WO2024024116A1

WO2024024116A1 - Employee resignation prediction device, employee resignation prediction learning device, method, and program

Info

Publication number: WO2024024116A1
Application number: PCT/JP2022/029396
Authority: WO
Inventors: 哲哉塩田; 方邦石井; 央倉沢; 奏山本
Original assignee: 日本電信電話株式会社
Priority date: 2022-07-29
Filing date: 2022-07-29
Publication date: 2024-02-01

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

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

The disclosed technology relates to a retirement prediction device, a retirement prediction learning device, a retirement prediction method, a retirement prediction learning method, and a program.

With the decline in the working population, the effective job openings-to-applicants ratio is on the rise, and recruitment costs are also on the rise. In a diversifying society, work is becoming more specialized, and it is more cost-effective to spend training costs to improve the skills of employees and have them demonstrate their performance in the company, rather than hiring more people blindly. . However, if an employee who has spent training costs resigns in a short period of time, it will be a loss to the company.

Therefore, in order to prevent experienced employees from resigning, methods have been proposed to predict whether or not employees will resign. This method predicts retirement using the relationship between employee evaluation and satisfaction level and turnover rate.

It tends to be difficult to retain employees by the time they notify the company of their desire to retire. Therefore, in order to have employees stay with the company, it is necessary to take appropriate measures to prevent employees from quitting at the timing when they start thinking about resigning, that is, several months before they retire.

However, although conventional technology predicts whether or not a person will retire in the future, it does not predict when the possibility of retirement will increase. Therefore, there is a high possibility that the appropriate timing to approach employees to prevent them from quitting may be missed, resulting in them resigning.

The disclosed technology was developed in view of the above points, and aims to predict retirement to support taking measures to prevent retirement at an appropriate time.

A first aspect of the present disclosure is a retirement prediction device, which includes a generation unit that generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of each of a current employee and a retiree; A predictive model that is trained using the feature vectors generated for each of the incumbents as positive examples and the feature vectors generated for each of the multiple incumbents as negative examples, and when the feature vectors for the incumbents are input, Using the prediction model that predicts the retirement probability indicating the possibility that the incumbent will retire at multiple points in the future, and the feature vector for the incumbent to be predicted, A prediction unit that predicts retirement probabilities at a plurality of points in time.

A second aspect of the present disclosure is a retirement prediction learning device, which includes a generation unit that generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of each of an incumbent and a retiree; The feature vectors generated for each of the incumbents are taken as positive examples, and the feature vectors generated for each of the multiple incumbents are taken as negative examples.If the feature vectors for the incumbents are input, and a learning unit that learns a predictive model that predicts a retirement probability indicating the possibility that an incumbent will retire.

A third aspect of the present disclosure is a retirement prediction method, in which the generation unit generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of each of the incumbent and the retiree, and the prediction unit , a predictive model trained using feature vectors generated for each of a plurality of retirees as positive examples and feature vectors generated for each of a plurality of incumbents as a negative example, where the feature vectors for the incumbents are input. The prediction model predicts the retirement probability indicating the possibility that the incumbent will retire at multiple points in the future when the incumbent is retired, and the feature vector for the incumbent The probability of retirement of a person at the plurality of future points is predicted.

A fourth aspect of the present disclosure is a retirement prediction learning method, in which the generation unit generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of each of the current employee and the retiree, and the learning unit The feature vectors generated for each of multiple retirees are taken as positive examples, and the feature vectors generated for each of multiple incumbents are taken as negative examples, and when the feature vectors for incumbents are input, future future A predictive model is learned that predicts a retirement probability indicating the likelihood that the incumbent will retire at multiple points in time.

A fifth aspect of the present disclosure is a program that causes a computer to function as each component of the retirement prediction device or retirement prediction learning device described above.

According to the disclosed technology, it is possible to predict retirement to support taking measures to prevent retirement at an appropriate time.

