A multimodal behavior data-driven intelligent recruitment matching platform and method

Abstract

The application discloses a multi-modal behavior data driven intelligent recruitment matching platform and method, relates to the technical field of human resource management and intelligent information processing, and is used for solving the problem that the system is difficult to balance between efficiency and accuracy due to the change of the number of links in the existing recruitment process. The technical means adopted is: constructing a multi-modal portrait of a post and a candidate, performing consistency checking and credibility correction between links, performing in-depth comparison on abnormal links when the number of links exceeds a threshold, simplifying the rest of the links, and continuously adjusting the threshold and parameters combined with the job performance feedback. Through the above technical means, the application effectively identifies false or contradictory information while ensuring the processing time, and improves the matching accuracy and the stability of the recruitment process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human resource management and intelligent information processing, and more particularly, relates to a multi-modal behavior data driven intelligent recruitment matching platform and method. BACKGROUND

[0002] Intelligent recruitment matching technology mainly uses information processing and data modeling methods to analyze and match job requirements and candidate characteristics to assist in recruitment decision-making. Such technology is widely used in enterprise recruitment, campus recruitment, talent evaluation and career planning scenarios, and can improve screening efficiency and reduce labor costs to a certain extent.

[0003] The prior art has the following disadvantages:

[0004] In the traditional intelligent recruitment process, candidates usually need to go through multiple links such as resume screening, questionnaire testing, and interview communication. With the increase in the number of links, if the same intensity of processing is used for all links, the overall time consumption is too long, and it is difficult to meet the response requirements of the system; if simplified processing is used uniformly, key links may be missed contradictions or abnormal information, resulting in inaccurate matching results. The prior art lacks a mechanism that can reasonably allocate processing depth among different links in combination with the number of links, importance and abnormal situations, so it is difficult to balance efficiency and accuracy. At the same time, thresholds and parameters are mostly set by artificial, and cannot be adjusted adaptively according to historical data and feedback results, resulting in insufficient flexibility of the system in different post or batch applications.

[0005] In view of the above problems, the present application provides a solution. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a multi-modal behavior data driven intelligent recruitment matching platform and method to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0008] A multi-modal behavior data driven intelligent recruitment matching method, comprising:

[0009] Receiving and obtaining the feature representation of the post and the candidate;

[0010] Implementing comparison mechanism switching and resource allocation, the system adaptively switches between the comprehensive comparison mechanism and the key comparison mechanism according to the number of links and the data complexity of the recruitment process, so as to balance the comparison accuracy and the calculation efficiency under different process scales;

[0011] The consistency check and dynamic credibility correction are performed. The system performs in-depth comparison on the links with contradictions or abnormalities, and uses simplified processing on the remaining links.

[0012] The abnormality detection and hierarchical early warning are performed. The system identifies abnormalities according to the deviation degree of the candidate characteristics, and triggers the suggestions of prompt, review or elimination according to the level.

[0013] The recruitment decision result is output. The system integrates the results of the post and the candidate characteristics, and calls the built-in matching score model to generate a comprehensive matching conclusion.

[0014] The performance feedback is introduced to carry out closed-loop optimization. The system collects performance feedback after the candidate is employed, and updates the matching score model based on the performance, compares the predicted results with the actual performance, and updates the related thresholds or parameters of the implementation comparison mechanism switching, resource allocation, consistency check, dynamic credibility correction, abnormality detection and hierarchical early warning.

[0015] In a preferred embodiment, the consistency check is performed based on a preset comparison threshold. When contradictions are found between the information before and after, evidence is recorded, and the credibility of the related information is corrected according to the severity of the information contradictions, the time sequence of the occurrence or the weight of the evidence.

[0016] In a preferred embodiment, the preset comparison threshold is determined by at least one of the following ways: based on the statistical results of historical recruitment data; based on comparative cases of employees with different performances; based on dynamic optimization of performance feedback after employment.

