Human resource intelligent information screening system and method based on data analysis

Abstract

The invention discloses a human resource intelligent information screening system and method based on data analysis, and particularly relates to the technical field of human resource information processing, and the method comprises the following steps: constructing a data offset weight recognition difference influence based on historical candidate samples, screening offset abnormal samples, and classifying the screened offset abnormal samples into an auxiliary verification set; after fields highly related to sensitive label fields are removed, non-sensitive attributes are extracted, candidate ability expression vectors are generated through semantic mapping, model training is completed, a disturbance comparison control mechanism is introduced, and finally, ability matching calculation is executed on a current applicant group, and an output result is calibrated in combination with an auxiliary verification set; according to the method, the quality of a training sample is controlled by constructing a data offset weight, sensitive features are eliminated based on statistical correlation to guarantee feature fairness, disturbance control and a result calibration mechanism are introduced to doubly improve structural fairness, and fairness and stability of intelligent screening are realized from a data source, a training process and a screening output full link.

Description

Technical Field

[0001] This invention relates to the field of human resource information processing technology, and more specifically, to a human resource intelligent information screening system and method based on data analysis. Background Technology

[0002] As human resource management transforms towards intelligence and digitalization, automated screening models built on historical recruitment data are gradually becoming a primary means for companies to initially identify candidates. These technologies typically extract attributes such as candidates' education, work experience, skill tags, and age, establishing a mapping relationship with existing hiring data to train a screening model for job matching. While offering significant advantages in processing efficiency and screening scale, existing methods generally suffer from insufficient identification of structural biases, leading to the gradual solidification of the model during practical application and amplifying unfair tendencies in the original data.

[0003] Specifically, in the data preprocessing stage, current screening models often lack mechanisms to identify implicit biases in historical data. For example, in past recruitment practices, due to factors such as job culture, subjective judgment, or implicit standards, certain groups (such as specific age groups or graduates from non-key universities) may be hired at a low rate. If the model does not label or remove such data differences during the training phase, the original bias will be treated as an "empirical rule for successful hiring" and incorporated into the model learning process, and further amplified by the algorithm during feature extraction and modeling. This structural bias, as the model is deployed and used in batches, may continue to affect candidate evaluation in a subtle way, resulting in inequality in the number of candidates receiving interviews or hiring opportunities in the real world.

[0004] Because existing model designs generally focus on metrics such as accuracy or recall, lacking end-to-end control over structural fairness, model outputs often fail to reflect the equal opportunity among diverse groups in recommendation results. Once systemic discrimination against sensitive groups is discovered externally, related companies may face multiple risks: first, public scrutiny; second, regulatory compliance reviews of discriminatory hiring practices; and third, distrust of screening criteria among internal employees or candidates, thereby impacting the company's brand reputation, hiring legitimacy, and overall human resource management stability. Therefore, this paper proposes a data-driven intelligent information screening system and method for human resources to address these issues. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: The data analysis-based intelligent information screening method for human resources includes the following steps: Attributes and hiring results are extracted from historical candidate samples. By performing statistical calculations on the joint distribution between attributes and hiring results, a set of data mapping parameters expressing differences in hiring preferences is obtained. This set is then used to construct data offset weights to identify the sources of difference in the training data. Based on the data offset weights, the training samples are grouped, and data whose offset weight values ​​exceed the limit range are included in the auxiliary verification sample set. In the training sample set formed after removing samples with excessive data offset weights, the statistical correlation index between the training feature fields and the sensitive label fields is identified. Training feature fields with correlation values ​​exceeding the preset judgment standard range are removed from the training sample set. The remaining training feature fields are used as a set of non-sensitive attribute fields that do not have sensitivity risks. Based on the semantic correspondence between job competency requirements and non-sensitive attribute fields, candidate competency expression vectors are generated through vector mapping in the retained training feature fields to construct a model sample set with a general competency structure for the job. Model training is performed by selecting models based on a sample set, and a perturbation contrast control mechanism is introduced to dynamically constrain the model's response to changes in input attributes during the training process. Using the trained screening model, candidate ability expression vectors are generated for the current applicant group, and matching calculations are performed based on the job standard ability structure. The screening output results are compared and verified by an auxiliary validation sample set, and the screening results are calibrated by combining the output probability distribution, so that the final output maintains the consistency of the recommendation probability distribution among different attribute groups.

[0006] In a preferred embodiment, the data mapping parameter set refers to: The historical candidate samples are classified and aggregated to construct a set of attribute combinations containing multiple preset dimensions, and the admission results corresponding to each attribute combination are labeled and statistically analyzed. Based on the aggregated attribute combinations and the admission results, the conditional admission probability of each attribute dimension under different combinations is obtained, and a joint probability matrix between each dimension and the admission results is generated to characterize the overall distribution pattern. The joint probability matrix is ​​compared with the ideal equilibrium distribution. The distribution offset distance method is used to score each attribute combination to obtain the recruitment tendency score. Finally, a set of data mapping parameters is generated to express the differences in recruitment tendency.

[0007] In a preferred embodiment, risk assessment refers to: For each candidate sample's attribute combination, the corresponding hiring preference score is extracted from the data mapping parameter set and compared with the global average hiring score to calculate the relative offset value. Based on the distribution density of relative offset values ​​in the attribute space, an attribute sensitivity factor is introduced to weight and correct the offset values, thereby reflecting the sensitivity of the sample to the overall distribution of the admission results. The weighted and corrected offset values ​​are normalized to generate numerical data offset weights that correspond one-to-one with each candidate sample, which serve as the basis for subsequent training sample selection and grouping.

[0008] In a preferred embodiment, the calculation process for the attribute sensitivity factor is as follows: For each candidate attribute dimension, the frequency of that attribute in all training samples is counted, and its proportion in the overall candidate group is calculated to measure the representativeness of that attribute in the dataset. The variance of the difference in recruitment tendency between each category and the recruitment results under this attribute dimension is calculated, and then processed according to the preset normalization rule. This variance is used as the attribute sensitivity factor to measure the recruitment sensitivity of this attribute in the training samples.

