Talent evaluation decision support system and method based on deep learning and dynamic weight

Abstract

The invention discloses a talent evaluation decision support system and method based on deep learning and dynamic weight, and the method comprises the steps: collecting and preprocessing multi-source talent data, and obtaining a standardized talent data set; forming a service scene context information data stream; calling a post hierarchy map and organization structure information, initializing an improved hierarchy variational auto-encoder, and outputting a latent factor set; performing prior consistency calibration on the latent factor set, and performing consistency alignment on latent factor distribution of each layer by adopting layered regularization constraint to generate a calibration latent factor set; generating a dynamic weight tensor according to the attention mechanism, the time sequence preference extraction network and the enhanced feedback signal; and outputting a capability radar map, a post matching degree matrix and a capability gap report according to the competency weighted score vector. According to the invention, intelligent recommendation of talent development paths, post optimization configuration and organization strategic talent reserve is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of talent evaluation, and in particular to a talent evaluation decision support system and method based on deep learning and dynamic weights. BACKGROUND

[0002] With the digital transformation of enterprises and the improvement of the intelligent level of human resource management, talent competency evaluation based on multi-source data has become an important research direction in the field of human resources. The existing talent competency evaluation methods mostly rely on static ability factor weighting, linear aggregation model or experience-driven scoring system. The existing talent competency evaluation methods usually score the ability of talents based on a single dimension or shallow features, which cannot effectively capture the multi-granularity ability characteristics of talents under different post levels and organizational structures. In addition, the existing methods lack systematic integration and modeling capability for heterogeneous data, sparse information and high-noise talent data, resulting in coarse granularity, semantic fragmentation and poor post migration adaptability of the evaluation results. With the rapid changes in organizational strategy and business environment, the traditional static weight mechanism cannot timely reflect the dynamic needs of the organization for talent competency structure, making it difficult to adapt to post demand hotspots and strategic adjustments, which easily leads to distorted competency evaluation and misallocation of talent resources.

[0003] In practical applications, the traditional method lacks effective modeling of the nonlinear coupling relationship between post levels, organizational ecology and individual ability, and the evaluation model has limited interpretability and generalizability, making it difficult to provide scientific basis for agile and refined talent allocation and development decisions for organizations. SUMMARY

[0004] One object of the present application is to provide a talent evaluation decision support system and method based on deep learning and dynamic weights. The present application realizes intelligent recommendation of talent development path, post optimization allocation and organizational strategic talent reserve.

[0005] A talent evaluation decision support method based on deep learning and dynamic weights according to an embodiment of the present application comprises the following steps:

[0006] Collecting multi-source talent data and preprocessing to obtain a standardized talent dataset;

[0007] Continuously acquiring operating key indicator flow, performance closed-loop indicator flow and strategic key indicator flow to form a business scenario context information data flow;

[0008] Calling a post level graph and organizational structure information, initializing an improved hierarchical variational autoencoder, and inputting the improved hierarchical variational autoencoder with the standardized talent dataset to learn individual latent factor vectors, post level latent factor vectors and organizational level latent factor vectors respectively, and outputting a latent factor set;

[0009] The prior consistency calibration is performed on the latent factor set, a hierarchical regularization constraint is used to align the consistency of the latent factor distribution of each layer, and a calibrated latent factor set is generated;

[0010] The calibrated latent factor set and the business scenario context information data stream are input into a dynamic weight decision module, a dynamic weight tensor is generated according to an attention mechanism, a time preference extraction network and a reinforcement feedback signal;

[0011] The calibrated latent factor set is taken as input to perform decoding operation to obtain a competence preliminary score vector, the dynamic weight tensor is used to weight and reconstruct the competence preliminary score vector to generate a competence weighted score vector, and an ability radar chart, a post matching degree matrix and an ability gap report are output according to the competence weighted score vector.

[0012] Optionally, the standardized talent data set construction comprises:

[0013] The multi-source talent data in the human resource information system, the learning management system, the office automation system, the performance management system and the knowledge management platform are collected, desensitized, formatted, completed and noise-suppressed through interface calling, log probes and text extraction to obtain the standardized talent data set.

[0014] Optionally, the formation of the business scenario context information data stream comprises:

[0015] Business indicators including operating key indicators, performance closed-loop indicators and strategic key indicators are collected;

[0016] The minimum-maximum normalization method is used for normalization processing on each business indicator value in the operating key indicators, the performance closed-loop indicators and the strategic key indicators;

[0017] The normalized operating key indicators, the performance closed-loop indicators and the strategic key indicators are respectively reorganized into time sequence matrices according to time sequence, and the business indicator values are organized in time steps to form a business scenario context time sequence tensor;

[0018] The business scenario context time sequence tensor is subjected to smoothing and denoising processing to obtain a stationary context time sequence tensor;

[0019] Based on the stationary context time sequence tensor, an input feature vector set for dynamic weight generation is constructed.

[0020] Optionally, the initialization of the improved hierarchical variational autoencoder and the output of the latent factor set comprise:

[0021] construct a competency semantic guidance matrix based on the post hierarchical atlas and the organizational structure information, the competency semantic guidance matrix being used to guide parameter initialization of an individual latent factor layer, a post layer latent factor layer and an organizational layer latent factor layer in the improved hierarchical variational autoencoder network;

[0022] input the standardized talent dataset into the individual latent factor layer fused with the competency semantic guidance matrix, the individual latent factor layer adopting a dual encoding mechanism of the fusion attention mechanism and the Gaussian inference network to learn and output an individual latent factor vector;

[0023] jointly input the individual latent factor vector and the competency semantic guidance matrix into the post layer latent factor layer, and learn and output a post layer latent factor vector through a semantic guidance interlayer variational attention mechanism;

[0024] jointly input the post layer latent factor vector and the competency semantic guidance matrix into the organizational layer latent factor layer, and learn and output an organizational layer latent factor vector by using a method combining an organizational competency structure embedding network and a hierarchical Gaussian variational inference network;

[0025] perform multi-scale semantic fusion on the individual latent factor vector, the post layer latent factor vector and the organizational layer latent factor vector, and output a latent factor set.

