Dynamic post demand model construction method based on fuzzy clustering

Abstract

The invention discloses a dynamic post demand model construction method based on fuzzy clustering. The method comprises the following steps: S1, collecting a comprehensive post demand data set related to post demands; s2, generating a feature vector set of the comprehensive post demand data set; s3, performing clustering analysis by applying an improved fuzzy C-means clustering algorithm to generate a category membership matrix; s4, designing and training a first-layer generative adversarial network by taking the key feature vector and the category membership matrix as input; s5, designing and training a second-layer generative adversarial network by taking the potential feature vector and post demand historical feedback data as input; and S6, dynamically optimizing the post demand model. According to the method, the reasonability and the accuracy of post demand characteristics generated by the model are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of job requirements, and particularly to a method for dynamically constructing a job requirements model based on fuzzy clustering. Background Art

[0002] With the rapid development of intelligent and big data technologies, enterprise recruitment has gradually transformed from the traditional mode to a data-driven and intelligent direction. The dynamic construction of the job requirements model has become one of the key technologies to improve recruitment efficiency and job matching accuracy. However, with the increase in enterprise business complexity and the rapid change of the market recruitment environment, the job requirements show the characteristics of diversity, uncertainty, and dynamic change, which pose higher requirements for traditional job requirements modeling methods.

[0003] Currently, most enterprises use static models or linear algorithms to construct job requirements models. Usually, feature extraction and statistical analysis with fixed rules are performed on historical recruitment data to generate feature descriptions matching the jobs. However, traditional methods have significant limitations: on the one hand, static models are difficult to adapt to the dynamic changes of job requirements and usually require manual adjustment of model parameters to adapt to market changes, resulting in lagging updates and low efficiency; on the other hand, linear algorithms perform poorly in dealing with complex, high-dimensional, and fuzzy data and are difficult to capture the non-linear relationships hidden in job requirements, such as the fuzzy expression of job skill requirements and the dynamic changes in the supply and demand of the recruitment market.

[0004] In recent years, technologies based on machine learning and clustering analysis have begun to be applied to job requirements modeling. Some studies have tried to use clustering algorithms to classify job data features and improve the prediction ability of the model through deep learning methods. However, existing technologies still have deficiencies in practical applications: traditional clustering algorithms perform limitedly in dealing with diverse and fuzzy job requirement features and are difficult to fully explore the hidden feature relationships between jobs; at the same time, although deep learning methods have improved in feature extraction, they cannot respond in a timely manner to the changes in the real-time recruitment environment due to the lack of a dynamic adjustment mechanism.

[0005] In summary, existing technologies have significant deficiencies in dealing with job requirement diversity, uncertainty, and dynamic changes. Specifically, they have weak processing ability for complex data, lack of dynamic response ability, and imperfect model real-time optimization mechanism. The defects of existing technologies directly affect the recruitment efficiency of enterprises and the accuracy of job matching. There is an urgent need for a new method that can integrate multiple data sources, dynamically adjust the job requirements model, and efficiently process fuzzy data to solve the above problems. Summary of the Invention

[0006] An object of the present invention is to propose a method for dynamically constructing a job requirements model based on fuzzy clustering, which significantly improves the rationality and accuracy of the job requirement features generated by the model.

[0007] A method for dynamically constructing a job demand model based on fuzzy clustering according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect a comprehensive job demand data set related to job demands;

[0009] S2. Clean the collected comprehensive job demand data set, remove noise data, perform normalization processing and standardization conversion on the comprehensive job demand data set, and generate a feature vector set of the comprehensive job demand data set through text vectorization technology;

[0010] S3. Based on the feature vector set of the comprehensive job demand data set, apply an improved fuzzy C-means clustering algorithm for clustering analysis to generate a category membership matrix, and extract key feature vectors of each category according to the category membership matrix;

[0011] S4. Taking the key feature vectors and the category membership matrix as inputs, design and train a first-layer generative adversarial network. Generate new latent feature vectors through the generator of the first-layer generative adversarial network, optimize the generator output through the discriminator of the first-layer generative adversarial network, and the generated latent feature vectors constitute an enhanced data set for further job demand modeling;

[0012] S5. Taking the latent feature vectors and historical feedback data of job demands as inputs, design and train a second-layer generative adversarial network. Generate a dynamically optimized job demand model through the generator of the second-layer generative adversarial network, evaluate the rationality of the job demand model through the discriminator of the second-layer generative adversarial network, and the generated job demand model is used as the output result of the current dynamic job demand model;

[0013] S6. Real-time collect newly added job demand data, market feedback data, and recruitment result data of the enterprise, input the real-time collected data into the corresponding modules of steps S1 to S5, update the category membership matrix of fuzzy clustering analysis, update the parameters of the first-layer and second-layer generative adversarial networks, and dynamically optimize the job demand model.

