Multi-dimensional talent recommendation method based on human resource big data

Abstract

The invention provides a multi-dimensional talent recommendation method based on human resource big data, and the method comprises the steps: obtaining post demand data issued by an enterprise, the post demand data comprising post responsibilities, skill requirements, working experience and educational background requirements of posts required by the enterprise; performing feature extraction on the post demand data, determining post feature data, performing feature mapping on the post feature data to obtain a post demand vector, generating a post portrait of a post required by an enterprise according to the post demand vector, and adding the post portrait into a post library; the post demand vector has the same dimension as the talent evaluation vector of the talent portrait of each job seeker in the talent pool; the talent portraits in the talent pool are matched with the post portraits of the posts required by the enterprise, and the adaptation degree between the talent portraits of the job seekers and the post portraits of the posts required by the enterprise is determined; and generating a talent recommendation list based on the adaptation degree between the talent portrait of the job seeker and the post portrait of the post required by the enterprise, and recommending the job seeker to the enterprise based on the talent recommendation list.

Description

Technical Field

[0001] This application relates to the technical field of human resource management. Specifically, it relates to a multi-dimensional talent recommendation method based on human resource big data. Background Art

[0002] In today's highly competitive market environment, the human resource management of enterprises faces many challenges. The traditional talent recruitment and evaluation methods have many limitations, such as single evaluation dimension, strong subjectivity, low efficiency, etc. During the recruitment process, it is difficult for enterprises to comprehensively and accurately understand the abilities and potentials of job seekers, resulting in a low matching degree between talents and positions, which not only increases the recruitment cost of enterprises, but also affects the operation efficiency and development of enterprises.

[0003] With the rise of big data technology, the human resource field has begun to try to use big data for talent management. However, most of the current solutions stay at the stage of data collection and simple analysis, lacking multi-dimensional in-depth evaluation of talents and precise matching with job requirements. Therefore, how to effectively utilize human resource big data and improve the scientificity and precision of talent management has become an urgent problem to be solved in the field of human resource management. Summary of the Invention

[0004] The purpose of the embodiments of this application is to provide a multi-dimensional talent recommendation method based on human resource big data to achieve precise matching between talents and positions and improve the talent recommendation effect for job requirements.

[0005] To achieve the above purpose, the embodiments of this application are implemented as follows:

[0006] In a first aspect, the embodiments of this application provide a multi-dimensional talent recommendation method based on human resource big data, including: obtaining job requirement data released by an enterprise, where the job requirement data includes job responsibilities, skill requirements, work experience, and educational requirements of the positions required by the enterprise; extracting features from the job requirement data to determine job feature data, and performing feature mapping on the job feature data to obtain a job requirement vector, and accordingly generating a job portrait of the positions required by the enterprise and adding it to the job library, where the dimension of the job requirement vector is the same as that of the talent evaluation vector of each job seeker's talent portrait in the talent library; matching the talent portraits in the talent library with the job portraits of the positions required by the enterprise to determine the matching degree between the talent portraits of the job seekers and the job portraits of the positions required by the enterprise; generating a talent recommendation list based on the matching degree between the talent portraits of the job seekers and the job portraits of the positions required by the enterprise, and recommending job seekers to the enterprise based on the talent recommendation list.

[0007] In combination with the first aspect, in the first possible implementation manner of the first aspect, before obtaining the job requirement data released by an enterprise, the method further includes the process of constructing a talent pool: obtaining a job seeker data set, where the job seeker data set includes multiple groups of job seeker-job data, each group of job seeker-job data includes job seeker data and job requirement data, and the job seeker has worked in the job required by the enterprise corresponding to the job requirement data for a set duration; performing data cleaning on the job seeker data in each group of job seeker-job data, and then performing feature extraction to obtain vectorized and uniformly dimensioned job seeker feature samples, so as to form a job seeker feature sample set; performing clustering analysis on all the job seeker feature samples in the job seeker feature sample set, determining M clustering centers, and calculating the distances between each job seeker feature sample and the M clustering centers to obtain an M-dimensional job seeker feature vector corresponding to the job seeker feature sample, which is used as the talent evaluation vector of the corresponding job seeker; generating a talent profile of the job seeker based on the talent evaluation vector of the job seeker and adding it to the talent pool to complete the construction of the talent pool.

[0008] In combination with the first possible implementation manner of the first aspect, in the second possible implementation manner of the first aspect, each job seeker data includes the basic information, work experience, skills and specialties, project achievements, personal evaluation, and professional quality assessment of the job seeker. Performing data cleaning on the job seeker data in each group of job seeker-job data, and then performing feature extraction to obtain vectorized and uniformly dimensioned job seeker feature samples, so as to form a job seeker feature sample set, including: for the job seeker data in each group of job seeker-job data: performing data cleaning on the job seeker data; performing key feature extraction on the basic information in the job seeker data after data cleaning to obtain a feature component of the first length; performing text feature vector extraction on the work experience, skills and specialties, project achievements, personal evaluation, and professional quality assessment in the job seeker data after data cleaning to obtain feature components of the second length, third length, fourth length, fifth length, and sixth length respectively; concatenating each feature component to obtain a uniformly dimensioned job seeker feature sample; constructing a job seeker feature sample set based on the job seeker feature samples corresponding to the job seeker data in each group of job seeker-job data.

