An employment information matching method and system based on data analysis

Abstract

The application discloses an employment information matching method and system based on data analysis, and particularly relates to the field of employment matching. Structured and unstructured data are collected, and BERT model and BiLSTM-CRF are used for semantic feature extraction. In the preprocessing stage, entity standardization is realized through a knowledge graph, and a job seeker portrait and a post model containing a skill matrix and a career development track are constructed. In the feature engineering stage, four core features, i.e. skill matching degree, salary expectation fitting degree, commuting tolerance and career development fitting degree, are extracted. The salary fitting degree is quantitatively evaluated through a bidirectional tolerance model, a dynamic weight is obtained by using a random forest model with time attenuation, feature weights are generated based on historical successful cases, and a real-time feedback mechanism is introduced to adjust the weight coefficient. Finally, a final matching degree function fuses weighted features and industry trend factors, and a three-level updating mechanism is used to continuously optimize the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of employment matching, more particularly, the present application relates to an employment information matching method and system based on data analysis. BACKGROUND

[0002] In today's society, the employment market presents a complex and dynamic situation. On the one hand, with the number of college graduates rising year by year, and the frequent flow of talents in various sectors of society, the size of job seekers is expanding, and their employment needs are becoming more diversified, covering industry selection, job type, salary, work location, career development prospects and other dimensions.

[0003] On the other hand, the demand for talent by enterprises is also changing. Under the background of industrial upgrading and technological innovation, the skill requirements of enterprises for positions are becoming increasingly refined and specialized, and there are significant differences in recruitment standards among different industries and different sizes of enterprises. However, the current employment information matching mode is difficult to meet the complex needs of both supply and demand.

[0004] However, it still has some shortcomings in actual use. Traditional employment information matching relies on simple keyword search. Job seekers filter information by inputting keywords such as job title and industry on recruitment platforms, lacking deep mining of multi-source heterogeneous data (such as industry trends, salary fluctuations, career development paths, etc.), and unable to respond to market changes in real time. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an employment information matching method and system based on data analysis to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] S1: Real-time acquisition of structured and unstructured employment data from enterprise recruitment databases, job seeker resume databases and industry public data platforms;

[0008] S2: Denoising, missing value filling, entity recognition and text vectorization processing of the collected data, generating standardized job seeker portraits and enterprise job models;

[0009] S3: Based on job requirements and job seeker capabilities, extracting four core features of skill matching degree, salary expectation fit degree, commuting tolerance and career development fit;

[0010] S4: Using historical matching success data to train a random forest model to generate dynamic weight coefficients for each feature;

[0011] S5: push the matching results to the user end in descending order of score, and update the weight model iteratively according to user behavior data.

[0012] Preferably, in the multi-source data collection of S1, structured data (such as job title, salary range, education requirement, etc. field information) in the enterprise recruitment database is captured in real time through a distributed crawler framework, and standardized fields (including work experience, skill certificate, education background, etc.) in the job seeker resume library are accessed; for the unstructured data part, a multi-modal data processing pipeline is adopted: first, the semantic block processing is carried out on the free text description (such as job responsibilities, job requirements, etc. paragraph text) in the enterprise recruitment brochure, and the long text is decomposed into independent semantic units through the text segmentation algorithm based on attention mechanism; second, the context correlation analysis is carried out on the project experience text in the job seeker resume, the key technology stack, business field and achievement index in the project are extracted by using named entity recognition technology, and the mapping relationship with the standardized skill tree is established; finally, the professional evaluation data of LinkedIn, Maimai and other platforms are obtained through the social network API interface, including anonymous evaluation text of employees, non-standard content such as industry discussion posts.

