Multi-dimensional talent team evaluation method and system

Abstract

The invention relates to the technical field of intelligent talent evaluation, and discloses a multi-dimension-based talent team evaluation method and system, and the method comprises the steps: collecting original talent data, and obtaining a standardized multi-source fusion talent data set through a deep semantic fusion method of a domain ontology; adopting a five-dimensional evaluation framework design and a dynamic weight learning method to obtain a multi-dimensional evaluation index system with a comprehensive initial weight; adopting a multi-factor adaptive weight adjustment method to obtain personalized dynamic weight configuration; a progressive mixed evaluation calculation method is adopted to obtain a comprehensive evaluation score; a multi-level time sequence evolution analysis and development potential prediction method is adopted to obtain a talent comprehensive value evaluation result; and an intelligent visual analysis and multi-dimensional decision support method is adopted to obtain a complete talent evaluation output result. According to the method, multi-dimensional dynamic evaluation and development potential prediction of talent teams can be realized, and scientific support is provided for talent management decision.

Description

Technical Field

[0001] This invention relates to the field of intelligent talent assessment technology, and more specifically, to a multi-dimensional talent team evaluation method and system. Background Technology

[0002] With the advent of the knowledge economy era, talent has become a key element of an organization's core competitiveness. Traditional talent evaluation methods mainly rely on single-dimensional performance appraisals and subjective evaluations, which are insufficient to comprehensively reflect a person's overall capabilities and development potential. Existing evaluation systems suffer from problems such as limited evaluation dimensions, rigid weighting, and a lack of dynamic adjustment mechanisms, making them unable to adapt to the differentiated evaluation needs of talent in different positions and at different development stages.

[0003] In existing technologies, talent evaluation data is typically scattered across multiple business systems, with inconsistent data formats and semantics, leading to difficulties in data integration. Furthermore, the determination of evaluation indicator weights largely relies on expert experience, lacking a data-driven dynamic learning mechanism, making it difficult to reflect the causal relationships and actual contributions between indicators. In addition, traditional evaluation methods focus on static competency assessment, lacking temporal analysis of talent development trajectories and prediction of development potential, thus failing to provide forward-looking guidance for talent cultivation and talent pipeline development.

[0004] Therefore, there is an urgent need for a talent evaluation method that can integrate multi-source heterogeneous data, construct a multi-dimensional evaluation index system, realize dynamic weight adjustment, integrate expert knowledge and data-driven approaches, and conduct time-series evolution analysis and potential prediction, in order to solve the technical problems of single evaluation dimensions, fixed weight configuration, lack of dynamic adaptability and development potential prediction capabilities in existing technologies. Summary of the Invention

[0005] This invention provides a multi-dimensional talent evaluation method and system, which solves the technical problems of single evaluation dimensions, lack of dynamic adaptability and development potential prediction ability in related technologies.

[0006] This invention provides a multidimensional talent evaluation method, comprising the following steps:

[0007] S1. Collect raw talent data and use a deep semantic fusion method based on domain ontology to obtain a standardized multi-source fusion talent dataset;

[0008] S2, obtain a standardized multi-source fusion talent dataset, and use a five-dimensional evaluation framework design and dynamic weight learning method to obtain a multi-dimensional evaluation index system with comprehensive initial weights;

[0009] S3. Based on a multi-dimensional evaluation index system with comprehensive initial weights, a multi-factor adaptive weight adjustment method is adopted to obtain personalized dynamic weight configuration.

[0010] S4. Based on the standardized multi-source fusion talent dataset and personalized dynamic weight configuration, a progressive hybrid evaluation calculation method is adopted to obtain the comprehensive evaluation score.

[0011] S5, obtain the comprehensive evaluation score, and use multi-level time series evolution analysis and development potential prediction method to obtain the comprehensive value assessment result of the talent;

[0012] S6 obtains the comprehensive evaluation score and the comprehensive talent value assessment results, and uses intelligent visualization analysis and multi-dimensional decision support methods to obtain complete talent evaluation output results.

[0013] In a preferred embodiment, the deep semantic fusion method employing domain ontology includes:

[0014] A graph structure is used to represent the core concept system and semantic relationship network in the field of talent evaluation. The graph nodes include personnel entities, organization entities, project entities, ability entities, contribution entities, and time entities, and the graph edges represent the semantic relationships between concepts.

[0015] Based on a pre-trained language model, metadata information of data fields and ontology concept nodes are encoded into semantic vector representations respectively. The mapping relationship between data fields and ontology concepts is discovered by calculating the semantic vector similarity.

[0016] When semantically ambiguous data of the same person is detected in different source systems, a multi-strategy joint disambiguation method is adopted, which combines a graph path-based inference method, a context consistency-based verification method, and a constraint satisfaction-based inference method.

[0017] In a preferred embodiment, the five-dimensional evaluation framework design and dynamic weight learning method include:

[0018] The indicators at each level in the evaluation index system are used as nodes in a Bayesian network, and directed edges are constructed based on the hierarchical relationship and causal dependency between the indicators.

[0019] The initial weights of each index are calculated using the group consensus pairwise comparison matrix and the analytic hierarchy process (AHP), and the initial weights are converted into conditional probability parameters in a Bayesian network.

[0020] Historical evaluation data is used as training samples, and the conditional probability parameters of the Bayesian network are learned using the expectation-maximization algorithm.

[0021] The causal contribution of each indicator to the comprehensive evaluation score is calculated using the causal effect estimation method, and the data-driven weights are obtained through intervention analysis.

[0022] The expert judgment weight, data-driven weight, and causal inference weight are used as independent sources of evidence. The posterior weight distribution is calculated using Bayes' theorem, and the expected value is extracted from the posterior distribution as the comprehensive initial weight.

[0023] In a preferred embodiment, the multi-factor adaptive weight adjustment method includes:

[0024] Based on the organizational strategic goal configuration parameters and strategic capability mapping matrix, calculate the strategic orientation weight adjustment coefficient for each evaluation dimension;

[0025] Based on the job category and rank information of the evaluated personnel, the job competency model library is queried to obtain the job differentiation weight configuration coefficient;

[0026] Based on the data timestamp information, the timeliness weight coefficient is calculated using the time decay function. Combined with the records of major events, the weight of the associated dimensions is enhanced to obtain the time-sensitive weight adjustment coefficient.

[0027] The strategic orientation adjustment coefficient, job differentiation coefficient, and time-sensitive adjustment coefficient are weighted and combined according to the influence intensity parameter, multiplied with the overall initial weight, and then normalized.

[0028] In a preferred embodiment, the progressive hybrid evaluation calculation method includes:

[0029] The number of valid evaluation records in the historical evaluation database is counted, and the data accumulation level is compared with the preset model training threshold to determine the current deployment stage of the system. The deployment stage is divided into the initial stage, the transition stage, and the mature stage.

[0030] In the initial stage, a meta-learning strategy based on few-shot learning is adopted to construct a meta-learning network model that includes a feature extractor and a task adapter.

[0031] During the transition phase, the model trained in the meta-learning phase is loaded as the initialization parameters, and the elastic weight consolidation algorithm is used to retain the memory of the old data patterns while learning the new data patterns.

[0032] In the mature stage, a gradient boosting-based tree model, a deep neural network-based representation learning model, and a graph neural network-based relation modeling model are constructed, and the prediction results of multiple models are integrated using stacked ensemble or weighted fusion methods.

[0033] In a preferred embodiment, the progressive hybrid evaluation calculation method further includes:

[0034] Based on a standardized multi-source fusion talent dataset and personalized dynamic weight configuration, a hierarchical weighted summation method is used to aggregate and calculate step by step to obtain an evaluation score based on the analytic hierarchy process.

[0035] If the machine learning model is available, the feature data of the person being evaluated is input into the model to obtain the predicted score of the machine learning model.

[0036] The fusion weights of the two methods are adaptively adjusted according to the system deployment phase. In the initial stage, the weight of the analytic hierarchy process (AHP) is a unit value. In the transition stage, the weight of the AHP gradually decreases while the weight of the machine learning method gradually increases. In the mature stage, the weights of the two methods are dynamically determined based on the model evaluation indicators.

[0037] The analytic hierarchy process (AHP) evaluation score and the machine learning model prediction score are weighted and summed according to the fusion weights. When the deviation between the two scores exceeds a preset threshold, the evaluation result review mechanism is triggered.

[0038] In a preferred embodiment, the method of employing multi-level time-series evolution analysis and development potential prediction includes:

[0039] The historical evaluation sequence of the evaluated personnel is extracted from the talent capability time series database, and the capability score time series is decomposed into multiple intrinsic mode functions and a residual trend term using the ensemble empirical mode decomposition algorithm.

[0040] A talent development model library containing multiple typical ability development models is constructed. The dynamic time warping algorithm is used to calculate the similarity distance between the ability development sequence of the evaluated person and each standard model template, and to identify the matching development model type.

[0041] A state-space model for talent capability development is constructed, with capability level, capability change rate, and capability acceleration as latent state variables. The Kalman filter algorithm is used for state estimation, and the Bayesian change point detection algorithm is used to identify significant change points in the capability development process.

[0042] By combining the results of multi-scale decomposition, growth pattern identification, and state estimation, a multi-level trend analysis result including short-term, medium-term, and long-term trends is generated.

[0043] In a preferred embodiment, the comprehensive talent value assessment results include:

[0044] The exponential smoothing model is used for short-term development forecasting, the combination of trend extrapolation and pattern matching is used for medium-term development forecasting, and scenario analysis is used for long-term development forecasting.

[0045] The standard error of the prediction is estimated based on the distribution characteristics of historical prediction errors, and the confidence interval of the prediction is constructed.

[0046] The comprehensive evaluation score is used as the current value score. The potential value score is calculated based on the growth trend factor and the development prediction factor. The current value score and the potential value score are then summed according to their respective weights.

[0047] In a preferred embodiment, the method of employing intelligent visualization analysis and multidimensional decision support includes:

[0048] Design a talent echelon health index system that includes echelon structure indicators, capability distribution indicators, development dynamic indicators, and risk warning indicators, and set warning thresholds and alarm thresholds for each health indicator;

[0049] The system aggregates evaluation data from all team members to calculate the current values ​​of each health indicator. When an indicator value triggers a warning threshold or alarm threshold, a warning message is generated.