[0009] In a preferred embodiment, the logic for partitioning the auxiliary verification sample set is as follows: Based on the normalized data offset weights corresponding to each training sample, a sample offset distribution sequence is constructed and arranged from low to high values ​​to form a continuous distribution interval. Set upper and lower bounds for the offset limit interval, filter out the sample set whose values ​​exceed the range by the interval boundary, and mark them as abnormal offset samples; Abnormal offset samples are removed from the training sample set and uniformly included in the auxiliary verification sample set.

[0010] In a preferred embodiment, the statistical correlation index is calculated using any one of the mutual information method, the Pearson correlation coefficient method, or the Spearman rank correlation coefficient method, and is used to quantify the degree of correlation between the training feature field and the sensitive label field.

[0011] In a preferred embodiment, the semantic correspondence between job competency requirements and non-sensitive attribute fields in the retained training feature fields is first measured by a semantic similarity score, which is determined based on the word vector similarity between the field name and the job competency label, the degree of keyword dictionary matching, or the matching rules in the standard competency label library. After obtaining the scoring results, for each job competency tag, fields with semantic similarity scores exceeding the preset similarity standard are selected from the non-sensitive attribute fields, and a unique or limited number of field-competency tag mapping relationships are established to construct the competency expression vector under the job competency dimension.

[0012] In a preferred embodiment, introducing a disturbance contrast control mechanism refers to: In the model sample set, the training feature fields are subjected to structured perturbation to construct a set of corresponding sample pairs of the original capability expression vector and the perturbed capability expression vector, which are used to detect the difference in the model's response to the training feature fields. The sample pairs are input to the model for screening. By comparing the difference in intermediate outputs generated by the two types of inputs in the sample pairs during training, a response evaluation parameter is formed to characterize the magnitude of the model's response. This parameter is then compared with the set control standard. During the model parameter update phase, constraints are imposed on the update direction and magnitude of the model based on the response evaluation parameters. That is, the response evaluation parameters are embedded as regularization terms into the model's loss function. The response evaluation parameters reflect the degree of output change before and after the perturbation and serve as control terms to limit the magnitude of weight updates, so that the output differences caused by perturbations of training feature fields during the training process meet the control standards.

[0013] In a preferred embodiment, the logic for putting the trained screening model into use is as follows: For the current applicant group, non-sensitive attribute fields are extracted, and candidate ability expression vectors are generated based on the semantic mapping relationship between job ability tags and fields. These vectors are used as input to the screening model to output the original recommendation score. Based on the set recruitment criteria, a preliminary group of candidates for recruitment is selected. The preliminary candidate group is divided into groups according to preset attribute fields, and the occurrence ratio of different value categories of each attribute field in the group is calculated; the calculated attribute distribution results are compared with the reference distribution of the corresponding attribute fields in the auxiliary verification sample set to obtain the recommendation deviation value of each attribute category. By combining the recommendation deviation value with the set tolerance range, an adjustment factor proportional to the recommendation deviation is generated for categories that exceed the tolerance range, and a set of adjustment factors is constructed. In the recommendation probability output stage, the original recommendation score of each candidate is weighted by the adjustment factor corresponding to its attribute category to obtain the adjusted final recommendation score, so as to achieve consistent control of the recommendation probability distribution among attribute groups. Under the same position, the candidate group is re-sorted according to the final recommendation score, and the final list of recruits is determined according to the set number of recruits or recommendation ratio.

[0014] In a preferred embodiment, the human resource intelligent information screening system based on data analysis specifically includes: The difference calculation module is used to extract attributes and recruitment results from historical candidate samples. By performing statistical calculations on the joint distribution between attributes and recruitment results, a set of data mapping parameters expressing differences in recruitment preferences is obtained. This set is then used to construct data offset weights to identify the sources of difference in the training data. The sample partitioning module is used to group training samples according to data offset weights, and to classify data whose offset weight values ​​exceed the limit range into the auxiliary verification sample set. The feature removal module is used to identify the statistical correlation index between training feature fields and sensitive label fields in the training sample set formed after removing samples with excessive data offset weights, and remove training feature fields whose correlation values ​​exceed the preset judgment standard range from the training sample set. The remaining training feature fields are used as a set of non-sensitive attribute fields that do not have sensitivity risks. The capability generation module is used to generate candidate capability expression vectors by vector mapping based on the semantic correspondence between job capability requirements and non-sensitive attribute fields in the retained training feature fields, and to build a model sample set with a general capability structure for the job. The model training module is used to select and train models based on the model sample set, and introduces a perturbation contrast control mechanism to dynamically constrain the model's response magnitude to changes in input attributes during the training process. The results generation module is used to generate candidate ability expression vectors for the current applicant group using the trained screening model, and to perform matching calculations based on the job standard ability structure. The results calibration module is used to compare and verify the screening output results with the auxiliary verification sample set, and to perform screening result calibration in combination with the output probability distribution, so that the final output maintains the consistency of the recommended probability distribution among different attribute groups.

[0015] The technical effects and advantages of this invention are as follows: This invention, by constructing a set of data mapping parameters and calculating data offset weights, can accurately identify the sources of attributes in training data that may cause differences in hiring preferences, enabling the model to proactively perceive and control structural biases in the early stages of training. During the training sample generation phase, this invention performs statistical analysis on historical candidate samples, extracts the joint distribution relationship between attributes and hiring results, quantifies it through the degree of offset, and forms a practically operable basis for characterizing bias, thus constructing data offset weights. This design not only helps identify historical patterns that may indicate uneven hiring but also groups training samples based on offset weights, assigning potentially biased samples to an auxiliary validation sample set, thereby preventing them from directly participating in model training. Through this mechanism, this invention performs quality screening of model input at the data level, preventing historical bias information from entering the main training process and ensuring the fairness and representativeness of the model's learning foundation from the source.

[0016] This invention performs statistical correlation analysis between training feature fields and sensitive label fields before training the screening model, and removes fields that exceed the judgment criteria, effectively reducing the model's structural dependence on sensitive attributes and ensuring the fairness of the model's decision-making process from the feature space level. After retaining the training samples, this invention further quantifies the association strength between training feature fields and sensitive label fields using statistical correlation indicators. By setting reasonable judgment criteria, fields with correlation values ​​higher than expected are removed from the training set, filtering out non-sensitive attribute fields without sensitivity risks, forming a controllable, usable, and interpretable training dataset. Simultaneously, based on the semantic correspondence between these non-sensitive attribute fields and job competency requirements, a candidate competency expression vector is constructed, making the model's learning objective more focused on job competency rather than implicit features related to personal background. This avoids model bias due to learning biased fields and effectively improves the compliance and social acceptability of the screening recommendation.