[0009] In combination with the first possible implementation manner of the first aspect, in the third possible implementation manner of the first aspect, performing clustering analysis on all the job seeker feature samples in the job seeker feature sample set, determining M clustering centers, and calculating the distances between each job seeker feature sample and the M clustering centers to obtain an M-dimensional job seeker feature vector corresponding to the job seeker feature sample, which is used as the talent evaluation vector of the corresponding job seeker, including: for the job seeker feature sample set P = {p 1 , p 2 ... p N}, calculate the similarity between every two job seeker feature samples, and determine the similarity matrix S, where N is the total number of job seeker feature samples; construct the attraction matrix R t and the membership matrix A t ; perform iteration on the attraction matrix R t and the membership matrix A t until the clustering is completed when the termination condition is met. Determine M clustering centers. The termination condition is: the clustering center after this iteration is the same as the clustering center after the previous iteration, or the set number of iterations is reached; calculate the distance between each job seeker feature sample and the M clustering centers to obtain the M-dimensional job seeker feature vector corresponding to the job seeker feature sample, which is used as the talent evaluation vector of the corresponding job seeker.

[0010] Combined with the first possible implementation manner of the first aspect, in the fourth possible implementation manner of the first aspect, after the construction of the talent pool is completed, the method further includes the construction process of the job pool: perform data cleaning on the job requirement data in each group of job seeker-job data, and then perform feature extraction to determine the job feature data that is vectorized and has a unified dimension, which is used as the job requirement feature sample to form a job requirement feature sample set; train a feature mapping model based on the job requirement feature sample set. The feature mapping model uses the LDA model, sets M output categories, and makes the dimension of the model output data be M; based on the trained feature mapping model, determine the job requirement vector corresponding to each job requirement feature sample, where the dimension of the job requirement vector is the same as the dimension of the talent evaluation vector, both being M; based on the job requirement vector, generate the job portrait of the enterprise's required jobs and add it to the job pool to complete the construction of the job pool.

[0011] Combined with the fourth possible implementation manner of the first aspect, in the fifth possible implementation manner of the first aspect, perform data cleaning on the job requirement data in each group of job seeker-job data, and then perform feature extraction to determine the job feature data that is vectorized and has a unified dimension, which is used as the job requirement feature sample to form a job requirement feature sample set, including: for the job requirement data in each group of job seeker-job data: perform data cleaning on the job requirement data; extract text feature vectors from the job responsibilities and skill requirements in the job requirement data after data cleaning to obtain feature components with the seventh length and the eighth length respectively, and extract features from the work experience and educational requirements to obtain feature components with the ninth length and the tenth length; splice each feature component to obtain a job requirement feature sample with a unified dimension; construct a job requirement feature sample set based on the job requirement feature samples corresponding to the job requirement data in each group of job seeker-job data.

[0012] Combined with the fourth possible implementation manner of the first aspect, in the sixth possible implementation manner of the first aspect, after the construction of the position library is completed, the method further includes: determining a talent weight vector based on the talent evaluation vectors of all job seekers in the talent library and M clustering centers; determining a position weight vector based on the M output categories of the feature mapping model; constructing an objective function for calculating the fitness; constructing a training data set based on the job seeker feature sample set and the position requirement feature sample set; training and optimizing the objective function based on the training data set to obtain an optimized fitness function.

[0013] Combined with the sixth possible implementation manner of the first aspect, in the seventh possible implementation manner of the first aspect, determining a talent weight vector based on the talent evaluation vectors of all job seekers in the talent library and M clustering centers includes: for the k-th clustering center among the M clustering centers: determining the job seeker feature samples in the job seeker feature sample set that belong to the k-th clustering center, and determining the maximum intra-cluster distance from the k-th clustering center The minimum intra-cluster distance and the average intra-cluster distance Calculate the distance weight reference value using the following formula

[0014]

[0015] Calculate the sample weight reference value using the following formula

[0016]

[0017] where p k is the number of job seeker feature samples in the job seeker feature sample set that belong to the k-th clustering center; based on the distance weight reference value and the sample weight reference value Calculate the talent weight parameter using the following formula:

[0018]

[0019] where is the talent weight parameter of the k-th clustering center; form a talent weight vector based on the talent weight parameters of the M clustering centers

[0020] Combined with the sixth possible implementation manner of the first aspect, in the eighth possible implementation manner of the first aspect, determining a position weight vector based on the M output categories of the feature mapping model includes: for the k-th output category of the feature mapping model: obtaining the intra-cluster similarity and the inter-cluster similarity The job weight parameter for the k-th output category is calculated using the following formula:

[0021]

[0022] where, is the job weight parameter for the k-th output category; based on the job weight parameters of M output categories, a job weight vector is formed

[0023] Combined with the sixth possible implementation manner of the first aspect, in the ninth possible implementation manner of the first aspect, the optimized fitness function is:

[0024]

[0025] where, P(AB) represents the fitness between the talent evaluation vector A and the job requirement vector B. The talent evaluation vector A = [A 1 , …, A k , …, A M , the job requirement vector B = [B 1 , …, B k , …, B M , M is the dimension number of the talent evaluation vector A and the job requirement vector B, δ is the adaptive weight vector, δ = [δ 1 , …, δ k , …, δ M , r A is the talent weight vector, r B is the job weight vector,

[0026] Beneficial effects: 1. This solution realizes the multi-dimensional evaluation of talents and jobs by constructing the talent portrait of job seekers and the job portrait of the positions required by enterprises. The talent portrait covers multiple aspects such as the basic information, work experience, skills and specialties, project achievements, personal evaluations, and professional quality assessments of job seekers, while the job portrait is based on job requirement data such as job responsibilities, skill requirements, work experience, and educational requirements. This multi-dimensional evaluation method can more comprehensively and accurately reflect the characteristics of job seekers and positions, thereby improving the matching degree between talents and positions. In the matching stage, efficient talent recommendation can be dynamically carried out according to the job requirements released by enterprises. Of course, based on this solution, the job recommendation port for talents can also be opened (although the job recommendation port for job seekers is currently open, based on the platform built by this system, it can be very easily implemented to provide a port for job seekers to recommend suitable jobs).

