated includes: learning rate hyperparameters preferably set to a continuous range from small to medium; tree depth hyperparameters preferably covering a discrete range from shallow to moderately deep; sampling ratio hyperparameters preferably covering a continuous range from medium to full sampling; regularization hyperparameters preferably covering a continuous range from no regularization to moderately strong regularization; iteration rounds hyperparameters preferably covering a discrete range from few to medium iterations; and class imbalance adjustment hyperparameters preferably covering a continuous range from no adjustment to moderate adjustment. The learning rate is set to 0.01 to 0.3, the tree depth to 3 to 12, the sampling ratio to 0.5 to 1.0, the regularization to 0 to 10 or 0 to 50, the iteration rounds to 50 to 800, and the class imbalance adjustment to 1 to 20. The specific processing logic for cross-validating the prediction model using the hyperparameter combination space to be evaluated includes: extracting the sample set of employee time window feature matrices corresponding to the batch window, and reading the corresponding label information according to the sample primary key; for each hyperparameter combination to be evaluated in the hyperparameter combination space, loading the hyperparameter combination to be evaluated as the hyperparameter configuration of the prediction model; within each batch window, using the sample set of employee time window feature matrices corresponding to the batch window as input, dividing the sample set into several non-overlapping subsets, and selecting one subset as the validation subset in each round of evaluation, while the remaining subsets are used as training subsets; to avoid information leakage caused by the same employee appearing in both the training and validation subsets in the same batch window, it is preferable to group the samples according to the employee identifier contained in the sample primary key, that is, selecting employees with the same employee identifier... The identified sample records are all assigned to the same subset. To ensure the comparability of label information in different subsets, the proportional structure of label information is kept consistent during grouping. A prediction model is trained on the training subset, and performance evaluation index values ​​are calculated on the validation subset. For the same hyperparameter combination to be evaluated, performance evaluation index values ​​are calculated for each round of cross-validation within the same batch window. The arithmetic mean of the performance evaluation index values ​​from each round is calculated and synthesized into the performance evaluation index value for the hyperparameter combination to be evaluated within that batch window. When the number of samples in each round of validation subsets is inconsistent, a weighted average of the performance evaluation index values ​​is performed based on the number of samples in each round of validation subsets. The performance evaluation index values ​​of each hyperparameter combination to be evaluated within each batch window are stored in association with the combination index and the batch window index for subsequent calculation of the variance contribution and parameter sensitivity of the hyperparameters to be evaluated. The weight of each round of the weighted average is equal to the number of samples in that round of validation subsets divided by the total number of samples in all validation subsets.

[0034] In this method, the prediction model is preferably a gradient boosting tree model, and the hyperparameters of the gradient boosting tree model are used as the hyperparameters of the prediction model for parameter sensitivity calculation, partitioning of the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set, and subsequent optimization by dynamic inertial weighted particle swarm optimization algorithm and neighborhood adaptive search algorithm; in other embodiments, the prediction model can be a random forest model, a logistic regression model or a support vector machine model, and the hyperparameters of the selected model are used as the hyperparameters of the prediction model in the processing flow of steps S2 to S4.

[0035] The sensitivity threshold is used to divide the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters. The preferred setting range is 0.03 to 0.08. The parameter sensitivity of this method is composed of the average value of the variance contribution of the hyperparameter to be evaluated under all batch windows. The variance contribution is calculated based on the cross-validation evaluation results in the continuous batch window set. Therefore, the setting of the sensitivity threshold needs to be matched with the size of the hyperparameter combination space to be evaluated, the number of windows in the continuous batch window set, and the statistical fluctuation of the performance evaluation index value, so that the optimization process can be performed on the set of statically sensitive hyperparameters and the set of dynamically sensitive hyperparameters respectively in step S4.

[0036] In a specific embodiment, the expression for calculating the variance contribution of the hyperparameter to be evaluated is as follows: ; in, For the first Within each batch window, the hyperparameters to be evaluated Parameter sensitivity, Index of hyperparameters to be evaluated. In the first Within each batch window, the random variable consists of the performance evaluation index values ​​of each hyperparameter combination to be evaluated under each batch window. For the first The representation of each hyperparameter to be evaluated as a random variable in the hyperparameter combination space. To remove Other combinations of hyperparameters, For the first Within each batch window, the performance evaluation index value Total variance For batch window index, For conditional variance, This is the expected operation.

[0037] Wherein, the conditional variance is the sum of the other hyperparameter combinations. When fixed, only the hyperparameters to be evaluated are considered. Caused by change Fluctuation magnitude, here the subscript Indicates to The variance of the changes is calculated; the expectation operation represents the variance of the remaining hyperparameter combinations. The average is calculated by taking different values ​​of the hyperparameter combination in the sample space. That is, averaging the values ​​of different hyperparameter combinations. Time The resulting conditional variance is then averaged across the population.

[0038] It should be noted that in current technologies, the space of hyperparameter combinations to be evaluated, representing the hyperparameter configuration, is usually obtained through full global search and manual experience. This invention creatively integrates sliding window sampling and joint sampling strategies into the calculation of the variance contribution of hyperparameters, obtaining corresponding static and dynamic sensitive hyperparameter sets based on whether the parameter sensitivity exceeds the sensitivity threshold. The data in these static and dynamic sensitive hyperparameter sets can simultaneously characterize the robustness of the hyperparameters' impact on the overall model performance and the sensitivity to temporal data window fluctuations. This solves the problems of excessive computational complexity and inability to perceive the temporal heterogeneity of parameters in traditional hyperparameter optimization. It lays a good foundation for the decoupling of static and dynamic parameters in the following text, which inputs the input into the dynamic inertial weighted particle swarm algorithm and the neighborhood adaptive search algorithm for hierarchical collaborative optimization. Thus, through the hierarchical decoupling parameter management method, an innovative effect that balances the global stability and local adaptability of the model is achieved.

