A data analysis-based intelligent matching model management method and system for positions and posts

Abstract

The application relates to the technical field of human position matching, in particular to a human position intelligent matching model management method and system based on data analysis, which comprises the following steps: collecting historical recruitment data to construct a job seeker characteristic vector and a post characteristic vector; extracting a job seeker characteristic vector and a post characteristic vector corresponding to a job seeking result to form a historical job seeking feedback data set, calculating a non-biased characteristic matching degree to construct a biased feedback data set; extracting non-biased feedback data based on the biased feedback data set and the historical job seeking feedback data set, and reconstructing and optimizing the historical job seeking feedback data set; constructing a non-biased human position matching data set, training a human position intelligent matching model by using the non-biased human position matching data set; applying the trained human position intelligent matching model to position recommendation, and collecting subsequent job seeking result data to optimize the model. The application can effectively eliminate biased factors in historical recruitment data and improve the fairness of matching results.

Description

Technical Field

[0001] This invention relates to the field of human-job matching technology, specifically to a human-job intelligent matching model management method and system based on data analysis. Background Technology

[0002] In today's digital employment market, intelligent job matching models play a crucial role in job recommendations. As companies increasingly prioritize efficiency in talent screening and job seekers demand precise job recommendations, the application of these models is becoming more widespread.

[0003] However, existing intelligent job matching models have significant flaws. In the process of collecting and utilizing historical recruitment and application data, numerous factors unrelated to job seekers' professional skills are often mixed in, such as age, gender, and place of origin. Past recruitment practices have shown that even if job seekers possess the core professional skills required for the position, they often fail in the job search process due to non-hard professional skill factors such as age not conforming to implicit company preferences, gender not meeting specific job stereotypes, or regional bias arising from place of origin.

[0004] When biased data is incorporated into the training set of a human-job matching model, the model will unconsciously perpetuate these biases when subsequently applied to job recommendations. For example, it may over-select job seekers based on specific age ranges, genders, or places of origin, while ignoring other candidates who are equally or even more qualified for the job. This causes the matching results to deviate significantly from the principles of fairness and impartiality, not only causing many job seekers to miss suitable development opportunities but also making it difficult for them to receive truly fair recommendations based on their professional abilities, thus reducing their efficiency in finding suitable positions in the job market.

[0005] To address this, a data-driven intelligent matching model management method and system for people and jobs is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a data analysis-based intelligent job matching model management method and system. It involves collecting historical recruitment data to construct job seeker feature vectors and job feature vectors; extracting the corresponding job seeker and job feature vectors from the job application results to form a historical job application feedback dataset; calculating the unbiased feature matching degree to construct a biased feedback dataset; extracting unbiased feedback data based on the biased feedback dataset and the historical job application feedback dataset; reconstructing and optimizing the historical job application feedback dataset; constructing an unbiased job matching dataset; using the unbiased job matching dataset to train the intelligent job matching model; applying the trained intelligent job matching model to job recommendation; and collecting subsequent job application result data to optimize the model. This invention can effectively eliminate bias factors in historical recruitment data and improve the fairness of the matching results.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A data-driven intelligent matching model management method for people and jobs includes:

[0009] Collect historical recruitment data; the historical recruitment data includes historical resume data, historical job posting data, and historical job application results data;

[0010] Construct a feature mapping dictionary, and construct job seeker feature vectors and job feature vectors based on the feature mapping dictionary and the historical recruitment data; the job seeker feature vectors and job feature vectors include hard requirement feature sub-vectors, soft preference feature sub-vectors, and potential bias feature sub-vectors;

[0011] Based on the historical job search results data, the corresponding job seeker feature vectors and job feature vectors are extracted to form a historical job search feedback dataset.

[0012] Calculate the unbiased feature matching degree based on the job seeker feature vector and the job feature vector, and construct a biased feedback dataset based on the unbiased feature matching degree and the historical job application feedback dataset.

[0013] Based on the biased feedback dataset and the historical job application feedback dataset, unbiased feedback data is extracted, and the historical job application feedback dataset is reconstructed and optimized according to the unbiased feedback data to obtain a reconstructed and optimized feedback dataset.

[0014] An unbiased person-job matching dataset is constructed based on the reconstructed and optimized feedback dataset, and the person-job intelligent matching model is trained using the unbiased person-job matching dataset.

[0015] The trained intelligent job matching model is applied to job recommendation, and subsequent job search results data are collected. The intelligent job matching model is then optimized based on the job search results data.

