A post selection whole-process intelligent management and control system based on a dynamic talent pool

Abstract

The application discloses a post selection whole-process intelligent management and control system based on a dynamic talent pool, belongs to the field of post management, and comprises a post analysis module, a talent integration module, a portrait generation module, a prediction and early warning module, a matching and scoring module, a recommendation and review module, a comprehensive evaluation module, a process cooperation module, a promotion management module, a storage and traceability module, a decision-making suggestion module and a publicizing and updating module. The application avoids manual repeated input and information fragmentation, makes talent portraits more comprehensive and real-time, can reflect changes in ability and potential trends, significantly improves the quality of post-talent matching, automatically updates weights and thresholds according to previous selection results, continuously optimizes the recommendation strategy, finally forms a special self-evolution selection model, strengthens the scientific nature and differentiated judgment ability of the investigation link, reduces human bias, and improves the transparency, fairness and consistency of the selection process.

Description

Technical Field

[0001] This invention relates to the field of job management, and in particular to an intelligent control system for the entire process of job selection based on a dynamic talent pool. Background Technology

[0002] As hospitals expand, their medical services grow, and their disciplines develop more deeply, the demand for high-quality talent is showing new characteristics, including structural improvement, cross-disciplinary integration, and dynamic changes. Meanwhile, traditional talent selection methods often rely on manual judgment, experience-based decisions, and static data, resulting in problems such as fragmented information, inconsistent standards, insufficient process traceability, difficulties in cross-departmental collaboration, and low efficiency in matching personnel to positions. In actual management, hospitals also face challenges such as a mismatch between their talent pool structure and strategic needs, unclear succession plans for key positions, insufficient identification of potential talent, and delayed early warning of high-risk personnel. Against the backdrop of digital transformation and the construction of smart hospitals, leveraging new technologies such as artificial intelligence, big data, blockchain, and process engines to build a comprehensive, quantifiable, and predictable talent selection and management system has become an important direction for hospitals to improve governance capabilities and strengthen the standardization of cadre management. To improve the scientific, objective, and forward-looking nature of talent management, it is necessary to construct an intelligent control system for the entire process of job selection based on a dynamic talent pool.

[0003] Existing intelligent management and control systems for the entire job selection process are prone to problems such as repetitive manual data entry and information fragmentation. They also fail to create comprehensive, real-time talent profiles that reflect changes in abilities and potential trends, thus reducing the quality of job-talent matching. Furthermore, the scientific rigor and ability to differentiate in the assessment process are low, resulting in significant human bias and reduced transparency, fairness, and consistency in the selection process. To address these issues, we propose an intelligent management and control system for the entire job selection process based on a dynamic talent pool. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent management and control system for the entire job selection process based on a dynamic talent pool.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A smart management and control system for the entire process of job selection based on a dynamic talent pool includes a job analysis module, a talent integration module, a profile generation module, a prediction and early warning module, a matching and scoring module, a recommendation and review module, a comprehensive evaluation module, a process collaboration module, a promotion management module, an evidence storage and traceability module, a decision suggestion module, and a public announcement and update module. The job analysis module generates structured job requirement tags for each candidate job based on the existing job qualification standard library. The talent integration module is used to automatically connect with various business systems and obtain full-dimensional information on on-the-job personnel in real time to build a dynamic talent resource pool. The profile generation module is used to build digital talent profiles for each on-the-job personnel based on a dynamic talent resource pool; The prediction and early warning module is used to predict the trend of changes in the abilities of each employee, the possibility of improving job suitability, and the probability of promotion / resignation. Based on job requirements, it generates high-potential talent recommendations and risk warnings. The matching and scoring module is used to match and evaluate job requirement tags with various digital talent profiles from multiple dimensions. The recommendation review module is used to automatically retrieve the educational background, professional title, assessment and resume information of the corresponding on-the-job personnel in the dynamic talent resource pool for qualification verification, and at the same time generate a recommendation ticket template and synchronize basic information. The comprehensive evaluation module generates a comprehensive evaluation report based on the behavioral data, physiological indicators, and voice and text emotion analysis data of each on-duty personnel, which is used for review and decision-making reference. The process collaboration module is used to push pending tasks and required data to relevant departments during process advancement, issue warnings for overdue nodes, identify the causes of process bottlenecks, and provide improvement suggestions. The promotion management module is used to trigger a special process when a case of exceptional promotion is detected, requiring the relevant on-duty personnel to upload the basis for the exceptional promotion and relevant supporting materials, and to approve it at each level, while recording the history of exceptional promotion. The evidence storage and traceability module is used to write the selection process for each position into the blockchain and provide authorized nodes to the auditing department for complete record review. The decision recommendation module is used to summarize various data and generate a comprehensive selection analysis file for candidates; The public announcement update module is used to generate public announcement content and filing materials after the appointment decision is released, and to update the appointment information, career path trajectory and development tags in the dynamic talent resource pool.

