Temporary worker flexible employment information matching method

Abstract

The invention relates to a temporary worker flexible employment information matching method, and the method comprises the steps: constructing a temporary worker skill portrait according to the historical employment information of a temporary worker; classifying the skills of the temporary workers according to the temporary worker skill portraits to obtain temporary worker subsets of different skill types; constructing a skill attenuation model for the temporary worker subset of each skill type; by analyzing the historical academic record and the training participation degree of the temporary worker, the learning ability coefficient of the temporary worker is calculated to serve as an input parameter of the attenuation model; according to the skill attenuation model, predicting the skill proficiency degree change condition of the temporary worker in a period of time in the future; dynamically acquiring latest skill use information of temporary workers, and updating input parameters; selecting a corresponding attenuation curve template to obtain a skill attenuation prediction result; combining various information to obtain a comprehensive matching degree; the prediction accuracy of the skill attenuation model is evaluated regularly; according to the invention, through the precise modeling skill attenuation process, the matching precision and timeliness of temporary workers and post demands are improved.

Description

Technical Field

[0001] The present disclosure relates to, in particular to, a method for matching temporary workers' flexible employment information. Background Art

[0002] In the flexible employment of temporary workers, the construction of skill decay models faces many technical challenges. Since temporary workers usually do not work for a long time in the same type of work and rarely have the opportunity to use a certain skill for a long time, their skills are likely to decay gradually due to long-term disuse. First, since the decay rates of different types of skills vary greatly, how to specifically quantify the decay rate of temporary workers' abilities is a major problem. The frequency and latest degree of use of programming language skills have a great impact on the decay rate, while customer service skills are more closely related to personal learning ability. Secondly, the time granularity of skill usage records also affects the accuracy of the decay curve. Too coarse granularity will lead to inaccurate estimation of the decay rate, and too fine granularity will make it difficult to obtain data. Furthermore, individuals' learning abilities vary greatly, and how to estimate this ability through historical data is also a technical challenge. In addition, temporary workers usually lack systematic on-the-job training and find it difficult to learn new knowledge and skills in a timely manner, which also accelerates the decay rate of their original skills. Finally, there may be correlations between skills, and the decay of certain skills may be affected by related skills. It is necessary to construct a skill correlation map to characterize this influence. In summary, to build a comprehensive and accurate skill attenuation model, we need to overcome a series of technical difficulties such as skill classification quantification, time granularity selection, personal learning ability estimation, skill association modeling, etc., and form a complete theoretical and algorithmic system to truly achieve the timeliness optimization of skill matching in flexible employment. Summary of the invention

[0003] In order to solve the problems existing in the above-mentioned prior art, the present invention aims to provide a method for matching temporary workers' flexible employment information.

[0004] The present disclosure discloses a method for matching temporary workers' flexible employment information, comprising:

[0005] S101, constructing a skill profile of temporary workers based on their historical employment information;

[0006] S102, classifying the skills of temporary workers according to their skill profiles to obtain subsets of temporary workers with different skill types;

[0007] S103, for each subset of temporary workers under each skill type, by analyzing their skill usage frequency and last usage time, a skill decay model is constructed to quantify the skill decay rate in different time periods;

[0008] S104, calculating the learning ability coefficient of the temporary worker by analyzing the historical learning performance and training participation, and using the learning ability coefficient and the skill usage frequency attribute as input parameters of the attenuation model;

[0009] S105, predicting the change in skill proficiency of temporary workers in the future based on the skill decay model;

[0010] S106, in the process of matching flexible employment information, dynamically obtain the latest skill usage information of temporary workers and update the input parameters in the skill decay model;

[0011] S107, constructing corresponding attenuation curve templates for different types of skills; selecting a corresponding attenuation curve template according to the skill type of the temporary worker, and adjusting the learning ability coefficient as a parameter to obtain a skill attenuation prediction result;

[0012] S108, in the process of skill matching, the skill proficiency, skill decay rate, and employment willingness of temporary workers are combined to obtain a comprehensive matching degree;

[0013] S109, regularly evaluate the prediction accuracy of the skill decay model, optimize the model parameters by collecting actual employment feedback information from temporary workers, and refine and adjust the model based on the differences in decay patterns of different skill types.

[0014] Preferably, the step of constructing a skill profile of a temporary worker based on the temporary worker's historical employment information includes:

[0015] Obtain the skill name involved in each job from the historical employment records of temporary workers. The skill name is obtained by matching the job description keywords in the preset skill vocabulary.

[0016] Calculate the last time each skill was used based on historical employment records, where the last time was determined by traversing the work history that included the skill and selecting the latest end date;

[0017] For each skill, the cumulative number of months of use in historical employment records is counted, and the cumulative number of months of use is divided by the total number of working months to obtain the skill use frequency value;

[0018] Combine the skill name, last used time and frequency of use into structured data to form a temporary worker's personal skill information table;

[0019] A multidimensional vector representation method is used to construct a skill profile of temporary workers. The multidimensional vector includes the skill name, last use time and frequency of use.

[0020] Preferably, the skills of temporary workers are classified according to their skill profiles to obtain subsets of temporary workers with different skill types, including:

[0021] Obtain all skill items in the skill profile of temporary workers, including programming languages ​​and customer service skills, and use the Word2Vec word embedding tool to convert the skill items into high-dimensional vector representations to obtain a skill vector set;

[0022] Perform dimensionality reduction processing based on the skill vector set, extract key features through principal component analysis tools, calculate eigenvalues ​​and eigenvectors, sort by eigenvalue size, accumulate eigenvalue contribution rates to a preset threshold, and retain the corresponding eigenvectors as dimensionality reduction results;

[0023] For the dimensionality reduction results, calculate the intra-cluster sum of squares corresponding to each K value within the preset K value range, draw a curve graph, and select the K value corresponding to the inflection point of the curve as the final number of clusters;

[0024] Use the final number of clusters, perform K-means clustering on the dimensionality reduction results, randomly select the initial center point, calculate the Euclidean distance from the sample to each center point, assign the sample to the nearest center point, recalculate the center point of each cluster, and repeat this process until the center point position no longer changes or the maximum number of iterations is reached;

[0025] According to the clustering results, a skill type label is assigned to each cluster. A skill type vector library is built based on the pre-established industry standard occupational classification. The cosine similarity between the cluster center and each skill type vector is calculated, and the top three types with the highest similarity are selected as the multi-labels of the cluster.

[0026] Preferably, for each skill type subset of temporary workers, by analyzing the skill usage frequency and the last usage time, a skill decay model is constructed to quantify the skill decay rate in different time periods, including:

[0027] Obtain the skill usage records of the temporary worker subset, and extract the usage frequency and last usage time of each skill based on the skill usage records;

[0028] The missing data in the frequency of use and the last time of use were processed by the median filling method, and the outliers in the frequency of use and the last time of use were eliminated by the 3σ principle to obtain a standardized skill use data set.

[0029] Based on the standardized skill usage data set, the last usage time is grouped by month using a fixed 30-day time window.

[0030] Calculate the average frequency of skill usage in each time window to form time series data;

[0031] Determine the initial form of the exponential decay function, the exponential decay function is f(t)=ae^(-bt), where a represents the initial frequency, b represents the decay coefficient, and t represents time;

[0032] The exponential decay function was fitted using the least squares method, with time as the independent variable and average frequency of use as the dependent variable;

[0033] The objective function is set to minimize the square difference between the predicted value and the actual value, and the optimal values ​​of the attenuation coefficient b and the initial frequency a are solved to construct a skill attenuation model.

[0034] Apply skill decay models to different skill types and calculate skill decay values ​​in each time period;

[0035] The decay rate is quantified by calculating the slope of the decay curve at different time points, and a decay curve is drawn to show the trend of skill changes over time.

[0036] Preferably, the above-mentioned calculating the learning ability coefficient of the temporary worker by analyzing the historical learning performance and training participation of the temporary worker, and using the learning ability coefficient and the skill usage frequency attribute as input parameters of the attenuation model includes:

[0037] Obtain learning performance and training participation records from the temporary workers' historical database, identify outliers, and use the mean to fill in missing data based on outlier judgments to obtain a standardized learning ability assessment data set;

[0038] Based on the standardized learning ability assessment data set, the hierarchical analysis method is used to determine the weight of the weighted average of multiple factors, and the weighted calculation is performed on historical learning performance, training participation, learning speed and knowledge retention rate to determine the learning ability coefficient of each temporary worker;

[0039] Obtain the learning ability coefficient and skill usage frequency of temporary workers, and use the correlation analysis tool to calculate the Pearson correlation coefficient and Spearman rank correlation coefficient between the two. If the correlation coefficient is greater than the preset threshold, it is judged whether it is reasonable to introduce the learning ability factor into the attenuation model.

[0040] If it is judged that it is reasonable to introduce the learning ability factor into the attenuation model, the mathematical expression of the original attenuation model is modified, the learning ability coefficient is used as a new input variable, combined with the original model by multiplication, and the model parameters are refitted through multiple linear regression;

[0041] The modified attenuation model is subjected to ten-fold cross validation, and the root mean square error of the model before and after modification is calculated. If the root mean square error of the modified model is smaller than that of the model before modification, it is determined that the modified attenuation model has higher prediction accuracy.