[0016] Preferably, the calculation steps for the unbiased feature matching degree are as follows:

[0017] Calculate the hard requirement matching degree based on the hard requirement feature subvectors of the job seeker and the position:

[0018]

[0019] Among them, M hard This indicates the degree of matching of the hard requirements; The weight of the k-th mandatory feature is represented by sim(); sim() represents the similarity calculation function. and These represent the vector components corresponding to the k-th hard requirement feature for both the job seeker and the position; n represents the number of hard requirement features.

[0020] Calculate the soft preference matching degree based on the soft preference feature subvectors of the job seeker and the job position:

[0021]

[0022] Among them, M soft This indicates the degree of matching of the soft preferences; V represents the weight of the l-th soft preference feature; l soft and V l ′ soft ...

[0023] The unbiased feature matching degree is calculated based on the hard requirement matching degree and the soft preference matching degree:

[0024] M nonbias =α·M hard +(1-α)·M soft ;

[0025] Among them, M nonbias α represents the unbiased feature matching degree; α represents the balance coefficient.

[0026] Preferably, constructing the biased feedback dataset based on the non-biased feature matching degree and the historical job application feedback dataset includes:

[0027] Define the biased feedback dataset by setting a non-biased matching threshold:

[0028] D bias ={d h |M nonbias (d h )>θ and Y h =0};

[0029] Among them, D bias This represents the bias feedback dataset; d h This represents the h-th historical job application feedback record; M nonbias (d h ) represents the unbiased feature matching degree corresponding to the h-th historical job application feedback data; θ represents the unbiased matching degree threshold; Y h Y represents the hiring result of the historical job application feedback data mentioned in item h. h =0 indicates not hired, Y h =1 indicates acceptance.

[0030] Preferably, extracting the unbiased feedback data based on the biased feedback dataset and the historical job application feedback dataset, and reconstructing and optimizing the historical job application feedback dataset based on the unbiased feedback data includes:

[0031] The biased feedback dataset is deleted from the historical job application feedback dataset to obtain the reconstructed job application feedback dataset;

[0032] Job success feedback data is extracted from the reconstructed job search feedback dataset, and a job-person matching similarity matrix is ​​calculated; each element in the job-person matching similarity matrix represents the job-person matching similarity between each job success feedback data and each biased feedback data.

[0033] The job application success feedback data with a job-person matching similarity greater than the job-person matching threshold are selected as the unbiased feedback data.

[0034] For each piece of unbiased feedback data, the number of times the person-job matching similarity is greater than the person-job matching threshold is counted, the bias weight is calculated, and the unbiased feedback data is sampled according to the bias weight to obtain an optimized unbiased feedback dataset;

[0035] The reconstructed job application feedback dataset and the optimized unbiased feedback dataset are combined to obtain the reconstructed optimized feedback dataset.

[0036] Preferably, a dual-tower model is used to construct the intelligent matching model between people and jobs, and the loss function is set as follows:

[0037] L = L match +β1·L bias +β2·L reg ;

[0038]

[0039] L reg =||W||2;

[0040] Where L represents the loss function; L match L represents the matching degree loss; bias Indicates bias control loss; L reg β1 and β2 represent the regularization loss; N represents the number of samples; y i The matching label value for sample i; m i This represents the matching degree of sample i in the model output; log() represents the logarithmic function; Let represent the unbiased feature matching degree of sample i; W represents the set of weights of the model; |||2 represents the Euclidean norm.

[0041] Preferably, the matching degree label values ​​for the sample include: a job seeker in the sample who did not apply for a position in the sample, corresponding to a matching degree label value of λ1; a job seeker in the sample who applied for a position in the sample but did not get an interview, corresponding to a matching degree label value of λ2; a job seeker in the sample who applied for a position in the sample and got an interview but did not pass the interview, corresponding to a matching degree label value of λ3; and a job seeker in the sample who applied for a position in the sample and was hired, corresponding to a matching degree label value of λ4.

[0042] A data analysis-based intelligent matching model management system for people and jobs includes:

[0043] The historical data acquisition module collects historical recruitment data, including historical resume data, historical job posting data, and historical job application results data.

[0044] The job feature construction module constructs a feature mapping dictionary and builds job feature vectors and job feature vectors based on the feature mapping dictionary and the historical recruitment data. The job feature vectors and job feature vectors include hard requirement feature sub-vectors, soft preference feature sub-vectors, and potential bias feature sub-vectors.

[0045] The feedback data extraction module extracts the corresponding job seeker feature vector and job feature vector based on the historical job search result data to form a historical job search feedback dataset.

[0046] The biased data extraction module calculates the non-biased feature matching degree based on the job seeker feature vector and the job feature vector, and constructs a biased feedback dataset based on the non-biased feature matching degree and the historical job application feedback dataset.

