Matching threshold adjustment method and system for dynamic updating of job profiles

Abstract

The present invention discloses a matching threshold adjustment method and system for dynamic updating of job profiles, which specifically relates to the field of intelligent recruitment technology. The method and system are used to solve the problem that low-quality candidate data pollutes the model and causes systematic deviation due to dynamic adjustment of thresholds on existing recruitment platforms; by obtaining the candidate skill test scores and performance data corresponding to the target job skill labels; calculating the KL divergence value of the skill test score distribution and the job benchmark curve; triggering a PID control algorithm based on the KL divergence threshold to generate a threshold correction coefficient; dynamically updating the matching threshold and generating a threshold range; adjusting the sampling weight strategy according to the deviation direction; and retraining the model to output a recommendation list, thereby achieving coordinated optimization of recruitment efficiency and model stability, and effectively suppressing the interference of data quality degradation on the algorithm benchmark.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recruitment technology, and more specifically, to a matching threshold adjustment method and system for dynamic updating of job profiles. Background Art

[0002] In existing technologies, recruitment platforms respond to market supply and demand fluctuations by dynamically adjusting the matching threshold between job requirements and candidates. When the skill requirements for a specific position surge or the talent supply is insufficient, the system usually expands the candidate pool by lowering the matching standards to ensure that job vacancies are filled in a timely manner. The threshold adjustment mainly relies on explicit indicators such as the number of resumes collected in real time and the distribution of skill tags, aiming to balance the efficiency of corporate recruitment and the accuracy of talent screening.

[0003] However, dynamic threshold adjustment will cause a systematic shift in the distribution of candidate abilities. When the system relaxes the requirements for specific skill dimensions due to insufficient supply, a large number of low-quality candidates enter the recommendation list and are included in the subsequent model training data, causing the algorithm's judgment criteria for true ability standards to continuously shift. This shift will further aggravate the error in threshold adjustment, forming a vicious cycle of low-quality data accumulation and degradation of matching standards, ultimately damaging the long-term reliability of the recommendation system. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a matching threshold adjustment method and system for dynamic updating of job profiles to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The matching threshold adjustment method for dynamic updating of job profiles includes the following steps:

[0007] S1. Obtain the candidate's skill test scores and post-employment performance data corresponding to the current skill tag of the target position;

[0008] S2. Calculate the KL divergence value between the distribution curve of the candidate's skill test scores within a preset period and the benchmark curve of the target position's skill requirements;

[0009] S3. When the KL divergence value exceeds the KL divergence threshold, a matching threshold correction coefficient is generated based on the PID control algorithm;

[0010] S4. Update the matching threshold of the current job skill tag according to the matching threshold correction coefficient to generate an updated threshold range;

[0011] S5. Determine a sampling weight adjustment strategy for the candidate's skill test score based on the deviation direction of the KL divergence value from the updated threshold range;

[0012] S6. Retrain the matching model based on the skill test score sampling weight adjustment strategy and output the updated candidate recommendation list.

[0013] In a preferred embodiment, the proportional coefficient of the PID control algorithm is dynamically adjusted according to the correlation coefficient between the candidate's skill test score and the performance data.

[0014] In a preferred embodiment, obtaining the candidate's skill test score and post-employment performance data corresponding to the current skill tag of the target position includes:

[0015] Obtain the target position's job description text, parse the job description text, and generate the target position's current skill tag;

[0016] Obtain the candidate's skill test scores corresponding to the current skill tags of the target position from the online assessment system;

[0017] Obtain the candidate's post-employment performance data in the position corresponding to the current skill tag of the target position from the human resources management system.

[0018] In a preferred embodiment, calculating the KL divergence value between the distribution curve of the candidate's skill test score within a preset period and the target position skill requirement benchmark curve includes:

[0019] The candidate's skill test scores are binned by skill label dimension within a preset period to generate a discretized skill test score distribution;

[0020] Dynamically adjust the bin width based on the distribution of the target job skill requirement benchmark curve to align the distribution of discretized skill test scores with the data interval of the benchmark curve;

[0021] Recalculate the discretized skill test score distribution based on the adjusted bin width and extract the kurtosis and skewness parameters of the benchmark curve;

[0022] The distribution function of the benchmark curve is modified based on the kurtosis parameter and skewness parameter to generate a dynamically adapted benchmark curve for target job skill requirements.

[0023] Calculate the KL divergence value between the discretized skill test score distribution and the dynamically adapted target job skill requirement benchmark curve.

[0024] In a preferred embodiment, when the KL divergence value exceeds the KL divergence threshold, a matching threshold correction coefficient is generated based on a PID control algorithm, including:

[0025] Obtain the correlation coefficient between the candidate's skill test score and performance data. The correlation coefficient is calculated using the Pearson product-moment correlation coefficient formula;

[0026] Dynamically adjust the proportional coefficient of the PID control algorithm based on the degree of deviation between the correlation coefficient and the preset proportional reference value;

[0027] The integral term is calculated based on the difference between the KL divergence value and the KL divergence threshold, and the differential term is calculated based on the rate of change of the KL divergence value;

[0028] The proportional coefficient, integral term and differential term are superimposed according to the preset weights to generate the matching threshold correction coefficient.

[0029] In a preferred embodiment, the matching threshold of the current job skill tag is updated according to the matching threshold correction coefficient to generate an updated threshold range, including:

[0030] Obtain the original matching threshold of the current job skill tag and the matching threshold correction coefficient generated in step S3;

[0031] Linearly adjust the original matching threshold according to the matching threshold correction coefficient to generate a preliminary adjustment threshold;

[0032] When the initial adjustment threshold exceeds the preset threshold range, it is truncated according to the preset threshold range boundary value to generate a compliance adjustment threshold;

[0033] Generate an updated threshold range based on the compliance adjustment threshold and the priority weight of the target position skill tag;

[0034] Write the updated threshold range into the job profile database and trigger the threshold effective timestamp record.

