A method and system for identifying low-intention candidates in a recruitment process

Abstract

The present invention discloses a method and system for identifying low-intent candidates in a recruitment process, the method comprising: online obtaining multiple attribute feature values ​​of candidates to be recruited and weight coefficients corresponding to the multiple attribute feature values; inputting the multiple attribute feature values ​​and weight coefficients into the following online mathematical model to obtain the probability y that the candidate to be recruited belongs to a low-intent candidate, and sending the orders of the candidates to be recruited whose probability y is greater than a set threshold as the recognition result to a manual customer service system or a telephone robot system. The present invention can accurately and efficiently filter out low-intent candidates, improve the communication efficiency of subsequent intelligent voice robots and manual customer service, and improve delivery efficiency.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method and system for identifying low-intent candidates in a recruitment process. Background Art

[0002] Online job search and recruitment platforms are a tripartite platform that brings together companies, job seekers, and headhunters. After receiving talent needs from corporate HR or headhunters, traditional online job search platforms will conduct a preliminary screening of candidates and select candidates with higher intentions to deliver to corporate HR or headhunters.

[0003] However, this model of manual customer service directly conducting preliminary screening of candidates has many disadvantages, such as manual customer service lacks patience to conduct a large number of repetitive preliminary screening work, cannot quickly and accurately obtain high-intention candidates, and has high labor costs. It not only wastes the time and energy of manual customer service, but also may harass high-intention customers and miss possible high-intention customers. Therefore, how to efficiently and cost-effectively screen out low-intention candidates and obtain high-intention candidate data before manual customer service or telephone robots start communication work has become one of the technical problems that online job search and recruitment platforms need to solve. Summary of the invention

[0004] In view of the above defects or deficiencies in the prior art, the present invention provides a method and system for identifying low-intent candidates in a recruitment process, which can partially or completely solve the above technical problems.

[0005] One aspect of the present invention provides a method for identifying low-intention candidates in a recruitment process, comprising:

[0006] Online acquisition of multiple attribute feature values ​​of the candidates to be recruited and weight coefficients corresponding to the multiple attribute feature values;

[0007] The multiple attribute feature values ​​and weight coefficients are input into the following online mathematical model to obtain the probability y that the candidate to be recruited is a low-intent candidate:

[0008]

[0009] Among them, ω T is the transposed matrix of the vector matrix ω, which is a 1×n column matrix [ω1ω2…ω n ],ω1,ω2,…,ω n Represents the weight coefficient; the vector matrix x is a 1×n matrix [x1x2…x n ],x1,x2,…,x n Represents the attribute characteristic value; ω T x represents ω1x1+ω2x2+…+ω n xn , n is the number of attribute eigenvalues, e is a natural constant, and b is a constant term;

[0010] The orders of the candidates to be recruited whose probability y is greater than the set threshold are taken as identification results and sent to the manual customer service system or the telephone robot system.

[0011] Furthermore, it also includes the steps of generating mathematical models:

[0012] Collect multiple attribute feature data of the candidates to be recruited, and perform data cleaning on the attribute feature data;

[0013] Perform feature processing on the attribute feature data after data cleaning to obtain a training set consisting of multiple attribute feature value samples;

[0014] Input the attribute feature value sample data in the training set into the following machine learning model:

[0015]

[0016] Among them, Y represents the preset probability that the candidate to be recruited belongs to the low-intention candidate; the vector matrix M is a 1×m matrix [M1 M2…M m ],M1,M2,…,M m Represents the input attribute feature value sample; W T is the transposed matrix of the vector matrix W, which is a 1×m-column matrix [W1W2…W m ],W1,W2,…,W m is the weight coefficient of the attribute feature value sample to be solved; W T M means W1M1+W2M2+…+W m M m , m is the number of attribute feature value samples, t is the constant term to be solved, and e is a natural constant;

[0017] Through the machine learning model, the weight coefficient and constant term corresponding to each attribute feature value sample are solved and used as the weight coefficient and constant term of the online mathematical model.

[0018] Furthermore, the data cleaning includes:

[0019] Discard or supplement missing attribute feature data;

[0020] Remove abnormal attribute feature data;

[0021] Standardize the data dimension of attribute feature data;

[0022] And / or, logarithmic smoothing is performed on the attribute feature data that causes the recognition result to show a gradually decreasing trend.

[0023] Furthermore, the feature processing includes:

[0024] Mapping categorical values ​​in attribute feature data to integer values;

[0025] Remove unimportant attribute feature data;

[0026] Perform feature dimension reduction on attribute feature data;

[0027] And / or, the attribute characteristic data is processed to obtain derived attribute characteristic data.