FIG. 2 is a block diagram showing the hardware configuration of a retirement prediction processing device. FIG. 2 is a block diagram showing an example of a functional configuration of a retirement prediction processing device. FIG. 3 is a diagram for explaining personnel-related information. FIG. 3 is a diagram for explaining the period of personnel-related information used for learning. FIG. 3 is a diagram for explaining an example of generation of feature amounts. FIG. 3 is a diagram for explaining the period of personnel-related information used for static items and statistical items. It is a figure showing an example of statistical processing of attendance data. FIG. 2 is a diagram for explaining time-series prediction of retirement probability. FIG. 3 is a diagram showing an example of a time-series prediction result of retirement probability. FIG. 3 is a diagram illustrating an example of the degree of contribution for each feature amount. It is a figure which shows an example of a correspondence table. FIG. 3 is a diagram for explaining the degree of similarity in contribution of feature amounts between an incumbent and a retired person. FIG. 7 is a diagram illustrating an example of a list of calculation results of contribution degrees of feature amounts. It is a figure showing an example of a prediction result list. 3 is a flowchart showing the flow of learning processing. It is a flowchart which shows the flow of prediction processing. FIG. 2 is a diagram schematically showing learning data, prediction data, and prediction results. It is a figure which shows an example of the verification result of the prediction accuracy of the prediction model in this embodiment. FIG. 3 is a diagram for explaining the degree of similarity of feature amounts between an incumbent and a retired person.

Hereinafter, an example of an embodiment of the disclosed technology will be described with reference to the drawings. In addition, the same reference numerals are given to the same or equivalent components and parts in each drawing. Furthermore, the dimensional ratios in the drawings are exaggerated for convenience of explanation and may differ from the actual ratios.

FIG. 1 is a block diagram showing the hardware configuration of a retirement prediction processing device 10 according to the present embodiment. As shown in FIG. 1, the retirement prediction processing device 10 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input section 15, a display section 16, and communication It has an I/F (interface) 17. Each configuration is communicably connected to each other via a bus 19.

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

The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a work area. The storage 14 is constituted by a storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores various programs including an operating system and various data.

The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to perform various inputs. The display unit 16 is, for example, a liquid crystal display, and displays various information. The display section 16 may employ a touch panel system and function as the input section 15.

The communication I/F 17 is an interface for communicating with other devices. For this communication, for example, a wired communication standard such as Ethernet (registered trademark) or FDDI, or a wireless communication standard such as 4G, 5G, or Wi-Fi (registered trademark) is used.

Next, the functional configuration of the retirement prediction processing device 10 will be explained. FIG. 2 is a block diagram showing an example of the functional configuration of the retirement prediction processing device 10. As shown in FIG. 2, the retirement prediction processing device 10 includes a retirement prediction learning section 20 and a retirement prediction section 40 as functional configurations. The retirement prediction learning section 20 further includes a filter section 21, a generation section 22, and a learning section 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 section 20 is an example of a retirement prediction learning device of the present invention, and the retirement prediction section 40 is an example of a retirement prediction device of the present invention. Each functional configuration is realized by the CPU 11 reading out a retirement prediction learning program and a retirement prediction program stored in the ROM 12 or the storage 14, loading them onto the RAM 13, and executing them.

First, the retirement prediction learning section 20 will be explained. The retirement prediction learning unit 20 is a functional configuration that functions when learning a prediction model 31, which will be described later.

The filter unit 21 acquires personnel-related information of the incumbent from the incumbent DB 51 and acquires personnel-related information of the retiree from the retiree DB 52. The personnel-related information includes information on data categories such as basic information, affiliation, salary, attendance, goal setting, and evaluation. The information such as basic information, affiliation, salary, attendance, goal setting, and evaluation further includes information on each item as shown in FIG. 3, for example. The example in FIG. 3 also shows "statistical processing presence/absence" indicating whether statistical processing is to be performed when generating feature amounts from the information of the items of each data category.

In each of the current employee DB 51 and retiree DB 52, personnel-related information from a predetermined start point (for example, when joining the company) to the present time or the time of retirement is stored in association with a personal code that is the identification information of each of the current employee and retiree. Information is stored every predetermined period (eg, one month). In addition, the retiree DB 52 also stores, for each retiree, the date of retirement and the reason for retirement obtained through an interview or the like at the time of retirement. Furthermore, the retiree DB 52 also stores, for each retiree, a feature vector generated from personnel-related information by a generation unit 22, which will be described later.

