Retirement prediction device, retirement prediction learning device, method, and program

Abstract

The purpose of a disclosed technology is to make employee resignation prediction for assisting in implementing an employee resignation prevention measure at an appropriate timing.　A generation unit (41) generates a feature vector including a plurality of feature quantities from personnel related information sets in a prescribed period of time for each active employee and each resigned employee. A prediction unit (42) predicts an employee resignation probability of an active employee subjected to prediction resigning at each of a plurality of time points in the future by using a feature vector of the active employee subjected to prediction and a prediction model 31 that is obtained by performing learning of feature vectors generated for a plurality of resigned employees as positive examples and feature vectors generated for a plurality of active employees as negative examples and that provides a prediction of, when the feature vector of the active employee is inputted, an employee resignation probability indicating the possibility of the active employee resigning at each of the plurality of time points in the future.

Description

Technical Field

[0001] The disclosed technology relates to a resignation prediction device, a resignation prediction learning device, a resignation prediction method, a resignation prediction learning method, and a program.

Background Art

[0002] With the decline of the working population, the effective job offer ratio has a tendency to increase, and accordingly, the recruitment cost also has a tendency to increase. In a diverse society, there is also an increase in the professionalism of business, and rather than increasing recruitment blindly, it is more cost-effective to spend on education costs to improve the skills of employees and have them perform at the company. However, if employees who have incurred education costs resign in a short period of time, it will be a loss to the company.

[0003] Therefore, in order to prevent the resignation of experienced employees, a method for predicting the resignation of employees has been proposed. This method performs resignation prediction using the relationship between the evaluation and satisfaction of employees and the resignation rate.

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] By the time an employee informs the company of their intention to resign, it tends to be difficult to retain them. Therefore, in order to retain employees, it is necessary to take appropriate measures to prevent resignation when the employee begins to consider leaving, that is, several months before the employee actually resigns.

[0006] However, while conventional technology can predict whether or not an employee will leave in the future, it cannot predict when the likelihood of resignation increases. Therefore, there is a high possibility that the appropriate timing for approaching employees with resignation prevention measures will be missed, ultimately leading to resignation.

[0007] The disclosure technology was developed in light of the above points and aims to provide employee turnover forecasts to help implement turnover prevention measures at the appropriate time. [Means for solving the problem]

[0008] A first aspect of this disclosure is a retirement prediction device comprising: a generation unit that generates a feature vector containing a plurality of features from personnel-related information for a predetermined period of time for both current employees and retirees; a prediction model that has been learned using the feature vectors generated for each of the plurality of retirees as positive examples and the feature vectors generated for each of the plurality of current employees as negative examples, and which, when a feature vector for a current employee is input, predicts the retirement probability indicating the likelihood of the current employee retiring at a plurality of future points in time; and a prediction unit that uses the feature vectors for the current employee to be predicted to predict the retirement probability of the current employee to be predicted at the plurality of future points in time.

[0009] A second aspect of this disclosure is a retirement prediction learning device, comprising: a generation unit that generates a feature vector containing multiple features from personnel-related information for a predetermined period of time for both current employees and retirees; and a learning unit that learns a prediction model that predicts the probability of retirement, indicating the likelihood of a current employee retiring at multiple points in the future, when a feature vector for a current employee is input, using the feature vectors generated for each of the multiple retirees as positive examples and the feature vectors generated for each of the multiple current employees as negative examples.

[0010] A third aspect of this disclosure is a retirement prediction method, wherein a generation unit generates a feature vector containing multiple features from personnel-related information for a predetermined period for each of current employees and retirees, and a prediction unit uses a prediction model, which has been learned with the feature vectors generated for each of the multiple retirees as positive examples and the feature vectors generated for each of the multiple current employees as negative examples, to predict the retirement probability of the current employee at the multiple future points in time when a feature vector for a current employee is input, and the feature vectors for the current employee to be predicted, to predict the retirement probability of the current employee at the multiple future points in time.

[0011] A fourth aspect of this disclosure is a retirement prediction learning method, wherein a generation unit generates a feature vector containing multiple features from personnel-related information for a predetermined period for each of current employees and retirees, and a learning unit learns a prediction model that predicts the probability of retirement, indicating the likelihood of a current employee retiring at multiple points in the future, when a feature vector for a current employee is input, using the feature vector generated for each of the multiple retirees as a positive example and the feature vector generated for each of the multiple current employees as a negative example.

[0012] A fifth aspect of this disclosure is a program that causes a computer to function as a component of the retirement prediction device or retirement prediction learning device described above. [Effects of the Invention]

[0013] According to the disclosure technology, it is possible to predict employee turnover to help implement turnover prevention measures at the appropriate time. [Brief explanation of the drawing]

[0014] [Figure 1] This is a block diagram showing the hardware configuration of the retirement prediction processing unit. [Figure 2] This is a block diagram showing an example of the functional configuration of a retirement prediction processing system. [Figure 3]It is a diagram for explaining personnel-related information. [Figure 4] It is a diagram for explaining the period of personnel-related information used for learning. [Figure 5] It is a diagram for explaining an example of the generation of feature quantities. [Figure 6] It is a diagram for explaining the period of personnel-related information used for static items and statistical items. [Figure 7] It is a diagram showing an example of the statistical processing of attendance data. [Figure 8] It is a diagram for explaining the time-series prediction of the retirement probability. [Figure 9] It is a diagram showing an example of the time-series prediction result of the retirement probability. [Figure 10] It is a diagram showing an example of the contribution degree for each feature quantity. [Figure 11] It is a diagram showing an example of a correspondence table. [Figure 12] It is a diagram for explaining the similarity of the contribution degrees of feature quantities between current employees and retirees. [Figure 13] It is a diagram showing an example of a list of calculation results of the contribution degrees of feature quantities. [Figure 14] It is a diagram showing an example of a prediction result list. [Figure 15] It is a flowchart showing the flow of learning processing. [Figure 16] It is a flowchart showing the flow of prediction processing. [Figure 17] It is a diagram showing an overview of learning data, prediction data, and prediction results. [Figure 18] It is a diagram showing an example of the verification result of the prediction accuracy of the prediction model in this embodiment. [Figure 19] It is a diagram for explaining the similarity of feature quantities between current employees and retirees.