[0006] As a further aspect of the present invention, the specific steps of the job analysis module in generating structured job requirement tags for each candidate job are as follows: S1.1: Based on the candidate job titles, the input text is then standardized and then a multi-strategy parallel retrieval method is used to search the hospital's existing job title standard database. Multiple candidate job title standard texts are returned in descending order of matching priority. S1.2: Standardize the character encoding and punctuation format of the job title standard texts for each position, then segment the standardized job title standard texts for each position, and perform word segmentation, part-of-speech tagging and syntactic dependency parsing on each segment. Based on the processing results, establish a list of paragraphs for the job title standard texts for each position, and add the original position index and basic linguistic annotation information to each paragraph. S1.3: The scoring criteria for paragraphs as candidate elements are based on name entity and pattern rule extraction, as well as the calculation of the cosine similarity between the job query vector and the paragraph vector. If a paragraph is selected by the rule extraction and the cosine similarity is higher than the preset threshold, its structured tag fragments are retained and merged. If only the cosine similarity is higher than the preset threshold, it is retained and marked with "semantic inference" for subsequent manual verification. Otherwise, if neither of the two extraction conditions is met, it is filtered out. S1.4: The paragraphs that have passed the screening and manual verification are added to the candidate pool. Each candidate paragraph is segmented using syntactic dependency information and semantic role labeling to generate multiple sets of phrases. Then, based on the part-of-speech and named entity results, each set of phrases generated after segmentation is mapped to specific fields in the core responsibilities, competency requirements, job qualifications and KPI elements. Then, through industry dictionaries, the institute's terminology database and word vector nearest neighbor retrieval, a set of candidate synonym phrases for each phrase is generated. Then, semantic similarity and word frequency information are used for screening, and each candidate synonym phrase is normalized to standard terminology. S1.5: Calculate the importance weight of each element based on text saliency, semantic relevance, and priority of demand strategy, and identify the indicator object, numerical target, time window, and calculation method description in each KPI description. At the same time, map the extracted information into structured KPI records, map the normalized phrases to the tag namespace, and generate job requirement tags based on preset tag specifications. When the number of candidates or importance density of any leaf-level tag is higher than the threshold, it is automatically split into finer sub-tags; otherwise, it is merged into a broader parent tag. S1.6: Based on the importance weight of each element, semantic similarity evidence, and KPI quantification completeness, calculate the overall confidence score of the job requirement label. If the overall confidence score is higher than the preset threshold, the job requirement label is directly output. If it is within the preset verification range, it is marked as pending verification; otherwise, the job requirement label is filtered out or archived as a candidate.

[0007] As a further aspect of the present invention, the specific steps for the talent integration module to construct a dynamic talent resource pool are as follows: S2.1: Based on the business systems being connected, identify the available APIs and database access methods one by one, create an access list for each system, establish credential management for each access endpoint, and configure key security storage and automatic rotation policies. Then, periodically check the connection status of each interface and record availability and latency. When a connection failure is detected, trigger an alarm and record the failure history for operation and maintenance analysis. After that, establish an access permission level mapping table and access approval flow, and configure the extraction mode for each system according to the access list. S2.2: Perform semantic modeling on personnel data from each source, establish a unified metadata dictionary, and save a mapping table from source fields to target fields. Based on preset conversion rules, perform format conversion on each mapped field. Then, for the remaining text fields, establish standardized pipelines for word segmentation, synonym replacement, and normalized terminology. At the same time, record the mapping confidence and unmatched fields and provide a manual supplementation interface. S2.3: Set the validation rules for each text field, perform batch validation on each field and generate a validation report. At the same time, mark the missing values ​​in each text field as pending manual entry. Then, use the box method to identify the outliers in each text field, delete the identified outliers and record them in the audit table. After that, perform cross-system cross-checking and mark the conflicting text fields as conflict records and trigger manual verification. S2.4: Establish a unified identity master index and set the corresponding authority identifier for the master index. Use the authority identifier to identify newly accessed data. If the corresponding authority identifier cannot be matched, perform fuzzy matching, calculate the similarity score of multiple candidate pairs generated by the matching, and merge the accessed data with the similarity score higher than the preset threshold with the corresponding identity. Otherwise, mark it as manual review and retain the source traceability information during the merging process. S2.5: Unify the timestamps of personnel data from different systems, build an event timeline for each personnel record, define a basic feature set according to the talent dimension, perform corresponding preprocessing on each type of personnel data, and store the preprocessed personnel data into the feature warehouse and label it with the current version number; S2.6: Set talent tag types and set corresponding rule templates for each type of tag. Calculate the tag trigger conditions for each person in batches and write them into the tag table. Record the tag source, generation time and confidence level to establish a corresponding talent pool. After receiving new data or extracting batches, compare the current talent pool records, calculate the change volume and change risk, write down changes that can be automatically merged and record audit items, automatically report changes with conflicts or impacts exceeding preset thresholds for manual review, and generate difference reports periodically. S2.7: When a missing, conflicting, or confidence label below a preset threshold is detected, a supplementary entry task is automatically generated and distributed to the responsible department or individual. At the same time, the operator, operation time, change description, and uploaded supporting documents are recorded during the supplementary entry. If the supplementary entry comes from an individual's application, the self-application materials are automatically required to be uploaded and the materials are reviewed. The reviewers verify the supplementary entry items and can accept, reject, or return them for modification. At the same time, the manual judgment results are written back to the talent pool.

[0008] As a further aspect of the present invention, the authoritative identifier mentioned in S2.4 specifically includes employee number, ID card number or personnel database ID, and when the authoritative identifier is missing, a multi-field aggregation matching rule is used for identification matching. The talent dimensions described in S2.5 specifically include personal information, education, professional title, work experience, performance evaluation score, research achievements, training hours, and disciplinary records.