[0013] Preferably, in S2, the system adopts a multi-level cleaning and conversion process to standardize the original employment data. First, the data denoising operation is carried out, and for structured data (such as salary range, work experience, etc. numerical fields), the outliers are eliminated through quartile range detection and Z-score standardization; for text data, regular expression is used to filter HTML tags, special characters and meaningless stop words, and a text cleaning model based on attention mechanism is used to identify and eliminate duplicate published job information or resume content; in the missing value processing link, different strategies are adopted according to the data type: numerical fields (such as jobs with undisclosed salary) are estimated by KNN nearest neighbor algorithm of the same industry and the same position; the category type field (such as missing education requirement) uses Bayesian network to infer the probability based on other known fields; the text type missing value (such as blank project description) is marked as "not specified" without filling to avoid introducing noise;

[0014] In the entity recognition stage, a hybrid NER model integrating domain knowledge was deployed: for enterprise job descriptions, a pre-trained legal-human resources joint model was used to identify welfare entities such as "five social insurances and one housing fund" and "equity incentives"; for resume text, a fine-grained recognizer based on BiLSTM-CRF was used to extract technical experience entities such as "3 years of SpringBoot experience" and "led projects with tens of millions of page views", and the ambiguity of easily confused skills such as "Java" and "JavaScript" was resolved by constructing an industry knowledge graph; all identified entities were mapped to a standardized skill ontology (e.g., "proficient in PS" was unified as "advanced Adobe Photoshop skills"), and structured labels with confidence scores were generated.

[0015] Preferably, in step S3, during the feature extraction stage, four core matching dimensions are constructed through quantitative analysis, and each dimension adopts a composite index calculation system.

[0016] In the skill matching degree assessment, a hierarchical evaluation model is adopted. First, the hard skill matching degree is calculated based on the knowledge graph. The specific calculation method is as follows:

[0017]

[0018] in, This is expressed as a hard skill match. Let the weight of the i-th skill in the job be denoted as . This represents the job seeker's skill level. This represents the required proficiency level for the position. This indicates the job seeker's actual level;

[0019] The soft skills section is calculated based on the semantic similarity between the resume text and the job description. The specific calculation method is as follows:

[0020]

[0021] Where B represents the soft skills match, Represented as a text vector of the job description. Represented as a vector of resume text;

[0022] The resume text and job description text are transformed into high-dimensional vectors through natural language processing.

[0023] The final skill match is a weighted harmonic average of hard and soft skills, calculated as follows:

[0024]

[0025] Where C represents the final skill match, A represents the hard skill match, and B represents the soft skill match. denotes the hard skill weight factor,

[0026] A bidirectional tolerance model is constructed, considering the salary expectation fit between the enterprise budget interval and the job seeker expectation interval, and the calculation method is as follows:

[0027]

[0028] wherein D denotes the salary expectation fit, denotes the lower limit of the job seeker's salary expectation, denotes the upper limit of the job seeker's salary expectation, denotes the lower limit of the enterprise budget, i.e. the minimum salary the enterprise is willing to pay; denotes the upper limit of the enterprise budget, i.e. the maximum salary the enterprise is willing to pay; denotes the market salary range span;

[0029] In the commuting tolerance, the space-time cost calculation based on geographic information system is calculated, and the calculation method is as follows:

[0030]

[0031] wherein F denotes the commuting tolerance, denotes the actual commuting time, denotes the maximum tolerable commuting threshold, denotes the urban traffic coefficient, denotes the remote work weight, denotes the proportion of remote office days.

[0032] Preferably, in the dynamic weight allocation stage of S4, a random forest ensemble learning model with time correction is constructed based on historical matching success data, and dynamic weight coefficients are generated through multi-dimensional feature importance analysis; the specific process is as follows: first, extract positive samples (including job seeker final employment and retention for more than 3 months of job matching records) from the historical successful matching case library, and construct a negative sample set (including high matching degree but being explicitly refused by the user or leaving the job within 30 days after employment), forming a training data set with time label , is a four-dimensional feature vector, i.e. ; denotes the matching success label, denotes the matching time stamp;

[0033] The model adopts time decay weighted sampling technology, and the sampling weight of recent samples increases exponentially, and the calculation method is as follows:

[0034] ​

[0035] wherein, denotes the sampling weight of a certain sample, K denotes the decay coefficient, denotes the timestamp corresponding to the dth sample, denotes the minimum time value in all samples, denotes the maximum time value in all samples;

[0036] The decay coefficient K = 2 ensures that the model pays more attention to recent market preference changes;

[0037] In the random forest training process, when each decision tree node is split, not only the traditional Gini impurity is calculated, but also a feature aging factor is introduced, and the calculation method is as follows:

[0038]

[0039] wherein, L denotes the adjusted impurity, g denotes the traditional Gini impurity, denotes the adjustment parameter, denotes the average effective period of the current feature in historical data, denotes the aging factor;

[0040] Through parallel training of 500 decision trees, the average importance score of each feature on the OOB data is output, and the base weight is obtained after Softmax normalization, and the calculation method is as follows:

[0041]

[0042] wherein, denotes the base weight of feature f, denotes the average importance score, denotes the sum of the exponential values of all features, denotes the original importance score of feature f Take the natural exponential.