[0014] Optionally, the step S1 includes the following sub-steps:

[0015] S11. Obtain a historical recruitment data set D formed during the recruitment process within the enterprise, including job descriptions, recruitment conditions, interview results, and final employment information;

[0016] S12. Extract a job description data set D from the existing job descriptions or newly added job demand information of the enterprise p , including job names, job responsibilities, skill requirements, and educational requirements;

[0017] S13. Obtain the market demand data set D from the external recruitment market m , including industry recruitment trends, job posting information, and market supply and demand analysis;

[0018] S14. Merge the historical recruitment data set, job description data set, and market demand data set to form a comprehensive job demand data set:

[0019] D = D h ∪D p ∪D m .

[0020] Optionally, the step S2 includes the following sub-steps:

[0021] S21. Clean each piece of data in the comprehensive job demand data set D, removing redundant data, outliers, and missing values, to form a cleaned comprehensive job demand data set;

[0022] S22. Normalize each data record in the cleaned comprehensive job demand data set, mapping the numerical attributes of the data to the interval [0, 1];

[0023] S23. Standardize the normalized comprehensive job demand data set, converting the numerical attributes to a standard normal distribution. The standardized results of all data records constitute the standardized comprehensive job demand data set D s ;

[0024] S24. Vectorize the non-numerical attributes in the standardized comprehensive job demand data set D s , and use the word embedding technique to convert the text fields into vector forms. Define the text vector set as:

[0025] F t = {f t1 , f t2 , …, f tn};

[0026] Among them, f tn represents the vector representation of the nth text record, and n is the total number of text records;

[0027] S25. Merge the text vector set with the comprehensive job demand data set to form a feature vector set of the comprehensive job demand data:

[0028] F = F t ∪D s .

[0029] Optionally, the step S3 includes the following sub-steps:

[0030] S31. Set the clustering parameters of the job requirement model, including the number of clustering categories c, the maximum number of iterations T, the fuzzy factor m > 1, the dynamic adaptation factor λ, and the membership threshold ∈, where c represents the target number of categories set according to the diversity characteristics of job requirements, m controls the degree of fuzziness of job requirement characteristics during the clustering process, λ is the dynamic adaptation factor used to adjust the sensitivity of the clustering result according to the real-time data stream, reflecting the adaptability of the job requirement model to real-time changes;

[0031] S32. Jointly process the comprehensive job requirement feature vector set F = {f 1 , f 2 , …, f n} and the real-time data stream feature vector set F r = {f r1 , f r2 , …, f rk} to generate a dynamically initialized membership matrix U (0) :

[0032] U (0) = [u ij ;

[0033]

[0034] Among them, u ij represents the initial membership of the comprehensive job requirement data f i to the j-th category, and φ(f i , f rj ) is the feature similarity function used to dynamically combine the real-time data stream to adjust the initialization result;

[0035] S33. Dynamically calculate the center vector of each category by combining the real-time data stream and the initialized membership matrix:

[0036]

[0037] Among them, represents the dynamic category center of the j-th category, is the membership value of the membership matrix, f i is the comprehensive job requirement feature vector, f r is the real-time data stream feature vector, and t 1 is the current iteration number;

[0038] S34. Update the membership matrix according to the dynamic category center

[0039]

[0040] Among them, ​Indicates the job requirement feature f i and the dynamic category center of the Euclidean distance;

[0041] S35. Repeat steps S33 and S34 until the following dynamic convergence conditions are met:

[0042]

[0043] S36. Based on the final membership matrix U (T) and the category center set V = {v 1 , v 2 , …, v c}, extract the key feature vector v c of the job requirement model, where v c represents the dynamic feature of each job category.

[0044] Optionally, step S4 includes the following sub-steps:

[0045] S41. Design the first-layer generative adversarial network, including the first-layer generator G 1 (z, U (T) ) and the first-layer discriminator D 1 (x), where G 1 (z, U (T) represents the first-layer generator with the random noise vector z and the membership matrix U (T) as inputs, used to generate the latent feature vector of the job requirement model, and D 1 (x) represents the first-layer discriminator for discriminating whether the input feature vector is a real job requirement feature vector;

[0046] S42. Initialize the parameters of the first-layer generator and the first-layer discriminator, including the weight matrix W g , the bias term b g , the weight matrix W d of the discriminator, and the bias term b d , and set the learning rate α and the optimization objective of the generative adversarial network;

[0047] S43. Input the category center set V and the random noise vector z into the first-layer generator to generate the latent feature vector set F g ;

[0048] S44. Input the latent feature vector set F g generated by the first-layer generator and the real feature vector set F into the first-layer discriminator D 1 (x) at the same time, and calculate the discrimination degree between the feature vectors generated by the generator and the real feature vectors through the first-layer discriminator;

[0049] S45. Define the loss function L of the first - layer generative adversarial network as the alternating adversarial optimization objective of the first - layer generator and the first - layer discriminator:

[0050]

[0051] Among them, P data is the distribution of real feature vectors, and P z is the distribution of random noise. By optimizing the parameters of the first - layer generator and the first - layer discriminator, update the weights and bias terms of the generator and the weights and bias terms of the first - layer discriminator;

[0052] S46. Take the set of latent feature vectors F g output by the optimized first - layer generator as the enhanced job - requirement feature dataset.