[0050] The matching score of candidates is calculated based on the matching rule matrix between project needs and talent capabilities. A heuristic search algorithm is used to generate team configuration optimization plan, identify common capability shortcomings of the team and match training resources.

[0051] In a preferred embodiment, a multidimensional talent evaluation system is used to perform the steps of the above-described multidimensional talent evaluation method, including:

[0052] The data fusion module is used to collect raw talent data and use a deep semantic fusion method based on domain ontology to obtain a standardized multi-source fused talent dataset.

[0053] The indicator construction module is used to obtain a standardized multi-source fusion talent dataset. It adopts a five-dimensional evaluation framework design and a dynamic weight learning method to obtain a multi-dimensional evaluation indicator system with comprehensive initial weights.

[0054] The weight adjustment module is used to obtain personalized dynamic weight configuration based on a multi-dimensional evaluation index system with comprehensive initial weights and a multi-factor adaptive weight adjustment method.

[0055] The evaluation calculation module is used to obtain a comprehensive evaluation score based on a standardized multi-source fusion talent dataset and personalized dynamic weight configuration, using a progressive hybrid evaluation calculation method.

[0056] The value assessment module is used to obtain a comprehensive evaluation score. It adopts a multi-level time-series evolution analysis and development potential prediction method to obtain the comprehensive value assessment results of talents.

[0057] The decision support module is used to obtain comprehensive evaluation scores and talent comprehensive value assessment results. It adopts intelligent visualization analysis and multi-dimensional decision support methods to obtain complete talent evaluation output results.

[0058] The beneficial effects of this invention are as follows:

[0059] By adopting a deep semantic fusion method based on domain ontology and a dynamic weight learning method based on Bayesian networks, intelligent fusion of multi-source heterogeneous data and weighted multi-source fusion of expert knowledge and data-driven data were realized. A multi-dimensional evaluation index system covering five dimensions, including professional ability, innovation potential, collaborative efficiency, growth trajectory and value contribution, was constructed. This solved the problems of single evaluation dimensions and fixed weight configuration in existing technologies, and improved the comprehensiveness and accuracy of talent evaluation.

[0060] By employing a multi-factor adaptive weight adjustment method, a progressive hybrid evaluation calculation method, and a multi-level time-series evolution analysis and development potential prediction method, this approach achieves personalized dynamic weight configuration that integrates strategic orientation, job characteristics, and time-series sensitivity; adaptive fusion of hierarchical analysis and machine learning; and growth pattern identification and development potential prediction based on multi-scale time-series decomposition. This solves the problem of the lack of dynamic adaptability and development potential prediction capabilities in existing technologies, and provides scientific support for talent management decisions. Attached Figure Description

[0061] Figure 1 This is a flowchart of a multi-dimensional talent evaluation method according to the present invention;

[0062] Figure 2 This is a flowchart of a multi-dimensional talent evaluation method according to the present invention. Detailed Implementation

[0063] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0064] At least one embodiment of the present invention discloses a multi-dimensional talent pool evaluation method, such as... Figures 1 to 2 As shown, it includes the following steps:

[0065] S1. Collect raw talent data and use a deep semantic fusion method based on domain ontology to obtain a standardized multi-source fusion talent dataset;

[0066] The process involves acquiring heterogeneous raw talent data from multiple sources, including human resource management systems, project management platforms, code hosting platforms, knowledge management systems, and collaborative office platforms. It then employs deep semantic understanding technology based on pre-trained language models, intelligent disambiguation methods based on domain ontology knowledge graphs, multi-dimensional data quality verification technology, and intelligent data completion technology based on multiple interpolation to obtain a standardized multi-source fused talent dataset that has undergone semantic fusion, quality verification, and missing data completion. Specifically, this includes the following steps:

[0067] S11, Construction of Multi-Source Data Acquisition Engine and Acquisition of Raw Data: Based on a distributed data acquisition architecture, a multi-source data acquisition engine is built using the adapter pattern to obtain the ability to acquire raw talent data covering various business systems.

[0068] Specifically, a structured data acquisition adapter is built for the human resource management system to acquire basic personnel information (personnel identifier, name, date of employment, position, job level, education, etc.) and career development history data (promotion records, job transfers, training records, performance appraisals, etc.), with a daily incremental acquisition frequency.

[0069] A project data acquisition adapter was built for the project management platform to obtain project participation records (project identifier, project type, participating roles, task list, completion status, quality evaluation, etc.), covering project records from the past three years.

[0070] Build a code activity data collection adapter for code hosting platforms to obtain code contribution data (commit records, lines of code, review participation, merge requests, technical documents, etc.) and permission level data.

[0071] A knowledge contribution data collection adapter is built for the knowledge management system to acquire knowledge output data (document writing, knowledge sharing, patent application, paper publication, scheme review, etc.) and influence indicators (read count, likes, citation count, etc.).

[0072] A collaborative behavior data collection adapter is built for collaborative office platforms to acquire collaborative communication data (message statistics, meeting participation, cross-departmental collaboration, task coordination, response time, etc.), and sensitive content is anonymized.

[0073] By executing five types of adapters in parallel, the raw data is aggregated into a unified buffer to form a multi-source raw talent data set.

[0074] S12, Intelligent data fusion and disambiguation processing based on deep semantic understanding; Based on the multi-source original talent data set output in step S11, a joint disambiguation method based on deep semantic understanding technology and domain ontology knowledge base based on pre-trained language model is adopted to obtain a high-quality fusion dataset with unified semantic expression.

[0075] Specifically, an ontology knowledge graph for the talent evaluation domain is constructed. This knowledge graph uses a graph structure to represent the core concept system and semantic relationship network of the talent evaluation domain. Graph nodes include top-level concept categories such as personnel entities, organizational entities, project entities, capability entities, contribution entities, and time entities. Under each top-level concept, a multi-level sub-concept tree is constructed. Under the personnel entity, sub-concepts such as personnel identifier, personnel attributes, personnel roles, and personnel status are defined; under the capability entity, sub-concepts such as professional skills, knowledge reserves, practical experience, and soft skills are defined. Graph edges represent semantic relationships between concepts, including membership, association, temporal, causal, and equivalence relationships. Each edge is assigned a relationship type label and a relationship strength weight.

[0076] A semantic mapping engine based on a pre-trained language model is constructed. A domain-specific language model pre-trained on a large-scale corpus of human resources and enterprise management texts is loaded, having learned rich semantic knowledge in the talent evaluation domain. For each data acquisition source system, metadata information of its data fields is extracted, including field name, field description, field type, and field value examples. This metadata information is encoded into semantic vector representations, and concept nodes in the ontology knowledge graph are also encoded into semantic vector representations. By calculating the similarity between field semantic vectors and concept semantic vectors, the semantic mapping relationship between data fields and ontology concepts is automatically discovered. For mapping relationships with similarity exceeding a threshold, a candidate mapping set is established; for cases with multiple candidate mappings, mapping disambiguation is performed by combining the contextual association information of the fields and the structural constraints of the ontology graph to select the optimal mapping relationship.

[0077] Perform semantic embedding-based data fusion processing. It iterates through each data record in the multi-source original dataset, converting source system field values ​​into ontology concept instances based on established semantic mapping relationships. During the conversion process, for text fields, a pre-trained language model is used to extract semantic embedding vectors of field values, and semantic similarity matching is used to achieve standardized mapping of field values; for numeric fields, unit standardization and dimension normalization are performed; for date fields, time zone unification and format standardization are performed; and for enumeration fields, semantic similarity matching is used to align enumeration values ​​from different source systems.

[0078] Intelligent disambiguation processing based on knowledge graph reasoning is performed. When semantically ambiguous data about the same person is detected in different source systems, a multi-strategy joint disambiguation process is initiated. A graph-based path reasoning method is used to find semantic paths between concept nodes corresponding to ambiguous data in the knowledge graph, and the semantic relevance between concepts is evaluated based on path length and path weight. A context-consistency-based verification method is used to extract contextual association information of the ambiguous data, including other field values ​​in the same record, historical data records of the same person, and data patterns of similar people, and the most reasonable semantic interpretation is selected through context-consistency scoring. A constraint-satisfaction-based reasoning method is used to eliminate candidate interpretations that do not meet the constraints, and retain the optimal interpretation that satisfies all constraints, utilizing the concept constraint rules and relation constraint rules defined in the ontology knowledge graph. For cases where a unique interpretation cannot be determined after multi-strategy joint disambiguation, the confidence score of each candidate interpretation is calculated, and the interpretation with the highest confidence score is selected as the disambiguation result, and the disambiguation confidence score is marked for subsequent quality assessment. Through deep semantic understanding and intelligent disambiguation processing, a high-quality fusion dataset with a unified ontology semantic expression is generated.

[0079] S13, Multi-dimensional data quality verification and anomaly handling: Based on the semantic fusion dataset output in step S12, multi-dimensional data quality verification technology is used to obtain a high-quality dataset that has undergone quality assessment and anomaly handling.

[0080] Specifically, the process includes: performing data integrity checks, defining lists of necessary and optional fields, checking valid field values, recording missing data, and calculating the integrity rate. Performing data consistency checks, establishing cross-source verification rules, verifying the consistency of personnel information, the reasonableness of the time sequence, and the validity of numerical ranges, and selecting master data values ​​based on the reliability of the data source. Performing data timeliness checks, evaluating update timestamps, marking overdue data with alerts, and setting timeliness thresholds based on data type differences. Performing data validity checks, verifying numerical ranges, enumerated values, and business logic constraints, and classifying data according to severity: minor issues are automatically corrected, moderate issues are marked with alerts, and severe issues are removed from the data.

[0081] Through multi-dimensional quality verification, a quality assessment report is generated for each record, including completeness rate, consistency score, timeliness level, and validity status.

[0082] S14, Intelligent data completion based on multiple interpolation: Based on the high-quality dataset and the queue to be completed output in step S13, intelligent data completion technology based on multiple interpolation algorithm is used to obtain a complete dataset with missing data reasonably filled.