[0028] Furthermore, it also includes:

[0029] Regularization constraints are applied to the solved weight coefficients through the L1 norm or the L2 norm, and the weight coefficients obtained after the regularization constraints are used as the weight coefficients of the online mathematical model.

[0030] Furthermore, it also includes the model monitoring step:

[0031] Regularly monitor the actual delivery of said identification results;

[0032] If the difference between the actual delivery of the recognition result and the average delivery of all recruitment candidates is within a preset range, the online mathematical model is optimized.

[0033] Another aspect of the present invention provides a system for identifying low-intention candidates in a recruitment process, comprising:

[0034] A data acquisition module is configured to acquire online a plurality of attribute characteristic values ​​of the candidates to be recruited and weight coefficients corresponding to the plurality of attribute characteristic values;

[0035] The identification module is configured to input the multiple attribute feature values ​​and weight coefficients into the following online mathematical model to obtain the probability y that the candidate to be recruited is a low-intent candidate:

[0036]

[0037] Among them, ω T is the transposed matrix of the vector matrix ω, which is a 1×n column matrix [ω1ω2…ω n ],ω1,ω2,…,ω n Represents the weight coefficient; the vector matrix x is a 1×n matrix [x1 x2…x n ],x1,x2,…,x n Represents the attribute characteristic value; ω T x represents ω1x1+ω2x2+…+ω n xn , n is the number of attribute eigenvalues, e is a natural constant, and b is a constant term;

[0038] The output module is configured to send the orders of the candidates to be recruited whose probability y is greater than a set threshold as identification results to the manual customer service system or the telephone robot system.

[0039] Furthermore, it also includes a model generation module, which is configured to:

[0040] Collect multiple attribute feature data of the candidates to be recruited, and perform data cleaning on the attribute feature data;

[0041] Perform feature processing on the attribute feature data after data cleaning to obtain a training set consisting of multiple attribute feature value samples;

[0042] Input the attribute feature value sample data in the training set into the following machine learning model:

[0043]

[0044] Among them, Y represents the preset probability that the candidate to be recruited belongs to the low-intention candidate; the vector matrix M is a 1×m matrix [M1 M2…M m ],M1,M2,…,M m Represents the input attribute feature value sample; W T is the transposed matrix of the vector matrix W, which is a 1×m column matrix [W1W2…W m ],W1,W2,…,W m is the weight coefficient of the attribute feature value sample to be solved; W T M means W1M1+W2M2+…+W m M m , m is the number of attribute feature value samples, t is the constant term to be solved, and e is a natural constant;

[0045] Through the machine learning model, the weight coefficient and constant term corresponding to each attribute feature value sample are solved and used as the weight coefficient and constant term of the online mathematical model.

[0046] Furthermore, the data cleaning includes: discarding or supplementing missing attribute feature data; removing abnormal attribute feature data; standardizing the data dimension of the attribute feature data; and / or logarithmically smoothing the attribute feature data that causes the recognition result to show a gradually decreasing trend;

[0047] The feature processing includes: mapping the classification values ​​in the attribute feature data to integer values; removing unimportant attribute feature data; performing feature dimension reduction on the attribute feature data; and / or processing the attribute feature data to obtain derived attribute feature data.

[0048] Furthermore, a model monitoring module is included, which is configured to:

[0049] Regularly monitor the actual delivery of said identification results;

[0050] If the difference between the actual delivery of the recognition result and the average delivery of all recruitment candidates is within a preset range, the online mathematical model is optimized.

[0051] The present invention provides a method and system for identifying low-intention candidates in the recruitment process. The method for identifying low-intention candidates provided by the present application is used to train with historical sample data, and then model the communication intention of the candidates. By filtering out low-intention candidates in this way, the communication efficiency of subsequent intelligent voice robots and manual customer service can be greatly improved, and the delivery efficiency can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings:

[0053] Figure 1 It is a logic flow chart of the generation and training of a machine learning model in a method for identifying low-intent candidates in a recruitment process provided by an embodiment of the present application;

[0054] Figure 2 It is a logic flow chart of the application stage of the online mathematical model in the method for identifying low-intention candidates in the recruitment process provided by one embodiment of the present application;

[0055] Figure 3 This is a logical structure diagram of a system for identifying low-intent candidates in a recruitment process provided by an embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0058] It should be understood that although the terms first, second, third, etc. may be used to describe the acquisition modules in the embodiments of the present invention, the acquisition modules should not be limited to these terms. These terms are only used to distinguish the acquisition modules from each other.