Mode for Carrying Out the Invention

[0015] An example of an embodiment of the disclosed technology will be described below with reference to the drawings. In each drawing, identical or equivalent components and parts are given the same reference numerals. Furthermore, the dimensional ratios in the drawings are exaggerated for illustrative purposes and may differ from actual ratios.

[0016] Figure 1 is a block diagram showing the hardware configuration of the retirement prediction processing device 10 according to this embodiment. As shown in Figure 1, the retirement prediction processing device 10 has a CPU (Central Processing Unit) 11, ROM (Read Only Memory) 12, RAM (Random Access Memory) 13, storage 14, input unit 15, display unit 16, and communication I / F (interface) 17. Each component is connected to the others so as to be able to communicate with each other via a bus 19.

[0017] The CPU 11 is a central processing unit that executes various programs and controls various parts. Specifically, the CPU 11 reads a program from the ROM 12 or storage 14 and executes the program using the RAM 13 as a working area. The CPU 11 controls each of the above components and performs various calculations according to the program stored in the ROM 12 or storage 14. In this embodiment, the ROM 12 or storage 14 stores a retirement prediction learning program for executing the learning process described later, and a retirement prediction program for executing the prediction process described later.

[0018] ROM12 stores various programs and data. RAM13 temporarily stores programs or data as a working area. Storage14 consists of storage devices such as HDD (Hard Disk Drive) and SSD (Solid State Drive) and stores various programs, including the operating system, and various data.

[0019] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used for various types of input. The display unit 16 is, for example, a liquid crystal display and displays various types of information. The display unit 16 may also function as the input unit 15 by employing a touch panel system.

[0020] Communication I / F17 is an interface for communicating with other devices. This communication may utilize wired communication standards such as Ethernet® or FDDI, or wireless communication standards such as 4G, 5G, or Wi-Fi®.

[0021] Next, the functional configuration of the retirement prediction processing device 10 will be described. Figure 2 is a block diagram showing an example of the functional configuration of the retirement prediction processing device 10. As shown in Figure 2, the retirement prediction processing device 10 includes a retirement prediction learning unit 20 and a retirement prediction unit 40 as its functional configuration. The retirement prediction learning unit 20 further includes a filter unit 21, a generation unit 22, and a learning unit 23. The retirement prediction unit 40 further includes a generation unit 41, a prediction unit 42, and an interpretation unit 43. Note that the retirement prediction learning unit 20 is an example of the retirement prediction learning device of the present invention, and the retirement prediction unit 40 is an example of the retirement prediction device of the present invention. Each functional configuration is realized by the CPU 11 reading the retirement prediction learning program and retirement prediction program stored in the ROM 12 or storage 14, expanding them in the RAM 13, and executing them.

[0022] First, let's explain the employee turnover prediction learning unit 20. The employee turnover prediction learning unit 20 is a functional configuration that operates during the training of the prediction model 31, which will be described later.

[0023] The filter unit 21 retrieves personnel-related information of current employees from the current employees database 51 and personnel-related information of former employees from the former employees database 52. The personnel-related information includes data categories such as basic information, department, salary, attendance, goal setting, and evaluation. Each of the basic information, department, salary, attendance, goal setting, and evaluation categories further includes information on each item, as shown in Figure 3, for example. In the example in Figure 3, the "Statistical Processing" indicator also shows whether or not statistical processing is performed when generating features from the information of each data category.

[0024] Each of the current employee database 51 and the former employee database 52 stores personnel-related information from a predetermined start date (e.g., at the time of hiring) to the present or at the time of retirement, associated with the personal code which is the identification information of each current employee and former employee, at predetermined intervals (e.g., one month). The former employee database 52 also stores, for each former employee, the date of retirement and the reason for retirement obtained through interviews or other means at the time of retirement. Furthermore, the former employee database 52 also stores, for each former employee, a feature vector generated from the personnel-related information by the generation unit 22, which will be described later.

[0025] The filter unit 21 extracts personnel-related information for the period used to generate feature vectors for training the predictive model 31 from the acquired personnel-related information for all periods of each employee, both former and current employees. Specifically, for former employees, the filter unit 21 extracts personnel-related information from a point in time up to one period prior to the time of retirement. For current employees, the filter unit 21 extracts personnel-related information from a point in time up to two periods prior to the point in time up to one period prior to the current time. Multiple periods can be set as this second period.

[0026] For example, let's consider the case where personnel-related information for training a prediction model 31 to predict the probability of an employee leaving N months later is extracted from monthly personnel-related information from the month of joining the company to the most recent month or the month of leaving. For employees who have left, as shown in the upper diagram of Figure 4, the filter unit 21 discards personnel-related information for N months prior to the month of leaving, and extracts personnel-related information from N months prior to the month of joining as the personnel-related information to be used for training. Similarly, for current employees, as shown in the lower diagram of Figure 4, the filter unit 21 discards personnel-related information for N months + X months prior to the month of leaving, and extracts personnel-related information from N months prior to the month of leaving to the month of joining as the personnel-related information to be used for training.