[0043] Preferably, in S5, after the matching degree is calculated, the result is first subjected to scenario processing: for enterprise HR users, the top 50 candidate simplified portraits are displayed in descending order of MatchScore, and a "job competitiveness analysis panel" is provided to intuitively show the deviation of the job from the market average level in terms of salary, skill requirements, etc.; for job seeker users, a progressive disclosure strategy is adopted, the core information (salary, commuting time, core skill matching degree) of the top 3 jobs with the highest matching degree is displayed on the first screen, and after sliding down, 20 recommended jobs ranked in descending order of MatchScore can be viewed, and a "matching advantage radar chart" is provided to visualize the scores in each dimension;

[0044] To construct an effective feedback loop, a three-level weight update mechanism was designed: short-term updates are fine-tuned based on real-time user behavior; when the click-through rate of a certain job category deviates from the historical average by ±15% for four consecutive hours, it is immediately updated. Adjust the corresponding feature weights, where, This is expressed as the adjustment amount for the feature weights. Represented as the learning rate, This represents the click-through rate for a specific job category within the current time window (a continuous 4-hour period). This is represented as the historical average click-through rate; mid-term updates are performed weekly, using successful matching cases from the past 7 days (defined as records of companies sending interview invitations after viewing resumes and job seekers accepting) to retrain the random forest model and optimize the cross-feature weights of commuting tolerance and career development; long-term updates are performed monthly in conjunction with the macro report on the human resources market, adjusting the weight allocation of each sub-item in the calculation formula of the industry trend factor λ through expert rules.

[0045] The technical effects and advantages of this invention are as follows:

[0046] This invention utilizes distributed web crawlers to collect real-time enterprise recruitment data and job seeker resumes. It employs the BERT model and BiLSTM-CRF technology to process unstructured text, constructing standardized talent profiles and job models. The core innovation lies in designing a four-dimensional feature system: skill matching, salary expectation alignment, commuting tolerance, and career development alignment. Feature weights are dynamically adjusted using a random forest model, and an industry trend factor λ is introduced for real-time calibration. Matching results are pushed using a tiered strategy, incorporating user behavior data to establish a three-level feedback mechanism: real-time fine-tuning, mid-term optimization, and long-term adjustment.

[0047] In terms of matching accuracy, a multi-level cleaning and transformation process is adopted to generate standardized profiles and models, extract four types of core features and quantify them, and combine them with a random forest model to generate dynamic weights, which greatly improves the accuracy of matching job positions with job seekers and reduces information mismatch.

[0048] The model is highly adaptable, incorporating time-decay weighted sampling, real-time feedback adjustment, and a three-level weight update mechanism, enabling the model to respond promptly to market changes and user behavior trends, dynamically optimize weights, and enhance its adaptability to the dynamics of the job market. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0050] Figure 2 This is a schematic diagram of the module connection of the present invention.

[0051] Figure 3 This is a schematic diagram illustrating the logic for extracting the four core features of this invention.

[0052] Figure 4 The user terminal recommendation logic schematic diagram of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0054] Please refer to Figure 1 As shown in the drawings, the present application provides an employment information matching method based on data analysis, and the method is as follows:

[0055] S1: Real-time acquisition of structured and unstructured employment data from enterprise recruitment database, job seeker resume database and industry public data platform;

