An insurance salesman intelligent screening method based on decision tree binning and logistic regression

Abstract

The present application relates to a kind of insurance business intelligence screening method based on decision tree binning and logistic regression, comprising: collecting historical data set including resume information, character test, interview evaluation and productivity label and preprocessing to generate standardized feature matrix;Using CART decision tree to perform single variable automatic binning on each continuous feature, generate binning data table and calculate each feature information value to screen key features;After evidence weight coding to key features, train logistic regression model, and according to regression coefficient, each feature binning interval is mapped to scorecard score;The data of the candidate to be evaluated are input into scorecard to obtain each stage and comprehensive score as productivity potential prediction result.Compared with prior art, the present application has the advantages of objective and unified screening standard, strong model interpretability and high prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing and intelligent evaluation technology, and in particular to an intelligent screening method for insurance agents based on decision tree binning and logistic regression. Background Technology

[0002] As sales representatives who directly face customers, the sales capabilities and performance of insurance company agents directly impact the company's market competitiveness and profitability. Therefore, precise candidate screening during the recruitment process is crucial. However, the current traditional recruitment model in the insurance industry suffers from numerous insurmountable drawbacks, severely hindering recruitment quality and efficiency.

[0003] In the resume screening stage, with the rapid development of the insurance industry, the number of applicants has surged, and recruiters often need to process hundreds or even thousands of resumes in a short period of time. Traditional manual screening is not only time-consuming and labor-intensive, but more importantly, it is difficult to accurately identify the core elements that significantly impact salesperson performance from massive amounts of information. For example, candidates with similar sales experience may have different levels of matching between their past sales product types, customer groups, and insurance business, which directly affects their subsequent performance. However, manual screening struggles to quantify this, often resulting in high-potential candidates being overlooked or unsuitable candidates advancing to later stages. In the personality testing stage, existing personality testing tools are mostly generic versions and not customized for the characteristics of insurance salesperson positions. Insurance sales work is characterized by high pressure, high communication frequency, and long-term follow-up, requiring practitioners to have strong resilience, empathy, and a long-term service mentality. However, the evaluation dimensions and scoring standards of generic personality tests cannot be directly linked to these core job requirements. Test results often only reflect general personality traits of candidates, lacking a mapping relationship with insurance performance indicators, making it difficult for interviewers to determine whether candidates are suitable for the position. During the interview evaluation phase, different interviewers exhibit significant differences in their evaluation criteria based on personal experience: some prioritize candidates' eloquence, while others focus on their resilience under pressure, lacking a unified evaluation framework. This inconsistency renders evaluation results incomparable, with the same candidate potentially receiving drastically different evaluations from different interviewers. Furthermore, interviewers' intuitive judgments often dominate the process, lacking a systematic consideration of performance-related factors, further reducing the accuracy of the screening. Moreover, the three screening stages mentioned above are fragmented; the results of resume screening, personality testing, and interview evaluation cannot be effectively linked and comprehensively analyzed, making it difficult to construct a complete profile of a candidate's abilities. This fragmented evaluation method makes it difficult for insurance companies to accurately predict candidates' performance potential during recruitment, resulting in high recruitment costs and significant fluctuations in recruitment quality.

[0004] In summary, existing methods for recruiting and screening insurance agents have technical flaws: resume screening lacks quantitative analysis tools for performance-related factors within massive amounts of information; personality test evaluation dimensions cannot establish a mapping relationship with the core requirements and performance indicators of the insurance agent position; interview evaluation lacks a unified quantitative assessment framework, with evaluation standards varying from person to person; and the data from the three stages of evaluation are fragmented and cannot be analyzed jointly. Therefore, how to uniformly quantify the multi-source, heterogeneous data from the entire process of resume screening, personality testing, and interview evaluation, and construct an interpretable and standardized evaluation model to accurately predict candidates' performance potential, is a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology by providing an intelligent screening method for insurance agents based on decision tree binning and logistic regression. By using CART decision trees to independently perform univariate automatic binning for each continuous feature with minimizing the Gini coefficient as the criterion and sample size proportion as a hard constraint, and combining IV value quantitative screening and VIF collinearity detection, the feature binning boundary and screening criteria are completely data-driven without the need for human experience intervention. This eliminates the bias caused by subjectively setting segmentation thresholds in traditional methods, and achieves objective and accurate quantitative extraction of performance-related factors from massive candidate information.

[0006] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, an intelligent screening method for insurance agents based on decision tree binning and logistic regression is provided, the specific steps of which include: S1. Collect historical datasets, which include resume information, personality test data, interview evaluation data, and corresponding post-employment productivity tags of insurance agents who have already joined the company; preprocess the historical datasets to generate a standardized feature matrix; S2. The CART classification decision tree algorithm is used to perform univariate automatic binning on each continuous feature in the standardized feature matrix, dividing the value of each continuous feature into several bin intervals, and calculating the capacity distribution statistics of each bin interval to generate a binning data table for each feature. S3. Calculate the information value of each feature based on the binning data table, and select key features from the standardized feature matrix according to the information value; encode the key features with evidence weights, and convert each binning interval of each key feature into an evidence weight value. S4. Using the key features encoded with evidence weights as independent variables and the production capacity label as the dependent variable, train a logistic regression model to obtain the regression coefficients of each key feature; based on the regression coefficients of the logistic regression model, map the different binning intervals of each key feature to the corresponding scorecard scores to establish a scorecard. S5. Input the raw data of the candidate to be evaluated into the scoring card, calculate the candidate's scores at each stage and the comprehensive score, and use them as the prediction result of the candidate's production capacity potential.