[0042] Preferably, the prediction of the change in skill proficiency of temporary workers in the future period of time based on the skill decay model includes:

[0043] Obtain the initial skill proficiency and attenuation coefficient, and substitute the initial skill proficiency and attenuation coefficient into the skill attenuation model formula for temporary workers;

[0044] According to the model formula, the Runge-Kutta method is used for numerical integration to obtain the predicted value of skill proficiency for each month;

[0045] The skill attenuation rate is calculated using the sliding window method through the predicted value of skill proficiency;

[0046] Obtain the proficiency prediction values ​​of two adjacent months from the sliding window, calculate the difference between the proficiency prediction values ​​of two adjacent months, and obtain the decay rate value of each month;

[0047] The K-means clustering algorithm is used to group the decay speeds of all skills. If the optimal number of clusters is determined by the silhouette coefficient method, the skills with faster decay speeds are clustered into one category.

[0048] For skill clusters with faster decay rates, the quartile method is used to select the upper quartile as the decay threshold. If the skill decay rate exceeds the decay threshold, the skill is included in the fast decay skill set.

[0049] Get the decay speed from the fast decay skill set and use the decay speed as the matching weight factor;

[0050] Obtain skill proficiency and timeliness weights, combine the matching weight factor, skill proficiency and timeliness weights, and use the linear weighted method to calculate the comprehensive matching score.

[0051] Preferably, in the flexible employment information matching process, dynamically obtaining the latest skill usage information of temporary workers and updating the input parameters in the skill decay model include:

[0052] Obtaining temporary workers’ skill usage information, which is extracted from work logs, project reports, and online collaboration platforms;

[0053] Update the skill's last usage time and usage frequency data based on skill usage information;

[0054] The individual learning ability coefficient is calculated using the online gradient descent method. The individual learning ability coefficient is obtained based on the training courses, test scores and skill certifications that the temporary workers have participated in;

[0055] Use the message queue system to transmit skill usage information and personal learning ability coefficient to the skill decay model;

[0056] If new skill usage information and personal learning ability coefficient are received, the skill decay model parameter update mechanism is triggered;

[0057] Use the online random forest algorithm to optimize the skill decay model, and dynamically adjust the model weight and decay coefficient during the optimization process;

[0058] Apply the skill decay model to the flexible employment information matching process, and the application process includes dynamically adjusting the matching weights;

[0059] The update and synchronization of multi-source data, including skill usage information, personal learning ability coefficients, and skill decay model parameters, are coordinated through data version control tools.

[0060] Preferably, for different types of skills, corresponding attenuation curve templates are constructed; according to the skill type of the temporary worker, the corresponding attenuation curve template is selected, and the learning ability coefficient is adjusted as a parameter to obtain the skill attenuation prediction result, including:

[0061] The historical data of different types of skills are obtained, and the attenuation curve template is constructed by the least squares method. The attenuation curve template is quantified into an exponential attenuation function f(t) = ae^(-bt); where a represents the initial proficiency, b represents the attenuation rate, and t represents time. According to the attenuation curve template, the K-means clustering algorithm is used to classify the skills of temporary workers; the classification process includes classifying the skills into predefined skill types based on skill descriptions and application scenarios; determining the optimal number of clusters by the silhouette coefficient method; and determining the attenuation curve template corresponding to each skill. For the attenuation curve template, the learning ability coefficient L is introduced for linear adjustment; the linear adjustment includes designing the parameter adjustment formula b' = b*(2-L); and dynamically adjusting the slope of the attenuation curve according to different learning ability levels. Using the ARIMA model, combined with the attenuation curve template and the adjusted parameters, the attenuation prediction of the specific skills of temporary workers is carried out; the attenuation prediction includes generating a skill proficiency change curve in the future. The root mean square error (RMSE) is used to evaluate the accuracy of the attenuation prediction results. The evaluation process includes comparing the deviation between the predicted value and the actual observed value. If the deviation exceeds the preset threshold, the model parameters are dynamically optimized. If the deviation does not exceed the preset threshold, the attenuation prediction result is determined to be a valid prediction result.

[0062] Preferably, the skill proficiency, skill decay rate, and employment willingness of the temporary worker are combined in the skill matching process to obtain a comprehensive matching degree, which includes:

[0063] The skill proficiency scores, skill decay rate prediction values ​​and employment intention survey results of temporary workers in the temporary worker database are obtained, and the employment intention survey results are converted into scores using the Likert scale; a judgment matrix is ​​constructed based on the skill proficiency scores, skill decay rate prediction values ​​and employment intention scores, and the eigenvectors are calculated to obtain weights for consistency test; the skill proficiency scores, skill decay rate prediction values ​​and employment intention scores are normalized using the weights to obtain the min-max normalization process to obtain the standardized scores; the comprehensive matching scores of temporary workers are calculated based on the standardized scores and corresponding weights; based on the comprehensive matching scores, the cosine similarity matching algorithm is used to match temporary workers with the needs of employers, and the temporary jobs with the highest similarity are selected as the recommendation results.

[0064] Preferably, the prediction accuracy of the skill decay model is regularly evaluated, and the model parameters are optimized by collecting actual employment feedback information of temporary workers; and the model is refined and adjusted according to the differences in decay rules of different skill types, including:

[0065] The performance evaluation data of temporary workers are obtained from employers, and the evaluation data include work quality, efficiency and skill application. The standard score is calculated by weighted average method based on the evaluation data. The ARIMA model is used to perform time series analysis on the standard score to obtain the predicted value. The prediction deviation and root mean square error between the predicted value and the standard score are calculated. The parameters of the skill decay model are optimized by stochastic gradient descent algorithm, and the parameters include decay coefficient and initial skill level. The K-means clustering algorithm is used to group skill types to obtain multiple skill clusters. An independent decay model is established for each skill cluster. The correlation matrix is ​​introduced to calculate the influence coefficient between skill clusters. The parameters of the skill decay model are adjusted according to the influence coefficient.

[0066] The method for matching temporary workers' flexible employment information disclosed in the present disclosure has the advantages that: the method disclosed in the present disclosure first constructs a skill profile based on the historical employment information of temporary workers, including the skill name, the last time of use, and the frequency of use. Then, different types of skills such as programming languages ​​and customer service skills are clustered and classified. For each skill type, a skill decay model is constructed, and an exponential decay function is used to quantify the decay rate of the skill over time. The model also introduces the factor of personal learning ability, and calculates the learning ability coefficient by analyzing historical learning performance and training participation. Based on this model, the future skill proficiency changes of temporary workers are predicted, and the prediction results are used for skill matching. During the matching process, the model input parameters are dynamically updated, including the latest skill usage information. Corresponding decay curve templates are established for different skill types, and adjusted according to personal learning ability. Finally, skill proficiency, decay rate and employment willingness are comprehensively considered, and the matching degree is obtained by weighted calculation, and temporary workers with high matching degree are recommended first.

[0067] The present invention improves the matching accuracy and timeliness of temporary workers and job requirements by accurately modeling the skill decay process. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flow chart of a method for matching temporary workers' flexible employment information described in the present invention. DETAILED DESCRIPTION

[0069] like Figure 1 As shown, a method for matching temporary workers' flexible employment information described in the present disclosure includes:

[0070] Step S101, based on the temporary workers' historical employment information, obtain the skill list of each temporary worker, including the skill name, the last time the skill was used, and the skill usage frequency attributes, and construct a skill profile of the temporary worker.

[0071] The names of the skills involved in each job are obtained from the historical employment records of temporary workers. The skill names are obtained by matching the job description keywords in the preset skill vocabulary. The last use time of each skill is calculated based on the historical employment records, where the last use time is determined by traversing the work experience that includes the skill and selecting the latest end date. For each skill, the cumulative number of months of use in the historical employment records is counted, and the cumulative number of months of use is divided by the total number of working months to obtain the skill use frequency value. The skill name, last use time and use frequency are combined into structured data to form a temporary worker personal skill information table. The multidimensional vector representation method is used to construct a temporary worker skill profile, and the multidimensional vector includes the skill name, last use time and use frequency dimensions.