[0035] In a preferred embodiment, determining a sampling weight adjustment strategy for a candidate's skill test score based on the deviation direction of the KL divergence value from the updated threshold range includes:

[0036] Determine the deviation direction of the KL divergence value from the updated threshold range. The deviation direction is divided into positive deviation from the upper threshold limit and negative deviation from the lower threshold limit.

[0037] When the deviation direction is a positive deviation towards the upper threshold limit, the sampling weight of the candidate skill test scores that are higher than the current upper threshold limit is reduced;

[0038] When the deviation direction is a negative deviation from the lower threshold, the sampling weight of the candidate skill test scores below the current lower threshold is increased;

[0039] Generate a dynamic sampling weight adjustment strategy based on the deviation direction type and deviation amplitude.

[0040] In a preferred embodiment, the sampling weight adjustment strategy includes a piecewise function mapping rule between the weight attenuation factor and the gain factor.

[0041] In a preferred embodiment, the matching model is retrained according to the skill test score sampling weight adjustment strategy, and an updated candidate recommendation list is output, including:

[0042] Extracting a weight decay factor and gain factor mapping table from a sampled weight adjustment strategy;

[0043] Adjusting the sampling weight distribution of the candidate's skill test scores based on the weight decay factor and the gain factor to generate a weighted training dataset;

[0044] The matching model is retrained using the weighted training dataset, and the matching model updates its parameters by minimizing the prediction error of the weighted samples;

[0045] The matching model output results are filtered according to the updated threshold range, and an updated candidate recommendation list is generated and appended with the threshold effective timestamp.

[0046] On the other hand, the present invention provides a matching threshold adjustment system for dynamic updating of job profiles, including the following modules:

[0047] Data acquisition module: used to obtain the candidate's skill test scores and post-employment performance data corresponding to the current skill tag of the target position;

[0048] Divergence calculation module: used to calculate the KL divergence value between the distribution curve of candidate skill test scores within a preset period and the benchmark curve of target position skill requirements;

[0049] Dynamic adjustment module: When the KL divergence value exceeds the KL divergence threshold, it is used to generate a matching threshold correction coefficient based on the PID control algorithm;

[0050] Threshold update module: used to update the matching threshold of the current job skill tag according to the matching threshold correction coefficient and generate the updated threshold range;

[0051] Strategy generation module: used to determine the sampling weight adjustment strategy of the candidate's skill test score based on the deviation direction of the KL divergence value and the updated threshold range;

[0052] Model optimization module: used to retrain the matching model based on the skill test score sampling weight adjustment strategy and output an updated candidate recommendation list.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. Dynamic optimization of matching thresholds and two-way control of data quality are achieved through a closed-loop feedback mechanism, effectively solving the model offset problem caused by threshold adjustment in traditional recruitment systems. By introducing the KL divergence value to quantify the degree of deviation between the candidate's ability distribution and the job benchmark in real time, and combining it with the PID control algorithm to dynamically correct the threshold parameters, the matching standard can be adaptively adjusted according to data quality feedback, avoiding error accumulation caused by fluctuations in a single indicator. The correlation between skill-performance data is embedded in the proportional coefficient adjustment logic, so that the threshold correction amplitude is strongly correlated with the candidate's actual performance, ensuring that the threshold adjustment not only responds to changes in market supply and demand, but also takes into account the accuracy of talent screening. This closed-loop control mechanism based on mathematical measurement significantly improves the rationality of threshold adjustment and the system's anti-interference ability.

[0055] 2. Through an asymmetric weight adjustment strategy and timing synchronization mechanism, the negative impact of low-quality samples on the model is suppressed at the source of the data. The sampling weights of high- and low-scoring candidates are adjusted differentially based on the direction of KL divergence deviation to proactively block data degradation paths. At the same time, the strict binding of the threshold effective timestamp and the model version is used to ensure the timing consistency of threshold updates, model training, and recommendation result output operations. The synergistic effect of various technical features ensures recruitment efficiency while maintaining the algorithm's stable judgment of true ability standards, ultimately achieving a dual improvement in talent matching accuracy and long-term system reliability, forming a technical closed loop that dynamically balances data quality and matching standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of the matching threshold adjustment method for dynamic updating of job profiles according to the present invention;

[0057] Figure 2 This is a structural diagram of the matching threshold adjustment system for dynamic updating of job profiles according to the present invention. DETAILED DESCRIPTION

[0058] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0059] Example 1: Figure 1 The present invention provides a matching threshold adjustment method for dynamic updating of job profiles, which includes the following steps:

[0060] S1. Obtain the candidate's skill test scores and post-employment performance data corresponding to the current skill tag of the target position;

[0061] S2. Calculate the KL divergence value between the distribution curve of the candidate's skill test scores within a preset period and the benchmark curve of the target position's skill requirements;

[0062] S3. When the KL divergence value exceeds the KL divergence threshold, a matching threshold correction coefficient is generated based on the PID control algorithm. The proportional coefficient of the PID control algorithm is dynamically adjusted according to the correlation coefficient between the candidate's skill test score and performance data;

[0063] S4. Update the matching threshold of the current job skill tag according to the matching threshold correction coefficient to generate an updated threshold range;

[0064] S5. Determine a sampling weight adjustment strategy for the candidate's skill test score based on the deviation direction of the KL divergence value from the updated threshold range;

[0065] S6. Retrain the matching model based on the skill test score sampling weight adjustment strategy and output the updated candidate recommendation list.

[0066] S1. Obtain the candidate's skill test scores and post-employment performance data corresponding to the current skill tag of the target position. The specific implementation is as follows:

[0067] The specific implementation method for obtaining the candidate's skill test score and post-employment performance data corresponding to the current skill tag of the target position is achieved through the following operations:

[0068] The original recruitment document for the target position is retrieved from the enterprise's human resources management platform. This document contains textual content written in natural language, including job descriptions, qualifications, and technical competences. A structured analysis is then performed on the text, including the following steps: identifying the start and end paragraph identifiers of the "Job Requirements" section of the text and extracting a set of complete sentences within that section; performing word segmentation on the sentence set, using a dictionary-based word segmentation algorithm to break the sentences into independent word sequences; filtering out stop words from the word sequences, retaining nouns and verb-object phrases containing technical competence descriptions as candidate keywords; and semantically matching the candidate keywords against a predefined standardized skill lexicon. Each skill tag in the skill lexicon is associated with at least three synonymous expressions. When "Proficient in Java language development" appears in the original text, it is normalized to the "Java skills" tag by matching it with the "Java development" tag in the lexicon, generating the current skill tag set for the target position. The skill lexicon is constructed by integrating an industry-standard skill classification framework and includes common technology stack tags and corresponding synonym mapping rules for the internet industry.