[0047] The data reconstruction and optimization module extracts unbiased feedback data based on the biased feedback dataset and the historical job application feedback dataset, and reconstructs and optimizes the historical job application feedback dataset based on the unbiased feedback data to obtain a reconstructed and optimized feedback dataset.

[0048] The model training and optimization module constructs an unbiased job-person matching dataset based on the reconstructed optimization feedback dataset, trains the intelligent job-person matching model using the unbiased job-person matching dataset, applies the trained intelligent job-person matching model to job recommendation, collects subsequent job search result data, and optimizes the intelligent job-person matching model based on the job search result data.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] 1. A job seeker and job feature vector was constructed, including hard requirement feature vectors, soft preference feature vectors, and potential bias feature vectors. This allows for a comprehensive and accurate quantification of job seeker and job information. By eliminating potential bias features and calculating the unbiased feature matching degree solely based on the hard requirement and soft preference feature vectors, the interference of non-professional factors on the matching results is effectively avoided. This makes the matching process more focused on the actual needs of the job and the matching of job seekers' abilities and qualities, improving the accuracy and fairness of job matching. It also provides a more scientific data foundation for model training and fundamentally reduces the impact of historical data bias on the matching results.

[0051] 2. A biased feedback dataset is constructed by setting thresholds based on unbiased feature matching to accurately identify job application feedback data that may be influenced by bias. By deleting biased feedback datasets and calculating a job-person matching similarity matrix, high-similarity successful cases are selected to form unbiased feedback data. This data is then optimized by sampling according to bias weights and finally combined with the reconstructed job application feedback dataset to obtain the reconstructed and optimized feedback dataset. By mining and utilizing historical data, biases in the original data are effectively corrected, improving the quality and representativeness of historical job application feedback datasets. This makes the training data more reflective of real job-person matching relationships, enhances the scientific rigor of model training, reduces the risk of unfair matching caused by bias in historical data, and promotes the performance improvement of the intelligent job-person matching model.

[0052] 3. A loss function is set up that includes matching degree loss, bias control loss, and regularization loss. The sample matching degree label values ​​are also finely configured to cover various scenarios: job seekers who did not apply, did not advance to the interview, failed the interview, or were hired. The loss function design allows the model to balance matching accuracy and fairness during training, effectively handling complex job-person matching relationships and preventing overfitting. Fine-tuned sample label settings provide more accurate training feedback, allowing the model to learn the matching degree of different states, thus making more detailed predictions, helping the model to continuously optimize, improving its reliability and adaptability in practical applications, and better addressing bias issues in job-person matching. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of a data analysis-based intelligent matching model management method for people and jobs according to the present invention.

[0054] Figure 2 This is a schematic diagram of the historical job application feedback dataset reconstruction and optimization process of the present invention;

[0055] Figure 3 This is a schematic diagram of the structure of a human-job intelligent matching model management system based on data analysis according to the present invention. Detailed Implementation

[0056] 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.

[0057] Please see Figures 1 to 3 This invention provides a data analysis-based intelligent matching model management method and system for people and jobs, the technical solution of which is as follows:

[0058] Example 1:

[0059] This embodiment addresses the bias in job recommendations for recent graduates on a certain recruitment platform. To provide fairer and more suitable job recommendations for each job seeker and help talent find the most suitable employment opportunities, a data-driven intelligent job matching model management method is applied, such as... Figure 1 As shown, it includes:

[0060] Collect historical recruitment data; the historical recruitment data includes historical resume data, historical job posting data, and historical job application results data;

[0061] Construct a feature mapping dictionary, and construct job seeker feature vectors and job feature vectors based on the feature mapping dictionary and the historical recruitment data; the job seeker feature vectors and job feature vectors include hard requirement feature sub-vectors, soft preference feature sub-vectors, and potential bias feature sub-vectors;

[0062] Based on the historical job search results data, the corresponding job seeker feature vectors and job feature vectors are extracted to form a historical job search feedback dataset.

[0063] Calculate the unbiased feature matching degree based on the job seeker feature vector and the job feature vector, and construct a biased feedback dataset based on the unbiased feature matching degree and the historical job application feedback dataset.

[0064] Based on the biased feedback dataset and the historical job application feedback dataset, non-biased feedback data is extracted, and the historical job application feedback dataset is reconstructed and optimized according to the non-biased feedback data to obtain a reconstructed and optimized feedback dataset.

[0065] An unbiased person-job matching dataset is constructed based on the reconstructed and optimized feedback dataset, and the person-job intelligent matching model is trained using the unbiased person-job matching dataset.