[0009] As a further aspect of the present invention, the specific steps of the prediction and early warning module in generating high-potential talent recommendations and risk warnings based on job requirements are as follows: S3.1: Collect historical time series data of each person in real time from the dynamic talent pool, unify the historical time series data to the same time step, imput missing values ​​for each group of historical time series data, and correct outliers of each historical time series data based on IQR or quantiles. Then, construct a corresponding time index sequence for each person and normalize each historical time series data. S3.2: Calculate the short-term and long-term sliding aggregation of each group of historical time series data to obtain the corresponding short-term fluctuations and long-term trends. Encode each discrete event in each historical time series data into an event indicator sequence. Construct an event window after the event occurs to obtain the changes in personnel behavior before and after the event in real time. Calculate the difference sequence and growth rate of each indicator. Based on the change rate and trend strength of the corresponding indicators, splice the obtained data to generate the corresponding personnel input sequence. S3.3: Based on the Transformer architecture, a corresponding push warning model is established. Then, each input sequence is input into the push warning model. After that, the push warning model passes each input sequence layer by layer through the forward propagation algorithm, and calculates the hidden state of each input sequence at each time step through the gating mechanism. Finally, the total loss value of each hidden state is calculated through the multi-task loss function. S3.4: The obtained total loss value is input into the push warning model based on the backpropagation algorithm, and the gradient value of the total loss value with respect to the parameters of each layer of the push warning model is calculated layer by layer. Then, the Adamn optimizer is used to adjust the parameters of each layer. The parameters of each layer of the push warning model are iteratively updated repeatedly until the total loss value of the push warning model converges to the preset range. S3.5: Utilize the trained push notification model to perform multi-step rolling predictions of the hidden state of each employee at a given current time point, outputting the ability score, job suitability, and promotion or resignation probability prediction sequence for each step. At the same time, the employee promotion or resignation probability output by the model is calibrated, and based on the prediction data of each group, the potential score and turnover risk score of each employee are calculated and classified as high, medium, and low according to preset thresholds. S3.6: For personnel with high potential or high turnover risk, trigger an alert and add the corresponding personnel to the pending review list for intervention by human resources or management. At the same time, record the triggering reason and time for subsequent audit and strategy optimization.

[0010] As a further aspect of the present invention, the historical time-series data mentioned in S3.1 includes quarterly performance scores, monthly learning hours, number of job change records, training pass rate, etc.

[0011] As a further aspect of the present invention, the specific steps of the matching and scoring module in matching and evaluating job requirement tags with various digital talent profiles are as follows: S4.1: Receive the job requirement tag set for each candidate position, break down each tag into coarse-grained terms, merge synonym terms through a thesaurus and terminology mapping table, then count the frequency of each tag in the job description, and perform preliminary weighting based on job level or strategic priority to obtain the initial strength of each job requirement tag, and normalize the initial strength of each job requirement tag to obtain the job requirement tag vector. S4.2: Extract various tag values ​​of candidates from the talent pool and preprocess them. Then, map the numerical and categorical features from different sources respectively. After that, assign initial trust weights to various tag values ​​according to the reliability of the source. Then, sum the various tag values ​​at the element level to generate a candidate talent profile vector. S4.3: Set up four scoring sub-modules: basic condition compliance, ability fit, performance relevance, and development potential. Each scoring sub-module uses each dimension of the job requirement tags to perform local matching scores. Then, each scoring sub-module calculates the similarity between the job requirement tag vector and the talent profile vector projected in the corresponding dimension, and obtains the score of each scoring sub-module through normalization. Based on the current weight of each scoring sub-module, the scores of each scoring sub-module are weighted and summed to obtain the final matching score. S4.4: Using the context of the current selection task and the matching score of the current candidate as input data for the environment state, the corresponding action space is established based on the adjustment of the sub-module weight vector and the screening threshold. The near-end strategy optimization method is adopted, and the corresponding reward signal is generated based on the recommendation results generated on the candidate talent pool by the current selection strategy. S4.5: Iterate and update the current selection strategy based on the generated reward signal until the cumulative reward converges to the preset range after multiple rounds of updates. Stop iterating and dynamically update the passing threshold based on the expected return. Then compare the final matching score with the passing threshold and sort the corresponding candidates who exceed the passing threshold from highest to lowest according to their final matching score, and generate a list of qualified candidates.

[0012] As a further aspect of the present invention, the specific steps for the comprehensive evaluation module to establish a comprehensive evaluation report are as follows: S5.1: Pull raw streaming data from various data sources and record it using a unified event format. Align all records according to a unified timeline, then map each modality to a unified observation window. Use interpolation to fill in the missing statistics for each modality within the observation window and record the filling flags to generate the original aligned view of each candidate. S5.2: Extract various types of data, including behavioral logs, physiological data, speech data, and text data, from each original aligned view. Based on these data types, obtain different modal features and output the corresponding modal meta-information. Then, calculate the quality score of each modal feature, perform scale normalization on each modal feature, and generate a trust vector for each modality. S5.3: Through the projection layer corresponding to each modality, the features of each modality are converted into representations of the same dimension. Then, the initial importance score of each modality feature is calculated. After that, the softmax function is used to normalize each importance score into attention weights. The corresponding fusion vector is obtained by weighted summation. Finally, the fusion vector and the corresponding modality contribution matrix are output. S5.4: The fusion vector is fed into three parallel scoring subnetworks for behavior, psychology, and ability. Each scoring subnetwork outputs the original score based on the fusion vector. Then, the output original scores are mapped and scaled. At the same time, the processed score of each dimension is decomposed into the corresponding driving factor contribution vector. Then, the uncertainty of each dimension score is calculated, and the scores, confidence intervals, main contribution modes, suggested examination focus, and quantifiable indicator tracking items for each dimension of behavior, psychology, and ability are generated based on each set of data.

