Remission prediction analysis method with interpretability
By obtaining multi-dimensional employee data, performing preprocessing and feature engineering, combining mainstream machine learning models and interpretability analysis of Shapash tools, the problem of inability to explain factors affecting employee turnover in the existing technology is solved, and the accuracy of turnover prediction is improved and targeted.
Patent Information
- Application Number
- CN202411348968.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-26
- Publication Date
- 2025-02-18
AI Technical Summary
The existing resignation prediction analysis methods cannot effectively explain the important factors that affect employee resignation, resulting in the inability to carry out targeted talent retention work.
Using an interpretable resignation prediction analysis method, by obtaining data from 8 dimensions of employees, performing data preprocessing and feature engineering, introducing four mainstream machine learning models for prediction, and using Shapash tool for interpretability analysis .
It achieves a good explanation of the reasons for employees' resignation, improves the accuracy of resignation forecasts, and can assist managers in implementing talent retention measures in a targeted manner.
Smart Images
Figure CN119476553A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to an analysis method, and more specifically, to an explainable resignation prediction analysis method, which belongs to the field of computer technology. Background Art
[0002] With the development of big data and artificial intelligence and the increasing importance of talent resources, employee turnover analysis is used to assist in retaining employees, especially core talents. At present, traditional turnover prediction analysis methods only predict the probability of employee turnover, but do not explain the important factors that affect each employee's turnover, so it is impossible to carry out targeted talent retention work. Summary of the invention
[0003] In order to solve the above-mentioned problems in the prior art, the present invention provides an explainable resignation prediction and analysis method having technical characteristics such as being able to assist managers in implementing targeted measures to retain talents.
[0004] In order to achieve the above object, the present invention is implemented by the following technical solutions:
[0005] The present invention provides an explainable resignation prediction analysis method, which comprises the following steps:
[0006] Step S1: Obtain data: Collect data in eight dimensions related to employees, including basic information, education information, organization and work content, position and grade, professional and technical title, performance, and salary;
[0007] Step S2: Data preprocessing: Data preprocessing includes information extraction and data conversion;
[0008] Specifically, the average performance score is obtained from the performance appraisal information; the relative actual salary, the actual salary growth rate, and whether the actual salary growth rate is higher than the average growth level are extracted from the salary information; the main job category of the family members is obtained from the family member information in the basic information; the type of graduation school, whether it is a junior college to undergraduate, and the academic degree code are extracted from the academic degree information in the education information; the branch, level, and line of work are obtained from the work information of the institution, work content, position, rank, and professional and technical title. After this process, the data set is enriched;
[0009] Step S3: Feature engineering: nonlinear features are introduced through feature engineering. In this process, the feature engineering method of variable binning and feature combination is used. The following features are obtained through variable binning: job level classification, line age binning, and performance score binning. The implementation process of feature combination is to combine level, line, job level, line age, education, performance, and marriage to obtain 12 features.
[0010] Step S4: Model training: In the model training stage, four mainstream machine learning models are introduced to predict whether an employee will resign. In the model training stage, the prediction effects of each model are compared through five-fold cross validation and AUC is used as the model evaluation index. Finally, the model with the highest AUC score is selected as the final prediction model for predicting whether an employee will resign.
[0011] Step S5: Explainability analysis: When using the Shapash tool to perform explainability analysis on the prediction results, the prediction results can be explained from multiple dimensions using a variety of charts.
[0012] Most methods are based on Shap analysis, while this method introduces a more advanced Shapash tool. The Shapash tool provides a web application interface that focuses on ease of use and interactivity in model interpretation. Shapash can use a variety of charts (scatter plots, bar charts, kernel density plots, line charts, etc.) to explain the prediction results from multiple dimensions, such as feature importance, single feature analysis, and multi-feature comparative analysis of multiple cases. With the help of the Shapash tool, the key factors that affect the resignation of each employee and the impact of the key factors can be obtained.
[0013] Preferably, the four mainstream machine learning models include logistic regression, random forest, XGBoostClassifier, and CatBoostClassifier.
[0014] Preferably, the data in step S1 contains 67 fields in total, and the time span of the data is 10 years.
[0015] Preferably, step S3 also includes deleting some features: 57 features are deleted through business understanding and correlation coefficients between features, and feature importance is introduced to filter out 4 features that are not helpful in improving the prediction effect of the model.
[0016] Preferably, the specific method for deleting some features is as follows: in the experiment, by training the model, outputting the feature importance of each model, and removing the features with the lowest feature importance in turn, when the AUC value of the model's evaluation index begins to decrease, stop removing features; after the above process, a data set containing 35 features and a target feature of "resignation or not" is obtained.
[0017] Beneficial effects: The method of the present invention uses characteristics of multiple dimensions of employees when predicting employee resignation analysis, which can better explain the reasons for employee resignation and improve the accuracy of predicting whether employees will resign; the key factors that affect the resignation of each employee and the impact of the key factors can be obtained, which can assist managers in implementing targeted measures to retain talents. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 It is the overall flow chart of the present invention.
[0019] Figure 2 This is a case diagram of the interpretability analysis of prediction results performed by the present invention. DETAILED DESCRIPTION
[0020] The present invention will be further described below in conjunction with the accompanying drawings, but the present invention is not limited to the following embodiments.