[0007] Furthermore, in step S1, the specific steps for preprocessing the historical dataset include: The missing rate of each feature dimension is calculated one by one. For feature dimensions with a missing rate exceeding 80%, mean imputation is used to fill in the missing values, which involves replacing the missing values ​​with the average value of the feature in all valid samples. For feature dimensions with a missing rate not exceeding 80%, the missing values ​​are retained and processed separately in subsequent binning steps. Outliers are detected for each continuous feature using box plots. Values ​​exceeding the range of the first quartile minus 1.5 times the interquartile range to the third quartile plus 1.5 times the interquartile range are truncated. The resume information data, personality test data, and interview evaluation data are horizontally concatenated according to candidate identifiers and combined with the corresponding productivity labels to form a standardized dataset.

[0008] Furthermore, in S2, the specific steps of using the CART classification decision tree algorithm to perform univariate automatic binning on each continuous feature in the standardized feature matrix include: For each continuous feature, using the current feature as the sole independent variable and the production capacity label as the target variable, a CART classification decision tree is trained independently. The decision tree algorithm automatically generates the optimal split boundary based on minimizing the Gini coefficient, dividing the value of the feature into K non-overlapping bins. Binning constraints are set, including: the number of bins K for a single feature is limited to 3 to 5 groups, and the sample size of each group is not less than 2% of the total sample size. For each bin obtained after binning, intermediate statistics are calculated, including the sample size of the interval, the proportion of the interval sample to the total sample size, the number of high-production-capacity samples in the interval, the number of low-production-capacity samples in the interval, the proportion of high-production-capacity samples in the interval to the total sample size of the interval, and the proportion of low-production-capacity samples in the interval to the total sample size of the interval. The output is a full-feature binning data table including the binning interval divisions of all features and the quantitative data of production capacity distribution in each interval.

[0009] Furthermore, in step S3, the specific steps of calculating the information value of each feature based on the binning data table and filtering key features according to the information value include: The total number of high-capacity samples and low-capacity samples in the entire historical data are counted separately. For the k-th bin interval of the j-th feature, the evidence weight value of the current interval is calculated based on the ratio of the number of high-capacity samples in the current interval to the total number of high-capacity samples in the entire data, and the ratio of the number of low-capacity samples in the current interval to the total number of low-capacity samples in the entire data. The information value of the feature is calculated by multiplying the difference between the proportion of high-capacity samples and the proportion of low-capacity samples in each bin interval by the corresponding evidence weight value and summing them. Features with information value greater than a preset threshold are selected as candidate features, and the variance inflation factor of the candidate features is calculated to perform multicollinearity detection. Features with variance inflation factor greater than 10 are removed, and the remaining features are retained as key features.

[0010] Furthermore, in step S3, the specific steps for performing evidence weight encoding on the key features include: For the j-th key feature of the i-th sample, the original value of the current feature is replaced with the evidence weight value corresponding to the binning interval according to the binning interval into which the feature value falls. After all key features are encoded, the evidence weight encoded dataset is obtained, and the evidence weight encoded dataset is used as the input of the logistic regression model.

[0011] Further, in step S4, the evidence weight encoded dataset is randomly divided into a training set and a test set according to the production capacity label. Using the key features encoded with evidence weights in the training set as independent variables and the production capacity label as the dependent variable, a logistic regression model is trained using maximum likelihood estimation to obtain the intercept term and the regression coefficients of each key feature. The trained logistic regression model is evaluated using the test set, calculating the accuracy and precision of the model's predictions relative to the actual production capacity label, and calculating the area under the receiver operating characteristic (ROC) curve. When the ROC curve area is greater than or equal to a preset threshold, the model is deemed to have discriminative power. Based on the intercept term of the logistic regression model and the regression coefficients of each key feature, combined with the key features... The evidence weight values ​​for each bin interval are used to calculate the contribution score of each key feature to the total score for each bin interval. For the key feature set corresponding to the resume screening stage, the contribution scores of each feature belonging to the current key feature set are added together to obtain the candidate's resume screening stage score. For the key feature set corresponding to the personality test stage, the contribution scores of each feature belonging to the current set are added together to obtain the candidate's personality test stage score. For the key feature set corresponding to the interview evaluation stage, the contribution scores of each feature belonging to the current set are added together to obtain the candidate's interview evaluation stage score. The resume screening stage score, personality test stage score, and interview evaluation stage score are added together to obtain the candidate's comprehensive score.