[0056] In the multi-source data collection of S1, structured data in the enterprise recruitment database (such as job title, salary range, education requirement, etc.) is captured in real time through a distributed crawler framework, and standardized fields in the job seeker resume database (including work experience, skill certificates, education background, etc.) are accessed. For unstructured data, a multi-modal data processing pipeline is used: first, the free text description in the enterprise recruitment brochure (such as job responsibilities, job requirements, etc.) is processed for semantic blocking, and the long text is divided into independent semantic units through a text segmentation algorithm based on attention mechanism; second, the project experience text in the job seeker's resume is analyzed in context, and the key technology stack, business domain and achievement index in the project are extracted using named entity recognition technology, and a mapping relationship with the standardized skill tree is established; finally, the professional evaluation data of LinkedIn, Weimei and other platforms are obtained through social network API interface, including anonymous employee evaluation text, industry discussion posts and other non-standard content. For these heterogeneous and unstructured data, a semantic feature extraction layer based on BERT model is deployed, including: domain adaptive pre-training of recruitment brochure text, fine-tuning of BERT token embedding layer on HR professional corpus to improve the representation accuracy of professional terminology; BiLSTM-CRF joint model is used for fine-grained entity relation extraction of resume project experience, to identify "technology-project-achievement" triplets; a sentiment analysis sub-unit is constructed for social media evaluation, and the sentiment polarity score and theme clustering features are output through the BERT [CLS] marker; all extracted semantic features are converted into 768-dimensional dense vectors and stored in the vector database of Elasticsearch, while the pointer index of the original text is retained for explainability analysis. The data collection module uses Kafka message queue to realize real-time stream processing, ensuring that the whole process from data acquisition to feature extraction is completed within 15 minutes for new recruitment information.

[0057] S2: Denoising, missing value filling, entity recognition and text vectorization processing are performed on the collected data to generate standardized job seeker portraits and enterprise job models;

[0058] In S2, the system adopts a multi-level cleaning and conversion process to standardize the original employment data. First, data denoising is performed. For structured data (such as salary range, work experience, etc.), abnormal values are eliminated through quartile range detection and Z-score standardization. For text data, regular expressions are used to filter HTML tags, special characters, and meaningless stop words. A text cleaning model based on attention mechanism is used to identify and remove duplicate job postings or resume content. In the missing value processing section, different strategies are used based on data type. For numerical fields (such as jobs with undisclosed salaries), KNN nearest neighbor algorithm is used to estimate the value based on similar industries and positions. For categorical fields (such as missing education requirements), Bayesian networks are used to infer the probability based on other known fields. For text missing values (such as blank project descriptions), they are marked as "not specified" without filling to avoid introducing noise.

[0059] In the entity recognition phase, a hybrid NER model that integrates domain knowledge is deployed. For job descriptions, a pre-trained legal-human resources joint model is used to identify welfare entities such as "five insurances and one fund" and "stock option". For resume texts, a fine-grained recognizer based on BiLSTM-CRF is used to extract technical experience entities such as "SpringBoot 3 years" and "led a project with over 10 million PV". An industry knowledge graph is constructed to solve the ambiguity problem of easily confused skills such as "Java" and "JavaScript". All recognized entities are mapped to a standardized skill ontology (such as "AdobePhotoshop advanced skill") and structured labels with confidence scores are generated.

[0060] In the text vectorization section, a hierarchical feature representation strategy is implemented. For technical term-intensive text segments (such as "familiar with distributed system architecture design" in job requirements), an SBERT model is used to generate domain-optimized sentence vectors. For descriptive content (such as company culture introduction), Doc2Vec is used to retain paragraph-level semantic features. Specifically, for key information such as salary and benefits, a special numerical perception embedding model is trained (such as encoding "15-20k" into a mixed vector containing quantile features). The final standardized candidate profile includes: basic attribute vector (education / experience, etc.), skill matrix (technology stack mastery score), career development vector (job / industry transition trajectory). The enterprise job model includes: demand core vector (hard requirements), preference auxiliary vector (bonus items), competitiveness indicators (salary / welfare quantitative values).

[0061] All pre-processed data is stored in a Neo4j database in a graph structure, with nodes containing entities and their relationship weights, and edge attributes saving co-occurrence frequency and temporal characteristics, providing an interpretable feature basis for subsequent matching calculations. The entire preprocessing process is scheduled by Airflow, and a quality monitoring unit is provided to automatically trigger reprocessing when data consistency is less than 95%.

[0062] S3: Based on the job requirements and the ability of job seekers, four core features are extracted, including skill matching degree, salary expectation fit degree, commuting tolerance and career development fit degree;

[0063] In S3, in the feature extraction stage, four core matching dimensions are constructed through quantitative analysis, and each dimension uses a composite index calculation system.