[0027] 2. In the stages of talent pool construction, job pool construction, and matching training (optimizing the objective function for fitness), the advantages of big data technology are fully utilized. Through means such as data cleaning, feature extraction, and clustering analysis, valuable information is mined from a large amount of job seeker - job data, providing a scientific basis for talent recommendation. In addition, through the training and optimization of the feature mapping model, the dimensional unity of the talent evaluation vector and the job requirement vector is achieved, laying a foundation for fitness calculation. Not only are talent pools and job pools constructed, but also a dynamic update mechanism for these pools is provided. With the addition of new job seeker data and job requirement data, the talent pools and job pools will continue to be enriched and improved. At the same time, through the training and optimization of the objective function, the fitness function can continuously adapt to the new data distribution, improving the accuracy and efficiency of talent recommendation. Moreover, in fitness calculation, an adaptive weight vector is introduced, enabling the importance of different dimensions in the matching process to be dynamically adjusted. This design not only considers the contribution degree differences of different dimensions to the matching result but also improves the flexibility and accuracy of the matching.

[0028] 3. The clustering algorithm is used to cluster the job seeker feature samples to form M clustering centers. Taking the distances between the job seeker feature samples and the M clustering centers as the talent evaluation vectors, potential features can be mined for matching. Additionally, the distances between the job seeker feature samples and the clustering centers (including the maximum distance, minimum distance, and average distance) are considered. These distance information can reflect the similarity between the job seeker feature samples and the clustering centers, and a distance weight reference value is introduced. Thus, it can guide the more accurate assignment of weights. At the same time, the probabilities of the job seeker feature samples being assigned to each clustering center are considered, and a sample weight reference value is introduced. A more reasonable and effective talent weight vector is comprehensively designed as the basis for matching, which can provide a reliable foundation for subsequent matching. The feature mapping model is used to map the job requirement vector into the space with the same dimension as the talent evaluation vector, making it possible to calculate the fitness between talent and job. The feature mapping model not only considers the inherent characteristics of job requirements but also reflects the similarities and differences between jobs through the output categories (i.e., clustering results). By using the within - cluster similarity and between - cluster similarity of the output categories, a job weight vector is designed, which can better measure the reflection ability of the job feature vector in the matching process. Also, the solution introduces an adaptive weight vector to further adjust the weights of each dimension in the talent evaluation vector and the job requirement vector, enhancing the flexibility and adaptability of the solution and enabling it to better cope with the complex and changeable recruitment environment.

[0029] To make the above - mentioned objects, features, and advantages of the present application more obvious and understandable, the following preferred embodiments are specifically given and described in detail in conjunction with the accompanying drawings as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the accompanying drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application, and thus should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0031] Figure 1 It is a flowchart for constructing a talent pool in the embodiments of the present application.

[0032] Figure 2 It is a flowchart for constructing a job pool in the embodiments of the present application.

[0033] Figure 3 It is a flowchart for matching by designing a fitness function in the embodiments of the present application.

[0034] Figure 4 It is a flowchart for a multi-dimensional talent recommendation method based on human resource big data provided in the embodiments of the present application. Detailed implementation manners

[0035] The following will describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application.

[0036] To facilitate the understanding of this solution, the preprocessing process for running the multi-dimensional talent recommendation method based on human resource big data (including the processes of constructing a talent pool, a job pool, and optimizing the fitness function) will be introduced here.

[0037] First, please refer to Figure 1 , Figure 1 It is a flowchart for constructing a talent pool in the embodiments of the present application.

[0038] In this embodiment, constructing a talent pool may include step S11, step S12, step S13, and step S14.

[0039] First, step S11 needs to be run.

[0040] Step S11: Obtain a job seeker dataset, where the job seeker dataset contains multiple groups of job seeker-job data. Each group of job seeker-job data contains job seeker data and job requirement data, and the job seeker has worked in the job required by the enterprise corresponding to the job requirement data for a set duration.

[0041] In this embodiment, in order to improve the quality of the talent pool, the quality of the job pool, and the effectiveness of matching, considering that it is not easy to match job seeker data and job data in conventional data, therefore, in this embodiment, during the data acquisition stage, data of job seekers who have worked in the jobs required by the enterprise corresponding to the job demand data for a set duration (such as 6 months, 1 year, 2 years, etc.) is screened to obtain their job seeker data and job demand data, forming a set of job seeker - job data, and based on this, a job seeker data set containing multiple sets of job seeker - job data is obtained. This method will filter out most of the data. Therefore, the data set screened from the database itself is not very large, but over time, there will be more and more available data later, and the scale of the job seeker data set will become larger and larger.

[0042] After obtaining the job seeker data set, step S12 can be executed.

[0043] Step S12: Clean the job seeker data in each set of job seeker - job data, and then perform feature extraction to obtain a vectorized and uniformly - dimensioned job seeker feature sample to form a job seeker feature sample set.

[0044] In this embodiment, each job seeker data includes the basic information of the job seeker (such as name, gender, age, education level, contact information, etc. Among them, job seeker privacy data such as name and contact information will be processed in this solution. On the one hand, irrelevant information is removed, and on the other hand, user privacy is protected to prevent accidental leakage. However, after being recommended to the enterprise and with the consent of the user, the complete information can be sent to the enterprise), work experience (such as work unit, position, working years, etc.), skills and specialties (such as professional skills, language ability, computer skills, etc.), project achievements (wide range and a large number of various projects, such as XX software development project, XX platform construction project, XX power system project, etc. in a broad sense, and detailed ones such as XXX function design based on XXX technology in XXX scenario), personal evaluation (such as colleague evaluation, superior evaluation, customer evaluation, etc.), and professional quality assessment (providing multi - dimensional assessment questionnaires: Enneagram test, MBTI career personality test, DISS personality test, SCL90 assessment, Holland vocational interest test, career anchor assessment, wvi assessment, etc. Generally, it is required to fill in one questionnaire, such as the MBTI career personality test).