[0072] Specifically, the skill names involved in each job are extracted based on the historical employment records of temporary workers. Keywords are identified from the job descriptions through word segmentation tools and matched with the preset skill word library to obtain a standardized skill name list. The preset skill word library is initially constructed through the collection of industry standard occupational classifications and enterprise job descriptions, and the word library content is regularly updated based on new employment records to maintain the timeliness and comprehensiveness of the skill word library. The last use time is calculated for each skill, the employment records are sorted by time, and all work experiences containing the skill are traversed, and the latest end date is selected as the last use time of the skill. The frequency counter is used to count the cumulative number of months of use of each skill in the historical employment records, divided by the total number of working months of the individual to obtain the skill use frequency value, and the frequency level is divided into three levels: high, medium and low. The name, last use time, use frequency and other information of each skill are combined into structured data to form a temporary worker personal skill information table. The skill profile of temporary workers is constructed by using multidimensional vector representation. The skill name, last use time and use frequency are used as different dimensions of the vector. The K-means clustering algorithm is used to classify similar skills. The average frequency of each skill category and the decay factor based on the last use time are calculated. The skill proficiency score of each category is comprehensively obtained, and finally a complete skill profile of temporary workers is formed. When extracting skill names from the historical employment records of temporary workers, the word segmentation tool can use the jieba library to perform Chinese word segmentation and extract keywords such as "Java development" and "data analysis". The preset skill vocabulary initially contains 1,000 common professional skills. New job descriptions are obtained from mainstream recruitment websites through crawlers every month, and new skills are extracted and updated. For a position such as "Java development engineer", multiple skills such as "Java", "Spr ing framework", and "MySQL database" may be extracted. When calculating the last time a skill was used, assume that a temporary worker has three job experiences: Java development from January 2020 to June 2021, Python development from July 2021 to December 2022, and full-stack development from January 2023 to the present. For the "Java" skill, the last time of use is June 2021; for the "Python" skill, the last time of use is December 2022; for the "full-stack development" skill, the last time of use is the current date. In the skill use frequency calculation, the cumulative working time is 48 months. If the "Java" skill is used for 18 months, 0 months, and 6 months in three jobs respectively, its use frequency is (18+0+6) / 48=0.5. The frequency level classification can be set as 0.7 or above for high frequency, 0.3 to 0.7 for medium frequency, and below 0.3 for low frequency. The structured data format can be {"Skill Name":"Java","Last Use Time":"2023-06-30","Usage Frequency":0.5,"Frequency Level":"Medium Frequency"}.When constructing a skill profile, the K-means algorithm is used to cluster skills. Assuming the k value is set to 5, skill categories such as "backend development", "frontend development", "data analysis", "project management", and "design" may be obtained. When calculating skill proficiency, a time decay factor can be introduced. For example, if the half-life is set to 2 years, the weight of the skill last used 2 years ago is halved. The final skill profile may be in the form of {"backend development": 0.8,"frontend development": 0.6,"data analysis": 0.4,"project management": 0.2,"design": 0.1}, indicating that the temporary worker is most proficient in the field of backend development, followed by front-end development, and relatively lacking in design.

[0073] Step S102, based on the skill profile of the temporary worker, for different types of skills, including programming languages ​​and customer service skills, a clustering algorithm is used to classify the skills of the temporary worker to obtain subsets of temporary workers with different skill types.

[0074] All skill items including programming languages ​​and customer service skills in the skill profile of temporary workers are obtained, and the Word2Vec word embedding tool is used to convert the skill items into high-dimensional vector representations to obtain a skill vector set; dimensionality reduction processing is performed based on the skill vector set, and key features are extracted through the principal component analysis tool, eigenvalues ​​and eigenvectors are calculated, sorted by eigenvalue size, eigenvalue contribution rates are accumulated to the preset threshold, and the corresponding eigenvectors are retained as dimensionality reduction results; for the dimensionality reduction results, the intra-cluster sum of squares corresponding to each K value is calculated within the preset K value range, a curve graph is drawn, and the K value corresponding to the inflection point of the curve is selected as the final number of clusters; using the final number of clusters, K-means clustering is performed on the dimensionality reduction results, the initial center point is randomly selected, the Euclidean distance from the sample to each center point is calculated, the sample is assigned to the nearest center point, the center point of each cluster is recalculated, and this process is repeated until the center point position no longer changes or the maximum number of iterations is reached; skill type labels are assigned to each cluster based on the clustering results, a skill type vector library is constructed based on the pre-established industry standard occupational classification, the cosine similarity between the cluster center and each skill type vector is calculated, and the top three types with the highest similarity are selected as the multi-labels of the cluster.

[0075] Specifically, all skills including programming languages ​​and customer service skills are extracted from the skill profiles of temporary workers. The Word2Vec word embedding tool is used to convert each skill into a high-dimensional vector representation. The Word2Vec model is trained through a large-scale skill corpus, and the window size is set to 5 and the vector dimension is set to 100 to obtain the skill vector set of each temporary worker. The skill vector set is reduced in dimension, and the key features are extracted through the principal component analysis tool. The eigenvalues ​​and eigenvectors are calculated, sorted by eigenvalue size, and the eigenvalue contribution rate is accumulated until the preset threshold of 95% is reached. The corresponding eigenvector is retained as the dimensionality reduction result. The optimal number of clusters K is determined using the elbow rule. Within the K value range of 1 to 20, the intra-cluster sum of squares corresponding to each K value is calculated, and a curve is drawn. The K value corresponding to the inflection point of the curve is selected as the final number of clusters. Based on the determined K value, K-means clustering is performed on the reduced feature vector. K initial center points are randomly selected, the Euclidean distance from each sample to each center point is calculated, the sample is assigned to the nearest center point, the center point of each cluster is recalculated, and this process is repeated until the center point position no longer changes or the maximum number of iterations is 100. According to the clustering results, skill type labels are assigned to each cluster. A predefined skill type vector library is built based on the industry standard occupational classification. The cosine similarity between the cluster center and each skill type vector is calculated. The top three types with the highest similarity are selected as the multi-label of the cluster, thereby obtaining a subset of temporary workers under different skill types. In the processing of temporary worker skill profiles, skill items such as "Java programming" and "customer communication" are first extracted from the skill profiles of 1,000 temporary workers, totaling 5,000 different skills. Using the Word2Vec tool, the model is trained by 10 million job description corpora to convert each skill into a 100-dimensional vector. For example, "Java programming" may be represented as [0.2, -0.1, 0.5, ..., 0.3]. For each temporary worker, all their skill vectors are merged to obtain a 500,000-dimensional sparse vector. Subsequently, the principal component analysis tool is used to reduce the dimension of these 1,000 500,000-dimensional vectors. After calculating the eigenvalues, it is found that the cumulative contribution rate of the first 50 principal components reaches 95.6%, so these 50 principal components are retained as the dimension reduction result. Next, the elbow rule is applied to determine the optimal number of clusters. The intra-cluster sum of squares is calculated in the K value range of 1 to 20, and the curve is plotted. It is found that there is an obvious inflection point when K=8, so 8 is selected as the final number of clusters. When performing K-means clustering, 8 initial center points are randomly selected, such as [1.2,-0.5,0.8,...,-0.3], etc. After 73 iterations, the center point position is stable, and 8 temporary worker skill clusters are obtained. Finally, 10 predefined skill type vectors are constructed based on industry standards, such as "software development" and "customer service".Calculate the cosine similarity between the 8 cluster centers and the 10 type vectors. For example, if the similarity between a cluster center and "software development" is 0.85, "customer service" is 0.72, and "data analysis" is 0.68, then assign these three labels to the cluster. Finally, 8 temporary worker skill subsets with multiple labels are obtained, such as the "software development + customer service + data analysis group" contains 150 temporary workers.

[0076] Step S103: for the subset of temporary workers under each skill type, a skill decay model is constructed by analyzing the skill usage frequency and the last usage time to quantify the skill decay rate in different time periods. The decay model is calculated using an exponential decay function.

[0077] The skill usage records of the subset of temporary workers were obtained, and the usage frequency and last usage time of each skill were extracted based on the skill usage records. The median filling method was used to handle the missing data in the usage frequency and the last usage time, and the outliers in the usage frequency and the last usage time were eliminated by the 3σ principle to obtain a standardized skill usage data set. According to the standardized skill usage data set, the last usage time was grouped by month using a fixed 30-day time window division method. The average usage frequency of the skill in each time window was calculated to form time series data. The initial form of the exponential decay function was determined, and the exponential decay function was f(t) = ae^(-bt), where a represents the initial frequency, b represents the decay coefficient, and t represents time. The exponential decay function was fitted using the least squares method, with time as the independent variable and the average usage frequency as the dependent variable. The objective function was set to minimize the square difference between the predicted value and the actual value, and the optimal values ​​of the decay coefficient b and the initial frequency a were solved to construct a skill decay model. The skill decay model was applied to different skill types to calculate the skill decay value in each time period. The decay rate was quantified by calculating the slope of the decay curve at different time points, and a decay curve was drawn to represent the trend of skill changes over time.