[0064] In the skill matching degree, a hierarchical evaluation model is used. First, the hard skill matching degree is calculated based on the knowledge graph. The calculation method is as follows:

[0065]

[0066] Wherein, represents the hard skill matching degree, represents the weight of the i-th skill of the post, represents the skill mastery value of the job seeker, represents the proficiency level required by the post, represents the actual level of the job seeker;

[0067] The soft skill part is calculated by the semantic similarity between the resume text and the job description. The calculation method is as follows:

[0068]

[0069] Wherein, B represents the soft skill matching degree, represents the text vector of the job description, represents the vector of the resume text;

[0070] The resume text and the job description text are processed through natural language processing to convert the text into a high-dimensional vector.

[0071] The final skill matching degree is the weighted harmonic mean of hard and soft skills. The calculation method is as follows:

[0072]

[0073] Wherein, C represents the final skill matching degree, A represents the hard skill matching degree, and B represents the soft skill matching degree, represents the hard skill weight factor, ;

[0074] A two-way tolerance model is constructed, which considers the alignment between the company's budget range and the job seeker's expected salary range. The specific calculation method is as follows:

[0075]

[0076] Where D represents the salary expectation fit. This represents the lower limit of a job seeker's salary expectations. This represents the upper limit of the job seeker's salary expectation. This represents the lower limit of a company's budget, i.e., the minimum salary a company is willing to pay; This represents the company's budget ceiling, which is the highest salary the company is willing to pay. This represents the range of market salaries.

[0077] For example, suppose a major internet company is hiring an AI architect, and the company's budget... , Salary expectations of job seekers , Market benchmark ;

[0078] Feasibility check: If the conditions are met, the main calculation process will begin.

[0079] Calculate the numerator:

[0080]

[0081] Calculate salary expectation fit: If the result is a perfect match, the output will be displayed.

[0082] Another example is corporate budgets. , Salary expectations of job seekers , Market benchmark ;

[0083] Feasibility check: If the conditions are met, the main calculation process will begin.

[0084] Calculate the numerator: 14

[0085]

[0086] Calculate salary expectation fit: If the calculated salary expectation match score is extremely low, then there is almost no room for negotiation.

[0087] Market salary range span is the interval width of the median salary of the same position in the same region, for example, the difference between the 25th percentile and the 75th percentile;

[0088] In the commuting tolerance, the space-time cost calculation based on geographic information system is calculated as follows:

[0089]

[0090] Where F represents the commuting tolerance, represents the actual commuting time, represents the maximum tolerable commuting threshold, represents the urban traffic coefficient, represents the remote work weight, represents the proportion of remote office days;

[0091] The urban traffic coefficient is 0.8 for first-tier cities, 0.5 for second-tier and third-tier cities, and 0.3 for fourth-tier and fifth-tier cities.

[0092] The career development fit is calculated using a three-dimensional evaluation model, and the calculation method is as follows:

[0093]

[0094] Where G represents the career development fit, represents the path growth value, represents the ability adaptation value, represents the industry inertia value.

[0095] The path growth value is calculated by the Pearson correlation coefficient of the historical job transition speed of the job seeker and the promotion path of the target position.

[0096] The ability adaptation value is calculated by the ratio of the job seeker's learning curve slope to the job technology iteration rate.

[0097] The calculation method of the industry inertia value is as follows:

[0098]

[0099] Where, represents the industry inertia value, represents the cumulative years of service in the target industry, represents the job demand, represents the distance between the job and the candidate's industry subfield, represents the maximum allowed subfield span threshold.

[0100] The preset fixed constant 1 is the maximum cross-field range upper limit that the enterprise can accept, and a value exceeding the upper limit is regarded as being out of professional scope; That is, the distance between the two sub-races is quantified, for example, if the "Internet finance" is , and the "traditional bank" is , then the distance is |0.8-0.2|=0.6;

[0101] All features are normalized to the interval [0, 1] by the sigmoid function, and the feature distribution deviation is monitored in real time.

