Precise talent matching and recommending system based on big data

Abstract

The invention relates to the field of big data processing, and discloses a big data-based accurate talent matching and recommendation system, which comprises a post demand modeling module used for converting post demand data into a multi-dimensional feature vector; the candidate feature modeling module is used for converting candidate data into multi-dimensional feature vectors; the matching cost construction module constructs a matching cost matrix; the optimal transmission matching module is used for solving an optimal transmission plan through an optimization algorithm and outputting a matching relationship between posts and candidates; and the matching result output module is used for outputting matching results and matching priorities of the candidates and the posts according to the optimal transmission plan. A post and candidate modeling modularization scheme is adopted, demands and features are converted into a multi-dimensional matrix, weight distribution is generated in combination with scores, and the post and candidate features are accurately represented. Compared with static rule or keyword matching, the problems of incomplete features and inflexible models are solved, and the problems of incomplete feature description and lack of flexibility of matching models are solved.

Description

Technical Field

[0001] The present invention relates to the field of big data processing, and in particular to an accurate talent matching and recommendation system based on big data. Background Art

[0002] In the field of human resource management, especially in the scenarios of recruitment and talent recommendation, how to achieve efficient matching between job requirements and candidate characteristics has always been the focus and difficulty of technology research and development. In the existing technology, recruitment systems or recommendation platforms mainly rely on the following two methods to achieve talent matching: one is based on simple keyword matching; the other is to recommend candidates through rule engines or heuristic algorithms. However, these traditional methods show obvious shortcomings when faced with complex job requirements and multi-dimensional candidate characteristics, and it is difficult to meet the needs of modern recruitment scenarios.

[0003] First, the traditional method based on keyword matching has the problem of insufficient feature expression. This method usually matches the keywords in the job description with the candidate's resume. The matching results often only stay at the semantic level, while ignoring the quantitative modeling of job and candidate characteristics. For example, for the skill requirements of a certain position, simple keyword matching may regard "proficient in Java" and "understanding Java" as the same match. This lack of precise representation leads to poor accuracy and interpretability of recommendation results. Especially when it is necessary to comprehensively evaluate the multi-dimensional characteristics of candidates (such as education, skills, and experience), traditional methods are even more inadequate.

[0004] Secondly, matching technologies based on rule engines or heuristic algorithms have problems with inefficiency and inflexibility. These technologies usually rely on manually set rules or weights (such as the importance weight of specific job skills). Although they can improve the matching quality to a certain extent, the rule adjustment process is time-consuming and lacks flexibility when facing complex changes in job requirements. In addition, these methods usually use greedy algorithms or simple heuristic optimization algorithms, which cannot achieve global optimal matching in multi-dimensional matching scenarios, especially when job requirements and candidate distribution are uneven, which can easily lead to unreasonable resource allocation and affect the recommendation effect.

[0005] Finally, existing technologies lack scientificity and comprehensiveness in the evaluation of matching results. Traditional methods mostly evaluate recommendation results based on a single indicator (such as matching score), which cannot effectively reflect the global matching effect and distribution differences of the system. For example, in large-scale job and candidate matching, the overall matching quality cannot be quantified based on local matching alone, resulting in unclear optimization direction and less than ideal matching effect. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a precise talent matching and recommendation system based on big data, which solves the problems in the prior art of insufficiently comprehensive feature expression, insufficient globality and efficiency of the matching algorithm, and a single evaluation method.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a precise talent matching and recommendation system based on big data, comprising: The job demand modeling module is used to convert job demand data into multi-dimensional feature vectors, form a job demand feature matrix, and generate job weight distribution based on job importance; The candidate feature modeling module is used to convert candidate data into multi-dimensional feature vectors to form a candidate feature matrix and generate candidate weight distribution based on the candidate's comprehensive score; A matching cost construction module is used to construct a matching cost matrix based on the feature differences between job requirement features and candidate features; The optimal transmission matching module is used to model the probabilistic transmission relationship between the job demand distribution and the candidate feature distribution according to the matching cost matrix, the job weight distribution and the candidate weight distribution, and solve the optimal transmission plan through the optimization algorithm to output the matching relationship between the job and the candidate; The matching result output module is used to output the matching results and matching priorities of candidates and positions according to the optimal transmission plan.