[0026] Optionally, the prior consistency calibration performed on the latent factor set comprises:

[0027] set a prior distribution constraint for the individual latent factor vector, the post layer latent factor vector and the organizational layer latent factor vector respectively;

[0028] define a prior consistency loss function for measuring information deviation between a posterior distribution of each layer latent factor vector and a prior distribution corresponding to the posterior distribution, and the information deviation is measured by using Kullback-Leibler divergence;

[0029] introduce a multi-level dynamic alignment regularization term to measure the ability transfer consistency between the post layer and the individual layer and between the organizational layer and the post layer;

[0030] the multi-level dynamic alignment regularization term is the Euclidean distance square between a mathematical expectation of the post layer latent factor vector of each post family and a mathematical expectation of the individual latent factor vectors of all talent individuals in the post family, multiplied by a regularization weight of an ability transfer consistency index, the Euclidean distance square between a mathematical expectation of the organizational layer latent factor vector of each organizational unit and a mathematical expectation of the post layer aggregated latent factor vectors of all post families in the organizational unit, multiplied by a regularization weight of an ability transfer consistency index, and the weighted results are obtained;

[0031] The information bias is added to the multi-level dynamic alignment regularization term to construct a joint optimization objective. The joint optimization objective is then optimized by backpropagation to adjust and improve the network parameters of each layer of the hierarchical variational autoencoder, and output the set of calibration latent factors after consistency calibration.

[0032] Optionally, generating the dynamic weight tensor includes:

[0033] The calibration latent factor set is combined with the business scenario context information data stream to construct a joint input tensor for dynamic weight generation;

[0034] The joint input tensor is input into the attention mechanism network, and the attention mechanism network is used to calculate the attention distribution vector of competency semantic weight perception.

[0035] The attention-focused representation is obtained by element-wise weighting the attention distribution vector with the joint input tensor.

[0036] The attention-focused input time series preference extraction network is used to extract the dynamic evaluation preference vectors of job families and organizational units in the time series based on the time window.

[0037] The dynamic weight generation input set is constructed by fusing attention focus representation, dynamic evaluation preference vector and reinforcement feedback signal, and then performing nonlinear mapping to output dynamic weight tensor.

[0038] Optionally, the step of outputting a competency radar chart, a job matching matrix, and a competency gap report based on a competency-weighted scoring vector includes:

[0039] The set of calibration potential factors is used as the input to the decoder. The decoder generates a preliminary competency score vector based on the individual potential factor vector, the job-level potential factor vector, and the organizational-level potential factor vector in the set of calibration potential factors.

[0040] Multiply the dynamic weight tensor element by element with the initial competency scoring vector to obtain the competency weighted scoring vector.

[0041] The competency radar chart, job matching matrix, and competency gap report are generated based on the competency weighted scoring vector.

[0042] Optionally, the capability radar chart is based on a competency-weighted scoring vector, mapping the scores of each competency label to the radius values ​​of each dimension of the radar chart.

[0043] The job matching matrix is ​​calculated based on the Euclidean distance or similarity function between each dimension of the competency weighted scoring vector and the competency profile standard of the target job, and outputs the matching score between each job and the current talent.

[0044] The competency gap report compares the competency weighted scoring vector with the job competency profile standard. It determines whether each competency tag meets the target requirements of the job profile. If it is lower than the target requirements, it is recorded as a competency gap. The report then calculates all competency gap tags and their sizes, and outputs a complete competency gap report.

[0045] A talent assessment decision support system based on deep learning and dynamic weights, used to execute a talent assessment decision support method based on deep learning and dynamic weights, includes:

[0046] The full-domain data collection and governance module is used to collect multi-source talent data and preprocess it to obtain a standardized talent dataset.

[0047] The business scenario context information acquisition module is used to collect and synchronize key operational indicators, performance closed-loop indicators, and strategic key indicators to obtain a stationary context time series tensor.

[0048] The hierarchical variational autoencoder module is improved. The network parameters are initialized based on the job level map and organizational structure information. The standardized talent dataset is used as input. The hierarchical encoding of individual potential factors, job level potential factors and organizational level potential factors is performed by integrating the competency semantic guidance matrix. The potential factor set is output and the prior consistency calibration of the potential factor set is performed to output the calibrated potential factor set.

[0049] The dynamic weight decision module concatenates the calibration latent factor set with the stationary context time series tensor into a joint input tensor, calculates the dynamic evaluation preference through the attention mechanism network and the temporal preference extraction network, and outputs the dynamic weight tensor.

[0050] The competency scoring and visualization analysis module takes the calibration potential factor set and dynamic weight tensor as input, decodes them to generate a preliminary competency scoring vector, and multiplies it with the dynamic weight tensor to obtain a weighted competency scoring vector. It outputs a competency radar chart, a job matching matrix, and a competency gap report to complete talent assessment decisions.

[0051] The beneficial effects of this invention are:

[0052] (1) This invention realizes highly interpretable representation of the multi-layer potential factor structure of job-organization-individual, improves the scientific nature of ability transfer and job adaptation, introduces an improved hierarchical variational autoencoder model guided by competency semantics to embed job hierarchy map and organizational structure information into the encoding and inference process of potential factors of individuals, job level, and organization level, realizes the structural integration and hierarchical expression of competency elements of knowledge, skills, traits, motivation and values ​​at different levels, and through the design of distributed prior, semantic alignment and dynamic regularization, the improved hierarchical variational autoencoder model can accurately reflect the differentiated needs of different job families and organizational units for competency spectrum, so that the talent competency representation has high interpretability and job transfer availability, significantly improving the scientific decision-making basis in job recommendation and organizational talent mobility.