[0039] In a specific embodiment, step S3 specifically includes: Step S31: In the standardized feature vector space, for any two adjacent batch windows, extract the corresponding first batch window sample set and second batch window sample set. Step S32: Divide the first batch of window sample sets and the second batch of window sample sets into corresponding organizational node sample subsets, calculate the difference in probability density distribution of organizational nodes in the first batch of window sample sets and the difference in topological distance under the organizational classification topology tree, and use the difference in topological distance as the difference weight to perform a weighted average of the difference in probability density distribution to obtain the batch data distribution drift. The batch data distribution drift is associated with the identifiers of two adjacent batch windows and stored to form a batch drift record table, which is then written into a modular data analysis object.

[0040] Specifically, any two adjacent batch windows are selected from the consecutive batch window set in window order as the first batch window and the second batch window. Based on the time batch identifier sequence contained in each of the two batch windows, sample records whose sample primary key time window identifiers fall into the corresponding time batch identifier sequence are selected from the employee time window feature matrix in the modular data analysis object. This yields the first batch window sample set and the second batch window sample set. Simultaneously, the organizational level code corresponding to the sample record is read from the organizational status information table according to the sample primary key, ensuring that the sample sets of the two batch windows can be further divided and calculated under the organizational classification topology tree.

[0041] The time batch identifier sequence refers to an ordered list of multiple time batch identifiers included in a batch window, arranged in chronological order. It is determined by the window identifier of the batch window and can be specifically represented as the arrangement of consecutive time batch identifiers covered by the batch window from earliest to latest. For example, when time batches are organized by month, a batch window can correspond to a time batch identifier sequence such as 2024-01, 2024-02, 2024-03. When filtering sample records from the employee time window feature matrix, the filtering condition is whether the time window identifier in the sample primary key falls into the time batch identifier sequence.

[0042] The first and second batches of window sample sets are divided into corresponding organizational node sample subsets. The specific processing logic includes reading the organizational level code corresponding to the primary key of each sample record in the first batch of window sample sets, locating the organizational node corresponding to the organizational level code in the organizational classification topology tree, and assigning the sample record to the organizational node sample subset corresponding to that organizational node. The same location and assignment operation is repeated for the second batch of window sample sets. When an organizational node does not have a sample record in either the first or second batch of window sample sets, it is preferable to create an empty organizational node sample subset for that organizational node to maintain the alignment of the two batch windows in the organizational node dimension, so that the subsequent calculation of probability density distribution differences and topological distance differences can be performed on the same organizational classification topology tree node set.

[0043] The specific processing logic for calculating the probability density distribution difference of an organizational node between the first batch window sample set and the second batch window sample set includes: for a given organizational node, extracting the employee time window feature matrix and feature vector sets from the organizational node sample subset corresponding to the first batch window and the organizational node sample subset corresponding to the second batch window, respectively; for each feature dimension, first determining the binning boundary, preferably generated according to the quantile of the feature dimension in the merged samples of the first and second batch window sample sets; counting the number of samples of the feature dimension falling into each bin in the organizational node sample subset, and normalizing the number of samples in each bin according to the total number of samples in the organizational node sample subset to obtain the proportion of each bin, forming the probability density distribution description of the feature dimension of the organizational node under the batch window; repeating the above processing for each feature dimension of the feature vector set to obtain the probability density distribution description set of the organizational node under the batch window; calculating the difference between the two based on the binning statistics as the probability density distribution difference of the organizational node; when there are multi-dimensional features, calculating the average value of the differences of each dimension and summing them to obtain the probability density distribution difference value at the organizational node level.

[0044] The specific processing logic for calculating the topological distance difference between organizational nodes in the first batch of window sample sets and the second batch of window sample sets under the organizational classification topology tree includes: for any two organizational nodes, the number of parent-child connections required to move from one organizational node to another along the node connection relationship of the organizational classification topology tree is the hierarchical path distance between the two organizational nodes; if the two organizational nodes are the same organizational node, the hierarchical path distance is zero; the sample proportion of each organizational node sample subset is counted in the first batch of window sample sets and the second batch of window sample sets respectively, forming the organizational node proportion distribution of the first batch of windows and the organizational node proportion distribution of the second batch of windows; then, based on the organizational node proportion distribution, the organizational node proportion from the first batch of windows to the second batch of windows is calculated. The migration amount is calculated and combined with the hierarchical path distance between organizational nodes. The sample proportion of each organizational node's sample subset is calculated in the first and second batch windows respectively, forming the organizational node proportion distribution of the two batch windows. Then, the change in the sample proportion of each organizational node between the two batch windows is calculated, and organizational nodes with increased sample proportion are matched with organizational nodes with decreased sample proportion. The migration amount of organizational nodes from decreasing nodes to increasing nodes is determined according to the matching results. Finally, for each pair of matched migration amounts, the hierarchical path distance between the two organizational nodes corresponding to the migration amount is taken as the cost and accumulated to obtain the topological distance difference. The topological distance difference is used to represent the degree of change in the organizational node distribution of the two batch windows under the constraints of the organizational classification topology tree structure.

[0045] The probability density distribution difference is weighted by using the topological distance difference as the difference weight, and the total topological distance is obtained by summing the topological distance differences of all organizational nodes. The proportion of the topological distance difference of each organizational node to the total topological distance is calculated as the weight of that node. Then, this weight is multiplied by the corresponding probability density distribution difference. Finally, the weighted results of all organizational nodes are summed to obtain the batch data distribution drift.

[0046] It should be noted that in current technologies, batch data distribution drift, representing the differences between batches of data, is usually calculated using purely statistical distribution indicators. This invention creatively integrates the topological distance differences under the organizational classification topology tree into the probability density distribution differences of the organizational node sample subsets. This yields a corresponding batch data distribution drift that can be weighted and calibrated according to the actual adjustment range of the enterprise's organizational structure. The data in this batch data distribution drift can simultaneously characterize the statistical heterogeneity of the bottom-level employee characteristic data and the degree of drastic change in the topology of the upper-level enterprise organizational structure. This solves the problems of lacking business semantic context and failing to accurately map the negative impact of organizational changes on data flow in traditional data drift detection. It lays a highly robust business quantification foundation for the neighborhood adaptive search model input to the dynamically sensitive hyperparameters as the core basis for iterative step size adjustment, thereby obtaining the innovative effect of deeply perceiving the laws of organizational evolution and achieving a deep correlation between data and business through topological weighted fusion.