[0013] As a further aspect of the present invention, the feature values ​​mentioned in S5.2 specifically include: calculating statistics such as event frequency, response delay distribution, collaborative network degree, session length distribution, and task completion rate based on behavior logs; calculating the mean heart rate, heart rate variability time domain features, skin conductance peak count, and recovery time based on physiological data during the stress simulation segment; simultaneously calculating short-term statistics; extracting fundamental frequency, energy, short-term entropy, speaking rate, emotion score, and semantic emotion tags from speech data; and extracting TF-IDF vectors, emotion scores, keyword coverage, and negative word ratios from text data.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention standardizes the input job titles, uses multi-strategy retrieval to filter high-quality text from a job qualification standard library, and extracts core responsibilities, competency requirements, qualifications, and KPI elements through steps such as segmentation, part-of-speech tagging, dependency parsing, and semantic matching. Finally, it generates structured job requirement tags and assesses their confidence level. It automatically connects to various business systems, performs unified modeling, standardization, cleaning, conflict resolution, and identity master index merging on personnel data to form a continuously updated dynamic talent pool. Tags are generated in batches according to rule templates, supporting manual data entry and review mechanisms. Afterwards, time-series data such as personnel historical performance, learning trajectory, and job experience are uniformly processed to construct the input sequence. A Transformer-based early warning model learns changes in ability, fit trends, and promotion / departure risks, outputting potential and risk assessments and automatically triggering early warnings in the job description. In the matching phase, job requirement tags are compared with talent profile vectors. A matching score is calculated through a multi-dimensional scoring module, and reinforcement learning is used to automatically optimize weights and thresholds based on historical selection results, enabling the matching strategy to continuously evolve. Subsequently, multimodal data such as behavioral logs, physiological indicators, voice, and text are integrated. After alignment, feature extraction, and attention fusion, a three-dimensional comprehensive evaluation result of behavior, psychology, and ability is generated, and the score, contribution modality, and assessment suggestions are output. This avoids repetitive manual data entry and information fragmentation, making the talent profile more comprehensive and real-time, reflecting changes in ability and potential trends, and significantly improving the quality of job-talent matching. It can automatically update weights and thresholds based on past selection results, continuously optimizing the recommendation strategy, and ultimately forming a unique self-evolving selection model. This enhances the scientific nature and differentiated judgment ability of the assessment process, reduces human bias, and improves the transparency, fairness, and consistency of the selection process. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a system block diagram of an intelligent management and control system for the entire process of job selection based on a dynamic talent pool, as proposed in this invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1, referring to Figure 1A smart management and control system for the entire process of job selection based on a dynamic talent pool includes a job analysis module, a talent integration module, a profile generation module, a prediction and early warning module, a matching and scoring module, a recommendation and review module, a comprehensive evaluation module, a process collaboration module, a promotion management module, an evidence storage and traceability module, a decision suggestion module, and a public announcement and update module.

[0019] The job analysis module generates structured job requirement tags for each candidate job based on an existing job qualification standard library.

[0020] Specifically, based on the candidate job titles, the input text is standardized and then a multi-strategy parallel retrieval method is used to search the hospital's existing job title standard database. Multiple sets of candidate job title standard texts are returned, sorted by matching priority from high to low. The character encoding and punctuation format of each job title standard text are standardized. The standardized job title standard texts are then segmented, and each segment undergoes word segmentation, part-of-speech tagging, and syntactic dependency parsing. Based on the processing results, a paragraph list of each job title standard text is created, and an original position index and basic linguistics are added to each paragraph. The annotation information is extracted based on named entity and pattern rules, and the cosine similarity between the job query vector and the paragraph vector is calculated. These criteria are used to score paragraphs as candidate elements. If a paragraph is selected by the rules and its cosine similarity is higher than a preset threshold, its structured tag fragments are retained and merged. If only the cosine similarity is higher than the preset threshold, it is retained and marked with a "semantic inference" tag for subsequent manual verification. Conversely, if neither extraction condition is met, it is filtered out. Paragraphs that pass both the filtering and manual verification are added to the candidate pool and segmented using syntactic dependency information and semantic role annotation. Candidate paragraphs are generated into multiple phrase sets. Then, based on part-of-speech and named entity recognition results, each phrase set is mapped to specific fields within the core responsibilities, competency requirements, job qualifications, and KPI element sets. Next, using industry dictionaries, the institute's terminology database, and word vector nearest neighbor retrieval, a set of candidate synonym phrases is generated for each phrase. Semantic similarity and word frequency information are then used for filtering, and each candidate synonym phrase is normalized to standard terminology. The importance weight of each element is calculated based on textual saliency, semantic relevance, and priority of demand strategies. Finally, the indicator objects, numerical targets, and timeframes in each KPI description are identified. The document describes the window and calculation criteria, mapping extracted information to structured KPI records. Normalized phrases are then mapped to a tag namespace, and job requirement tags are generated based on preset tag specifications. If the number of candidates or importance density of any leaf-level tag exceeds a threshold, it is automatically split into finer sub-tags; otherwise, it is merged into a broader parent tag. The overall confidence score of the job requirement tag is calculated based on the importance weight of each element, semantic similarity evidence, and KPI quantification completeness. If the overall confidence score is higher than a preset threshold, the job requirement tag is directly output. If it falls within a preset verification range, it is marked as pending review; otherwise, the job requirement tag is filtered out or archived as a candidate.

[0021] The talent integration module is used to automatically connect with various business systems and obtain full-dimensional information on on-the-job personnel in real time to build a dynamic talent resource pool.