[0017] This invention introduces a perturbation contrast control mechanism during model training and combines it with an auxiliary validation sample set for screening result calibration during the inference output stage. This dual-path approach ensures that the model maintains accuracy in identifying job capabilities while achieving consistency in recommendation probabilities among different attribute groups, significantly improving the structural fairness and reliable stability of the model's output. During model training, the perturbation contrast control mechanism applies structured perturbations to the training feature fields and learns the differences between the capability expression vectors formed by the original input and the perturbed input. This guides the model to dynamically suppress over-response to minor changes in input attributes, improving its robustness to interference. During the model inference stage, this invention uses the pre-defined auxiliary validation sample set as an evaluation benchmark. By comparing the distribution of group attributes in the preliminary screening results, a recommendation bias value is generated, and a set of adjustment factors is constructed to structurally adjust the output recommendation score. This combined strategy ensures fairness control during model training and achieves consistency in the distribution of group recommendation probabilities through result calibration, simultaneously improving the fairness and transparency of human resource screening from both the model mechanism and the final effect perspectives. Attached Figure Description

[0018] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of the intelligent information screening method for human resources based on data analysis in this invention.

[0019] Figure 2 This is a schematic diagram of the intelligent information screening system for human resources based on data analysis in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 - Figure 2 The following examples were obtained: Example 1: A data analysis-based intelligent information screening method for human resources, comprising the following steps: By extracting attributes and hiring results from historical candidate samples, and statistically calculating the joint distribution between attributes and hiring results, a set of data mapping parameters expressing differences in hiring preferences is obtained. This set is then used to construct data offset weights to identify the sources of difference in the training data. By constructing a joint statistical mapping between historical hiring behavior and attribute characteristics, the degree of hiring preference of candidates under different attribute combinations is characterized, thus providing a quantitative basis for identifying potential bias factors in the training data. Data offset weights, as a benchmark indicator for selecting training samples, help to identify structural biases that affect the fairness of the model.

[0022] Based on the data offset weights, the training samples are grouped, and data with offset weight values ​​exceeding the limit range are assigned to the auxiliary validation sample set. By distinguishing the numerical range of the offset weights, training samples that may have abnormal distribution due to historical bias are removed from the main training set and used for subsequent model performance validation. This prevents the model from absorbing biased samples during the training phase, thereby improving the structural fairness and representativeness of the training data source.

[0023] After removing samples with excessive data offset weights from the training sample set, statistical correlation indicators between training feature fields and sensitive label fields are identified. Training feature fields with correlation values ​​exceeding the preset judgment criteria are removed from the training sample set. The remaining training feature fields are designated as the set of non-sensitive attribute fields without sensitivity risks. A structural review of the training sample fields is conducted, and statistical analysis is used to eliminate training fields highly correlated with sensitive attributes. This reduces the model's dependence on socially sensitive attributes during the learning process, ensuring that the model's input dimensions are structurally neutral at the feature level, which is conducive to generating fair and robust expression vectors. Sensitive label fields are a preset set of fields in the system. Alternatively, the probability distribution of attribute fields in the auxiliary verification sample set can be compared with the distribution of the same attribute field in the entire sample set. When the probability of an attribute field appearing in the verification set exceeds a set threshold for the probability of the corresponding attribute field appearing in the entire set, it is considered a candidate sensitive label field, and its inclusion in the sensitive label field set can be confirmed according to system settings.

[0024] In the retained training feature fields, based on the semantic correspondence between job competency requirements and non-sensitive attribute fields, candidate competency expression vectors are generated through vector mapping to construct a model sample set with a general competency structure for the job. By establishing semantic associations between job requirements and fields, this step transforms the original training data into expression vectors with a unified competency structure. This allows the model to understand job competency without relying on the sensitive attribute dimension during the learning process, thus constructing a standardized candidate expression method based on actual competency requirements and improving the model's accuracy in judging job matching.

[0025] Model training is performed using a sample set of selected models, and a perturbation contrast control mechanism is introduced to dynamically constrain the model's response magnitude to changes in input attributes during training. By introducing perturbation inputs and responses during model training, a set of response constraint paths is constructed to ensure that the model's output changes are dynamically controlled when faced with field perturbations. This mechanism effectively suppresses the model's tendency to overfit to potentially sensitive paths, thereby enhancing the model's fairness and stability across multiple populations and reducing the risk of structural discrimination.

[0026] Using a trained screening model, candidate ability expression vectors are generated for the current applicant pool, and matching calculations are performed based on the job's standard ability structure. The screening output results are compared and verified using an auxiliary validation sample set, and the screening results are calibrated based on the output probability distribution to ensure consistency in the recommendation probability distribution across different attribute groups. This step constitutes the model deployment and application phase. Candidates are scored for job matching after expressing themselves through non-sensitive fields, and bias comparisons are performed using the auxiliary validation sample set to achieve structural calibration of the recommendation results. By introducing a readjustment operation for recommendation scores, the final recruitment output maintains statistical consistency across attribute groups, improving the social fairness and algorithm credibility of the overall screening process.

[0027] In one specific embodiment of the present invention, the data mapping parameter set refers to a set of parameters that express the pattern of differences in recruitment, generated by performing distribution analysis, recruitment result statistics, and recruitment tendency scoring on historical candidate samples. The construction process includes the following steps: classifying and aggregating historical candidate samples to construct a set of attribute combinations containing multiple preset dimensions, and annotating and statistically analyzing the recruitment results corresponding to each attribute combination; wherein the attributes include common human resource evaluation dimensions such as gender, age range, education level, professional category, work experience range, and job category; by combining the values ​​of each candidate on each dimension, multiple attribute combinations with composite feature expression capabilities are formed; and based on this, the number of recruitment results corresponding to each attribute combination is statistically analyzed, thereby obtaining basic statistical data reflecting the differences in performance of different attribute combinations in the historical recruitment process. This statistical data provides data support for subsequent conditional recruitment probability calculations.