[0069] The specific implementation process for obtaining skill test scores corresponding to a candidate's current skill tags for the target position from the online assessment system is as follows: Based on the generated set of current skill tags for the target position, a skill assessment request is sent to the assessment system interface. The request includes the skill tag name and the corresponding assessment level parameter. The online assessment system selects theoretical questions and practical tasks from a pre-set question bank based on the skill tag name. For example, for the "Java skills" tag, a multi-threaded programming question and a JVM tuning scenario task are invoked. The difficulty level of the theoretical questions matches the assessment level parameter, and the design standards of the practical tasks are set based on the actual engineering requirements of the target position. After the candidate completes the assessment, the online assessment system calculates the skill test score for each skill tag according to pre-set scoring rules. The scoring rules include a dynamic proportional allocation of weights for theoretical question accuracy, code execution efficiency, and scenario task completion. The score results are stored in an intermediate database in a structured format of "candidate ID-skill tag-skill test score." A database indexing mechanism ensures data relevance for subsequent steps.

[0070] The specific steps for obtaining post-employment performance data from the human resources management system include: performing a joint query on the employee profile database in the human resources management system based on the current skill tag set for the target position. Query criteria include an exact match between the employee's current position title and the target position title, the employee's skill profile record containing the current target skill tag, and the employee's time on board exceeding the preset assessment cycle threshold. Performance data includes quantitative metrics automatically collected by the system and qualitative scores manually assessed by direct supervisors. Quantitative metrics include code submission defect rate, project delivery on-time rate, and average task completion time, while qualitative scores include technical solution innovation and teamwork skills. During data extraction, employee identity and skill tags are linked to ensure that each performance data entry corresponds to the assessment result under a specific skill tag. The data update mechanism is synchronized with the company's performance assessment cycle. After each performance assessment, a data synchronization interface is triggered, and the latest performance data is written to the skill-performance association database in the format of "employee ID-skill tag-performance data."

[0071] The associated storage of skill test scores and performance data is implemented by establishing a skill-performance mapping table in a relational database. The mapping table contains a candidate unique identification field, a skill tag field, a skill test score field, and a performance data field. The associated storage process is completed through the following steps: the skill test score file exported by the online assessment system is combined into a joint primary key according to the candidate ID and skill tag; the performance data file exported by the human resources management system is matched according to the same joint primary key; for successfully matched records, the skill test score and performance data are merged and written into the mapping table; for unmatched records, exception logging is performed and a manual review process is triggered. A foreign key constraint is set for the skill tag field in the database table structure design. The foreign key constraint points to the current skill tag set of the target position to ensure the consistency of the skill tags. When a skill tag mismatch or missing data is detected during the writing process, the exception handling process is triggered and a log record is generated. The administrator manually reviews the data source and re-executes the associated storage operation.

[0072] S2. Calculate the KL divergence between the distribution curve of the candidate's skill test scores within the preset period and the benchmark curve of the target position's skill requirements. The specific implementation is as follows:

[0073] The specific process of binning candidate skill test scores according to skill label dimensions within a preset period to generate a discretized skill test score distribution is as follows: extract the candidate skill test score data corresponding to the current skill label of the target position from the associated storage database of step S1, and classify and organize them into independent data sets according to the skill label name. Perform data preprocessing on the skill test scores under each skill label. The preprocessing includes removing data items that exceed the preset valid range and filling missing values. The preset valid range is set according to the historical data distribution characteristics of the skill label. For example, the upper limit of the skill test score is 1.2 times the theoretical maximum score, and the lower limit is 0.8 times the theoretical minimum score. After preprocessing is completed, the skill test scores are discretized using the equal-width binning method. The initial bin boundaries are dynamically set according to the historical score distribution characteristics of the skill label. When the skill test scores are normally distributed, uniform binning is used. When the score distribution is obviously skewed, the number of bins is increased in the dense interval. After the binning is completed, the proportion of candidates in each interval is counted to generate a discretized skill test score distribution table containing the binning intervals, frequency numbers, and proportions. The number of binning intervals does not exceed the preset upper limit, which is set according to the scoring granularity requirements of the skill label. For example, when the theoretical maximum score is 100 points, the upper limit is 10 bins.

[0074] The specific process for dynamically adjusting the bin width based on the distribution of the target position skill requirement benchmark curve to align the discretized skill test score distribution with the data range of the benchmark curve is as follows: The initial discretized skill test score distribution table generated in step S2 and the preset target position skill requirement benchmark curve are obtained. The distribution characteristics of the benchmark curve are analyzed. When the benchmark curve exhibits a high-density distribution in a specific score range, the bin width of the corresponding range is reduced to 50% of the original width to increase the number of bins. When the benchmark curve exhibits a low-density distribution in a specific range, the bin width of the corresponding range is expanded to 150% of the original width to reduce the number of bins. The adjusted bin boundaries are strictly aligned with the data density distribution trend of the benchmark curve, ensuring that the numerical range of each bin interval fully matches the key data segment of the benchmark curve. During the dynamic adjustment process, the number of bins is kept within a preset upper limit, which is set based on the scoring granularity requirements of the skill label. For example, if the theoretical maximum score is 100, the upper limit is 10 bins. The adjusted bin interval boundaries are synchronized with the benchmark curve through a database transaction lock mechanism to prevent bin misalignment caused by data contention.