[0022] Specifically, based on the business systems being connected, available APIs and database access methods are systematically reviewed, and an access list is created for each system. Credential management is established for each access endpoint, and key security storage and automatic rotation policies are configured. Then, the connection status of each interface is periodically checked, and availability and latency are recorded. When a connection failure is detected, an alarm is triggered, and the failure history is recorded for operations and maintenance analysis. Next, an access permission level mapping table and access approval flow are established, and extraction modes are configured for each system according to the access list. Semantic modeling is performed on personnel data from each source, a unified metadata dictionary is established, and a mapping table from source fields to target fields is saved. Based on preset conversion rules, the format of each mapped field is converted, and then the remaining text fields are... Establish standardized pipelines for word segmentation, synonym replacement, and normalized terminology. Record mapping confidence levels and unmatched fields, and provide a manual supplementation interface. Set validation rules for each text field, perform batch validation of each field, and generate validation reports. Mark missing values ​​in each text field as requiring manual supplementation. Then, use box-cutting to identify outliers in each text field, delete the identified outliers, and record them in the audit table. Perform cross-system cross-checking, marking conflicting text fields as conflict records and triggering manual verification. Establish a unified identity master index and set a corresponding authority identifier for the master index. Use the authority identifier to identify newly accessed data; if a match with the corresponding authority identifier cannot be found, perform fuzzy matching. The process involves calculating similarity scores for multiple candidate pairs generated through matching, merging access data with similarity scores exceeding a preset threshold with their corresponding identities, and marking those with scores below a preset threshold as requiring manual review. During the merging process, source tracing information is preserved, timestamps for personnel data from different systems are standardized, and an event timeline is constructed for each personnel record. A basic feature set is then defined based on talent, followed by preprocessing of personnel data according to different types. The preprocessed personnel data is stored in a feature repository and labeled with the current version number. Talent tag types are set, and corresponding rule templates are defined for each tag type. Tag triggering conditions for each personnel are calculated in batches and written to a tag table, recording the tag source, generation time, and confidence level to establish... Corresponding to the talent pool, after receiving new data or extracting batches, the system compares the current talent pool records, calculates the change volume and change risk, and directly writes and records audit entries for changes that can be automatically merged. Changes with conflicts or impacts exceeding preset thresholds are automatically reported for manual review, and discrepancy reports are generated periodically. When missing, conflicting, or confidence labels below preset thresholds are detected, a supplementary entry task is automatically generated and distributed to the responsible department or individual. During the supplementary entry, the operator, operation time, change description, and uploaded supporting documents are recorded. If the supplementary entry comes from an individual's application, the system automatically requires the upload of self-submitted materials and conducts material approval. Reviewers verify the supplementary entry items and can accept, reject, or return them for modification. At the same time, the manual judgment results are written back to the talent pool.

[0023] The profile generation module is used to build digital talent profiles for each employee based on a dynamic talent resource pool.

[0024] Example 2, refer to Figure 1 A smart management and control system for the entire process of job selection based on a dynamic talent pool includes a job analysis module, a talent integration module, a profile generation module, a prediction and early warning module, a matching and scoring module, a recommendation and review module, a comprehensive evaluation module, a process collaboration module, a promotion management module, an evidence storage and traceability module, a decision suggestion module, and a public announcement and update module.

[0025] The prediction and early warning module is used to predict the trend of changes in the abilities of each employee, the possibility of improved job suitability, and the probability of promotion / departure. Based on job requirements, it generates recommendations for high-potential talents and risk warnings.

[0026] Specifically, historical time-series data of each individual is collected in real time from a dynamic talent pool. All historical time-series data are unified to the same time step. Missing values ​​are imputed for each set of historical time-series data, and outliers are corrected based on IQR or quantiles. A corresponding time index sequence is then constructed for each individual. The historical time-series data are then normalized, and short-term and long-term sliding aggregates are calculated for each set of historical time-series data to obtain corresponding short-term fluctuations and long-term trends. Discrete events in each historical time-series data are encoded as event indicator sequences, and an event window is constructed after an event occurs to obtain real-time changes in individual behavior before and after the event. The difference sequences and growth rates of each indicator are calculated. Based on the rate of change and trend strength of the corresponding indicators, the acquired data are concatenated to generate the corresponding input sequence for each individual. A corresponding push notification and early warning model is built based on the Transformer architecture. Each input sequence is then input into the push notification and early warning model, which then propagates the input sequences layer by layer through a forward propagation algorithm and calculates the corresponding values ​​through a gating mechanism. The hidden state at each time step of the input sequence is calculated, and the total loss value of each hidden state is calculated using a multi-task loss function. The obtained total loss value is then input into the push notification warning model based on the backpropagation algorithm. The gradient value of the total loss value with respect to the parameters of each layer of the push notification warning model is calculated layer by layer. Then, the Adamn optimizer is used to adjust the parameters of each layer. The parameters of each layer of the push notification warning model are iteratively updated repeatedly until the total loss value of the push notification warning model converges to a preset range. The trained push notification warning model is used to perform multi-step rolling prediction of the hidden state of each employee at a given current time point. The model outputs the ability score, job suitability, and the predicted sequence of promotion or resignation probability for each step. At the same time, the employee promotion or resignation probability output by the model is calibrated. Based on the prediction data of each group, the potential score and turnover risk score of each employee are calculated and divided into high, medium, and low according to preset thresholds. For employees with high potential or high turnover risk, an alert is triggered and the corresponding employees are added to the review list for human resources or management intervention. The triggering reason and time are recorded for subsequent auditing and strategy optimization.

[0027] The matching and scoring module is used to match and evaluate job requirement tags with various digital talent profiles from multiple dimensions.