[0012] Furthermore, the method further includes a closed-loop verification step, which includes: The actual productivity data of new employees selected and hired through the scoring card is collected after their onboarding. The actual productivity data includes first-year premium income and first-year performance rate. Simultaneously, the scores of the new employees during the recruitment stage, including resume screening, personality test, interview evaluation, and comprehensive scores, are extracted. The scores at each stage are mapped one-to-one with the actual productivity data according to the candidate identifier to construct a score-productivity correspondence database. The correlation coefficient between the comprehensive score and the first-year premium income is calculated. If the correlation coefficient is greater than a preset threshold, the model is deemed to have predictive ability. New employees are divided into several scoring strata according to the comprehensive score range. The average first-year premium and average performance rate of new employees in each stratum are calculated to verify the productivity differences between the strata.

[0013] Furthermore, the method further includes an iterative optimization step, which includes: When the correlation coefficient is lower than a preset threshold or business needs change, collect the original data of newly added candidates and their actual productivity data after joining the company, and add the new data to the historical dataset; re-execute the preprocessing, CART decision tree automatic binning, information value screening key features, evidence weight encoding, logistic regression model training and score card score mapping steps on the updated historical dataset to generate an updated score card; include the newly added feature dimensions that can identify the potential of new employees in the feature collection scope, participate in feature screening and model training when the above steps are re-executed, and update the feature set and score configuration of the score card.

[0014] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0015] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention constructs a unified standardized feature matrix by collecting resume information data, personality test data and interview evaluation data, and uses the CART classification decision tree algorithm to independently perform univariate automatic binning on each continuous feature in the standardized feature matrix. The optimal segmentation boundary of each feature is automatically generated with the Gini coefficient minimization as the criterion and the sample size ratio not less than 2% as a hard constraint. Combined with IV value quantitative screening and VIF collinearity detection, the feature binning process and screening criteria are completely driven by data. There is no need to rely on human experience to set segmentation thresholds and subjective judgment of feature importance. This eliminates the technical defects of arbitrary manual binning boundaries and inconsistent feature screening criteria in traditional methods. It realizes the objective and repeatable quantitative extraction of performance-related factors in massive candidate information, thereby enabling the screening model to accurately focus on the core driving factors that truly affect the productivity of new employees, effectively improving the prediction accuracy of recruitment screening.

[0017] (2) This invention encodes the selected key features using WOE and then inputs them into the logistic regression model for training. The capacity difference information of each bin interval obtained by CART binning is quantified and retained through WOE values. Then, the logistic regression coefficients are automatically mapped to the contribution scores of each feature and each bin interval to the total score. The entire scoring card construction process, from determining the bin boundaries and setting the feature selection threshold to calculating the score, is automatically completed by the algorithm based on historical data. There is no need to manually set weights or subjectively adjust the scores. At the same time, the linear structure of the logistic regression model and the traceability of WOE encoding make the source of each score traceable back to the candidate's original feature value along the chain of "bin interval → WOE value → regression coefficient → contribution score". This not only eliminates the problem of inconsistent evaluation caused by the difference in personal standards of interviewers in traditional interview evaluation, but also solves the defect that black box models cannot explain the scoring basis. This allows interviewers to clearly understand the reasons for the candidate's scores in each dimension, enhancing the credibility and acceptability of the screening results.

[0018] (3) This invention collects the actual productivity data of new employees who have been screened by the scoring card and joined the company, and constructs a one-to-one corresponding scoring and productivity database with the scores of each stage of the recruitment process. It calculates the correlation coefficient between the comprehensive score and the first year's premium income and verifies the productivity differences between each level according to the scoring. When the correlation coefficient is lower than the threshold or the business needs change, the new data will be re-executed to re-execute the full process of mapping from CART binning to scoring card. This enables the model to automatically perceive the decay of the screening effect based on objective productivity feedback and adaptively adjust the binning boundary and feature weights. There is no need to manually redesign the evaluation criteria. This solves the limitation of traditional screening methods where the evaluation criteria are difficult to update with changes in the market environment and business needs once they are determined. This enables the screening model to maintain a long-term match with the actual business and continuously and stably deliver business talents with high productivity potential to insurance companies. Attached Figure Description