[0066] The trained intelligent job matching model is applied to job recommendation, and subsequent job search results data are collected. The intelligent job matching model is then optimized based on the job search results data.

[0067] Table 1 shows a partial example of the feature mapping dictionary, constructed based on industry-accepted common sense. Table 2 shows an example of historical job application feedback data, specifically including the job seeker's feature vector, the job feature vector of the position applied for by the job seeker, and the corresponding job application results. The table displays specific information about the corresponding job application results.

[0068] Table 1 Example of Feature Mapping Dictionary

[0069] Feature type Feature Name Hard requirements characteristics Education Hard requirements characteristics major Hard requirements characteristics Professional skills Soft preference characteristics Project Experience Soft preference characteristics Awards Potential bias characteristics age Potential bias characteristics Place of origin

[0070] Table 2 Example of Historical Job Application Feedback Data

[0071] Person-job matching ID Job seeker ID Job ID Interview required Interview results Final result M001 R001 J001 yes pass Recruitment M002 R002 J001 yes pass Recruitment M003 R003 J001 no none Not hired M004 R004 J002 yes pass Recruitment M005 R005 J003 yes Not passed Not hired

[0072] Furthermore, the calculation steps for the unbiased feature matching degree are as follows:

[0073] Calculate the hard requirement matching degree based on the hard requirement feature subvectors of the job seeker and the position:

[0074]

[0075] Among them, M hard This indicates the degree of matching of the hard requirements; The weight of the k-th mandatory feature is represented by sim(); sim() represents the similarity calculation function. and These represent the vector components corresponding to the k-th hard requirement feature for both the job seeker and the position; n represents the number of hard requirement features.

[0076] Calculate the soft preference matching degree based on the soft preference feature subvectors of the job seeker and the job position:

[0077]

[0078] Among them, M soft This indicates the degree of matching of the soft preferences; V represents the weight of the l-th soft preference feature; l soft and V l ′ soft ...

[0079] The unbiased feature matching degree is calculated based on the hard requirement matching degree and the soft preference matching degree:

[0080] M nonbias =α·M hard +(1-α)·M soft;

[0081] Among them, M nonbias α represents the unbiased feature matching degree; α represents the balance coefficient.

[0082] Furthermore, the similarity calculation function adopts different calculation methods according to the feature type. For text description features, word embedding technology is used to convert the text description into vectors, and cosine similarity is used as the similarity calculation function. For classification features, one-hot encoding is used, and the similarity is determined by judging whether the encoded vectors are equal. When they are equal, the similarity value is 1, and when they are not equal, the similarity value is 0. For numerical features, numerical standardization is used, and the similarity is calculated based on the difference between the standardized numerical values.

[0083] The unbiased feature matching degree is calculated based on the feature subvectors of hard requirements and soft preferences. This clearly distinguishes the core matching factors between the job and the job seeker and the possible bias factors, effectively avoiding excessive interference of non-professional skill factors on the matching results, improving the initial accuracy and fairness of job matching, and providing more reliable data for subsequent screening and training.

[0084] Furthermore, constructing the biased feedback dataset based on the non-biased feature matching degree and the historical job application feedback dataset includes:

[0085] Define the biased feedback dataset by setting a non-biased matching threshold:

[0086] D bias ={d h |M nonbias (d h )>θ and Y h =0};

[0087] Among them, D bias This represents the bias feedback dataset; d h This represents the h-th historical job application feedback record; M nonbias (d h ) represents the unbiased feature matching degree corresponding to the h-th historical job application feedback data; θ represents the unbiased matching degree threshold; Y h Y represents the hiring result of the historical job application feedback data mentioned in item h. h =0 indicates not hired, Y h =1 indicates acceptance.

[0088] Each historical job application feedback data entry includes the job seeker's feature vector, the job feature vector of one of the positions the job seeker applied for, and the corresponding job application result.

[0089] By constructing a biased feedback dataset based on a threshold set according to the non-biased feature matching degree, it is possible to quickly and accurately identify the parts of historical job application feedback data that may be influenced by bias. This allows subsequent data processing and model optimization to address the bias issues in historical data in a targeted manner, reduce the negative impact on model training, improve the model's ability to learn fair and impartial matching, and enhance the overall performance of the model.

[0090] Table 3 records an example of the calculation results for the unbiased feature matching degree, where the balance coefficient is set to 0.6. The unbiased matching degree threshold is set to 0.8. As can be seen from Tables 2 and 3, the unbiased feature matching degree of the historical job application feedback data with person-job matching ID M003 is greater than the unbiased matching degree threshold, and the job application result is not hired. Therefore, the historical job application feedback data M003 is regarded as a biased feedback data.