[0028] Specifically, the system receives a set of job requirement tags for each candidate position. Each tag is broken down into coarse-grained terms and merged using a thesaurus and terminology mapping table. The frequency of each tag in the job description is then calculated, and preliminary weighting is applied based on job level or strategic priority to obtain the initial strength of each job requirement tag. This initial strength is then normalized to obtain a tag vector for each job requirement. Various tag values ​​for candidates are extracted from the talent pool and preprocessed. Numerical and categorical features from different sources are mapped separately. Initial trust weights are assigned to various tag values ​​according to the reliability of the source. Finally, the tag values ​​are summed element-wise to generate a candidate talent profile vector. Four scoring sub-modules are set up: basic condition compliance, ability fit, performance relevance, and development potential. Each scoring sub-module uses different dimensions of each job requirement tag for local matching and scoring. Each scoring submodule calculates the similarity between the job requirement label vector and the talent profile vector projected in the corresponding dimension, and obtains the score of each scoring submodule through normalization. Based on the current weight of each scoring submodule, the scores of each scoring submodule are weighted and summed to obtain the final matching score. The context of the current selection task and the matching score of the current candidate are used as input data for the environmental state. Based on adjusting the submodule weight vector and the screening threshold, a corresponding action space is established. A proximate strategy optimization method is adopted. Based on the recommendation results generated on the candidate talent pool by the current selection strategy, a corresponding reward signal is generated. The current selection strategy is iteratively updated according to the generated reward signal until the cumulative reward after multiple rounds of updates converges to a preset interval. Then, the iteration stops and the passing threshold is dynamically updated based on the expected return. After that, the final matching score is compared with the passing threshold, and the corresponding candidates who exceed the passing threshold are arranged from highest to lowest according to the final matching score, and a list of qualified personnel is generated.

[0029] The recommendation review module automatically retrieves the educational background, professional title, performance evaluation, and resume information of the corresponding on-the-job personnel from the dynamic talent resource pool for qualification verification, while generating recommendation ticket templates and synchronizing basic information; the comprehensive evaluation module establishes a comprehensive evaluation report based on the behavioral data, physiological indicators, and voice and text emotion analysis data of each on-the-job personnel for reference in assessment and decision-making.

[0030] Specifically, raw streaming data is pulled from various data sources and recorded using a unified event format. All records are aligned along a unified timeline, and each modality is mapped to a unified observation window. Interpolation is used to fill in missing statistics for each modality within the observation window, and the filling flag is recorded. This generates raw aligned views for each candidate. Behavioral logs, physiological data, speech data, and text data are extracted from each raw aligned view, and different modal features are obtained based on these data types. Corresponding modal metadata is also output. Then, the quality score for each modal feature is calculated, and scale normalization is performed on each modality feature. A trust vector for each modality is generated. Finally, through the projection layer corresponding to each modality, the modal features are converted into representations of the same dimension. Next, the initial importance score of each modality feature is calculated. Then, the softmax function is used to normalize each importance score into attention weights, and the corresponding fusion vector is obtained by weighted summation. Then, the fusion vector and the corresponding modality contribution matrix are output. The fusion vector is fed into three parallel scoring subnetworks for behavior, psychology and ability. Each scoring subnetwork outputs the original score based on the fusion vector. Then, the output original scores are mapped and scaled. At the same time, the processed score of each dimension is decomposed into the corresponding driving factor contribution vector. Then, the uncertainty of each dimension score is calculated, and the scores, confidence intervals, main contributing modalities, suggested examination focus and quantifiable indicator tracking items for each dimension of behavior, psychology and ability are generated based on each set of data.

[0031] The process collaboration module is used to push pending tasks and required data to relevant departments during process advancement, issue warnings for overdue nodes, identify the causes of process bottlenecks, and provide improvement suggestions.

[0032] The promotion management module is used to trigger a special process when a case of exceptional promotion is detected, requiring the relevant on-duty personnel to upload the basis for the exceptional promotion and relevant supporting materials, and to approve it at each level, while recording the history of exceptional promotion; the evidence storage and traceability module is used to write the selection operations of each position into the blockchain and provide the audit department with authorized nodes for complete record review.

[0033] The decision-making suggestion module is used to summarize various data and generate a comprehensive selection analysis file for candidates; the public announcement update module is used to generate public announcement content and filing materials after the appointment decision is released, and update the appointment information, career path trajectory and development tags in the dynamic talent resource pool.

Claims

1. A smart management and control system for the entire job selection process based on a dynamic talent pool, characterized in that, It includes modules for job analysis, talent integration, profile generation, prediction and early warning, matching and scoring, recommendation review, comprehensive evaluation, process collaboration, promotion management, evidence storage and traceability, decision-making suggestions, and public announcement and update. The job analysis module generates structured job requirement tags for each candidate job based on the existing job qualification standard library. The talent integration module is used to automatically connect with various business systems and obtain full-dimensional information on on-the-job personnel in real time to build a dynamic talent resource pool. The profile generation module is used to build digital talent profiles for each on-the-job personnel based on a dynamic talent resource pool; The prediction and early warning module is used to predict the trend of changes in the abilities of each employee, the possibility of improving job suitability, and the probability of promotion / resignation. Based on job requirements, it generates high-potential talent recommendations and risk warnings. The matching and scoring module is used to match and evaluate job requirement tags with various digital talent profiles from multiple dimensions. The recommendation review module is used to automatically retrieve the educational background, professional title, assessment and resume information of the corresponding on-the-job personnel in the dynamic talent resource pool for qualification verification, and at the same time generate a recommendation ticket template and synchronize basic information. The comprehensive evaluation module generates a comprehensive evaluation report based on the behavioral data, physiological indicators, and voice and text emotion analysis data of each on-duty personnel, which is used for review and decision-making reference. The process collaboration module is used to push pending tasks and required data to relevant departments during process advancement, issue warnings for overdue sections from multiple dimensions, identify the causes of process bottlenecks, and provide improvement suggestions. The promotion management module is used to trigger a special process when a case of exceptional promotion is detected, requiring the relevant on-duty personnel to upload the basis for the exceptional promotion and relevant supporting materials, and to approve it at each level, while recording the history of exceptional promotion. The evidence storage and traceability module is used to write the selection process for each position into the blockchain and provide authorized nodes to the auditing department for complete record review. The decision recommendation module is used to summarize various data and generate a comprehensive selection analysis file for candidates; The public announcement update module is used to generate public announcement content and filing materials after the appointment decision is released, and to update the appointment information, career path trajectory and development tags in the dynamic talent resource pool.