[0078] Specifically, based on the skill usage records of the temporary workers subset, the usage frequency and last usage time of each skill are extracted. The missing data are processed by the median filling method, and the 3σ principle is used to eliminate outliers to obtain a standardized skill usage data set. The last usage time of the skill is grouped by month using a fixed 30-day time window division method, and the average usage frequency of the skill in each time window is calculated to form time series data. Based on domain knowledge and pre-experimental results, the initial form of the exponential decay function is determined to be f(t) = ae^(-bt), where a represents the initial frequency, b represents the decay coefficient, and t represents time. The exponential decay function is fitted using the least squares method, with time as the independent variable and the average usage frequency as the dependent variable. The objective function is set to minimize the square difference between the predicted value and the actual value. By taking the derivative and setting the derivative to zero, the optimal value of the decay coefficient b and the initial frequency a is solved to construct a skill decay model. The decay model is applied to different skill types, and the skill decay value in each time period is calculated. The decay rate is quantified by calculating the slope of the decay curve at different time points, and the decay curve is drawn to represent the trend of skill changes over time. The decay model of different skill types was validated by cross-validation method, and the root mean square error between the predicted value and the actual value was calculated to evaluate the accuracy and reliability of the model. In the process of constructing the skill decay model, data was first extracted from the skill usage records of 10,000 temporary workers, and each temporary worker had an average of 20 skills. The median filling method was used to handle about 5% of the missing data, such as filling the missing usage frequency of the "Java programming" skill with the median value of the skill, 3 times / month. Outliers were eliminated by the 3σ principle, such as records with a usage frequency greater than the mean plus 3 times the standard deviation (15 times / month) were considered abnormal and deleted, and finally 198,000 valid data were obtained. A time window of 30 days was used, and the data of the last 730 days was divided into 24 windows. The average usage frequency of the skill in each window was calculated, such as the average frequency of "Java programming" in the first window was 4.2 times / month, the second window was 3.8 times / month, and so on. Based on the preliminary experimental results, the initial form of the exponential decay function was determined to be f(t)=ae^(-bt). The least squares fitting function is used, with time t (unit: month) as the independent variable and the average frequency of use f(t) as the dependent variable. The initial frequency of the skill "Java programming" is a = 5.1 times / month, and the attenuation coefficient is b = 0.03 / month. The attenuation model is applied to 10 different skill types, such as the average attenuation coefficient of the "software development" type is 0.025 / month, and the "customer service" type is 0.015 / month. By calculating the slope of the attenuation curve at t = 6 months, the attenuation rate of the "software development" type is -0.12 times / month2, and the "customer service" type is -0.07 times / month2.The 5-fold cross-validation method was used to evaluate the model and calculate the root mean square error. The average error for the "software development" type was 0.3 times / month, and the average error for the "customer service" type was 0.2 times / month, verifying the accuracy and reliability of the model.

[0079] Step S104, introduce the individual learning ability factor into the skill decay model. For temporary workers with strong learning ability, their learning ability coefficient is calculated by analyzing the temporary workers' historical learning performance and training participation, and the learning ability coefficient and skill usage frequency attribute are used as input parameters of the decay model.

[0080] The learning performance and training participation records are obtained from the temporary workers' historical database, outliers are identified, and the mean is used to fill the missing data according to the outliers to obtain a standardized learning ability assessment data set. According to the standardized learning ability assessment data set, the hierarchical analysis method is used to determine the weight of the multi-factor weighted average, and the weighted calculation is performed for historical learning performance, training participation, learning speed and knowledge retention rate to determine the learning ability coefficient of each temporary worker. The learning ability coefficient and skill usage frequency of temporary workers are obtained, and the Pearson correlation coefficient and Spearman rank correlation coefficient between the two are calculated by the correlation analysis tool. If the correlation coefficient is greater than the preset threshold, the rationality of introducing the learning ability factor into the attenuation model is judged. If it is judged that the learning ability factor is introduced into the attenuation model, the mathematical expression of the original attenuation model is modified, and the learning ability coefficient is used as a new input variable. It is combined with the original model by multiplication, and the model parameters are refitted by multivariate linear regression. The modified attenuation model is cross-validated by ten-fold, and the root mean square error of the model before and after the modification is calculated. If the root mean square error of the modified model is less than that of the model before the modification, it is determined that the modified attenuation model has higher prediction accuracy.

[0081] Specifically, the historical learning performance and training participation records of each temporary worker were extracted from the temporary worker history database. The Z-score method was used to identify and process outliers, and the mean was used to fill in missing data to obtain a standardized learning ability assessment data set. The hierarchical analysis method was used to determine the weight distribution of multi-factor weighted average, and the historical learning performance, training participation, learning speed and knowledge retention rate were used as evaluation factors. The learning speed was calculated by the number of knowledge points mastered per unit time, and the knowledge retention rate was calculated by the score change rate of regular tests. The learning ability coefficient of each temporary worker was obtained comprehensively. The Pearson correlation coefficient and Spearman rank correlation coefficient between the learning ability coefficient and the skill use frequency were calculated using the correlation analysis tool to determine the degree of correlation between the two and determine the rationality of introducing the learning ability factor into the attenuation model. According to the results of the correlation analysis, the mathematical expression of the original attenuation model was modified, and the learning ability coefficient was used as a new input variable. It was combined with the original model by multiplication, and the model parameters were refitted by multivariate linear regression to obtain an improved skill attenuation model that takes into account individual learning ability. The modified attenuation model was cross-validated by ten-fold, and the root mean square error of the model before and after the modification was compared to evaluate the prediction accuracy of the new model. In the process of introducing the individual learning ability factor into the skill decay model, the learning records were first extracted from the historical database of 10,000 temporary workers. The Z-score method was used to identify outliers, and data exceeding the mean plus or minus 3 standard deviations were considered abnormal and deleted, accounting for about 2% of the total data. For missing data, such as blank items in historical learning performance, the average score of the item was 80 points to fill in. The hierarchical analysis method was used to determine the weights, and the weights of historical learning performance, training participation, learning speed, and knowledge retention were obtained. The weight was 0.3, 0.2, 0.25, and 0.25. The learning speed was calculated by the number of knowledge points mastered per hour, with an average of 5 points / hour; the knowledge retention rate was calculated by the score change rate of the retest after 3 months, with an average of 85%. The comprehensive calculation resulted in the learning ability coefficient of each temporary worker, which was distributed between 0.6 and 1.4. The Pearson correlation coefficient between the learning ability coefficient and the skill use frequency calculated by SPSS tools was 0.72, and the Spearman rank correlation coefficient was 0.68, indicating that the two had a strong positive correlation. The learning ability coefficient L is multiplied by the original attenuation model f(t) = ae^(-bt) to obtain a new model f'(t) = aLe^(-bt). The parameters are refitted using multiple linear regression to obtain a new model of Java programming skills f'(t) = 5.1Le^(-0.028t). Through ten-fold cross validation, the root mean square error of the new model is 0.25 times / month, which is 16.7% lower than the original model's 0.3 times / month, proving that the introduction of the learning ability factor improves the prediction accuracy of the model.

[0082] Step S105, predicting the skill proficiency changes of temporary workers in the future based on the skill decay model, and using the prediction results as the timeliness basis for skill matching, with priority given to skills that decay quickly.

[0083] Obtain the initial skill proficiency and attenuation coefficient, and substitute the initial skill proficiency and attenuation coefficient into the skill attenuation model formula for temporary workers; according to the model formula, use the Runge-Kutta method for numerical integration to obtain the predicted skill proficiency value for each month; use the sliding window method to calculate the skill attenuation rate based on the predicted skill proficiency value; obtain the predicted proficiency values ​​of two adjacent months from the sliding window, calculate the difference between the predicted proficiency values ​​of two adjacent months, and obtain the attenuation rate value for each month; use the K-means clustering algorithm to group the attenuation rates of all skills, and if the optimal number of clusters is determined by the silhouette coefficient method, the skills with faster attenuation rates are clustered into one category; for the clustering of skills with faster attenuation rates, use the quartile method to select the upper quartile as the attenuation threshold, and if the skill attenuation rate exceeds the attenuation threshold, the skill is included in the fast-decaying skill set; obtain the attenuation rate from the fast-decaying skill set, and use the attenuation rate as the matching weight factor; obtain the skill proficiency and timeliness weight, combine the matching weight factor, skill proficiency and timeliness weight, and use the linear weighted method to calculate the comprehensive matching score.

[0084] Specifically, according to the skill decay model of temporary workers, the prediction period is set to the next 6 months, the initial skill proficiency and decay coefficient are substituted into the model formula, and the Runge-Kutta method is used for numerical integration to solve the predicted value of skill proficiency for each month. The sliding window method is used to calculate the skill decay rate. With a window size of 1 month, the difference between the predicted proficiency values ​​of two adjacent months is calculated to obtain the decay rate value of each month. The window size is adjusted according to the characteristics of the skill type. The decay rates of all skills are grouped using the K-means clustering algorithm. The optimal number of clusters is determined by the silhouette coefficient method. Skills with faster decay rates are clustered into one category. The quartile method is used to select the upper quartile as the decay threshold, and the skill set with a decay rate exceeding the threshold is screened out. For the screened fast decay skill set, a priority matching rule is designed, and the decay rate is used as a matching weight factor. It is combined with the skill proficiency and timeliness weights, and the linear weighted method is used to calculate the comprehensive matching score. The timeliness weight is determined according to the proficiency change range within the prediction period, and the final skill matching priority is determined according to the score. The decay characteristics of different skill types are analyzed, and the weight coefficients of each type of skill in the comprehensive matching score are adjusted to ensure that the differences in skill types are taken into account during the matching process. In the process of predicting the changes in the proficiency of temporary workers' skills, the initial proficiency of Java programming skills is set to 85 points (out of 100 points), and the decay coefficient is 0.05 / month. The fourth-order Runge-Kutta method is used for numerical integration for 6 months, and the monthly proficiency prediction values ​​are 85, 80.75, 76.71, 72.88, 69.23, and 65.77 points respectively. The sliding window method with a window size of 1 month is used to calculate the decay rate, and 5 decay rate values ​​are obtained: 4.25, 4.04, 3.83, 3.65, and 3.46 points / month. K-means clustering is performed on the decay rates of various skills of 1,000 temporary workers. By calculating the silhouette coefficients of different K values ​​2-10, the optimal cluster number K=3 is determined, and the skills are divided into three categories: fast decay, medium decay, and slow decay. The upper quartile of the decay rate calculated using the quartile method is 3.8 points / month as the threshold to filter out the fast decay skill set. In the priority matching rule, the decay rate weight is set to 0.4, the skill proficiency weight is set to 0.4, and the timeliness weight is set to 0.2. The timeliness weight is determined based on the decline in proficiency within 6 months (22.6%). For Java programming skills, the calculated comprehensive matching score is 0.4×(4.25 / 5)+0.4×(85 / 100)+0.2×0.226=0.7252. Considering different skill types, such as increasing the weight of programming skills by 10% and reducing the weight of customer service skills by 5%, the final adjusted matching score of Java programming skills is 0.7252×1.1=0.79772. Based on this score, the matching priority of Java programming skills among fast decay skills is determined.