[0019] Figure 1 This is a flowchart of an intelligent screening method for insurance agents based on decision tree binning and logistic regression. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] In the recruitment and screening of insurance agents, candidate evaluation data typically comes from three independent channels: resume information, personality tests, and interview evaluations. These data types differ significantly in their data structure, dimensions, and data types, making direct joint analysis and unified quantitative assessment difficult. Current technologies, when processing such multi-source heterogeneous data, often rely on subjective segmentation of continuous features, such as years of work experience and number of work experience segments, before incorporating them into the evaluation system. This approach is highly dependent on personal experience, lacks objective and unified statistical basis for segmentation boundaries, and the segmentation standards for the same feature often differ among different evaluators or in different scenarios, leading to poor comparability and repeatability of subsequent evaluation results. Furthermore, when selecting key factors related to job performance from numerous candidate features, existing methods often rely on qualitative judgments or simple single-factor analysis, lacking a systematic quantitative screening mechanism. This makes it difficult to objectively measure the distinguishing ability and contribution of each feature to productivity results, easily introducing redundant features or omitting truly predictive core features. Furthermore, existing assessment models are mostly black-box structures. While they may possess some predictive ability, they cannot clearly explain to interviewers the specific source and calculation logic of each score. Interviewers struggle to conduct targeted follow-up questions and make comprehensive judgments based on the model's output, reducing the model's usability and credibility in real-world recruitment scenarios. To address these issues, this application constructs a standardized multi-source data acquisition and preprocessing framework to achieve unified quantitative representation of three types of heterogeneous data. It employs CART decision tree automatic binning technology for data-driven optimal segmentation of continuous features and combines this with an IV value quantitative screening mechanism to objectively extract key features. A white-box scoring card is constructed using WOE encoding and logistic regression, ensuring that each score can be traced back to its original feature value. This approach guarantees predictive ability while achieving full transparency and explainability in the assessment process, providing an objective, unified, and practical quantitative decision-making tool for insurance agent recruitment and screening.

[0022] like Figure 1 As shown in this embodiment, an intelligent screening method for insurance agents based on decision tree binning and logistic regression is provided. The specific steps include: S1. Collect historical datasets, which include resume information, personality test data, interview evaluation data, and corresponding post-employment productivity tags of insurance agents who have already joined the company; preprocess the historical datasets to generate a standardized feature matrix. S2. The CART classification decision tree algorithm is used to perform univariate automatic binning on each continuous feature in the standardized feature matrix, dividing the value of each continuous feature into several bin intervals, and calculating the capacity distribution statistics of each bin interval to generate a binning data table for each feature. S3. Calculate the information value of each feature based on the binning data table, and select key features from the standardized feature matrix according to the information value; encode the key features with evidence weights, and convert each binning interval of each key feature into an evidence weight value. S4. Using the key features encoded with evidence weights as independent variables and the production capacity label as the dependent variable, train a logistic regression model to obtain the regression coefficients of each key feature; based on the regression coefficients of the logistic regression model, map the different binning intervals of each key feature to the corresponding scorecard scores to establish a scorecard. S5. Input the raw data of the candidate to be evaluated into the scoring card, calculate the candidate's scores at each stage and the comprehensive score, and use them as the prediction results of the candidate's production capacity potential.

[0023] Before collecting historical datasets in this embodiment, it is necessary to pre-define the field collection specifications for three types of data: resume information data, personality test data, and interview evaluation data. Resume information data comes from electronic resumes submitted by candidates or offline forms, and the values ​​of each field are extracted after structured parsing. Personality test data is automatically generated from candidates completing standardized online assessment questionnaires, with raw scores for each dimension automatically generated. The scores for each dimension in the assessment questionnaire are linearly summed and normalized to a range of 0-100. Interview evaluation data is entered by interviewers through an electronic scoring system after the interview. The preset threshold for the post-employment productivity label is determined based on the insurance company's annual business targets and the historical productivity distribution of each business line; the specific value can be configured by the business unit according to actual needs.

[0024] The resume data includes the candidate's age, gender, education, type of university, major, total years of work experience, years of work experience in the insurance industry, number of banking and insurance work experience segments, percentage of total working hours in the most recent job, type of industry in the most recent job, whether the candidate holds an insurance practitioner qualification certificate, a financial planner certificate, or a financial analyst certificate, and self-evaluation text data; personality test data includes the candidate's assessment scores in the dimensions of efficient execution, customer service, teamwork, stress resistance, and empathy; interview evaluation data includes the interviewer's scores in the dimensions of the candidate's expression ability, logical thinking, stress resistance performance, and appearance and demeanor; the productivity label is determined as high or low productivity based on whether the candidate's premium income in the first year after joining the company exceeds a preset threshold.

[0025] In S1, the specific steps for preprocessing the historical dataset include: The missing rate of each feature dimension is calculated one by one. For feature dimensions with a missing rate exceeding 80%, mean imputation is used to fill in the missing values, which involves replacing the missing values ​​with the average value of the feature in all valid samples. For feature dimensions with a missing rate not exceeding 80%, the missing values ​​are retained and processed separately in subsequent binning steps. Outliers are detected for each continuous feature using box plots. Values ​​exceeding the range of the first quartile minus 1.5 times the interquartile range to the third quartile plus 1.5 times the interquartile range are truncated. The resume information data, personality test data, and interview evaluation data are horizontally concatenated according to candidate identifiers and combined with the corresponding productivity labels to form a standardized dataset.

[0026] Mean imputation is performed independently for each feature dimension. The imputed value is calculated only based on the valid values ​​of the non-missing samples of that feature, without introducing information from other feature dimensions. In the box plot method, the first quartile is the value at the 25th percentile after sorting all valid samples of the current feature in ascending order, the third quartile is the value at the 75th percentile, and the interquartile range is the difference between the third quartile and the first quartile. During truncation, values ​​less than the first quartile minus 1.5 times the interquartile range are reset to this lower bound, and values ​​greater than the third quartile plus 1.5 times the interquartile range are reset to this upper bound. The specific method for horizontally concatenating resume information data, personality test data, and interview evaluation data by candidate identifier is as follows: using the candidate's unique identifier as the connection key, records belonging to the same candidate in the three categories of data are merged into the same feature record. If a candidate has multiple records in a certain category of data, the latest record generated is used for concatenation.