[0083] Specifically, a missing pattern analysis is performed on the data to be completed, distinguishing between three patterns: completely random missing, random missing, and non-random missing. Statistical imputation, similarity group imputation, and domain rule-based imputation methods are employed for each pattern. A talent feature similarity model is constructed, extracting multi-dimensional feature vectors such as job category, job level, and years of service. Personnel similarity is calculated, and a reference group is identified. Multiple imputation processing is performed, generating multiple candidate imputation values ​​based on random sampling of the reference group data distribution. The most representative value is selected, and the imputation confidence is calculated. A posterior validation is performed on the imputation results to verify business logic constraints and data consistency. Values ​​that fail the validation are rolled back and marked as unimputable. A standardized multi-source fusion talent dataset is output, including the original data, the imputed data, and confidence indicators.

[0084] Furthermore, a conditional imputation method based on domain knowledge graph constraints can be adopted to ensure that the completed data conforms to business logic constraints. Specifically, a domain knowledge graph for talent data is constructed, and rules for inter-field dependency constraints and value constraints are defined. These constraints are used as prior conditions for the imputation model, generating a set of candidate completed values ​​that satisfy the constraints. The optimal completed value is then selected through constraint satisfaction evaluation.

[0085] S2, obtain a standardized multi-source fusion talent dataset, and use a five-dimensional evaluation framework design and dynamic weight learning method to obtain a multi-dimensional evaluation index system with comprehensive initial weights;

[0086] Obtain the standardized multi-source fusion talent dataset output from step S1. Combine this dataset with an organizational capability model library (which is pre-built and includes a job classification system, core capability requirements for each job category, career development stage models, and the theoretical foundation of talent evaluation, defining the capabilities that talents at different positions, levels, and development stages should possess and their relative importance) and historical evaluation data. Employ a five-dimensional evaluation framework structured decomposition method, group decision aggregation technology, and Bayesian network-based dynamic weight learning and multi-source fusion technology to obtain a multi-dimensional evaluation index system and its comprehensive initial weight configuration covering professional capability, innovation potential, collaboration effectiveness, growth trajectory, and value contribution dimensions. Specifically, this includes the following steps:

[0087] S21, Five-dimensional evaluation framework design and index decomposition: Based on talent evaluation theory and organizational capability model library, a five-dimensional evaluation framework is designed using a structured decomposition method, resulting in a complete evaluation index system structure containing multi-level indicators.

[0088] Specifically, the professional competence dimension includes four secondary indicators: mastery of core skills (skill certification, assessment results, application frequency), depth of professional knowledge (educational background, training duration, assessment results), technical application ability (solution output, adoption rate, problem-solving efficiency), and problem-solving ability (handling difficult problems, timeliness of resolution, and solution quality).

[0089] The dimension of design innovation potential includes four secondary indicators: activity of innovative thinking (creative proposals, adoption rate, scope of influence), originality of technical solutions (patent applications, number of innovative points, number of citations), knowledge transfer ability (cross-domain projects, learning speed, breadth of technology stack), and sensitivity to cutting-edge technologies (cutting-edge learning, trend analysis, introduction of new technologies).

[0090] The collaboration effectiveness dimension is designed to include four secondary indicators: communication and coordination ability (meeting facilitation, communication frequency, feedback and evaluation), resource integration ability (application success rate, utilization efficiency, sharing contribution), cross-departmental collaboration efficiency (project participation, completion time, satisfaction), and knowledge sharing contribution (document writing, training and delivery, question answering).

[0091] The design of the growth trajectory dimension includes four secondary indicators: learning progress curve (learning duration, skill improvement, and grade improvement rate), ability improvement rate (promotion interval, score change rate, and task difficulty trajectory), historical path stability (frequency of job changes, consistency of direction, and performance fluctuation), and career stage matching degree (ability attainment rate, task completion rate, and goal achievement rate).

[0092] The design value contribution dimension includes four secondary indicators: strategic goal contribution (participation in strategic projects, proportion of key tasks, and relevance of results), core project participation (number of projects, importance of roles, and contribution evaluation), key technology breakthrough contribution (overcoming difficulties, patent authorization, and achievement transformation), and talent cultivation radiation impact (guiding newcomers, training effectiveness, and role model influence).

[0093] Through structured decomposition, a complete evaluation indicator system is formed, comprising five primary dimensions, twenty secondary indicators, and sixty tertiary indicators.

[0094] S22, Construction of pairwise comparison matrix for group decision-making: Based on the evaluation index system structure output in step S21, the group decision-making aggregation technology is used to collect and integrate the judgments of multiple experts to obtain the pairwise comparison matrix of indicators at each level of group consensus.

[0095] Specifically, an expert committee on the weighting of evaluation indicators was established, composed of senior experts in human resource management, technology R&D management, and organizational strategic planning. A pairwise comparison questionnaire was designed, generating pairwise comparison items for each level of indicators. A nine-level scaling method was used to collect expert judgments on the relative importance of the indicators. Individual judgment consistency tests were performed, calculating consistency indices and consistency ratios. If a threshold was exceeded, feedback was provided to experts for correction until the test passed. Group judgment aggregation was performed, using the geometric mean method to aggregate pairwise comparison matrices from multiple experts. A consistency test was then performed again on the aggregated group matrix. The group consensus pairwise comparison matrices for each level were output.

[0096] S23, Dynamic weight learning and multi-source fusion: Based on the group consensus pairwise comparison matrix and historical evaluation data output in step S22, Bayesian network modeling technology and multi-source evidence fusion method are used to obtain a comprehensive initial weight configuration that takes into account expert knowledge, data patterns and causal relationships.

[0097] Specifically, a Bayesian network model of evaluation indicators is constructed. Each level of the evaluation indicator system is used as a node in the Bayesian network, and directed edges are constructed based on the hierarchical and causal relationships between the indicators. The network structure reflects the probabilistic dependencies between indicators: parent nodes represent influencing factors, and child nodes represent the influenced evaluation results. For each node, its conditional probability distribution is defined, representing the probability of each state of that node given the state of its parent node. The top-level node of the network is the comprehensive evaluation score, the bottom-level nodes are the various tertiary indicators, and the intermediate-level nodes are the various secondary indicators and the primary dimension.

[0098] Perform network parameter initialization based on expert knowledge. Using the group consensus pairwise comparison matrix output in step S22, calculate the initial weights of each indicator using the analytic hierarchy process (AHP). Convert the initial weights into conditional probability parameters in a Bayesian network; the larger the weight of an indicator, the higher its probability of influencing the state of its parent node. Simultaneously, invite domain experts to label key causal dependencies, identify direct causal influences and indirect transmission effects between indicators, and encode these causal relationships into the network's edge structure and conditional probability constraints.

[0099] Perform network parameter learning based on historical data. Extract complete evaluation record samples from the historical evaluation database. Each record contains observed values ​​of indicators at each level and the final comprehensive evaluation score. Use historical data as training samples and employ the expectation-maximization algorithm or variational inference algorithm to learn the conditional probability parameters of the Bayesian network. The learning process automatically discovers the actual contribution patterns of each indicator to the comprehensive evaluation result, including direct contributions, indirect contributions, and interactive contributions. For missing values ​​in the data, utilize the probabilistic inference capability of the Bayesian network to estimate the probability of missing values, avoiding parameter learning bias caused by missing data.

[0100] Weight extraction based on causal inference is performed again. In the trained Bayesian network, a causal effect estimation method is used to calculate the causal contribution of each indicator to the overall evaluation score. Causal contribution differs from simple correlation; it reflects the true impact of indicator changes on the evaluation results. Through intervention analysis, the expected impact of changing a certain indicator value on the overall evaluation score is simulated, and the strength of the causal effect is used as the data-driven weight for that indicator. Causal inference methods can effectively identify and eliminate the influence of confounding factors, obtaining a more accurate assessment of indicator importance.

[0101] A Bayesian fusion of multi-source evidence is performed. Expert judgment weights, data-driven weights, and causal inference weights are treated as three independent evidence sources, and a Bayesian evidence fusion framework is used to integrate the weights. A confidence parameter is assigned to each evidence source, determined based on the quality and reliability of the source. Initially, expert judgments have higher confidence; as historical data accumulates, the confidence of data-driven and causal inferences gradually increases. The posterior weight distribution is calculated using Bayes' theorem. The posterior weights integrate information from multiple sources and quantify the uncertainty of weight estimation. The expected value is extracted from the posterior distribution as the initial weight, and the variance is extracted as the confidence index for the weights.

[0102] Perform consistency checks and sensitivity analyses on the weight configuration. Verify that the initial weights meet business logic constraints, including that the sum of weights at each level equals the unit value, weight values ​​are within the valid range, and the weight distribution conforms to the expected priority relationship. Perform sensitivity analysis to assess the impact of the weight configuration on the evaluation results and identify the key indicators that have the most significant impact on the evaluation results. For anomalies or indicators with excessively high sensitivity found during the verification, adjust them and record the adjustment log. Output the verified initial weight configuration of the multi-dimensional evaluation indicator system, along with the confidence index of the weights and a sensitivity analysis report.

[0103] S3. Based on a multi-dimensional evaluation index system with comprehensive initial weights, a multi-factor adaptive weight adjustment method is adopted to obtain personalized dynamic weight configuration.

[0104] Obtain the multidimensional evaluation index system and its comprehensive initial weight configuration output in step S2. Combine the organizational strategic goal parameters, job characteristic data, evaluation time series parameters and major event records, and adopt the strategic orientation weight adjustment, job differentiation weight configuration, time series sensitive weight adjustment and multi-factor composite weight calculation methods to obtain a personalized dynamic weight configuration that integrates strategic orientation, job characteristics and time series sensitivity.

[0105] S31, Calculation of organizational strategic orientation weight adjustment coefficients: Based on the current organizational strategic goal configuration parameters and strategic capability mapping matrix, the strategic goal analysis and capability element correlation analysis method is used to obtain the strategic orientation weight adjustment coefficients for each evaluation dimension.