[0047] In a specific embodiment, step S4 specifically includes: Step S41: Using the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set as optimization objects, perform a global search of the static sensitive hyperparameter set using the dynamic inertia weighted particle swarm algorithm to obtain the optimal solution for the static sensitive hyperparameters. Step S42: Using the optimal solution of statically sensitive hyperparameters as a fixed input, and combining the batch data distribution drift and parameter sensitivity, perform neighborhood adaptive search on the set of dynamically sensitive hyperparameters to obtain the optimal solution of dynamically sensitive hyperparameters. Step S43: Merge the optimal solutions of statically sensitive hyperparameters and dynamically sensitive hyperparameters to obtain the final hyperparameter set, and use the final hyperparameter set to train the prediction model to obtain the optimal prediction model.

[0048] Specifically, the processing logic for training the prediction model using the final hyperparameter set to obtain the optimal prediction model includes: merging the optimal solutions of statically sensitive hyperparameters and dynamically sensitive hyperparameters to form the final hyperparameter set, and loading the final hyperparameter set as the hyperparameter configuration of the prediction model; selecting sample records corresponding to the primary keys of samples belonging to the continuous batch window set from the employee time window feature matrix as candidate training sample sets; aligning the candidate training sample set with label information according to the sample primary keys, and deleting sample records with missing label information; performing sample validity filtering on the candidate training sample set according to the on-duty status field in the organization status information table, and deleting sample records that do not meet the on-duty status consistency rule; performing feature alignment and standardization processing on the candidate training sample set in the same way as step S1 to ensure that the feature dimension and feature order are consistent; and combining the processed feature vectors with their corresponding label information to form the employee time window feature matrix sample set and its corresponding label information used for training. Within the range of consecutive batch windows, samples are sorted according to time batch identifiers. Sample records corresponding to earlier time batch identifiers are assigned to the training subset, and sample records corresponding to later time batch identifiers are assigned to the validation subset. When the number of samples under the same time batch identifier is large, random sampling based on the sample primary key is allowed within that time batch identifier, but sample records with the same employee identifier must not appear in both the training and validation subsets simultaneously. During training, the training state information of each iteration on the training subset is recorded, and performance evaluation index values ​​are periodically calculated on the validation subset as training process records. When the performance evaluation index values ​​recorded for several consecutive times fail to meet the preset improvement range requirement, training is stopped; or when the number of training iterations reaches the preset maximum number of training iterations, training is stopped. After training stops, the corresponding prediction model state and the final hyperparameter set are written together as the optimal prediction model into the modular data analysis object. In this method, the preset improvement range requirement is preferably set to an improvement in the performance evaluation index value between a number of consecutive records that is not less than a small threshold, preferably set to 0.001–0.005. When computational resources are limited and sample fluctuations are large within the batch window, the small threshold is preferably set to 0.001. When the sample size is large and the performance evaluation index value fluctuates little, the small threshold is preferably set to 0.003–0.005. The number of consecutive records is 10–20, used to determine if the performance evaluation index value of a number of consecutive records has not met the preset improvement range requirement. The preset maximum number of training iterations is preferably set to 200–1000.

[0049] In a specific embodiment, step S41, which involves performing a global search of the statically sensitive hyperparameter set using the dynamic inertia-weighted particle swarm optimization algorithm, specifically includes: Initialize the particle positions and velocities of the particle swarm in a statically sensitive hyperparameter set; In each iteration, the dynamic inertia weight is gradually decayed and updated. Based on the updated dynamic inertia weight, the particle velocity and particle position are updated. The static sensitive hyperparameter combination corresponding to each updated particle is cross-validated and evaluated on a continuous batch window set to obtain the performance evaluation index value of the updated particle in each batch window. The particle fitness is calculated, and the individual optimal particle and the global optimal particle are updated according to the particle fitness. In continuous iterations, if the improvement in the particle fitness of the globally optimal particle does not exceed the early termination threshold, the iteration is stopped. Output the optimal solution of static sensitive hyperparameters corresponding to the globally optimal particle, and write the optimal solution of static sensitive hyperparameters into a modular data analysis object.

[0050] Specifically, the initialization logic for particle positions and velocities includes reading the value range and value type constraints of each hyperparameter in the statically sensitive hyperparameter set, mapping each hyperparameter in the statically sensitive hyperparameter set to a dimension of the particle position, ensuring the number of particle position dimensions matches the number of statically sensitive hyperparameters; when the hyperparameter corresponding to this dimension has a continuous value constraint, generating a real number within the value range and performing boundary constraint processing after updating; when the hyperparameter corresponding to this dimension has an integer value constraint, generating a real number within the value range, converting it to an integer according to a preset rounding rule, and performing boundary constraint processing; when the hyperparameter corresponding to this dimension has a discrete value constraint, selecting a candidate value from the discrete candidate value set according to a preset sampling method, and performing set member constraint processing after updating; when joint constraints exist, performing joint constraint processing after generating particle positions. Constraint validity checks and necessary corrections are performed. Within the range of values, static sensitive hyperparameter values ​​are generated dimension-by-dimensionally for each particle to form the initial particle position. Discrete and integer hyperparameter values ​​are generated by rounding, and components exceeding the range are subject to boundary constraints. Then, an initial particle velocity with the same dimension as the particle position is generated for each particle. The initial particle velocity is preferably randomly generated within a preset velocity range, and components that do not meet the preset velocity constraints are subject to boundary constraints. After initializing the particle position and particle velocity, the initial particle position of each particle is used as the initial record of the individual optimal particle for the corresponding particle. The particle position and particle velocity are written into the historical optimization trajectory record of the modular data analysis object for subsequent iterations when updating particle velocity, updating particle position, and updating individual optimal particles and global optimal particles.