[0091] Table 3 Examples of Unbiased Feature Matching Degrees

[0092] Person-job matching ID Hard requirement for matching degree Soft preference matching degree Unbiased feature matching degree M001 0.95 0.85 0.91 M002 0.90 0.82 0.87 M003 0.85 0.75 0.81 M004 0.98 0.88 0.94 M005 0.82 0.70 0.77

[0093] Furthermore, such as Figure 2 As shown, the extraction of unbiased feedback data based on the biased feedback dataset and the historical job application feedback dataset, and the reconstruction and optimization of the historical job application feedback dataset based on the unbiased feedback data, include:

[0094] The biased feedback dataset is deleted from the historical job application feedback dataset to obtain the reconstructed job application feedback dataset;

[0095] Job success feedback data is extracted from the reconstructed job search feedback dataset, and a job-person matching similarity matrix is ​​calculated; each element in the job-person matching similarity matrix represents the job-person matching similarity between each job success feedback data and each biased feedback data.

[0096] The job application success feedback data with a job-person matching similarity greater than the job-person matching threshold are selected as the unbiased feedback data.

[0097] For each piece of unbiased feedback data, count the number of times the person-job matching similarity is greater than the person-job matching threshold, calculate the bias weight, and sample the unbiased feedback data according to the bias weight to obtain an optimized unbiased feedback dataset;

[0098] The reconstructed job application feedback dataset and the optimized unbiased feedback dataset are combined to obtain the reconstructed optimized feedback dataset.

[0099] Furthermore, the job matching similarity is calculated using cosine similarity; the calculation of bias weight includes: counting the number of times the job matching similarity of each unbiased feedback data is greater than the job matching threshold, dividing the number by the number of biased feedback data to obtain the initial weight, and normalizing the initial weight of all unbiased feedback data to obtain the bias weight.

[0100] By reconstructing and optimizing the historical job application feedback dataset, we can fully explore and utilize the effective information in the historical data, correct the biases of the original data, and make the reconstructed and optimized dataset more accurately reflect the objective laws of job matching. This will improve data quality, enhance the effectiveness of model training, and reduce the possibility of incorrect matching caused by biases in historical data.

[0101] Furthermore, a dual-tower model is used to construct a human-job intelligent matching model, and the loss function is set as follows:

[0102] L = L match +β1·L bias +β2·L reg ;

[0103]

[0104] L reg =||W||2;

[0105] Where L represents the loss function; L match L represents the matching degree loss; bias Indicates bias control loss; L reg β1 and β2 represent the regularization loss; N represents the number of samples; y i The matching label value for sample i; m i This represents the matching degree of sample i in the model output; log() represents the logarithmic function; Let represent the unbiased feature matching degree of sample i; W represents the set of weights of the model; |||2 represents the Euclidean norm.

[0106] By adopting a dual-tower model and setting a comprehensive loss function that includes matching degree loss, bias control loss, and regularization loss, the model can comprehensively consider the accuracy, fairness, and prevention of overfitting during training. The dual-tower model structure helps to better handle the feature information of job seekers and positions, and the loss function can improve the model's adaptability and decision-making ability, ensuring that the matching results output by the model not only meet the actual job requirements but also reflect the principles of fairness and impartiality.

[0107] Further, the matching degree label values ​​for the samples are set as follows: A job seeker in the sample who did not apply for a position in the sample has a matching degree label value of λ1; a job seeker in the sample who applied for a position in the sample but did not receive an interview has a matching degree label value of λ2; a job seeker in the sample who applied for a position in the sample and received an interview but did not pass the interview has a matching degree label value of λ3; a job seeker in the sample who applied for a position in the sample and was hired has a matching degree label value of λ4. In this embodiment, λ1 = 0, λ2 = 0.3, λ3 = 0.7, and λ4 = 1.

[0108] By refining the sample matching label values ​​to cover various states of job seekers during the job search process, the model is provided with richer and more accurate training feedback information. This enables the model to learn more deeply the characteristics and patterns of different job search stages, further improving the model's understanding and judgment of the person-job matching relationship, enhancing the model's prediction accuracy and reliability, optimizing the model's training effect, and improving the model's performance in practical applications.