[0028] Based on the attribute combinations obtained from classification and aggregation, and the admission results, the conditional admission probability of each attribute dimension under different combinations is obtained, and a joint probability matrix between each dimension and the admission results is generated to characterize the overall distribution pattern. The classification and aggregation method is not specified, as it is a prior art well known to those skilled in the art and will not be elaborated upon. Each conditional admission probability is obtained by calculating the ratio of the number of admissions corresponding to a certain attribute combination to the total number of candidates under that combination. The joint probability matrix is ​​generated by arranging the conditional admission probabilities of all attribute dimensions in dimensional order, thus forming a multidimensional statistical structure that can reflect the joint distribution relationship between attributes and admission results. This statistical structure is used to capture the overall performance shift of different attribute combinations in the admission process, providing a complete distribution expression for subsequent difference calculation.

[0029] A discrepancy analysis was performed between the joint probability matrix and the ideal equilibrium distribution. The distribution offset distance measure was used to score each attribute combination to obtain a hiring preference score. The ideal equilibrium distribution describes the hiring probability distribution that each attribute combination should exhibit under conditions of no structural bias. The distribution offset distance measure is used to calculate the difference between the hiring probability of each attribute combination in the joint probability matrix and the corresponding position in the ideal equilibrium distribution. This difference is quantified by summing the absolute differences or by accumulating the difference intervals in segments, thus obtaining the degree of deviation of each attribute combination from a fair hiring state. The greater the deviation, the more likely the attribute combination is to have been structurally influenced in historical hiring behavior, and therefore the higher the hiring preference score. This calculation method transforms raw statistical data into a quantitative indicator that reflects the hiring preference of attribute combinations.

[0030] After obtaining the hiring preference score, a data mapping parameter set is generated to express the differences in hiring preferences. This set associates and stores each attribute combination with its corresponding hiring preference score, which is used to extract the corresponding hiring preference score during the risk assessment of subsequent training samples and serves as the basis for determining the degree of bias in candidate samples. Through this data mapping parameter set, this invention can effectively identify the sources of abnormal hiring biases during the training data source stage, providing a reliable data foundation for subsequent data bias weight construction, auxiliary validation sample set partitioning, and fair screening model training. This ensures that the model has the ability to identify historical biases and suppress unfair structures during the overall training process.

[0031] In one specific embodiment of the present invention, risk assessment refers to extracting the hiring tendency features of each candidate sample from the data mapping parameter set, comprehensively correcting them by combining distribution density and attribute sensitivity, and finally generating data offset weights for subsequent training sample screening and grouping. The implementation process includes the following steps: For each candidate sample's attribute combination, extract the corresponding hiring tendency score from the data mapping parameter set and compare it with the global average hiring score to calculate the relative offset value. The candidate sample's attribute combination includes a joint expression composed of values ​​across multiple dimensions such as gender, age range, education level, professional category, work experience range, and job category. By matching these values ​​with each attribute combination recorded in the data mapping parameter set, the hiring tendency score corresponding to that combination is extracted. This hiring tendency score represents the probability of that attribute combination being hired in historical samples. Then, the hiring tendency scores of all candidate samples are averaged to obtain the global average hiring score. The difference between each candidate sample's score and this average is calculated to obtain a value used to measure the relative offset of its hiring probability, i.e., the relative offset value.

[0032] Based on the distribution density of relative offset values ​​in the attribute space, an attribute sensitivity factor is introduced to weight and correct the offset values, thereby reflecting the sensitivity of a sample to the overall distribution and its impact on the admission results. The distribution density of the attribute space refers to the frequency of occurrence of a particular attribute combination in all samples and the representativeness of that attribute dimension across all dimensions. The weighted offset values ​​are then normalized to generate numerical data offset weights corresponding to each candidate sample, serving as the basis for subsequent training sample selection and grouping. The specific normalization process includes: determining the minimum and maximum values ​​based on the set of weighted corrected offset values ​​for all samples, and performing interval scaling on the weighted corrected offset value of each sample to map it to a preset standard numerical range. The normalized data offset weights retain the relative ranking of the offset degree and provide a unified dimensional basis for each sample weight, facilitating uniform threshold selection during training data processing.

[0033] By binding data offset weights to candidate samples, a training sample set with difference-determining attributes is formed. In subsequent steps, this weight is used as the partitioning criterion to identify candidate samples whose offset weight values ​​exceed a certain limit, and these are then assigned to the auxiliary validation sample set. This process ensures that samples that might amplify historical biases are removed from the training sample set, preventing the inheritance and reinforcement of potential discriminatory structures during model learning, while retaining representative data for fairness detection and model calibration. This risk assessment method constructs a quantitative path for identifying recruitment biases from historical data, providing a stable and effective support mechanism for the structural fairness objective of this invention.

[0034] In one specific embodiment of the present invention, the calculation process of the attribute sensitivity factor is as follows: For each candidate attribute dimension, the frequency of the attribute appearing in all training samples is counted, and its proportion in the overall candidate group is calculated to measure the representativeness of the attribute in the dataset. For example, in a certain training dataset, there are multiple attribute dimensions such as gender, age, education, major, and years of work experience. The "gender" dimension includes two value categories: "male" and "female." Statistics show that "male" samples appear 700 times, and "female" samples appear 300 times. The total number of candidates is 1,000. Then, the frequency of the "gender" attribute dimension appearing in the samples is 1,000 times, accounting for the complete proportion of the total attribute combination. The occurrence proportions of the single category are 70% and 30%, respectively. This proportion can be used to measure the representativeness of the attribute dimension in the dataset. The stronger the representativeness, the more significant the impact of the attribute on the overall bias.

[0035] The variance of the difference in hiring probabilities between each category and the hiring results under this attribute dimension is calculated to assess the degree of fluctuation of each category under the two outcomes of being hired and not being hired. For example, in the "Education Level" dimension, there are three value categories: "High School," "Bachelor's," and "Master's." Among the samples of the "Bachelor's" category, 60% were hired; among the samples of the "Master's" category, 80% were hired; while the hiring rate of the "High School" category was only 20%. Therefore, the hiring probabilities of different categories under the "Education Level" dimension differ significantly. By calculating the squared difference between the hiring probability of each category and the average hiring probability of this dimension, and taking the average, the variance of this attribute dimension is obtained. The larger the variance, the higher the sensitivity of this attribute dimension to the hiring results.