[0051] The value range and value type constraints for each hyperparameter include: max_depth, which is an integer value constraint, preferably ranging from 3 to 12; min_child_weight, which is a continuous value constraint, preferably ranging from 1 to 10; subsample, which is a continuous value constraint, preferably ranging from 0.5 to 1.0; colsample_bytree, which is a continuous value constraint, preferably ranging from 0.5 to 1.0; and gamma, which is a continuous value constraint, preferably ranging from 0 to 10.

[0052] The preset speed range is for hyperparameters constrained by integer values, and the preferred preset speed range for the corresponding dimension is [-2, 2]. For hyperparameters with continuous value constraints and a value range of [0, 1], the preferred preset speed range for the corresponding dimension is [-0.1, 0.1]. For hyperparameters with continuous value constraints and a relatively wide value range, the preferred preset speed range for the corresponding dimension is ±20% of its value range span. The preset velocity constraint is that if the particle velocity in any dimension exceeds the preset velocity range for that dimension, the particle velocity in that dimension is truncated to the boundary value of the preset velocity range. After the particle position is updated, the corresponding dimension is rounded to the nearest integer, and further boundary constraints are applied according to the value range [3,12]. If a boundary constraint occurs, it is preferable to set the particle velocity in that dimension to 0. After the particle position is updated, if the corresponding dimension exceeds the boundary of the value range, it is preferable to truncate the particle position in that dimension to the boundary value and set the particle velocity in that dimension to 0. If the particle velocity or particle position has a non-numerical or infinite value, it is preferable to reset the particle velocity in the corresponding dimension to 0 and reset the particle position in the corresponding dimension to the median value of the value range for that dimension.

[0053] The calculation expression for the gradual decay update of the dynamic inertia weight is as follows: ; in, This represents the current iteration number. The maximum number of iterations, and These are the initial inertia weights and the final inertia weights, respectively. This represents the iteration progress ratio, used to normalize the iteration process to a range of 0 to 1, thereby controlling the decay rate of the dynamic inertia weight. This represents the total decay of the dynamic inertia weight from initial to final.

[0054] The overall update logic follows Increase Increase, therefore from The amount subtracted gradually increases, making According to the iteration progress Gradual decay approaching .

[0055] The specific processing logic for updating particle velocity and position based on the updated dynamic inertia weights includes: representing the combination of static sensitive hyperparameters corresponding to each particle as the particle position, and representing the particle velocity as the change in particle position between adjacent iterations; in each iteration, reading the dynamic inertia weight of the current iteration, and using the dynamic inertia weight as the inertia term weight in the particle velocity update process; calculating the new velocity of the particle based on the current particle velocity, the individual best particle corresponding to the particle's historical best position, and the global best particle corresponding to the current group best position; after obtaining the new particle velocity, synthesizing and updating the particle's new position with the particle's current position; performing boundary constraint processing on the new particle position, so that the value of each dimension of the hyperparameter in the new particle position satisfies the value range constraint of the set of static sensitive hyperparameters, and using the processed new particle position as the particle position input for the next iteration.

[0056] The specific processing logic for cross-validation evaluation of the static sensitive hyperparameter combination corresponding to each updated particle on a continuous batch window set includes resolving the updated particle position into a static sensitive hyperparameter combination, loading the static sensitive hyperparameter combination into the static sensitive hyperparameter values ​​in the prediction model hyperparameter configuration, setting the remaining prediction model hyperparameters that do not belong to the static sensitive hyperparameter set to preset fixed values, with the learning rate preferably set to 0.1, the value of the regularization-related hyperparameters for L2 regularization set to 1, the value of the value for L1 regularization set to 0, the class imbalance adjustment-related hyperparameter set to 1, the maximum number of leaves in a single tree set to the default value, and the number of training epochs set to 300. For each batch window in the continuous batch window set, the sample set of employee time window feature matrix corresponding to that batch window is extracted from the modular data analysis object, and the corresponding label information is read according to the sample primary key. Within that batch window, the sample set is divided into a training subset and a validation subset according to a preset cross-validation partitioning rule. A prediction model is trained on the training subset, and a performance evaluation index value is calculated on the validation subset. The arithmetic mean of the cross-validation evaluation results within the same batch window is calculated and summarized as the updated particle performance evaluation index value under that batch window. The above evaluation and summarization process is repeated for all batch windows in the continuous batch window set to form the updated particle performance evaluation index value set for each batch window, which is used for subsequent calculation of particle fitness.

[0057] The preset cross-validation partitioning rule is to perform K-fold cross-validation partitioning on the employee time window feature matrix sample set within each batch window, where K is preferably 5. During partitioning, the employee identifier in the sample primary key is preferably used as the grouping key, ensuring that sample records corresponding to the same employee identifier do not appear simultaneously in the training subset and validation subset. When there is class imbalance in the label information, the preset cross-validation partitioning rule preferably includes a hierarchical partitioning constraint, that is, while satisfying the employee identifier grouping constraint, the label information ratio structure of each fold training subset and validation subset is made as consistent as possible. When the number of samples in the batch window is insufficient to support K-fold partitioning, K is preferably adjusted to 3, and this partitioning method is recorded in the modular data analysis object for experiment reproduction.

[0058] The specific processing logic for calculating particle fitness includes taking the updated set of performance evaluation index values ​​for each batch window as input and outputting the particle fitness according to a preset fitness calculation rule. The preset fitness calculation rule includes a central tendency description and a fluctuation level description of the performance evaluation index values, where the central tendency description is the mean of the set of performance evaluation index values, and the fluctuation level description is the dispersion of the set of performance evaluation index values. The central tendency description is subtracted from the fluctuation level description, and a weighted sum is performed to synthesize the particle fitness. The particle fitness is then written into the iteration record corresponding to that particle for updating the individual optimal particle and the globally optimal particle, as well as for determining early termination conditions. The weighting of the summation is shown in Table 1. When the central tendency weight is 0.7 and the volatility weight is 0.3, the fitness values ​​of the three test particles are relatively evenly distributed, indicating that the algorithm is robust to different initial positions. The convergence iteration count is 35 rounds, which takes into account search efficiency. The final fitness value is 0.847, which is not the highest, but it performs best in the subsequent cross-window generalization test with the smallest performance drop.