[0109] This invention constructs a data analysis-based intelligent job matching model management method. First, by collecting and analyzing historical recruitment data, a feature mapping dictionary is constructed, including hard requirements, soft preferences, and potential biases, to generate feature vectors for job seekers and positions, laying the data foundation for subsequent matching. Then, based on the feature vectors, the unbiased feature matching degree is calculated. A biased feedback dataset is filtered out by setting a threshold, and reconstructed and optimized using historical feedback data. Samples are extracted from highly similar successful cases, and a reconstructed and optimized feedback dataset is generated through weight calculation. This reduces the influence of historical biased data while improving the representativeness and fairness of the data. Finally, the reconstructed and optimized feedback dataset is used to construct an unbiased job matching dataset to train the intelligent job matching model. This invention can effectively eliminate bias factors in historical recruitment data, improve the fairness of job recommendation results, and significantly enhance the matching accuracy and applicability of the model.

[0110] Example 2:

[0111] This invention also provides a data analysis-based intelligent matching model management system for people and jobs, such as... Figure 3 As shown, it includes:

[0112] The historical data acquisition module collects historical recruitment data, including historical resume data, historical job posting data, and historical job application results data.

[0113] The job feature construction module constructs a feature mapping dictionary and builds job feature vectors and job feature vectors based on the feature mapping dictionary and the historical recruitment data. The job feature vectors and job feature vectors include hard requirement feature sub-vectors, soft preference feature sub-vectors, and potential bias feature sub-vectors.

[0114] The feedback data extraction module extracts the corresponding job seeker feature vector and job feature vector based on the historical job search result data to form a historical job search feedback dataset.

[0115] The biased data extraction module calculates the non-biased feature matching degree based on the job seeker feature vector and the job feature vector, and constructs a biased feedback dataset based on the non-biased feature matching degree and the historical job application feedback dataset.

[0116] The data reconstruction and optimization module extracts unbiased feedback data based on the biased feedback dataset and the historical job application feedback dataset, and reconstructs and optimizes the historical job application feedback dataset based on the unbiased feedback data to obtain a reconstructed and optimized feedback dataset.

[0117] The model training and optimization module constructs an unbiased job-person matching dataset based on the reconstructed optimization feedback dataset, trains the intelligent job-person matching model using the unbiased job-person matching dataset, applies the trained intelligent job-person matching model to job recommendation, collects subsequent job search result data, and optimizes the intelligent job-person matching model based on the job search result data.

[0118] Furthermore, the calculation steps for the unbiased feature matching degree are as follows:

[0119] Calculate the hard requirement matching degree based on the hard requirement feature subvectors of the job seeker and the position:

[0120]

[0121] Among them, M hard This indicates the degree of matching of the hard requirements; The weight of the k-th mandatory feature is represented by sim(); sim() represents the similarity calculation function. and These represent the vector components corresponding to the k-th hard requirement feature for both the job seeker and the position; n represents the number of hard requirement features.

[0122] Calculate the soft preference matching degree based on the soft preference feature subvectors of the job seeker and the job position:

[0123]

[0124] Among them, M softThis indicates the degree of matching of the soft preferences; V represents the weight of the l-th soft preference feature; l soft and V l ′ soft ...

[0125] The unbiased feature matching degree is calculated based on the hard requirement matching degree and the soft preference matching degree:

[0126] M nonbias =α·M hard +(1-α)·M soft ;

[0127] Among them, M nonbias α represents the unbiased feature matching degree; α represents the balance coefficient.

[0128] Furthermore, constructing the biased feedback dataset based on the non-biased feature matching degree and the historical job application feedback dataset includes:

[0129] Define the biased feedback dataset by setting a non-biased matching threshold:

[0130] D bias ={d h |M nonbias (d h )>θ and Y h =0};

[0131] Among them, D bias This represents the bias feedback dataset; d h This represents the h-th historical job application feedback record; M nonbias (d h ) represents the unbiased feature matching degree corresponding to the h-th historical job application feedback data; θ represents the unbiased matching degree threshold; Y h Y represents the hiring result of the historical job application feedback data mentioned in item h. h =0 indicates not hired, Y h =1 indicates acceptance.

[0132] Furthermore, the extraction of unbiased feedback data based on the biased feedback dataset and the historical job application feedback dataset, and the reconstruction and optimization of the historical job application feedback dataset based on the unbiased feedback data, include:

[0133] The biased feedback dataset is deleted from the historical job application feedback dataset to obtain the reconstructed job application feedback dataset;

[0134] Job success feedback data is extracted from the reconstructed job search feedback dataset, and a job-person matching similarity matrix is ​​calculated; each element in the job-person matching similarity matrix represents the job-person matching similarity between each job success feedback data and each biased feedback data.

[0135] The job application success feedback data with a job-person matching similarity greater than the job-person matching threshold are selected as the unbiased feedback data.