[0053] (2) The dynamic weight tensor generation mechanism of the present invention adaptively integrates multi-source scenario data, effectively eliminating the evaluation time lag and weight rigidity. By designing dynamic weights with attention mechanism, time-series preference extraction network and enhanced feedback signal as the core, it can perceive changes in business scenario context, performance feedback and job profile in real time and dynamically adjust the weights of each dimension of competency and scoring criteria. By calibrating the deep integration of latent factor set with historical performance and business indicators, it solves the defect of static weighting in the traditional evaluation system that cannot reflect organizational strategy and business adjustment in a timely manner, and reduces the competency evaluation error and mismatch probability caused by environmental changes.

[0054] (3) In the decoding stage, the present invention combines dynamic weight tensor to reconstruct the latent factor scoring vector, outputs the competency weighted scoring vector, and generates a competency radar chart, job matching matrix and competency gap report. It can automatically identify the relative advantages and disadvantages of talents on each key competency label, quantify job suitability and competency shortcomings, and realize intelligent recommendation of talent development path, job optimization and allocation and organizational strategic talent reserve. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 This is a flowchart of a talent evaluation decision support system and method based on deep learning and dynamic weights proposed in this invention. Detailed Implementation

[0057] Example 1:

[0058] refer to Figure 1 A talent assessment decision support method based on deep learning and dynamic weights includes the following steps:

[0059] Collect and preprocess talent data from multiple sources to obtain a standardized talent dataset;

[0060] Continuously acquire key operational metrics, performance closed-loop metrics, and strategic priority metrics to form a business scenario context information data stream;

[0061] Call the job level map and organizational structure information to initialize the improved hierarchical variational autoencoder. The improved hierarchical variational autoencoder takes the standardized talent dataset as input and learns the individual latent factor vector, job level latent factor vector and organizational level latent factor vector respectively, and outputs the latent factor set.

[0062] Perform prior consistency calibration on the latent factor set, and use hierarchical regularization constraints to align the distribution of latent factors in each layer to generate a calibration latent factor set;

[0063] The calibration latent factor set and the business scenario context information data stream are input into the dynamic weight decision module, and a dynamic weight tensor is generated based on the attention mechanism, the temporal preference extraction network and the reinforcement feedback signal.

[0064] Using the calibration potential factor set as input, a decoding operation is performed to obtain the preliminary competency score vector. The preliminary competency score vector is then reconstructed by weighting using a dynamic weight tensor to generate a weighted competency score vector. Based on the weighted competency score vector, a competency radar chart, a job matching matrix, and a competency gap report are output.

[0065] In this implementation, the construction of the standardized talent dataset includes:

[0066] By using API calls, log probes, and text extraction, multi-source talent data from human resource information systems, learning management systems, office automation systems, performance management systems, and knowledge management platforms are collected, anonymized, formatted, filled in missing data, and noise suppressed to obtain a standardized talent dataset.

[0067] In this embodiment, the formation of the business scenario context information data stream includes:

[0068] Collect business metrics including key operational metrics streams, performance closed-loop metrics streams, and strategic priority metrics streams;

[0069] During the data collection process, a unified time sampling window is set up to assign a timestamp, indicator value, indicator source system identifier, and business indicator category label to each business indicator record. The initial data stream of business indicators based on a unified time axis is realized through timestamp alignment. The number of business indicator records is the total number of records. The timestamp is used to mark the time of business indicator collection, the indicator value is the corresponding business measurement result, the indicator source system identifier is used to distinguish the source of the business system, and the business indicator category label is used to distinguish different business scenarios.

[0070] The initial data streams of business indicators are classified and processed according to the category labels to which the business indicators belong. The key operational indicator stream, the performance closed-loop indicator stream, and the strategic key indicator stream are extracted to ensure that the key operational indicator stream, the performance closed-loop indicator stream, and the strategic key indicator stream do not overlap. All initial data streams of business indicators are obtained by merging the key operational indicator stream, the performance closed-loop indicator stream, and the strategic key indicator stream.

[0071] The values ​​of each business indicator in the key operational indicators stream, the performance closed-loop indicators stream, and the strategic key indicators stream are normalized using the min-max normalization method.

[0072] The min-max normalization method subtracts the minimum business indicator value under the corresponding category from each business indicator value and divides it by the difference between the maximum and minimum business indicator values ​​under the corresponding category, normalizing all business indicator values ​​to the interval [0, 1]. The normalization process ensures that the key business indicator stream, the performance closed-loop indicator stream, and the strategic key indicator stream have a unified dimension.

[0073] The normalized operational key indicator stream, performance closed-loop indicator stream, and strategic key indicator stream are reassembled into a time series matrix according to time order, and the business indicator values ​​are organized into a business scenario context time series tensor by time step.

[0074] The feature dimensions of the business scenario context time series tensor are equal to the sum of the feature dimensions of key operational indicators, performance closed-loop indicators, and strategic priority indicators.

[0075] Smoothing and denoising are performed on the business scenario context time series tensor to obtain a stationary context time series tensor.

[0076] The smoothing and denoising process involves using a moving average method to locally smooth the business indicator values ​​at each time step in the business scenario context time series tensor, and using an exponentially weighted moving average method to perform trend smoothing on the business indicator values ​​at each time step in the business scenario context time series tensor.

[0077] Construct a set of input feature vectors for dynamic weight generation based on stationary context time series tensors;

[0078] The input feature vector set provides a structured description of the business scenario context information data flow in units of time steps.

[0079] In this embodiment, the improvement to the initialization and latent factor set output of the hierarchical variational autoencoder includes:

[0080] A competency semantic guidance matrix is ​​constructed based on job level map and organizational structure information. The competency semantic guidance matrix is ​​used to guide the parameter initialization of individual latent factor layer, job level latent factor layer and organizational level latent factor layer in the improved hierarchical variational autoencoder network.