[0106] Specifically, an organizational strategic goal configuration model is established, decomposing strategic goals into strategic type labels such as technology innovation-driven, scale expansion, efficiency optimization, talent reserve, and risk prevention, and configuring the activation status and intensity parameters for each type. A mapping matrix between strategic types and evaluation dimensions is constructed, defining the influence coefficient of each strategic type on the five evaluation dimensions. Strategic orientation weight adjustment coefficients are calculated to obtain the activated strategic types and intensity parameters, and the influence coefficients are obtained by querying the mapping matrix. The influence coefficients for the same dimension are weighted, summed, and normalized. Smoothing of the adjustment coefficients is performed, introducing a time smoothing mechanism to avoid drastic weight fluctuations, and the current adjustment coefficients are weighted and merged with previous coefficients. The smoothed strategic orientation weight adjustment coefficients for each evaluation dimension are output.

[0107] S32, Calculation of differentiated weight allocation coefficients for specific positions: Based on the job category, job level information and job competency model library of the evaluated personnel, the differentiated weight allocation coefficients for specific positions are obtained by using the job feature matching and competency requirement mapping method.

[0108] Specifically, a job competency model library is constructed, classifying jobs according to functional sequences (technical research, technical development, technical management, technical support, etc.) and levels (junior, intermediate, senior, expert, etc.), and defining core competency requirements for each job category. Job characteristic identification is performed on the evaluated personnel, retrieving job information from the dataset and mapping it to the job category matrix. The competency model is queried to obtain the competency requirement configuration, retrieving the importance weight definitions for the five evaluation dimensions for that job category. The job differentiation weight configuration coefficient is calculated, using the weights of each dimension in the competency requirement configuration as benchmark values ​​and standardizing them. The job differentiation weight configuration coefficient is then output.

[0109] Furthermore, a dynamic competency profile generation method based on job description text analysis can be adopted to improve the flexibility of job competency requirement identification. Specifically, the job description and job requirements text are obtained, competency keywords are extracted using natural language processing technology, semantically matched with a concept library of five evaluation dimensions, the frequency and importance of keywords in each dimension are statistically analyzed, competency requirement profiles are automatically generated, and job differentiation coefficients are calculated.

[0110] S33, Calculation of time-sensitive weight adjustment coefficient: Based on evaluation time parameters, data timestamp information and major event records, time decay modeling and event triggering enhancement methods are used to obtain weight adjustment coefficients that reflect time sensitivity.

[0111] Specifically, a data time decay model is designed. A time decay function is defined, which takes the interval between the data generation time and the current evaluation time as input and outputs the timeliness weight coefficient of the data. The time decay function adopts an exponential decay form, and the decay rate parameter is set according to the differences in data types. The decay rate of project outcome data is slower, while the decay rate of collaborative behavior data is faster. The value range of the timeliness weight coefficient is from the benchmark value to the unit value.

[0112] Calculate the time-series weighting factor for each dimension of data. For each indicator data point under each evaluation dimension, calculate its timeliness weight coefficient based on the data timestamp. Aggregate the timeliness weight coefficients of each indicator data within the same dimension using a weighted average method, with the weight being the initial weight percentage of each indicator. The aggregation result serves as the time-series weighting factor for that dimension.

[0113] Perform weight enhancement processing triggered by significant events. Query the significant event records of the evaluated personnel to identify key contribution events that occurred within the evaluation period. Significant event types include breakthroughs in core technologies, successful delivery of important projects, resolution of major issues, and acquisition of industry honors. Define the associated evaluation dimensions and weight enhancement magnitude for each event type. When a significant event is detected, enhance the time-series weighting factor of the associated dimensions; the enhancement magnitude is determined based on the importance of the event.

[0114] The time-sensitive weight adjustment coefficients are calculated comprehensively. The time-weighted factors and event-triggered enhancement results are merged to generate the time-sensitive weight adjustment coefficients for each evaluation dimension. The adjustment coefficients are then normalized to ensure their distribution is within a reasonable range. The time-sensitive weight adjustment coefficients are then output.

[0115] S34, Comprehensive dynamic weight configuration calculation; Based on the comprehensive initial weight output in step S2, the strategic orientation adjustment coefficient output in step S31, the job differentiation coefficient output in step S32, and the time-sensitive adjustment coefficient output in step S33, a multi-factor composite weight calculation method is adopted to obtain a personalized dynamic weight configuration for specific talents at a specific evaluation point.

[0116] Specifically, the influence intensity parameters for each adjustment factor are set. Based on the organization's evaluation strategy preferences, the relative influence intensity of strategic-oriented factors, job-differentiating factors, and time-sensitive factors is configured. The sum of the influence intensity parameters for the three factors is a unit value. Under the default configuration, job-differentiating factors have the highest influence intensity, followed by strategic-oriented factors, and time-sensitive factors have the lowest.

[0117] Perform a composite calculation of multi-factor adjustment coefficients. For each evaluation dimension, the strategic orientation adjustment coefficient, job differentiation coefficient, and time-series sensitivity adjustment coefficient are weighted and combined according to their respective influence intensity parameters. The weighted combination result serves as the comprehensive adjustment factor for that dimension.

[0118] Calculate the dynamic weight configuration. Multiply the initial weights by the adjustment factor to obtain the adjusted original weights. Normalize the adjusted original weights so that the sum of the dynamic weights of each dimension is a unit value. The normalized result is the dynamic weight of each evaluation dimension.

[0119] Dynamic weights at the dimension level are passed down to the indicator level. The dynamic weight of each second-level indicator is equal to the dynamic weight of its parent dimension multiplied by the local weight of that indicator within the parent dimension. The dynamic weight of each third-level indicator is equal to the dynamic weight of its parent second-level indicator multiplied by the local weight of that indicator within the parent indicator. Through this layer-by-layer transmission, a complete personalized dynamic weight configuration is generated, outputting the dynamic weight values ​​of each level of evaluation indicator.

[0120] S4. Based on the standardized multi-source fusion talent dataset and personalized dynamic weight configuration, a progressive hybrid evaluation calculation method is adopted to obtain the comprehensive evaluation score.

[0121] Obtain the standardized multi-source fusion talent dataset output from step S1 and the personalized dynamic weight configuration output from step S3. Combine historical evaluation data and system deployment stage parameters, and employ methods such as original indicator value extraction and standardized score mapping, analytic hierarchy process (AHP) structured evaluation calculation, meta-learning-based adaptive evaluation model construction, and adaptive fusion of AHP results and machine learning results to obtain the sub-item scores and weighted comprehensive evaluation scores for each evaluation dimension. Specifically, the steps include:

[0122] S41, Extraction of raw indicator values ​​and standardized scoring mapping: Based on the standardized multi-source fusion talent dataset and the definition of a multi-dimensional evaluation indicator system, the standardized scores of each third-level evaluation indicator are obtained by using indicator value extraction rules and scoring mapping functions.

[0123] Specifically, a rule library for extracting indicator values ​​is established. For each level-three evaluation indicator, rules for extracting its raw values ​​are defined, including data source fields, data filtering conditions, numerical aggregation methods, and time range limitations. For example, the extraction rule for the skill certification record quantity indicator is: from the skill certification data table, filter records with valid certification status, count the number of records, and the time range is the past three years. The extraction rule for the code submission quantity indicator is: from the code submission record table, filter the submission records of the evaluated personnel, accumulate the number of lines of code changes, and the time range is the past year.

[0124] The process involves extracting the raw values ​​of the performance indicators. It iterates through all tertiary indicators in the evaluation indicator system and extracts the corresponding raw data from the standardized talent dataset according to the rules defined in the indicator value extraction rule base. The extracted raw data is then subjected to the rule-defined aggregation calculations to obtain the raw value for each tertiary indicator. For indicators that fail to be extracted, the reason for the failure is recorded and marked as missing.

[0125] Establish a scoring mapping function library. For each tertiary evaluation indicator, select an appropriate mapping function type and configure function parameters based on the indicator's business characteristics. Mapping function types include linear mapping functions, logarithmic mapping functions, sigmoid mapping functions, and piecewise mapping functions. Linear mapping functions are suitable for indicators where the original value and evaluation significance have a linear relationship; logarithmic mapping functions are suitable for indicators with diminishing marginal effects; sigmoid mapping functions are suitable for indicators that are sensitive in the middle and saturated at both ends; piecewise mapping functions are suitable for indicators with clear level divisions. The parameters of the mapping functions include the benchmark value, standard deviation, upper limit value, and lower limit value, and the parameter values ​​are determined based on the historical data distribution characteristics and the experience of business experts.

[0126] Perform score mapping calculations. For each tertiary indicator with extracted raw values, call the corresponding mapping function to convert the raw values ​​into standardized scores. The standardized scores range from zero to the maximum score. For indicators with missing raw values, handle them according to the missing value handling strategy: if the indicator has a low weight, use the default median score to fill in the missing values; if the indicator has a high weight, mark the reliability of the evaluation result as limited. Output the standardized scores of each tertiary evaluation indicator.

[0127] S42, Structured evaluation calculation using the analytic hierarchy process (AHP); Based on the standardized scores of each of the three-level indicators output in step S41 and the personalized dynamic weight configuration output in step S3, the hierarchical weighted summation method is used to obtain the evaluation scores of each level based on the AHP.

[0128] Specifically, the aggregation calculation from the scores of the tertiary indicators to the secondary indicators is performed. For each secondary indicator, the standardized scores and dynamic weights of all its constituent tertiary indicators are obtained. The score of the secondary indicator is calculated by multiplying the standardized scores of each tertiary indicator by its dynamic weight and summing the results. The sum is then normalized by dividing the result by the sum of the weights to obtain the standardized score of the secondary indicator.

[0129] Perform aggregate calculations from secondary indicator scores to primary dimensions. For each primary evaluation dimension, obtain the standardized scores and dynamic weights of all its included secondary indicators. Calculate the standardized score for that primary dimension using the same weighted summation method as in the previous step.

[0130] Perform aggregate calculation from the first-level dimension scores to the overall score. Obtain the standardized scores and dynamic weights of the five first-level evaluation dimensions, and calculate the overall evaluation score using a weighted summation method. The overall evaluation score is the structured evaluation result based on the analytic hierarchy process (AHP).

[0131] The evaluation results are quality-labeled. Based on the number of missing indicators and their respective weights in the evaluation calculation process, a reliability index is calculated. The reliability index reflects the degree of data completeness of the evaluation results. The evaluation scores and reliability indices for each level are output using the analytic hierarchy process (AHP).