The filter unit 21 extracts the personnel-related information for the period used to generate the feature vector for learning the prediction model 31 from the acquired personnel-related information for the entire period of each of the retirees and current employees. Specifically, for retirees, the filter unit 21 extracts personnel-related information prior to a first period before the time of retirement. In addition, for the current employee, the filter unit 21 extracts personnel-related information from a point in time that is a first period before the present time and a point in time that is a second period before the present time. A plurality of periods are set as this second period.

For example, a case will be described in which personnel-related information for learning the prediction model 31 for predicting the probability of retirement N months later is extracted from personnel-related information for each month from the month of employment to the latest month or month of retirement. For retirees, as shown in the upper diagram of FIG. 4, the filter unit 21 discards personnel-related information for N months from the month of retirement as unused, out of personnel-related information from the month of employment to the month of retirement. Extract personnel-related information from the previous month up to the month of joining the company as personnel-related information to be used for learning. In addition, for the current employee, as shown in the lower part of FIG. shall be used. Then, the filter unit 21 extracts personnel-related information from N months +

For the current employee, we will explain the reason why personnel-related information for X months is not used in addition to N months' worth of personnel information going back from the latest month. If N months from the latest month are simply not used, a model that predicts the probability of retiring in the month corresponding to the latest month will be learned. That is, the leaked feature amount is included in the feature amount generated from the extracted personnel-related information. For example, if the latest month is December and N=3, if N months from the latest month are simply not used, all personnel related information up to September will be used for the current employee. For example, when a feature quantity called "number of holidays" is generated from attendance data included in personnel-related information by statistical processing described later, the value of the feature quantity called "number of holidays" for all incumbents will be the same, As a result, it becomes a leak feature.

Therefore, in addition to N months' worth of personnel-related information going back from the latest month, X months' worth of personnel-related information is also not used. Note that a plurality of arbitrary numbers greater than or equal to 0 are prepared for X, and allocated to each incumbent. The allocation of X to incumbents may be random, or may be correlated to years of service, for example. In the latter case, if the number of years of service is short, set X to a small value, and if the number of years of service is long, set X to a large value to eliminate the loss caused by the period of N+X months being larger than the period from the month of joining to the latest month. It is possible to utilize as much personnel-related information as possible without causing any problems.

The filter unit 21 sets the first periods to N months, N+1 months, . . . , N+L months (L is an arbitrary integer) and extracts personnel-related information according to each first period. The filter unit 21 passes the personnel-related information for the period extracted for each of retirees and current employees to the generation unit 22.

The generation unit 22 generates a feature vector including a plurality of feature quantities from the personnel-related information passed from the filter unit 21. For example, the generation unit 22 generates information about items whose values do not change during the period extracted by the filter unit 21 or items whose values change irregularly (hereinafter referred to as “static items”) among personnel-related information. generates feature amounts using item values as they are or by converting item values into categorical variables. For example, for items of personnel-related information whose values are numerical and whose values are changed regularly (hereinafter referred to as "statistical items"), the generation unit 22 statistically calculates the values for the most recent predetermined period. The feature amount is generated by processing.

More specifically, as shown in FIG. 5, the generation unit 22 generates data for each item included in the personnel-related information according to the data category to which the item belongs and the type of value (numeric value, categorical variable, text, etc.). Generate features by sequentially applying processing. In the example of FIG. 5, the generation unit 22 first removes abnormal values from the values of each item based on a predetermined rule. Next, the generation unit 22 executes a name matching process, for example, when a plurality of pieces of personnel-related information exist for the same person. Next, the generation unit 22 performs statistical processing (details will be described later) for numerical statistical items such as salary, attendance, goals, and evaluation. Next, the generation unit 22 standardizes items whose values are numerical values by converting them so that the values are in the range of 0 to 1. Next, for items whose values are numerical values, for example, people with the same employment category are grouped together, and then group standardization processing is executed to standardize the values on a group-by-group basis. Next, the generation unit 22 performs one-hot encoding on the items whose values are categorical variables. Next, the generation unit 22 separates the text items into words, and weights each word using an index (for example, TF-IDF) regarding the appearance rate of the word in all personnel-related information. Next, if there is a missing value for each item, the generation unit 22 complements the missing value using the average value, mode, etc. of the item.

Note that not all of the processes shown in FIG. 5 are essential. For example, outlier removal, group standardization, and missing value interpolation may be omitted. Further, some of the processes may be executed with the order changed.