[0059] The word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0060] It should be noted that the directional words such as "upper", "lower", "left", and "right" described in the embodiments of the present invention are described at the angles shown in the drawings and should not be understood as limiting the embodiments of the present invention. In addition, in the context, it should also be understood that when it is mentioned that an element is formed "on" or "under" another element, it can not only be formed directly "on" or "under" another element, but also be formed "on" or "under" another element indirectly through an intermediate element.

[0061] An embodiment of the present invention provides a method for identifying low-intent candidates in the recruitment process, which identifies low-intent job seekers (including unintended job seekers) by constructing a mathematical model, filters out unintended job seekers, and then pushes the pre-screened candidates to an intelligent voice robot for further communication and screening. Through two-level screening, it can help manual customer service to lock in high-intent candidates, and then manual customer service can communicate and promote manually to improve delivery efficiency and avoid harassment of high-intent job seekers.

[0062] Specifically, we fit the final delivery (i.e. the headhunter delivers qualified candidates to the recruiter) based on factors such as the candidate's own attributes, the candidate's job-seeking intention, the candidate's proactive behavior, the candidate's popularity, the degree of match between the candidate's expectations and the communicated position, and explore and find out the main factors affecting the candidate's delivery. When the demand from the company's HR or headhunters is initiated, we use the model to predict the candidate's delivery intention and screen out people with low intention.

[0063] The method for identifying low-intent candidates in the recruitment process disclosed in this embodiment includes a generation and training stage of a machine learning model and a practical application stage of an online mathematical model.

[0064] See also Figure 1 ,The generation and training phase of the machine learning model includes steps such as data collection, data cleaning, feature processing, and model building. Specifically, it includes the following processes:

[0065] Step S101, collecting multiple attribute feature data of candidates to be recruited, and performing data cleaning on the attribute feature data.

[0066] Specifically, the machine learning model needs to be trained using historical data, so whether the candidate accepts the communication and gives feedback is used as the dependent variable (output variable). The relevant attribute characteristics of the candidate are used as independent variables, mainly including the categories shown in Table 1:

[0067] Table 1

[0068]

[0069] First, based on the understanding of the recruitment business, make preliminary assumptions about the independent variables, and then determine the storage location and acquisition method of the attribute feature data corresponding to these independent variables to obtain the most original basic data.

[0070] After the above basic data is prepared, data cleaning is performed first to prepare for modeling.

[0071] Data cleaning includes descriptive statistics for each attribute feature data, such as data volume, number of missing values, maximum value, minimum value, enumeration value, etc., so as to have a basic understanding of each attribute feature data.

[0072] 1. Processing of missing values

[0073] When the missing values ​​of attribute feature data account for a high proportion and have little correlation with the dependent variable, the attribute feature data can be discarded. If the missing values ​​account for a small proportion, missing values ​​can be supplemented, such as by replacing them with 0 or the mean.

[0074] For example, if the educational background data of some candidates to be recruited is empty, you can consider replacing it with the option with the highest educational background distribution ratio.

[0075] 2. Basic visualization methods

[0076] That is, look at the distribution of the data and identify whether there are outliers in time. Box plots are a common identification method. For example, if a user (a candidate to be recruited) has modified his resume abnormally many times recently, it is likely due to abnormal log sending or user behavior. If there are few such users, these samples need to be removed.

[0077] 3. Standardize different data dimensions

[0078] (1) min-max standardization:

[0079]

[0080] Where x is the attribute feature variable to be normalized, x min is the minimum value of the attribute feature variable, x max is the maximum value of the attribute feature variable, x * It is a new attribute feature variable after normalization, and its range is between 0 and 1.

[0081] (2) Z-score standardization

[0082] If the number of times the candidate's information has been viewed and downloaded by HR / headhunters in the recent period is quite different in magnitude, Z-score standardization is used to convert all attribute feature variables to the same scale range. The standardization formula is:

[0083]

[0084] where x * is the normalized new attribute feature variable, μ is the mean of the attribute feature variable, σ is the standard deviation of the attribute feature variable, and x is the attribute feature variable to be normalized. If the distribution of x is close to the normal distribution, the normalized new attribute feature variable is approximately a standard normal distribution with a mean of 0 and a variance of 1.