[0085] Step S106, during the flexible employment information matching process, dynamically obtain the latest skill usage information of temporary workers and update the input parameters in the skill decay model. The input parameters also include the last time the skill was used, the frequency of use, and the personal learning ability coefficient.

[0086] Obtain temporary workers' skill usage information, which is extracted from work logs, project reports and online collaboration platforms; update the skill last usage time and usage frequency data based on the skill usage information; use the online gradient descent method to calculate the personal learning ability coefficient, which is obtained based on the training courses, test scores and skill certification status of the temporary workers; use the message queue system to transmit the skill usage information and the personal learning ability coefficient to the skill decay model; if new skill usage information and personal learning ability coefficient are received, the skill decay model parameter update mechanism is triggered; use the online random forest algorithm to optimize the skill decay model, and dynamically adjust the model weights and decay coefficients during the optimization process; apply the skill decay model to the flexible employment information matching process, and the application process includes dynamically adjusting the matching weights; coordinate the update and synchronization of multi-source data through data version control tools, and the multi-source data includes skill usage information, personal learning ability coefficients and skill decay model parameters.

[0087] Specifically, the named entity recognition tool is used to monitor the skill usage of temporary workers, skill application records are extracted from work logs, project reports and online collaboration platforms, and the last time and frequency of skill usage data are updated. At the same time, data desensitization technology is used to protect the personal privacy of temporary workers. The online gradient descent method is used to dynamically adjust the personal learning ability coefficient, and the learning ability evaluation index is updated in real time according to the training courses, test scores and skill certifications participated by temporary workers. The message queue system is used for data transmission, and the newly acquired skill usage information and the updated personal learning ability coefficient are transmitted to the skill decay model to trigger the real-time update mechanism of the model parameters. Based on the updated input parameters, the online random forest algorithm is used to incrementally optimize the skill decay model, dynamically adjust the model weights and decay coefficients, and improve the model's sensitivity to skill changes and prediction accuracy. The updated skill decay model is applied to the flexible employment information matching process, and the matching weights are dynamically adjusted to ensure the consistency and reliability of data from different sources. The data version control tool is used to coordinate the update and synchronization of multi-source data. In the flexible employment information matching process, the BERT-based named entity recognition tool is used to monitor the skill usage of temporary workers in real time. Taking Java programming skills as an example, relevant descriptions such as "Completed Java backend interface development" are identified from the work log, and the last use time of the skill is updated to June 15, 2023, and the frequency of use is increased by 1 time / week. At the same time, the temporary worker ID is replaced with a hash value to protect personal privacy. Using the online gradient descent method, based on the temporary worker's recent participation in the "Advanced Java Programming" training course, the score is 90 points out of 100 points, and the Java skill learning ability coefficient is updated from 0.8 to 0.85. Through the RabbitMQ message queue, the new skill usage information and learning ability coefficient are transmitted to the skill decay model in real time to trigger the update mechanism. The skill decay model uses an online random forest algorithm to dynamically adjust based on the newly input parameters, such as reducing the decay coefficient of Java skills from 0.05 / month to 0.04 / month. The updated model is applied to flexible employment matching, and the matching weight of Java skills is increased from 0.8 to 0.85. Git is used for data version control to ensure the consistency of multi-source data such as work logs and training records, and the manual review process is automatically triggered when data conflicts are detected.

[0088] Step S107, for different types of skills, corresponding attenuation curve templates are constructed, including the attenuation curve of programming languages ​​and the attenuation curve of customer service skills. When performing skill matching, the corresponding attenuation curve template is selected according to the skill type of the temporary worker, and the learning ability coefficient is adjusted as a parameter to obtain the skill attenuation prediction result.

[0089] The historical data of different types of skills are obtained, and the attenuation curve template is constructed by the least squares method. The attenuation curve template is quantified into an exponential attenuation function f(t) = ae^(-bt); where a represents the initial proficiency, b represents the attenuation rate, and t represents time. According to the attenuation curve template, the K-means clustering algorithm is used to classify the skills of temporary workers; the classification process includes classifying the skills into predefined skill types based on skill descriptions and application scenarios; determining the optimal number of clusters by the silhouette coefficient method; and determining the attenuation curve template corresponding to each skill. For the attenuation curve template, the learning ability coefficient L is introduced for linear adjustment; the linear adjustment includes designing the parameter adjustment formula b' = b*(2-L); and dynamically adjusting the slope of the attenuation curve according to different learning ability levels. Using the ARIMA model, combined with the attenuation curve template and the adjusted parameters, the attenuation prediction of the specific skills of temporary workers is carried out; the attenuation prediction includes generating a skill proficiency change curve in the future. The root mean square error (RMSE) is used to evaluate the accuracy of the attenuation prediction results. The evaluation process includes comparing the deviation between the predicted value and the actual observed value. If the deviation exceeds the preset threshold, the model parameters are dynamically optimized. If the deviation does not exceed the preset threshold, the attenuation prediction result is determined to be a valid prediction result.

[0090] Specifically, based on historical data, the least squares method is used to construct decay curve templates for different types of skills, and the decay characteristics of skill types such as programming languages ​​and customer service skills are quantified as exponential decay functions f(t) = ae^(-bt), where a represents the initial proficiency, b represents the decay rate, and t represents time. The K-means clustering algorithm is used to classify the skills of temporary workers. Based on the skill description and application scenario, the skills are classified into predefined skill types. The optimal number of clusters is determined by the silhouette coefficient method, and the decay curve template corresponding to each skill is determined. The learning ability coefficient L is introduced into the decay curve template through the linear adjustment function, and the parameter adjustment formula b' = b*(2-L) is designed to dynamically adjust the slope of the decay curve according to different learning ability levels. The ARIMA model is used to predict the decay of the specific skills of temporary workers, combined with the selected decay curve template and the adjusted parameters, and the skill proficiency change curve in the future period is generated. The root mean square error (RMSE) is used to evaluate the accuracy of the decay prediction results, compare the deviation between the predicted value and the actual observed value, and dynamically optimize the model parameters. According to the differences in the decay characteristics of different types of skills such as programming languages ​​and customer service skills, decay curves are constructed and adjusted respectively. For example, programming languages ​​use a larger initial decay rate, while customer service skills use a smaller initial decay rate. In the process of constructing the skill decay curve template, Java programming and customer service skills are used as examples to illustrate. First, the least squares method is used to fit the historical data, and the decay function of Java programming is f(t) = 95e^(-0.05t), and the decay function of customer service skills is f(t) = 90e^(-0.03t), where t is in months. Then, the K-means clustering algorithm is used to classify the 50,000 skills of 10,000 temporary workers. By calculating the silhouette coefficient, the optimal number of clusters is determined to be 8, Java programming is classified into the "programming language" category, and customer service is classified into the "soft skills" category. The learning ability coefficient L is introduced to adjust the decay curve. Assuming that the Java skill learning ability coefficient of a temporary worker is 1.2, the adjusted decay rate b'=0.05*(2-1.2)=0.04, and the new decay function f(t)=95e^(-0.04t) is obtained. The ARIMA(1,1,1) model is used to predict the skill proficiency in the next 6 months. Combined with the adjusted decay function, the Java skill proficiency of the temporary worker in 6 months is predicted to be 87.3 points. Finally, by comparing the predicted value with the actual observed value, the RMSE is calculated to be 2.1 points. Based on this, the model parameters are fine-tuned. Considering the differences in the decay characteristics of programming languages ​​and customer service skills, different initial decay rates and learning ability coefficient adjustment methods are used to more accurately reflect the decay laws of different types of skills.

[0091] For two different skill types, programming languages ​​and customer service skills, we collected a large amount of skill usage data in related fields, including skill usage frequency and skill usage time interval. We used the data to fit an exponential decay curve that can reflect the decay characteristics of this skill type. This curve serves as the decay curve template for this type of skill. Different curve templates are established according to the decay speed and rules of different skill types.