[0045] Then, for the job seeker data in each set of job seeker - job data: First, the job seeker data can be cleaned (such as cleaning, de - duplication, standardization, etc. processing to improve data quality and prepare for subsequent analysis. The cleaning process includes operations such as identifying and correcting errors in the data and filling in missing values).

[0046] After that, key features can be extracted from the basic information in the job seeker data after data cleaning to obtain feature components of the first length. Since the basic information is similar to semi-structured data, key features can be conveniently extracted through key feature extraction. For example, these features can be extracted by one-hot encoding to form feature components of the first length.

[0047] Then, text feature vectors need to be extracted from the work experience, skills and specialties, project achievements, personal evaluations, professional quality assessments, etc. in the job seeker data after data cleaning, respectively obtaining feature components of the second length, third length, fourth length, fifth length, and sixth length.

[0048] In this embodiment, for the text feature vector extraction of information such as work experience, skills and specialties, project achievements, personal evaluations, professional quality assessments, etc., methods such as the bag-of-words model, TF-IDF (an improvement of the bag-of-words model), word embeddings (such as Word2Vec, GloVe, BERT, etc.), sentence embeddings, LDA (a topic model for text analysis), etc. can be used. In this embodiment, the Sentence-BERT sentence embedding model is selected, which can convert the entire sentence or document into a vector representation of a fixed dimension, facilitating unified dimensions. Sentence-BERT is based on the pre-trained Transformer architecture (BERT) and is fine-tuned to generate high-quality sentence embeddings, capable of capturing the overall semantic information of sentences or documents and having consistent output feature dimensions. Accordingly, feature components of the second length, third length, fourth length, fifth length, and sixth length can be obtained respectively.

[0049] It should be noted that for information such as work experience, skills and specialties, project achievements, personal evaluations, professional quality assessments, etc., they are relatively independent, that is, each type of information needs to be processed separately once, rather than being integrated and input uniformly. To improve efficiency, 5 Sentence-BERTs can be arranged to run in parallel, which can achieve different output lengths, but the deployment cost is relatively high. It is also possible to use one Sentence-BERT, which needs to mark multiple modes (such as 5 modes, respectively used to extract the feature components of work experience, skills and specialties, project achievements, personal evaluations, and professional quality assessments) to achieve different lengths, but the processing efficiency is relatively low. Considering the cost, and the construction of the talent pool can be regarded as a one-time task, this mode is adopted in this embodiment to save costs. And Sentence-BERT is not the improvement point of the present invention and can be implemented using existing models. Open-source code can be found in the forum and will not be elaborated here.

[0050] After obtaining the characteristic components of the first length, second length, third length, fourth length, fifth length, and sixth length, splicing can be performed. For example, direct splicing can be used to obtain a job seeker characteristic sample with a unified dimension (the length is the sum of the first length, second length, third length, fourth length, fifth length, and sixth length).

[0051] Accordingly, the job seeker data in each group of job seeker-position data in the job seeker dataset is processed to obtain a job seeker characteristic sample corresponding to each job seeker data, forming a job seeker characteristic sample set.

[0052] After obtaining the job seeker characteristic sample set, step S13 can be executed.

[0053] Step S13: Perform clustering analysis on all job seeker characteristic samples in the job seeker characteristic sample set to determine M clustering centers, and calculate the distances between each job seeker characteristic sample and the M clustering centers to obtain an M-dimensional job seeker characteristic vector corresponding to the job seeker characteristic sample, which is used as the talent evaluation vector of the corresponding job seeker.

[0054] In this embodiment, the affinity propagation algorithm is used for clustering analysis. For the job seeker characteristic sample set P = {p 1 , p 2 ... p N}, the similarity between every two job seeker characteristic samples can be calculated to determine a similarity matrix S, where N is the total number of job seeker characteristic samples. Then, an attractiveness matrix R t and a membership matrix A t are constructed. The attractiveness matrix R t and the membership matrix A t are iterated. Each iteration will update the clustering centers until the clustering is completed when the termination condition is met, and M clustering centers are determined. The termination condition is: the clustering centers after this iteration are the same as those after the previous iteration, or the set number of iterations is reached. The affinity propagation algorithm is already a mature clustering analysis technology and will not be elaborated here.

[0055] After the clustering analysis is completed, step S14 can be executed.

[0056] Step S14: Generate a talent portrait of the job seeker based on the talent evaluation vector of the job seeker and add it to the talent pool to complete the construction of the talent pool.

[0057] In this embodiment, the distances between each job seeker feature sample and the M clustering centers can be calculated to obtain an M-dimensional job seeker feature vector corresponding to the job seeker feature sample, which serves as the talent evaluation vector for the corresponding job seeker. Then, based on the talent evaluation vector of the job seeker, a talent portrait of the job seeker (associating the job seeker data of the job seeker with its corresponding talent evaluation vector as the talent portrait) is generated and added to the talent pool. After processing each job seeker's data, the construction of the talent pool can be completed.

[0058] After the construction of the talent pool is completed, it is necessary to further construct a job position pool. Please refer to Figure 2 , and the construction process of the job position pool may include step S21, step S22, step S23, and step S24.

[0059] Based on the obtained job seeker data set, step S21 is run.

[0060] Step S21: Clean the job requirement data in each group of job seeker-job position data, and then perform feature extraction to determine the vectorized and uniformly dimensioned job feature data as the job requirement feature sample to form a job requirement feature sample set.

[0061] In this embodiment, the job requirement data in each group of job seeker-job position data can be cleaned (such as cleaning, deduplication, standardization, etc. to improve data quality and prepare for subsequent analysis, where the cleaning process includes operations such as identifying and correcting errors in the data and filling missing values), and then feature extraction is performed. Considering that the job requirement data includes information such as job responsibilities, skill requirements, work experience, and educational requirements of the positions required by the enterprise, in this embodiment, a text feature extraction scheme can be used to perform text feature extraction on job responsibilities and skill requirements to obtain feature components of the seventh length and the eighth length. Similar to the text feature extraction scheme described above, Sentence-BERT is also used to implement it, and different length feature components can also be obtained by designing tags to distinguish text feature extraction for different information, which will not be elaborated here.