[0085] 4. Logarithmic smoothing

[0086] Logarithmic smoothing is suitable for scenarios where the impact of changes in the independent variable on the dependent variable is gradually decreasing. For example, logarithmic smoothing is required for the number of days without login in a recent period of time.

[0087] The smoothing formula is:

[0088] x * =log 10 (x+1)

[0089] Among them, x * is the new attribute feature variable after normalization, and x is the attribute feature variable to be normalized.

[0090] Step S102, performing feature processing on the attribute feature data after data cleaning to obtain a training set consisting of a plurality of attribute feature value samples.

[0091] Specifically, for categorical variables in attribute feature variables, One-Hot encoding is used to process them, so that an enumeration variable with K enumeration values ​​is converted into K new binary variables. For example, candidate attribute features such as gender have two enumerations, male and female. Using One-Hot encoding, it is converted into two variables, 1-male or not, 2-female or not, and the corresponding values ​​are 0 / 1.

[0092] Then, feature selection is performed. In order to ensure statistically significant results, the data required by the model needs to grow exponentially with the increase in dimension. Therefore, when there are too many input attribute feature variables, some unimportant features need to be removed, or feature dimensionality reduction can be performed. Methods such as correlation coefficient, variance analysis, and principal component analysis can be used.

[0093] (1) Correlation coefficient: The Pearson correlation coefficient is used to quantify the linear correlation between two continuous variables. Therefore, the first step is to calculate the correlation between each attribute feature and the target variable, so that only features with a high correlation with the target variable are retained.

[0094] (2) Variance analysis: A small variance of a variable means that the data distribution corresponding to this feature has a small variation range. When selecting features, the features are first normalized and then the variance is compared for screening.

[0095] (3) Principal component analysis: It transforms a series of related variables into linearly independent variables or principal components through orthogonal transformation, selects several principal components with larger variance, and retains more information in the original data with fewer variables.

[0096] In the process of model building in this embodiment, the correlation coefficients between the attribute characteristics and the dependent variables are calculated, and then through comparison, it is found that among the matching characteristics of the candidate's expectations and the communication position, the correlation coefficient between whether the industry matches and whether the final delivery can be made is relatively small, only 0.09, while the correlation coefficients of other variables are between 0.4-0.78. Therefore, when modeling, the attribute characteristic "whether the industry matches" is discarded.

[0097] In some scenarios, it is also necessary to use the original data for processing to generate a comprehensive index as an input feature. In the process of model construction in this embodiment, the degree of completion of different content items of the resume is used to process the comprehensive index of resume completeness, which is used to reflect the job-seeking intention of the candidate to be recruited to a certain extent. A complete job application resume contains personal information, educational experience, work experience, project experience, career development intention, language ability, self-evaluation, additional information and other parts, each of which contains more detailed content. The completion ratio of the resume is calculated based on the filled content. It is assumed that the more complete a person fills out the resume, the more willing he is to fully display more information about himself, reflecting his higher desire to find a job. For example, one person only fills in personal information and educational experience, completing 10%, and another person carefully fills in all the information of the resume, completing 100%. The recruitment platform is more willing to believe that the person who completes 100% has a stronger desire to find a job.

[0098] Step S103, input the attribute feature value sample data in the training set into the following machine learning model,

[0099]

[0100] Among them, Y represents the preset probability that the candidate to be recruited belongs to the low-intention candidate; the vector matrix M is a 1×m matrix [M1 M2…M m ],M1,M2,…,M m Represents the input attribute feature value sample; W T is the transposed matrix of the vector matrix W, which is a 1×m-column matrix [W1W2…W m ],W1,W2,…,W m is the weight coefficient of the attribute feature value sample to be solved; W T M means W1M1+W2M2+…+W m M m , m is the number of attribute feature value samples, t is the constant term to be solved, and e is a natural constant.

[0101] Specifically, after data cleaning and feature processing, the model building step begins. The dependent variable, whether the candidate has successfully communicated, is a binary variable (yes / no), so a binary classification model is used to fit the data. At the same time, the confusion matrix is ​​combined with important indicators for classification model evaluation such as precision, recall, ROC curve, and AUC in model selection.

[0102] The confusion matrix is ​​a visualization tool in supervised learning that is used to compare classification results and the true information of an instance. Each row in the confusion matrix represents the predicted category of an instance, and each column represents the true category of the instance.

[0103] Precision measures how many of the samples predicted as positive are actually positive.