[0008] Preferably, the job requirement feature matrix includes skill requirement weights, work experience requirements, industry category codes and geographic location features in the job description.

[0009] Preferably, the candidate feature modeling module extracts candidate feature data through a semantic embedding model to generate a candidate feature vector, wherein the semantic embedding model includes a natural language processing model based on deep learning, which is used to extract skill descriptions, work experience, and educational background in the candidate's resume text and represent them as standardized multi-dimensional feature vectors.

[0010] Preferably, the matching cost matrix is ​​calculated based on a geometric distance between the job requirement feature matrix and the candidate feature matrix, including one of a Euclidean distance and a cosine distance.

[0011] Preferably, the position weight distribution is generated based on the importance score of the position, and the candidate weight distribution is generated based on the comprehensive score of the candidate, and the comprehensive score is further calculated based on the candidate's education, work experience and skill matching degree.

[0012] Preferably, the optimal transmission matching module adopts an entropy regularization optimization algorithm for solving, and improves numerical stability by introducing a regularization term.

[0013] Preferably, the weight of the regularization term is controlled by a regularization coefficient, and the regularization coefficient is dynamically adjusted according to the sparsity of the matching cost matrix.

[0014] Preferably, the entropy regularization optimization algorithm adopts the Sinkhorn-Knopp algorithm to output the optimal transfer matrix by iteratively updating the position weights and candidate weights.

[0015] Preferably, the matching result output module generates a matching relationship through the allocated weight value of each position and candidate in the optimal transmission matrix, and prioritizes the matching relationship based on the allocated weight value, and the allocated weight value is determined by the optimal transmission matching module when solving the probabilistic transmission relationship between the position demand distribution and the candidate feature distribution.

[0016] Preferably, the matching result evaluation method is performed through the following two indicators: The calculation of total matching cost is used to quantify the overall matching cost between job requirements and candidate characteristics; The calculation of Wasserstein distance is used to quantify the overall difference between the job requirement distribution and the candidate feature distribution.

[0017] The present invention provides a precise talent matching and recommendation system based on big data. It has the following beneficial effects: 1. The present invention adopts a modularized technical solution for job requirement modeling and candidate feature modeling, converts job requirement data and candidate feature data into a multidimensional feature matrix, and generates a weight distribution based on job importance scores and candidate comprehensive scores, thereby achieving the technical effect of accurately characterizing job requirements and candidate features. Compared with the technical solution in the prior art that only relies on static rules or simple keyword matching, the present invention solves the problem of incomplete description of job and candidate features and lack of flexibility in matching models.

[0018] 2. The present invention adopts the entropy regularization optimization algorithm combined with the optimization solution technology of the Sinkhorn-Knopp algorithm, introduces the regularization term in the matching process to improve the numerical stability of the calculation, and realizes efficient solution through iterative updating, thus achieving the technical effect of global optimal matching. Compared with the technical solutions based on greedy algorithms or heuristic algorithms in the prior art, the problems of low matching efficiency and insufficient global optimality are solved.

[0019] 3. The present invention uses a dual-level evaluation method of total matching cost and Wasserstein distance to conduct a comprehensive quantitative analysis of the matching results from the two perspectives of matching cost and distribution similarity, achieving the technical effect of improving the accuracy of system evaluation and optimizing feedback effects. Compared with the technical solutions in the prior art that lack global distribution evaluation or rely only on a single indicator, this solves the problem of a single evaluation dimension and insufficient optimization guidance. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] In order to better understand the present invention, the above contents are described in detail below in conjunction with specific embodiments.

[0022] The embodiment of the present invention provides a precise talent matching and recommendation system based on big data, including: The job demand modeling module is used to convert job demand data into multi-dimensional feature vectors, form a job demand feature matrix, and generate job weight distribution based on job importance; In this embodiment, the job requirement modeling module converts information such as skill requirements, work experience, industry category and geographical location involved in the job description into a multi-dimensional feature vector by extracting and processing the job requirement data, and constructs a job requirement feature matrix; at the same time, based on the job importance score (such as the urgency of the job, recruitment priority, etc.), a corresponding job weight distribution is generated to represent the weights of different jobs in the matching process.