[0021] like Figure 1-2 The figure shows a specific embodiment of an interpretable resignation prediction analysis method. The embodiment is an interpretable resignation prediction analysis method, and the resignation prediction analysis method includes the following steps:
[0022] Step S1: Obtain data: Collect data in eight dimensions related to employees, including basic information, education information, organization and work content, position and grade, professional and technical title, performance, and salary; the data contains 67 fields in total, and the time span of the data is 10 years;
[0023] Step S2: Data preprocessing: Data preprocessing includes information extraction and data conversion;
[0024] Specifically, the average performance score is obtained from the performance appraisal information; the relative actual salary, the actual salary growth rate, and whether the actual salary growth rate is higher than the average growth level are extracted from the salary information; the main job category of the family members is obtained from the family member information in the basic information; the type of graduation school, whether it is a junior college to undergraduate, and the academic degree code are extracted from the academic degree information in the education information; the branch, level, and line of work are obtained from the work information of the institution, work content, position, rank, and professional and technical title. After this process, the data set is enriched;
[0025] Step S3: Feature engineering: nonlinear features are introduced through feature engineering. In this process, the feature engineering method of variable binning and feature combination is used. The following features are obtained through variable binning: job level classification, line age binning, and performance score binning. The implementation process of feature combination is to combine level, line, job level, line age, education, performance, and marriage to obtain 12 features.
[0026] It also includes the deletion of some features: through business understanding and the correlation coefficient between features, 57 features were deleted, and the feature importance was introduced to screen out 4 features that were not helpful in improving the model prediction effect; the specific method of deleting some features is as follows: in the experiment, by training the model, the feature importance of each model is output, and the features with the lowest feature importance are eliminated in turn. When the AUC value of the model's evaluation indicator begins to decrease, the feature elimination is stopped; after the above process, a data set containing 35 features and a target feature of "resignation or not" is obtained.
[0027] Step S4: Model training: In the model training stage, four mainstream machine learning models are introduced to predict whether an employee will resign (the four mainstream machine learning models include logistic regression, random forest, XGBoostClassifier, and CatBoostClassifier). In the model training stage, the prediction effects of each model are compared through five-fold cross validation and AUC is used as the model evaluation indicator. Finally, the model with the highest AUC score is selected as the final prediction model for predicting whether an employee will resign;
[0028] Step S5: Explainability analysis: When using the Shapash tool to perform explainability analysis on the prediction results, the prediction results can be explained from multiple dimensions using a variety of charts. In the current prior art, most methods are based on Shap for analysis, and this method introduces a more advanced Shapash tool. The Shapash tool provides a web application interface that focuses on the ease of use and interactivity of model interpretation. Shapash can use a variety of charts (scatter plots, bar charts, kernel density plots, line charts, etc.) to explain the prediction results from multiple dimensions, such as feature importance, single feature analysis, and multi-feature comparative analysis of multiple cases. With the help of the Shapash tool, the key factors that affect the resignation of each employee and the size of the impact of the key factors can be obtained.
[0029] Finally, it should be noted that the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by ordinary technicians in this field should be considered as the protection scope of the present invention.
Claims
1. An interpretable resignation prediction analysis method, characterized by The resignation prediction analysis method includes the following steps: Step S1: Obtain data: Collect data in eight dimensions related to employees, including basic information, education information, organization and work content, position and grade, professional and technical title, performance, and salary; Step S2: Data preprocessing: Data preprocessing includes information extraction and data conversion; Specifically, the average performance score is obtained from the performance appraisal information; the relative actual salary, the actual salary growth rate, and whether the actual salary growth rate is higher than the average growth level are extracted from the salary information; the main job category of the family members is obtained from the family member information in the basic information; the type of graduation school, whether it is a junior college to undergraduate, and the academic degree code are extracted from the academic degree information in the education information; the branch, level, and line of work are obtained from the work information of the institution, work content, position, rank, and professional and technical title. After this process, the data set is enriched; Step S3: Feature engineering: nonlinear features are introduced through feature engineering. In this process, the feature engineering method of variable binning and feature combination is used. The following features are obtained through variable binning: job level classification, line age binning, and performance score binning. The implementation process of feature combination is to combine level, line, job level, line age, education, performance, and marriage to obtain 12 features. Step S4: Model training: In the model training stage, four mainstream machine learning models are introduced to predict whether an employee will resign. In the model training stage, the prediction effects of each model are compared through five-fold cross validation and AUC is used as the model evaluation index. Finally, the model with the highest AUC score is selected as the final prediction model for predicting whether an employee will resign. Step S5: Explainability analysis: When using the Shapash tool to perform explainability analysis on the prediction results, the prediction results can be explained from multiple dimensions using a variety of charts.
2. The explainable resignation prediction analysis method according to claim 1, characterized in that: The four mainstream machine learning models include logistic regression, random forest, XGBoostClassifier, and CatBoostClassifier.
3. The explainable resignation prediction analysis method according to claim 1 or 2, characterized in that: The data in step S1 contains 67 fields in total, and the time span of the data is 10 years.
4. The explainable resignation prediction analysis method according to claim 1, characterized in that: Step S3 also includes the deletion of some features: 57 features were deleted through business understanding and the correlation coefficient between features, and feature importance was introduced to screen out 4 features that were not helpful in improving the prediction effect of the model.
5. The explainable resignation prediction analysis method according to claim 4, characterized in that: The specific method of deleting some features is as follows: in the experiment, by training the model, output the feature importance of each model, and remove the features with the lowest feature importance in turn. When the AUC value of the model evaluation index begins to decrease, stop removing features; After the above process, a data set containing 35 features and a target feature of "resignation or not" is obtained.