[0104] Recall, which measures how many positive samples are predicted to be positive;

[0105] ROC curve, also known as receiver operating characteristic curve. ROC curve is derived based on confusion matrix. The two main indicators in ROC curve are true positive rate and false positive rate. Among them, the horizontal axis is false positive rate, and the vertical axis is true positive rate. The closer the ROC curve is to the upper left corner, and the steeper the curve is, the better the model performance is.

[0106] AUC is the area under the ROC curve. The area under the curve comprehensively measures the effects of all possible classification thresholds. The steeper the ROC curve, the better, so the ideal value of AUC is 1. For a square, the worst random judgment is 0.5, so the general AUC value is between 0.5 and 1, and the closer to 1, the better.

[0107] In the model building scenario of this embodiment, in order to prevent the omission of deliverable candidates caused by the telephone robot's calling, more attention is paid to the accuracy of identifying low-intent candidates, and the model selection is carried out in combination with the comprehensive indicator AUC.

[0108] In this embodiment, the final consideration is whether a candidate is willing to communicate, which is a binary classification problem. Based on the evaluation results of the model (the evaluation process is not described in detail), the interpretability of the model and its popularity in product and operation applications are considered, and the logistic regression model is finally selected as the machine learning model.

[0109] The logistic regression function is an S-shaped curve whose value is between [0,1]. When it is far away from 0, the function value will quickly approach 0 or 1. This feature is very important for solving binary classification problems. The hypothesis function (machine learning model) of logistic regression is as follows:

[0110]

[0111] Among them, Y represents the preset probability that the candidate to be recruited belongs to the low-intention candidate; the vector matrix M is a 1×m matrix [M1 M2…M m ],M1,M2,…,M m Represents the input attribute feature value sample; W T is the transposed matrix of the vector matrix W, which is a 1×m-column matrix [W1W2…W m ],W1,W2,…,W m is the weight coefficient of the attribute feature value sample to be solved; W T M means W1M1+W2M2+…+W mM m , m is the number of attribute feature value samples, t is the constant term to be solved, and e is a natural constant;

[0112] Step S104, using a machine learning model, solve the weight coefficient and constant term corresponding to each attribute feature value sample, and use them as the weight coefficient and constant term of the online mathematical model.

[0113] Specifically, consider Y as the probability that Y is a positive example when M is determined, that is,

[0114]

[0115] Among them, M is the matrix of the sample of the relevant attribute characteristics of the candidates selected in the first step, W is the required weight parameter, and t is the required constant term. In other words, through the training of a large number of samples, the ultimate goal of the above machine learning model is to find the appropriate W and t, and use the final W and t as the weight coefficient and constant term of the online mathematical model.

[0116] Furthermore, after the mathematical form of the machine learning model (logistic regression model) is determined, the maximum likelihood estimation is often used in statistics to estimate the model parameters. In practical applications, in order to avoid overfitting and improve the generalization ability of the model, the present embodiment further adopts a regularization method to control the complexity of the model. Specifically, the weight coefficient W can be constrained by L1 paradigm or L2 paradigm. L1 paradigm regularization increases the sum of the absolute values ​​of all weight coefficients W, forcing more W to be zero, that is, becoming sparse, and can actually achieve automatic selection of features at the same time. In L2 paradigm regularization, the sum of the squares of all weight coefficients W is increased, forcing all W to approach zero as much as possible but not zero (the derivative of L2 approaches zero). Finally, the weight coefficient W obtained after the regularization constraint is used as the weight coefficient of the formally used online mathematical model.

[0117] The purpose of modeling the online mathematical model is to find the probability that Y is equal to 1 under the given conditions of W and M (W is obtained through the machine learning model). The meaning of Y equal to 1 in this scenario is that the candidate to be recruited is a low-intention candidate.

[0118] See also Figure 2 ,The application phase of the online mathematical model includes the following steps:

[0119] Step S201, obtaining multiple attribute feature values ​​of the candidates to be recruited and weight coefficients corresponding to the multiple attribute feature values ​​online;

[0120] After a lot of training of the machine learning model, the weight coefficient W and constant term t that meet practical requirements can be obtained, and they and multiple attribute feature values ​​of the candidates to be recruited in the recruitment project are substituted into the online mathematical model as input values.

[0121] Step S202: Input the multiple attribute feature values ​​and weight coefficients into the following online mathematical model to obtain the probability y that the candidate to be recruited is a low-intent candidate:

[0122]

[0123] Among them, ω T is the transposed matrix of the vector matrix ω, which is a 1×n column matrix [ω1ω2…ω n ],ω1,ω2,…,ω n Represents the weight coefficient; the vector matrix x is a 1×n matrix [x1x2…x n ],x1,x2,…,x n Represents the attribute characteristic value; ω T x represents ω1x1+ω2x2+…+ω n x n , n is the number of attribute eigenvalues, e is a natural constant, and b is a constant term.