[0023] The candidate feature modeling module is used to convert candidate data into multi-dimensional feature vectors to form a candidate feature matrix and generate candidate weight distribution based on the candidate's comprehensive score; In this embodiment, the candidate feature modeling module obtains data from the candidate's resume, and then uses a semantic embedding model to convert the skill description, work experience, and educational background information in the candidate's resume into a multi-dimensional feature vector, thereby constructing a candidate feature matrix; then, based on the candidate's comprehensive score (including education, work experience, and skill matching factors), a candidate weight distribution is generated to characterize the priority of different candidates in the matching process.

[0024] A matching cost construction module is used to construct a matching cost matrix based on the feature differences between job requirement features and candidate features; In this embodiment, the matching cost construction module constructs a matching cost matrix by calculating the feature differences between the job requirement feature matrix and the candidate feature matrix, wherein the feature differences are calculated based on geometric distance, usually using Euclidean distance or cosine distance, to quantify the degree of mismatch between job requirements and candidate features.

[0025] The optimal transmission matching module is used to model the probabilistic transmission relationship between the job demand distribution and the candidate feature distribution according to the matching cost matrix, the job weight distribution and the candidate weight distribution, and solve the optimal transmission plan through the optimization algorithm to output the matching relationship between the job and the candidate; In this embodiment, the optimal transmission matching module introduces a matching cost matrix, a job weight distribution, and a candidate weight distribution to model the probabilistic transmission relationship between the job demand distribution and the candidate feature distribution, and uses this to construct an optimization objective function, aiming to minimize the global matching cost; this module uses an entropy regularization optimization algorithm to iteratively calculate the transmission weights between jobs and candidates, and ultimately solves the optimal transmission plan, generates a matching relationship between jobs and candidates, and provides accurate recommendation results.

[0026] The matching result output module is used to output the matching results and matching priorities of candidates and positions according to the optimal transmission plan.

[0027] In this embodiment, the matching result output module extracts the matching relationship between each position and the candidate according to the optimal transmission plan generated by the optimal transmission matching module, and prioritizes the matching relationship by the assigned weight value, and finally outputs the matching result between the candidate and the position and its priority.

[0028] The job requirement feature matrix includes the skill requirement weights, work experience requirements, industry category codes and geographical location characteristics in the job description.

[0029] In this embodiment, the job requirement feature matrix includes the following feature information: First, the skill requirement weights in the job description are an important part of the feature matrix. The skill requirement weights are used to quantify the degree to which a job requires a specific skill. For example, for a software development position, it may be necessary to give a higher weight value, such as 0.8, to the "Java" skill, and a lower weight value, such as 0.3, to the "Python" skill. The skill requirement weights can be implemented by combining keyword extraction technology in the job description with natural language processing algorithms (such as the BERT model). Specifically, the frequency or semantic importance of each skill in the job description can be calculated to determine the weight value. As an option, these weights can be further normalized to ensure the numerical stability of the feature matrix.

[0030] Secondly, work experience requirements are another key dimension of the job requirement feature matrix. Generally speaking, work experience requirements can be extracted from the numbers clearly stated in the job description, such as "more than 3 years of experience" or "5 years of work experience". In one possible implementation, work experience requirements can be quantified by setting discrete intervals (such as 0-2 years, 3-5 years, more than 6 years) to generate corresponding experience values. The experience value can also be combined with the importance weight of the position for weighted processing.

[0031] Finally, the industry category code is used to identify the industry category to which the position belongs. This code can be annotated by the standard classification system of the industry to which the position belongs (such as the ISIC standard). Specifically, the industry category code is a multi-category feature vector, with each category corresponding to a one-hot encoding value. For example, for "software development industry", its encoding can be [1,0,0,0,…], while for "manufacturing industry", its encoding may be [0,1,0,0,…]. This encoding method helps to adjust the industry relevance weight in the subsequent matching cost calculation.