[0075] The specific process for recalculating the discretized skill test score distribution based on the adjusted bin width and extracting the kurtosis and skewness parameters of the baseline curve is as follows: re-dividing the skill test score intervals according to the dynamically adjusted bin width, calculating the percentage of candidates within each new interval, and generating an updated discretized skill test score distribution table. The statistical distribution parameters of the target position skill requirement baseline curve are extracted. The kurtosis parameter quantifies the sharpness of the distribution shape and is calculated based on the numerical results of the standardized fourth-order central moment. The skewness parameter quantifies the symmetry of the distribution shape and is calculated based on the numerical results of the standardized third-order central moment. The calculation process for the kurtosis and skewness parameters uses a sliding window method to traverse the data points of the baseline curve to perform local feature extraction. The window width is dynamically adjusted based on the total amount of data in the baseline curve, and the window width is set according to the square root of the total amount of data. The calculation results are stored in the feature parameter database by skill label for subsequent steps.

[0076] The distribution function of the baseline curve is modified based on the kurtosis and skewness parameters to generate a dynamically adapted benchmark curve for target job skill requirements. The specific process involves comparing the kurtosis parameter of the baseline curve with the preset kurtosis threshold of the job skill standard. When the kurtosis parameter exceeds the threshold, the baseline curve's distribution function is subjected to a kurtosis reduction operation, which is achieved by reducing the probability density of the distribution function in the peak region. The skewness parameter of the baseline curve is compared with the preset skewness threshold of the job skill standard. When the skewness parameter indicates left or right skewness exceeding the threshold, the distribution function is subjected to a skewness correction operation, which is achieved by shifting the center point of the distribution function or adjusting the slope of the probability density function. The corrected distribution function is then refitted to the skill test score distribution of historically successful candidates to generate a dynamically adapted benchmark curve for target job skill requirements. The morphological characteristics of this curve reflect both the theoretical distribution characteristics of job skill requirements and the dynamic trends of the actual distribution of current candidates. A version control mechanism is used to record historical changes to the revised benchmark curve, supporting data backtracking and error analysis.

[0077] The specific process of calculating the KL divergence value of the discretized skill test score distribution and the dynamically adapted target position skill requirement benchmark curve is as follows: the frequency proportion data in the updated discretized skill test score distribution table is converted into a discrete probability distribution. The conversion rule is to divide the frequency proportion of each bin interval by the interval width to eliminate the bin scale difference.

[0078] The dynamically adapted target position skill requirement baseline curve is discretized into a baseline probability distribution with the same bin width, ensuring that the numerical range of the bin intervals is fully consistent with the skill test score distribution. The logarithm of the ratio of the candidate distribution probability to the baseline curve probability is calculated bin by bin, and the calculated results for each bin are weighted and summed according to the candidate distribution probability to generate the final KL divergence value. Bin intervals with zero probability are ignored during the calculation process to prevent mathematical invalidation. The KL divergence results are categorized by skill label and stored in the analysis database for subsequent access. A visualization tool is used to generate a distribution comparison chart of the calculated results to assist in manual review of the data's rationality.

[0079] S3. When the KL divergence value exceeds the KL divergence threshold, a matching threshold correction coefficient is generated based on the PID control algorithm. The proportional coefficient of the PID control algorithm is dynamically adjusted according to the correlation coefficient between the candidate's skill test score and performance data. The specific implementation is as follows:

[0080] Obtain the correlation coefficient between the candidate's skill test score and performance data. The specific process for calculating the correlation coefficient using the Pearson product-moment correlation coefficient formula is as follows: Extract the candidate's skill test scores and post-employment performance data corresponding to the target position's current skill tag from the associated storage database in step S1, and organize them into independent data sets by skill tag name. Perform data alignment on the skill test scores and performance data under each skill tag. This data alignment operation includes removing invalid candidate records that do not have both skill test scores and performance data, and sorting the skill test scores and performance data of the remaining records by timestamp to ensure time series consistency of the data.

[0081] Calculate the covariance between the skill test scores and performance data. Covariance is calculated by subtracting the product of the two means from the mean of the product of the corresponding items in the two data sets. Calculate the standard deviation of the skill test scores and performance data separately. The standard deviation is calculated as the square root of the mean of the squared differences between the data items and the means. Divide the covariance by the product of the standard deviation of the skill test scores and the standard deviation of the performance data to generate the standardized Pearson product-moment correlation coefficient. The results are stored in the analysis database by skill label for subsequent access.

[0082] The specific process of dynamically adjusting the proportional coefficient of the PID control algorithm based on the degree of deviation between the correlation coefficient and the preset proportional reference value is as follows: obtaining the correlation coefficient calculated in step S3 and the preset proportional reference value. The proportional reference value is set according to the long-term correlation level between the skill test scores and the performance data in the historical successful candidate data of the target position. The specific setting rule is to select the moving average of the correlation coefficient in the historical data as the reference value.

[0083] When the correlation coefficient is greater than the proportional baseline value, the proportional coefficient is incremented according to a preset linear gain rule. The increment is the difference between the correlation coefficient and the baseline value multiplied by a preset gain factor, which is set based on the target position's adjustment sensitivity requirements. When the correlation coefficient is less than the baseline value, the proportional coefficient is decremented according to a preset exponential decay rule. The decrement is the difference between the baseline value and the correlation coefficient multiplied by the decay factor. The dynamic adjustment range of the proportional coefficient is limited to a preset interval. The lower limit of the interval is set according to the minimum skill requirement of the position, and the upper limit is set according to the maximum skill tolerance of the position. During the adjustment process, the proportional coefficient is detected in real time to see if it exceeds the interval range and the boundary truncation is performed.