[0027] We will now explain why, for current employees, we will not use personnel-related information for X months in addition to the N months preceding the most recent month. Simply excluding N months from the most recent month would result in a model being trained to predict the probability of resignation in the month corresponding to the most recent month. In other words, leaked features would be included in the features generated from the extracted personnel-related information. For example, if the most recent month is December and N=3, simply excluding N months from the most recent month would mean using personnel-related information up to September for all current employees. For example, if a feature called "number of holidays" is generated from attendance data included in the personnel-related information using statistical processing described later, the value of the "number of holidays" feature for all current employees would be the same, resulting in a leaked feature.

[0028] Therefore, in addition to N months of personnel-related information going back from the most recent month, X months of personnel-related information will also be left unused. Note that X can be any number greater than or equal to 0, and will be assigned to each current employee. The assignment of X to current employees can be random, or it can be correlated with, for example, the length of service. In the latter case, by setting a small value for X if the length of service is short and a large value for X if the length of service is long, it is possible to use as much personnel-related information as possible without causing gaps due to the N+X month period being longer than the period from the month of joining the company to the most recent month.

[0029] The filter unit 21 sets the first period to N months, N+1 months, ..., N+L months (where L is any integer) and extracts personnel-related information corresponding to each first period. The filter unit 21 then passes the extracted personnel-related information for each period for retired and current employees to the generation unit 22.

[0030] The generation unit 22 generates a feature vector containing multiple features from the personnel-related information received from the filter unit 21. For example, for items in the personnel-related information whose values ​​have not changed during the period extracted by the filter unit 21, or items whose values ​​change irregularly (hereinafter referred to as "static items"), the generation unit 22 generates features by using the item values ​​as they are, or by converting the item values ​​to categorical variables. Also, for example, for items in the personnel-related information whose values ​​are numerical and change periodically (hereinafter referred to as "statistical items"), the generation unit 22 generates features by statistically processing the values ​​for the most recent predetermined period.

[0031] More specifically, as shown in Figure 5, the generation unit 22 generates features by sequentially applying processing to each item included in the personnel-related information according to the data category to which the item belongs and the type of value (numerical, categorical variable, text, etc.). In the example in Figure 5, the generation unit 22 first removes outliers from the values ​​of each item based on predetermined rules. Next, the generation unit 22 performs a data matching process, for example, when there is multiple personnel-related information for the same person. Next, the generation unit 22 performs statistical processing (details will be described later) on numerical statistical items such as salary, attendance, goals, and evaluations. Next, the generation unit 22 standardizes items with numerical values ​​by converting them so that the values ​​are in the range of 0 to 1. Next, for items with numerical values, it performs group standardization processing, for example, by grouping people with the same employment category and then standardizing the values ​​on a group basis. Next, the generation unit 22 one-hot encodes the values ​​of items with categorical variables. Next, the generation unit 22 segments the text items into individual words and weights each word using an index (e.g., TF-IDF) that reflects the frequency of word occurrences in all personnel-related information. Then, if there are missing values ​​for any item, the generation unit 22 imputes those missing values ​​using the mean, mode, etc., of that item.

[0032] Note that not all of the processes shown in Figure 5 are mandatory. For example, outlier removal, group standardization, and missing value imputation may be omitted. Also, the order in which some processes are performed may be changed.

[0033] Here, we will explain the scope of use of personnel-related information for static and statistical items, respectively. For example, for static items such as years of service, qualification level, department, employment category, and performance details, the entire period (shaded area) extracted by the filter unit 21 is used, as shown in Figure 6. For items whose values ​​have changed within the period, a feature may be generated by combining those multiple values, or a feature may be generated with a flag indicating that the value has changed set for the latest value.

[0034] For statistical items whose values ​​fluctuate monthly, such as salary and attendance, the most recent M months from the period extracted by the filter unit 21 are used, as shown in Figure 6. That is, for employees who have left the company, M months are used, starting from N months prior to the month of their departure; for current employees, M months are used, starting from N+X months prior to the month of their departure. Note that items such as goals and evaluations often change, for example, annually or semi-annually, so the period for using personnel-related information may not be limited to M months, but the entire period may be used.

[0035] Figure 7 shows an example of statistical processing of the "total working hours" item in attendance data. The upper part of Figure 7 shows the data for the total working hours per day for M months (in the example in Figure 7, M=3). The generation unit 22 performs statistical processing on these M months of total working hours to generate features by calculating the minimum, maximum, sum, mean, median, standard deviation, etc. for each month.

[0036] The generation unit 22 generates a feature vector for each current employee and former employee, using the multiple feature quantities generated as described above as its elements. The generation unit 22 also generates a feature vector for each piece of personnel-related information extracted for the first period, which is set to N months, N+1 months, ..., N+L months (where L is any integer). The generation unit 22 passes the generated feature vectors to the learning unit 23. The generation unit 22 also stores the feature vectors of former employees in the former employee DB 52.

[0037] The learning unit 23 uses the feature vectors passed from the generation unit 22 as training data to learn a prediction model 31 that predicts the probability of an employee leaving the company at multiple points in the future, given that feature vectors for current employees are input. Specifically, the learning unit 23 learns an N-month-later

[0038] The learning unit 23 similarly learns about each of the N+1 month retirement prediction models 31N+1, ..., which predict the probability of retirement after N+1 months, and the N+L month retirement prediction model 31N+L which predicts the probability of retirement after N+L months. The learning unit 23 stores the prediction model 31, which consists of the N month retirement prediction model 31N, the N+1 month retirement prediction model 31N+1, ..., and the N+L month retirement prediction model 31N+L, in a predetermined memory area of ​​the retirement prediction processing device 10.