[0102] S4: Use the history matching success data to train the random forest model to generate dynamic weight coefficients of each feature;

[0103] In S4, in the dynamic weight distribution stage, a random forest ensemble learning model with time correction is constructed based on the history matching success data, and dynamic weight coefficients are generated through multi-dimensional feature importance analysis. The specific process is as follows: first, extract the positive samples (including the job seeker's final job placement record and the retention of more than 3 months) from the history successful matching case library, and construct the negative sample set (including the cases of high matching degree but being explicitly refused by the user or leaving the job within 30 days after employment), form a training data set with time label , is a four-dimensional feature vector, that is ; is represented as a matching success label, is represented as a matching time stamp;

[0104] The model uses time decay weighted sampling technology, and the sampling weight of recent samples increases exponentially. The calculation method is as follows:

[0105]

[0106] Among them, is the sampling weight of a certain sample, K is the decay coefficient, is the time stamp corresponding to the dth sample, is the minimum time value in all samples, is the maximum time value in all samples;

[0107] The decay coefficient K=2 ensures that the model pays more attention to recent market preference changes;

[0108] In the random forest training process, when each decision tree node is split, not only the traditional Gini impurity is calculated, but also a feature time factor is introduced, and the calculation method is as follows:

[0109]

[0110] Wherein, L represents the adjusted impurity, g represents the traditional Gini impurity, represents the adjustment parameter, represents the average effective period of the current feature in the historical data, represents the aging factor;

[0111] Through parallel training of 500 decision trees, the average importance score of each feature on the OOB data is output, and the basic weight is obtained after Softmax normalization. The calculation method is as follows:

[0112]

[0113] Wherein, represents the basic weight of feature f, represents the average importance score, represents the sum of the exponential values of all features, represents the original importance score of feature f Take the natural exponential;

[0114] 0 is not an exponential operation here, 0 is a label and does not participate in operation;

[0115] To further enhance the adaptability of the model, a real-time feedback adjustment mechanism is introduced: when it is detected that a certain type of feature presents a significant change trend in the recent user behavior (within 30 days) (such as the click-through rate of commuting tolerance decreases by more than 15%), local retraining is automatically triggered, and the new data in the sliding time window is used to update the sub-tree of the affected feature, and the adjusted dynamic weight is generated. The calculation method is as follows:

[0116]

[0117] Wherein, represents the maximum weight of feature f at time , represents the basic weight of feature f, represents the adjustment coefficient, represents the change slope of feature f in the time window;

[0118] , 0 is not an exponential operation here, 0 is a label and does not participate in operation;

[0119] According to the weighted feature, the final matching degree score function is constructed, and the calculation method of the final matching degree score is as follows:

[0120] ​

[0121] wherein Q represents the final matching score, represents the value of feature f, represents the final weight of feature f at time represents the final weight of feature f at time represents the industry adjustment coefficient, represents the industry trend adjustment term, if the demand for talents in a certain industry increases recently, q takes a positive value, otherwise, q takes a negative value;

[0122] The features f include skill matching degree, salary expectation matching degree, commuting tolerance, and career development matching degree.

[0123] S5: push the matching results to the user end in descending order of scores, and update the weight model according to user behavior data iteratively;

[0124] In S5, after the matching degree is calculated, first, the results are processed according to the scene: for enterprise HR users, the top 50 candidates are displayed in descending order of MatchScore, and a “job competition analysis panel” is provided to intuitively show the deviation of the job from the market average level in terms of salary and skill requirements; for job seeker users, a progressive disclosure strategy is adopted, the core information (salary, commuting time, core skill matching degree) of the top 3 jobs with the highest matching degree is displayed on the first screen, and after sliding down, 20 recommended jobs ranked in descending order of MatchScore can be viewed, and a “matching advantage radar chart” is provided to visualize the scores in each dimension;

[0125] To build an effective feedback loop, a three-level weight update mechanism is designed: short-term update based on user real-time behavior fine-tuning, when it is detected that the click-through rate of a certain type of job deviates from the historical mean by ± 15% for 4 consecutive hours, the corresponding feature weight is immediately adjusted by , wherein represents the adjustment amount of the feature weight, represents the learning rate, represents the click-through rate of a certain type of job in the current time window (4 consecutive hours), represents the historical mean click-through rate; medium-term update is performed once a week, using the successful matching cases in the past 7 days (defined as the record of the enterprise viewing the resume and initiating an interview invitation, and the job seeker accepting) to retrain the random forest model, and optimize the cross-feature weights of commuting tolerance and career development; long-term update is performed every month, combined with the human resource market macro report, the weight distribution of each sub-item in the calculation formula of the industry trend factor λ is adjusted through expert rules;

[0126] All update operations are verified through A / B testing, the new weight model is first published in gray scale to 5% of the traffic, and when the conversion rate improvement is statistically significant, it is put into full-scale operation.