[0084] The integral term is calculated based on the difference between the KL divergence value and the KL divergence threshold, and the differential term is calculated based on the rate of change of the KL divergence value. The specific process is as follows: obtain the KL divergence value calculated in step S2 and the preset KL divergence threshold, and calculate the difference between the current KL divergence value and the threshold as the input for the integral term. The integral term is calculated by accumulating the weighted sum of the historical difference sequence. The weight distribution of the weighted sum adopts a time decay strategy. The difference value in the most recent calculation cycle is weighted as 1, and the weight of the previous cycle is the product of the decay factor and the inverse of the time interval. The decay factor is set based on the skill stability requirements of the position. The differential term is calculated based on the rate of change of the KL divergence value within a preset time window. The width of the time window is set according to the adjustment response speed of the target position. The rate of change is calculated as the difference between the current KL divergence value and the KL divergence value at the start of the window divided by the length of the time interval, which is calculated in days and is accurate to one decimal place. The calculation results of the integral and differential terms are stored in the control parameter database according to skill label.

[0085] The specific process for generating the matching threshold correction coefficient by superimposing the proportional coefficient, integral term, and differential term according to preset weights is as follows: The proportional coefficient, integral term, and differential term for the current skill tag are extracted from the control parameter database and linearly superimposed according to the preset weight distribution ratio. The preset weights are set based on the skill priority of the target position. The weight distribution of core skill tags is biased towards the proportional term and differential term, while the weight distribution of auxiliary skill tags is biased towards the integral term and differential term.

[0086] Before the superposition calculation, the integral and differential terms are normalized, mapping the integral term to the range of 0 to 1 and the differential term to the range of -1 to 1. The linear superposition result is smoothed using a sigmoid function to generate a matching threshold correction coefficient. The correction coefficient range is limited to 0.5 to 2.0, and values ​​outside this range are truncated at the boundary value. The correction coefficients are categorized by skill label and stored in the output database for recall in step S4 and written to a log file for subsequent audit tracking.

[0087] S4. Update the matching threshold of the current job skill tag according to the matching threshold correction coefficient to generate an updated threshold range. The specific implementation is as follows:

[0088] The specific process of obtaining the original matching threshold of the current job skill label and the matching threshold correction coefficient generated in step S3 is as follows: read the original matching threshold of the current skill label of the target job from the matching threshold configuration table of the job portrait database. The original matching threshold is set as follows: based on the skill test score distribution of successful candidates in the historical recruitment data of the target job, select the 70th percentile value as the initial threshold benchmark.

[0089] For example, when the skill test scores of historical successful candidates under a certain skill tag are arranged in ascending order, the 70th percentile corresponds to a score of 78 points, and the original matching threshold is set to 78 points. The percentile calculation method uses linear interpolation to ensure the continuity of the score distribution. The matching threshold correction coefficient that completely matches the current skill tag name is extracted from the output database of step S3. The correction coefficient is generated by the linear superposition calculation of step S3 and stored in conjunction with the unique code of the skill tag. A consistency check of the skill tag name is performed during data extraction. The check rule is: when the characters of the skill tag name in the threshold configuration table and the correction coefficient table are exactly the same (including uppercase and lowercase letters and spaces), it is considered a match; otherwise, an abnormal alarm is triggered and the process is terminated.

[0090] The original matching threshold is adjusted based on the matching threshold correction factor. For example, if the matching threshold is 78.00 and the correction factor is 1.15, the initial adjustment threshold is 89.70; if the correction factor is 0.92, the initial adjustment threshold is 71.76. The calculation results are stored in a temporary threshold table in the intermediate database by skill tag name. The temporary threshold table structure includes fields for skill tag name, original threshold, correction factor, and initial adjustment threshold. When writing data, the database row-level locking mechanism is enabled to prevent data overwriting or loss caused by concurrent writes.

[0091] If the initial adjustment threshold exceeds the preset threshold range, it is truncated according to the preset threshold range boundary value. The specific process for generating the compliance adjustment threshold is as follows: the preset threshold range boundary values ​​are obtained from the threshold configuration policy table. The boundary value setting logic is as follows: based on the rigidity of the target position's skill requirements, skill tags are divided into core skills and auxiliary skills. The lower limit of the threshold range for core skills is set at the 30th percentile value of the skill test scores of historical successful candidates, and the upper limit is set at the 90th percentile value. The lower limit of the threshold range for auxiliary skills is set at the 20th percentile value, and the upper limit is set at the 80th percentile value. For example, if the 30th percentile score of historical successful candidates for the core skill tag is 65 and the 90th percentile score is 92, the threshold range is set to 65-92. If the initial adjustment threshold exceeds the upper limit, the threshold is set to the upper limit; if it falls below the lower limit, the threshold is set to the lower limit. For example, if the initial adjustment threshold is 95.00 and the upper limit is 92.00, the compliance adjustment threshold is truncated to 92.00. If the initial adjustment threshold is 58.00 and the lower limit is 65.00, it is adjusted to 65.00. The truncated compliance adjustment threshold is written to the review queue and awaits priority weight fusion.

[0092] The updated threshold range is generated based on the compliance adjustment threshold and the priority weights of the target position's skill tags. The following steps are used to extract the priority weights of the target position's skill tags from the weight configuration table in the position profile database. The weights are set based on the importance classification of skill items in the job description. The weights of core skill tags are dynamically calculated based on the tag level: those marked "required" receive a weight of 0.75, and those marked "priority" receive a weight of 0.60. The weights of auxiliary skill tags are uniformly set to 0.30. The compliance adjustment threshold and priority weights are weighted and combined to generate the updated threshold range. The weighted combination rule is: using the compliance adjustment threshold as the baseline value, the floating range is negatively correlated with the weight value, and the calculation formula is: floating range = baseline value × (1 - weight) × adjustment factor. The adjustment factor is set based on the urgency of the job recruitment and is generally 0.10. For example, when the compliance adjustment threshold for core skills is 85.00 points and the weight is 0.75, the floating range = 85.00×(1-0.75)×0.10=2.13 points, and the updated threshold range is 82.87-87.13 points; when the compliance adjustment threshold for auxiliary skills is 70.00 points and the weight is 0.30, the floating range = 70.00×(1-0.30)×0.10=4.90 points, and the threshold range is 65.10-74.90 points.