[0039] Furthermore, the learning unit 23 generates an interpretation model 32 based on the prediction model 31, which calculates the contribution of each feature included in the feature vector to the prediction result of the prediction model 31. For example, the learning unit 23 generates the interpretation model 32 using SHAP (SHapley Additive exPlanations). The learning unit 23 stores the generated interpretation model 32 in a predetermined memory area of ​​the retirement prediction processing device 10.

[0040] Next, the retirement prediction unit 40 will be described. The retirement prediction unit 40 is a functional configuration that operates when predicting the probability of retirement and the reasons for retirement for current employees who are the target of the prediction.

[0041] The generation unit 41 acquires personnel-related information of the current employee to be predicted, which is input to the retirement prediction processing device 10. The personnel-related information of the current employee to be predicted is the same as the personnel-related information described in the retirement prediction learning unit 20. Similar to the generation unit 22 of the retirement prediction learning unit 20, the generation unit 41 generates feature vectors from the personnel-related information of the current employee to be predicted. In this case, the generation unit 41 uses personnel-related information for the period going back from the most recent month.

[0042] The prediction unit 42 predicts the probability of a current employee leaving the company at multiple future points in time by inputting the characteristic vectors of the target employee into the prediction model 31. Specifically, as shown in the upper part of Figure 8, the prediction unit 42 inputs the characteristic vectors of the target employee into each of the prediction model 31: the N-month retirement prediction model 31N, the N+1-month retirement prediction model 31N+1, ..., and the N+L-month retirement prediction model 31N+L. As a result, the prediction unit 42 obtains the retirement probability at N months, the retirement probability at N+1 months, ..., and the retirement probability at N+L months.

[0043] In this way, by predicting the probability of employee turnover at multiple points in the future, it becomes possible to implement turnover prevention measures in a timely manner. For example, based on the prediction results, it becomes possible to determine that an employee may have few signs of leaving in the short term, but whose probability of leaving will increase if the current situation continues, and who will show signs of leaving in the long term, thus requiring early turnover prevention measures. Also, for example, if the probability of leaving is trending upward in three months, it is possible to implement turnover prevention measures such as gradually reducing workload over three months, thereby preventing employee turnover.

[0044] Furthermore, the prediction unit 42 may create a graph showing the time-series change in the probability of resignation from the probability of resignation after N months, the probability of resignation after N+1 months, ..., and the probability of resignation after N+L months, as shown in the lower part of Figure 8. Figure 9 shows the time-series change in the predicted probability predicted in March, with N=6 and L=5, i.e., the probability of resignation from 6 months later (September) to 11 months later (February). In the example in Figure 9, a threshold for determining whether the probability of resignation is high or low is also shown. By outputting such a graph, it is possible to make judgments such as, "The probability of resignation has risen sharply in January, so some kind of resignation prevention measures are needed before then." The prediction unit 42 passes the characteristic vector of the current employee to be predicted, as well as the probability of resignation after N months, the probability of resignation after N+1 months, ..., and the probability of resignation after N+L months, to the interpretation unit 43.

[0045] The interpretation unit 43 uses the prediction model 31, the feature vector of the current employee to be predicted, and the interpretation model 32 to calculate the contribution of each feature included in the feature vector to the prediction result. If the contribution of all features were presented to a person in charge, such as a human resources staff member or a manager of the current employee, it would be difficult for the person in charge to review if there are many features. Therefore, the interpretation unit 43 presents the features whose contribution is equal to or greater than a predetermined value, or the top predetermined number of features in descending order of contribution, as the basis for the prediction result that the prediction model 31 predicted.

[0046] Figure 10 shows an example of the contribution of each feature. In the example in Figure 10, the horizontal axis represents the contribution (0.0 to 1.0), and the numbers in parentheses for each feature are the values ​​of that feature. By checking the values ​​of features with high contribution, the person in charge can determine the reasons for predicting the departure of current employees who have a high probability of leaving.

[0047] For example, in the example in Figure 10, the person in charge can see that the employee being predicted to leave has not taken any special leave such as year-end / New Year holidays or summer holidays, and that their total working hours a month ago were relatively low, indicating that they tend to take time off. The person in charge can also see that the employee being predicted to leave worked a shift that started at 2 PM, from which they can infer that the employee is not a morning person or has morning errands. Furthermore, the person in charge can see that the employee being predicted to leave was late twice three months ago, with a total lateness of 272 minutes, from which they can infer that the employee is experiencing health problems. The person in charge can also see that the employee has been with the company for a relatively long time, more than three years. Furthermore, the person in charge can see that the employee's total working hours fluctuate significantly, and that they were late three times a month ago, indicating that their absences and lateness are increasing. By considering all of these factors, the person in charge can determine that the predicted reason for the employee's resignation is "poor health".

[0048] Furthermore, even if only the highly contributing features and their values ​​are considered, the content may be difficult for humans to understand. Therefore, by converting the combination of features and their values ​​into natural language for interpretation, human interpretation becomes easier even for those unfamiliar with machine learning. Specifically, the interpretation unit 43 refers to a correspondence table 33 that stores the correspondence between the combination of features and their values ​​and the natural language text that interprets that combination, and converts the highly contributing features, i.e., the basis for the prediction, into natural language.

[0049] Figure 11 shows an example of the correspondence table 33. In the example in Figure 11, the feature name, the feature value, the natural language for interpretation, and the reason for predicted retirement are stored in association with each other. The interpretation unit 43 obtains the natural language for interpretation and the reason for predicted retirement from the correspondence table 33, corresponding to the combination of the highly contributing features extracted using the interpretation model 32 and the value of those features.