Here, the scope of use of personnel-related information for each of static items and statistical items will be explained. For example, for static items such as years of employment, qualification grade, department, employment category, and performance details, the entire period (shaded area) extracted by the filter section 21 is used, as shown in FIG. For items whose values have changed within the period, you can generate a feature by combining multiple values, or you can create a feature by setting a flag to the latest value to indicate that the value has changed. may be generated.

As shown in FIG. 6, for statistical items such as salary and attendance whose values change every month, the most recent M months of the period extracted by the filter unit 21 are used. That is, for retirees, the amount is M months from N months from the month of retirement, and for current employees, it is M months from N+X months from the month of retirement. Note that items such as goals and evaluations are often changed, for example, every fiscal year or every half year, so the period for using personnel-related information is not limited to M months, and even if the entire period is used. good.

FIG. 7 shows an example of statistical processing of the item "total working hours" of attendance data. The upper diagram of FIG. 7 shows data on the total working hours for each day of each month for M months (M=3 in the example of FIG. 7). The generation unit 22 performs statistical processing to obtain monthly minimum values, maximum values, totals, averages, medians, standard deviations, etc. for the total working hours for M months to generate feature quantities.

The generation unit 22 generates a feature vector whose elements are the plurality of feature amounts generated as described above for each of the incumbent and retired person. Furthermore, the generation unit 22 generates a feature vector for each piece of personnel-related information extracted by setting the first period to N months, N+1 months, . . . , N+L months (L is an arbitrary integer). The generation unit 22 passes the generated feature vector to the learning unit 23. Further, the generation unit 22 stores the retiree's feature vector in the retiree DB 52.

The learning unit 23 uses the feature vectors received from the generating unit 22 as learning data, and when the feature vectors about the incumbent are input, the learning unit 23 generates retirement information indicating the possibility of the incumbent retiring at multiple points in the future. A prediction model 31 that predicts probabilities is learned. Specifically, the learning unit 23 uses a feature vector generated from personnel-related information extracted with the first period as N months to create a retirement prediction model 31N after N months that predicts the probability of retirement after N months. learn. At this time, the learning unit 23 uses the feature vectors generated for each of the plurality of retirees as a positive example (“retirement probability correct answer = 1”), and the feature vector generated for each of the plurality of incumbents as a negative example (“retirement probability correct answer = 1”). "Correct answer for retirement probability = 0"). That is, the learning unit 23 outputs the retirement prediction model 31N after N months when the feature vector of the positive example is inputted is close to 1, and when the feature vector of the negative example is inputted, the output of the retirement prediction model 31N after N months is input. The parameters of the N-month retirement prediction model 31N are learned so that the output of 31N becomes close to 0.

The learning unit 23 similarly learns each of the N+1 month later retirement prediction models 31N+1, . . The learning unit 23 stores the prediction model 31 composed of the N months later retirement prediction model 31N, the N+1 months later retirement prediction model 31N+1, . . . , the N+L months later retirement prediction model 31N+L in a predetermined storage area of the retirement prediction processing device 10. to be memorized.

Furthermore, the learning unit 23 generates an interpretation model 32 based on the prediction model 31, which calculates the degree of 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 storage area of the retirement prediction processing device 10.

Next, the retirement prediction unit 40 will be explained. The retirement prediction unit 40 is a functional configuration that functions when predicting the probability of retirement and the reason for retirement for an incumbent to be predicted.

The generation unit 41 acquires the personnel-related information of the current employee to be predicted, which is input into the retirement prediction processing device 10. The personnel-related information of the incumbent who is the prediction target is the same as the personnel-related information described in the retirement prediction learning section 20. Similar to the generation unit 22 of the retirement prediction learning unit 20, the generation unit 41 generates a feature vector from the personnel-related information of the incumbent who is the prediction target. At this time, the generation unit 41 uses personnel-related information for a period dating back from the latest month.