[0092] Web page information containing programming language and customer service skill usage data from online recruitment platforms and social networks is obtained, and skill usage frequency and time interval data are extracted based on the web page information; based on the skill usage frequency and time interval data, the Z-score method is used to identify outliers, and duplicates are removed through the hash algorithm to obtain a skill data set in a unified format; for the skill data set, the Levenberg-Marquardt algorithm is used for nonlinear regression fitting, and the exponential decay function y=ae^(-bx) is selected as the basic model, where a represents the initial skill level, b represents the decay rate, and x represents time, and the best fitting parameters are determined through iterative optimization; based on the best fitting parameters, the decay curve templates of programming languages ​​and customer service skills are constructed, and the slope of the decay curve is calculated by derivation to obtain the average decay rate of the two skill types; for the decay curve template, the K-fold cross-validation method is used for verification and evaluation, and the root mean square error is calculated to determine whether the stability and generalization ability of the decay curve template meet the preset threshold.

[0093] Specifically, the Scrapy framework is used to crawl the usage data of two skills, programming language and customer service skills, from online recruitment platforms and social networks, extract the frequency and time interval information of skill usage, and construct a structured data set containing skill type, usage frequency, and time interval. The Z-score method is used to identify and remove outliers, and the hash algorithm is used to remove duplicates. The data format is unified, the skill usage frequency is standardized to the number of times per month, and the time interval is converted to days. The Levenberg-Marquardt algorithm is used to perform nonlinear regression fitting on the processed data, and the exponential decay function y=ae^(-bx) is selected as the basic model, where a represents the initial skill level, b represents the decay rate, and x represents time. The best fitting parameters are determined by iterative optimization. According to the fitting results, the decay curve templates of programming language and customer service skills are constructed respectively. The slope of the decay curve is calculated by derivation to obtain the average decay rate of the two skill types, and the decay curve parameters are stored in the template library. The constructed decay curve template is verified and evaluated using the K-fold cross-validation method, and the root mean square error is calculated to ensure the stability and generalization ability of the model. Compare the differences between programming languages ​​and customer service skills in data characteristics, fitting parameters and decay rules. For example, the initial skill level a of programming languages ​​is higher but the decay rate b is larger, while the initial skill level a of customer service skills is lower but the decay rate b is smaller. In the process of constructing the skill decay curve template, Java programming and customer service skills are used as examples to illustrate. First, the Scrapy framework is used to crawl 500,000 data from platforms such as Linked In and GitHub. The skill usage information is extracted through the BERT model. The average usage frequency of Java programming is 15 times / month, the time interval is 3 days, and the frequency of customer service skills is 50 times / month, and the time interval is 0.5 days. The Z-score method (threshold is set to 3) is used to identify outliers, and about 2% of the data is eliminated. The MD5 hash algorithm is used to remove 1.5% of duplicates. The Levenberg-Marquardt algorithm is used for nonlinear regression fitting. The decay function of Java programming is y=95e^(-0.05x), and that of customer service skills is y=85e^(-0.02x), where x is in months. By taking the derivative to calculate the decay rate, the initial decay rate of Java programming is -4.75 points / month, and that of customer service skills is -1.7 points / month. Using a 5-fold cross-validation evaluation model, the root mean square error of Java programming is 2.3 points, and that of customer service skills is 1.8 points. Comparing the two skills, it is found that the initial skill level of Java programming is high (95 points) but decays quickly, while the initial level of customer service skills is low (85 points) but decays slowly, reflecting the typical decay characteristics of hard skills and soft skills. These decay curve templates lay the foundation for subsequent skill matching and prediction.

[0094] Determine the type of skills they possess, select the attenuation curve template of the corresponding skill type, substitute the personal learning ability coefficient of the temporary worker into the attenuation curve template, adjust the curve, and obtain the corrected attenuation coefficient by combining the personal learning ability coefficient with the skill type attenuation curve. Substitute the corrected attenuation coefficient into the attenuation curve template to calculate the predicted value of the temporary worker's skill proficiency at different time points.

[0095] The skill description information of temporary workers is obtained, and the skill description information includes multiple skill contents; the skill description information is classified and processed by using a support vector machine algorithm, wherein the support vector machine algorithm uses TF-IDF feature extraction and RBF kernel function; the parameters C and gamma of the support vector machine algorithm are optimized by a cross-validation method, and multiple skill contents are classified into predefined skill types according to the optimized parameters; the historical learning records and training performance data of temporary workers are received, and the personal learning ability coefficient is calculated by a multi-factor scoring method according to the historical learning records and training performance data; after obtaining the personal learning ability coefficient, the personal learning ability coefficient is combined with the attenuation curve corresponding to the predefined skill type through a linear adjustment function to obtain a revised attenuation coefficient; the fourth-order Runge-Kutta method is used to substitute the revised attenuation coefficient into the preset attenuation curve template to calculate the skill proficiency prediction value sequence at different time points.

[0096] Specifically, the support vector machine algorithm is used to classify the skill descriptions of temporary workers. TF-IDF feature extraction and RBF kernel function are used. Through cross-validation optimization of parameters C and gamma, each skill is classified into a predefined skill type, and the corresponding type of decay curve template is selected from the decay curve template library. According to the historical learning records and training results of temporary workers, the multi-factor scoring method is used to calculate the personal learning ability coefficient. The scoring items include learning speed (weight 0.3), knowledge retention rate (weight 0.3), and application ability (weight 0.4). The score range of each item is 0-100, and the weighted average is calculated to obtain the learning ability coefficient. The personal learning ability coefficient is combined with the skill type decay curve through the linear adjustment function f(x) = ax + b, where a and b are preset parameters and x is the original decay coefficient, and the modified decay coefficient is obtained. Using the fourth-order Runge-Kutta method, the modified decay coefficient is substituted into the decay curve template y = A*e^(-kt), the initial proficiency A, the prediction time range T and the step size h are set, the skill proficiency prediction values ​​at different time points are calculated, and the prediction result sequence is generated. Use Matplotlib to draw a graph of skill proficiency over time, with time as the horizontal axis and proficiency as the vertical axis. Different skills are represented by different colored curves. For multiple skills, the weighted average of the proficiency of each skill is calculated. The weight is determined based on the importance of the skill to obtain the comprehensive skill decay prediction result. In the process of skill decay prediction, Java programming and customer service are used as examples to illustrate. First, the support vector machine algorithm is used to classify 10,000 skill descriptions. After 5-fold cross validation, the optimal parameters C=10 and gamma=0.1 are selected, and the classification accuracy reaches 95%. Java programming is classified as "programming language" and the corresponding decay curve template y=100e^(-0.05t) is selected; customer service is classified as "soft skills" and the template y=90e^(-0.02t) is selected. When calculating the learning ability coefficient of a temporary worker, the learning speed score is 85 (weight 0.3), the knowledge retention score is 90 (weight 0.3), and the application ability score is 80 (weight 0.4). The weighted average is 84.5. The learning ability coefficient is combined with the attenuation curve using the linear adjustment function f(x) = 0.5x + 0.5. The attenuation coefficient of Java programming is adjusted from 0.05 to 0.04625, and the customer service is adjusted from 0.02 to 0.0185. Using the fourth-order Runge-Kutta method, the initial proficiency Java is set to 95 and the customer service is set to 85. The prediction time range is 12 months and the step length is 1 month. The proficiency of Java programming after 12 months is calculated to be 72.3 and the customer service is 74.6. Use matplotlib to draw the attenuation curve, with Java programming represented by the blue line and customer service represented by the red line.Assuming that the weight of Java programming is 0.6 and the weight of customer service is 0.4, the comprehensive skill proficiency is calculated to be 73.2 after 12 months, which provides data support for temporary worker matching.

[0097] Step S108, in the skill matching process, the skill proficiency, skill decay rate, and employment willingness of the temporary workers are combined, and a weighted calculation method is adopted to obtain a comprehensive matching degree. Temporary workers with high matching degrees will be recommended to employers first.

[0098] The skill proficiency scores, skill decay rate prediction values ​​and employment intention survey results of temporary workers in the temporary worker database are obtained, and the employment intention survey results are converted into scores using the Likert scale; a judgment matrix is ​​constructed based on the skill proficiency scores, skill decay rate prediction values ​​and employment intention scores, and the eigenvectors are calculated to obtain weights for consistency test; the skill proficiency scores, skill decay rate prediction values ​​and employment intention scores are normalized using the weights to obtain the min-max normalization process to obtain the standardized scores; the comprehensive matching scores of temporary workers are calculated based on the standardized scores and corresponding weights; based on the comprehensive matching scores, the cosine similarity matching algorithm is used to match temporary workers with the needs of employers, and the temporary jobs with the highest similarity are selected as the recommendation results.