[0017] In a preferred embodiment, the system counts the actual number of links in the recruitment process when implementing the comparison mechanism switching and resource allocation, and judges by referring to the preset link number threshold and data complexity index. When the number of links does not exceed the threshold, the comprehensive comparison mechanism is used; when the number of links exceeds the threshold, the key comparison mechanism is used, so as to realize the adaptive switching of the mechanism.

[0018] In a preferred embodiment, the system performs consistency check and credibility correction on all recruitment links under the comprehensive comparison mechanism, so as to ensure the integrity of the information and the accuracy of the matching; under the key comparison mechanism, only the key links are deeply compared to improve the overall processing efficiency.

[0019] In a preferred embodiment, the judgment of the key links is determined according to the information gain and abnormal weight of each recruitment link to determine the priority of the link. The information gain is used to reflect the information contribution of the link to the overall judgment result of the candidate, and the abnormal weight is used to reflect the strength of the data contradictions or abnormal signs in the link. The system selects the links with high information value or significant abnormalities for deep comparison, and the remaining links are subjected to simplified verification, so as to realize the dynamic allocation and optimal utilization of computing resources.

[0020] In a preferred embodiment, the abnormality detection adopts standard score and threshold comparison for numerical features, adopts reasonable value set determination for categorical features, and divides the abnormality results into three levels of mild, moderate and severe, respectively triggering prompt, review and elimination suggestions.

[0021] In a preferred embodiment, the accuracy, recall rate and comprehensive index are calculated after collecting the on-the-job performance, and the updated parameters include consistency comparison threshold, abnormality determination threshold and the weight of each recruitment link in matching calculation.

[0022] In a preferred embodiment, the update of the parameters adopts a regression calibration method, linearly corrects the mapping between the recruitment stage features and the on-the-job performance, and updates the regression coefficients and intercepts in batches after accumulating feedback samples.

[0023] A multi-modal behavior data driven intelligent recruitment matching platform, comprising:

[0024] A data processing module for obtaining job requirements and candidate data and generating feature representations;

[0025] A resource allocation module for implementing comparison mechanism switching and resource allocation;

[0026] A consistency verification module for performing consistency verification and dynamic credibility correction;

[0027] An abnormality warning module for abnormality detection and grading warning;

[0028] A decision output module for outputting recruitment decision results;

[0029] A feedback optimization module for introducing performance feedback to carry out closed-loop optimization.

[0030] In a preferred embodiment, the resource allocation module is used to issue the generated comparison plan to the consistency verification module for execution, and the feedback optimization module is used to write back the updated parameters to the consistency verification module, resource allocation module, abnormality warning module and decision output module to form a closed-loop optimization.

[0031] The technical effects and advantages of the multi-modal behavior data driven intelligent recruitment matching platform and method of the present application are:

[0032] The application introduces a new processing method based on the number and importance of links in the recruitment process, that is, when the number of links is large, the system can preferentially perform in-depth comparison on key or abnormal links, and use simplified processing for general links, so as to balance the overall efficiency and identification of key contradictory information within a limited time. At the same time, the system uses the multi-modal portrait of the candidate to perform consistency check and credibility correction, reduces the interference of false information on the result; and collects performance feedback after the candidate is employed, dynamically adjusts the threshold and parameters, so that the processing strategy can be continuously optimized. Through the above design, the application improves the matching accuracy between the post and the candidate, and ensures the efficiency and stability of the recruitment process. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The intelligent recruitment matching method flowchart provided for the embodiment of the application is as follows:

[0034] Figure 2 The intelligent recruitment matching platform module structure diagram provided for the embodiment of the application is as follows:

[0035] Figure 3 The comparison mechanism switching and resource allocation flowchart of the application is as follows. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all embodiments.

[0037] The system aims to solve the problem that the system is difficult to balance between efficiency and accuracy due to the change of the number of links in the existing recruitment process. The existing process lacks cross-link correlation check and real-time correction, and once false information is accepted in the early stage, it will continue to affect subsequent judgment; when the number of process links is large, full comparison is time-consuming, and when the number of links is small, key information is easy to be ignored, so that efficiency and accuracy are difficult to balance. The application provides a multi-modal behavior data driven intelligent recruitment matching method, which refers to Figure 1 , Figure 1 The intelligent recruitment matching method flowchart provided for the embodiment of the application is as follows:

[0038] Step S101, constructing a post portrait, that is, the system analyzes the post demand, converts the natural language description into a standardized feature vector, and forms a structured representation of the post.