[0127] Please refer to Figure 2 In the embodiment, it is necessary to be specifically explained that the application provides an employment information matching system based on data analysis, comprising the following modules:

[0128] Multi-source data acquisition module: real-time acquisition of structured and unstructured employment data from enterprise recruitment database, job seeker resume database and industry public data platform;

[0129] Data preprocessing module: denoising, missing value filling, entity recognition and text vectorization processing are performed on the collected data to generate standardized job seeker portrait and enterprise post model;

[0130] Multi-dimensional feature extraction module: based on job requirements and job seeker ability, four types of core features are extracted, including skill matching degree, salary expectation fit degree, commuting tolerance and career development fit degree;

[0131] Dynamic weight distribution module: use historical matching success data to train random forest model to generate dynamic weight coefficients of each feature;

[0132] Intelligent recommendation and feedback optimization module: push the matching results to the user end in descending order of score, and update the weight model according to the user behavior data.

[0133] Finally: the above is only the preferred embodiment of the application and is not used to limit the application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A data analysis-based employment information matching method, characterized by, Comprise: S1: Real-time acquisition of structured and unstructured employment data from enterprise recruitment databases, job seeker resume databases, and industry public data platforms; S2: denoising, missing value filling, entity recognition and text vectorization processing are performed on the collected data to generate standardized job seeker portraits and enterprise post models; S3: based on the post demand and the job seeker's ability, four core features including skill matching degree, salary expectation fit degree, commuting tolerance and career development fit degree are extracted; in the feature extraction stage of S3, four core matching dimensions are constructed through quantitative analysis, and a composite index calculation system is used for each dimension; in the skill matching degree, a hierarchical evaluation model is used, first, the hard skill matching degree is calculated based on the knowledge graph, the calculation method is specifically: wherein, represents the hard skill matching degree, represents the weight of the i-th skill of the post, represents the skill mastery value of the job seeker, represents the proficiency level required by the post, represents the actual level of the job seeker; the soft skill part is calculated by the semantic similarity between the resume text and the post description, the calculation method is specifically: wherein, B represents the soft skill matching degree, represents the text vector of the post description, represents the vector of the resume text; the final skill matching degree is the weighted harmonic mean of hard and soft skills, the calculation method is specifically: wherein, represents the hard skill weight factor; a two-way tolerance model is constructed, considering the salary expectation fit degree of the enterprise budget interval and the job seeker's expectation interval, the calculation method is specifically: wherein D represents the salary expectation fit degree, represents the lower limit of the job seeker's salary expectation, represents the upper limit of the job seeker's salary expectation, represents the lower limit of the enterprise budget, i.e. the minimum salary the enterprise is willing to pay; represents the upper limit of the enterprise budget, i.e. the maximum salary the enterprise is willing to pay; represents the market salary range span; in the commuting tolerance, the time and space cost calculation based on geographic information system is used, the calculation method is specifically: wherein, F represents the commuting tolerance, represents the actual commuting time, represents the maximum tolerable commuting threshold, represents the urban traffic coefficient, represents the remote work weight, represents the proportion of remote office days; the career development fit degree is calculated by a three-dimensional evaluation model, the calculation method is specifically: wherein, G represents the career development fit degree, represents the path growth value, is expressed as a capability adaptation value, is expressed as an industry inertia value; the path growth value is calculated by the Pearson correlation coefficient of the historical job transition speed of the job seeker and the promotion path of the target job; the capability adaptation value is calculated by the ratio of the learning curve slope of the job seeker and the technical iteration rate of the job; the calculation method of the industry inertia value is as follows: wherein, is expressed as an industry inertia value, is expressed as the cumulative length of service in the target industry, is expressed as the job demand, is expressed as the distance between the job and the industry subdivision field of the candidate, is expressed as the maximum allowed subdivision field span threshold; S4: using historical matching success data to train a random forest model to generate dynamic weight coefficients of each feature; in the dynamic weight distribution stage, a random forest ensemble learning model with time correction is constructed based on historical matching success data, and dynamic weight coefficients are generated through multi-dimensional feature importance analysis; the specific process is as follows: first, extract positive samples from the historical successful matching case library, construct a negative sample set at the same time, and form a training data set with time label , is a four-dimensional feature vector, i.e. ; is expressed as a matching success label, is expressed as a matching time stamp; the model uses time decay weighted sampling technology, and the sampling weight of recent samples increases exponentially, and the calculation method is as follows: wherein, is expressed as the sampling weight of a certain sample, K is expressed as the decay coefficient, is expressed as the time stamp corresponding to the dth sample, is expressed as the minimum time value in all samples, is expressed as the maximum time value in all samples; in the random forest training process, for each decision tree node splitting, calculate the traditional Gini impurity, and introduce the feature time factor; through the parallel training of 500 decision trees, the average importance score of each feature on the OOB data is output, and the basic weight is obtained after Softmax normalization, and the calculation method is as follows: wherein, is expressed as the basic weight of feature f, is expressed as the average importance score, is expressed as the sum of the exponential values of all features, is expressed as the original importance score of feature f Take the natural index; when a certain feature is detected to present a significant change trend in the recent user behavior, automatically trigger local retraining, use the new data in the sliding time window to incrementally update the sub-tree of the affected feature, generate the adjusted dynamic weight, and the calculation method is as follows: Wherein, The maximum weight of the feature f at time is represented as, The base weight of the feature f is represented as, is represented as the adjustment coefficient, is represented as the change slope of the feature f in the time window; S5: push the matching result to the user end in descending order of score, and update the weight model according to the user behavior data.