[0062] For the work experience and educational requirements in the job requirement data after data cleaning, the extraction of feature components can be achieved through simple feature extraction schemes such as one-hot encoding to obtain feature components of the ninth length and the tenth length.

[0063] After obtaining the characteristic components of the seventh length, eighth length, ninth length, and tenth length, the characteristic components can be concatenated. For example, concatenating them in sequence to obtain a job demand characteristic sample with a unified dimension (the length is the sum of the seventh length, eighth length, ninth length, and tenth length). After processing the job demand data in each group of job seeker - job data in the job seeker dataset, each job demand data has a corresponding job demand characteristic sample, and based on this, a job demand characteristic sample set can be formed.

[0064] To improve the matching accuracy, more - dimensional data can also be considered to be introduced into the job demand data, such as quality requirements (for example, reflecting the requirement for the stress - resistance ability of job seekers), business - trip competence (whether they can complete business - trip tasks), etc. Feature extraction is performed through corresponding text feature extraction methods, and then concatenated to construct a job demand characteristic sample, which is not limited here.

[0065] After that, step S22 can be executed.

[0066] Step S22: Train a feature mapping model based on the job demand characteristic sample set. Among them, the feature mapping model adopts the LDA model, sets M output categories, and makes the dimension of the model output data be M.

[0067] In this embodiment, a feature mapping model can be trained based on the job demand characteristic sample set. The feature mapping model adopts the LDA (Linear Discriminant Analysis) model, sets M output categories, and makes the dimension of the model output data be M. Since LDA requires supervised training, the job demand characteristic samples need to use the job name as a label for training the feature mapping model. In addition, since this solution does not improve LDA itself, it will not be elaborated here and only a simple description is given: Because LDA is a supervised linear dimensionality - reduction algorithm, its core idea is to project high - dimensional data into a low - dimensional space through linear transformation while maintaining the class information between data points, so that the data points after dimensionality reduction are as easy to distinguish as possible. Therefore, LDA will find a projection direction or projection matrix such that the projection points of the same - class samples are as close as possible, and the projection points of different - class samples are as far apart as possible. Two concepts are involved, within - cluster similarity and between - cluster similarity. The within - cluster similarity reflects the similarity between samples within the same cluster, while the between - cluster similarity reflects the similarity between different clusters; or it can be measured by within - cluster divergence (for example, the opposite of within - cluster similarity) and between - cluster divergence (for example, the opposite of between - cluster similarity). In this embodiment, within - cluster similarity and between - cluster similarity are taken as examples.

[0068] After completing the training of the feature mapping model, step S23 can be executed.

[0069] Step S23: Based on the trained feature mapping model, determine the job requirement vectors corresponding to each job requirement feature sample, where the dimension of the job requirement vector is the same as that of the talent evaluation vector, both being M.

[0070] In this embodiment, each job requirement feature sample can be input into the trained feature mapping model, and the feature mapping is performed by the feature mapping model to obtain the job requirement vectors (with a dimension of M) corresponding to each job requirement feature sample, which is the same as the dimension of the talent evaluation vector, ensuring that the job requirement vectors and the talent evaluation vectors can be used for similarity calculation.

[0071] After that, step S24 can be executed.

[0072] Step S24: Based on the job requirement vectors, generate the job portraits of the jobs required by the enterprise and add them to the job library to complete the construction of the job library.

[0073] In this embodiment, based on the job requirement vectors, the job portraits of the jobs required by the enterprise can be generated (for example, the job requirement data is associated with its corresponding job requirement vector), and then added to the job library to complete the construction of the job library.

[0074] After the construction of the talent library and the job library is completed, the design of the matching module is required, that is, to determine a suitable fitness calculation scheme to achieve the matching of job seekers and jobs.

[0075] Please refer to Figure 3 , the process of designing a fitness function for job seeker and job matching can include step S31, step S32, step S33, step S34 and step S35.

[0076] First, step S31 can be executed.

[0077] Step S31: Based on the talent evaluation vectors of all job seekers in the talent library and M clustering centers, determine the talent weight vectors.

[0078] In this embodiment, for the k-th clustering center among the M clustering centers:

[0079] First, determine the job seeker feature samples in the job seeker feature sample set that belong to the k-th clustering center, and determine the maximum intra-cluster distance the minimum intra-cluster distance and the average intra-cluster distance

[0080] Then, use the following formula to calculate the distance weight reference value

[0081]

[0082] Calculate the reference value of the sample weight using the following formula

[0083]

[0084] where p k is the number of job seeker feature samples belonging to the k-th cluster center in the job seeker feature sample set.

[0085] Then, based on the reference value of the distance weight and the reference value of the sample weight calculate the talent weight parameter using the following formula:

[0086]

[0087] where is the talent weight parameter of the k-th cluster center.

[0088] After that, based on the talent weight parameters of the M cluster centers, a talent weight vector can be formed:

[0089]

[0090] Of course, the talent weight vector r A is also M-dimensional.

[0091] After determining the talent weight vector, step S32 can be executed.

[0092] Step S32: Determine the job weight vector based on the M output categories of the feature mapping model.

[0093] For the k-th output category of the feature mapping model:

[0094] The within-cluster similarity and the between-cluster similarity of the k-th output category can be obtained. Then, calculate the job weight parameter of the k-th output category using the following formula:

[0095]

[0096] where is the job weight parameter of the k-th output category.