[0050] Furthermore, the interpretation unit 43 extracts a predetermined number of retirees from among multiple retirees, based on whether the similarity between the contribution of each feature of the retiree and the contribution of each feature of the current employee being predicted is above a predetermined value, or in order of highest similarity. Specifically, the interpretation unit 43 obtains feature vectors of multiple retirees from the retiree DB 52, and as shown in Figure 12, uses the feature vector of each retiree and the interpretation model 32 to calculate the contribution of each feature for each retiree, and lists the calculation results. Figure 13 shows an example of a list of the calculation results of the contributions. The interpretation unit 43 creates a contribution vector with the contribution of each feature of the current employee being predicted and multiple retirees as elements, and calculates the Euclidean distance, etc., of the contribution vectors between the current employee being predicted and the retiree as the similarity. Then, the interpretation unit 43 sorts the retirees in descending order of similarity with the contribution vector of the current employee being predicted, that is, in descending order of the distance between the contribution vectors, and extracts a predetermined number of retirees based on whether the similarity is above a predetermined value, or in order of highest similarity.

[0051] The interpretation unit 43 retrieves the reasons for resignation stored in the resignation database 52 for the extracted resignations. Alternatively, the interpretation unit 43 may identify the feature with the highest contribution for the extracted resignations from a list such as that shown in Figure 13, and retrieve the reasons for resignation corresponding to the combination of that feature and its value from the correspondence table 33. This is based on the idea that features with a high contribution among resignations whose feature contribution is similar to that of the current employee being predicted are likely to have influenced the reasons for resignation. In the example in Figure 13, the interpretation unit 43 retrieves the reasons for resignation corresponding to feature 1 of person Y, whose feature contribution is the highest among the current employee being predicted.

[0052] The interpretation unit 43 creates a prediction result list containing the information interpreted as described above. Figure 14 shows an example of a prediction result list. Figure 14 is an example of a list of prediction results for current employees whose probability of resignation after N months, as predicted by the prediction unit 42, is above a predetermined value. The prediction result list in Figure 14 includes the "personal code" of the current employee in question, along with the "probability of resignation after N months," "predicted reason for resignation," "top K basis for prediction," and "past employees with similar trends who have resigned."

[0053] "Probability of resignation in N months" is the probability of resignation in N months, as passed from the prediction unit 42. "Resignation prediction reason" is the resignation prediction reason obtained from the correspondence table 33 based on the combination of the most contributing features and their values. "Top K prediction basis" is the text converted into natural language by referring to the correspondence table 33, using the top K features (K=3 in the example in Figure 14) in order of contribution and their values. In the example in Figure 14, the standardized contribution values ​​of each feature are also included as reference information, so that the sum of the contribution values ​​of the K features equals 1. "Past resignations with similar trends" are the personal codes of resignations extracted based on the similarity of the contribution values ​​of the features, and the reasons for their resignation. In the example in Figure 14, only the information of one resignation is listed, but information of multiple resignations may be listed. By listing the prediction results and interpreted information in this way, current employees can be filtered by resignation probability and resignation reason, allowing for quick searching of current employees who require focused attention.

[0054] The interpretation unit 43 further generates similar lists of predicted results for N+1 months, ..., and N+L months. This allows for flexible use, such as current employees who need immediate action checking the list of predicted results for N months, and current employees who need long-term action checking the list of predicted results for N+L months.

[0055] The interpretation unit 43 outputs the created list of prediction results. The interpretation unit 43 may also output a graph showing the time series of the retirement probability described above. By presenting the retirement probability at multiple future points in time, along with the basis for the prediction, in an easily interpretable format, it becomes possible to understand why the current employees being predicted are considering retirement, and to implement appropriate retirement prevention measures.

[0056] Next, the operation of the retirement prediction processing device 10 will be explained. Figure 15 is a flowchart showing the flow of the learning process by the retirement prediction processing device 10. The CPU 11 reads the retirement prediction learning program from the ROM 12 or storage 14, loads it into the RAM 13, and executes it, thereby performing the learning process. Note that the learning process is an example of the retirement prediction learning method of the present invention.

[0057] In step S11, the CPU 11, acting as a filter unit 21, retrieves personnel-related information of current employees from the current employees DB 51 and personnel-related information of retired employees from the retired employees DB 52.

[0058] Next, in step S12, the CPU 11, acting as a filter unit 21, extracts personnel-related information for former employees from a point in time up to one period prior to their retirement date. The filter unit 21 also extracts personnel-related information for current employees from a point in time up to two periods prior to the point in time up to one period prior to the current date.

[0059] Next, in step S13, the CPU 11, acting as a generation unit 22, performs various processes such as statistical processing and conversion to categorical variables on each item of the personnel-related information of current and former employees extracted in step S12, to generate a feature vector containing multiple features.

[0060] Next, in step S14, the CPU 11, as the learning unit 23, trains a prediction model 31 using the feature vectors of retired employees generated in step S13 as positive examples and the feature vectors of current employees as negative examples. Next, in step S15, the CPU 11, as the learning unit 23, generates an interpretation model 32 based on the trained prediction model 31, which calculates the contribution of each feature included in the feature vectors to the prediction result of the prediction model 31, and the learning process ends.

[0061] Figure 16 is a flowchart showing the flow of prediction processing by the retirement prediction processing device 10. The CPU 11 reads the retirement prediction program from the ROM 12 or storage 14, loads it into the RAM 13, and executes it to perform the prediction processing. Note that this prediction processing is an example of the retirement prediction method of the present invention.

[0062] In step S21, the CPU 11, acting as a generation unit 41, acquires personnel-related information of the current employee to be predicted, which is input to the retirement prediction processing device 10. Next, in step S22, the CPU 11, acting as a generation unit 41, generates feature vectors from the personnel-related information of the current employee to be predicted.