2. The method for matching employment information based on data analysis according to claim 1, characterized in that: In the multi-source data collection of S1, structured data in the enterprise recruitment database is real-time crawled through a distributed crawler framework, and standardized fields in the job seeker resume database are accessed. For the unstructured data part, a multi-modal data processing pipeline is adopted: first, the semantic block processing of free text description in the enterprise recruitment brochure is performed, and the long text is decomposed into independent semantic units through a text segmentation algorithm based on attention mechanism; second, the context correlation analysis of the project experience text in the job seeker resume is performed, and the key technology stack, business field and achievement index in the project are extracted using named entity recognition technology, and a mapping relationship with the standardized skill tree is established.

3. The method of claim 1, wherein: In S2, a multi-level cleaning and conversion process is used to standardize the original employment data. First, data denoising is performed. For structured data, abnormal values are eliminated through quartile range detection and Z-score standardization. For text data, HTML tags, special characters and meaningless stop words are filtered using regular expressions, and duplicate published job information or resume content is identified and removed using a text cleaning model based on attention mechanism. In the missing value processing section, different strategies are used according to the data type: numerical fields are estimated using KNN nearest neighbor algorithm of the same industry and same position; categorical fields are inferred using Bayesian network based on other known fields; text type missing values are marked as "not specified" without filling to avoid introducing noise.

4. The method for matching employment information based on data analysis according to claim 1, characterized in that: In S5, after the matching degree calculation is completed, the results are first processed according to the scenario: for enterprise HR users, the top 50 candidates are displayed in descending order of MatchScore, and a "job competition analysis panel" is provided to show the deviation of the job in terms of salary and skill requirements from the market average level; for job seeker users, an incremental disclosure strategy is used, with the core information of the top 3 jobs with the highest matching degree displayed on the first screen, and the recommended jobs ranked in descending order of MatchScore displayed after scrolling down, with a "matching advantage radar chart" to visualize the scores in each dimension.

5. A data analysis based employment information matching system using the data analysis based employment information matching method according to any one of claims 1 to 4, characterized by: Multi-source data collection module: real-time acquisition of structured and unstructured employment data from enterprise recruitment databases, job seeker resume databases, and industry public data platforms; Data preprocessing module: denoising, missing value filling, entity recognition and text vectorization processing of the collected data to generate standardized job seeker portraits and enterprise job models; Multi-dimensional feature extraction module: based on job requirements and job seeker capabilities, four core features are extracted: skill matching degree, salary expectation fit degree, commuting tolerance, and career development fit degree; dynamic weight allocation module: using historical matching success data to train a random forest model to generate dynamic weight coefficients for each feature; intelligent recommendation and feedback optimization module: pushing the matching results to the user end in descending order of score, and updating the weight model iteratively according to user behavior data.

Images