[0081] The competency semantic guidance matrix is ​​jointly constructed based on the job-competency requirement weight matrix and the organization-competency requirement weight matrix. The job-competency requirement weight matrix is ​​structured with job families as rows and competency tags as columns. The element values ​​are obtained by comprehensively analyzing job descriptions, expert annotations, and statistical analysis of historical evaluation data. The organization-competency requirement weight matrix is ​​structured with organizational units as rows and competency tags as columns. The element values ​​are obtained by comprehensively assessing the requirement intensity based on business strategy, organizational culture, and talent planning information. The competency semantic guidance matrix is ​​generated using a weighted linear fusion method, guided by the multidimensional requirements of job families and organizational units. This ensures that each element of the competency semantic guidance matrix expresses the overall requirement intensity of the job and organization in the competency dimension.

[0082] The standardized talent dataset is input into the individual latent factor layer that integrates the competency semantic guidance matrix. The individual latent factor layer learns and outputs the individual latent factor vector using a dual encoding mechanism that integrates attention mechanism and Gaussian inference network.

[0083] The fusion attention mechanism weights and aggregates knowledge, skills, traits, motivations, and values ​​in the standardized talent dataset based on the job hierarchy map, organizational structure information, and the weight distribution of the competency semantic guidance matrix to obtain an attention aggregation feature vector. The attention aggregation feature vector reflects the relative importance of individual competency indicators in job and organizational contexts.

[0084] The Gaussian inference network takes the attention aggregation feature vector as input, learns the mean and variance parameters of the multidimensional Gaussian distribution to obtain the individual latent factor vector, which is used to represent the implicit structural features of talents in the dimensions of knowledge, skills, traits, motivation and value competence.

[0085] Individual latent factor vectors and competency semantic guidance matrices are jointly input into the job-level latent factor layer. The job-level latent factor vectors are learned and output through a semantically guided inter-layer variational attention mechanism.

[0086] The job-level latent factor encoder network built into the job-level latent factor layer performs convergence calculation on all individual latent factor vectors within the same job family using an inter-layer variational attention mechanism guided by job semantics. In this embodiment, one calculation method based on the inter-layer variational attention mechanism guided by job semantics is as follows:

[0087] The matching correlation between the individual latent factor vector of each talent in the job family and the job characteristics of the job family is used as the attention weight of the job layer to perform weighted summation on the individual latent factor vectors in the job family, so as to obtain the job layer aggregate representation. The job layer aggregate representation is parametrically modeled by variational inference method to output the job layer latent factor vector.

[0088] Job-level latent factor vectors are used to characterize the structural requirements of a specific job family for individual talent latent factor vectors, reflecting the similarity and competency differences among individual latent factor vectors within the job family.

[0089] The job-level potential factor vector and the competency semantic guidance matrix are jointly input into the organizational-level potential factor layer. The organizational-level potential factor vector is learned and output by combining the organizational competency structure embedding network and the hierarchical Gaussian variational inference network.

[0090] In the organizational competency structure embedding network, the job-level latent factor vector is mapped to a structural embedding representation based on organizational structure information and the synergistic relationship between job families. The structural embedding representation is used to reflect the coupling strength of competency requirements between different job families in the organization. In the hierarchical Gaussian variational inference network, the job-level latent factor vector and the structural embedding representation are jointly input, and the organizational-level latent factor vector is probabilistically modeled by Gaussian variational distribution in the hierarchical Gaussian variational inference network to output the organizational-level latent factor vector. The organizational-level latent factor vector is used to reveal the structural requirements of the organization as a whole for the job-level latent factor vector, and represents the mutual coupling relationship of competency structure between job families within the organization.

[0091] The individual potential factor vector, the job-level potential factor vector, and the organizational-level potential factor vector are semantically fused at multiple scales to output a potential factor set.

[0092] Example 1 achieves highly interpretable representation of a multi-layered latent factor structure of job-organization-individual, enhancing the scientific nature of competency transfer and job adaptation. An improved hierarchical variational autoencoder model guided by competency semantics embeds job hierarchy maps and organizational structure information into the encoding and inference process of latent factors at the individual, job, and organizational levels. This enables the structural fusion and hierarchical expression of competency elements such as knowledge, skills, traits, motivation, and values ​​at different levels. Through the design of distributed priors, semantic alignment, and dynamic regularization, the improved hierarchical variational autoencoder model can accurately reflect the differentiated needs of different job families and organizational units for competency spectrums, making talent competency representation highly interpretable and usable for job transfer, significantly improving the scientific decision-making basis for job recommendation and organizational talent mobility.

[0093] In this embodiment, prior consistency calibration is performed on the latent factor set, including:

[0094] Prior distribution constraints are set for individual potential factor vectors, job-level potential factor vectors, and organizational-level potential factor vectors, respectively.

[0095] Define a prior consistency loss function to measure the information deviation between the posterior distribution of each latent factor vector and its corresponding prior distribution;

[0096] Information bias is measured using Kullback-Leibler divergence. The posterior distribution of individual latent factor vectors is obtained by passing the standardized talent dataset through the individual latent factor layer encoder. The posterior distribution of job-level latent factor vectors is obtained by passing the individual latent factor vectors through the job-level latent factor layer. The posterior distribution of organizational-level latent factor vectors is obtained by passing the job-level latent factor vectors through the organizational-level latent factor layer. The prior consistency loss function is equal to the sum of the Kullback-Leibler divergences between the posterior distributions of individual latent factor vectors, job-level latent factor vectors, and organizational-level latent factor vectors and their respective prior distributions.

[0097] A multi-level dynamic alignment regularization term is introduced to measure the consistency of capability transfer between the job level and the individual level, and between the organizational level and the job level.