[0132] S43, Construction and Evolution of Adaptive Evaluation Model Based on Meta-learning: Based on the accumulation of historical evaluation data and system deployment stage parameters, a meta-learning framework and model adaptive evolution strategy are adopted to obtain an intelligent evaluation model with rapid adaptability and continuous learning ability.

[0133] Specifically, the historical data accumulation status is assessed and a task distribution is constructed. The number of valid evaluation records in the historical evaluation database is counted, and a data accumulation level index is calculated. The data accumulation level is compared with a preset model training threshold to determine the current deployment stage of the system. The deployment stage is divided into three categories: initial stage, transition stage, and mature stage. Simultaneously, based on characteristics such as the job category, job level, and evaluation time of the evaluation objects, the historical evaluation data is divided into multiple sub-tasks, each corresponding to an evaluation problem in a specific scenario. The construction of the task distribution provides a diverse set of training tasks for subsequent meta-learning.

[0134] Differentiated model building strategies are implemented based on the deployment phase. In the initial phase, historical data is insufficient to support effective model training; therefore, a meta-learning strategy based on few-shot learning is adopted. A meta-learning network model is constructed, comprising two core components: a feature extractor and a task adapter. The feature extractor learns general talent feature representations, while the task adapter learns the ability to quickly adapt to new tasks. Using a meta-training task set constructed from publicly available talent evaluation datasets and similar organization datasets, the meta-learning model is trained to learn how to quickly learn evaluation patterns from a small number of samples. In practical applications, only a small number of labeled samples from the organization are needed; the meta-learning model can quickly adjust its parameters through the task adapter to adapt to the organization's evaluation characteristics.

[0135] During the transition phase, a hybrid strategy combining transfer learning and continuous learning is employed. The model trained in the meta-learning phase is loaded as initial parameters; this model already possesses good feature representation capabilities and rapid adaptability. Continuous learning training utilizes historical evaluation data accumulated within the organization, employing elastic weight consolidation algorithms or progressive neural network architectures to retain the memory of old data patterns while learning new ones, thus avoiding catastrophic forgetting. During continuous learning, model parameters gradually evolve from general patterns to organization-specific patterns, while preserving cross-scenario generalization capabilities.

[0136] In the mature stage, with ample historical data accumulation, a multi-model fusion strategy based on ensemble learning is adopted. Multiple heterogeneous evaluation models are constructed, including a gradient boosting-based tree model, a deep neural network-based representation learning model, and a graph neural network-based relationship modeling model. Tree models excel at capturing non-linear interactions and piecewise decision rules between indicators; deep neural network models excel at learning complex representations of high-dimensional features and end-to-end mapping relationships; and graph neural network models excel at modeling collaborative relationship networks and knowledge propagation paths among personnel. Multimodal feature data, including structured indicator values, textual descriptions, and relationship network data, are extracted from historical evaluation data. Each heterogeneous model is trained separately, and the optimal model hyperparameter configuration is selected through cross-validation. The prediction results of multiple models are integrated using stacked ensemble or weighted fusion methods, with the fusion weights dynamically determined based on the performance of each model on the validation set.

[0137] Perform model quality assessment and interpretability analysis. Evaluate the performance of the trained model using the retained test dataset, calculating multi-dimensional evaluation metrics such as prediction accuracy, mean squared error, correlation coefficient, and ranking consistency. Use SHAP value analysis or attention weight visualization methods to explain the model's decision-making process and identify the key features that most significantly impact the evaluation results. Generate a model interpretability report, including global feature importance ranking, local prediction explanations, and decision path visualization. When the model evaluation metrics reach the preset usability threshold and interpretability meets the requirements, mark the model as usable; otherwise, analyze the reasons for insufficient performance, adjust the model structure or training strategy, and retrain.

[0138] Construct a model adaptive evolution mechanism. Save the trained model parameters to a model library, recording metadata such as training time, training data volume, evaluation metrics, and interpretability reports. The model library supports multi-version management and A / B testing, facilitating model rollback and performance tracking. Establish a model performance monitoring mechanism to continuously track the model's predictive accuracy and user feedback satisfaction in real-world applications. When a decline in model performance or a significant change in data distribution is detected, automatically trigger the model update process, using incremental learning methods to update model parameters with new data, achieving adaptive model evolution. Output usable intelligent evaluation models, quality assessment reports, and interpretability analysis reports.

[0139] Furthermore, since ensemble learning methods, while improving prediction accuracy, increase model complexity and computational overhead, an automated model optimization method based on neural architecture search can be employed. The goal is to automatically discover the optimal model architecture best suited for a specific evaluation scenario, reducing model complexity while maintaining high accuracy. Specifically, a search space for model architectures is defined, including the number of network layers, layer types, connection methods, and hyperparameter ranges. Reinforcement learning or evolutionary algorithms are used to explore candidate architectures in the search space, and the performance of each candidate architecture is evaluated using historical evaluation data. The search process simultaneously optimizes both prediction accuracy and model efficiency, seeking the Pareto optimal architecture configuration. After the search is complete, the architecture with the best overall performance is selected for full training, resulting in a customized evaluation model. This method can automatically adapt to the data characteristics and evaluation needs of different organizations, reducing the workload of manual parameter tuning.

[0140] S44, Adaptive fusion of analytic hierarchy process (AHP) results and machine learning results;

[0141] Based on the analytic hierarchy process (AHP) evaluation score output in step S42 and the machine learning model prediction score output in step S43, an adaptive weight fusion strategy is adopted to obtain a comprehensive evaluation score that combines the advantages of the two methods.

[0142] Specifically, the machine learning model is used for prediction calculations. If the machine learning model is available, the feature data of the person being evaluated is input into the model, and the predicted evaluation score output by the model is obtained. The prediction results include the individual predicted scores for each dimension and the overall predicted score. If the model is unavailable, this step is skipped.

[0143] Determine the fusion weights for the two methods. These weights are adaptively adjusted based on the system deployment phase. In the initial phase, the analytic hierarchy process (AHP) weight is a single value, and the machine learning weight is zero. In the transition phase, the AHP weight gradually decreases from a higher value, while the machine learning weight gradually increases. In the mature phase, the weights of the two methods tend to balance, with the specific ratio dynamically determined based on model evaluation metrics. The better the model evaluation metrics, the higher the proportion of machine learning weight.

[0144] Perform the fusion evaluation score calculation. The overall evaluation score is calculated by multiplying the analytic hierarchy process (AHP) score by its fusion weight, and then adding the machine learning prediction score multiplied by its fusion weight. The same fusion calculation method is used for the scores of each dimension. The fused score is the final evaluation output.

[0145] Perform a consistency check on the evaluation results. Calculate the degree of deviation between the analytic hierarchy process (AHP) score and the machine learning prediction score. When the deviation exceeds a preset alarm threshold, trigger the evaluation result review mechanism, mark the evaluation record as pending manual review, and record the deviation details for review reference.

[0146] Output the fusion evaluation results. The output includes the standardized scores for each tertiary indicator, the aggregated scores for each secondary indicator, the summary scores for each primary dimension, the fused comprehensive evaluation score, the evaluation credibility index, and the review status indicator. The fusion evaluation results also retain details of the analytic hierarchy process (AHP) scores and machine learning prediction scores, supporting subsequent analysis and comparison.

[0147] S5, obtain the comprehensive evaluation score, and use multi-level time series evolution analysis and development potential prediction method to obtain the comprehensive value assessment result of the talent;

[0148] Obtain the current comprehensive evaluation score and historical multi-period evaluation score series output from step S4. Employ growth pattern recognition based on multi-scale time series decomposition, capability inflection point detection, multi-level development potential prediction, and multi-dimensional value assessment model calculation methods to obtain a comprehensive talent value assessment result including current value score, potential value score, and attached confidence intervals. Specifically, this includes the following steps:

[0149] S51, Construction and updating of talent capability time series database; Based on the current evaluation score and evaluation time information, the talent capability time series database is constructed and updated using time series data management technology to obtain historical evaluation data sequences that support time series analysis.

[0150] Specifically, a data model for a time-series database of talent capabilities is designed. The data model comprises two core table structures: a master evaluation record table and a detailed evaluation score table. The master evaluation record table stores metadata for each evaluation, including fields such as a unique identifier for the evaluation record, the identifier of the person being evaluated, the evaluation time, the evaluation period type, and the overall evaluation score. The detailed evaluation score table stores evaluation score details for each level of indicator, including fields such as the evaluation record association identifier, indicator identifier, indicator level, indicator score, and data quality identifier. The two tables are linked through the evaluation record identifier.

[0151] Perform the database entry process for the current evaluation results. Transform the evaluation results output in step S4 according to the data model format. Generate record items in the main evaluation record table, including the current time point and the overall evaluation score. Iterate through the evaluation scores of each level of indicators to generate multiple record items in the detailed evaluation score table. Perform a database insertion operation to persistently store the transformed data in the time-series database.

[0152] Establish a time-series index to support efficient time-series queries. Create a composite index based on personnel identification and evaluation time on the main evaluation record table. This index supports fast retrieval of historical evaluation record sequences for specific personnel by time range. Simultaneously, establish an auxiliary index based on evaluation period type to support data aggregation analysis at different evaluation period granularities.

[0153] Perform lifecycle management of historical data. Based on a pre-defined data retention strategy, archive historical evaluation data that has exceeded its retention period. Migrate archived data to a cold storage area, retaining data summary information to support long-term trend analysis; detailed data is loaded on demand. Output an updated talent capability time-series database to support subsequent time-series analysis and processing.

[0154] S52, Growth pattern identification and trend prediction based on multi-scale time series decomposition; Based on the talent capability time series database output in step S51, multi-scale time series decomposition technology and pattern recognition algorithm are used to obtain the deep growth patterns and multi-level trend analysis results of each evaluation dimension.

[0155] Specifically, historical evaluation sequences of the evaluated personnel are extracted from the time-series database. Based on a pre-defined analysis time span, all historical evaluation records for that personnel within the analysis time range are queried. These records are then sorted in ascending order by evaluation time to form a time-series evaluation data sequence. For each primary evaluation dimension, the corresponding dimension score sequence is extracted. Simultaneously, contextual information related to the evaluation time point is extracted, including event data such as project participation, training experience, job changes, and organizational environment changes within the evaluation period.