[0039] Step S102, constructing a candidate multi-modal image, that is, the system extracts features immediately after the end of the resume, online questionnaire, communication record, and interview performance, and converts data from different sources into structured vectors, and gradually combines them to form a multi-modal image of the candidate. The so-called multi-modal image is a unified representation generated by collecting and integrating multiple feature data. This image serves as a standardized input, providing direct support for subsequent consistency verification and anomaly early warning.

[0040] Step S103, implementing comparison mechanism switching and resource allocation. The system counts the actual number of loops of the current recruitment process in the running stage, and judges by referring to the number of loop thresholds determined in the initialization stage and the data complexity index, to realize adaptive switching between the two comparison mechanisms. When the number of loops does not exceed the threshold, the comprehensive comparison mechanism is enabled, and consistency verification and credibility correction are performed on all loops; when the number of loops exceeds the threshold, the key comparison mechanism is enabled, the information gain and anomaly weight of each loop are calculated, and the key loop is selected into deep comparison, and the remaining loops are executed for simplified verification, wherein the information gain and anomaly weight reflect the importance and data credibility of the loop. Through the above switching and allocation strategy, the system can balance the comparison accuracy and calculation efficiency under different process scale and complexity conditions, and realize dynamic balance and optimal utilization of resource allocation.

[0041] Step S104, performing consistency verification and dynamic credibility correction. The system compares the information of different loops with each other, records evidence when contradictions are found, and automatically reduces the credibility of related loops according to the severity and time sequence.

[0042] Step S105, performing anomaly detection and grading early warning. The system divides the deviation degree of the candidate in education, experience, skill and other characteristics into three grades of mild, moderate and severe, and triggers the suggestion of prompt, review or elimination according to the grade, to realize early interception in the middle of the process.

[0043] Step S106, outputting the recruitment decision result. After completing consistency verification, comparison mechanism switching and resource allocation, and anomaly detection and grading early warning, the system calls the built-in matching score model to comprehensively evaluate the post image and the candidate image, and generate the matching result of the post and the candidate. The model performs quantitative analysis based on the credibility result, matching index and anomaly level output in the previous stage, and outputs a structured matching conclusion. The output result includes matching level, main basis explanation and recommendation suggestion, which can be directly called by the enterprise decision-making end, or used as input for the subsequent closed-loop feedback optimization stage for model updating and threshold correction.

[0044] Step S107, introduce performance feedback to carry out closed-loop optimization. After the candidate is employed, the system will collect the actual work performance, compare the real feedback results with the matching score model output in the recruitment stage, and then dynamically adjust the consistency threshold, abnormality judgment threshold and weight of each link to make the screening strategy continue to evolve over time.

[0045] The execution subject of the above steps can be a platform or a server-side process, but is not limited thereto.

[0046] The present application realizes instant collection in each recruitment link, and then compares the new data with the previous information, and dynamically corrects the credibility when contradictions are found. In the case of a large number of links, the system uses a key comparison mechanism: key or abnormal links enter deep comparison, and ordinary links perform simplified processing, so that potential risk information can still be identified within a limited time. The platform also has an abnormal grading warning function, which prompts, reviews or eliminates different degrees of deviation; the performance results after employment will be introduced to drive the update of the threshold and weight. Through this closed-loop design, false information or distorted data can be effectively suppressed, the efficiency and accuracy of recruitment matching are taken into account, and the system can maintain stable applicability in various post and industry scenarios.