[0099] Specifically, the skill proficiency score, skill decay rate prediction value and employment intention survey results of each temporary worker were extracted from the temporary worker database, where the employment intention was converted into a score of 1-5 using the Likert scale. For multiple skills, the weighted average score was calculated, and the weight was determined based on the importance of the skill. The weights of the three factors of skill proficiency, skill decay rate and employment intention were determined using the hierarchical analysis method, and a judgment matrix was constructed. The eigenvectors were calculated to obtain the weights, and consistency tests were performed to ensure the reliability of the results. The skill proficiency score, skill decay rate prediction value and employment intention score were converted into standardized scores in the range of 0-1 using the min-max normalization method. According to the weighted calculation model, the standardized scores of the three factors were multiplied and summed with the corresponding weights to obtain the comprehensive matching score of each temporary worker. The temporary workers were sorted from high to low according to the scores to generate a recommendation list. A matching algorithm based on cosine similarity was designed to match the recommendation list with the needs of the employer, and the top N temporary workers with the highest similarity were selected as the final recommendation results. For missing data, the mean filling method was used, and outliers were identified and processed using the 3σ principle to ensure the reliability and fairness of the calculation results. In the process of skill matching, the Java development engineer position is used as an example. The information of 1,000 temporary workers is extracted from the database, among which the average Java skill proficiency score is 85 points (out of 100 points), the decay rate prediction value is 0.05 points / month, and the employment intention survey result is 4.2 points on an average scale of 1-5. For multi-skilled temporary workers, such as those who have both Java and Python skills, the weighted average score is calculated according to the weight of 6:4. The weights are determined by the hierarchical analysis method. After constructing the judgment matrix, the skill proficiency weight is 0.5, the decay rate weight is 0.3, and the employment intention weight is 0.2. The consistency ratio CR is 0.02, which is less than 0.1 and passes the consistency test. The min-max normalization is used to convert the three factors into 0-1 scores. For example, the normalized scores of a temporary worker are 0.9, 0.8, and 0.95 respectively. The comprehensive matching degree is calculated to be 0.9×0.5+0.8×0.3+0.95×0.2=0.88. The 1,000 temporary workers are sorted and a recommendation list is generated. The employer demand vector is [0.9, 0.8, 0.7], which represents the requirements for skill proficiency, stability, and enthusiasm. The cosine similarity algorithm is used to calculate the matching degree, and the top 10 are selected as the final recommendation. In data processing, 5% of missing values ​​are filled with the mean, and the 3σ principle identifies 2% of outliers and replaces them with the median to ensure the reliability and fairness of the results.

[0100] Step S109, regularly evaluate the prediction accuracy of the skill decay model, optimize the model parameters by collecting actual employment feedback information of temporary workers, including post-employment work performance evaluation, and refine and adjust the model according to the differences in decay laws of different skill types.

[0101] The performance evaluation data of temporary workers are obtained from employers, and the evaluation data include work quality, efficiency and skill application. The standard score is calculated by weighted average method based on the evaluation data. The ARIMA model is used to perform time series analysis on the standard score to obtain the predicted value. The prediction deviation and root mean square error between the predicted value and the standard score are calculated. The parameters of the skill decay model are optimized by stochastic gradient descent algorithm, and the parameters include decay coefficient and initial skill level. The K-means clustering algorithm is used to group skill types to obtain multiple skill clusters. An independent decay model is established for each skill cluster. The correlation matrix is ​​introduced to calculate the influence coefficient between skill clusters. The parameters of the skill decay model are adjusted according to the influence coefficient.

[0102] Specifically, the performance evaluation data of temporary workers, including work quality, efficiency and skill application, are obtained from employers through RESTful API, and the evaluation results are quantified into standard scores of 0-100 using the weighted average method. The ARIMA model is used for time series analysis to compare the difference between the actual performance scores of temporary workers and the predicted values ​​of the skill decay model, calculate the prediction deviation and root mean square error, and evaluate the prediction accuracy of the model. The stochastic gradient descent algorithm is used to optimize the parameters of the skill decay model, setting the initial learning rate to 0.01 and the maximum number of iterations to 1000. The decay coefficient and initial skill level are adjusted according to the prediction deviation to minimize the prediction error. The K-means clustering algorithm is used to group the skill types, and an independent decay model is established for each skill cluster. The performance of the refined model is evaluated by the 5-fold cross-validation method. The model evaluation and optimization is set to be performed once a quarter, and the changes in the skill decay trend are analyzed by comparing the model performance in different quarters. When dealing with the mutual influence between different skill types, the correlation matrix is ​​introduced to calculate the influence coefficient between skills. To ensure the stability and convergence of the model optimization process, the early stopping method is used. When the model performance does not improve after 5 consecutive iterations, the optimization is stopped. In the process of skill decay model evaluation and optimization, Java programming skills are used as an example. The quarterly performance evaluation of 1,000 temporary workers is obtained from 50 partner companies through RESTful API. The evaluation dimensions include code quality, development efficiency, and problem-solving ability, with weights of 0.4, 0.3, and 0.3 respectively. The weighted average method is used to calculate the standard score. For example, the score of a temporary worker is 85 points. The ARIMA (1,1,1) model is used to compare the actual score with the predicted value, and the root mean square error RMSE is calculated to be 7.5 points. The stochastic gradient descent algorithm uses RMSE as the objective function and the initial learning rate is set to 0.01. After 500 iterations, the decay coefficient of Java skills is optimized from 0.05 / month to 0.042 / month. K-means clustering divides 1,000 temporary workers into three skill clusters, including 350 senior, 450 mid-level, and 200 junior, and establishes decay models for each. 5-fold cross-validation shows that the average RMSE of the refined model is reduced to 6.2 points. In the quarterly evaluation, it was found that the decay trend of Java skills slowed down, and the annual decay rate dropped from 6% to 5%. Skill correlation analysis shows that the correlation coefficient between Java and Python is 0.7, indicating that there is a strong correlation between the decay of these two skills. During the model optimization process, the early stopping mechanism is triggered when the improvement of RMSE for 5 consecutive iterations is less than 0.1%, ensuring the stability and efficiency of the optimization process.

[0103] In the description of the present disclosure, it is necessary to understand that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present disclosure and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the scope of protection of the present disclosure.

[0104] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present disclosure.

Claims

1. A method for matching temporary workers' flexible employment information, characterized in that: include: S101, constructing a skill profile of temporary workers based on their historical employment information; S102, classifying the skills of temporary workers according to their skill profiles to obtain subsets of temporary workers with different skill types; S103, for each subset of temporary workers under each skill type, by analyzing their skill usage frequency and last usage time, a skill decay model is constructed to quantify the skill decay rate in different time periods; S104, calculating the learning ability coefficient of the temporary worker by analyzing the historical learning performance and training participation, and using the learning ability coefficient and the skill usage frequency attribute as input parameters of the attenuation model; S105, predicting the change in skill proficiency of temporary workers in the future based on the skill decay model; S106, in the process of matching flexible employment information, dynamically obtain the latest skill usage information of temporary workers and update the input parameters in the skill decay model; S107, for different types of skills, construct corresponding attenuation curve templates; According to the skill type of temporary workers, the corresponding attenuation curve template is selected, and the learning ability coefficient is adjusted as a parameter to obtain the skill attenuation prediction result; S108, in the process of skill matching, the skill proficiency, skill decay rate, and employment willingness of temporary workers are combined to obtain a comprehensive matching degree; S109, regularly evaluate the prediction accuracy of the skill decay model, optimize the model parameters by collecting actual employment feedback information from temporary workers, and refine and adjust the model based on the differences in decay patterns of different skill types.

2. The method for matching temporary workers' flexible employment information according to claim 1, characterized in that: The construction of a temporary worker skill profile based on the temporary worker's historical employment information includes: Obtain the skill name involved in each job from the historical employment records of temporary workers. The skill name is obtained by matching the job description keywords in the preset skill vocabulary. Calculate the last time each skill was used based on historical employment records, where the last time was determined by traversing the work history that included the skill and selecting the latest end date; For each skill, the cumulative number of months of use in historical employment records is counted, and the cumulative number of months of use is divided by the total number of working months to obtain the skill use frequency value; Combine the skill name, last used time and frequency of use into structured data to form a temporary worker's personal skill information table; A multidimensional vector representation method is used to construct a skill profile of temporary workers. The multidimensional vector includes the skill name, last use time and frequency of use.

3. The method for matching temporary workers' flexible employment information according to claim 1, characterized in that: According to the skill profile of temporary workers, the skills of temporary workers are classified, and the subsets of temporary workers with different skill types are obtained, including: Obtain all skill items in the skill profile of temporary workers, including programming languages ​​and customer service skills, and use the Word2Vec word embedding tool to convert the skill items into high-dimensional vector representations to obtain a skill vector set; Perform dimensionality reduction processing based on the skill vector set, extract key features through principal component analysis tools, calculate eigenvalues ​​and eigenvectors, sort by eigenvalue size, accumulate eigenvalue contribution rates to a preset threshold, and retain the corresponding eigenvectors as dimensionality reduction results; For the dimensionality reduction results, calculate the intra-cluster sum of squares corresponding to each K value within the preset K value range, draw a curve graph, and select the K value corresponding to the inflection point of the curve as the final number of clusters; Use the final number of clusters, perform K-means clustering on the dimensionality reduction results, randomly select the initial center point, calculate the Euclidean distance from the sample to each center point, assign the sample to the nearest center point, recalculate the center point of each cluster, and repeat this process until the center point position no longer changes or the maximum number of iterations is reached; According to the clustering results, a skill type label is assigned to each cluster. A skill type vector library is built based on the pre-established industry standard occupational classification. The cosine similarity between the cluster center and each skill type vector is calculated, and the top three types with the highest similarity are selected as the multi-labels of the cluster.