[0032] And, finally, the geographical location features in the job description are also an important part of the feature matrix. In general, geographical location features can be generated through a standard geographic coding system (such as ISO country and region codes). For example, the geographical location features of Beijing's positions can be encoded as "China-Beijing-Chaoyang District", and further converted into corresponding numerical or vector forms. In some embodiments, in order to improve computing efficiency, geographical location features can be represented by hierarchical locations as hierarchical coding vectors of countries, provinces, and cities, such as [China, Beijing, Chaoyang District]->[1,0,0,…]. In addition, the geographical location can also be combined with the business trip frequency requirements in the job requirements to generate a composite geographical weight.

[0033] All dimensional information of the job requirement feature matrix can be unified into one In matrix form, is the number of positions, is the number of feature dimensions. Each row of the matrix corresponds to a position, and its element value is the standardized feature value. For example, assuming that the skill requirement weight of a position is [0.8, 0.3], the work experience requirement is 5, the industry category code is [1, 0, 0], and the geographical location feature is [1, 0], then its feature vector can be expressed as:

[0034] In this case, the construction of the feature matrix can be expressed as:

[0035] in, All elements are normalized.

[0036] In order to further improve the matching adaptability of the feature matrix, in some embodiments, weight values ​​may be assigned to different dimensions of the feature matrix to generate a weighted feature matrix. For example, for some skill-oriented positions, the skill requirement weight may be increased, while for cross-industry positions, the industry category code may have a higher priority.

[0037] In a possible extension technology, the feature matrix can also be updated in real time to adapt to dynamic job requirements. For example, when a job requirement changes (such as adding a new skill requirement or changing the work location), the job requirement modeling module can partially update the feature matrix through incremental learning methods without regenerating the entire matrix.

[0038] Through the above implementation method, the job requirement feature matrix can comprehensively and accurately represent the multi-dimensional information of each job, providing efficient support for subsequent matching cost calculation and global optimization matching.

[0039] The candidate feature modeling module extracts candidate feature data through a semantic embedding model to generate a candidate feature vector. The semantic embedding model includes a natural language processing model based on deep learning, which is used to extract skill descriptions, work experience, and educational background in the candidate's resume text and represent it as a standardized multi-dimensional feature vector.

[0040] In this embodiment, the candidate feature modeling module extracts candidate feature data and generates a candidate feature vector through a semantic embedding model. The semantic embedding model includes a natural language processing model based on deep learning, specifically, a pre-trained language model such as BERT is used to extract semantic information from the candidate resume.

[0041] The semantic embedding model first performs preprocessing operations such as word segmentation and stop word removal on the candidate's resume text. Word segmentation can divide the text into basic units. For example, "proficient in using Python for data analysis" can be divided into "proficient", "use", "Python", "conduct", and "data analysis". As an option, the semantic embedding model can convert the above words into vector representations based on word embedding technology. For example, the vector corresponding to the word "Python" may be .

[0042] The pre-trained model can capture the deep semantic information of texts such as skill descriptions and work experience by training on a large-scale corpus. Key phrases in skill descriptions, such as "proficient in Java" and "proficient in SQL", can be encoded by the BERT model to obtain multi-dimensional vector representations. These vectors are concatenated or weighted averaged to form an overall representation of the candidate in terms of skill characteristics.

[0043] The extraction of work experience is usually based on the time series information of the context. For example, "Worked as a data analyst in Company A from 2018 to 2020" can be extracted through the natural language processing model to obtain the time information "2018-2020", the company name "Company A", the job title "Data Analyst" and the relevant task description. As an implementation method, the time information can be encoded through the time slice embedding technology, while the job title and task description generate vectors through semantic embedding.

[0044] The processing of educational background mainly focuses on the extraction of information such as school name, major, and academic level. For example, "Master of Computer Science from Peking University" can be decomposed into "school = Peking University", "major = computer science", and "education level = Master". In this implementation, the academic level can be discretized, such as "Master = 3, Bachelor = 2, Associate = 1".