[0059] The specific processing logic for updating the individual optimal particle and the global optimal particle based on particle fitness includes: maintaining the optimal particle fitness and corresponding particle position in the historical iterations for each particle, which is used as the individual optimal particle record; in each iteration, comparing the particle fitness of the current iteration with the optimal particle fitness in the individual optimal particle record, and updating the current particle position to the position corresponding to the individual optimal particle and simultaneously updating the optimal particle fitness of the individual optimal particle; after updating the individual optimal particles of all particles, comparing the current particle fitness of all particles with the historical optimal particle fitness of the group in the current iteration, selecting the particle position with the best particle fitness as the position corresponding to the global optimal particle, and updating the optimal particle fitness of the global optimal particle; finally, writing the updated individual optimal particle and global optimal particle into the optimized trajectory record for use in the next iteration when updating particle velocity and particle position.

[0060] In a specific embodiment, step S42, performing neighborhood adaptive search on the set of dynamically sensitive hyperparameters, specifically includes: Historical optimization trajectories are read from modular data analysis objects to construct an autocovariance matrix, and the mean of historical optimization trajectories is calculated as the benchmark center of the dynamic sensitive hyperparameter set; A Gaussian distribution model is constructed by multiplying the scaling factor by the autocovariance matrix, where the scaling factor is calculated as follows: ; in, Scaling factor For expansion coefficient, This is the drift adjustment coefficient. This refers to the drift degree of the batch data distribution. Specifically, the batch data distribution drift is normalized, and its value range is limited to [0,1].

[0061] During the iteration process, Monte Carlo sampling is performed on the Gaussian distribution model to generate a candidate parameter set; For each candidate parameter in the candidate parameter set, cross-validation is performed in a continuous batch window set to obtain the performance evaluation index value corresponding to the candidate parameter. The surrogate model is updated according to the candidate parameter and the corresponding performance evaluation index value. The expected improvement criterion value of the candidate parameter set is calculated based on the surrogate model. The candidate parameter with the largest expected improvement criterion value is taken as the candidate optimal solution of the dynamic sensitive hyperparameter in the current iteration until the preset iteration stopping condition is met. The optimal solution of the dynamic sensitive hyperparameter is then output and written into the modular data analysis object.

[0062] The specific processing logic for constructing the autocovariance matrix from historical optimization trajectories in the modular data analysis object includes: reading historical optimization trajectory records corresponding to the set of dynamically sensitive hyperparameters from the modular data analysis object; the historical optimization trajectory records include the accepted combinations of dynamically sensitive hyperparameter values ​​in multiple iterations and their corresponding performance evaluation index values; filtering the combinations of dynamically sensitive hyperparameter values ​​in the historical optimization trajectory records, deleting missing values ​​and values ​​outside the range; calculating the dispersion of each hyperparameter value and the common variation relationship between dimensions on the filtered combinations of dynamically sensitive hyperparameter values, and forming the autocovariance matrix accordingly; when the number of filtered records is insufficient to stably estimate the common variation relationship between dimensions, only a diagonal form of the autocovariance matrix is ​​constructed, and the diagonal elements are set to the preset scale corresponding to the range of each hyperparameter value or the variance estimate calculated from a small number of existing records; when the autocovariance matrix does not meet the numerical stability constraints of the subsequent Gaussian distribution model, it is preferable to perform diagonal regularization on the autocovariance matrix before using it for subsequent calculations.

[0063] When calculating the mean of historical optimized trajectories as the baseline center of the dynamic sensitive hyperparameter set, the combination of dynamic sensitive hyperparameter values ​​in the historical optimized trajectory records is read from the modular data analysis object, and the mean value of each dynamic sensitive hyperparameter is calculated to obtain the baseline center of the dynamic sensitive hyperparameter set.

[0064] If historical optimization trajectories are initially unavailable or insufficient in number, a preset benchmark value is used as the benchmark center, and this benchmark center is written into the modular data analysis object as the initial recording basis for subsequent historical optimization trajectories. The preset benchmark value preferably has a learning rate of 0.1, a value of 1 for L2 regularization, a value of 0 for L1 regularization, and a value of 1 for class imbalance adjustment hyperparameters.

[0065] The mathematical expression for the Gaussian distribution model is: ; in, Candidate parameters generated when performing Monte Carlo sampling on a set of dynamically sensitive hyperparameters. As the benchmark center for the set of dynamically sensitive hyperparameters, It is the autocovariance matrix. Scaling factor It follows a Gaussian distribution.

[0066] Before using the Gaussian distribution model, The numerical stability is checked, and if necessary, the autocovariance matrix is ​​diagonally regularized to satisfy the numerical constraints of the Monte Carlo sampling process.

[0067] The specific processing logic for generating candidate parameter sets by performing Monte Carlo sampling on the Gaussian distribution model includes: using the dimension of the dynamic sensitive hyperparameter set as the sampling dimension, and generating multiple combinations of dynamic sensitive hyperparameter values ​​from the Gaussian distribution model according to a preset number of candidates; performing value range constraint processing on each sampled dynamic sensitive hyperparameter value combination to ensure that the values ​​of each dimension satisfy the value range and value type constraints of the dynamic sensitive hyperparameter set, wherein discrete and integer values ​​of hyperparameters are rounded; then performing deduplication processing on the candidate parameter set to delete duplicate candidates and candidates that do not meet the training constraints of the prediction model; finally, writing the processed candidate parameter set into the candidate record area of ​​the modular data analysis object for subsequent cross-validation evaluation and surrogate model update calls.