[0156] A multi-scale time series decomposition based on empirical mode decomposition (EMD) is performed. An ensemble EMD algorithm is used to decompose the capability score time series into multiple intrinsic mode functions (IMFs) and a residual trend term. Each IMF represents a fluctuation component at a different time scale: high-frequency IMFs reflect short-term random fluctuations and noise interference, mid-frequency IMFs reflect periodic capability fluctuation patterns, and low-frequency IMFs reflect medium- to long-term capability development trends. The residual trend term reflects the long-term fundamental trend of capability development. Through multi-scale decomposition, the complex time series signal is separated into multiple components with clear physical meaning, facilitating the analysis of variation patterns at different scales.

[0157] Perform growth pattern identification based on dynamic time warping. Construct a talent growth pattern library containing various typical ability development patterns, such as linear growth, exponential growth, S-shaped growth, fluctuating upward growth, plateau, and decline. Each pattern is represented by a standardized time-series template. Calculate the similarity distance between the evaluated individual's ability development sequence and each standard pattern template using the dynamic time warping algorithm. The dynamic time warping algorithm can handle nonlinear scaling and local shifts in the time-series sequence, achieving more robust pattern matching. Identify the growth pattern type most similar to the current individual's development sequence and calculate the confidence score of the pattern match. For cases with low matching confidence, indicating that the individual's development pattern is unique, a clustering algorithm is used to classify it as a new growth pattern type, enriching the diversity of the pattern library.

[0158] Trend estimation and change point detection based on the state-space model are performed again. A state-space model of talent capability development is constructed, with capability level, capability change rate, and capability acceleration as latent state variables, and observed evaluation scores as observed variables. Kalman filtering or particle filtering algorithms are used for state estimation, recursively calculating the optimal state estimate and estimation uncertainty at each time point. The filtered state sequence eliminates the influence of observation noise and random fluctuations, providing a smoother and more reliable trend curve. Based on the state estimation, a Bayesian change point detection algorithm is used to identify significant change points in the capability development process. The change point detection algorithm calculates the posterior probability of each time point as a change point; when the posterior probability exceeds a threshold, it is marked as a significant change point. The location of the change point corresponds to the transition moment of the capability development mode, such as from a plateau period to a rapid growth period or from a growth period to a decline period.

[0159] Perform trend attribution explanation based on causal analysis. Combine extracted contextual event data to analyze the causal relationship between capability trend changes and external events. Use the Granger causality test to examine whether the occurrence of a specific event has a predictive effect on capability changes. For detected significant change points, search for relevant events within the time window before and after the change point, and calculate the temporal proximity and causal association strength between the events and the change point. The event with the strongest causal association strength is taken as the likely attribution of the change point, providing a business explanation for the trend change. Simultaneously, use counterfactual inference to estimate the expected change in the capability development trajectory if a certain event had not occurred, quantifying the actual impact of the event on capability development.

[0160] Perform a multi-level trend state comprehensive assessment. Integrate multi-scale decomposition results, growth pattern identification results, state estimation results, and causal attribution analysis results to generate a multi-level trend analysis report. The report includes three levels: short-term trend assessment, medium-term trend assessment, and long-term trend assessment. Short-term trends are based on high-frequency and mid-frequency mode function analysis, reflecting recent capability fluctuations; medium-term trends are based on low-frequency mode functions and state change rate analysis, reflecting the current capability development stage; long-term trends are based on residual trend terms and growth pattern identification results, reflecting the overall capability development trajectory. For each evaluation dimension, output its multi-scale decomposition components, matched growth pattern type, state estimation sequence, list of identified change points, attribution event analysis, and multi-level trend state labels.

[0161] Furthermore, since traditional time series analysis methods primarily focus on the evolution of univariate time series, they may overlook the co-evolutionary patterns among multiple evaluation dimensions. A multidimensional co-evolutionary trend analysis method based on tensor decomposition can be employed. The aim is to simultaneously model the time series evolution of multiple evaluation dimensions and their interrelationships, discovering co-evolutionary change patterns among dimensions. Specifically, the time series data of multiple evaluation dimensions are organized into a third-order tensor structure, with the three dimensions being personnel, evaluation dimensions, and time. The tensor is decomposed into a combination of a core tensor and a factor matrix using Tucker decomposition or CP decomposition algorithms. The factor matrix captures the latent representations of personnel characteristics, dimensional characteristics, and time characteristics, while the core tensor describes the interaction relationships among these three characteristics. By analyzing the evolutionary patterns of the factor matrix, co-evolutionary change trends among multiple evaluation dimensions are identified, such as the synchronous rise or inverse substitution relationships of certain dimensions. This method can provide a more comprehensive analysis of capability development trends.

[0162] S53, Identification of Inflection Points in Capabilities: Based on the growth trend analysis results output in step S52, the change point detection algorithm is used to identify key turning points in the talent capability development process, and the capability inflection point identification results are obtained.

[0163] Specifically, the time series of evaluation scores is preprocessed. First-order difference sequences between adjacent time points are calculated; these difference sequences reflect the rate of change in capability. The difference sequences are then standardized to eliminate the influence of differences in dimensions and scales.

[0164] Significant changes are identified using the cumulative sum change point detection method. The cumulative sum statistic is calculated for the difference series, starting from zero and sequentially summing the standardized differences at each time point. A cumulative sum curve is plotted; points where the slope of the curve changes are potential inflection points. A significance threshold is set; when the change in the slope of the cumulative sum curve exceeds the threshold, that time point is marked as a significant change point.

[0165] Perform a change point classification. For each identified significant change point, analyze the changes in trend characteristics before and after the change point. Based on the changes in the direction and intensity of the trend before and after the change point, classify the change points into types such as capability breakthrough points, capability bottleneck points, capability recovery points, and capability decline points. A capability breakthrough point is characterized by a sudden shift from a low-growth or stable state to a high-growth state; a capability bottleneck point is characterized by a sudden shift from a high-growth state to a low-growth or stable state; a capability recovery point is characterized by a shift from a declining state to a stable or rising state; and a capability decline point is characterized by a shift from a stable or rising state to a declining state.

[0166] Perform business attribution analysis on the turning points. Based on the time of the turning point, query relevant event records of the evaluated personnel within that time period, including job changes, project participation, and training experience. Attempt to identify business events associated with the turning points to provide explanatory references for their causes. Output a list of capability inflection points across various dimensions, including the inflection point time, inflection point type, magnitude of change, and possible attributable events.

[0167] S54, Multi-level development potential prediction: Based on the growth trend analysis results output in step S52 and the capability inflection point information output in step S53, a multi-level prediction strategy is adopted to predict talent development potential, and the development potential prediction results with confidence intervals are obtained.

[0168] Specifically, short-term development forecasts are implemented. These forecasts employ a time-series extrapolation method, with the timeframe set to one to two future evaluation periods. Based on the evaluation score series from the most recent periods, an exponential smoothing model is used for forecast calculations. The smoothing coefficient of the exponential smoothing model is adaptively selected according to the volatility characteristics of the series. Short-term forecasts focus on continuous changes in capabilities and are suitable for predicting the continuation of recent performance.

[0169] A mid-term development forecast is executed. The mid-term forecast employs a combination of trend extrapolation and pattern matching, with a timeframe set for the next three to five evaluation periods. Based on the comprehensive growth rate obtained from growth trend analysis, a baseline forecast trajectory is generated by linear extrapolation according to the current trend. Simultaneously, historical cases similar to the current personnel development pattern are retrieved from the historical database, and their subsequent development is analyzed to generate case reference forecast trajectories. The mid-term development forecast result is generated by combining the baseline forecast and the case reference forecasts.

[0170] Execute long-term development forecasts. Long-term forecasts employ scenario analysis, with a timeframe set at five to ten evaluation periods. Three development assumptions are defined: optimistic, neutral, and conservative. The optimistic scenario assumes continued growth at historically best rates; the neutral scenario assumes stable growth at average rates; and the conservative scenario assumes a gradual slowdown in growth rates to the industry average. Development trajectory predictions are calculated for each of the three scenarios to form a range of possible long-term development outcomes.

[0171] Calculate the confidence intervals for the forecast results. For short-term and medium-term forecasts, estimate the standard error of the forecast based on the distribution characteristics of historical forecast errors. Construct the confidence interval for the forecast with the predicted value as the center and a radius equal to a multiple of the standard error. The further away from the current forecast time, the wider the confidence interval, reflecting the increase in forecast uncertainty. Output the short-term, medium-term, and long-term scenario forecasts for each dimension, along with their corresponding confidence intervals.

[0172] Furthermore, since traditional time-series forecasting methods struggle to capture the non-linear, abrupt changes in talent development, a deep time-series forecasting method based on recurrent neural networks can be employed. The aim is to improve the predictive ability for complex development patterns. Specifically, a time-series forecasting model based on long short-term memory networks is constructed, capable of learning long-term dependencies in time-series data. Historical evaluation score sequences and related feature data are used as model inputs, with evaluation scores for several future periods serving as the prediction target. The model is trained using historical data from all personnel within the organization to learn general talent development patterns. For specific individuals being evaluated, their individual historical data is used for model fine-tuning to generate personalized development trajectory predictions.

[0173] S55, Calculation of Comprehensive Talent Value Assessment;

[0174] Based on the current evaluation score output in step S4 and the development potential prediction results output in step S54, a multi-dimensional value assessment model is adopted to obtain a comprehensive talent value assessment result that reflects current capabilities and future potential.

[0175] Specifically, a dimensional framework for comprehensive talent value assessment is defined. Value assessment comprises two core dimensions: current value and potential value. Current value reflects an individual's immediate contribution to the organization based on their existing capabilities; potential value reflects an individual's future contribution to the organization based on development expectations.

[0176] Calculate the current value score. The current value score is directly derived from the comprehensive evaluation score output in step S4. This score is calculated based on the actual performance of the multi-dimensional evaluation indicators and reflects the talent's current comprehensive ability level.