[0093] The specific process for writing the updated threshold range to the job profile database and triggering the threshold effective timestamp recording is as follows: The generated updated threshold range is updated to the matching threshold configuration table in the job profile database by skill tag name, overwriting the original threshold data. Before executing the update operation, a database transaction snapshot is created. The snapshot includes the original threshold data, operation time, and operation session ID, ensuring rollback to the previous version in the event of an exception. A threshold effective timestamp record is inserted into the transaction log table. The record fields include the operation time (accurate to milliseconds), operator ID (a unique ID automatically generated by the system), skill tag name, old threshold range, and new threshold range. The generation of the timestamp record and the threshold update are ensured consistent through atomic database transactions. Atomic transactions are implemented by locking all write operations in the relevant data tables before the transaction commits and releasing the locks after the log record and threshold data are successfully written. The effective threshold range is automatically activated at midnight on the next calendar day. During activation, the new threshold range is verified to be compatible with the skill tag status of the current job profile. If there is a conflict (e.g., the lower threshold limit is higher than the current candidate's highest score), a manual review process is triggered. The old threshold version is migrated to the historical version library, and the expiration timestamp and operator identification are marked during migration. The historical version data is retained for five years for compliance audits and data analysis.

[0094] S5. Based on the deviation direction of the KL divergence value and the updated threshold range, determine the sampling weight adjustment strategy for the candidate's skill test score. The specific implementation is as follows:

[0095] The specific process for determining the direction of deviation of the KL divergence value from the updated threshold range is as follows: extract the KL divergence value of the current skill tag from the KL divergence value calculation result in step S2 and compare it with the updated threshold range generated in step S4. This comparison is implemented through a sign function operation: when the KL divergence value is greater than the upper limit of the updated threshold range, it is marked as a positive deviation from the upper threshold; when the KL divergence value is less than the lower limit of the updated threshold range, it is marked as a negative deviation from the lower threshold. The input of the sign function is the difference between the KL divergence value and the upper and lower limits of the threshold range, and the output is a deviation direction identifier. A positive number represents a positive deviation, and a negative number represents a negative deviation. The deviation direction determination result is stored in the deviation record table of the analysis database according to the skill tag name. The fields in the table include the skill tag name, KL divergence value, upper and lower limits of the threshold range, deviation direction identifier, and timestamp.

[0096] When the deviation is positive toward the upper threshold, the sampling weights for candidates whose skill test scores exceed the upper threshold are reduced as follows: The upper threshold value of the updated threshold range and the candidate skill test score data are obtained, and candidate records with skill test scores above the upper threshold are screened. The sampling weights of these records are then attenuated. The attenuation rule is: a weight attenuation factor is set based on the magnitude of the deviation from the upper threshold. The deviation ranges are categorized as 0-5 points, 5-10 points, and 10 points or more, with corresponding attenuation factors of 0.9, 0.7, and 0.5. For example, if a candidate's skill test score is 95, the upper threshold is 90, and the deviation is 5 points, the attenuation factor is 0.7, and the original weight of 1.0 is adjusted to 0.7. The adjusted weight data is stored in the sampling weight configuration table by candidate ID. The weight values ​​are verified to be within the valid range of 0 to 1 during storage. If they exceed the range, they are automatically truncated to the boundary value.

[0097] When the deviation is negative from the lower threshold, the sampling weight for candidates whose skill test scores fall below the current lower threshold is increased by obtaining the lower limit of the updated threshold range and the candidate skill test score data. The sampling weights of these candidates are then increased. The gain factor is set based on the magnitude of the deviation from the lower threshold. The deviation ranges are categorized as 0-5, 5-10, and 10 or more, with corresponding gain factors of 1.1, 1.3, and 1.6. For example, if a candidate's skill test score is 58 and the lower threshold is 60, the gain factor is 1.1, and the original weight of 1.0 is adjusted to 1.1. The adjusted weight data is stored in the sampling weight configuration table by candidate ID. The weight value is verified to be within the valid range (0 to 2) during storage. If it exceeds the range, it is automatically truncated to the boundary value.

[0098] A dynamic sampling weight adjustment strategy is generated based on the deviation direction type and deviation magnitude. This strategy includes piecewise function mapping rules for weight decay factors and gain factors. The specific process involves constructing a piecewise mapping table between deviation magnitude and weight factors. This mapping rule is dynamically adjusted based on the skill stability requirements of the target position. For positions with high stability requirements (such as core technical positions), a steeper factor gradient is adopted. For example, a positive deviation magnitude of 0-5 points in the upper threshold corresponds to a decay factor of 0.8 (previously 0.9), and 5-10 points corresponds to 0.6 (previously 0.7). For ordinary positions, the default gradient is maintained. The mapping table is stored in a policy library by skill tag name. The policy library structure includes fields for skill tag name, deviation direction, deviation magnitude range, weight factor, and effective timestamp. A version control mechanism is triggered during policy generation. Each update generates a new version and retains historical version data, supporting policy rollbacks and comparative analysis of effectiveness.

[0099] S6. Retrain the matching model based on the skill test score sampling weight adjustment strategy and output an updated candidate recommendation list. The specific implementation is as follows:

[0100] The specific process for extracting the weight decay factor and gain factor mapping table from the sampled weight adjustment strategy is as follows: The weight decay factor and gain factor mapping table for the target position's current skill tag is read from the policy library generated in step S5. The mapping table is categorized and stored by skill tag name and deviation direction. During extraction, an exact match check is performed on the skill tag name to ensure that only mapping table entries that fully match the active skill tags in the current position profile are extracted. For example, if the target position's active skill tags are "Java skills" and "distributed system development," mapping table data for inactive tags such as "front-end development" is excluded. The mapping table data structure includes a skill tag name field, a deviation direction field (upper positive deviation or lower negative deviation), a deviation range field (e.g., 0-5 points, 5-10 points, etc.), and a corresponding weight factor field. The extracted mapping table is sorted by skill tag name and loaded into the training parameter pool in memory. Data integrity is verified during loading, and exception handling is triggered if any fields are missing or values ​​are out of bounds.