[0136] For each piece of unbiased feedback data, the number of times the person-job matching similarity is greater than the person-job matching threshold is counted, the bias weight is calculated, and the unbiased feedback data is sampled according to the bias weight to obtain an optimized unbiased feedback dataset;

[0137] The reconstructed job application feedback dataset and the optimized unbiased feedback dataset are combined to obtain the reconstructed optimized feedback dataset.

[0138] Furthermore, a dual-tower model is used to construct a human-job intelligent matching model, and the loss function is set as follows:

[0139] L = L match +β1·L bias +β2·L reg ;

[0140]

[0141] L reg =||W||2;

[0142] Where L represents the loss function; L match L represents the matching degree loss; bias Indicates bias control loss; L reg β1 and β2 represent the regularization loss; N represents the number of samples; y i The matching label value for sample i; m i This represents the matching degree of sample i in the model output; log() represents the logarithmic function; Let represent the unbiased feature matching degree of sample i; W represents the set of weights of the model; |||2 represents the Euclidean norm.

[0143] Furthermore, the matching degree label values ​​for the sample include: a job seeker in the sample who did not apply for a position in the sample, corresponding to a matching degree label value of λ1; a job seeker in the sample who applied for a position in the sample but did not get an interview, corresponding to a matching degree label value of λ2; a job seeker in the sample who applied for a position in the sample and got an interview but did not pass the interview, corresponding to a matching degree label value of λ3; and a job seeker in the sample who applied for a position in the sample and was hired, corresponding to a matching degree label value of λ4.

[0144] Table 4 compares the job-person matching effect before and after reconstructing and optimizing the historical job application feedback dataset using the method of this invention, using multiple indicators. The data shows that eliminating biased data not only increased job seeker participation but also significantly improved the final job application success rate.

[0145] Table 4 Matching Effect Analysis

[0146] Performance indicators Before optimization After optimization Improvement range Resume submission rate 15％ 28％ +13％ Interview invitation rate 20％ 35％ +15％ Job application success rate 18％ 32％ +14％

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data-driven intelligent matching model management method for personnel and positions, characterized in that, include: Collect historical recruitment data; the historical recruitment data includes historical resume data, historical job posting data, and historical job application results data; Construct a feature mapping dictionary, and construct job seeker feature vectors and job feature vectors based on the feature mapping dictionary and the historical recruitment data; the job seeker feature vectors and job feature vectors include hard requirement feature sub-vectors, soft preference feature sub-vectors, and potential bias feature sub-vectors; Based on the historical job search results data, the corresponding job seeker feature vectors and job feature vectors are extracted to form a historical job search feedback dataset. Calculate the unbiased feature matching degree based on the job seeker feature vector and the job feature vector; construct a biased feedback dataset based on the unbiased feature matching degree and the historical job application feedback dataset; set an unbiased matching degree threshold and define the biased feedback dataset. D bias = {d h | M nonbias (d h ) > θ and Y h = 0}; Among them, D bias This represents the bias feedback dataset; d h This represents the h-th historical job application feedback record; M nonbias (d h ) represents the unbiased feature matching degree corresponding to the h-th historical job application feedback data; θ represents the unbiased matching degree threshold; Y h Y represents the hiring result of the historical job application feedback data mentioned in item h. h =0 indicates not hired, Y h =1 indicates acceptance; Extracting unbiased feedback data from the biased feedback dataset and the historical job application feedback dataset, and then reconstructing and optimizing the historical job application feedback dataset based on the unbiased feedback data includes: deleting the biased feedback dataset from the historical job application feedback dataset to obtain a reconstructed job application feedback dataset; extracting successful job application feedback data from the reconstructed job application feedback dataset and calculating a job-person matching similarity matrix; each element in the job-person matching similarity matrix represents the job-person matching similarity between each successful job application feedback data and each biased feedback data; selecting successful job application feedback data with a job-person matching similarity greater than a job-person matching threshold as unbiased feedback data; for each unbiased feedback data, counting the number of times the job-person matching similarity is greater than the job-person matching threshold, calculating a bias weight, and sampling the unbiased feedback data based on the bias weight to obtain an optimized unbiased feedback dataset; and combining the reconstructed job application feedback dataset and the optimized unbiased feedback dataset to obtain a reconstructed optimized feedback dataset. An unbiased person-job matching dataset is constructed based on the reconstructed and optimized feedback dataset, and the person-job intelligent matching model is trained using the unbiased person-job matching dataset. The trained intelligent job matching model is applied to job recommendation, and subsequent job search results data are collected. The intelligent job matching model is then optimized based on the job search results data.