[0047] In step S101, the system first converts the job requirements provided by the enterprise from natural language description to structured feature representation. This structured representation is called job portrait. The job portrait refers to the standardized feature set formed after analyzing and splitting the job requirements, which is used to uniformly describe the recruitment requirements. The job portrait includes three types of elements: text description features, hard conditions and soft qualities. Text description features refer to the semantic information contained in the job description; hard conditions include educational background, professional background, professional qualification and work experience, etc. objective requirements; soft qualities cover communication skills, learning ability and teamwork, etc. subjective qualities. The system establishes a vector representation for each of these three parts of information, and fuses them in the vector space to finally obtain the complete job portrait vector .

[0048] For example, the vector splicing method can be used to form the job portrait:

[0049]

[0050] , wherein, represents the job text semantic vector, represents the job hard condition vector, represents the job soft quality vector, represents connecting each part of the vector into a larger vector. Through this structured modeling, the originally ambiguous job requirements are converted into a machine-understandable and computable unified representation, laying the foundation for subsequent matching with candidate information.

[0051] After the generation of the job profile, the system also sets threshold parameters using historical recruitment data and industry benchmarks as standards for screening and evaluation. For example, based on the education distribution and experience length of past excellent employees in the same position, the education range threshold and experience length threshold can be set. In some embodiments, the threshold can not only be a fixed interval, but also be extended to a probability threshold or a fuzzy interval to enhance flexibility and adaptability.

[0052] In step S102, the system extracts the candidate's data immediately at the end of each link and converts information from different sources into structured vectors. The link vectors are stored and continuously expanded one by one after collection is completed, gradually forming the candidate's multi-modal profile.

[0053] For example, in an embodiment, the recruitment process includes three links: resume, questionnaire, and interview. The system extracts the education, experience, and skill fields in the resume link, analyzes the professional knowledge and personality test results in the questionnaire link, and extracts the voice tone and non-verbal behavior in the interview link. The data of each link is immediately structured and converted into a standardized feature vector after the end, gradually building a multi-modal profile of the candidate. In this embodiment, the multi-modal profile of the candidate can be uniformly represented as:

[0054]

[0055] where, represents the feature vector of the resume link, represents the feature vector of the questionnaire link, represents the feature vector of the interview link, represents vector splicing.

[0056] It should be noted that the above three links are only examples, and the number and type of actual recruitment process links can be flexibly configured according to job requirements, such as background investigation, practical operation test, or group interview links. In this case, the multi-modal profile can be extended to the form of , where represents the feature vector of the new link.

[0057] Step S103, implement comparison mechanism switching and resource allocation. The system counts the actual number of links in the current recruitment process The system then uses the threshold number of steps and data complexity metrics determined during the initialization phase to determine the optimal switching mechanism between the two comparison methods, enabling adaptive switching. For a clearer understanding of the logical relationship between comparison mechanism switching and resource allocation, please refer to [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the comparison mechanism switching and resource allocation process of the present invention.

[0058] The system first initializes itself before running, determining the threshold number of stages based on historical recruitment data. This threshold is used to reflect the preset calculation time limit. Within this range, the system can complete the maximum number of steps for comprehensive comparison. Let the average comparison time for a single step be... Then the constraint relationship is satisfied:

[0059] In other words, when the number of steps in the recruitment process does not exceed When the number of process steps N exceeds a certain threshold, the system will activate a comprehensive comparison mechanism to perform consistency checks and reliability corrections on all steps. At that time, a key comparison mechanism will be activated.

[0060] The system introduces information gain and anomaly weighting mechanisms to select key stages for in-depth processing. Information gain is defined as:

[0061] ,in, This indicates uncertainty when all stages are included. This means removing the first... Uncertainty following each step. If The larger the value, the higher the contribution of this step to reducing uncertainty, and it should be retained as a priority.

[0062] If a contradiction is found in a certain step during consistency verification, the system will assign it an abnormal weight. That is, the first The abnormal weight values ​​corresponding to each step. The final comprehensive score is:

[0063] ,in, For information gain coefficient, For abnormal weighting coefficients, Indicates the first Abnormal weights in each stage, Indicates the first The contribution of each step to reducing uncertainty. This represents the stage index. The system sorts stages by score, prioritizing stages with higher scores for in-depth comparison, while other stages undergo only lightweight detection. This mechanism ensures accuracy in short processes and conserves computational power in complex processes.