[0177] Calculate the potential value score. The potential value score comprehensively considers growth trend and development potential prediction. Weights are assigned to the growth trend factor and the development prediction factor. The growth trend factor is calculated based on the overall growth rate and trend status; a sustained upward trend receives a higher factor value, a stable trend receives a medium factor value, and a downward trend receives a lower factor value. The development prediction factor is calculated based on the expected increase in the medium term; the greater the expected increase, the higher the factor value. The potential value score equals the growth trend factor multiplied by its weight plus the development prediction factor multiplied by its weight, and the result is standardized.

[0178] Calculate the overall talent value score. Set relative weights for current value and potential value; these weights can be adjusted based on the organization's talent management strategy. The overall value score equals the current value score multiplied by its weight, plus the potential value score multiplied by its weight.

[0179] Generate a comprehensive talent value assessment report. The report includes the current value score, detailed current scores for each dimension, potential value score, a summary of growth trend analysis, a summary of development potential prediction, the overall value score, and an overall value level label. The overall value level is determined by setting a threshold based on the score distribution, categorizing talent into levels such as Excellent, Good, Satisfactory, and Needs Improvement. Output the complete comprehensive talent value assessment results.

[0180] S6: Obtain the comprehensive evaluation score and the comprehensive talent value assessment results, and use intelligent visualization analysis and multi-dimensional decision support methods to obtain complete talent evaluation output results;

[0181] Obtain the comprehensive evaluation score output in step S4 and the comprehensive talent value assessment result output in step S5. Employ methods such as multi-dimensional visualization analysis chart generation, intelligent talent profile construction based on multimodal generation, real-time monitoring and early warning of talent pipeline health, and intelligent decision support suggestion generation to obtain a complete talent evaluation output including visualization analysis charts, personalized talent profile reports, talent pipeline health monitoring reports, and decision support suggestions. Specifically, this includes the following steps:

[0182] S61, Generation of multi-dimensional visualization analysis charts; Based on evaluation score data and time series analysis results, data visualization technology is used to generate various types of analysis charts, resulting in visual analysis results that intuitively present the talent's ability status.

[0183] Specifically, a multi-dimensional competency radar chart is generated for each individual. Five primary evaluation dimensions serve as the five axes of the radar chart, with the standardized scores of each dimension acting as coordinate values ​​along those axes. Connecting the five coordinate points forms a closed polygon; the shape and area of ​​the polygon visually reflect the distribution characteristics and overall level of an individual's competencies. Within the same radar chart, competency polygons from the same historical period can be overlaid to create a comparative view of competency changes. Simultaneously, standard competency polygons required for the job can be overlaid to visually demonstrate the difference in match between an individual's competencies and job requirements.

[0184] Generate a team capability heatmap. Construct a capability score matrix with team members as the row dimension and evaluation metrics as the column dimension. Map the scores to color shades, displaying high scores in darker colors and low scores in lighter colors. The heatmap provides an overall view of the distribution of team members' capabilities across various metrics. The color distribution pattern allows for quick identification of the team's strengths and weaknesses, as well as individual members with outstanding strengths or significant shortcomings.

[0185] Generate a personal growth trend comparison chart. Using time as the horizontal axis and evaluation scores as the vertical axis, plot trend curves showing the changes in scores for each evaluation dimension over time. Different dimensions are distinguished by different colors. Mark the identified ability inflection points on the chart and add labels explaining the type of inflection point. The trend comparison chart visually displays the development trajectory and key change nodes of an individual's abilities in each dimension.

[0186] Generate a development potential prediction trajectory map. Plot a development potential prediction curve with the future time period as the horizontal axis and the predicted score as the vertical axis. The confidence interval range of the prediction is indicated by shaded areas around the curve. For long-term predictions, the prediction trajectories for optimistic, neutral, and conservative scenarios are displayed as bifurcated curves. The potential prediction map helps managers intuitively understand the future development possibilities and uncertainties of talent.

[0187] S62, intelligent talent profile construction and narrative presentation based on multimodal generation; based on evaluation data and analysis results, it automatically generates personalized talent profile reports using multimodal content generation technology and narrative expression methods, resulting in a comprehensive talent profile document that combines data insights and readability.

[0188] Specifically, a multi-layered talent profile knowledge graph is constructed. Talent evaluation data, historical trajectories, ability characteristics, development patterns, and contextual information are organized into a graph structure. Graph nodes include talent entity nodes, ability attribute nodes, event nodes, project nodes, and evaluation result nodes. Graph edges represent the relationships between nodes, including ownership, participation, influence, and evolutionary relationships. Each edge is assigned a timestamp and relationship strength attributes, supporting the expression of temporal relationships. Through the structured representation of the knowledge graph, scattered evaluation data is integrated into a semantically related knowledge network, providing a unified knowledge foundation for subsequent intelligent analysis and content generation.

[0189] Deep mining of competency features based on graph neural networks is performed. On the constructed talent profile knowledge graph, a graph neural network model is used for node representation learning and relational reasoning. The graph neural network aggregates neighborhood information of nodes through a message passing mechanism, learning low-dimensional embedding representations of nodes. Node embedding vectors fuse node attributes and graph structure context information, capturing implicit competency features and association patterns. In the learned node embedding space, a clustering algorithm is used to identify talent groups with similar competency features, locating the evaluated personnel to the corresponding competency groups. By comparing and analyzing the differences between the evaluated personnel and the typical features of the groups, personalized competency tags are automatically extracted. Competency tags include not only explicit high-score and low-score dimensions, but also potential strengths and hidden weaknesses discovered through graph reasoning.

[0190] This system implements intelligent positioning of career development stages based on contrastive learning. A multi-dimensional feature space for career development stages is constructed, with dimensions including years of work experience, job level, ability level, growth rate, project experience complexity, and knowledge breadth. A contrastive learning method is used to train a development stage classification model. The model learns to map individuals at the same development stage to neighboring regions in the feature space, and individuals at different development stages to distant regions. Within the trained feature space, the distance between the evaluated individual and the prototypes of each development stage is calculated, positioning them to the nearest development stage category. Development stages are categorized as novice, growth, maturity, expert, and authority. Simultaneously, the maturity score of the evaluated individual at the current stage is calculated. This maturity score reflects their development progress at the current stage, providing a basis for predicting stage transitions.

[0191] Personalized narrative generation based on a generative pre-trained model is executed again. A generative pre-trained language model, fine-tuned with data from the talent evaluation domain, automatically generates narrative text content for talent profiles. A structured prompt template is constructed, containing key information such as basic talent information, capability assessment data, growth trend analysis, development stage positioning, and capability tags. After filling the prompt template with specific data, it is input into the generation model, which outputs fluent and natural narrative text, including descriptions of capability characteristics, summaries of growth trajectories, explanations of strengths and highlights, and evaluations of development potential. The generated narrative text is not a simple data listing, but rather transforms data into a logically coherent and readable story-like expression, facilitating managers' quick understanding of the overall talent profile.

[0192] Implement personalized development path planning based on reinforcement learning. Construct a Markov decision process model for talent development, where the state space represents the talent's ability state, the action space represents available development actions (such as participating in training, undertaking projects, job rotation, etc.), and the reward function represents the improvement in ability and the achievement of career goals. Employ reinforcement learning algorithms to learn the optimal development strategy, which recommends the most beneficial sequence of actions based on the talent's current ability state. During training, historical talent development data is used as experience replay to learn the long-term effects of different development paths. For the evaluated individual, based on their current ability state and development goals, the learned optimal strategy is applied to generate a personalized development path plan. The development path plan includes short-term action suggestions, mid-term development milestones, and long-term career goals, with each suggestion accompanied by an expected effect assessment and priority ranking.

[0193] Perform multimodal content integration and interactive presentation. Integrate narrative text, visual charts, development path planning, and interactive analysis components to generate multimodal talent profile reports. The reports feature a responsive layout design, supporting adaptive display on different devices. Embedded interactive components allow users to click to view detailed data, adjust analysis parameters, and explore hypothetical scenarios. For example, users can adjust action choices in the development path, and the system predicts the adjusted capability development trajectory in real time. The reports support multiple export formats, including PDF documents, web page formats, and data interfaces. Output personalized intelligent talent profile reports, including narrative content, visual analysis, development path planning, and interactive exploration functions.

[0194] S63, Real-time monitoring and early warning of talent pipeline health; Based on the aggregation of talent evaluation data at the team level, a pipeline health index system and threshold early warning mechanism are adopted to obtain the monitoring and risk warning results of the talent pipeline health status.

[0195] Specifically, a talent pipeline health indicator system is designed. The health indicators include four categories: pipeline structure indicators, capability distribution indicators, development dynamic indicators, and risk warning indicators. Pipeline structure indicators assess the staffing ratio at each job level and the depth of reserves for core positions; capability distribution indicators assess the overall capability level of the team and the coverage of capability gaps; development dynamic indicators assess the team's growth rate and talent turnover rate; and risk warning indicators identify the risks of key talent loss and capability gaps.

[0196] Perform team health index calculations. Aggregate evaluation data for all team members from the talent evaluation database. Calculate and statistically analyze the current values ​​of each health index according to its calculation rules. The calculation rules are predefined, including the data selection scope, aggregation method, and result standardization method.

[0197] Perform threshold-based early warning determination. Set early warning and alarm thresholds for each health indicator. A yellow early warning message is generated when the early warning threshold is triggered, indicating that the indicator needs attention; a red alarm message is generated when the alarm threshold is triggered, indicating that the indicator needs immediate action. Iterate through the current values ​​of all indicators and compare them with the corresponding thresholds to identify the indicator that triggered the early warning or alarm.

[0198] Generate and push early warning information. For indicators that trigger early warnings or alarms, generate early warning information, including the indicator name, current value, threshold, degree of deviation, possible cause analysis, and suggested precautions. Sort the early warning information according to urgency and push it to relevant management personnel through system messages, emails, or instant messaging.

[0199] Generate a talent pipeline health dashboard view. Visualize each health indicator in a dashboard format, with indicator status distinguished by color as normal, warning, and alert. The dashboard supports drill-down viewing of detailed indicator data and trend changes. Output a talent pipeline health monitoring report and a list of warning messages.

[0200] S64, Intelligent Decision Support Suggestion Generation: Based on talent evaluation results and talent pipeline analysis data, a decision support algorithm is used to generate team configuration optimization suggestions and training needs identification results, resulting in intelligent suggestion output to assist management decisions.