2. The intelligent management and control system for the entire process of job selection based on a dynamic talent pool as described in claim 1, characterized in that, The specific steps by which the job analysis module generates structured job requirement tags for each candidate job are as follows: S1.1: Based on the candidate job titles, the input text is then standardized and then a multi-strategy parallel retrieval method is used to search the hospital's existing job title standard database. Multiple candidate job title standard texts are returned in descending order of matching priority. S1.2: Standardize the character encoding and punctuation format of the job title standard texts for each position, then segment the standardized job title standard texts for each position, and perform word segmentation, part-of-speech tagging and syntactic dependency parsing on each segment. Based on the processing results, establish a list of paragraphs for the job title standard texts for each position, and add the original position index and basic linguistic annotation information to each paragraph. S1.3: The scoring criteria for paragraphs as candidate elements are based on name entity and pattern rule extraction, as well as the calculation of the cosine similarity between the job query vector and the paragraph vector. If a paragraph is selected by the rule extraction and the cosine similarity is higher than the preset threshold, its structured tag fragments are retained and merged. If only the cosine similarity is higher than the preset threshold, it is retained and marked with "semantic inference" for subsequent manual verification. Otherwise, if neither of the two extraction conditions is met, it is filtered out. S1.4: The paragraphs that have passed the screening and manual verification are added to the candidate pool. Each candidate paragraph is segmented using syntactic dependency information and semantic role labeling to generate multiple sets of phrases. Then, based on the part-of-speech and named entity results, each set of phrases generated after segmentation is mapped to specific fields in the core responsibilities, competency requirements, job qualifications and KPI elements. Then, through industry dictionaries, the institute's terminology database and word vector nearest neighbor retrieval, a set of candidate synonym phrases for each phrase is generated. Then, semantic similarity and word frequency information are used for screening, and each candidate synonym phrase is normalized to standard terminology. S1.5: Calculate the importance weight of each element based on text saliency, semantic relevance, and priority of demand strategy, and identify the indicator object, numerical target, time window, and calculation method description in each KPI description. At the same time, map the extracted information into structured KPI records, map the normalized phrases to the tag namespace, and generate job requirement tags based on preset tag specifications. When the number of candidates or importance density of any leaf-level tag is higher than the threshold, it is automatically split into finer sub-tags; otherwise, it is merged into a broader parent tag. S1.6: Calculate the overall confidence score of the job requirement label based on the importance weight of each element, semantic similarity evidence, and KPI quantification completeness. If the overall confidence score is higher than the preset threshold, the job requirement label is directly output. If it is within the preset verification range, it is marked as pending verification. Otherwise, the job requirement label is filtered out or archived as a candidate.

3. The intelligent management and control system for the entire process of job selection based on a dynamic talent pool as described in claim 1, characterized in that, The specific steps for the talent integration module to construct a dynamic talent resource pool are as follows: S2.1: Based on the business systems being connected, identify the available APIs and database access methods one by one, create an access list for each system, establish credential management for each access endpoint, and configure key security storage and automatic rotation policies. Then, periodically check the connection status of each interface and record availability and latency. When a connection failure is detected, trigger an alarm and record the failure history for operation and maintenance analysis. After that, establish an access permission level mapping table and access approval flow, and configure the extraction mode for each system according to the access list. S2.2: Perform semantic modeling on personnel data from each source, establish a unified metadata dictionary, and save a mapping table from source fields to target fields. Based on preset conversion rules, perform format conversion on each mapped field. Then, for the remaining text fields, establish standardized pipelines for word segmentation, synonym replacement, and normalized terminology. At the same time, record the mapping confidence and unmatched fields and provide a manual supplementation interface. S2.3: Set the validation rules for each text field, perform batch validation on each field and generate a validation report. At the same time, mark the missing values ​​in each text field as pending manual entry. Then, use the box method to identify the outliers in each text field, delete the identified outliers and record them in the audit table. After that, perform cross-system cross-checking and mark the conflicting text fields as conflict records and trigger manual verification. S2.4: Establish a unified identity master index and set the corresponding authority identifier for the master index. Use the authority identifier to identify newly accessed data. If the corresponding authority identifier cannot be matched, perform fuzzy matching, calculate the similarity score of multiple candidate pairs generated by the matching, and merge the accessed data with the similarity score higher than the preset threshold with the corresponding identity. Otherwise, mark it as manual review and retain the source traceability information during the merging process. S2.5: Unify the timestamps of personnel data from different systems, build an event timeline for each personnel record, define a basic feature set according to the talent dimension, perform corresponding preprocessing on each type of personnel data, and store the preprocessed personnel data into the feature warehouse and label it with the current version number; S2.6: Set talent tag types and set corresponding rule templates for each type of tag. Calculate the tag trigger conditions for each person in batches and write them into the tag table. Record the tag source, generation time and confidence level to establish a corresponding talent pool. After receiving new data or extracting batches, compare the current talent pool records, calculate the change volume and change risk, write down the changes that can be automatically merged and record the audit items, and automatically report changes that conflict or have an impact exceeding the preset threshold for manual review. At the same time, issue regular difference reports. S2.7: When a missing, conflicting, or confidence label below a preset threshold is detected, a supplementary entry task is automatically generated and distributed to the responsible department or individual. At the same time, the operator, operation time, change description, and uploaded supporting documents are recorded during the supplementary entry. If the supplementary entry comes from an individual's application, the self-application materials are automatically required to be uploaded and the materials are reviewed. The reviewers verify the supplementary entry items and can accept, reject, or return them for modification. At the same time, the manual judgment results are written back to the talent pool.