[0064] For example, when the recruitment process contains eight steps, the threshold is five. The system will select the five most important steps to perform deep comparison, and the remaining three steps will only be checked quickly, thereby reducing the time-consuming of calculation while ensuring the reliability of the results.

[0065] In step S104, the system performs consistency check on the information provided by the candidate in different steps, and dynamically corrects the credibility score when inconsistencies are found. Consistency check refers to comparing the information provided by the candidate in the resume, questionnaire, communication and interview steps one by one to identify inconsistencies. Credibility score is a quantitative evaluation of the authenticity and reliability of the candidate's information. If the system finds conflicts in the candidate's skills, experience or timeline during the comparison process, it will immediately mark and adjust the credibility of the relevant steps.

[0066] The system finds inconsistencies through consistency check, for example, the candidate claims to be proficient in a certain skill in the resume, but scores very low on this skill in the interview answer, which is judged as inconsistent. In the above embodiment, in order to quantify consistency, the system uses similarity calculation method. If the resume skill vector is represented as and the interview skill vector is represented as , the similarity calculation is:

[0067]

[0068] When the similarity is lower than the consistency threshold , the system determines that the information is inconsistent. The threshold is set including four sources. First, statistics historical recruitment data to find the difference in similarity distribution between candidates with good and poor performance after being hired to determine the critical value. Second, combined with industry and post benchmark to make corrections, the consistency requirements of different posts are significantly different. Third, select the threshold that makes the macro average or Youden index maximum through cross-validation. Fourth, continuously update according to the job performance in the closed-loop feedback stage.

[0069] For inconsistencies, the system introduces a contradiction intensity coefficient , a time decay coefficient and a global correction coefficient to calculate the correction coefficient: , where , the value range is , used to quantify the intensity of inconsistency between fields; is the global correction coefficient, the value range is ; is a constant; is the time decay coefficient, generally represents the time interval from the contradiction occurrence to the present, represents the time decay rate.

[0070] and the revised credibility is calculated as: wherein represents the initial credibility score, represents the revised credibility score. In this way, the credibility score of the candidate will change with new evidence and be closer to the true situation.

[0071] In step S105, the system performs anomaly detection on the multi-modal feature data of the candidate and classifies the degree of deviation, while triggering the corresponding early warning mechanism. The goal of anomaly detection is to identify whether the candidate significantly deviates from the job requirements or industry conventions in terms of education, experience, skill level, or behavior performance. The classification process divides the anomaly into three levels of mild, moderate, and severe, so as to distinguish and respond: mild anomaly only for prompt, moderate anomaly requires manual review, and severe anomaly directly triggers elimination suggestion, so as to exclude obvious unqualified candidates in advance when the process is not completed.

[0072] The system uses the standard score method to quantify the degree of deviation: wherein, is the feature value of the candidate, is the industry average, is the standard deviation. When exceeds the anomaly determination threshold , the feature is marked as an anomaly. For categorical features, the system sets a reasonable value set, and if the candidate information is not within the set range, it is also considered as an anomaly.

[0073] The setting of the threshold includes three sources: one is based on statistical analysis of historical recruitment data to determine the normal interval; two is to refer to the comparison cases of high and low performance employees to find the critical point to reduce the misjudgment rate; three is to continuously optimize the threshold value by using the performance results in the closed-loop feedback of step S107, so as to ensure that the threshold value can be dynamically adjusted with data accumulation. In this way, anomaly detection has both statistical basis and adaptability to different job and industry scenarios. The system calculates the comprehensive anomaly score of each feature and classifies the anomaly into three levels of mild, moderate, and severe according to the deviation amplitude. Different levels trigger prompt, review, and elimination suggestion, respectively, to realize real-time risk control of the recruitment process.