[0053] Optionally, the step S5 includes the following sub - steps:

[0054] S51. Construct an adaptively optimized second - layer generative adversarial network model, including a second - layer generator G 2 (F g , H, A) and a second - layer discriminator D 2 (M, A), where H is the set of historical feedback data on job requirements, reflecting the historical information of the past recruitment of the enterprise matching job requirements, A is the set of job - requirement adaptability parameter, reflecting the sensitivity and adaptability of the job - requirement model to market dynamic changes, and M is the generated dynamic optimized job - requirement model;

[0055] S52. Initialize the set of job - requirement adaptability parameters A, including:

[0056] The job - requirement change rate a rate , indicating the speed of job - requirement change:

[0057]

[0058] Among them, H t and H t-1 represent the job - requirement feedback data at the current and previous moments respectively, and t is the time interval;

[0059] The job - skill deviation a skill , indicating the deviation degree between job - skill requirements and market skill supply:

[0060]

[0061] Among them, and represent the job - requirement skills and market actual skills respectively;

[0062] The job - supply - demand weight amarket , representing the weight ratio between the market job demand and the candidate supply:

[0063]

[0064] Among them, job_openings and applicants respectively represent the number of job postings and the number of job seekers;

[0065] S53. Input the set of potential feature vectors F g , the set of historical feedback data of job demands H, and the set of job demand adaptability parameters A into the second-layer generator G 2 , to generate a dynamically optimized job demand model M:

[0066] M = G 2 (F g , H, A);

[0067] S54. Input the job demand model M output by the second-layer generator and the adaptability parameter A into the second-layer discriminator D 2 , to discriminate the rationality and dynamic adaptability of the job demand model, and output the discrimination result:

[0068] D 2 (M, A) = σ(W d2 · M + W a · A + b d2 );

[0069] Among them, W d2 and b d2 are the weights and bias terms of the discriminator, which determine the sensitivity of the discrimination process to the model features. W a is the adaptability parameter weight, and σ is the activation function;

[0070] S55. Define the loss function L of the second-layer generative adversarial network 2 , and add an adaptability weight constraint on the basis of the generative adversarial loss:

[0071]

[0072] Among them, P real represents the distribution of the true job demand model, P g represents the distribution of the potential feature vectors, γ is the adaptability weight adjustment factor, and A target is the ideal value of the adaptability parameter;

[0073] S56. Use the job demand model M output by the optimized second-layer generator as the final dynamic optimization result for job demand matching and recruitment optimization.

[0074] The beneficial effects of the present invention are:

[0075] (1) The present invention improves the traditional fuzzy C-means clustering algorithm. By introducing a real-time data dynamic adaptation mechanism and a category membership degree self-adaptive adjustment strategy, the algorithm can effectively handle the ambiguity and uncertainty problems existing in job demand data. By dynamically adjusting the fuzzy factor and the adaptation parameter, the response ability to the dynamically changing job demand data is enhanced. Combining the real-time market feedback data can more accurately extract the key features of the job, significantly improving the model's ability to depict the diversity and fuzzy boundaries of job categories. Compared with the traditional static clustering algorithm, the accuracy and generalization ability of job demand feature extraction are improved.

[0076] (2) The present invention designs a two-layer generative adversarial network structure. The first-layer generative adversarial network solves the problem of data sparsity of job demand feature data by combining the fuzzy clustering results to generate potential feature vectors. The second-layer generative adversarial network takes the potential feature vectors and historical feedback data as inputs and generates a real-time optimized job demand model by dynamically adjusting the set of adaptation parameters, which can adapt to the rapidly changing recruitment environment. By introducing an adaptive constraint weight, the second-layer generative adversarial network can fully consider the impact of market changes when optimizing the job demand model, significantly improving the rationality and accuracy of the job demand features generated by the model.

[0077] (3) The present invention proposes a set of self-adaptive job demand parameter systems, including the job demand change rate, the job skill deviation degree, and the job supply-demand weight, which are used to capture the dynamic changes in the market and the adjustment direction of job demands in real time. By combining the dynamic optimization mechanism of the generative adversarial network, the efficient integration of real-time market data and the historical feedback of enterprise recruitment is realized in the construction of the job demand model, thereby dynamically adjusting the model output results, significantly enhancing the adaptability of the job demand model to market trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0079] Figure 1 is a flowchart of a method for dynamically constructing a job demand model based on fuzzy clustering proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0080] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0081] Refer to Figure 1 , a method for dynamically constructing a job demand model based on fuzzy clustering, including the following steps:

[0082] S1. Collect a comprehensive job demand dataset related to job requirements;