[0098] Calculate the squared Euclidean distance between the expected value of the job-level potential factor vector for each job family and the expected value of the potential factor vectors of all individual talents within the job family, and multiply by the regularization weight of the ability transfer consistency index. For each organizational unit, calculate the squared Euclidean distance between the expected value of the organizational-level potential factor vector of the organizational unit and the expected value of the aggregated potential factor vectors of all job levels within the organizational unit, and multiply by the regularization weight of the ability transfer consistency index.

[0099] ;

[0100] in, For the first Talent individual potential factor vectors under job families For the first Job family job layer potential factor vector, For the first Organizational unit job level aggregation potential factor vector, For the first Organizational unit's organizational layer latent factor vector, This represents the mathematical expectation of the corresponding distribution. The regularization weights for the capability transfer consistency index are used to adjust the strength of inter-layer capability transfer consistency within job families and organizational units, respectively. For job groups A collection of organizational units.

[0101] The information bias is added to the multi-level dynamic alignment regularization term to construct a joint optimization objective. The joint optimization objective is then optimized by backpropagation to adjust and improve the network parameters of each layer of the hierarchical variational autoencoder, and output the set of calibration latent factors after consistency calibration.

[0102] Adjusting and improving the network parameters of each layer of the hierarchical variational autoencoder is used to achieve dynamic consistency alignment of the distribution of potential factors of individual talents, job families, and organizational units under different business scenarios and organizational strategies. The joint optimization objective enables the calibration potential factor set to inherit the distribution pattern of the historical high-performance competency spectrum and to adaptively adjust with changes in business and strategy.

[0103] In this embodiment, generating the dynamic weight tensor includes:

[0104] The calibration latent factor set is combined with the business scenario context information data stream to construct a joint input tensor for dynamic weight generation;

[0105] The splicing operation is used to merge the calibration latent factor set with the business scenario context information data stream in a dimension to generate a joint input tensor.

[0106] The joint input tensor is input into the attention mechanism network, and the attention mechanism network is used to calculate the attention distribution vector of competency semantic weight perception.

[0107] Each element in the attention distribution vector represents the attention score of the corresponding feature dimension in the joint input tensor. All attention scores are summed to 1, and each attention score takes a value between 0 and 1. The attention distribution vector is used to reflect the importance of each input feature dimension to the dynamic weight generation in the current business context.

[0108] The attention-focused representation is obtained by element-wise weighting the attention distribution vector with the joint input tensor.

[0109] Element-wise weighting involves multiplying each input feature dimension by its corresponding attention score to obtain an attention-focused representation. This attention-focused representation is used to highlight the most representative feature dimension in the current business context.

[0110] The attention-focused input time series preference extraction network is used to extract the dynamic evaluation preference vectors of job families and organizational units in the time series based on the time window.

[0111] The temporal preference extraction network analyzes the changes in input features at different time steps and outputs the degree of preference for each competency dimension at the current moment. The dynamic evaluation preference vector is used to describe the trend of the focus of job families and organizational units on competency evaluation at different time stages over time.

[0112] The dynamic weight generation input set is constructed by fusing attention focus representation, dynamic evaluation preference vector and reinforcement feedback signal, and then performing nonlinear mapping to output dynamic weight tensor;

[0113] The enhanced feedback signal is composed of the competency assessment error of the previous cycle and the deviation of actual business performance. It is used to adjust the direction and intensity of the dynamic weight generation in the current cycle. The dynamic weight generation input set is input into a deep neural network to achieve nonlinear mapping and output a dynamic weight tensor. The rows of the dynamic weight tensor represent the number of competency dimensions, and the columns represent the rating dimension or job role dimension.

[0114] The dynamic weight tensor reflects the weighted relationship between each competency dimension and the scoring dimension in the current context. The real-time generation of the dynamic weight tensor enables each competency assessment result to respond promptly to organizational strategic guidance, changes in job requirements, and historical performance feedback, thereby improving the adaptability and accuracy of the competency assessment model.

[0115] The dynamic weight tensor generation mechanism in Example 1 adaptively integrates multi-source scenario data, effectively eliminating evaluation lag and weight rigidity. By designing dynamic weights with attention mechanisms, temporal preference extraction networks, and reinforcement feedback signals as the core, it can perceive changes in business scenario context, performance feedback, and job profiles in real time and dynamically adjust the weights of each dimension of competency and scoring criteria. By calibrating the deep integration of latent factor set with historical performance and business indicators, it solves the defect of static weighting in traditional evaluation systems that cannot reflect organizational strategy and business adjustments in a timely manner, and reduces competency evaluation errors and mismatch probability caused by environmental changes.

[0116] In this implementation, based on the competency-weighted scoring vector, a competency radar chart, a job matching matrix, and a competency gap report are output, including:

[0117] The set of calibration potential factors is used as the input to the decoder. The decoder generates a preliminary competency score vector based on the individual potential factor vector, the job-level potential factor vector, and the organizational-level potential factor vector in the set of calibration potential factors.

[0118] Each dimension of the preliminary competency scoring vector corresponds to a competency label, representing the predicted talent ability score under each competency label.

[0119] Multiply the dynamic weight tensor element by element with the initial competency scoring vector to obtain the competency weighted scoring vector.

[0120] The rows of the dynamic weight tensor correspond one-to-one with the dimensions of the initial competency scoring vector, and the columns correspond to the job role dimension or scoring dimension. The product of each element represents the weighted score of the corresponding competency dimension in the current business context. The competency weighted scoring vector is used to reflect the fusion effect of the calibration potential factor set and the business scenario context.

[0121] The competency radar chart, job matching matrix, and competency gap report are generated based on the competency weighted scoring vector.