[0097] After that, based on the job weight parameters of the M output categories, a job weight vector can be formed:

[0098]

[0099] Similarly, the job weight vector r B is also M-dimensional.

[0100] Determine the talent weight vector r A and the position weight vector r B After that, it is necessary to construct the objective function, at which point step S33 can be executed.

[0101] Step S33: construct an objective function for calculating the degree of fitness.

[0102] In this embodiment, the objective function is designed using the idea of ​​cosine similarity:

[0103]

[0104] Where P(AB) is the objective function, which represents the degree of fit between the talent assessment vector A and the job demand vector B, and is used to calculate the degree of fit between the talent assessment vector and the job demand vector; Talent assessment vector A = [A 1 ,…,A k ,…,A M ], job demand vector B = [B 1 ,…,B k ,…,B M ], M is the number of dimensions of the talent evaluation vector A and the job demand vector B; δ is the adaptive weight vector, δ = [δ 1 ,…,δ k ,…,δ M ], and the Adam optimization algorithm is used for optimization; ‖A‖ is the modulus of the talent evaluation vector A, and ‖B‖ is the modulus of the job demand vector B.

[0105] After the objective function is determined, step S34 may be executed.

[0106] Step S34: construct a training data set based on the job seeker feature sample set and the job requirement feature sample set.

[0107] In this embodiment, to optimize the adaptive weight vector, it is necessary to combine the job seeker feature sample set and the job requirement feature sample set. Since this solution screens groups of job seeker-job data when obtaining the job seeker data set, there is a strict one-to-one correspondence, which means that the job seeker feature samples in the job seeker feature sample set and the job requirement feature samples in the job requirement feature sample set have a one-to-one correspondence. Here, based on these one-to-one correspondences, each job seeker feature sample and the corresponding job requirement feature sample can be combined to form N groups of training data, and a training data set containing N groups of training data can be constructed accordingly.

[0108] Afterwards, step S35 can be executed.

[0109] Step S35: Perform training optimization on the objective function based on the training data set to obtain an optimized fitness function.

[0110] In this embodiment, N sets of training data in the training dataset can be used, and the Adam optimization algorithm can be used for training optimization, which is achieved by minimizing the loss function Loss. The loss function Loss can be 1 - P(AB), thereby realizing the optimization of the adaptive weight vector δ, and finally obtaining the optimized fitness function:

[0111]

[0112] The fitness function is identical in form to formula (7), except that the adaptive weight vector δ is optimized by the Adam optimization algorithm.

[0113] After completing these pre - processing steps, the multi - dimensional talent recommendation method based on human resource big data can be run. Please refer to Figure 4 , the multi - dimensional talent recommendation method based on human resource big data may include steps S41, S42, S43, and S44.

[0114] First, step S41 can be run.

[0115] Step S41: Obtain the job requirement data released by the enterprise. Among them, the job requirement data includes the job responsibilities, skill requirements, work experience, and educational requirements of the positions required by the enterprise.

[0116] In this embodiment, when the enterprise has job requirements, it can release the job requirement data of the positions required by the enterprise on the system, which needs to include the job responsibilities, skill requirements, work experience, and educational requirements of the positions required by the enterprise.

[0117] After that, step S42 can be run.

[0118] Step S42: Extract features from the job requirement data to determine the job feature data, and perform feature mapping on the job feature data to obtain a job requirement vector. Accordingly, generate a job portrait of the position required by the enterprise and add it to the job library. Among them, the dimension of the job requirement vector is the same as that of the talent evaluation vector of the talent portrait of each job seeker in the talent library.

[0119] In this embodiment, data cleaning and feature extraction can be performed on the job requirement data (refer to the previous text, which will not be elaborated here) to obtain the job feature data. Then, use the feature mapping model to perform feature mapping on the job feature data to obtain an M - dimensional job requirement vector. Accordingly, a job portrait of the position required by the enterprise can be generated and added to the job library (similar to the dynamic update of the job library), and the dimension of the job requirement vector is the same as that of the talent evaluation vector of the talent portrait of each job seeker in the talent library, both being M - dimensional.

[0120] At this time, step S43 can be run.

[0121] Step S43: Match the talent portraits in the talent pool with the job portraits of the positions required by the enterprise to determine the fitness between the talent portraits of the job seekers and the job portraits of the positions required by the enterprise.

[0122] In this embodiment, the talent portraits of all job seekers in the talent pool can be matched with the job portraits of the positions required by the enterprise, and the fitness between the talent portraits of the job seekers and the job portraits of the positions required by the enterprise can be calculated by using formula (8) to calculate the fitness between the talent evaluation vector associated with the talent portrait of each job seeker and the job requirement vector associated with this job portrait.

[0123] After that, step S44 can be executed.

[0124] Step S44: Generate a talent recommendation list based on the fitness between the talent portraits of the job seekers and the job portraits of the positions required by the enterprise, and recommend job seekers to the enterprise based on the talent recommendation list.

[0125] In this embodiment, a talent recommendation list (hiding the privacy information of the job seekers) can be generated based on the fitness between the talent portraits of the job seekers and the job portraits of the positions required by the enterprise, but other information of the job seekers is displayed, such as other basic information, work experience, skills and specialties, project achievements, personal evaluations, professional quality assessments, etc. If an enterprise is interested in a certain job seeker in the talent recommendation list, a request can be generated. After the job seeker agrees, all the information of the job seeker (including name, contact information, etc.) can be displayed to achieve talent recommendation.

[0126] Of course, although the job recommendation for job seekers is not currently open in this embodiment, it is very easy to implement based on the system built in this embodiment. The job recommendation processing flow is only slightly different (the data obtained needs to be job seeker data, and then relevant processing of the job seeker data is performed to form a talent evaluation vector. Then, the fitness between the talent evaluation vector and all the job requirement vectors in the job pool is calculated, and a job recommendation list is formed based on the fitness. However, the relevant information of the enterprise also needs to be hidden, and after obtaining the consent of the enterprise, the complete enterprise information including the job requirement data, contact information, etc. is displayed to the job seeker to achieve job recommendation for job seekers). Therefore, this should not be regarded as a limitation to this application.