[0036] The aforementioned variance values ​​are processed according to a preset normalization rule and used as an attribute sensitivity factor to measure the sensitivity of the attribute in the training samples for recruitment. The normalization rule involves mapping the variance values ​​of each attribute dimension to a range, using the maximum and minimum values ​​of the variance across all attribute dimensions as boundaries, ensuring that the sensitivity factor is within a uniform weight range, for example, between 0% and 100%. The normalized sensitivity factor balances representativeness and recruitment volatility, reflecting both the distribution strength of the attribute in the dataset and quantifying the risk of discrepancies in recruitment results, becoming a key parameter for subsequent bias-weighted correction. The attribute sensitivity factor constructed using this method enables precise modeling of multidimensional human resource attributes in determining structural biases, thus providing a solid parameter foundation for achieving highly fair data screening and model training.

[0037] In practical implementation, to further enhance the model's ability to identify and intervene in sensitive samples during the training phase, samples with significant shift characteristics can be uniformly included in an auxiliary validation sample set, serving as a structural risk control set distinct from the model's main training data, based on the constructed data shift weights. This process specifically includes the following steps: constructing a sample shift distribution sequence based on the normalized data shift weights corresponding to each candidate sample calculated previously, and arranging all training samples in ascending order of shift weight values ​​to form a continuous distribution interval covering the overall sample shift distribution characteristics. In practical applications, the cumulative density curve of the shift weights within the percentile intervals can be used for visualization to determine the boundary of the sample's shift distribution in the overall data structure.

[0038] Based on the sensitivity of the training task to model fairness requirements, upper and lower boundary values ​​are set for the offset limit interval. For example, in a specific job scenario, samples with offset weights below 5% or above 95% of the overall distribution are defined as an anomaly interval, and this interval boundary is used as the sample partitioning benchmark. Further, all samples with offset weights exceeding this interval are screened from the sample offset distribution sequence and identified as the anomaly offset sample set. These identified anomaly offset samples are removed from the original training sample set and uniformly incorporated into the auxiliary validation sample set. This set does not participate in the main model training process but is used in subsequent stages to verify the stability of the model's response to the screening results under different offset structures and the fairness of the recommendation. This method achieves proactive screening of the training sample structure, effectively preventing structural biases from being absorbed by the model learning.

[0039] Before constructing the candidate ability representation vector, to ensure that the feature fields in the input training samples do not interfere with the fair learning process of the screening model due to excessive correlation with the sensitive label fields, correlation identification and removal operations can be performed on the training feature fields retained in the training samples. The quantification process is achieved using correlation index calculation, including the following steps: Extracting a set of field pairs between training feature fields and sensitive label fields from the training sample set, and constructing a joint distribution structure of fields for all training samples. For example, training feature fields may include "years of work experience," "number of awards," "major," "project experience," etc., while sensitive label fields may include "gender," "ethnicity," "age group," "educational level," etc. By constructing an observation sample set between all field pairs, a distributional foundation is laid for subsequent calculations. For each of the above field pairs, one of the following methods—mutual information calculation, Pearson correlation coefficient calculation, or Spearman rank correlation coefficient calculation—is selected to quantify the dependence between the training feature fields and the sensitive label fields. Taking the Pearson correlation coefficient as an example, if the linear correlation coefficient between the "Number of Project Experiences" field and the "Gender" field exceeds one-third of the range, it can be preliminarily determined that the field may have a strong structural relationship with gender and should not be directly used in model training.

[0040] Based on the constraint strength of fairness control imposed by the training task, a judgment standard range for relevance indicators is set. For example, in the screening of positions where sensitive interventions need to be avoided, a threshold of one-third or more of the absolute value of the relevance indicator is used for elimination. The calculation results for all fields are sorted, and training feature fields with relevance values ​​exceeding the set standard are identified as fields that may lead to structural biases in the model. Training feature fields with relevance indicators exceeding the judgment standard range are removed from the training sample set and are no longer used as the basis for generating candidate ability expression vectors. Simultaneously, the remaining fields are marked as non-sensitive attribute fields and participate in subsequent semantic matching and job ability mapping. Through this approach, a structural feature review mechanism based on training samples is effectively implemented, ensuring the suppression of interference from sensitive dimensions during the model training phase and providing technical support for achieving fairness in recommendations within positions.

[0041] In constructing the competency expression vector, to ensure that candidates can be expressed and compared in a consistent manner across different attribute groups, a stable and quantifiable competency structure mapping mechanism needs to be established based on the semantic correspondence between training feature fields retained in the training sample set and job competency labels. Specifically, this includes the following steps: performing semantic analysis on all field names identified as non-sensitive attribute fields in the training sample set and matching them with a pre-defined set of job competency labels. Job competency labels may include standard competency dimensions such as "logical thinking," "technical mastery," "teamwork," "language expression," and "cross-departmental communication." The semantic relationship between each field name and the job competency label is measured in three ways: first, calculating the semantic similarity between word vectors; second, comparing the degree of semantic overlap based on a keyword dictionary; and third, identifying the functional overlap between fields and labels by referring to semantic matching rules in the job competency label library. The semantic relationship between each field name and all job competency labels can be quantitatively characterized using any of these methods. To improve the stability and interpretability of the scoring, different semantic similarity scores can be pre-set for word vector distance, keyword overlap, or label rule matching levels during the scoring process.

[0042] Based on semantic similarity scores, a set of fields with semantic similarity scores exceeding a set similarity standard is selected for each job competency label, and a one-to-one or many-to-one field label mapping relationship is established. In most cases, each competency label corresponds to one or two high-scoring fields to ensure that the mapping relationship has both interpretability and generalization ability. If multiple fields have similar scores, fields with stronger information completeness, broader time coverage, or closer alignment with the competency requirements in the job description are prioritized, such as establishing a priority mapping between the "leadership project experience" field and the "teamwork" competency label. Based on the above mapping relationship, the values ​​of non-sensitive attribute fields are converted into a standard expression format under the job competency dimension, constructing a competency expression vector under the job competency dimension. This competency expression vector has a fixed dimensional structure, ensuring that candidate features are understood and processed in a uniform way in the model, thus providing structurally fair input semantic representation during subsequent model training, and also providing an expressive foundation for introducing perturbation control mechanisms and output fairness calibration in the later stages of the model.