The prediction unit 42 predicts the retirement probability of the incumbent to be predicted at multiple points in the future by inputting the feature vector of the incumbent to be predicted to the prediction model 31. Specifically, as shown in the upper diagram of FIG. 8, the prediction unit 42 converts the feature vector of the incumbent to be predicted into the N months later retirement prediction model 31N and the N+1 months later retirement prediction model that constitute the prediction model 31. 31N+1, . . . , N+L months later retirement prediction model 31N+L. Thereby, the prediction unit 42 obtains the probability of retirement after N months, the probability of retirement after N+1 months, . . . the probability of retirement after N+L months.

In this way, by predicting the probability of retirement at multiple points in the future, it becomes possible to take timely measures to prevent retirement. For example, based on the prediction results, there are few signs of retirement in the short term, but if the current situation continues, the probability of retirement will increase, and there are signs of retirement in the long term, so early retirement prevention measures are necessary. This makes it possible to determine whether a person is a current employee. Furthermore, for example, since the probability of quitting after three months is on the rise, it is possible to prevent quitting by taking measures to prevent quitting, such as gradually reducing work hours over a three-month period.

Further, as shown in the lower diagram of FIG. 8, the prediction unit 42 indicates a time-series change in the retirement probability from the retirement probability after N months, the retirement probability after N+1 months, ..., the retirement probability after N+L months. You may also create a graph. Figure 9 shows a graph of the time-series change in the predicted probability predicted as of March, with N=6 and L=5, that is, the retirement probability from 6 months later (September) to 11 months later (February). shows. The example in FIG. 9 also shows a threshold value for determining whether the retirement probability is high or low. By outputting such a graph, it can be determined that the probability of retirement has increased significantly in January and that some kind of retirement prevention measure is necessary before that happens. The prediction unit 42 passes the feature vector of the incumbent to be predicted, the probability of retirement in N months, the probability of retirement in N+1 months, . . . the probability of retirement in N+L months to the interpretation unit 43.

The interpretation unit 43 uses the prediction model 31, the feature vector of the incumbent to be predicted, and the interpretation model 32 to calculate the degree of contribution of each feature included in the feature vector to the prediction result. If the contribution of all feature quantities is presented to a person in charge, such as a human resources department employee or a current manager, if the number of feature quantities is large, it will be difficult for the person in charge to check. Therefore, the interpretation unit 43 presents the feature quantities whose degree of contribution is greater than or equal to a predetermined value, or the predetermined number of feature quantities whose degree of contribution is higher than the predetermined value, as the basis on which the prediction model 31 has predicted the prediction result.

FIG. 10 shows an example of the degree of contribution for each feature amount. In the example of FIG. 10, the horizontal axis is the degree of contribution (0.0 to 1.0), and the numerical value shown in parentheses for each feature is the value of that feature. By checking the values of the feature values with high contribution degrees, the person in charge can determine the reason for predicting retirement for incumbents who have a high probability of retiring.

For example, in the example shown in Figure 10, the person in charge knows that the current employee to be predicted does not take special leave such as year-end and New Year vacation or summer vacation, and that the total number of working hours one month ago was relatively low and he/she tends to take days off. I can understand things, etc. In addition, the person in charge knows that the incumbent who is the target of the prediction was engaged in a shift that started at 2:00 p.m., and from this, the person in charge makes a prediction that the incumbent is not good at mornings or has something to do in the morning. be able to. In addition, the person in charge learned that the incumbent who was the target of the prediction was late twice three months ago, and that the total tardiness time was 272 minutes, which was extremely long. It can be predicted that the patient is feeling unwell. In addition, the person in charge can understand that the current employee who is the target of prediction has been employed for a relatively long period of three years or more. In addition, the person in charge can predict that the incumbent's total working hours are subject to large fluctuations, and that the incumbent has been late three times in the past month, and that the number of days off and tardiness is increasing. By comprehensively considering these factors, the person in charge can determine that the reason for the predicted retirement of the current employee to be predicted is "poor health."

Note that the content may be difficult for a human to understand using only the feature quantities with a high degree of contribution and the values of the feature quantities. Therefore, by converting combinations of feature quantities and feature quantity values into natural language for interpretation, interpretation by humans is facilitated even by those who are not familiar with machine learning. Specifically, the interpretation unit 43 refers to the correspondence table 33 that stores combinations of feature quantities and feature quantity values in association with texts written in natural language in which the combinations are interpreted, and selects the combinations that have a high degree of contribution. Convert features, or the basis for prediction, into natural language.