[0045] Therefore, the output feature vector of the semantic embedding model can be expressed as:

[0046] in, represents the skill description feature vector; represents the work experience feature vector; Represents the educational background feature vector.

[0047] Each dimension of the feature vector is normalized and limited to the range of [0,1] to ensure the numerical stability of subsequent matching calculations. In some embodiments, normalization can be achieved through Min-Max Scaling. For example, for the work experience dimension, assuming that the experience value range is 1 to 10 years, the normalization formula is:

[0048] For example, if the current experience is 5 years, the normalized value is:

[0049] Finally, in the above implementation, the final form of the candidate feature vector is a standardized multi-dimensional vector, expressed as:

[0050] Among them, each dimension corresponds to a specific feature, such as skill weight, industry experience, academic qualifications, etc.

[0051] As an extension, the semantic embedding model can also combine additional information such as the candidate's social network data, publicly published papers or project experience. This information can further enrich the dimensions of the candidate's feature vector, making it more comprehensive. For example, the star rating of a GitHub project can be used as a supplement to the candidate's skill features, and the number of paper citations can be used as one of the features of the academic background. Therefore, the candidate feature modeling module can convert complex resume text information into structured feature vectors, providing basic data support for subsequent matching cost calculation and optimized matching.

[0052] The matching cost matrix is ​​calculated based on the geometric distance between the job requirement feature matrix and the candidate feature matrix, including one of the Euclidean distance and the cosine distance.

[0053] In this embodiment, the matching cost matrix is ​​calculated based on the geometric distance between the job requirement feature matrix and the candidate feature matrix. The geometric distance is used to quantify the difference between the job and the candidate in the multidimensional feature space, and specifically includes two types: Euclidean distance and cosine distance.

[0054] Generally, Euclidean distance is used to measure the absolute difference between job and candidate characteristics. The job requirement feature vector is expressed as , the candidate feature vector is expressed as , the Euclidean distance between the two The calculation formula is:

[0055] Among them, it means job With the candidates The matching cost between ; and Represents the position and candidate in The eigenvalues ​​in the dimensions; The number of dimensions of the feature vector.

[0056] As an option, the matching cost matrix can also be calculated based on the cosine distance. The cosine distance reflects the matching relationship by measuring the directional similarity between feature vectors. Its formula is:

[0057] in, represents the inner product between eigenvectors; and Represent the modulus of the position and candidate feature vectors respectively.

[0058] In one possible implementation, the matching cost matrix It can be expressed as:

[0059] in, is the number of positions; is the number of candidates; indicates the position The cost of matching with candidate j.

[0060] Through the above methods, the calculation of the matching cost matrix can fully reflect the multi-dimensional differences between job requirements and candidate characteristics, providing reliable data input for the subsequent optimal transmission matching module. This matrix plays a core role in accurate matching recommendations, ensuring the accuracy and efficiency of matching results.

[0061] The position weight distribution is generated based on the position importance score, and the candidate weight distribution is generated based on the candidate's comprehensive score, which is further calculated based on the candidate's education, work experience, and skill match.

[0062] In this embodiment, the position weight distribution is generated based on the importance score of the position. Specifically, the importance score of each position is set by the system based on factors such as the urgency of the enterprise's needs, the priority of the position, or the business impact. As an option, an initial score can be assigned to each position, for example, a high priority position is scored as 100, a medium priority position is scored as 80, and a low priority position is scored as 60. After the scoring is completed, the weight distribution is normalized and converted into a probability distribution form:

[0063] in, Indicates position The weight of For the position Importance rating; The total number of positions.

[0064] For example, if there are three positions, and their importance scores are 100, 80, and 60 respectively, then the corresponding weight distribution after normalization is:

[0065]

[0066]

[0067] The candidate weight distribution is generated based on the candidate's comprehensive score. The comprehensive score is further calculated based on the candidate's education, work experience and skill match. In general, the comprehensive score is obtained by weighting multiple sub-items:

[0068] in, , and Represent the weight coefficients of academic qualifications, work experience and skill scores respectively; and the weight coefficients satisfy =1.