[0043] After the candidate competency expression vectors are constructed and the model sample set is formed, to prevent the screening model from becoming overly sensitive to changes in specific fields during training, thus preventing the propagation of structural bias, a perturbation contrast control mechanism is introduced during the model training phase. This mechanism dynamically constrains the model's response amplitude to input field perturbations, thereby controlling the fairness bias of the screening process. The implementation of this mechanism may include the following steps: In the model sample set, structural perturbation processing is performed on the training feature fields retained in the training samples to construct a set of sample pairs that correspond one-to-one between the original competency expression vectors and the perturbed competency expression vectors. Perturbation processing includes introducing slight changes to numerical fields while maintaining semantic stability, performing synonym replacement or partial masking on textual fields, and randomly sampling categorical fields from the set of alternatives. For example, if "project leadership experience" is a field with an original value of five years, it can be perturbed to four or six years to form a new sample. The perturbed competency expression vectors maintain structural consistency but have limited differences in content from the original vectors.

[0044] The original capability representation vector and its corresponding perturbation capability representation vector are input into the screening model, and during model training, the vector representation outputs of the two types of inputs at key intermediate layers are recorded. The hidden representation of a certain layer in the model is used as a measure of the degree of response to input features. By calculating the distance difference between the original input and the perturbation input at that layer's output, a response evaluation parameter is formed to quantify the model's sensitivity to input perturbations. For example, if a perturbation of a field causes the output difference to exceed twice the standard deviation of the average perturbation response, that field is identified as a high-risk response item. The response evaluation parameter is compared with preset control standards to determine whether there is a risk of model over-response. The control standards include the maximum tolerance of response amplitude, the balance constraint of the perturbation field, and the standard distribution requirement of the output difference between sample pairs. If the evaluation parameter exceeds the standard, the value is used as a feedback signal during the training phase to correct and control the model update behavior.

[0045] During the training process of the selection model, after pairing the original capability representation vector with the perturbation capability representation vector and obtaining the intermediate output difference, the response evaluation parameter formed by this difference needs to be further introduced into the model's loss calculation to achieve structural fairness control during the training phase. To control the model's sensitivity to perturbations in training feature fields, the response evaluation parameter can be used as a moderating factor in the model parameter update process in the following way: For each pair of samples input into the model, a distance metric is used to measure the intermediate output representation. Specifically, the vector output is extracted from the key representation layer of the model, the sum of squared differences of the vectors before and after perturbation across all dimensions is calculated, and the square root of the result is taken to obtain the absolute value response evaluation parameter representing the degree of response of the sample to the perturbation. The larger the parameter value, the more sensitive the model is to perturbations in that field, and vice versa. To ensure that this parameter participates in the overall model training process, it needs to be introduced as a moderating factor into the original model training objective function. The original objective function typically includes an error measure between the model output and the actual sample label. To introduce perturbation response control, a moderating term is added to the original error measure. This adjustment term is the sum of the response evaluation parameters for all training sample pairs, or the average value calculated as needed by the model, forming the final loss value. For example, suppose there is a set of training sample pairs where the intermediate output differences between the original input and the perturbation input are three, four, and five, respectively. Without the adjustment term, the model update is only affected by the prediction error; however, after introducing this adjustment term, the model will also consider how to reduce these differences, thereby forcing the model to mitigate the amplification effect of small changes in the input field when updating parameters.

[0046] In each training epoch, the model parameters are backpropagated and adjusted based on the updated overall loss function value. Since the adjustment term is included in the loss calculation, the model will optimize the update direction to converge the response evaluation parameters to a lower range, ultimately suppressing the magnitude of field perturbations. To avoid the adjustment term interfering with the model's primary training objective, further optimization can be implemented: the weight of this term is gradually increased in the initial training phase. Specifically, a progress factor for the training epochs is set, with the adjustment term having a lower weight in the initial stage and gradually increasing to a level comparable to the prediction error as the training epochs progress. This achieves a dynamic transition from accuracy-first to fairness-first, ensuring that the final model has both good screening accuracy and structural fairness constraints.

[0047] In this implementation, after the screening model completes training, it enters the actual screening stage, making screening decisions for the current applicant group to ensure the structural fairness of the recommendation results. This process includes the following steps: For the current applicant group, non-sensitive attribute fields are extracted for each candidate. These non-sensitive attribute fields are the set of fields remaining in the model sample set after sensitive labels have been removed, such as professional direction, years of work experience, job experience, and project capabilities. Subsequently, the non-sensitive attribute fields are matched one-to-one with preset job capability labels, generating candidate capability expression vectors based on the semantic mapping relationship between the job capability labels and the fields. For example, when the job capability label is "project coordination ability," if a candidate has the field "experience in managing multiple cross-departmental projects" and a word vector similarity score of 90 or higher, a mapping relationship is established. Thus, each candidate can form a set of capability expression vectors as input. The trained screening model calculates and outputs an initial recommendation score, and a preliminary candidate group is selected based on the set recruitment criteria, which can be set to a recommendation score higher than 80.

[0048] The initial pool of candidates is divided into groups based on predefined attribute fields. Each attribute field is a categorical attribute retained during the initial model training phase but not used as a sensitive label, such as gender, ethnicity, and age group. After segmentation, the different value categories of each attribute field are statistically analyzed, and their proportion in the current initial pool of candidates is calculated. For example, if the attribute field is gender, males account for 70% and females for 30% of the initial pool; if it is ethnicity, a certain minority group accounts for only 5%. Then, this distribution result is compared with the reference distribution of the corresponding attribute field in the auxiliary validation sample set. The reference distribution of the auxiliary validation sample set serves as the baseline for fair distribution, and the recommendation bias value for each attribute category is calculated. The recommendation bias value is the difference in proportion between the two distributions; for example, the bias value for males is positive 20%, and the bias value for females is negative 20%.