[0122] In this embodiment, the competency radar chart is based on the competency weighted scoring vector, and maps the scores of each competency label to the radius values ​​of each dimension of the radar chart, so as to visualize the distribution of talents in different competency dimensions.

[0123] The job matching matrix is ​​calculated based on the Euclidean distance or similarity function between each dimension of the competency weighted scoring vector and the competency profile standard of the target job, and outputs the matching score between each job and the current talent. The job matching score is used to determine the suitability of the talent for different jobs.

[0124] The competency gap report compares the competency weighted scoring vector with the job competency profile standard. It determines whether each competency tag meets the target requirements of the job profile. If it is lower than the target requirements, it is recorded as a competency gap. The report then calculates all competency gap tags and their sizes, and outputs a complete competency gap report.

[0125] A talent assessment decision support system based on deep learning and dynamic weights, used to execute a talent assessment decision support method based on deep learning and dynamic weights, includes:

[0126] The full-domain data collection and governance module is used to collect multi-source talent data and preprocess it to obtain a standardized talent dataset.

[0127] The business scenario context information acquisition module is used to collect and synchronize key operational indicators, performance closed-loop indicators, and strategic key indicators to obtain a stationary context time series tensor.

[0128] The hierarchical variational autoencoder module is improved. The network parameters are initialized based on the job level map and organizational structure information. The standardized talent dataset is used as input. The hierarchical encoding of individual potential factors, job level potential factors and organizational level potential factors is performed by integrating the competency semantic guidance matrix. The potential factor set is output and the prior consistency calibration of the potential factor set is performed to output the calibrated potential factor set.

[0129] The dynamic weight decision module concatenates the calibration latent factor set with the stationary context time series tensor into a joint input tensor, calculates the dynamic evaluation preference through the attention mechanism network and the temporal preference extraction network, and outputs the dynamic weight tensor.

[0130] The competency scoring and visualization analysis module takes the calibration potential factor set and dynamic weight tensor as input, decodes them to generate a preliminary competency scoring vector, and multiplies it with the dynamic weight tensor to obtain a weighted competency scoring vector. It outputs a competency radar chart, a job matching matrix, and a competency gap report to complete talent assessment decisions.

[0131] Example 1 combines dynamic weight tensors to reconstruct the latent factor scoring vector during the decoding stage, outputs a competency weighted scoring vector, and generates a competency radar chart, job matching matrix, and competency gap report. It can automatically identify the relative strengths and weaknesses of talents in each key competency label, quantify job suitability and competency shortcomings, and realize intelligent recommendations for talent development paths, job optimization and allocation, and organizational strategic talent reserves.

[0132] Example 2:

[0133] In the pilot assessment of the organization's job competency upgrade, the implementers selected a multi-position mixed team of 48 employees. The multi-position mixed team involved 5 job families and 2 organizational units. The members included front-line positions, management positions and multi-skill positions. During the data collection phase, the system automatically accessed the employees' performance appraisal records, project delivery logs, learning curves from training and learning platforms, collaboration behavior logs, job rotation information and self-commitment of abilities to collect a data table matrix of employees × indicators. It covered 12 competency tags, including investment analysis, risk management, team collaboration, learning and growth, and innovation ability, as well as quantifiable business indicators such as quarterly project success rate, number of job rotations and completion of competency improvement courses. Each employee also had KPI scores for the past 8 quarters. The average number of collaboration behavior log entries was about 52, and the average number of project delivery records was about 6 per person.

[0134] During the formation of the data stream within the business scenario context, the implementer sets up a business indicator collection window with a quarterly sampling period to obtain key strategic indicators such as the quarterly revenue growth of business lines, the speed of new business development, and the response time of management processes. The data stream is then automatically subjected to minimum-maximum normalization and outlier removal. In Example 2, the original distribution of the new business development speed indicator within this sampling period was 1.2-2.5, which, after normalization, falls within the range of [0.0, 1.0]. The original range of the project success rate was 0.60-0.97, which, after normalization, falls within the range of [0.14, 1.0], facilitating unified analysis across different job roles.

[0135] The system dynamically generates a job-competency requirement weight matrix and an organization-competency requirement weight matrix based on the latest job hierarchy map. In Example 2, job family A has a requirement weight of 0.75 for innovation capability and 0.58 for risk management; job family B has a requirement weight of 0.82 for teamwork and 0.65 for investment analysis; organizational unit 1 has a requirement weight of 0.72 for learning and growth, and organizational unit 2 has a requirement weight of 0.77 for management decision-making ability. The competency semantic guidance matrix is ​​generated by linearly fusing the weights of job families and organizational units.

[0136] During the hierarchical variational autoencoder operation phase, the system outputs an individual latent factor vector for each employee using the aforementioned standardized data as input. The vector length is 12, and each component is mapped to 12 competency labels. Taking employee A's individual latent factors as an example: the innovation capability dimension is 0.63, the risk management dimension is 0.71, and the teamwork dimension is 0.81. The job-level latent factor vector is obtained by aggregating all individual latent factors under the job family through job requirement weights: the job family A's job-level latent factor aggregation value in the innovation capability dimension is 0.67. The organizational-level latent factor vector is calculated by aggregating all job-level latent factors under the organizational unit and organizational requirement weights. In Example 2, the organizational-level latent factor of organizational unit 1 in the learning and growth dimension is 0.74.

[0137] The system simultaneously uses the historical high-performing employee distribution as a prior distribution constraint to compare the current potential factor distribution of employees with the historical high-performing distribution. In the conditional Kullback-Leibler divergence measurement stage, the implementer observed that the KL divergence between the individual potential factor distribution and the historical high-performing distribution within job family A was 0.021, in job family B it was 0.027, in organizational unit 1 it was 0.018, and in organizational unit 2 it was 0.024. By introducing a capability transfer consistency regularization term, the transfer consistency index between potential factors at the job level and the organizational level was less than 0.04, thus meeting the structural alignment requirements.