[0027] In S2, the specific steps of using the CART classification decision tree algorithm to perform univariate automatic binning on each continuous feature in the standardized feature matrix include: For each continuous feature, using the current feature as the sole independent variable and the production capacity label as the target variable, a CART classification decision tree is trained independently. The decision tree algorithm automatically generates the optimal split boundary based on minimizing the Gini coefficient, dividing the value of the feature into K non-overlapping bins. Binning constraints are set, including: the number of bins K for a single feature is limited to 3 to 5 groups, and the sample size of each group is not less than 2% of the total sample size. For each bin obtained after binning, intermediate statistics are calculated, including the sample size of the interval, the proportion of the interval sample to the total sample size, the number of high-production-capacity samples in the interval, the number of low-production-capacity samples in the interval, the proportion of high-production-capacity samples in the interval to the total sample size of the interval, and the proportion of low-production-capacity samples in the interval to the total sample size of the interval. The output is a full-feature binning data table including the binning interval divisions of all features and the quantitative data of production capacity distribution in each interval.

[0028] CART classification decision trees use the Gini coefficient as a measure of the purity of node splits. A smaller Gini coefficient indicates higher class purity of samples within a node. During splitting, all possible values ​​of the current feature are traversed as candidate split points. The optimal split boundary is selected as the value with the smallest weighted sum of the Gini coefficients of the child nodes after the split, while also satisfying the sample size constraint. The binning constraint ensures that each bin contains at least 2% of the total sample size to guarantee the stability of statistics within each binning interval and prevent unreliable binning results due to insufficient samples. In the intermediate statistics, the number of high-capacity samples within an interval refers to the number of samples with a capacity label of 1 within that interval, and the number of low-capacity samples within an interval refers to the number of samples with a capacity label of 0 within that interval. The proportion of high-capacity samples within an interval to the total number of samples in that interval is the ratio of the number of high-capacity samples in that interval to the sample size of that interval, and the proportion of low-capacity samples within an interval to the total number of samples in that interval is the ratio of the number of low-capacity samples in that interval to the sample size of that interval. A key benefit of this step is that by automatically determining the splitting boundaries through an algorithm and combining them with statistical significance constraints, the binning results reflect both the inherent patterns of the data and possess statistical reliability, eliminating the subjective arbitrariness of manual partitioning.

[0029] In S3, the specific steps for calculating the information value of each feature based on the binning data table and filtering key features according to the information value include: The total number of high-capacity samples and low-capacity samples in the entire historical data are counted separately. For the k-th bin of the j-th feature, the evidence weight value of the current bin is calculated based on the ratio of the number of high-capacity samples in the current bin to the total number of high-capacity samples in the entire data, and the ratio of the number of low-capacity samples in the current bin to the total number of low-capacity samples in the entire data. The information value of the feature is calculated by multiplying the difference between the proportion of high-capacity samples and the proportion of low-capacity samples in each bin by the corresponding evidence weight value and summing them. Features with information value greater than a preset threshold are selected as candidate features, and the variance inflation factor of the candidate features is calculated to perform multicollinearity detection. Features with a variance inflation factor greater than 10 are removed, and the remaining features are retained as key features.

[0030] In S3, the specific steps for weighted evidence encoding of key features include: For the j-th key feature of the i-th sample, the original value of the current feature is replaced with the evidence weight value corresponding to the binning interval according to the binning interval into which the feature value falls. After all key features are encoded, the evidence weight encoded dataset is obtained, and the evidence weight encoded dataset is used as the input of the logistic regression model.

[0031] Specifically, for the k-th binning interval of the j-th feature, the formula for calculating the evidence weight is based on the distribution difference between high-productivity and low-productivity samples within that interval: when the proportion of high-productivity samples is higher than the proportion of low-productivity samples within that interval, the evidence weight is positive, indicating that candidates falling into that interval are more likely to be high-productivity; conversely, the evidence weight is negative, indicating a greater tendency towards low-productivity. The information value is calculated by weighting and summing the evidence weight values ​​of each binning interval according to the difference between the proportions of high-productivity and low-productivity. A higher information value indicates a more significant difference in the product productivity distribution across different value intervals, and a stronger ability to distinguish product productivity results. The variance inflation factor (VIF) is calculated by performing a linear regression with the current candidate feature as the dependent variable and the remaining candidate features as independent variables, obtaining the determination coefficient R0 of the regression model. 2 VIF equals 1 divided by 1 minus R 2 The difference between the VIF values ​​indicates that the feature is more collinear with other features. A key benefit of this step is that using the VIF value as a unified quantitative screening threshold makes the feature screening process objective, repeatable, and auditable. Different operators performing the same steps on the same dataset will obtain consistent screening results. At the same time, the VIF detection removes strongly collinear features, improving the stability of subsequent logistic regression model coefficient estimation.