[0063] Next, in step S23, the CPU 11, acting as the prediction unit 42, inputs the feature vectors of the employees to be predicted into the prediction model 31 to predict the probability of the employees to retire at multiple future points in time.

[0064] Next, in step S24, the CPU 11, acting as an interpreter 43, uses the prediction model 31, the feature vectors of the current employees and retirees to be predicted, and the interpreter 32 to calculate the contribution of each feature included in the feature vector to the prediction result. Next, in step S25, the CPU 11, acting as an interpreter 43, refers to the correspondence table 33 to obtain the natural language for interpretation and the reason for retirement prediction corresponding to the combination of the feature with the highest contribution among the features of the current employees to be predicted and the value of that feature.

[0065] Next, in step S26, the CPU 11, acting as the interpretation unit 43, extracts a predetermined number of former employees from among multiple former employees, based on whether the similarity between the contribution of each feature of the former employee and the contribution of each feature of the current employee being predicted is above a predetermined value, or in order of the highest similarity. Next, in step S27, the CPU 11, acting as the interpretation unit 43, retrieves the reasons for retirement stored in the former employee DB 52 for the extracted former employees.

[0066] Next, in step S28, the CPU 11, as the interpretation unit 43, creates a prediction result list that includes the predicted retirement probability calculated in step S23 for current employees whose retirement probability after the first period is equal to or greater than a predetermined value. The CPU 11, as the interpretation unit 43, also includes the natural language and retirement prediction reasons obtained in step S25, the personal codes of the retirees extracted in step S26, and the retirement reasons of the retirees obtained in step S27 in the prediction result list. The CPU 11, as the interpretation unit 43, outputs the created prediction result list, and the prediction process ends.

[0067] As described above, the retirement prediction unit of the retirement prediction processing device according to this embodiment generates a feature vector containing multiple features from personnel-related information for a predetermined period for both current employees and retirees. The prediction model is trained using the feature vectors generated for each of the multiple retirees as positive examples and the feature vectors generated for each of the multiple current employees as negative examples. When a feature vector for a current employee is input, the prediction model predicts the retirement probability indicating the likelihood of a current employee retiring at multiple future points in time. The prediction unit then uses the feature vector for the target current employee to predict the retirement probability at multiple future points in time. This enables retirement prediction to support the implementation of retirement prevention measures at the appropriate time.

[0068] Furthermore, according to the retirement prediction unit of the retirement prediction processing device according to this embodiment, as shown in Figure 17A, some employees who were currently employed at the time of learning may be predicted to have a possibility of retirement in N months at the time of prediction. Also, as shown in Figure 17B, the retirement probability and reason for retirement are predicted for currently employed employees who were not employed at the time of learning.

[0069] Furthermore, the employee turnover prediction unit may output the probability of employee turnover at multiple future points in time as a graph showing the time-series change in the probability of turnover. This allows for an appropriate determination of when to implement measures to prevent employee turnover.

[0070] Furthermore, the employee turnover prediction unit may use a prediction model, a feature vector of the current employee being predicted, and an interpretation model that calculates the contribution of each feature to the prediction result of the prediction model. The unit may then present features whose contribution is greater than or equal to a predetermined value, or the top predetermined number of features in descending order of contribution, as the basis for the prediction result made by the prediction model. This allows for an understanding of the basis for predicting the probability of employee turnover, and enables the consideration of appropriate employee turnover prevention measures.

[0071] Furthermore, the employee turnover prediction unit may extract and present a predetermined number of employees from among multiple employees who have left the company, based on whether the similarity between the feature vector of the employee who left and the feature vector of the employee being predicted is above a predetermined value, or in order of highest similarity, or based on whether the similarity between the contribution of the feature of the employee who left and the contribution of the feature of the employee being predicted is above a predetermined value, or in order of highest similarity. This allows for the consideration of appropriate employee turnover prevention measures by referring to information on employees who have similar tendencies to the employee being predicted. In particular, when using the similarity of the contribution of feature quantities, it is assumed that the features that led to the employee's departure are similar, allowing for a more appropriate assessment of the reasons for departure and the consideration of more appropriate employee turnover prevention measures.

[0072] Furthermore, the employee turnover prediction unit may present pre-stored reasons for departure for the extracted employees. This allows for the appropriate anticipation of reasons for departure and the consideration of more appropriate employee turnover prevention measures.

[0073] Furthermore, the retirement prediction unit may refer to a correspondence table that stores combinations of features and reference values ​​for those features, along with natural language text interpreting those combinations, and then convert the features to be presented into natural language. This allows personnel, managers, and other staff members who are not familiar with machine learning to easily understand the basis for the predictions made by the prediction model.

[0074] Furthermore, the employee retirement prediction unit may refer to a correspondence table in which combinations of features and reference values ​​of those features are further associated with the reasons for employee retirement that the combinations represent, and then present the reasons for employee retirement corresponding to the features being presented. This makes it easy to understand the reasons for employee retirement corresponding to the features that form the basis of the prediction.

[0075] Furthermore, the employee turnover prediction unit may generate feature vectors by either keeping the feature values ​​as they are or converting them into categorical variables for items in the personnel-related information that include multiple items, where the values ​​do not change over a predetermined period, or where the values ​​change irregularly. For items in the personnel-related information where the values ​​are numerical and change regularly, the unit may generate feature vectors by statistically processing the values ​​for the most recent predetermined period. This allows for the generation of features from static items such as qualifications and departments that can learn from past experience, and features from statistical items such as attendance and salary that enable learning that takes recent circumstances into account.