[0138] After splicing and calibrating the latent factor set and the business scenario context information data stream, the system inputs the dynamic weight decision module. The attention mechanism network adaptively adjusts the weights of capability labels under different job families and business scenarios. In Example 2, for job family A, the innovation capability weight is adjusted from a static 0.75 to 0.80, and the risk management weight is adjusted from 0.58 to 0.62, reflecting the dynamic changes in capability requirements of business strategy. The time-series preference extraction network outputs a dynamic evaluation preference vector, showing that in the past four quarters, the learning and growth weight of job family A has gradually increased from 0.52 to 0.61. In the reinforcement feedback signal, the job recommendation error in the previous period was 0.11, and the actual business performance deviation was 0.07. Based on this, the system fine-tunes the weights of each label in the current period.

[0139] During the competency assessment phase, the system takes the calibration potential factor set as input, and the decoder generates a preliminary competency assessment vector. In Example 2, employee B's preliminary assessment scores are 0.78 in teamwork and 0.67 in innovation. Subsequently, the dynamic weight tensor is multiplied element-wise with the preliminary assessment vector to obtain the weighted competency assessment vector. In Example 2, employee B's teamwork score increases to 0.83 and innovation score increases to 0.70 after weighting in the current period. For the composite positions of job family B, the dynamic weighting mechanism automatically reduces the risk control ability score weight of some employees due to business mistakes in the past two quarters, adjusting it from 0.68 to 0.62, which is more in line with the actual business needs.

[0140] The system automatically generates a capability radar chart, a job matching matrix, and a capability gap report. In Example 2, the capability radar chart of employee C shows that his communication, collaboration, learning, growth, and innovation capabilities all exceed 0.75. In the job matching matrix, his fit with job family B is 0.92 and his fit with job family A is 0.85. The capability gap report indicates that employee C's score on the risk management tag is 0.61, which is lower than the target position's score of 0.72, with a gap of 0.11. The system automatically pushes a risk management-specific training recommendation.

[0141] In the comparative analysis, the system performed a full-process result verification between the method of this invention and the traditional static weighted method. Under the traditional method, the average job matching degree of team members was 0.77, while after adopting the present invention, the average value increased to 0.86. The accuracy rate of capability gap identification under the traditional method was 68%, while that under the present invention was 89%. Under the traditional method, 5 management personnel were recommended to unsuitable positions due to insufficient granularity of the capability model, and their performance evaluation score dropped by an average of 4.1 points after six months. After the present invention, the job matching recommendation error rate was reduced to 1 / 5, and the average performance score of employees transferred to other positions increased by 6.3 points within six months. Further statistics revealed that there was a lag of more than 1.5 weeks between the assessment and actual business adjustments under the traditional method. The dynamic weighting mechanism of the present invention supports real-time data stream adjustments, shortening the assessment-recommendation-adjustment cycle to 1-2 days. In the quarter with the highest weight of innovation capability in the business scenario, the number of team innovation projects increased by 22% year-on-year, while the control team using the traditional method only increased by 7%.

[0142] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A talent evaluation decision support method based on deep learning and dynamic weights, characterized in that, include: Collect and preprocess talent data from multiple sources to obtain a standardized talent dataset; Continuously acquire key operational metrics, performance closed-loop metrics, and strategic priority metrics to form a business scenario context information data stream; Call the job level map and organizational structure information to initialize the improved hierarchical variational autoencoder. The improved hierarchical variational autoencoder takes the standardized talent dataset as input and learns the individual latent factor vector, job level latent factor vector and organizational level latent factor vector respectively, and outputs the latent factor set. Perform prior consistency calibration on the latent factor set, and use hierarchical regularization constraints to align the distribution of latent factors in each layer to generate a calibration latent factor set; The calibration latent factor set and the business scenario context information data stream are input into the dynamic weight decision module, and a dynamic weight tensor is generated based on the attention mechanism, the temporal preference extraction network and the reinforcement feedback signal. Using the calibration potential factor set as input, a decoding operation is performed to obtain the preliminary competency score vector. The preliminary competency score vector is then reconstructed by weighting using a dynamic weight tensor to generate a weighted competency score vector. Based on the weighted competency score vector, a competency radar chart, a job matching matrix, and a competency gap report are output.

2. The talent evaluation decision support method based on deep learning and dynamic weights according to claim 1, characterized in that, The construction of the standardized talent dataset includes: By using API calls, log probes, and text extraction, multi-source talent data from human resource information systems, learning management systems, office automation systems, performance management systems, and knowledge management platforms are collected, anonymized, formatted, filled in missing data, and noise suppressed to obtain a standardized talent dataset.

3. The talent evaluation decision support method based on deep learning and dynamic weights according to claim 1, characterized in that, The formation of the business scenario context information data stream includes: Collect business metrics including key operational metrics streams, performance closed-loop metrics streams, and strategic priority metrics streams; The values ​​of each business indicator in the key operational indicators stream, the performance closed-loop indicators stream, and the strategic key indicators stream are normalized using the min-max normalization method. The normalized operational key indicator stream, performance closed-loop indicator stream, and strategic key indicator stream are reassembled into a time series matrix according to time order, and the business indicator values ​​are organized into a business scenario context time series tensor by time step. Smoothing and denoising are performed on the business scenario context time series tensor to obtain a stationary context time series tensor. Construct a set of input feature vectors for dynamic weight generation based on stationary context time series tensors.