[0201] Specifically, the team configuration optimization analysis is performed. Project requirement information for the team to be formed is obtained, including project type, technical field, difficulty level, and timeline. A matching rule matrix between project requirements and talent capabilities is defined, describing the weight of each capability dimension for different project types. Based on the matching rule matrix, a matching score is calculated for each candidate. The matching score is equal to the sum of the weighted products of the candidate's capability scores for each dimension and their corresponding requirement weights.

[0202] Perform team composition optimization calculations. Under the constraint of team size, find the personnel combination scheme that achieves the best overall team fit and optimal skill complementarity. Skill complementarity assessment evaluates the coverage of skill gaps among team members to avoid common skill gaps within the team. Several candidate team composition schemes are generated using a greedy algorithm or heuristic search algorithm.

[0203] Perform training needs identification and analysis. Summarize the competency evaluation data of each team member to identify common competency gaps. Common competency gaps refer to situations where a significant proportion of members score low in that competency dimension. For the identified common competency gaps, match corresponding training courses or learning resource entries. Simultaneously, for significant individual competency gaps, generate personalized training recommendation lists.

[0204] The report integrates and outputs decision support recommendations. It includes team configuration optimization plans, plan evaluation explanations, training needs identification results, and a list of recommended training resources. For team configuration plans, multiple candidate plans are provided for managers to choose from. For training needs, common team needs and individual specific needs are differentiated and presented in order of priority. A complete intelligent decision support recommendations report is output.

[0205] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multidimensional talent evaluation method, characterized in that, Includes the following steps: S1. Collect raw talent data and use a deep semantic fusion method based on domain ontology to obtain a standardized multi-source fusion talent dataset; S2, obtain a standardized multi-source fusion talent dataset, and use a five-dimensional evaluation framework design and dynamic weight learning method to obtain a multi-dimensional evaluation index system with comprehensive initial weights; S3. Based on a multi-dimensional evaluation index system with comprehensive initial weights, a multi-factor adaptive weight adjustment method is adopted to obtain personalized dynamic weight configuration. S4. Based on the standardized multi-source fusion talent dataset and personalized dynamic weight configuration, a progressive hybrid evaluation calculation method is adopted to obtain the comprehensive evaluation score. S5, obtain the comprehensive evaluation score, and use multi-level time series evolution analysis and development potential prediction method to obtain the comprehensive value assessment result of the talent; S6 obtains the comprehensive evaluation score and the comprehensive talent value assessment results, and uses intelligent visualization analysis and multi-dimensional decision support methods to obtain complete talent evaluation output results.

2. The multidimensional talent evaluation method according to claim 1, characterized in that, The deep semantic fusion method using domain ontology includes: A graph structure is used to represent the core concept system and semantic relationship network in the field of talent evaluation. The graph nodes include personnel entities, organization entities, project entities, ability entities, contribution entities, and time entities, and the graph edges represent the semantic relationships between concepts. Based on a pre-trained language model, metadata information of data fields and ontology concept nodes are encoded into semantic vector representations respectively. The mapping relationship between data fields and ontology concepts is discovered by calculating the semantic vector similarity. When semantically ambiguous data of the same person is detected in different source systems, a multi-strategy joint disambiguation method is adopted, which combines a graph path-based inference method, a context consistency-based verification method, and a constraint satisfaction-based inference method.

3. The multidimensional talent evaluation method according to claim 1, characterized in that, The five-dimensional evaluation framework design and dynamic weight learning method include: The indicators at each level in the evaluation index system are used as nodes in a Bayesian network, and directed edges are constructed based on the hierarchical relationship and causal dependency between the indicators. The initial weights of each index are calculated using the group consensus pairwise comparison matrix and the analytic hierarchy process (AHP), and the initial weights are converted into conditional probability parameters in a Bayesian network. Historical evaluation data is used as training samples, and the conditional probability parameters of the Bayesian network are learned using the expectation-maximization algorithm. The causal contribution of each indicator to the comprehensive evaluation score is calculated using the causal effect estimation method, and the data-driven weights are obtained through intervention analysis. The expert judgment weight, data-driven weight, and causal inference weight are used as independent sources of evidence. The posterior weight distribution is calculated using Bayes' theorem, and the expected value is extracted from the posterior distribution as the comprehensive initial weight.

4. The multidimensional talent evaluation method according to claim 1, characterized in that, The multi-factor adaptive weight adjustment method includes: Based on the organizational strategic goal configuration parameters and strategic capability mapping matrix, calculate the strategic orientation weight adjustment coefficient for each evaluation dimension; Based on the job category and rank information of the evaluated personnel, the job competency model library is queried to obtain the job differentiation weight configuration coefficient; Based on the data timestamp information, the timeliness weight coefficient is calculated using the time decay function. Combined with the records of major events, the weight of the associated dimensions is enhanced to obtain the time-sensitive weight adjustment coefficient. The strategic orientation adjustment coefficient, job differentiation coefficient, and time-sensitive adjustment coefficient are weighted and combined according to the influence intensity parameter, multiplied with the overall initial weight, and then normalized.

5. The multidimensional talent evaluation method according to claim 1, characterized in that, The progressive hybrid evaluation calculation method includes: The number of valid evaluation records in the historical evaluation database is counted, and the data accumulation level is compared with the preset model training threshold to determine the current deployment stage of the system. The deployment stage is divided into the initial stage, the transition stage, and the mature stage. In the initial stage, a meta-learning strategy based on few-shot learning is adopted to construct a meta-learning network model that includes a feature extractor and a task adapter. During the transition phase, the model trained in the meta-learning phase is loaded as the initialization parameters, and the elastic weight consolidation algorithm is used to retain the memory of the old data patterns while learning the new data patterns. In the mature stage, a gradient boosting-based tree model, a deep neural network-based representation learning model, and a graph neural network-based relation modeling model are constructed, and the prediction results of multiple models are integrated using stacked ensemble or weighted fusion methods.

6. The multidimensional talent evaluation method according to claim 1, characterized in that, The progressive hybrid evaluation calculation method also includes: Based on a standardized multi-source fusion talent dataset and personalized dynamic weight configuration, a hierarchical weighted summation method is used to aggregate and calculate step by step to obtain an evaluation score based on the analytic hierarchy process. If the machine learning model is available, the feature data of the person being evaluated is input into the model to obtain the predicted score of the machine learning model. The fusion weights of the two methods are adaptively adjusted according to the system deployment phase. In the initial stage, the weight of the analytic hierarchy process (AHP) is a unit value. In the transition stage, the weight of the AHP gradually decreases while the weight of the machine learning method gradually increases. In the mature stage, the weights of the two methods are dynamically determined based on the model evaluation indicators. The analytic hierarchy process (AHP) evaluation score and the machine learning model prediction score are weighted and summed according to the fusion weights. When the deviation between the two scores exceeds a preset threshold, the evaluation result review mechanism is triggered.

7. The multidimensional talent evaluation method according to claim 1, characterized in that, The method employing multi-level temporal evolution analysis and development potential prediction includes: The historical evaluation sequence of the evaluated personnel is extracted from the talent capability time series database, and the capability score time series is decomposed into multiple intrinsic mode functions and a residual trend term using the ensemble empirical mode decomposition algorithm. A talent development model library containing multiple typical ability development models is constructed. The dynamic time warping algorithm is used to calculate the similarity distance between the ability development sequence of the evaluated person and each standard model template, and to identify the matching development model type. A state-space model for talent capability development is constructed, with capability level, capability change rate, and capability acceleration as latent state variables. The Kalman filter algorithm is used for state estimation, and the Bayesian change point detection algorithm is used to identify significant change points in the capability development process. By combining the results of multi-scale decomposition, growth pattern identification, and state estimation, a multi-level trend analysis result including short-term, medium-term, and long-term trends is generated.

8. The multidimensional talent evaluation method according to claim 1, characterized in that, The results of the comprehensive talent value assessment include: The exponential smoothing model is used for short-term development forecasting, the combination of trend extrapolation and pattern matching is used for medium-term development forecasting, and scenario analysis is used for long-term development forecasting. The standard error of the prediction is estimated based on the distribution characteristics of historical prediction errors, and the confidence interval of the prediction is constructed. The comprehensive evaluation score is used as the current value score. The potential value score is calculated based on the growth trend factor and the development prediction factor. The current value score and the potential value score are then summed according to their respective weights.

9. The multidimensional talent evaluation method according to claim 1, characterized in that, The method employing intelligent visualization analysis and multidimensional decision support includes: Design a talent echelon health index system that includes echelon structure indicators, capability distribution indicators, development dynamic indicators, and risk warning indicators, and set warning thresholds and alarm thresholds for each health indicator; The system aggregates evaluation data from all team members to calculate the current values ​​of each health indicator. When an indicator value triggers a warning threshold or alarm threshold, a warning message is generated. The matching score of candidates is calculated based on the matching rule matrix between project needs and talent capabilities. A heuristic search algorithm is used to generate team configuration optimization plan, identify common capability shortcomings of the team and match training resources.

10. A multidimensional talent evaluation system, used to perform the steps of a multidimensional talent evaluation method as described in any one of claims 1-9, characterized in that, include: The data fusion module is used to collect raw talent data and use a deep semantic fusion method based on domain ontology to obtain a standardized multi-source fused talent dataset. The indicator construction module is used to obtain a standardized multi-source fusion talent dataset. It adopts a five-dimensional evaluation framework design and a dynamic weight learning method to obtain a multi-dimensional evaluation indicator system with comprehensive initial weights. The weight adjustment module is used to obtain personalized dynamic weight configuration based on a multi-dimensional evaluation index system with comprehensive initial weights and a multi-factor adaptive weight adjustment method. The evaluation calculation module is used to obtain a comprehensive evaluation score based on a standardized multi-source fusion talent dataset and personalized dynamic weight configuration, using a progressive hybrid evaluation calculation method. The value assessment module is used to obtain a comprehensive evaluation score. It adopts a multi-level time-series evolution analysis and development potential prediction method to obtain the comprehensive value assessment results of talents. The decision support module is used to obtain comprehensive evaluation scores and talent comprehensive value assessment results. It adopts intelligent visualization analysis and multi-dimensional decision support methods to obtain complete talent evaluation output results.

Images