[0083] S2. Clean the collected comprehensive job demand dataset, remove noise data, perform normalization processing and standardization transformation on the comprehensive job demand dataset, and generate a feature vector set of the comprehensive job demand dataset through text vectorization technology;

[0084] S3. Based on the feature vector set of the comprehensive job demand dataset, apply an improved fuzzy C-means clustering algorithm for clustering analysis, generate a class membership matrix, and extract the key feature vectors of each class according to the class membership matrix;

[0085] S4. Using the key feature vectors and the class membership matrix as inputs, design and train the first-layer generative adversarial network. Generate new latent feature vectors through the generator of the first-layer generative adversarial network, optimize the generator output through the discriminator of the first-layer generative adversarial network, and the generated latent feature vectors constitute an enhanced dataset for further job demand modeling;

[0086] S5. Using the latent feature vectors and historical feedback data of job demands as inputs, design and train the second-layer generative adversarial network. Generate a dynamically optimized job demand model through the generator of the second-layer generative adversarial network, evaluate the rationality of the job demand model through the discriminator of the second-layer generative adversarial network, and the generated job demand model is used as the output result of the current dynamic job demand model;

[0087] S6. Real-time collect newly added job demand data, market feedback data, and recruitment result data of the enterprise, input the real-time collected data into the corresponding modules of steps S1 to S5, update the class membership matrix of fuzzy clustering analysis, update the parameters of the first-layer and second-layer generative adversarial networks, and dynamically optimize the job demand model.

[0088] In this embodiment, step S1 includes the following sub-steps:

[0089] S11. Obtain the historical recruitment data set D formed during the recruitment process within the enterprise, including job descriptions, recruitment conditions, interview results, and final employment information;

[0090] S12. Extract the job description data set D p , including job names, job responsibilities, skill requirements, and educational requirements;

[0091] S13. Obtain the market demand data set D from the external recruitment market m , including industry recruitment trends, job posting information, and market supply and demand analysis;

[0092] S14. Merge the historical recruitment data set, the job description data set, and the market demand data set to form a comprehensive job demand data set:

[0093] D = D h ∪D p ∪D m 。

[0094] In this embodiment, step S2 includes the following sub-steps:

[0095] S21. Clean each piece of data in the comprehensive job demand data set D, removing redundant data, outliers, and missing values, to form a cleaned comprehensive job demand data set;

[0096] S22. Normalize each data record in the cleaned comprehensive job demand data set, mapping the numerical attributes of the data to the interval [0, 1];

[0097] S23. Standardize the normalized comprehensive job demand data set, converting the numerical attributes to a standard normal distribution. The standardized results of all data records constitute the standardized comprehensive job demand data set D s ;

[0098] S24. Vectorize the non-numerical attributes in the standardized comprehensive job demand data set D s Use the word embedding technique to convert the text fields into vector form, and define the text vector set as:

[0099] F t ={f t1 ,f t2 ,…,f tn};

[0100] where f tn represents the vector representation of the nth text record, and n is the total number of text records;

[0101] S25. Merge the text vector set with the comprehensive job demand data set to form a feature vector set of the comprehensive job demand data:

[0102] F = F t ∪D s 。

[0103] In this embodiment, step S3 includes the following sub-steps:

[0104] S31. Set the clustering parameters of the job demand model, including the number of clustering categories c, the maximum number of iterations T, the fuzzy factor m > 1, the dynamic adaptation factor λ, and the membership threshold ∈. Here, c represents the target number of categories set according to the diversity characteristics of job demands, m controls the degree of fuzziness of job demand characteristics during the clustering process, λ is the dynamic adaptation factor used to adjust the sensitivity of the clustering result according to the real-time data stream, reflecting the adaptability of the job demand model to real-time changes;

[0105] S32. Jointly process the comprehensive job demand feature vector set F = {f 1 , f 2 , …, f n} and the real-time data stream feature vector set F r = {f r1 , f r2 , …, f rk} to generate a dynamically initialized membership matrix U (0) :

[0106] U (0) = [u ij ;

[0107]

[0108] where u ij represents the initial membership of the comprehensive job demand data f i to the j-th category, and φ(f i , f rj ) is the feature similarity function used to dynamically combine the real-time data stream to adjust the initialization result;

[0109] S33. Dynamically calculate the center vector of each category by combining the real-time data stream and the initialized membership matrix:

[0110]

[0111] where, represents the dynamic category center of the j-th category, is the membership value of the membership matrix, f i is the comprehensive job demand feature vector, f r is the real-time data stream feature vector, and t 1 is the current iteration number;

[0112] S34. Update the membership matrix according to the dynamic category center

[0113]

[0114] where, ​Indicates the job requirement feature f i and the dynamic category center of the Euclidean distance;

[0115] S35. Repeat steps S33 and S34 until the following dynamic convergence condition is met:

[0116]

[0117] S36. Based on the final membership matrix U (T) and the category center set V = {v 1 , v 2 , …, v c} extract the key feature vector v c of the job requirement model, where v c represents the dynamic feature of each job category.