[0032] In S4, the evidence weight-coded dataset is stratified and randomly divided into training and test sets according to the production capacity label. Using the key features encoded with evidence weights in the training set as independent variables and the production capacity label as the dependent variable, a logistic regression model is trained using maximum likelihood estimation to obtain the intercept term and regression coefficients of each key feature. The trained logistic regression model is evaluated using the test set, calculating the accuracy and precision of the model's predictions relative to the actual production capacity label, and calculating the area under the receiver operating characteristic (ROC) curve. A model is considered to have discriminative ability when the ROC curve area is greater than or equal to a preset threshold. Based on the intercept term and regression coefficients of each key feature of the logistic regression model, combined with the binning of each key feature... The evidence weight values ​​for each interval are used to calculate the contribution score of each bin interval of each key feature to the total score. For the key feature set corresponding to the resume screening stage, the contribution scores of each feature belonging to the current key feature set are added together to obtain the candidate's resume screening stage score. For the key feature set corresponding to the personality test stage, the contribution scores of each feature belonging to the current set are added together to obtain the candidate's personality test stage score. For the key feature set corresponding to the interview evaluation stage, the contribution scores of each feature belonging to the current set are added together to obtain the candidate's interview evaluation stage score. The resume screening stage score, personality test stage score, and interview evaluation stage score are added together to obtain the candidate's comprehensive score.

[0033] After all P key features are encoded, the original 92-dimensional feature vector of each sample is compressed into a P-dimensional WOE encoded vector. The value of each dimension in this vector is the WOE value of the bin interval where the current value of the corresponding key feature is located. WOE encoding has the function of quantifying the correlation strength between features and production capacity. The sign and magnitude of its value directly reflect the direction and degree of contribution of the bin interval to production capacity.

[0034] The method in this embodiment further includes a closed-loop verification step, which includes: The actual productivity data of new employees selected through a scorecard and hired after onboarding is collected. This data includes first-year premium income and first-year performance rate. Simultaneously, scores from the resume screening, personality test, interview evaluation, and overall performance of new employees during the recruitment phase are extracted. These scores are then mapped one-to-one with the actual productivity data according to candidate identifiers, constructing a score-productivity database. The correlation coefficient between the overall score and first-year premium income is calculated; if the correlation coefficient exceeds a preset threshold, the model is deemed to have predictive power. New employees are divided into several score strata based on their overall score ranges, and the average first-year premium and average performance rate within each stratum are calculated to verify the productivity differences between strata.

[0035] The specific method for stratified random partitioning based on capacity labels is as follows: First, all samples are divided into two subsets based on capacity label values ​​of 0 and 1. Samples are then randomly drawn from each subset in a 7:3 ratio to form the training set and the test set, ensuring that the ratio of high-capacity samples to low-capacity samples in the training and test sets is approximately consistent with the ratio in the full dataset. Accuracy is the ratio of the number of samples in the test set whose predicted labels match the actual labels to the total number of samples in the test set. Precision is the ratio of the number of samples in the test set that were predicted as having high capacity and actually had high capacity to the total number of samples predicted as having high capacity. The area under the curve (AUC) ranges from 0.5 to 1. An AUC of 0.5 indicates that the model has no discriminative ability, while an AUC closer to 1 indicates stronger discriminative ability. A preset threshold of 0.7 is set; that is, when an AUC is greater than or equal to 0.7, the model is considered to have acceptable discriminative ability, and if it is lower than this threshold, the model needs to be readjusted. A key benefit of this step is that by comprehensively evaluating multiple indicators such as accuracy, precision, and AUC, the model can be deployed only after it has a stable ability to identify high-capacity candidates on historical data, thus preventing ineffective models from entering actual business scenarios.

[0036] The method in this embodiment further includes an iterative optimization step, which includes: When the correlation coefficient falls below a preset threshold or business needs change, collect the original data of newly added candidates and their actual productivity data after joining the company, and add the new data to the historical dataset; re-execute the preprocessing, CART decision tree automatic binning, information value screening of key features, evidence weight encoding, logistic regression model training and score card score mapping steps on the updated historical dataset to generate an updated score card; include the newly added feature dimensions that can identify the potential of new employees in the feature collection scope, participate in feature screening and model training when the above steps are re-executed, and update the feature set and score configuration of the score card.

[0037] The scoring system maps scores to a logarithmic linear relationship between scores and probability ratios using pre-defined scoring scale parameters. The probability ratio is the ratio of the probability of high productivity to the probability of low productivity. The scoring scale parameters include the baseline score corresponding to the baseline probability ratio and the number of points the score decreases for each doubling of the probability ratio. The logarithmic probability ratio of the current candidate is calculated from the intercept term and regression coefficients of the logistic regression model, and then substituted into the scoring formula to obtain the comprehensive score. Scores for each stage are calculated based on the feature subsets corresponding to each stage. These feature subsets are pre-defined by business experts before model construction based on the stage to which the data originates. The contribution scores of all key features belonging to the same stage are summed to obtain the score for that stage. The final comprehensive score ranges from 0 to 100 points; a higher score indicates a greater probability that the candidate is predicted to be high-productivity, providing interviewers with an intuitive and quantitative screening basis. A key benefit of this step is that the entire score mapping process, from regression coefficients to scoring scales, is calculated automatically by the algorithm without relying on any manual weight settings, ensuring the objectivity and consistency of the scoring. At the same time, the white-box nature of logistic regression allows interviewers to trace back along the scoring chain to the candidate's original feature values ​​and accurately answer the reasons for the scoring.