[0101] The specific process for generating a weighted training dataset by adjusting the sampling weight distribution of candidate skill test scores based on weight decay and gain factors is as follows: Candidate skill test score data is extracted from the associated storage database in step S1, and each candidate's sampling weight is adjusted according to the sampling weight adjustment strategy in step S5. For candidates whose skill test scores exceed the current upper threshold, their weight is reduced according to the decay factor in the mapping table. For example, if a candidate's "Java Skills" score is 95 (upper threshold 90), and the decay factor is 0.7, their weight is adjusted from 1.0 to 0.7. For candidates whose scores are below the current lower threshold, their weight is increased according to the gain factor. For example, if a candidate's "Distributed Systems Development" score is 55 (lower threshold 60), and the gain factor is 1.3, their weight is adjusted from 1.0 to 1.3. The adjusted weight data is stored in the weighted training dataset using a composite primary key composed of the candidate ID and skill tag. The structure of the dataset includes candidate ID, skill tag name, skill test score, adjusted weight, original weight and timestamp fields. During storage, a database transaction lock mechanism is used to prevent concurrent write conflicts, and data version control is enabled to support historical tracing.

[0102] The matching model is retrained using the weighted training dataset. The matching model updates its parameters by minimizing the prediction error of weighted samples. The specific process is as follows: a previously trained matching model is loaded and the weighted training dataset is fed into the model for incremental training. During training, the loss function weight for each sample is dynamically scaled based on the adjusted weights. The total error in the loss function is calculated as the sum of the weighted errors for all samples, where the weighted error is the squared difference between the predicted and actual values ​​multiplied by the adjusted weights. Model parameters are updated using the stochastic gradient descent algorithm, with the learning rate dynamically decaying based on the current training round. The initial learning rate is set to 0.01, and after every 10 training rounds, the learning rate is multiplied by a factor of 0.9, with a minimum learning rate of no less than 0.0001. After training is complete, the updated model parameters are stored in the model repository by skill tag name. The model version number, training timestamp, and weighting strategy version number used are recorded to ensure traceability of the model and weight adjustment strategy version association.

[0103] The specific process for filtering the matching model output based on the updated threshold range to generate an updated candidate recommendation list and append a threshold effective timestamp is as follows: The retrained matching model is applied to the current candidate pool to generate an original recommendation score list. The recommendation score list is filtered based on the updated threshold range written to the job profile database in step S4, retaining only candidates whose skill test scores fall within the threshold range. For example, if the updated threshold range for "Java skills" is 85-95, candidates with scores below 85 or above 95 are filtered out. The filtered recommendation list is sorted by skill tag name, with the following sorting rule: candidates with the same skill tag are sorted in descending order by recommendation score, and candidates with the same score are sorted in ascending order by candidate ID. The threshold effective timestamp appended to the recommendation list is extracted from the database transaction log in step S4. During extraction, the timestamp is verified to correspond to the current effective threshold version. If a timestamp conflict exists, an alarm is triggered and manual intervention is initiated. The resulting recommendation list is written to the recruitment platform's display queue. The display queue's data structure contains fields for candidate ID, skill tag name, recommendation score, threshold range, and effective timestamp. At the same time, the filtered candidate data will be migrated to the pool for review. Each record in the pool for review will be marked with the filtering reason (such as "exceeding the upper limit" or "below the lower limit") and the threshold version number for review by the recruitment specialist.

[0104] This embodiment dynamically optimizes the job matching threshold by building a closed-loop feedback mechanism. In traditional recruitment systems, threshold adjustment typically relies on explicit indicators such as the number of resumes and the market supply-demand ratio. This lacks quantitative control over candidate skill distribution deviations, which can easily lead to a vicious cycle of low-quality data contaminating the model. This embodiment introduces the KL divergence value to quantify the degree of deviation between the candidate skill test score distribution and the job benchmark curve. A PID control algorithm is then used to dynamically adjust the matching threshold correction coefficient. The proportional coefficient is linked to the correlation coefficient between skill-performance data, allowing the threshold correction amplitude to adaptively adjust with data quality feedback, thereby suppressing error accumulation. Furthermore, an asymmetric adjustment is performed on the sampling weights of high- and low-scoring candidates based on the deviation direction (positive deviations reduce weights, negative deviations increase weights), proactively mitigating the negative impact of low-quality data on model training. Furthermore, by strictly binding the threshold effective timestamp to the model version number, the timing consistency of threshold updates, model training, and result filtering is ensured, avoiding data version conflicts. Compared to existing technologies, this systematic integration of distribution deviation detection, dynamic control theory, weight inverse adjustment, and timing synchronization significantly improves talent matching accuracy and the long-term stability of the recommendation system while ensuring recruitment efficiency.

[0105] Example 2: Figure 2 The present invention provides a structural diagram of a matching threshold adjustment system for dynamic updating of job profiles. The matching threshold adjustment system for dynamic updating of job profiles includes the following modules:

[0106] Data acquisition module: used to obtain the candidate's skill test scores and post-employment performance data corresponding to the current skill tag of the target position;

[0107] Divergence calculation module: used to calculate the KL divergence value between the distribution curve of candidate skill test scores within a preset period and the benchmark curve of target position skill requirements;

[0108] Dynamic adjustment module: When the KL divergence value exceeds the KL divergence threshold, it is used to generate a matching threshold correction coefficient based on the PID control algorithm;

[0109] Threshold update module: used to update the matching threshold of the current job skill tag according to the matching threshold correction coefficient and generate the updated threshold range;

[0110] Strategy generation module: used to determine the sampling weight adjustment strategy of the candidate's skill test score based on the deviation direction of the KL divergence value and the updated threshold range;

[0111] Model optimization module: used to retrain the matching model based on the skill test score sampling weight adjustment strategy and output an updated candidate recommendation list.

[0112] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to actual conditions.

[0113] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0114] Those skilled in the art will appreciate that the modules and algorithm steps of each example 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 performed in hardware or software depends on the specific application of the technical solution and the invention constraints. Professional and technical personnel 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.