[0069] The academic qualification score can be completed by mapping different academic qualification levels to numerical values, such as "Master = 3, Bachelor = 2, Associate = 1". The experience score is quantified according to the range of working years, such as "0-2 years = 1, 3-5 years = 2, 6 years or more = 3". The skill score calculates the skill matching degree between the candidate and the position by matching the skill feature dimension in the cost matrix, and the matching score range is [0,1].

[0070] In one possible implementation, assume that a candidate has a bachelor's degree, has 5 years of work experience, and has a skill match of 0.75 with the position. The weight coefficient is , , , the comprehensive score is calculated as follows:

[0071] The candidate weight distribution also needs to be normalized to generate a probability distribution form:

[0072] in, Indicates the candidate The weight of For candidates The overall score of is the total number of candidates.

[0073] For example, suppose there are three candidates with comprehensive scores of 1.875, 2.5, and 2.0, respectively. The corresponding weight distribution after normalization is:

[0074]

[0075]

[0076] As an extension, the comprehensive scoring can be combined with a dynamic adjustment strategy. For example, when the needs of the enterprise or the job requirements change, the candidate weight distribution can be updated in real time to ensure the dynamic adaptability of the system. In some embodiments, external data sources (such as the candidate's reputation in the industry or social media activity) can also be introduced to further supplement the scoring mechanism.

[0077] Therefore, through the above calculation method, the job weight distribution and the candidate weight distribution are uniformly modeled as probability distribution, providing basic support for the global optimization of the matching cost matrix and the optimal transmission matching module.

[0078] The optimal transmission matching module adopts the entropy regularization optimization algorithm to solve, and the numerical stability is improved by introducing the regularization term. The weight of the regularization term is controlled by the regularization coefficient, and the regularization coefficient is dynamically adjusted according to the sparsity of the matching cost matrix. The entropy regularization optimization algorithm adopts the Sinkhorn-Knopp algorithm to output the optimal transmission matrix by iteratively updating the position weight and candidate weight.

[0079] In this embodiment, the entropy regularization optimization algorithm adds a regularization term to the objective function of the optimal transmission problem, and rewrites the objective function as follows:

[0080] in, Indicates position With the candidates The transmission weight of Represents the positions in the matching cost matrix With the candidates The matching cost; is the regularization coefficient, which is used to control the weight of the regularization term; it represents the boundary constraints of the job weight distribution and the candidate weight distribution; and They are the boundary constraints for job weight distribution and candidate weight distribution respectively; represents the set of all transmission matrices that satisfy the above boundary conditions.

[0081] In general, the introduction of regularization terms can effectively improve the numerical stability of sparse matrices and accelerate the convergence of optimization algorithms. As an option, the regularization coefficient It can be dynamically adjusted according to the sparsity of the matching cost matrix. Specifically, the sparsity can be measured by the proportion of non-zero elements in the cost matrix, for example:

[0082] in is the number of positions, is the number of candidates. When the sparsity is high, a smaller regularization coefficient can be selected to enhance the flexibility of the model; when the sparsity is low, a larger regularization coefficient can be selected to improve numerical stability. In one possible implementation, the regularization coefficient The adjustment formula is:

[0083] in, To adjust the parameters; the sparsity is dynamically calculated based on the sparsity of the matching cost matrix.

[0084] The solution of the entropy regularization optimization algorithm is implemented by the Sinkhorn-Knopp algorithm. The algorithm uses an iterative calculation method of bilateral normalization to gradually update the job weights and candidate weights. Specifically, the kernel matrix is ​​initialized for:

[0085] And initialize the vector and is a vector of all 1s. The iterative update formula is:

[0086]

[0087] in, is the job weight distribution; Weight distribution for candidates; and Respectively represent the results of matrix-vector multiplication; This updating process is repeated until and Convergence. After the convergence conditions are met, the final optimal transmission matrix is ​​calculated as:

[0088] in, and Represents vectors and Convert to a diagonal matrix.

[0089] As an extension, in some embodiments, the transmission matrix The non-zero elements of can be further screened to eliminate matching relationships with low transmission weights to improve the accuracy of recommendation results. For example, a threshold can be set Keep only matching relationship.