2. The data analysis-based intelligent matching model management method for personnel and positions according to claim 1, characterized in that, The steps for calculating the unbiased feature matching degree are as follows: Calculate the hard requirement matching degree based on the hard requirement feature subvectors of the job seeker and the position: Among them, M hard This indicates the degree of matching of the hard requirements; The weight of the k-th mandatory feature is represented by sim(); sim() represents the similarity calculation function. and These represent the vector components corresponding to the k-th hard requirement feature for both the job seeker and the position; n represents the number of hard requirement features. Calculate the soft preference matching degree based on the soft preference feature subvectors of the job seeker and the job position: Among them, M soft This indicates the degree of matching of the soft preferences; V represents the weight of the l-th soft preference feature; l soft and V l ′ soft ... The unbiased feature matching degree is calculated based on the hard requirement matching degree and the soft preference matching degree: M nonbias =α·M hard +(1-a)·M soft ; Among them, M nonbias α represents the unbiased feature matching degree; α represents the balance coefficient.

3. The data analysis-based intelligent matching model management method for personnel and positions according to claim 1, characterized in that, A dual-tower model is used to construct an intelligent person-job matching model, and the loss function is set as follows: L=L match +β1·L bias +β2·L reg ; L reg =||W||2; Where L represents the loss function; L match L represents the matching degree loss; bias Indicates bias control loss; L reg β1 and β2 represent the regularization loss; N represents the number of samples; y i The matching label value for sample i; m i This represents the matching degree of sample i in the model output; log() represents the logarithmic function; Let represent the unbiased feature matching degree of sample i; W represents the set of weights of the model; |||2 represents the Euclidean norm.

4. The data analysis-based intelligent matching model management method for personnel and positions according to claim 3, characterized in that, The matching degree label values ​​for the sample are set as follows: For job seekers in the sample who did not apply for any of the positions in the sample, the matching degree label value is λ1; for job seekers in the sample who applied for any of the positions in the sample but did not receive an interview, the matching degree label value is λ2; for job seekers in the sample who applied for any of the positions in the sample and received an interview but did not pass the interview, the matching degree label value is λ3; for job seekers in the sample who applied for any of the positions in the sample and were hired, the matching degree label value is λ4.

5. A data analysis-based intelligent matching model management system for people and jobs, characterized in that, include: The historical data acquisition module collects historical recruitment data, including historical resume data, historical job posting data, and historical job application results data. The job feature construction module constructs a feature mapping dictionary and builds job feature vectors and job feature vectors based on the feature mapping dictionary and the historical recruitment data. The job feature vectors and job feature vectors include hard requirement feature sub-vectors, soft preference feature sub-vectors, and potential bias feature sub-vectors. The feedback data extraction module extracts the corresponding job seeker feature vector and job feature vector based on the historical job search result data to form a historical job search feedback dataset. The bias data extraction module calculates the non-biased feature matching degree based on the job seeker feature vector and the job feature vector, and constructs a bias feedback dataset based on the non-biased feature matching degree and the historical job application feedback dataset; it sets a non-biased matching degree threshold and defines the bias feedback dataset. Among them, D bias This represents the bias feedback dataset; d h This represents the h-th historical job application feedback record; M nonbias (d h ) represents the unbiased feature matching degree corresponding to the h-th historical job application feedback data; θ represents the unbiased matching degree threshold; Y h Y represents the hiring result of the historical job application feedback data mentioned in item h. h =0 indicates not hired, Y h =1 indicates acceptance; The data reconstruction and optimization module extracts unbiased feedback data based on the biased feedback dataset and the historical job application feedback dataset, and reconstructs and optimizes the historical job application feedback dataset based on the unbiased feedback data, including: deleting the biased feedback dataset from the historical job application feedback dataset to obtain a reconstructed job application feedback dataset; extracting successful job application feedback data from the reconstructed job application feedback dataset and calculating a job-person matching similarity matrix; each element in the job-person matching similarity matrix represents the job-person matching similarity between each successful job application feedback data and each biased feedback data; selecting successful job application feedback data with a job-person matching similarity greater than a job-person matching threshold as unbiased feedback data; for each unbiased feedback data, counting the number of times the job-person matching similarity is greater than the job-person matching threshold, calculating a bias weight, sampling the unbiased feedback data based on the bias weight to obtain an optimized unbiased feedback dataset; and combining the reconstructed job application feedback dataset and the optimized unbiased feedback dataset to obtain a reconstructed optimized feedback dataset. The model training and optimization module constructs an unbiased job-person matching dataset based on the reconstructed optimization feedback dataset, trains the intelligent job-person matching model using the unbiased job-person matching dataset, applies the trained intelligent job-person matching model to job recommendation, collects subsequent job search result data, and optimizes the intelligent job-person matching model based on the job search result data.

Images