[0074] In step S106, the system decision output module calls the matching score model to comprehensively analyze the matching degree between the job and the candidate by integrating the job portrait, the candidate portrait, the consistency verification result, and the anomaly detection result. The matching score model is established in the system initialization stage to fuse multi-source features and calculate the comprehensive matching score. The model takes the job portrait vector and candidate multimodal profile vector as the core input, while introducing a revised credibility of the consistency verification stage and the anomaly score of the anomaly detection stage as auxiliary inputs.

[0075] The model performs weighted fusion on different information sources according to the importance of each feature in the fusion calculation stage, and forms a comprehensive matching score. The comprehensive matching score can be represented as: wherein, represents the semantic similarity between the post and the candidate profile; is the revised credibility result, is the anomaly score of the anomaly detection stage; , , is a weight coefficient for controlling the relative influence of each part, and The higher the comprehensive score, the stronger the overall matching degree between the candidate and the post.

[0076] In the running process of the model, the feature vectors of the post and the candidate are first normalized and fused to extract their semantic similarity and condition difference; then the key features are weighted calculated combined with the credibility weight of each link; finally, the comprehensive matching score and semantic explanation are generated according to the anomaly level and feature deviation. The output results of the model include matching level, main basis and recommendation, which can be directly called by the enterprise decision-making end, or used as input for the closed-loop feedback optimization stage for subsequent parameter updating and threshold correction.

[0077] In step S107, the system uses the performance results of the candidate after entering the job to establish a closed-loop optimization mechanism from prediction to verification and then to revision, so that the recruitment model can continuously calibrate itself in the long-term operation. Performance data comes from the enterprise's evaluation system, including probation KPI, daily work indicators, supervisor scores and team collaboration feedback. The system will convert these data into standardized numerical values and match them one by one with the prediction results in the recruitment stage to form samples that can be used for evaluation and optimization. The system compares these performance data with the output results of the matching score model in the recruitment stage, evaluates the differences between the prediction results and the actual performance, and automatically adjusts the weight distribution and threshold setting of the matching score model accordingly. After each round of feedback optimization is completed, the system will return the updated parameters and decision logic to the consistency verification, comparison mechanism switching and resource allocation, and anomaly detection and grading early warning stages, so as to realize the adaptive closed loop from data collection, matching calculation to feedback learning.

[0078] In order to measure the reliability of the prediction results, the system counts the number of samples of the prediction results and the actual performance. the number of samples predicted as matching and performance satisfactory, the number of samples predicted as matching but performance unsatisfactory, the number of samples predicted as not matching but performance satisfactory. Based on these statistics, the system calculates three core indicators. The accuracy rate reflects how many of the candidates determined as matching actually perform satisfactorily, with the formula ; the recall rate indicates how many of all the candidates who perform satisfactorily are correctly identified and recommended by the system, with the formula ; to balance between reducing false positives and avoiding false negatives, the system further calculates the comprehensive indicator , with the formula , where represents the accuracy rate, represents the recall rate, and the larger the value, the more robust the prediction result.

[0079] In terms of parameter optimization, the system adjusts the consistency threshold and the anomaly determination threshold according to feedback data. The goal of optimization is to maximize the value on historical feedback samples. The system can try different threshold combinations in discrete space through grid search, or use Bayesian optimization in continuous space to improve efficiency. The update mechanism is subject to the following conditions: , , where represents the number of available feedback samples, represents the minimum sample size requirement, represents the comprehensive indicator before updating, represents the comprehensive indicator under the candidate threshold, represents the minimum improvement rate.

[0080] In addition to the threshold, the system also dynamically optimizes the weights of each link. The optimization goal is to minimize the loss function composed of prediction bias , with the update rule , where represents the importance coefficient of the th link; represents the link index; represents the learning rate, used to control the step size of weight update; quantifies the degree of prediction bias. If the correlation between a link and performance increases, its weight is increased; if the correlation decreases or there is a risk of misleading, its weight is decreased.