[0118] In this embodiment, step S4 includes the following sub-steps:

[0119] S41. Design the first-layer generative adversarial network, including the first-layer generator G 1 (z, U (T) ) and the first-layer discriminator D 1 (x), where G 1 (z, U (T) represents the first-layer generator with the random noise vector z and the membership matrix U (T) as inputs, which is used to generate the latent feature vector of the job requirement model, and D 1 (x) represents the first-layer discriminator for discriminating whether the input feature vector is a real job requirement feature vector;

[0120] S42. Initialize the parameters of the first-layer generator and the first-layer discriminator, including the weight matrix W g , the bias term b g , the weight matrix W d of the discriminator, and the bias term b d , and set the learning rate α and the optimization objective of the generative adversarial network;

[0121] S43. Input the category center set V and the random noise vector z into the first-layer generator to generate the latent feature vector set F g ;

[0122] S44. Input the latent feature vector set F g generated by the first-layer generator and the real feature vector set F into the first-layer discriminator D 1 (x) at the same time, and calculate the discrimination degree between the feature vector generated by the generator and the real feature vector through the first-layer discriminator;

[0123] S45. Define the loss function L of the first-layer generative adversarial network as the alternating adversarial optimization objective of the first-layer generator and the first-layer discriminator:

[0124]

[0125] Among them, P data is the distribution of real feature vectors, and P z is the distribution of random noise. By optimizing the parameters of the first-layer generator and the first-layer discriminator, update the weights and bias terms of the generator and the weights and bias terms of the first-layer discriminator;

[0126] S46. Use the set F g of latent feature vectors output by the optimized first-layer generator as the enhanced job demand feature dataset.

[0127] In this embodiment, step S5 includes the following sub-steps:

[0128] S51. Construct an adaptively optimized second-layer generative adversarial network model, including a second-layer generator G 2 (F g , H, A) and a second-layer discriminator D 2 (M, A), where H is the set of historical feedback data on job demands, reflecting the historical information of the enterprise's past recruitment matching job demands, A is the set of job demand adaptability parameters, reflecting the sensitivity and adaptability of the job demand model to market dynamic changes, and M is the generated dynamic optimized job demand model;

[0129] S52. Initialize the set A of job demand adaptability parameters, including:

[0130] The job demand change rate a rate , indicating the speed of job demand change:

[0131]

[0132] Among them, H t and H t-1 represent the job demand feedback data at the current and previous moments respectively, and t is the time interval;

[0133] The job skill deviation a skill , indicating the degree of deviation between job skill requirements and market skill supply:

[0134]

[0135] Among them, and represent the job demand skills and the actual market skills respectively;

[0136] Job supply - demand weight a market , representing the weight ratio between the market job demand and candidate supply:

[0137]

[0138] where job_openings and applicants represent the number of job postings and the number of job seekers respectively;

[0139] S53. Input the set of potential feature vectors F g , the set of historical feedback data of job demands H, and the set of job demand adaptability parameters A into the second - layer generator G 2 to generate an optimized dynamic job demand model M:

[0140] M = G 2 (F g , H, A);

[0141] S54. Input the job demand model M and the adaptability parameter A output by the second - layer generator into the second - layer discriminator D 2 to judge the rationality and dynamic adaptability of the job demand model and output the judgment result:

[0142] D 2 (M, A)=σ(W d2 ·M + W a ·A + b d2 );

[0143] where W d2 and b d2 are the weights and bias terms of the discriminator, which determine the sensitivity of the discrimination process to the model features. W a is the adaptability parameter weight, and σ is the activation function;

[0144] S55. Define the loss function L of the second - layer generative adversarial network 2 , and add an adaptability weight constraint on the basis of the generative adversarial loss:

[0145]

[0146] where P real represents the distribution of the real job demand model, P g represents the distribution of the potential feature vectors, γ is the adaptability weight adjustment factor, and A target is the ideal adaptability parameter value;

[0147] S56. Use the job demand model M output by the optimized second - layer generator as the final dynamic optimization result for job demand matching and recruitment optimization.

[0148] Example 1:

[0149] A large Internet enterprise is located in Area A. Its core business covers software development, cloud computing, and artificial intelligence. During the enterprise's expansion process, in order to cope with the rapid development of different technical fields, the company plans to complete the recruitment tasks for 50 positions within the next three months, covering positions such as software development engineers, algorithm engineers, and product managers. Due to the diverse types of positions and the rapid changes in requirements, coupled with fierce market competition, the traditional method of modeling job requirements can no longer meet the company's efficient recruitment needs. The existing methods perform poorly in dealing with the ambiguity and dynamic adjustment of job skill requirements, resulting in low recruitment efficiency and inaccurate candidate matching. Therefore, the company decides to adopt the method for dynamically constructing job requirement models based on fuzzy clustering and double-layer generative adversarial networks of the present invention, aiming to improve recruitment efficiency and job matching accuracy.