[0076] Furthermore, the retirement prediction learning unit of the retirement prediction processing device according to this embodiment generates a feature vector containing multiple features from personnel-related information for a predetermined period for both current employees and retirees. Using the feature vectors generated for each of the multiple retirees as positive examples and the feature vectors generated for each of the multiple current employees as negative examples, the unit learns a prediction model that predicts the probability of retirement, indicating the likelihood of current employees retiring at multiple points in the future, when a feature vector for current employees is input. This makes it possible to generate a prediction model that can perform retirement predictions to support the implementation of retirement prevention measures at the appropriate time.

[0077] Furthermore, the retirement prediction learning unit may use the following HR-related information to generate features for training a prediction model that predicts the probability of retirement from the current time to the first period after retirement: for retired employees, HR-related information from the time one period prior to retirement is extracted; for current employees, HR-related information from the time two periods prior to the time one period prior to retirement is extracted; multiple periods are set as the second period; and the extracted HR-related information is used as HR-related information for a predetermined period to generate feature vectors. This makes it possible to suppress the generation of identical features based on the values ​​of items common to all current employees, which would result in leak features.

[0078] Here, Figure 18 shows an example of the verification results of the prediction accuracy of the prediction model in this embodiment. In Figure 18, the recall and precision for each employment category of current employees are shown as verification results for the prediction model that predicts the probability of resignation after 6 months. Note that A to H in Figure 18 are data that has been masked to represent employment categories such as part-time workers and full-time employees. Although there are differences for each employment category, high recall rates are obtained for certain employment categories. Even in employment category G, which has the lowest recall rate, it can be seen that approximately one in four current employees can be identified as showing signs of resignation.

[0079] <Variation> In the above embodiment, the case of predicting the probability of resignation after N months was explained using the example of an integer greater than 0 for N, but any number can be set, such as N=1.5.

[0080] Furthermore, in the above embodiment, we described a case where the similarity of the contribution of features is used when extracting retired employees who are similar to the current employees being predicted. However, as shown in Figure 19, the similarity of feature vectors may also be used.

[0081] Furthermore, although the above embodiment describes a case where a prediction model and an interpretation model are provided separately, a machine learning model capable of calculating the importance of features, such as LightGBM, may be used as the prediction model.

[0082] Furthermore, while the above embodiment describes a case where the probability of resignation and information interpreting the prediction results, such as the reasons for resignation, are output together, it is also possible to output only the probability of resignation.

[0083] Furthermore, although the above embodiment describes a case where the employee retirement prediction learning unit and the employee retirement prediction unit are implemented on a single computer (employee retirement prediction processing unit), they may also be implemented on different computers.

[0084] Furthermore, the retirement prediction learning process and retirement prediction process, which the CPU reads and executes in the above embodiment, may also be executed by various processors other than the CPU. Examples of such processors include PLDs (Programmable Logic Devices) such as FPGAs (Field-Programmable Gate Arrays) whose circuit configuration can be changed after manufacturing, and dedicated electrical circuits, which are processors with circuit configurations specifically designed to execute specific processes, such as ASICs (Application Specific Integrated Circuits). The retirement prediction learning process and retirement prediction process may also be executed by one of these various processors, or by a combination of two or more processors of the same or different types (for example, multiple FPGAs, and a combination of a CPU and an FPGA). More specifically, the hardware structure of these various processors is an electrical circuit that combines circuit elements such as semiconductor elements.

[0085] Furthermore, although the above embodiment describes a configuration in which the retirement prediction learning program and the retirement prediction program are pre-stored (installed) in ROM 12 or storage 14, the invention is not limited thereto. The program may be provided in a form stored on a non-transitory storage medium such as a CD-ROM (Compact Disk Read Only Memory), DVD-ROM (Digital Versatile Disk Read Only Memory), and USB (Universal Serial Bus) memory. Alternatively, the program may be provided in a form that can be downloaded from an external device via a network.

[0086] The following additional information is disclosed regarding the embodiments described above.

[0087] (Additional note 1) Memory and At least one processor connected to the memory, Includes, The aforementioned processor, A feature vector containing multiple features is generated from personnel-related information for a specified period for both current and former employees. A predictive model trained with feature vectors generated for each of multiple retirees as positive examples and feature vectors generated for each of multiple current employees as negative examples, wherein when feature vectors for current employees are input, the predictive model predicts the probability of the current employee retiring at multiple future points in time, and the predictive model uses the feature vectors for the current employee to be predicted to predict the probability of the current employee to retire at multiple future points in time. A retirement prediction device configured in such a way.

[0088] (Additional note 2) Memory and At least one processor connected to the memory, Includes, The aforementioned processor, A feature vector containing multiple features is generated from personnel-related information for a specified period for both current and former employees. The model learns a predictive model that, given the feature vectors for current employees, predicts the probability of a current employee leaving at multiple points in the future, using the feature vectors generated for each of multiple former employees as positive examples and the feature vectors generated for each of multiple current employees as negative examples. A retirement prediction learning device configured in such a way.

[0089] (Additional note 3) A non-temporary recording medium that stores a program executable by a computer to perform retirement prediction processing, The aforementioned retirement prediction process is: A feature vector containing multiple features is generated from personnel-related information for a specified period for both current and former employees. A predictive model trained with feature vectors generated for each of multiple retirees as positive examples and feature vectors generated for each of multiple current employees as negative examples, wherein when feature vectors for current employees are input, the predictive model predicts the probability of the current employee retiring at multiple future points in time, and the predictive model uses the feature vectors for the current employee to be predicted to predict the probability of the current employee to retire at multiple future points in time. A non-temporary recording medium that includes the following.