[0038] The correlation coefficient is the Pearson correlation coefficient, ranging from -1 to 1. A correlation coefficient greater than a preset threshold and a corresponding p-value less than 0.05 is considered statistically significant; the preset threshold is set to 0.3. Several scoring stratifications are established based on the comprehensive score, divided into five equally spaced intervals: 0-20, 20-40, 40-60, 60-80, and 80-100 points, from low to high. After calculating the average first-year premium and average success rate for newcomers within each stratum, the model's discriminative stability across different scoring intervals is verified by comparing the mean productivity differences between adjacent strata. A key benefit of this step is that by establishing a data link between the score and actual productivity, and using correlation coefficients and stratification comparisons as quantitative verification methods, the model's effectiveness is objectively judged, no longer relying on subjective feelings or cursory observation.

[0039] When collecting new data, the newly added features that identify the potential of new recruits are derived from business feedback from frontline recruiters or from business experts' redefinition of the characteristics of high-productivity new recruits in the current market environment. After being included in the feature collection scope, the new features must be defined using the same data type and value range as the original features to ensure compatibility during subsequent re-execution of preprocessing and binning. When re-executing all steps, the original key features and scorecard scores are completely discarded. The updated dataset independently regenerates the binning intervals, IV ranking results, input feature set, and feature score configurations for all features. A key benefit of this step is that when changes in the business environment cause shifts in the original patterns, simply adding new data samples and re-executing the entire process allows the model to automatically learn the new patterns and adjust binning boundaries and feature weights, eliminating the need for manual redesign of the scoring criteria. This enables the screening model to self-evolve and adapt to changes in business conditions.

[0040] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0041] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0042] Multiple components in the device are connected to an I / O interface, including: input units such as a keyboard, mouse, etc.; output units such as various types of displays, speakers, etc.; storage units such as disks, optical disks, etc.; and communication units such as network interface cards, modems, wireless transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or the communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method of the present invention described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the method of the present invention by any other suitable means (e.g., by means of firmware).

[0043] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0044] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0045] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0046] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent screening of insurance agents based on decision tree binning and logistic regression, characterized in that, The specific steps include: S1. Collect historical datasets, which include resume information, personality test data, interview evaluation data, and corresponding post-employment productivity tags of insurance agents who have already joined the company; preprocess the historical datasets to generate a standardized feature matrix; S2. The CART classification decision tree algorithm is used to perform univariate automatic binning on each continuous feature in the standardized feature matrix, dividing the value of each continuous feature into several bin intervals, and calculating the capacity distribution statistics of each bin interval to generate a binning data table for each feature. S3. Calculate the information value of each feature based on the binning data table, and select key features from the standardized feature matrix according to the information value; encode the key features with evidence weights, and convert each binning interval of each key feature into an evidence weight value. S4. Using the key features encoded with evidence weights as independent variables and the production capacity label as the dependent variable, train a logistic regression model to obtain the regression coefficients of each key feature; based on the regression coefficients of the logistic regression model, map the different binning intervals of each key feature to the corresponding scorecard scores to establish a scorecard. S5. Input the raw data of the candidate to be evaluated into the scoring card, calculate the candidate's scores at each stage and the comprehensive score, and use them as the prediction result of the candidate's production capacity potential.

2. The intelligent screening method for insurance agents based on decision tree binning and logistic regression as described in claim 1, characterized in that, In step S1, the specific steps for preprocessing the historical dataset include: The missing rate of each feature dimension is calculated one by one. For feature dimensions with a missing rate exceeding 80%, mean imputation is used to fill in the missing values, which involves replacing the missing values ​​with the average value of the feature in all valid samples. For feature dimensions with a missing rate not exceeding 80%, the missing values ​​are retained and processed separately in subsequent binning steps. Outliers are detected for each continuous feature using box plots. Values ​​exceeding the range of the first quartile minus 1.5 times the interquartile range to the third quartile plus 1.5 times the interquartile range are truncated. The resume information data, personality test data, and interview evaluation data are horizontally concatenated according to candidate identifiers and combined with the corresponding productivity labels to form a standardized dataset.