[0150] The company imported the historical recruitment data of the past year from the human resource management system, including 15,000 recruitment records. Each record contains job name, recruitment requirements, skill requirements, salary range, and employment situation. At the same time, the company extracted the existing job description data by combining job specifications and collected the job posting information and market supply and demand data of competing companies in the industry through recruitment platforms. After data cleaning and normalization processing, a comprehensive job requirement dataset was generated, which contains a total of 18,000 data records.

[0151] After normalizing and standardizing the cleaned data, the data is converted into the form of feature vectors. The feature vector of a certain job requirement data is:

[0152] {Job name = Algorithm engineer, Python = 1, C++ = 0.8, TensorFlow = 0.9, Monthly salary =

[0153] 20,000 - 30,000, Education = Master};

[0154] Through vectorization technology, these non-numerical features are mapped into the vector space to generate a set of feature vectors for the comprehensive job requirement data.

[0155] Apply the improved fuzzy C-means clustering algorithm to the set of feature vectors, set the number of clustering categories to 6, and cluster different jobs. The algorithm dynamically combines real-time market feedback data, sets the fuzzy factor m to 1.8, and the dynamic adaptation factor to 0.5. The final obtained category membership matrix shows that the membership degree of a certain job requirement data to the "Algorithm engineer" category is 0.85, and the membership degree to the "Software development engineer" category is 0.15.

[0156] The clustering results show that the key feature vectors of different categories include the following information:

[0157] Category 1 (Algorithm Engineer): The main skill requirements are Python and TensorFlow. A master's degree is required, and more than 3 years of work experience is needed.

[0158] Category 2 (Software Development Engineer): The main skill requirements are Java and Spring. A bachelor's degree is required, and more than 2 years of work experience is needed.

[0159] Category 3 (Product Manager): Mainly focus on user requirement analysis ability and project management experience.

[0160] Using the fuzzy clustering results, construct the first-layer generative adversarial network. The generator takes a random noise vector and a category membership matrix as inputs, generates a set of latent feature vectors for enhancing sparse data. A certain feature vector simulated and generated by the generator is:

[0161] {Job Title = Algorithm Engineer, PyTorch = 0.75, TensorFlow = 0.85, Monthly Salary = 28000}

[0162] The first-layer discriminator evaluates the generated feature vectors and compares them with the real data to optimize the output quality of the generator. After multiple rounds of iteration, the latent feature vectors generated by the generator have a high similarity with the actual job requirement features.

[0163] Combining the latent feature vectors generated in the first layer and the enterprise historical recruitment feedback data, input them into the second-layer generator. Historical feedback data: The screening rate of a certain job in the interview stage is 70%, and the employment rate is 25%. The adaptability parameter set includes the job requirement change rate and the job skill deviation degree. The job skill deviation degree is calculated as:

[0164]

[0165] The generator generates a dynamically optimized job requirement model under the guidance of the adaptability parameters. The optimized job requirement for the algorithm engineer position is:

[0166] {Python = 1, TensorFlow = 0.9, PyTorch = 0.8, Monthly Salary = 28000 - 32000, Education =

[0167] Master}

[0168] The second-layer discriminator ensures the rationality and adaptability of the generated job requirement model through comparative evaluation.

[0169] The experiment was carried out in the scenario of a large job fair in Area A, divided into two groups of tests: the traditional method and the method of the present invention. The recruitment period was 2 months, and the number of recruitment positions was 50. The following are the comparison results:

[0170]

[0171] In this embodiment, by simulating a recruitment scenario, the significant advantages of the method of the present invention in processing dynamic modeling of job requirements are demonstrated. The experimental results show that compared with traditional methods, the method of the present invention can more efficiently capture the dynamic changes of job requirement characteristics, improve the accuracy of candidate matching, and significantly shorten the recruitment cycle, verifying the practical feasibility and superiority of the present invention.

[0172] The present invention improves the traditional fuzzy C-means clustering algorithm. By introducing a real-time data dynamic adaptation mechanism and a category membership degree self-adaptive adjustment strategy, the algorithm can effectively handle the ambiguity and uncertainty problems existing in job requirement data. By dynamically adjusting the fuzzy factor and the adaptation parameter, the response ability to dynamically changing job requirement data is enhanced. Combining real-time market feedback data can more accurately extract the key characteristics of the job, significantly improving the model's ability to depict the diversity and fuzzy boundaries of job categories. Compared with traditional static clustering algorithms, the accuracy and generalization ability of job requirement feature extraction are improved.