[0124] Specifically, the weight coefficient W obtained through training is used as ω, the constant term t obtained through training is used as the constant term b, and multiple attribute feature values ​​of the candidates to be recruited in the recruitment project are used as x, and the probability y that the candidate to be recruited is a low-intent candidate can be calculated online.

[0125] Step S203: The orders of the candidates to be recruited whose probability y is greater than the set threshold are taken as identification results and sent to the manual customer service system or the telephone robot system.

[0126] The probability y of each candidate for recruitment being a low-intent candidate can be calculated through an online mathematical model. By comparing the probability y with the set threshold, the candidates for recruitment whose probability y is greater than the set threshold can be obtained. For example, if the threshold is set to 0.7, the candidates for recruitment whose probability y is greater than 0.7 are regarded as low-intent candidates. The orders corresponding to the predicted low-intent candidates are sent to the telephone robot system as the recognition results. The telephone robot system can exclude the above orders and only make telephone inquiries for orders with unexpected recognition results, thereby efficiently obtaining high-intent candidate data.

[0127] Furthermore, it also includes:

[0128] Step S204, regularly monitoring the actual delivery of the recognition results; if the difference between the actual delivery of the recognition results and the average delivery of all recruitment candidates is within a preset range, optimizing the online mathematical model.

[0129] After the online mathematical model is deployed, it is necessary to regularly monitor whether its online performance is consistent with expectations. The effect of the model will also change with the change of the basic data distribution. Therefore, it is necessary to monitor in time. Once the effect is lower than the acceptable threshold, the model needs to be adjusted and optimized and redeployed.

[0130] Specifically, telephone inquiries are also conducted on the low-intention candidates for recruitment in the identification results, and the actual delivery of these candidates identified by the online mathematical model is counted. For example, among these identified low-intention candidates for recruitment, nearly 40% are finally selected and can be delivered to the recruiting companies, while 45% of all candidates for recruitment are finally selected and delivered to the recruiting companies. The ratios of 40% and 45% are very close, which shows that the low-intention candidates identified by the online mathematical model are not particularly accurate low-intention candidates, and the recognition ability of the model is poor, which needs to be adjusted and optimized. For another example, among these identified low-intention candidates for recruitment, nearly 10% are finally selected and can be delivered to the recruiting companies, while 60% of all candidates for recruitment are finally selected and delivered to the recruiting companies. There is a significant difference between 10% and 60%, which shows that the low-intention candidates identified by the online mathematical model are relatively accurate, and there is no need to adjust and optimize the online mathematical model.

[0131] Out of prudence, this model adopts a gradual expansion approach in actual business. The first step is to push candidate orders with a predicted value greater than 0.7 to the robot. At present, the actual accuracy rate is 88%, which is better than the model effect. Next, we will consider gradually expanding the scale to increase the coverage ratio to a certain extent and reduce the workload of customer service.

[0132] See also Figure 3 Another embodiment of the present invention further provides a system 300 for identifying low-intention candidates in a recruitment process, comprising a data acquisition module 301, an identification module 302, and an output module 303. The system 300 can execute each step in the above method embodiment.

[0133] Specifically:

[0134] The data acquisition module 301 is configured to acquire online a plurality of attribute characteristic values ​​of the candidates to be recruited and weight coefficients corresponding to the plurality of attribute characteristic values;

[0135] The identification module 302 is configured to input the multiple attribute feature values ​​and weight coefficients into the following online mathematical model to obtain the probability y that the candidate to be recruited is a low-intent candidate:

[0136]

[0137] Among them, ω Tis the transposed matrix of the vector matrix ω, which is a 1×n column matrix [ω1ω2…ω n ],ω1,ω2,…,ω n Represents the weight coefficient; the vector matrix x is a 1×n matrix [x1 x2…x n ],x1,x2,…,x n Represents the attribute characteristic value; ω T x represents ω1x1+ω2x2+…+ω n x n , n is the number of attribute eigenvalues, e is a natural constant, and b is a constant term;

[0138] The output module 303 is configured to send the orders of the candidates to be recruited whose probability y is greater than a set threshold as identification results to the manual customer service system or the telephone robot system.