[0068] The specific processing logic for cross-validation evaluation of each candidate parameter in the candidate parameter set within a continuous batch window set includes: taking the candidate parameter as a dynamic sensitive hyperparameter value and combining it with the optimal solution of the static sensitive hyperparameter to form the prediction model parameter configuration; then, for each batch window in the continuous batch window set, extracting the employee time window feature matrix sample set corresponding to that batch window from the modular data analysis object, and reading the corresponding label information according to the sample primary key; within that batch window, dividing the sample set into a training subset and a validation subset according to a preset cross-validation partitioning rule, training the prediction model on the training subset and calculating the performance evaluation index value on the validation subset, calculating the arithmetic mean of the cross-validation evaluation results and summarizing them as the performance evaluation index value of the candidate parameter under that batch window; repeating the above process for all batch windows in the continuous batch window set to obtain the performance evaluation index value set of the candidate parameter under each batch window, and associating the candidate parameter with its performance evaluation index value set and writing it into the modular data analysis object for subsequent proxy model updates and expected improvement criterion value calculation.

[0069] In step S42, the surrogate model refers to a model used to characterize the correspondence between candidate parameters and performance evaluation index values. The surrogate model is fitted with the candidate parameters as input and the performance evaluation index values ​​obtained by cross-validation evaluation of the candidate parameters on a continuous batch window set as output, and provides a predictive basis for the subsequent calculation of the expected improvement criterion value of the candidate parameter set based on the surrogate model.

[0070] When updating the surrogate model, it is preferable to summarize the candidate parameters and their corresponding performance evaluation index values ​​recorded in the modular data analysis object into a training dataset, and perform incremental updates and retraining updates on the training dataset. The incremental update includes adding the candidate parameters and their performance evaluation index values ​​newly evaluated in this round to the training dataset of the surrogate model and refitting. The retraining update includes refitting the surrogate model on all historical candidate parameter records. The specific implementation of the surrogate model in this method preferably adopts a Gaussian process regression model, and in the update process, the input candidate parameters are checked for their value range in the same way as in step S1 to ensure that the input of the surrogate model is consistent with the value constraints of the dynamic sensitive hyperparameter set.

[0071] The expression for calculating the expected improvement criterion value is: ; in, Candidate parameters The expected improvement criterion value, These are the currently recorded optimal candidate parameters. Candidate parameters The corresponding performance evaluation index values, The currently recorded optimal candidate parameters The corresponding performance evaluation index values, Candidate parameters Performance evaluation index value Less than or equal to the current optimal value If the difference is negative or zero, and the maximum value of 0 is taken, the expected improvement criterion value is 0; if Greater than If the difference is positive, the expected improvement criterion value is the expectation of that positive difference.

[0072] The preset iteration stopping condition is that if the improvement in the expected improvement criterion value of the current iteration, compared to the improvement in the expected improvement criterion value of several consecutive iterations, is less than the expected improvement threshold, then the optimization process is considered to have converged, and iteration stops; or, if the number of iterations reaches the maximum number of iterations, iteration terminates. When either of the above conditions is met, the candidate parameter with the largest expected improvement criterion value is selected as the candidate optimal solution for the dynamic sensitive hyperparameter in the current iteration, and the final output is the optimal solution for the dynamic sensitive hyperparameter. The maximum number of iterations is preferably set to 50, and the expected improvement threshold is preferably set to 1*10. -4 This means that when the expected improvement explored in several consecutive iterations has become negligible, it can be concluded that the current solution has reached a relatively stable local optimum in the context of data distribution drift.

[0073] It should be noted that traditional techniques for adaptive neighborhood search of dynamically sensitive hyperparameters typically obtain candidate parameter sets through sampling with a fixed scale and a single time-decreasing Gaussian distribution. This approach cannot cope with data fluctuations caused by sudden organizational changes and is prone to getting trapped in local optima during parameter tuning. This invention fuses the mean of historical optimization trajectories and the autocovariance matrix into a Gaussian distribution model and dynamically calculates the resulting scaling factor. Traditional scaling factors are usually updated using preset constants and empirical decay methods, leaving the technical problem of difficulty in flexibly adjusting the search range according to the actual data flow evolution. However, the nonlinear exponential update method adopted in this invention, which fuses the drift degree of batch data distribution, can adaptively amplify the search neighborhood of the Gaussian distribution as the data drift degree increases. This can eliminate frequent blind trial and error, obtain the optimal solution of dynamically sensitive hyperparameters that can quickly track new data patterns, and significantly improve the adaptive convergence speed and prediction accuracy of the model in complex dynamic environments.

[0074] In a specific embodiment, step S5 specifically includes: Select the employee time window feature matrix corresponding to the target batch window as the inference input, load the optimal prediction model and perform prediction calculation on the inference input to obtain the prediction output corresponding to each sample primary key, associate the prediction output with the organization status information table according to the sample primary key, and generate the prediction result table. The prediction result table is grouped and aggregated according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table. The prediction result table and the organizational summary result table are then written into a modular data analysis object.

[0075] Specifically, the process of loading the optimal prediction model and performing prediction calculations on the inference input includes: reading and loading the optimal prediction model and its corresponding final hyperparameter set from the modular data analysis object, and reading the standardized processing configuration consistent with that used during the training of the optimal prediction model; selecting the employee time window feature matrix corresponding to the target batch window as the inference input, and performing the same standardization and feature alignment processing as in step S12 on the inference input to ensure that the feature dimensions and feature order of the inference input are consistent with the training input of the optimal prediction model; after completing the standardization processing, sending the inference input into the optimal prediction model to perform prediction calculations, obtaining the prediction output corresponding to each sample primary key; associating the prediction output with the sample primary key to generate a prediction result table, and writing the prediction result table into the modular data analysis object.