4. The talent evaluation decision support method based on deep learning and dynamic weights according to claim 1, characterized in that, The initialization and output of the latent factor set of the improved hierarchical variational autoencoder include: A competency semantic guidance matrix is ​​constructed based on job level map and organizational structure information. The competency semantic guidance matrix is ​​used to guide the parameter initialization of individual latent factor layer, job level latent factor layer and organizational level latent factor layer in the improved hierarchical variational autoencoder network. The standardized talent dataset is input into the individual latent factor layer that integrates the competency semantic guidance matrix. The individual latent factor layer learns and outputs the individual latent factor vector using a dual encoding mechanism that integrates attention mechanism and Gaussian inference network. Individual latent factor vectors and competency semantic guidance matrices are jointly input into the job-level latent factor layer. The job-level latent factor vectors are learned and output through a semantically guided inter-layer variational attention mechanism. The job-level potential factor vector and the competency semantic guidance matrix are jointly input into the organizational-level potential factor layer. The organizational-level potential factor vector is learned and output by combining the organizational competency structure embedding network and the hierarchical Gaussian variational inference network. The individual potential factor vector, the job-level potential factor vector, and the organizational-level potential factor vector are semantically fused at multiple scales to output a potential factor set.

5. The talent evaluation decision support method based on deep learning and dynamic weights according to claim 1, characterized in that, The prior consistency calibration of the latent factor set includes: Prior distribution constraints are set for individual potential factor vectors, job-level potential factor vectors, and organizational-level potential factor vectors, respectively. Define a prior consistency loss function to measure the information bias between the posterior distribution of each latent factor vector and its corresponding prior distribution. The information bias is measured by the Kullback-Leibler divergence. A multi-level dynamic alignment regularization term is introduced to measure the consistency of capability transfer between the job level and the individual level, and between the organizational level and the job level. The multi-level dynamic alignment regularization term is calculated by multiplying the squared Euclidean distance between the expected value of the job-level potential factor vector of each job family and the expected value of the potential factor vector of all individual talents within the job family by the regularization weight of the ability transfer consistency index. For each organizational unit, the squared Euclidean distance between the expected value of the organizational-level potential factor vector of the organizational unit and the expected value of the aggregated potential factor vector of all job levels within the organizational unit is calculated by multiplying the squared Euclidean distance between the organizational-level potential factor vector of the organizational unit and the expected value of the aggregated potential factor vector of all job levels within the organizational unit by the regularization weight of the ability transfer consistency index, and then weighted. The information bias is added to the multi-level dynamic alignment regularization term to construct a joint optimization objective. The joint optimization objective is then optimized by backpropagation to adjust and improve the network parameters of each layer of the hierarchical variational autoencoder, and output the set of calibration latent factors after consistency calibration.

6. The talent evaluation decision support method based on deep learning and dynamic weights according to claim 1, characterized in that, The generation of the dynamic weight tensor includes: The calibration latent factor set is combined with the business scenario context information data stream to construct a joint input tensor for dynamic weight generation; The joint input tensor is input into the attention mechanism network, and the attention mechanism network is used to calculate the attention distribution vector of competency semantic weight perception. The attention-focused representation is obtained by element-wise weighting the attention distribution vector with the joint input tensor. The attention-focused input time series preference extraction network is used to extract the dynamic evaluation preference vectors of job families and organizational units in the time series based on the time window. The dynamic weight generation input set is constructed by fusing attention focus representation, dynamic evaluation preference vector and reinforcement feedback signal, and then performing nonlinear mapping to output dynamic weight tensor.

7. The talent evaluation decision support method based on deep learning and dynamic weights according to claim 1, characterized in that, The output of the competency radar chart, job matching matrix, and competency gap report based on the competency-weighted scoring vector includes: The set of calibration potential factors is used as the input to the decoder. The decoder generates a preliminary competency score vector based on the individual potential factor vector, the job-level potential factor vector, and the organizational-level potential factor vector in the set of calibration potential factors. Multiply the dynamic weight tensor element by element with the initial competency scoring vector to obtain the competency weighted scoring vector. The competency radar chart, job matching matrix, and competency gap report are generated based on the competency weighted scoring vector.

8. The talent evaluation decision support method based on deep learning and dynamic weights according to claim 1, characterized in that, The capability radar chart is based on a competency-weighted scoring vector, mapping the scores of each competency label to the radius values ​​of each dimension of the radar chart. The job matching matrix is ​​calculated based on the Euclidean distance or similarity function between each dimension of the competency weighted scoring vector and the competency profile standard of the target job, and outputs the matching score between each job and the current talent. The competency gap report compares the competency weighted scoring vector with the job competency profile standard. It determines whether each competency tag meets the target requirements of the job profile. If it is lower than the target requirements, it is recorded as a competency gap. The report then calculates all competency gap tags and their sizes, and outputs a complete competency gap report.

9. A talent evaluation decision support system based on deep learning and dynamic weights, used to execute the talent evaluation decision support method based on deep learning and dynamic weights as described in claims 1-8, characterized in that, include: The full-domain data collection and governance module is used to collect multi-source talent data and preprocess it to obtain a standardized talent dataset. The business scenario context information acquisition module is used to collect and synchronize key operational indicators, performance closed-loop indicators, and strategic key indicators to obtain a stationary context time series tensor. The hierarchical variational autoencoder module is improved. The network parameters are initialized based on the job level map and organizational structure information. The standardized talent dataset is used as input. The hierarchical encoding of individual potential factors, job level potential factors and organizational level potential factors is performed by integrating the competency semantic guidance matrix. The potential factor set is output and the prior consistency calibration of the potential factor set is performed to output the calibrated potential factor set. The dynamic weight decision module concatenates the calibration latent factor set with the stationary context time series tensor into a joint input tensor, calculates the dynamic evaluation preference through the attention mechanism network and the temporal preference extraction network, and outputs the dynamic weight tensor. The competency scoring and visualization analysis module takes the calibration potential factor set and dynamic weight tensor as input, decodes them to generate a preliminary competency scoring vector, and multiplies it with the dynamic weight tensor to obtain a weighted competency scoring vector. It outputs a competency radar chart, a job matching matrix, and a competency gap report to complete talent assessment decisions.

Images