[0090] Through the above implementation method, the entropy regularization optimization algorithm not only improves the numerical stability of the optimal transmission matching module, but also speeds up the calculation speed of the algorithm, providing efficient technical support for the accurate matching of positions and candidates.

[0091] The matching result output module generates matching relationships through the allocation weight values ​​of each position and candidate in the optimal transmission matrix, and prioritizes the matching relationships based on the allocation weight values. The allocation weight values ​​are determined by the optimal transmission matching module when solving the probabilistic transmission relationship between the position demand distribution and the candidate feature distribution.

[0092] In this embodiment, the matching result output module uses the optimal transmission matrix The allocation weight values ​​in generate the matching relationship between positions and candidates. Specifically, the optimal transfer matrix is a The matrix of is the number of positions, is the number of candidates, each element in the matrix Indicates position With the candidates The transmission weight value is determined by the optimal transmission matching module in solving the job demand distribution. Distribution of candidate characteristics The probability transmission relationship between them is determined.

[0093] Generally, the generation of matching relationships is based on the extraction of non-zero elements in the transfer matrix. , by filtering Candidates , a list of matching candidates for the position can be generated. As an option, the system can set a threshold , only keep , thereby improving the accuracy of recommendations. , , , and set , then only keep and matching relationship.

[0094] Specifically, the priority ranking is based on the size of the assigned weight value. The priority of the matching relationship can be expressed as: Prioritization:

[0095] in:

[0096] In one possible implementation, for each position, the system sorts the candidates according to the assigned weight value from large to small. The candidate weights are , , , then the priority after sorting is , , .

[0097] As an extension, the system can also output the matching probability distribution between candidates and positions to quantify the credibility of each matching relationship. The matching probability can be obtained by normalizing the assigned weight value, and the formula is:

[0098] in, Indicates position For candidates The matching probability of Indicates position The total weight distribution value.

[0099] For example, if the allocation weight value of a position is , , , then the corresponding matching probability is:

[0100]

[0101]

[0102] In some embodiments, the matching result output module can combine the matching results of multiple positions to generate a global matching priority. , the global matching priority is calculated based on the product of the position weight and the matching relationship weight:

[0103] in, is the weight distribution value of the position, For candidates About the position The allocation weight value of .

[0104] Therefore, the matching result output module can accurately generate matching relationships and optimize recommendation results through priority sorting and probability distribution, providing intuitive and reliable technical support for the accurate matching of positions and candidates.

[0105] The matching results are evaluated using the following two indicators: The calculation of total matching cost is used to quantify the overall matching cost between job requirements and candidate characteristics; The calculation of Wasserstein distance is used to quantify the overall difference between the job requirement distribution and the candidate feature distribution.

[0106] In this embodiment, the matching result evaluation is achieved through the following two indicators: total matching cost and Wasserstein distance.

[0107] Generally speaking, the total cost of matching is used to measure the overall matching cost between job requirements and candidate characteristics. The calculation formula is:

[0108] in, is the position in the optimal transmission matrix and candidates The allocation weight value of To match the positions in the cost matrix and candidates The matching cost; and are the number of positions and candidates, respectively.

[0109] Specifically, the total matching cost reflects the overall matching cost of all positions and candidates. As an implementation method, if the number of positions is 3 and the number of candidates is 3, the transmission matrix and cost matrix are as follows:

[0110]

[0111] The total matching cost is calculated as:

[0112]

[0113] As an extension, the total matching cost can be further subdivided into skill cost, experience cost, etc., so as to perform dimensional analysis on matching quality. For example, the skill cost calculates the matching cost of the skill dimension separately to evaluate the matching of candidates and positions in terms of skills.

[0114] Wasserstein distance is used to quantify the overall difference between job demand distribution and candidate feature distribution. is defined as:

[0115] in, Indicates the job weight distribution; represents the weight distribution of candidates; The transfer weight between positions and candidates; are the elements of the matching cost matrix.