[0081] In terms of mapping between features and performance, the system uses linear regression for calibration, with the formula , where characteristic score representing the recruitment stage, performance score representing the post-hiring, regression coefficient, intercept. As new data accumulates, the system will update the regression coefficient and intercept in batches to ensure that the prediction results are consistent with the actual performance. When the number of feedback samples is insufficient, the system will postpone the update to avoid instability caused by insufficient data.

[0082] Through this design, the closed-loop feedback not only corrects the short-term errors of the model, but also forms a mechanism for self-adaptive evolution over time. It can continuously reflect the real performance signal back to the consistency verification of S104, the resource allocation of S103, and the anomaly detection of S105, so that the entire system gradually approaches the optimal in the long-term operation.

[0083] The overall role of the platform is to realize steps S101 to S107 in a modular collaborative manner. Referring to Figure 2 , Figure 2 the intelligent recruitment matching platform module structure provided by the embodiment of the application, the platform is composed of a data processing module, a resource allocation module, a consistency verification module, an anomaly early warning module, a decision output module, and a feedback optimization module. Each module runs independently while maintaining sequential connection.

[0084] The data processing module directly supports steps S101 and S102. This module completes job requirement analysis and generates a job portrait vector , and at the same time, collects candidate data and constructs a candidate multi-modal portrait vector immediately after each link is completed. The two types of vectors are standardized inputs for subsequent calculations.

[0085] The resource allocation module corresponds to step S103. This module monitors the number of links and data complexity, and selects a comprehensive comparison mechanism or a key comparison mechanism according to the threshold and upper time limit .

[0086] The consistency verification module corresponds to step S104. This module receives and related features, completes cross-link comparison according to the abnormal judgment threshold , and related features, updates the candidate credibility score according to the correction coefficient , and outputs the verified credibility result for the next step.

[0087] The abnormality early warning module corresponds to step S105. The module calculates the standard score and compares it with the threshold to complete the mild, moderate and severe abnormality classification, and trigger the early warning actions such as prompt, review and elimination according to the level, and output the classification label and processing suggestion.

[0088] The decision output module corresponds to step S106, and is used to call the matching scoring model to output the recruitment decision result.

[0089] The feedback optimization module corresponds to step S107. The module accesses the performance result, calculates the accuracy , recall rate and comprehensive index , updates the consistency threshold , abnormality determination threshold and link weight accordingly, corrects the mapping from feature to performance based on the relationship , and outputs the latest parameters to rewrite the foregoing modules.

[0090] The above modules correspond to each other in steps and form a closed loop. The core is concentrated in the four links of consistency check, dynamic resource allocation, abnormality early warning and closed loop feedback, which guarantees the landing and adaptability of the method in the recruitment scene.

[0091] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

[0092] Those skilled in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application of the technical solutions and the constraints of the application. Professional technicians can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0093] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0094] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0095] Finally, the above merely provides the preferred embodiments of the present application, but is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A multi-modal behavioral data-driven intelligent recruitment matching method, characterized in that, The method comprises the following steps: Receiving and obtaining the feature representations of the positions and candidates; Implementing comparison mechanism switching and resource allocation, and the system switches between the comprehensive comparison mechanism and the key comparison mechanism according to the number of recruitment process loops and the data complexity, so as to balance the comparison accuracy and the calculation efficiency under different process scales; Performing consistency checking and dynamic credibility correction, the system compares the information of the same candidate in at least two recruitment links, and corrects the credibility of the related information when contradictions are found between the information before and after; Performing abnormality detection and hierarchical early warning, the system identifies abnormalities according to the deviation degree of the candidate features and triggers suggestions for prompt, review or elimination according to the level; wherein the deviation degree of the candidate features is used to represent the difference between the candidate features and the position requirements; Outputting recruitment decision results, the system calls the built-in matching score model to generate a comprehensive matching conclusion by weighted fusion according to the semantic similarity between the position and the candidate portrait, the corrected credibility result and the abnormality score in the abnormality detection stage; The semantic similarity between the position and the candidate portrait is used to represent the matching degree between the position and the candidate feature representation; the corrected credibility result is used to represent the credibility of the candidate information after the consistency checking and dynamic credibility correction; and the abnormality score in the abnormality detection stage is used to represent the quantitative result of the deviation degree of the candidate features; Introducing performance feedback to carry out closed-loop optimization, the system collects performance feedback after the candidate is employed, compares the prediction results with the actual performance based on the performance, updates the matching score model, and then updates and optimizes the related thresholds or parameters of the comparison mechanism switching and resource allocation, the consistency checking and dynamic credibility correction, and the abnormality detection and hierarchical early warning.