[0173] The present invention designs a two-layer generative adversarial network structure. The first layer of the generative adversarial network solves the problem of data sparsity of job requirement feature data by combining fuzzy clustering results to generate potential feature vectors. The second layer of the generative adversarial network takes the potential feature vectors and historical feedback data as inputs and generates a real-time optimized job requirement model by dynamically adjusting the set of adaptation parameters, which can adapt to the rapidly changing recruitment environment. By introducing an adaptive constraint weight, the second layer of the generative adversarial network can fully consider the impact of market changes when optimizing the job requirement model, significantly improving the rationality and accuracy of the job requirement features generated by the model.

[0174] The present invention proposes a set of adaptive job requirement parameter systems, including job requirement change rate, job skill deviation degree, and job supply-demand weight, which are used to capture the dynamic changes of the market and the adjustment direction of job requirements in real time. By combining the dynamic optimization mechanism of the generative adversarial network, the efficient integration of real-time market data and the historical feedback of enterprise recruitment is realized in the construction of the job requirement model, thereby dynamically adjusting the model output results, significantly enhancing the adaptability of the job requirement model to market trends.

[0175] As mentioned above, the above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for dynamically constructing a job demand model based on fuzzy clustering, characterized in that: The steps include: S1. Collect comprehensive job demand data sets related to job requirements; S2. Clean the collected comprehensive job demand data set, remove noise data, normalize and standardize the comprehensive job demand data set, and generate a feature vector set of the comprehensive job demand data set through text vectorization technology; S3. Based on the feature vector set of the comprehensive job demand data set, the improved fuzzy C-means clustering algorithm is used to perform cluster analysis, generate a category membership matrix, and extract the key feature vectors of each category according to the category membership matrix; S4. Taking the key feature vector and the category membership matrix as input, design and train the first-layer generative adversarial network, generate new potential feature vectors through the generator of the first-layer generative adversarial network, optimize the generator output through the discriminator of the first-layer generative adversarial network, and the generated potential feature vectors constitute an enhanced data set for further job requirement modeling; S5. Taking the potential feature vector and the historical feedback data of job requirements as input, design and train the second-layer generative adversarial network, generate a dynamically optimized job requirement model through the generator of the second-layer generative adversarial network, evaluate the rationality of the job requirement model through the discriminator of the second-layer generative adversarial network, and use the generated job requirement model as the output result of the current dynamic job requirement model; S6. Collect the company's newly added job demand data, market feedback data and recruitment result data in real time, input the real-time collected data into the modules corresponding to steps S1 to S5, update the category membership matrix of the fuzzy clustering analysis, update the parameters of the first and second layers of the generative adversarial network, and dynamically optimize the job demand model.

2. According to claim 1, a method for dynamically constructing a job demand model based on fuzzy clustering is characterized in that: The step S1 comprises the following sub-steps: S11. Obtain the historical recruitment data set D formed during the recruitment process within the enterprise, including job descriptions, recruitment conditions, interview results, and final recruitment information; S12. Extract job description data set D from the company's existing job descriptions or real-time new job demand information p , including job title, job responsibilities, skill requirements and educational requirements; S13. Obtain market demand data set D from the external recruitment market m , including industry recruitment trends, job posting information and market supply and demand analysis; S14. Combine the historical recruitment data set, job description data set and market demand data set to form a comprehensive job demand data set: D=D h ∪D p ∪D m 。 3. The method for dynamically constructing a job requirement model based on fuzzy clustering according to claim 1 is characterized in that: The step S2 comprises the following sub-steps: S21. Clean each piece of data in the comprehensive job demand data set D, remove redundant data, outliers and missing values, and form a cleaned comprehensive job demand data set; S22. normalize each data record in the cleaned comprehensive job demand data set, and map the numerical attribute of the data to the interval [0,1]; S23. Standardize the normalized comprehensive job demand data set, convert the numerical attributes into standard normal distribution, and the standardized results of all data records constitute the standardized comprehensive job demand data set D s ; S24. Standardized comprehensive job demand data set D s The non-numeric attributes in the text are vectorized, and the word embedding technology is used to convert the text field into a vector form. The text vector set is defined as: F t ={f t1 ,f t2 ,…,f tn }; Among them, f tn The vector representation of the nth text record, where n is the total number of text records; S25. Merge the text vector set with the comprehensive job demand data set to form a feature vector set of the comprehensive job demand data: F=F t ∪D s 。 4. The method for dynamically constructing a job requirement model based on fuzzy clustering according to claim 1 is characterized in that: The step S3 comprises the following sub-steps: S31. Set the clustering parameters of the job demand model, including the number of clustering categories c, the maximum number of iterations T, the fuzzy factor m>1, the dynamic adaptation factor λ, and the membership threshold ∈, where c represents the number of target categories set according to the diversity of job demand characteristics, m controls the degree of fuzziness of job demand characteristics in the clustering process, and λ is the dynamic adaptation factor, which is used to adjust the sensitivity of the clustering results according to the real-time data stream, reflecting the adaptability of the job demand model to real-time changes; S32. For the comprehensive job demand feature vector set F = {f1,f2,…,f n } and the real-time data stream feature vector set F r ={f r1 ,f r2 ,…,f rk } perform joint processing to generate a dynamically initialized membership matrix U (0) : IN (0) =[in ij ]; Among them, u ij Represents comprehensive job demand data f i For the initial membership of the jth class, φ(f i ,f rj ) is a feature similarity function, which is used to dynamically adjust the initialization result in combination with the real-time data stream; S33. Combine the real-time data stream and the initialized membership matrix to dynamically calculate the center vector of each category: in, represents the dynamic category center of the jth category, is the membership value of the membership matrix, f i For comprehensive job demand characteristics, f r is the real-time data stream feature vector, t1 is the current iteration number; S34. Based on dynamic category center Update the membership matrix in, Indicates the job requirement characteristics f i With dynamic category center The Euclidean distance of S35. Repeat steps S33 and S34 until the following dynamic convergence conditions are met: or t1 ≥ T; S36. Based on the final membership matrix U (T) and the category center set V = {v1,v2,…,v c } Extract the key feature vector v of the job demand model c , where v c Represents the dynamic characteristics of each job category.