[0076] The specific processing logic for grouping and aggregating the prediction result table based on the organizational hierarchy code in the organizational status information table includes: first, associating the prediction result table with the organizational status information table by sample primary key to obtain an association result containing sample primary key, prediction output, and organizational hierarchy code; grouping the association result using the organizational hierarchy code as the grouping key, and calculating the mean of the prediction output within each group; when the organizational summary result table needs to be output simultaneously by time batch identifier, further performing grouping and aggregation by using both the time batch identifier and the organizational hierarchy code as the grouping key; and finally, writing the organizational summary result table into a modular data analysis object.

[0077] Example 2, refer to Figure 2 This provides a human resources data intelligent analysis system, including a data acquisition module, a hierarchical module, a difference calculation module, an optimization module, and a prediction output module; The data acquisition module is used to acquire human resources data, construct modular data analysis objects, standardize the modular data analysis objects, and generate a standardized feature vector space. Modular data analysis objects include employee time window feature matrices and organizational status information tables; The hierarchical module is used to calculate the parameter sensitivity of hyperparameters using sliding window sampling, thereby dividing the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters. The difference calculation module is used to calculate the difference in probability density distribution and topological distance between adjacent batch windows to obtain the batch data distribution drift. The optimization module is used to optimize the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set respectively using the dynamic inertial weighted particle swarm algorithm and the neighborhood adaptive search algorithm, and train to obtain the optimal prediction model; The prediction output module is used to predict the employee time window feature matrix using the optimal prediction model, generate a prediction result table, and perform grouping and aggregation operations on the prediction results according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table.

[0078] This invention achieves a dual-track approach to parameters, combining static and dynamic methods, through sliding window sampling and variance contribution calculation. It employs a dynamic inertial weighted particle swarm optimization algorithm to lock in the model's static infrastructure, while leveraging neighborhood adaptive search to empower dynamic hyperparameters. This hierarchical optimization strategy significantly compresses the search space for high-dimensional parameters, ensuring both robust underlying performance and agile response to instantaneous data changes. An organizational classification topology tree is introduced. By calculating the topological distance differences between organizational nodes within the tree structure and using this as a heterogeneity weight, the differences in probability density distribution are weighted. This imbues abstract data drift with a concrete business context. Individual employee temporal characteristics are integrated and encapsulated with macro-level organizational status information, creating a standardized feature vector space and modular data analysis objects. This achieves decoupling and dynamic binding between micro-level feature data and upper-level management hierarchy coding, enabling prediction results to be accurately grouped and aggregated along the tree-like administrative structure.

[0079] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0080] 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, and all such modifications or substitutions should be covered within the protection scope of the present invention. All data acquisition actions in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located and with the authorization granted by the owner of the corresponding device.

Claims

1. A method for intelligent analysis of human resource data, characterized in that, Includes the following steps: Step S1: Obtain human resources data and construct a modular data analysis object. Standardize the modular data analysis object to generate a standardized feature vector space. The modular data analysis objects include an employee time window feature matrix and an organizational status information table; Step S2: Calculate the parameter sensitivity of hyperparameters using sliding window sampling, and divide the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters. Step S3: Calculate the probability density distribution difference and topological distance difference between adjacent batch windows to obtain the batch data distribution drift. Step S4: Based on the batch data distribution drift, the dynamic inertial weighted particle swarm algorithm and the neighborhood adaptive search algorithm are used to optimize the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set respectively, and the optimal prediction model is trained. Step S5: Use the optimal prediction model to predict the employee time window feature matrix, generate a prediction result table, and perform grouping and aggregation operations on the prediction results according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table.

2. The human resource data intelligent analysis method of claim 1, wherein, Step S1 specifically includes: Step S11: Obtain human resources data, which includes employee master data, organizational structure data, attendance data, salary data, and performance data; Human resources data is batch-organized according to time batches, and the sample primary key is determined by the employee identifier and time window identifier in the employee master data. The human resources data is aggregated and calculated according to the sample primary key to generate the employee time window feature matrix. Extract the horizontal departmental affiliation relationships and vertical management hierarchy chains from the organizational structure data, construct an organizational classification topology tree consisting of the group, business unit, first-level department and second-level department, and generate an organizational status information table based on the organizational classification topology tree. The organizational status information table includes organizational level code, on-the-job status field, organizational structure version identifier and time batch identifier. The employee time window feature matrix and the organizational status information table are encapsulated into a modular data analysis object according to the same sample primary key; Step S12: Standardize the modular data analysis object to obtain a standardized feature vector space; The standardization process includes sample alignment, feature alignment, and feature standardization.

3. The method of claim 2, wherein the human resource data is analyzed by using a machine learning algorithm. Calculate parameter sensitivity, including: Construct a continuous batch window set based on a preset sliding window sampling method; Based on the joint sampling strategy, cooperative perturbation sampling is performed within the range of values ​​of the hyperparameters to be evaluated to construct the combination space of the hyperparameters to be evaluated; In each batch window of the continuous batch window set, the prediction model is cross-validated using the hyperparameter combination space to be evaluated, and the performance evaluation index value of each hyperparameter combination to be evaluated is obtained in each batch window. The variance contribution of each hyperparameter to be evaluated is calculated based on the performance evaluation index value of each combination of hyperparameters to be evaluated under each batch window, and the average value of the variance contribution under all batch windows is used as the parameter sensitivity of the corresponding hyperparameter to be evaluated. If the parameter sensitivity of the hyperparameter to be evaluated does not exceed the sensitivity threshold, it is added to the static sensitive hyperparameter set; if the parameter sensitivity of the hyperparameter to be evaluated exceeds the sensitivity threshold, it is added to the dynamic sensitive hyperparameter set. The set of statically sensitive hyperparameters, the set of dynamically sensitive hyperparameters, and the parameter sensitivity of each hyperparameter to be evaluated are written into a modular data analysis object.