[0116] In one possible implementation, if the job weight distribution is , the candidate weight distribution is , the corresponding matching cost matrix is:

[0117] The transmission matrix solved by the optimal transmission matching algorithm is:

[0118] Then the Wasserstein distance is:

[0119] As an option, the Wasserstein distance can be calculated hierarchically in combination with job categories or skill dimensions to evaluate the global match between different categories of jobs and candidates.

[0120] Through the comprehensive evaluation of the above two indicators, the system can quantify the quality of matching results from both local and global levels, providing a scientific basis for the optimization and adjustment of recommendation results. The total matching cost reflects the matching efficiency of the system, while the Wasserstein distance evaluates the overall similarity between the job demand distribution and the candidate feature distribution. This evaluation method provides comprehensive data support for subsequent matching optimization.

[0121] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A precise talent matching and recommendation system based on big data, characterized in that: include: The job demand modeling module is used to convert job demand data into multi-dimensional feature vectors, form a job demand feature matrix, and generate job weight distribution based on job importance; The candidate feature modeling module is used to convert candidate data into multi-dimensional feature vectors to form a candidate feature matrix and generate candidate weight distribution based on the candidate's comprehensive score; A matching cost construction module is used to construct a matching cost matrix based on the feature differences between job requirement features and candidate features; The optimal transmission matching module is used to model the probabilistic transmission relationship between the job demand distribution and the candidate feature distribution according to the matching cost matrix, the job weight distribution and the candidate weight distribution, and solve the optimal transmission plan through the optimization algorithm to output the matching relationship between the job and the candidate; The matching result output module is used to output the matching results and matching priorities of candidates and positions according to the optimal transmission plan.

2. According to the big data-based accurate talent matching and recommendation system of claim 1, it is characterized in that: The job requirement feature matrix includes skill requirement weights, work experience requirements, industry category codes and geographical location features in the job description.

3. According to the big data-based accurate talent matching and recommendation system of claim 1, it is characterized in that: The candidate feature modeling module extracts candidate feature data through a semantic embedding model to generate a candidate feature vector. The semantic embedding model includes a natural language processing model based on deep learning, which is used to extract skill descriptions, work experience, and educational background in the candidate's resume text and represent them as standardized multi-dimensional feature vectors.

4. According to claim 1, a precise talent matching and recommendation system based on big data is characterized in that: The matching cost matrix is ​​calculated based on the geometric distance between the job requirement feature matrix and the candidate feature matrix, including one of the Euclidean distance and the cosine distance.

5. According to the big data-based accurate talent matching and recommendation system of claim 1, it is characterized in that: The position weight distribution is generated based on the importance score of the position, and the candidate weight distribution is generated based on the comprehensive score of the candidate, which is further calculated based on the candidate's education, work experience and skill matching degree.

6. According to claim 1, a precise talent matching and recommendation system based on big data is characterized in that: The optimal transmission matching module adopts an entropy regularization optimization algorithm for solving, and improves numerical stability by introducing a regularization term.

7. According to claim 6, a precise talent matching and recommendation system based on big data is characterized in that: The weight of the regularization term is controlled by the regularization coefficient, and the regularization coefficient is dynamically adjusted according to the sparsity of the matching cost matrix.

8. The accurate talent matching and recommendation system based on big data according to claim 6 is characterized in that: The entropy regularization optimization algorithm adopts the Sinkhorn-Knopp algorithm to output the optimal transfer matrix by iteratively updating the position weights and candidate weights.

9. The accurate talent matching and recommendation system based on big data according to claim 8 is characterized in that: The matching result output module generates a matching relationship through the allocation weight value of each position and candidate in the optimal transmission matrix, and prioritizes the matching relationship based on the allocation weight value. The allocation weight value is determined by the optimal transmission matching module when solving the probabilistic transmission relationship between the position demand distribution and the candidate feature distribution.

10. The accurate talent matching and recommendation system based on big data according to claim 9, characterized in that: The matching result evaluation method is performed through the following two indicators: The calculation of total matching cost is used to quantify the overall matching cost between job requirements and candidate characteristics; The calculation of Wasserstein distance is used to quantify the overall difference between the job requirement distribution and the candidate feature distribution.

Images