3. The intelligent screening method for insurance agents based on decision tree binning and logistic regression according to claim 1, characterized in that, In step S2, the specific steps of using the CART classification decision tree algorithm to perform univariate automatic binning on each continuous feature in the standardized feature matrix include: For each continuous feature, using the current feature as the sole independent variable and the production capacity label as the target variable, a CART classification decision tree is trained independently. The decision tree algorithm automatically generates the optimal split boundary based on minimizing the Gini coefficient, dividing the value of the feature into K non-overlapping bins. Binning constraints are set, including: the number of bins K for a single feature is limited to 3 to 5 groups, and the sample size of each group is not less than 2% of the total sample size. For each bin obtained after binning, intermediate statistics are calculated, including the sample size of the interval, the proportion of the interval sample to the total sample size, the number of high-production-capacity samples in the interval, the number of low-production-capacity samples in the interval, the proportion of high-production-capacity samples in the interval to the total sample size of the interval, and the proportion of low-production-capacity samples in the interval to the total sample size of the interval. The output is a full-feature binning data table including the binning interval divisions of all features and the quantitative data of production capacity distribution in each interval.

4. The intelligent screening method for insurance agents based on decision tree binning and logistic regression according to claim 1, characterized in that, In step S3, the specific steps for calculating the information value of each feature based on the binning data table and filtering key features according to the information value include: The total number of high-capacity samples and low-capacity samples in the entire historical data are counted separately. For the k-th bin interval of the j-th feature, the evidence weight value of the current interval is calculated based on the ratio of the number of high-capacity samples in the current interval to the total number of high-capacity samples in the entire data, and the ratio of the number of low-capacity samples in the current interval to the total number of low-capacity samples in the entire data. The information value of the feature is calculated by multiplying the difference between the proportion of high-capacity samples and the proportion of low-capacity samples in each bin interval by the corresponding evidence weight value and summing them. Features with information value greater than a preset threshold are selected as candidate features, and the variance inflation factor of the candidate features is calculated to perform multicollinearity detection. Features with variance inflation factor greater than 10 are removed, and the remaining features are retained as key features.

5. The intelligent screening method for insurance agents based on decision tree binning and logistic regression according to claim 1, characterized in that, In step S3, the specific steps for performing evidence weight encoding on the key features include: For the j-th key feature of the i-th sample, the original value of the current feature is replaced with the evidence weight value corresponding to the binning interval according to the binning interval into which the feature value falls. After all key features are encoded, the evidence weight encoded dataset is obtained, and the evidence weight encoded dataset is used as the input of the logistic regression model.

6. The intelligent screening method for insurance agents based on decision tree binning and logistic regression according to claim 1, characterized in that, In step S4, the evidence weight encoding dataset is randomly divided into a training set and a test set according to the production capacity label. The key features encoded with evidence weight in the training set are used as independent variables and the production capacity label is used as the dependent variable. The maximum likelihood estimation method is used to train the logistic regression model to obtain the intercept term and the regression coefficients of each key feature. The trained logistic regression model is evaluated using a test set. The accuracy and precision of the model's predictions are calculated against the actual productivity labels. The area under the receiver operating characteristic (ROC) curve is also calculated. If the ROC curve is greater than or equal to a preset threshold, the model is considered to have discriminative ability. Based on the intercept term of the logistic regression model and the regression coefficients of each key feature, combined with the evidence weights of each bin interval for each key feature, the contribution score of each bin interval for each key feature to the total score is calculated. For the key feature set corresponding to the resume screening stage, the contribution scores of each feature belonging to the current key feature set are summed to obtain the candidate's resume screening stage score. For the key feature set corresponding to the personality test stage, the contribution scores of each feature belonging to the current set are summed to obtain the candidate's personality test stage score. For the key feature set corresponding to the interview evaluation stage, the contribution scores of each feature belonging to the current set are summed to obtain the candidate's interview evaluation stage score. The resume screening stage score, personality test stage score, and interview evaluation stage score are summed to obtain the candidate's comprehensive score.

7. The intelligent screening method for insurance agents based on decision tree binning and logistic regression according to claim 1, characterized in that, The method further includes a closed-loop verification step, which includes: The actual productivity data of new employees selected and hired through the scoring card is collected after their onboarding. The actual productivity data includes first-year premium income and first-year performance rate. Simultaneously, the scores of the new employees during the recruitment stage, including resume screening, personality test, interview evaluation, and comprehensive scores, are extracted. The scores at each stage are mapped one-to-one with the actual productivity data according to the candidate identifier to construct a score-productivity correspondence database. The correlation coefficient between the comprehensive score and the first-year premium income is calculated. If the correlation coefficient is greater than a preset threshold, the model is deemed to have predictive ability. New employees are divided into several scoring strata according to the comprehensive score range. The average first-year premium and average performance rate of new employees in each stratum are calculated to verify the productivity differences between the strata.

8. The intelligent screening method for insurance agents based on decision tree binning and logistic regression according to claim 1, characterized in that, The method further includes an iterative optimization step, which includes: When the correlation coefficient is lower than a preset threshold or business needs change, collect the original data of newly added candidates and their actual productivity data after joining the company, and add the new data to the historical dataset; re-execute the preprocessing, CART decision tree automatic binning, information value screening key features, evidence weight encoding, logistic regression model training and score card score mapping steps on the updated historical dataset to generate an updated score card; include the newly added feature dimensions that can identify the potential of new employees in the feature collection scope, participate in feature screening and model training when the above steps are re-executed, and update the feature set and score configuration of the score card.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images