[0127] In summary, the embodiment of the present application provides a multi-dimensional talent recommendation method based on human resource big data. By constructing a talent profile of job seekers and a job profile of the positions required by enterprises, a multi-dimensional evaluation of talents and positions is achieved. The talent profile covers multiple aspects such as the basic information, work experience, skills and specialties, project achievements, personal evaluations, and professional quality assessments of job seekers, while the job profile is based on job requirement data such as job responsibilities, skill requirements, work experience, and educational requirements. This multi-dimensional evaluation method can more comprehensively and accurately reflect the characteristics of job seekers and positions, thereby improving the matching degree between talents and positions. In the matching stage, efficient talent recommendations can be made dynamically according to the job requirements released by enterprises. Of course, based on this solution, the job recommendation port for talents can also be opened (although the job recommendation port for job seekers is currently open, based on the platform built on this system, it is very easy to implement a port for job seekers to recommend suitable positions).

[0128] In the stages of talent pool construction, job pool construction, and matching training (optimizing the objective function for adaptability), the advantages of big data technology are fully utilized. Through means such as data cleaning, feature extraction, and clustering analysis, valuable information is mined from a large amount of job seeker-position data, providing a scientific basis for talent recommendation. In addition, through the training and optimization of the feature mapping model, the dimensional unity of the talent evaluation vector and the job requirement vector is achieved, laying a foundation for the calculation of adaptability. Not only are talent pools and job pools constructed, but also a dynamic update mechanism for these pools is provided. With the addition of new job seeker data and job requirement data, the talent pools and job pools will continue to be enriched and improved. At the same time, through the training and optimization of the objective function, the adaptability function can continuously adapt to the new data distribution, improving the accuracy and efficiency of talent recommendation. Moreover, in the calculation of adaptability, an adaptive weight vector is introduced, enabling the importance of different dimensions to be dynamically adjusted during the matching process. This design not only considers the contribution degree differences of different dimensions to the matching result but also improves the flexibility and accuracy of the matching.

[0129] Using the clustering algorithm to cluster the job seeker feature samples to form M clustering centers, and taking the distances between the job seeker feature samples and the M clustering centers as the talent evaluation vectors, potential features can be mined for matching. In addition, the distances between the job seeker feature samples and the clustering centers (including the maximum distance, minimum distance, and average distance) are considered. These distance information can reflect the similarity between the job seeker feature samples and the clustering centers, and a distance weight reference value is introduced to guide a more accurate weight allocation; at the same time, the probabilities of the job seeker feature samples being assigned to each clustering center are considered, and a sample weight reference value is introduced Comprehensively design a more reasonable and effective talent weight vector as the basis for matching, which can provide a reliable foundation for subsequent matching. Use the feature mapping model to map the job requirement vector into the space with the same dimension as the talent evaluation vector, making it possible to calculate the fitness between talent and job. The feature mapping model not only considers the inherent characteristics of job requirements, but also reflects the similarities and differences between jobs through the output categories (i.e., clustering results). By using the within-cluster similarity and between-cluster similarity of the output categories, design the job weight vector, which can better measure the reflection ability of the job feature vector in the matching process. In addition, the solution introduces an adaptive weight vector to further adjust the weights of each dimension in the talent evaluation vector and the job requirement vector, enhancing the flexibility and adaptability of the solution, enabling it to better cope with the complex and changeable recruitment environment.

[0130] In this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0131] The above are only examples of the present application and are not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-dimensional talent recommendation method based on human resources big data, characterized in that: include: Obtaining job demand data published by enterprises, where job demand data includes job responsibilities, skill requirements, work experience, and educational requirements of the positions required by the enterprises; Extract features from job demand data, determine job feature data, and perform feature mapping on the job feature data to obtain a job demand vector. Based on this, a job profile of the job required by the enterprise is generated and added to the job database. The job demand vector has the same dimension as the talent assessment vector of the talent profile of each job seeker in the talent database. Match the talent profiles in the talent pool with the job profiles of the positions required by the company to determine the degree of compatibility between the talent profiles of job seekers and the job profiles of the positions required by the company; Based on the degree of compatibility between the talent profile of job seekers and the job profile of the positions required by the enterprise, a talent recommendation list is generated, and job seekers are recommended to the enterprise based on the talent recommendation list.

2. The multi-dimensional talent recommendation method based on human resources big data according to claim 1 is characterized in that: Before obtaining the job demand data released by the enterprise, the method also includes a talent pool construction process: Obtain a job seeker data set, wherein the job seeker data set includes multiple sets of job seeker-position data, each set of job seeker-position data includes job seeker data and position requirement data, and the job seeker has worked in the position required by the enterprise corresponding to the position requirement data for a set period of time; Performing data cleaning on the job seeker data in each set of job seeker-position data, and then performing feature extraction to obtain a vectorized and uniformly dimensioned job seeker feature sample to form a job seeker feature sample set; Perform cluster analysis on all the job seeker feature samples in the job seeker feature sample set, determine M cluster centers, and calculate the distance between each job seeker feature sample and the M cluster centers to obtain the M-dimensional job seeker feature vector corresponding to the job seeker feature sample as the talent evaluation vector of the corresponding job seeker; Based on the talent assessment vector of the job applicant, a talent profile of the job applicant is generated and added to the talent pool to complete the construction of the talent pool.