2. The method of claim 1, wherein, The consistency checking is performed based on a preset comparison threshold, evidence is recorded when contradictions are found between the information before and after, and the credibility of the related information is corrected according to the severity of the information contradictions, the time sequence of the occurrence or the weight of the evidence.

3. The method of claim 2, wherein, The preset comparison threshold is determined by at least one of the following methods: based on the statistical results of historical recruitment data; based on comparative cases of employees with different performances; and dynamically optimized based on the performance feedback after employment.

4. The method of claim 1, wherein, When implementing the comparison mechanism switching and resource allocation, the system counts the actual number of recruitment process loops, and judges by referring to the preset loop number threshold and the data complexity index, adopts the comprehensive comparison mechanism when the loop number does not exceed the threshold, and adopts the key comparison mechanism when the loop number exceeds the threshold, so as to realize the adaptive switching of the mechanism.

5. The method of claim 4, wherein, The system performs consistency checking and credibility correction on all recruitment links under the comprehensive comparison mechanism, so as to ensure the information integrity and matching accuracy; and only performs in-depth comparison on key links under the key comparison mechanism, so as to improve the overall processing efficiency.

6. The method of claim 5, wherein, The determination of the key link determines the link priority according to the information gain and the abnormal weight of each recruitment link; the information gain is used to reflect the information contribution of the link to the overall judgment result of the candidate, and the abnormal weight is used to reflect the strength of the data contradiction or abnormal sign in the link, and the system selects the link with high information value or obvious abnormality to enter deep comparison, and the remaining links perform simplified verification, realizing dynamic allocation and optimal utilization of computing resources.

7. The method of claim 1, wherein, The abnormality detection and grading early warning system quantifies the deviation degree of the candidate's characteristics to obtain an abnormality score and generate an abnormality result; the quantification method has two kinds, the standard score is adopted for the numerical value type characteristics and the abnormality determination threshold is compared, the reasonable value set determination is adopted for the classification type characteristics; and the abnormality result is divided into three grades of mild, moderate and severe, which respectively triggers the suggestion, review and elimination suggestion; wherein, the formula of the standard score is: wherein, is the characteristic value of the candidate, is the industry average value, is the standard deviation.​ 8. The method of claim 1, wherein, After collecting the onboarding performance, the accuracy, recall rate and comprehensive index are calculated, and the updated parameters include consistency comparison threshold, abnormality judgment threshold and the weight of each recruitment link in matching calculation.

9. The method of claim 1, wherein, The updating of the parameters adopts the regression calibration method, linearly corrects the mapping between the recruitment stage characteristics and the onboarding performance, and updates the regression coefficients and intercepts in batches after the accumulation of feedback samples.

10. A multi-modal behavioral data-driven intelligent recruitment matching platform for implementing the multi-modal behavioral data-driven intelligent recruitment matching method of any one of claims 1-9, characterized in that, It includes: a data processing module for obtaining job requirements and candidate data and generating feature representation; a resource allocation module for implementing comparison mechanism switching and resource allocation; a consistency verification module for performing consistency verification and dynamic credibility correction; an abnormality early warning module for abnormality detection and grading early warning; a decision output module for outputting recruitment decision results; a feedback optimization module for introducing performance feedback to carry out closed-loop optimization; The resource allocation module is used to issue the generated comparison plan to the consistency verification module for execution, and the feedback optimization module is used to write back the updated parameters to the consistency verification module, resource allocation module, abnormality early warning module and decision output module to form a closed-loop optimization.

Images