5. The method for dynamically constructing a job requirement model based on fuzzy clustering according to claim 1 is characterized in that: The step S4 comprises the following sub-steps: S41. Design the first layer of the generative adversarial network, including the first layer generator G1(z,U (T) ) and the first layer discriminator D1(x), where G1(z,U (T) Represents a random noise vector z and a membership matrix U (T) is the first-layer generator of the input, used to generate the potential feature vector of the job requirement model, and D1(x) represents the first-layer discriminator that determines whether the input feature vector is the real job requirement feature vector; S42. Initialize the parameters of the first layer generator and the first layer discriminator, including the weight matrix W g , bias term b g , the weight matrix W of the discriminator d and the bias term b d , and set the learning rate α and optimization target of the generative adversarial network; S43. Input the category center set V and the random noise vector z into the first layer generator to generate a potential feature vector set F g ; S44. The potential feature vector set F generated by the first layer generator g The first layer discriminator D1(x) is input into the first layer discriminator D1(x) at the same time as the real feature vector set F, and the first layer discriminator calculates the degree of distinction between the feature vector generated by the generator and the real feature vector; S45. Define the loss function L of the first-layer generative adversarial network as the alternating adversarial optimization objective of the first-layer generator and the first-layer discriminator: Among them, P data is the distribution of the true eigenvector, P z For the distribution of random noise, the weights and bias items of the generator and the weights and bias items of the first layer of discriminator are updated by optimizing the parameters of the first layer of generator and the first layer of discriminator; S46. The potential feature vector set F output by the optimized first layer generator g As an enhanced job requirement feature dataset.

6. The method for dynamically constructing a job requirement model based on fuzzy clustering according to claim 1, characterized in that: The step S5 comprises the following sub-steps: S51. Construct an adaptively optimized second-layer generative adversarial network model, including the second-layer generator G2 (F g ,H,A) and the second layer discriminator D2(M,A), where H is the historical feedback data set of job requirements, reflecting the historical information of the matching between the company's past recruitment and job requirements, A is the set of job demand adaptability parameters, reflecting the sensitivity and adaptability of the job demand model to market dynamic changes, and M is the generated dynamic optimization job demand model; S52. Initialize the job demand adaptability parameter set A, including: Job demand change ratea rate , indicating how fast job requirements change: Among them, H t and H t-1 They represent the job demand feedback data at the current and previous moments respectively, and t is the time interval; Job Skill Deviationa skill , indicating the degree of deviation between job skill requirements and market skill supply: in, and They represent job-required skills and market-practical skills respectively; Job supply and demand weighta market , which represents the weight ratio between market job demand and candidate supply: Among them, job_openings and applicants represent the number of job postings and the number of job seekers respectively; S53. The potential feature vector set F g , the job demand historical feedback data set H and the job demand adaptability parameter set A are input into the second-layer generator G2 to generate the dynamically optimized job demand model M: M=G2(F g ,H,A); S54. Input the job requirement model M and adaptability parameter A output by the second-layer generator into the second-layer discriminator D2 to discriminate the rationality and dynamic adaptability of the job requirement model and output the discrimination result: D2(M,A)=σ(W d2 ·M+W a ·A+b d2 ); Among them, W d2 and b d2 is the weight and bias of the discriminator, which determines the sensitivity of the discrimination process to the model features. a is the adaptive parameter weight, σ is the activation function; S55. Define the loss function L2 of the second-layer generative adversarial network, and add an adaptive weight constraint based on the generative adversarial loss: Among them, P real represents the distribution of the real job demand model, P g represents the distribution of potential feature vectors, γ is the adaptive weight adjustment factor, and A target is the ideal adaptability parameter value; S56. The job requirement model M output by the optimized second-layer generator is used as the final dynamic optimization result for job requirement matching and recruitment optimization.

Images