3. The multi-dimensional talent recommendation method based on human resources big data according to claim 2 is characterized in that: Each job seeker data includes the basic information, career experience, skills, project achievements, personal evaluation, and professional quality assessment of the job seeker. The job seeker data in each set of job seeker-position data is cleaned and then feature extracted to obtain a vectorized and unified dimension job seeker feature sample to form a job seeker feature sample set, including: For each set of job seeker-position data: Perform data cleaning on job applicant data; Extract key features from basic information in the job seeker data after data cleaning to obtain feature components of a first length; Extract text feature vectors from the job seeker data of professional experience, skills, project achievements, personal evaluation, and professional quality assessment after data cleaning, and obtain feature components of the second length, third length, fourth length, fifth length, and sixth length respectively; Combine the various feature components to obtain a sample of job seeker features of a unified dimension; Based on the job seeker feature samples corresponding to the job seeker data in each set of job seeker-position data, a job seeker feature sample set is constructed.

4. The multi-dimensional talent recommendation method based on human resources big data according to claim 2 is characterized in that: Perform cluster analysis on all the job seeker feature samples in the job seeker feature sample set, determine M cluster centers, and calculate the distance between each job seeker feature sample and the M cluster centers to obtain the M-dimensional job seeker feature vector corresponding to the job seeker feature sample as the talent evaluation vector of the corresponding job seeker, including: For the sample set of job seeker features P = {p1, p2...p N }, calculate the similarity between every two job seeker feature samples and determine the similarity matrix S, where N is the total number of job seeker feature samples; Construct attraction matrix R t And the membership matrix A t ; The attraction matrix R t And the membership matrix A t Iterate, and update the cluster center each time until clustering is completed when the termination condition is met, and M cluster centers are determined. The termination condition is: the cluster center after this iteration is the same as the cluster center after the previous iteration, or the set number of iterations is reached; The distance between each job applicant feature sample and the M cluster centers is calculated to obtain the M-dimensional job applicant feature vector corresponding to the job applicant feature sample as the talent evaluation vector of the corresponding job applicant.

5. The multi-dimensional talent recommendation method based on human resources big data according to claim 2 is characterized in that: After completing the construction of the talent pool, the method further includes the construction process of the position pool: The job requirement data in each group of job seeker-job data is cleaned, and then feature extraction is performed to determine the quantized and unified-dimensional job feature data as the job requirement feature samples, so as to form a job requirement feature sample set; Train the feature mapping model based on the job requirement feature sample set. Among them, the feature mapping model adopts the LDA model, sets M output categories, and makes the dimension of the model output data M; Based on the trained feature mapping model, determine the job requirement vector corresponding to each job requirement feature sample, where the dimension of the job requirement vector is the same as the dimension of the talent assessment vector, both of which are M; Based on the job requirement vector, a job profile of the job required by the enterprise is generated and added to the job database to complete the construction of the job database.

6. The multi-dimensional talent recommendation method based on human resources big data according to claim 5 is characterized in that: The job requirement data in each group of job seeker-job data is cleaned, and then feature extraction is performed to determine the quantized and unified-dimensional job feature data as the job requirement feature sample to form a job requirement feature sample set, including: For each group of job seekers-job data, job demand data: Clean the job demand data; Extract text feature vectors from job responsibilities and skill requirements in the job demand data after data cleaning to obtain feature components of the seventh and eighth lengths, respectively; and extract features from work experience and educational requirements to obtain feature components of the ninth and tenth lengths; Combine the various feature components to obtain a sample of job requirement features of a unified dimension; Based on the job requirement feature samples corresponding to the job requirement data in each set of job seeker-job data, a job requirement feature sample set is constructed.

7. The multi-dimensional talent recommendation method based on human resources big data according to claim 5 is characterized in that: After the construction of the job database is completed, the method further includes: Based on the talent evaluation vectors of all job seekers in the talent pool and M cluster centers, a talent weight vector is determined; Based on the M output categories of the feature mapping model, determine the position weight vector; Constructing an objective function for calculating the fitness; Construct a training data set based on the sample set of job seeker characteristics and the sample set of job requirement characteristics; The objective function is trained and optimized based on the training data set to obtain the optimized fitness function.

8. The multi-dimensional talent recommendation method based on human resources big data according to claim 7 is characterized in that: Based on the talent evaluation vectors of all job seekers in the talent pool and M cluster centers, a talent weight vector is determined, including: For the kth cluster center among M cluster centers: Determine the candidate feature samples that belong to the kth cluster center in the candidate feature sample set, and determine the maximum distance within the cluster to the kth cluster center Minimum distance within a cluster and the average distance within the cluster The distance weight reference value is calculated using the following formula: The sample weight reference value is calculated using the following formula: Among them, p k is the number of job seeker feature samples belonging to the kth cluster center in the job seeker feature sample set; Based on distance weight reference value and sample weight reference value The talent weight parameter is calculated using the following formula: in, is the talent weight parameter of the kth cluster center; Based on the talent weight parameters of M cluster centers, a talent weight vector is formed.

9. The multi-dimensional talent recommendation method based on human resources big data according to claim 7 is characterized in that: Based on the M output categories of the feature mapping model, the position weight vector is determined, including: For the kth output category of the feature mapping model: Get the intra-cluster similarity of the k-th output category and inter-cluster similarity The following formula is used to calculate the job weight parameter of the kth output category: in, is the job weight parameter of the kth output category; Based on the job weight parameters of M output categories, a job weight vector is formed 10. The multi-dimensional talent recommendation method based on human resources big data according to claim 7 is characterized in that: The optimized fitness function is: Among them, P(AB) represents the degree of fit between the talent assessment vector A and the job demand vector B. The talent assessment vector A = [A1,…,A k ,…,A M ], job demand vector B = [B1,…,B k ,…,B M ], M is the number of dimensions of the talent evaluation vector A and the job demand vector B, δ is the adaptive weight vector, δ=[δ1,…,δ k ,…,δ M ],r A is the talent weight vector, r B is the position weight vector,

Images