[0049] The recommendation deviation value for each attribute category is compared with a set tolerance range, which can be set to ±10%. If the recommendation deviation value exceeds the tolerance range, the recommendation result is considered structurally inconsistent. A set of adjustment factors is constructed based on this deviation value. These adjustment factors are used to adjust the original recommendation score. The construction rule is as follows: the absolute value of the deviation value is linearly converted into a percentage weight, and the corresponding adjustment factor is constructed. For example, if the recommendation deviation value for a certain female group is negative 20%, and the adjustment factor ratio is set to 0.2, all candidates in this group will receive a positive correction in their final recommendation score, increasing their final recommendation score by the corresponding weight ratio. Each adjustment factor is bound to a corresponding attribute category, forming a set of adjustment factors used for subsequent recommendation result correction.

[0050] In the recommendation probability output stage, the original recommendation score for each candidate is weighted and calculated by the adjustment factor corresponding to its attribute category, generating the adjusted final recommendation score. A linear weighted model can be used, where the final recommendation score equals the original score multiplied by the adjustment coefficient. For example, if the original recommendation score is 85 and the adjustment factor is 1.1, the final recommendation score is 93.5. Based on the final recommendation scores, candidates are re-ranked for the same position. The final list of successful candidates is determined according to the set number of openings or recommendation ratio for each position. For example, if a position has five openings, the top five candidates with the highest final recommendation scores are the final successful candidates. This ensures consistent control of the recommendation probability distribution among different attribute groups while guaranteeing job competency matching, thus improving overall structural fairness.

[0051] Example 2: A data-driven intelligent information screening system for human resources, specifically including: The difference calculation module is used to extract attributes and recruitment results from historical candidate samples. By performing statistical calculations on the joint distribution between attributes and recruitment results, a set of data mapping parameters expressing differences in recruitment preferences is obtained. This set is then used to construct data offset weights to identify the sources of difference in the training data. The sample partitioning module is used to group training samples according to data offset weights, and to classify data whose offset weight values ​​exceed the limit range into the auxiliary verification sample set. The feature removal module is used to identify the statistical correlation index between training feature fields and sensitive label fields in the training sample set formed after removing samples with excessive data offset weights, and remove training feature fields whose correlation values ​​exceed the preset judgment standard range from the training sample set. The remaining training feature fields are used as a set of non-sensitive attribute fields that do not have sensitivity risks. The capability generation module is used to generate candidate capability expression vectors by vector mapping based on the semantic correspondence between job capability requirements and non-sensitive attribute fields in the retained training feature fields, and to build a model sample set with a general capability structure for the job. The model training module is used to select and train models based on the model sample set, and introduces a perturbation contrast control mechanism to dynamically constrain the model's response magnitude to changes in input attributes during the training process. The results generation module is used to generate candidate ability expression vectors for the current applicant group using the trained screening model, and to perform matching calculations based on the job standard ability structure. The results calibration module is used to compare and verify the screening output results with the auxiliary verification sample set, and to perform screening result calibration in combination with the output probability distribution, so that the final output maintains the consistency of the recommended probability distribution among different attribute groups.

[0052] It should be noted that before performing operations such as constructing data offset weights, calculating attribute sensitivity factors, identifying sensitive label fields, and calibrating the screening output on the candidate training samples in this invention, a set of standard attribute requirements for each position must be established in advance. This set is used to identify attribute fields with clear hard restrictions in the recruitment requirements for that position. This set of fields serves as a rigid attribute constraint set for the position, and the attribute fields it contains do not participate in the subsequent bias identification and structural fairness adjustment process. In other words, if a position explicitly requires male candidates, the "gender" field should not be identified as a sensitive field that may cause discrimination risks, nor should it participate in the construction of sensitivity factors. Similarly, if the minimum requirement for a position is a master's degree, then the "education" field is the hard screening criterion for the position and is exempted from subsequent data perturbation processing, model fairness constraints, and output recommendation score calibration. This mechanism enables precise data correction and fairness control while ensuring that the basic recruitment requirements for positions are not misjudged as structural biases. This ensures that the model output achieves a balanced recommendation probability for different attribute groups within the "candidates who meet the job requirements," effectively preventing job matching errors caused by generalization correction. The model achieves structural fairness while maintaining a high degree of consistency and adaptability to the job's business requirements.

[0053] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0054] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0055] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0056] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0057] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data analysis-based intelligent information screening method for human resources, characterized in that, Includes the following steps: Attributes and hiring results are extracted from historical candidate samples. By performing statistical calculations on the joint distribution between attributes and hiring results, a set of data mapping parameters expressing differences in hiring preferences is obtained. This set is then used to construct data offset weights to identify the sources of difference in the training data. Based on the data offset weights, the training samples are grouped, and data whose offset weight values ​​exceed the limit range are included in the auxiliary verification sample set. In the training sample set formed after removing samples with excessive data offset weights, the statistical correlation index between the training feature fields and the sensitive label fields is identified, and the training feature fields whose correlation values ​​exceed the preset judgment standard range are removed from the training sample set. The remaining training feature fields are used as a set of non-sensitive attribute fields that do not have sensitivity risks. Based on the semantic correspondence between job competency requirements and non-sensitive attribute fields, candidate competency expression vectors are generated through vector mapping in the retained training feature fields, and a model sample set with a general competency structure for the job is constructed. Model training is performed by selecting models based on a sample set, and a perturbation contrast control mechanism is introduced to dynamically constrain the model's response to changes in input attributes during the training process. Using the trained screening model, candidate ability expression vectors are generated for the current applicant group, and matching calculations are performed based on the job standard ability structure. The screening output results are compared and verified by an auxiliary validation sample set, and the screening results are calibrated by combining the output probability distribution, so that the final output maintains the consistency of the recommendation probability distribution among different attribute groups.