4. The intelligent management and control system for the entire process of job selection based on a dynamic talent pool as described in claim 2, characterized in that, The specific steps of the prediction and early warning module in generating high-potential talent recommendations and risk warnings based on job requirements are as follows: S3.1: Collect historical time series data of each person in real time from the dynamic talent pool, unify the historical time series data to the same time step, imput missing values ​​for each group of historical time series data, and correct outliers of each historical time series data based on IQR or quantiles. Then, construct a corresponding time index sequence for each person and normalize each historical time series data. S3.2: Calculate the short-term and long-term sliding aggregation of each group of historical time series data to obtain the corresponding short-term fluctuations and long-term trends. Encode each discrete event in each historical time series data into an event indicator sequence. Construct an event window after the event occurs to obtain the changes in personnel behavior before and after the event in real time. Calculate the difference sequence and growth rate of each indicator. Based on the change rate and trend strength of the corresponding indicators, splice the obtained data to generate the corresponding personnel input sequence. S3.3: Based on the Transformer architecture, a corresponding push warning model is established. Then, each input sequence is input into the push warning model. After that, the push warning model passes each input sequence layer by layer through the forward propagation algorithm, and calculates the hidden state of each input sequence at each time step through the gating mechanism. Finally, the total loss value of each hidden state is calculated through the multi-task loss function. S3.4: The obtained total loss value is input into the push warning model based on the backpropagation algorithm, and the gradient value of the total loss value with respect to the parameters of each layer of the push warning model is calculated layer by layer. Then, the Adamn optimizer is used to adjust the parameters of each layer. The parameters of each layer of the push warning model are iteratively updated repeatedly until the total loss value of the push warning model converges to the preset range. S3.5: Utilize the trained push notification model to perform multi-step rolling predictions of the hidden state of each employee at a given current time point, outputting the ability score, job suitability, and promotion or resignation probability prediction sequence for each step. At the same time, the employee promotion or resignation probability output by the model is calibrated, and based on the prediction data of each group, the potential score and turnover risk score of each employee are calculated and classified as high, medium, and low according to preset thresholds. S3.6: For personnel with high potential or high turnover risk, trigger an alert and add the corresponding personnel to the pending review list for intervention by human resources or management. At the same time, record the triggering reason and time for subsequent audit and strategy optimization.

5. The intelligent management and control system for the entire process of job selection based on a dynamic talent pool as described in claim 1, characterized in that, The specific steps of the matching and scoring module in matching and evaluating job requirement tags with various digital talent profiles are as follows: S4.1: Receive the job requirement tag set for each candidate position, break down each tag into coarse-grained terms, merge synonym terms through a thesaurus and terminology mapping table, then count the frequency of each tag in the job description, and perform preliminary weighting based on job level or strategic priority to obtain the initial strength of each job requirement tag, and normalize the initial strength of each job requirement tag to obtain the job requirement tag vector. S4.2: Extract various tag values ​​of candidates from the talent pool and preprocess them. Then, map the numerical and categorical features from different sources respectively. After that, assign initial trust weights to various tag values ​​according to the reliability of the source. Then, sum the various tag values ​​at the element level to generate a candidate talent profile vector. S4.3: Set up four scoring sub-modules: basic condition compliance, ability fit, performance relevance, and development potential. Each scoring sub-module uses each dimension of the job requirement tags to perform local matching scores. Then, each scoring sub-module calculates the similarity between the job requirement tag vector and the talent profile vector projected in the corresponding dimension, and obtains the score of each scoring sub-module through normalization. Based on the current weight of each scoring sub-module, the scores of each scoring sub-module are weighted and summed to obtain the final matching score. S4.4: Using the context of the current selection task and the matching score of the current candidate as input data for the environment state, the corresponding action space is established based on the adjustment of the sub-module weight vector and the screening threshold. The near-end strategy optimization method is adopted, and the corresponding reward signal is generated based on the recommendation results generated on the candidate talent pool by the current selection strategy. S4.5: Iterate and update the current selection strategy based on the generated reward signal until the cumulative reward converges to the preset range after multiple rounds of updates. Stop iterating and dynamically update the passing threshold based on the expected return. Then compare the final matching score with the passing threshold and sort the corresponding candidates who exceed the passing threshold from highest to lowest according to their final matching score, and generate a list of qualified candidates.

6. The intelligent management and control system for the entire process of job selection based on a dynamic talent pool as described in claim 5, characterized in that, The specific steps for the comprehensive evaluation module to generate a comprehensive evaluation report are as follows: S5.1: Pull raw streaming data from various data sources and record it using a unified event format. Align all records according to a unified timeline, then map each modality to a unified observation window. Use interpolation to fill in the missing statistics for each modality within the observation window and record the filling flags to generate the original aligned view of each candidate. S5.2: Extract various types of data, including behavioral logs, physiological data, speech data, and text data, from each original aligned view. Based on these data types, obtain different modal features and output the corresponding modal meta-information. Then, calculate the quality score of each modal feature, perform scale normalization on each modal feature, and generate a trust vector for each modality. S5.3: Through the projection layer corresponding to each modality, the features of each modality are converted into representations of the same dimension. Then, the initial importance score of each modality feature is calculated. After that, the softmax function is used to normalize each importance score into attention weights. The corresponding fusion vector is obtained by weighted summation. Finally, the fusion vector and the corresponding modality contribution matrix are output. S5.4: The fusion vector is fed into three parallel scoring subnetworks for behavior, psychology, and ability. Each scoring subnetwork outputs the original score based on the fusion vector. Then, the output original scores are mapped and scaled. At the same time, the processed score of each dimension is decomposed into the corresponding driving factor contribution vector. Then, the uncertainty of each dimension score is calculated, and the scores, confidence intervals, main contribution modes, suggested examination focus, and quantifiable indicator tracking items for each dimension of behavior, psychology, and ability are generated based on each set of data.

Images