4. The method of claim 3, wherein, The expression for calculating the variance contribution of the hyperparameter to be evaluated is as follows: ； in, For the first Within each batch window, the hyperparameters to be evaluated Parameter sensitivity, Index of hyperparameters to be evaluated. In the first Within each batch window, the random variable consists of the performance evaluation index values ​​of each hyperparameter combination to be evaluated under each batch window. For the first The representation of each hyperparameter to be evaluated as a random variable in the hyperparameter combination space. To remove Other combinations of hyperparameters, For the first Within each batch window, the performance evaluation index value Total variance For batch window index, For conditional variance, This is the expected operation.

5. The method of claim 4, wherein, Step S3 specifically includes: Step S31: In the standardized feature vector space, for any two adjacent batch windows, extract the corresponding first batch window sample set and second batch window sample set. Step S32: Divide the first batch of window sample sets and the second batch of window sample sets into corresponding organizational node sample subsets, calculate the difference in probability density distribution of organizational nodes in the first batch of window sample sets and the difference in topological distance under the organizational classification topology tree, and use the difference in topological distance as the difference weight to perform a weighted average of the difference in probability density distribution to obtain the batch data distribution drift. The batch data distribution drift is associated with the identifiers of two adjacent batch windows and stored to form a batch drift record table, which is then written into a modular data analysis object.

6. The intelligent analysis method of human resource data according to claim 5, wherein, Step S4 specifically includes: Step S41: Using the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set as optimization objects, perform a global search of the static sensitive hyperparameter set using the dynamic inertia weighted particle swarm algorithm to obtain the optimal solution for the static sensitive hyperparameters. Step S42: Using the optimal solution of statically sensitive hyperparameters as a fixed input, and combining the batch data distribution drift and parameter sensitivity, perform neighborhood adaptive search on the set of dynamically sensitive hyperparameters to obtain the optimal solution of dynamically sensitive hyperparameters. Step S43: Merge the optimal solutions of statically sensitive hyperparameters and dynamically sensitive hyperparameters to obtain the final hyperparameter set, and use the final hyperparameter set to train the prediction model to obtain the optimal prediction model.

7. The method of claim 6, wherein the method further comprises: In step S41, the global search of the statically sensitive hyperparameter set using the dynamic inertia-weighted particle swarm optimization algorithm specifically includes: Initialize the particle positions and velocities of the particle swarm in a statically sensitive hyperparameter set; In each iteration, the dynamic inertia weight is gradually decayed and updated. Based on the updated dynamic inertia weight, the particle velocity and particle position are updated. The static sensitivity hyperparameters corresponding to each updated particle are cross-validated and evaluated on a continuous batch window set to obtain the performance evaluation index value of the updated particle in each batch window. The particle fitness is calculated, and the individual optimal particle and the global optimal particle are updated according to the particle fitness. In continuous iterations, if the improvement in the particle fitness of the globally optimal particle does not exceed the early termination threshold, the iteration is stopped. Output the optimal solution of static sensitive hyperparameters corresponding to the globally optimal particle, and write the optimal solution of static sensitive hyperparameters into a modular data analysis object.

8. The intelligent analysis method of human resource data according to claim 7, wherein, In step S42, the neighborhood adaptive search for the set of dynamically sensitive hyperparameters specifically includes: Historical optimization trajectories are read from modular data analysis objects to construct an autocovariance matrix, and the mean of historical optimization trajectories is calculated as the benchmark center of the dynamic sensitive hyperparameter set; A Gaussian distribution model is constructed by multiplying the scaling factor by the autocovariance matrix, where the scaling factor is calculated as follows: ； wherein, is a scaling factor, is an expansion coefficient, is a drift adjustment coefficient, is a batch data distribution drift degree; During the iteration process, Monte Carlo sampling is performed on the Gaussian distribution model to generate a candidate parameter set; For each candidate parameter in the candidate parameter set, cross-validation is performed in a continuous batch window set to obtain the performance evaluation index value corresponding to the candidate parameter. The surrogate model is updated according to the candidate parameter and the corresponding performance evaluation index value. The expected improvement criterion value of the candidate parameter set is calculated based on the surrogate model. The candidate parameter with the largest expected improvement criterion value is taken as the candidate optimal solution of the dynamic sensitive hyperparameter in the current iteration until the preset iteration stopping condition is met. The optimal solution of the dynamic sensitive hyperparameter is then output and written into the modular data analysis object.

9. The human resource data intelligent analysis method of claim 8, wherein, Step S5 specifically includes: Select the employee time window feature matrix corresponding to the target batch window as the inference input, load the optimal prediction model and perform prediction calculation on the inference input to obtain the prediction output corresponding to each sample primary key, associate the prediction output with the organization status information table according to the sample primary key, and generate the prediction result table. The prediction result table is grouped and aggregated according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table. The prediction result table and the organizational summary result table are then written into a modular data analysis object.

10. A human resource data intelligent analysis system applied to the human resource data intelligent analysis method of any one of claims 1-9, characterized in that, It includes a data acquisition module, a layering module, a difference calculation module, an optimization module, and a prediction output module; The data acquisition module is used to acquire human resources data, construct modular data analysis objects, standardize the modular data analysis objects, and generate a standardized feature vector space. The modular data analysis objects include an employee time window feature matrix and an organizational status information table; The hierarchical module is used to calculate the parameter sensitivity of hyperparameters using sliding window sampling, thereby dividing the hyperparameters of the prediction model into a set of statically sensitive hyperparameters and a set of dynamically sensitive hyperparameters. The difference calculation module is used to calculate the probability density distribution difference and topological distance difference between adjacent batch windows to obtain the batch data distribution drift degree. The optimization module is used to optimize the static sensitive hyperparameter set and the dynamic sensitive hyperparameter set respectively using the dynamic inertial weighted particle swarm algorithm and the neighborhood adaptive search algorithm, and train to obtain the optimal prediction model. The prediction output module is used to predict the employee time window feature matrix using the optimal prediction model, generate a prediction result table, and perform grouping and aggregation operations on the prediction results according to the organizational hierarchy code in the organizational status information table to generate an organizational summary result table.

Images