[0139] Furthermore, it also includes a model generation module 304, which is configured to:

[0140] Collect multiple attribute feature data of the candidates to be recruited, and perform data cleaning on the attribute feature data;

[0141] Perform feature processing on the attribute feature data after data cleaning to obtain a training set consisting of multiple attribute feature value samples;

[0142] Input the attribute feature value sample data in the training set into the following machine learning model:

[0143]

[0144] Among them, Y represents the preset probability that the candidate to be recruited belongs to the low-intention candidate; the vector matrix M is a 1×m matrix [M1 M2…M m ],M1,M2,…,M m Represents the input attribute feature value sample; W T is the transposed matrix of the vector matrix W, which is a 1×m column matrix [W1W2…W m ],W1,W2,…,W m is the weight coefficient of the attribute feature value sample to be solved; W T M means W1M1+W2M2+…+W m M m , m is the number of attribute feature value samples, t is the constant term to be solved, and e is a natural constant;

[0145] Through the machine learning model, the weight coefficient and constant term corresponding to each attribute feature value sample are solved and used as the weight coefficient and constant term of the online mathematical model.

[0146] The data cleaning includes: discarding or supplementing missing attribute feature data; removing abnormal attribute feature data; standardizing the data dimension of the attribute feature data; and / or logarithmically smoothing the attribute feature data that causes the recognition result to show a gradually decreasing trend;

[0147] The feature processing includes: mapping the classification values ​​in the attribute feature data to integer values; removing unimportant attribute feature data; performing feature dimension reduction on the attribute feature data; and / or processing the attribute feature data to obtain derived attribute feature data.

[0148] It also includes a model monitoring module 305, which is configured to regularly monitor the actual delivery of the recognition results. If the difference between the actual delivery of the recognition results and the average delivery of all recruitment candidates is within a preset range, the online mathematical model is optimized.

[0149] It should be noted that the system 300 for identifying low-intention candidates in the recruitment process provided in this embodiment corresponds to a technical solution that can be used to execute each method embodiment, and its implementation principle and technical effect are similar to the method, which will not be repeated here.

[0150] The above description is only a preferred embodiment of the present invention. Those skilled in the art should understand that the disclosure scope involved in the present invention is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present invention (but not limited to) by each other to form a technical solution.

Claims

1. A method for identifying low-intention candidates in a recruitment process, characterized in that: include: Online acquisition of multiple attribute characteristic values ​​that can reflect the job-seeking intention of the candidate to be recruited and weight coefficients corresponding to the multiple attribute characteristic values; wherein the job-seeking intention indicates the possibility that the candidate to be recruited is willing to communicate; the multiple attribute characteristic values ​​include the candidate's own attributes, the candidate's job-seeking intention, the candidate's proactive behavior, the candidate's popularity, the degree of match between the candidate's expectations and the communication position, and the candidate's recent performance in other orders; The multiple attribute feature values ​​and weight coefficients are input into the following online mathematical model to obtain the probability y that the candidate to be recruited is a low-intent candidate: Among them, ω T is the transposed matrix of the vector matrix ω, which is a 1×n column matrix [ω1ω2…ω n ],ω1,ω2,…,ω n Represents the weight coefficient; the vector matrix x is a 1×n matrix [x1 x2…x n ],x1,x2,…,x n Represents the attribute characteristic value; ω T x represents ω1x1+ω2x2+…+ω n x n , n is the number of attribute eigenvalues, e is a natural constant, and b is a constant term; The orders of the candidates to be recruited whose probability y is greater than the set threshold are taken as identification results and sent to the manual customer service system or the telephone robot system; The weight coefficient and the constant term are calculated as follows: Collect multiple attribute feature data of the candidates to be recruited, and perform data cleaning on the attribute feature data; Perform feature processing on the attribute feature data after data cleaning to obtain a training set consisting of multiple attribute feature value samples; Input the attribute feature value sample data in the training set into the following machine learning model: Among them, Y represents the preset probability that the candidate to be recruited belongs to the low-intention candidate; the vector matrix M is a 1×m matrix [M1M2…M m ],M1,M2,…,M m Represents the input attribute feature value sample; W T is the transposed matrix of the vector matrix W, which is a 1×m column matrix [W1W2…W m ],W1,W2,…,W m is the weight coefficient of the attribute feature value sample to be solved; W T M means W1M1+W2M2+…+W m M m , m is the number of attribute feature value samples, t is the constant term to be solved, and e is a natural constant; Through the machine learning model, the weight coefficient and constant term corresponding to each attribute feature value sample are solved; Regularization constraints are applied to the solved weight coefficients through the L1 norm or the L2 norm, and the weight coefficients obtained after the regularization constraints are used as the weight coefficients of the online mathematical model.