[0115] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0116] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0117] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0118] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A matching threshold adjustment method for dynamic updating of job profiles, characterized in that: The steps include: S1. Obtain the candidate's skill test scores and post-employment performance data corresponding to the current skill tag of the target position; S2. Calculate the KL divergence between the distribution curve of the candidate's skill test scores within the preset period and the benchmark curve of the target position's skill requirements, including: The candidate's skill test scores are binned by skill label dimension within a preset period to generate a discretized skill test score distribution; Dynamically adjust the bin width based on the distribution of the target job skill requirement benchmark curve to align the distribution of discretized skill test scores with the data interval of the benchmark curve; Recalculate the discretized skill test score distribution based on the adjusted bin width and extract the kurtosis and skewness parameters of the benchmark curve; The distribution function of the benchmark curve is modified based on the kurtosis parameter and skewness parameter to generate a dynamically adapted benchmark curve for target job skill requirements. Calculate the KL divergence value between the discretized skill test score distribution and the dynamically adapted target job skill requirement benchmark curve; S3. When the KL divergence value exceeds the KL divergence threshold, a matching threshold correction coefficient is generated based on the PID control algorithm; S4. Update the matching threshold of the current job skill tag according to the matching threshold correction coefficient to generate an updated threshold range; S5. Based on the deviation direction of the KL divergence value from the updated threshold range, determine the sampling weight adjustment strategy for the candidate's skill test score, including: Determine the deviation direction of the KL divergence value from the updated threshold range. The deviation direction is divided into positive deviation from the upper threshold limit and negative deviation from the lower threshold limit. When the deviation direction is a positive deviation towards the upper threshold limit, the sampling weight of the candidate skill test scores that are higher than the current upper threshold limit is reduced; When the deviation direction is a negative deviation from the lower threshold, the sampling weight of the candidate skill test scores below the current lower threshold is increased; Generate a dynamic sampling weight adjustment strategy based on the deviation direction type and deviation magnitude; S6. Retrain the matching model based on the skill test score sampling weight adjustment strategy and output the updated candidate recommendation list.

2. The matching threshold adjustment method for dynamic updating of job profiles according to claim 1 is characterized in that: The proportional coefficient of the PID control algorithm is dynamically adjusted based on the correlation coefficient between the candidate's skill test score and performance data.

3. The matching threshold adjustment method for dynamic updating of job profiles according to claim 1 is characterized in that: Obtain candidate skill test scores and post-employment performance data corresponding to the target position's current skill tags, including: Obtain the target position's job description text, parse the job description text, and generate the target position's current skill tag; Obtain the candidate's skill test scores corresponding to the current skill tags of the target position from the online assessment system; Obtain the candidate's post-employment performance data in the position corresponding to the current skill tag of the target position from the human resources management system.

4. The matching threshold adjustment method for dynamic updating of job profiles according to claim 1 is characterized in that: When the KL divergence value exceeds the KL divergence threshold, a matching threshold correction coefficient is generated based on the PID control algorithm, including: Obtain the correlation coefficient between the candidate's skill test score and performance data. The correlation coefficient is calculated using the Pearson product-moment correlation coefficient formula; Dynamically adjust the proportional coefficient of the PID control algorithm based on the degree of deviation between the correlation coefficient and the preset proportional reference value; The integral term is calculated based on the difference between the KL divergence value and the KL divergence threshold, and the differential term is calculated based on the rate of change of the KL divergence value; The proportional coefficient, integral term and differential term are superimposed according to the preset weights to generate the matching threshold correction coefficient.

5. The matching threshold adjustment method for dynamic updating of job profiles according to claim 1 is characterized in that: Update the matching threshold of the current job skill tag based on the matching threshold correction coefficient to generate the updated threshold range, including: Obtain the original matching threshold of the current job skill tag and the matching threshold correction coefficient generated in step S3; Linearly adjust the original matching threshold according to the matching threshold correction coefficient to generate a preliminary adjustment threshold; When the initial adjustment threshold exceeds the preset threshold range, it is truncated according to the preset threshold range boundary value to generate a compliance adjustment threshold; Generate an updated threshold range based on the compliance adjustment threshold and the priority weight of the target position skill tag; The updated threshold range is written into the job profile database and the threshold effective timestamp record is triggered.

6. The matching threshold adjustment method for dynamic updating of job profiles according to claim 1 is characterized in that: The sampling weight adjustment strategy includes a piecewise function mapping rule between the weight attenuation factor and the gain factor.

7. The matching threshold adjustment method for dynamic updating of job profiles according to claim 1 is characterized in that: Retrain the matching model based on the skill test score sampling weight adjustment strategy and output an updated candidate recommendation list, including: Extracting a weight decay factor and gain factor mapping table from a sampled weight adjustment strategy; Adjusting the sampling weight distribution of the candidate's skill test scores based on the weight decay factor and the gain factor to generate a weighted training dataset; The matching model is retrained using the weighted training dataset, and the matching model updates its parameters by minimizing the prediction error of the weighted samples; The matching model output results are filtered according to the updated threshold range, and an updated candidate recommendation list is generated and appended with the threshold effective timestamp.

8. A matching threshold adjustment system for dynamic updating of job profiles, used to implement the matching threshold adjustment method for dynamic updating of job profiles according to any one of claims 1 to 7, characterized in that: Includes the following modules: Data acquisition module: used to obtain the candidate's skill test scores and post-employment performance data corresponding to the current skill tag of the target position; Divergence calculation module: used to calculate the KL divergence value between the distribution curve of candidate skill test scores within a preset period and the benchmark curve of target position skill requirements; Dynamic adjustment module: When the KL divergence value exceeds the KL divergence threshold, it is used to generate a matching threshold correction coefficient based on the PID control algorithm; Threshold update module: used to update the matching threshold of the current job skill tag according to the matching threshold correction coefficient and generate the updated threshold range; Strategy generation module: used to determine the sampling weight adjustment strategy of the candidate's skill test score based on the deviation direction of the KL divergence value and the updated threshold range; Model optimization module: used to retrain the matching model based on the skill test score sampling weight adjustment strategy and output an updated candidate recommendation list.

Images