[0090] (Additional note 4) A non-temporary recording medium that stores a program executable by a computer to perform retirement prediction learning processing, The aforementioned retirement prediction learning process is, A feature vector containing multiple features is generated from personnel-related information for a specified period for both current and former employees. The model learns a predictive model that, given the feature vectors for current employees, predicts the probability of a current employee leaving at multiple points in the future, using the feature vectors generated for each of multiple former employees as positive examples and the feature vectors generated for each of multiple current employees as negative examples. A non-temporary recording medium that includes the following. [Explanation of Symbols]

[0091] 10 Retirement prediction processing device 11 CPU 12 ROM 13 RAM 14 Storage 15 Input section 16 Display section 17 Communication I / F 19 bus 20 Retirement Prediction Learning Department 21 Filter section 22 Generation part 23 Learning Department 31 Predictive Models 32 Interpretation Models 33 Correspondence Table 40 Retirement Forecast Department 41 Generation part 42 Prediction Section 43 Interpretation Section 51 Current Officials Database 52 Retired Employees Database

Claims

1. A filter unit that, as personnel-related information used to generate features for training a predictive model that predicts the probability of retirement from the present time to the first period after retirement, extracts personnel-related information for retired employees from a point in time up to the first period before the time of retirement, and for current employees, extracts personnel-related information for a point in time up to the second period before the time up to the first period before the time of current employees, and sets a plurality of periods as the second period, A generation unit that generates a feature vector containing multiple feature quantities from personnel-related information for a predetermined period for each of the aforementioned current employees and former employees, the generation unit that generates the feature vector using the personnel-related information extracted by the filter unit as personnel-related information for the predetermined period, A prediction model trained using feature vectors generated for each of multiple retirees as positive examples and feature vectors generated for each of multiple current employees as negative examples, wherein when feature vectors for current employees are input, the prediction model predicts the probability of a current employee retiring at multiple future points in time; and a prediction unit that uses the feature vectors for the current employee to be predicted to predict the probability of the current employee to retire at multiple future points in time. Retirement prediction device including

2. The retirement prediction device according to claim 1, which includes an interpretation unit that uses the prediction model, the feature vector of the current employee to be predicted, and an interpretation model that calculates the contribution of each feature to the prediction result of the prediction model, to present a predetermined number of features whose contribution is equal to or greater than a predetermined value, or the top predetermined number of features in order of their contribution, as the basis for the prediction result that the prediction model predicted.

3. The retirement prediction device according to claim 1, which includes an interpretation unit that uses the prediction model, the feature vectors of each of the plurality of retirees, the feature vectors of the current employees to be predicted, and an interpretation model that calculates the contribution of each feature to the prediction result of the prediction model, to extract and present a predetermined number of retirees from the plurality of retirees in order of the similarity between the contribution of the feature of the retiree and the contribution of the feature of the current employees to be predicted, which is equal to or greater than a predetermined value.

4. A filter unit that, as personnel-related information used to generate features for training a predictive model that predicts the probability of retirement from the present time to the first period after retirement, extracts personnel-related information for retired employees from a point in time prior to the first period before retirement, and for current employees, extracts personnel-related information for a point in time prior to the second period from a point in time prior to the first period before retirement, and sets a plurality of periods as the second period, A generation unit that generates a feature vector containing multiple feature quantities from personnel-related information for a predetermined period for each of the aforementioned current employees and former employees, the generation unit that generates the feature vector using the personnel-related information extracted by the filter unit as personnel-related information for the predetermined period, A learning unit learns a prediction model that, when a feature vector for a current employee is input, predicts the probability of the current employee leaving at multiple points in the future, using the feature vectors generated for each of the multiple former employees as positive examples and the feature vectors generated for each of the multiple current employees as negative examples. A retirement prediction learning device that includes this.

5. The filter unit extracts personnel-related information used to generate features for learning a predictive model that predicts the probability of retirement from the present time to the first period later, for retired employees, personnel-related information from the time of retirement to the time before the first period earlier, and for current employees, personnel-related information from the time before the first period earlier to the time before the second period earlier, and sets a plurality of periods as the second period. The generation unit generates a feature vector containing multiple feature quantities from personnel-related information for a predetermined period for each of the current employees and the former employees, and generates the feature vector using the personnel-related information extracted by the filter unit as personnel-related information for the predetermined period. The prediction unit uses the following to predict the retirement probability of the target current employee at the aforementioned future points in time: the prediction model, which has been learned by using the feature vectors generated for each of the multiple retirees as positive examples and the feature vectors generated for each of the multiple current employees as negative examples, and which, when a feature vector for a current employee is input, predicts the retirement probability indicating the likelihood of the current employee retiring at multiple future points in time; and the feature vector for the target current employee. Methods for predicting retirement.

6. The filter unit extracts personnel-related information used to generate features for training a predictive model that predicts the probability of retirement from the present time to the first period later, for retired employees, personnel-related information from the time of retirement to the time before the first period earlier, and for current employees, personnel-related information from the time before the first period earlier to the time before the second period earlier, and sets a plurality of periods as the second period. The generation unit generates a feature vector containing multiple feature quantities from personnel-related information for a predetermined period for each of the current employees and the former employees, and generates the feature vector using the personnel-related information extracted by the filter unit as personnel-related information for the predetermined period. The generation unit generates a feature vector containing multiple features from personnel-related information for a predetermined period for both current and retired employees. The learning unit uses the feature vectors generated for each of the multiple former employees as positive examples and the feature vectors generated for each of the multiple current employees as negative examples to learn the prediction model, which predicts the probability of a current employee leaving the company at multiple points in the future, given that a feature vector for a current employee is input. A learning method for predicting retirement.

7. A program for causing a computer to function as one of the components constituting the retirement prediction device described in any one of claims 1 to 3, or as one of the components constituting the retirement prediction learning device described in claim 4.

Images