2. A method for identifying low-intention candidates in a recruitment process according to claim 1, characterized in that: The data cleaning includes: Discard or supplement missing attribute feature data; Remove abnormal attribute feature data; Standardize the data dimension of attribute feature data; And / or, logarithmic smoothing is performed on the attribute feature data that causes the recognition result to show a gradually decreasing trend.

3. A method for identifying low-intention candidates in a recruitment process according to claim 1, characterized in that: The feature processing includes: Mapping categorical values ​​in attribute feature data to integer values; Remove unimportant attribute feature data; Perform feature dimension reduction on attribute feature data; And / or, the attribute characteristic data is processed to obtain derived attribute characteristic data.

4. The method for identifying low-intention candidates in a recruitment process according to claim 1, characterized in that: It also includes the model monitoring step: Regular monitoring of the actual delivery of said identification results; If the difference between the actual delivery of the recognition result and the average delivery of all recruitment candidates is within a preset range, the online mathematical model is optimized.

5. A system for identifying low-intention candidates in a recruitment process, characterized in that: include: The data acquisition module is configured to acquire online a plurality of attribute characteristic values ​​that can reflect the job-seeking intention of the candidate to be recruited and the weight coefficients corresponding to the plurality of attribute characteristic values; wherein the job-seeking intention indicates the possibility that the candidate to be recruited is willing to communicate; the plurality of attribute characteristic values ​​include the candidate's own attributes, the candidate's job-seeking intention, the candidate's proactive behavior, the candidate's popularity, the degree of match between the candidate's expectations and the communication position, and the candidate's recent performance in other orders; The identification module is configured to input the multiple attribute feature values ​​and weight coefficients into the following online mathematical model to obtain the probability y that the candidate to be recruited is a low-intent candidate: Among them, ω T is the transposed matrix of the vector matrix ω, which is a 1×n column matrix [ω1ω2…ω n ],ω1,ω2,…,ω n Represents the weight coefficient; the vector matrix x is a 1×n matrix [x1 x2…x n ],x1,x2,…,x n Represents the attribute characteristic value; ω T x represents ω1x1+ω2x2+…+ω n x n , n is the number of attribute eigenvalues, e is a natural constant, and b is a constant term; An output module is configured to send the orders of the candidates to be recruited whose probability y is greater than a set threshold as identification results to a manual customer service system or a telephone robot system; The parameter acquisition module is configured to collect multiple attribute feature data of the candidates to be recruited, perform data cleaning on the attribute feature data; perform feature processing on the attribute feature data after data cleaning to obtain a training set consisting of multiple attribute feature value samples; input the attribute feature value sample data in the training set into the following machine learning model, Among them, Y represents the preset probability that the candidate to be recruited belongs to the low-intention candidate; the vector matrix M is a 1×m matrix [M1M2…M m ],M1,M2,…,M m Represents the input attribute feature value sample; W T is the transposed matrix of the vector matrix W, which is a 1×m column matrix [W1W2…W m ],W1,W2,…,W m is the weight coefficient of the attribute feature value sample to be solved; W T M means W1M1+W2M2+…+W m M m , m is the number of attribute feature value samples, t is the constant term to be solved, and e is the natural constant; the weight coefficient and constant term corresponding to each attribute feature value sample are solved through the machine learning model; the solved weight coefficient is regularized and constrained through the L1 norm or the L2 norm, and the weight coefficient obtained after the regularization constraint is used as the weight coefficient of the online mathematical model.

6. A system for identifying low-intention candidates in a recruitment process according to claim 5, characterized in that: The data cleaning includes: discarding or supplementing missing attribute feature data; removing abnormal attribute feature data; standardizing the data dimension of the attribute feature data; and / or logarithmically smoothing the attribute feature data that causes the recognition result to show a gradually decreasing trend; The feature processing includes: mapping the classification values ​​in the attribute feature data to integer values; removing unimportant attribute feature data; performing feature dimension reduction on the attribute feature data; and / or processing the attribute feature data to obtain derived attribute feature data.

7. A system for identifying low-intention candidates in a recruitment process according to claim 5, characterized in that: It also includes a model monitoring module, which is configured to regularly monitor the actual delivery of the recognition results. If the difference between the actual delivery of the recognition results and the average delivery of all recruitment candidates is within a preset range, the online mathematical model is optimized.

Images