2. The intelligent human resource information screening method based on data analysis according to claim 1, characterized in that, The data mapping parameter set refers to: The historical candidate samples are classified and aggregated to construct a set of attribute combinations containing multiple preset dimensions, and the admission results corresponding to each attribute combination are labeled and statistically analyzed. Based on the aggregated attribute combinations and the admission results, the conditional admission probability of each attribute dimension under different combinations is obtained, and a joint probability matrix between each dimension and the admission results is generated. The joint probability matrix is ​​compared with the ideal equilibrium distribution. The distribution offset distance method is used to score each attribute combination to obtain the recruitment tendency score. Finally, a set of data mapping parameters is generated to express the differences in recruitment tendency.

3. The intelligent human resource information screening method based on data analysis according to claim 2, characterized in that, Risk assessment refers to: For each candidate sample's attribute combination, the corresponding hiring preference score is extracted from the data mapping parameter set and compared with the global average hiring score to calculate the relative offset value. Based on the distribution density of relative offset values ​​in the attribute space, an attribute sensitivity factor is introduced to weight and correct the offset values. The weighted and corrected offset values ​​are normalized to generate numerical data offset weights that correspond one-to-one with each candidate sample.

4. The intelligent human resource information screening method based on data analysis according to claim 3, characterized in that, The calculation process for the attribute sensitivity factor is as follows: For each candidate attribute dimension, the frequency of that attribute in all training samples is counted, and its proportion in the overall candidate population is calculated. The variance of the difference in hiring preference between each category and the hiring results under this attribute dimension is calculated, and then processed according to the preset normalization rule, and used as the attribute sensitivity factor.

5. The intelligent human resource information screening method based on data analysis according to claim 4, characterized in that, The logic for partitioning the auxiliary verification sample set is as follows: Based on the normalized data offset weights corresponding to each training sample, a sample offset distribution sequence is constructed and arranged from low to high values ​​to form a continuous distribution interval. Set upper and lower bounds for the offset limit interval, filter out the sample set whose values ​​exceed the range by the interval boundary, and mark them as abnormal offset samples; Abnormal offset samples are removed from the training sample set and uniformly included in the auxiliary verification sample set.

6. The intelligent human resource information screening method based on data analysis according to claim 5, characterized in that, The statistical correlation index can be calculated using any one of the following methods: mutual information method, Pearson correlation coefficient method, or Spearman rank correlation coefficient method.

7. The intelligent human resource information screening method based on data analysis according to claim 6, characterized in that, Among the retained training feature fields, the semantic correspondence between job competency requirements and non-sensitive attribute fields is first measured by semantic similarity scoring. The scoring is determined based on the word vector similarity between the field name and the job competency label, the degree of keyword dictionary matching, or the matching rules in the standard competency label library. After obtaining the scoring results, for each job competency tag, fields with semantic similarity scores exceeding the preset similarity standard are selected from the non-sensitive attribute fields, and a unique or limited number of field-competency tag mapping relationships are established.

8. The intelligent human resource information screening method based on data analysis according to claim 7, characterized in that, The introduction of a disturbance contrast control mechanism refers to: In the model sample set, the training feature fields are subjected to structured perturbation to construct a set of corresponding sample pairs of the original capability expression vector and the perturbed capability expression vector, which are used to detect the difference in the model's response to the training feature fields. The sample pairs are input to the model for screening. By comparing the difference in intermediate outputs generated by the two types of inputs in the sample pairs during training, a response evaluation parameter is formed to characterize the magnitude of the model's response. This parameter is then compared with the set control standard. During the model parameter update phase, constraints are imposed on the update direction and magnitude of the model based on the response evaluation parameters. That is, the response evaluation parameters are embedded as regularization terms into the model's loss function, so that the output differences caused by perturbations of training feature fields during the training process meet the control standards.

9. The intelligent human resource information screening method based on data analysis according to claim 8, characterized in that, The logic for putting a trained screening model into use is as follows: For the current applicant group, non-sensitive attribute fields are extracted, and candidate ability expression vectors are generated based on the semantic mapping relationship between job ability tags and fields. These vectors are used as input to the screening model to output the original recommendation score. Based on the set recruitment criteria, a preliminary group of candidates for recruitment is selected. The preliminary candidate group is divided into groups according to preset attribute fields, and the occurrence ratio of different value categories of each attribute field in the group is calculated; the calculated attribute distribution results are compared with the reference distribution of the corresponding attribute fields in the auxiliary verification sample set to obtain the recommendation deviation value of each attribute category. By combining the recommendation deviation value with the set tolerance range, an adjustment factor proportional to the recommendation deviation is generated for categories that exceed the tolerance range, and a set of adjustment factors is constructed. In the recommendation probability output stage, the original recommendation score of each candidate is weighted and calculated with the adjustment factor corresponding to its attribute category to obtain the final adjusted recommendation score.

10. A human resource intelligent information screening system based on data analysis, based on the human resource intelligent information screening method based on data analysis as described in any one of claims 1-9, characterized in that, Specifically, it includes: The difference calculation module is used to extract attributes and recruitment results from historical candidate samples. By performing statistical calculations on the joint distribution between attributes and recruitment results, a set of data mapping parameters expressing differences in recruitment preferences is obtained. This set is then used to construct data offset weights to identify the sources of difference in the training data. The sample partitioning module is used to group training samples according to data offset weights, and to classify data whose offset weight values ​​exceed the limit range into the auxiliary verification sample set. The feature removal module is used to identify the statistical correlation index between training feature fields and sensitive label fields in the training sample set formed after removing samples with excessive data offset weights, and remove training feature fields whose correlation values ​​exceed the preset judgment standard range from the training sample set. The remaining training feature fields are used as a set of non-sensitive attribute fields that do not have sensitivity risks. The capability generation module is used to generate candidate capability expression vectors by vector mapping based on the semantic correspondence between job capability requirements and non-sensitive attribute fields in the retained training feature fields, and to build a model sample set with a general capability structure for the job. The model training module is used to select and train models based on the model sample set, and introduces a perturbation contrast control mechanism to dynamically constrain the model's response magnitude to changes in input attributes during the training process. The results generation module is used to generate candidate ability expression vectors for the current applicant group using the trained screening model, and to perform matching calculations based on the job standard ability structure. The results calibration module is used to compare and verify the screening output results with the auxiliary verification sample set, and to perform screening result calibration in combination with the output probability distribution, so that the final output maintains the consistency of the recommended probability distribution among different attribute groups.

Images