4. The method for matching temporary workers' flexible employment information according to claim 1 is characterized in that: For each subset of temporary workers under each skill type, a skill decay model is constructed by analyzing the skill usage frequency and the last usage time, and the skill decay speed in different time periods is quantified, including: Obtain the skill usage records of the temporary worker subset, and extract the usage frequency and last usage time of each skill based on the skill usage records; The missing data in the frequency of use and the last time of use were processed by the median filling method, and the outliers in the frequency of use and the last time of use were eliminated by the 3σ principle to obtain a standardized skill use data set. Based on the standardized skill usage data set, the last usage time is grouped by month using a fixed 30-day time window. Calculate the average frequency of skill usage in each time window to form time series data; Determine the initial form of the exponential decay function, the exponential decay function is f(t)=ae^(-bt), where a represents the initial frequency, b represents the decay coefficient, and t represents time; The exponential decay function was fitted using the least squares method, with time as the independent variable and average frequency of use as the dependent variable; The objective function is set to minimize the square difference between the predicted value and the actual value, and the optimal values ​​of the attenuation coefficient b and the initial frequency a are solved to construct a skill attenuation model. Apply skill decay models to different skill types and calculate skill decay values ​​in each time period; The decay rate is quantified by calculating the slope of the decay curve at different time points, and a decay curve is drawn to show the trend of skill changes over time.

5. The method for matching temporary workers' flexible employment information according to claim 1, characterized in that: The above mentioned method of calculating the learning ability coefficient of temporary workers by analyzing their historical learning performance and training participation, and using the learning ability coefficient and skill usage frequency attribute as input parameters of the attenuation model includes: Obtain learning performance and training participation records from the temporary workers' historical database, identify outliers, and use the mean to fill in missing data based on outlier judgments to obtain a standardized learning ability assessment data set; Based on the standardized learning ability assessment data set, the hierarchical analysis method is used to determine the weight of the weighted average of multiple factors, and the weighted calculation is performed on historical learning performance, training participation, learning speed and knowledge retention rate to determine the learning ability coefficient of each temporary worker; Obtain the learning ability coefficient and skill usage frequency of temporary workers, and use the correlation analysis tool to calculate the Pearson correlation coefficient and Spearman rank correlation coefficient between the two. If the correlation coefficient is greater than the preset threshold, it is judged whether it is reasonable to introduce the learning ability factor into the attenuation model. If it is judged that it is reasonable to introduce the learning ability factor into the attenuation model, the mathematical expression of the original attenuation model is modified, the learning ability coefficient is used as a new input variable, combined with the original model by multiplication, and the model parameters are refitted through multiple linear regression; The modified attenuation model is subjected to ten-fold cross validation, and the root mean square error of the model before and after modification is calculated. If the root mean square error of the modified model is smaller than that of the model before modification, it is determined that the modified attenuation model has higher prediction accuracy.

6. The method for matching temporary workers' flexible employment information according to claim 1, characterized in that: According to the skill decay model, the prediction of the skill proficiency changes of temporary workers in the future period includes: Obtain the initial skill proficiency and attenuation coefficient, and substitute the initial skill proficiency and attenuation coefficient into the skill attenuation model formula for temporary workers; According to the model formula, the Runge-Kutta method is used for numerical integration to obtain the predicted value of skill proficiency for each month; The skill attenuation rate is calculated using the sliding window method through the predicted value of skill proficiency; Obtain the proficiency prediction values ​​of two adjacent months from the sliding window, calculate the difference between the proficiency prediction values ​​of two adjacent months, and obtain the decay rate value of each month; The K-means clustering algorithm is used to group the decay speeds of all skills. If the optimal number of clusters is determined by the silhouette coefficient method, the skills with faster decay speeds are clustered into one category. For skill clusters with faster decay rates, the quartile method is used to select the upper quartile as the decay threshold. If the skill decay rate exceeds the decay threshold, the skill is included in the fast decay skill set. Get the decay speed from the fast decay skill set and use the decay speed as the matching weight factor; Obtain skill proficiency and timeliness weights, combine the matching weight factor, skill proficiency and timeliness weights, and use the linear weighted method to calculate the comprehensive matching score.

7. The method for matching temporary workers' flexible employment information according to claim 2, characterized in that: In the flexible employment information matching process, the latest skill usage information of temporary workers is dynamically obtained, and the input parameters in the skill decay model are updated, including: Obtaining temporary workers’ skill usage information, which is extracted from work logs, project reports, and online collaboration platforms; Update the skill's last usage time and usage frequency data based on skill usage information; The individual learning ability coefficient is calculated using the online gradient descent method. The individual learning ability coefficient is obtained based on the training courses, test scores and skill certifications that the temporary workers have participated in; Use the message queue system to transmit skill usage information and personal learning ability coefficient to the skill decay model; If new skill usage information and personal learning ability coefficient are received, the skill decay model parameter update mechanism is triggered; Use the online random forest algorithm to optimize the skill decay model, and dynamically adjust the model weight and decay coefficient during the optimization process; Apply the skill decay model to the flexible employment information matching process, and the application process includes dynamically adjusting the matching weights; The update and synchronization of multi-source data, including skill usage information, personal learning ability coefficients, and skill decay model parameters, are coordinated through data version control tools.

8. The method for matching temporary workers' flexible employment information according to claim 2 is characterized in that: For different types of skills, corresponding attenuation curve templates are constructed; According to the skill type of temporary workers, the corresponding attenuation curve template is selected, and the learning ability coefficient is adjusted as a parameter. The skill attenuation prediction results include: The historical data of different types of skills are obtained, and the attenuation curve template is constructed by the least squares method. The attenuation curve template is quantified into an exponential attenuation function f(t) = ae^(-bt); where a represents the initial proficiency, b represents the attenuation rate, and t represents time. According to the attenuation curve template, the K-means clustering algorithm is used to classify the skills of temporary workers; the classification process includes classifying the skills into predefined skill types based on skill descriptions and application scenarios; determining the optimal number of clusters by the silhouette coefficient method; and determining the attenuation curve template corresponding to each skill. For the attenuation curve template, the learning ability coefficient L is introduced for linear adjustment; the linear adjustment includes designing the parameter adjustment formula b' = b*(2-L); and dynamically adjusting the slope of the attenuation curve according to different learning ability levels. Using the ARIMA model, combined with the attenuation curve template and the adjusted parameters, the attenuation prediction of the specific skills of temporary workers is carried out; the attenuation prediction includes generating a skill proficiency change curve in the future. The root mean square error (RMSE) is used to evaluate the accuracy of the attenuation prediction results. The evaluation process includes comparing the deviation between the predicted value and the actual observed value. If the deviation exceeds the preset threshold, the model parameters are dynamically optimized. If the deviation does not exceed the preset threshold, the attenuation prediction result is determined to be a valid prediction result.

9. The method for matching temporary workers' flexible employment information according to claim 2, characterized in that: In the skill matching process, the skill proficiency, skill decay rate, and employment willingness of temporary workers are combined to obtain a comprehensive matching degree including: The skill proficiency scores, skill decay rate prediction values ​​and employment intention survey results of temporary workers in the temporary worker database are obtained, and the employment intention survey results are converted into scores using the Likert scale; a judgment matrix is ​​constructed based on the skill proficiency scores, skill decay rate prediction values ​​and employment intention scores, and the eigenvectors are calculated to obtain weights for consistency test; the skill proficiency scores, skill decay rate prediction values ​​and employment intention scores are normalized using the weights to obtain the min-max normalization process to obtain the standardized scores; the comprehensive matching scores of temporary workers are calculated based on the standardized scores and corresponding weights; based on the comprehensive matching scores, the cosine similarity matching algorithm is used to match temporary workers with the needs of employers, and the temporary jobs with the highest similarity are selected as the recommendation results.

10. The method for matching temporary workers' flexible employment information according to claim 2, characterized in that: The prediction accuracy of the skill decay model is evaluated regularly, and the model parameters are optimized by collecting actual employment feedback information of temporary workers; the model is refined and adjusted according to the differences in the decay laws of different skill types, including: The performance evaluation data of temporary workers are obtained from employers, and the evaluation data include work quality, efficiency and skill application. The standard score is calculated by weighted average method based on the evaluation data. The ARIMA model is used to perform time series analysis on the standard score to obtain the predicted value. The prediction deviation and root mean square error between the predicted value and the standard score are calculated. The parameters of the skill decay model are optimized by stochastic gradient descent algorithm, and the parameters include decay coefficient and initial skill level. The K-means clustering algorithm is used to group skill types to obtain multiple skill clusters. An independent decay model is established for each skill cluster. The correlation matrix is ​​introduced to calculate the influence coefficient between skill clusters. The parameters of the skill decay model are adjusted according to the influence coefficient.

Images