FIG. 11 shows an example of the correspondence table 33. In the example of FIG. 11, the feature name, the value of the feature, the natural language for interpretation, and the retirement prediction reason are stored in association with each other. The interpretation unit 43 acquires from the correspondence table 33 the natural language for interpretation and the retirement prediction reason corresponding to the combination of the highly contributing feature extracted using the interpretation model 32 and the value of the feature.

Further, the interpretation unit 43 determines whether, from among the plurality of retirees, the degree of similarity between the degree of contribution of each feature quantity of the retiree and the degree of contribution of each feature quantity of the incumbent person to be predicted is equal to or greater than a predetermined value, or the degree of similarity A predetermined number of retirees are extracted from the top in descending order of their values. Specifically, the interpretation unit 43 acquires feature vectors of a plurality of retirees from the retiree DB 52, and uses the feature vectors of each retiree and the interpretation model 32 as shown in FIG. The contribution of each feature is calculated for each feature, and the calculation results are listed. FIG. 13 shows an example of a list of contribution calculation results. The interpretation unit 43 creates a contribution vector whose elements are the contribution degrees for each characteristic value of the incumbent and the retirees to be predicted, and creates a contribution vector between the incumbent and the retirees to be predicted. The Euclidean distance between the two is calculated as the degree of similarity. Then, the interpretation unit 43 sorts the retirees in descending order of similarity to the contribution vector of the incumbent to be predicted, that is, in descending order of the distance between the contribution vectors. A predetermined number of top retirees are extracted in descending order of degree.

The interpretation unit 43 acquires the reason for retirement stored in the retiree DB 52 for the extracted retiree. Furthermore, the interpretation unit 43 identifies the feature with the highest degree of contribution for the extracted retiree from the list as shown in FIG. You may also obtain it from This is based on the idea that features with high contributions for retirees whose feature values are similar to those of incumbents to be predicted are likely to have an influence on the reason for retirement. In the example of FIG. 13, the interpretation unit 43 acquires the reason for retirement corresponding to the feature quantity 1 of Mr. Y whose feature quantity has the highest degree of contribution to the current incumbent who is the prediction target.

The interpretation unit 43 creates a prediction result list that describes each piece of information interpreted as described above. FIG. 14 shows an example of the prediction result list. FIG. 14 is an example of a list of prediction results for incumbents whose probability of retirement after N months predicted by the prediction unit 42 is equal to or greater than a predetermined value. The prediction result list in Figure 14 includes information such as ``Probability of retiring after N months'', ``Retirement prediction reasons'', ``Top K prediction grounds'', and ``Retirement probability in N months'', and ``Top K prediction grounds'', in association with the ``Personal code'' of the relevant incumbent. Includes "past retirees."

The "probability of retirement after N months" is the probability of retirement after N months, which is passed from the prediction unit 42. “Retirement prediction reason” is a retirement prediction reason obtained from the correspondence table 33 based on a combination of a feature amount with a high degree of contribution and a value of the feature amount. “Top K prediction grounds” is a combination of the top K features (in the example of FIG. 14, K=3) and the values of the features in descending order of degree of contribution, by referring to the correspondence table 33. This is text that has been converted into a language. In the example of FIG. 14, a value obtained by standardizing the contribution degree of each feature amount is also written as reference information so that the sum of the contribution degrees of each of the K feature amounts is 1. "Past retirees with similar tendencies" is the personal code of the retiree and the reason for retirement of the retiree, which are extracted based on the similarity of the degree of contribution of the feature amount. Note that in the example of FIG. 14, only information on one retiree is described, but information on a plurality of retirees may be recorded. By creating a list of prediction results and interpreted information in this way, it is possible to filter incumbents based on their probability of leaving or reason for leaving, and quickly search for incumbents who require priority attention.

The interpretation unit 43 further similarly creates prediction result lists for N+1 months later, . . . , N+L months later. This allows incumbents who need to take immediate action to check the list of predicted results for N months from now, while incumbents who need to take action over the long term can check the list of predicted results for N+L months. can do.

The interpretation unit 43 outputs the created prediction result list. Furthermore, the interpretation unit 43 may also output a graph showing the time series of the retirement probabilities described above. In this way, by presenting the basis for the prediction in an easy-to-interpret format along with the probability of retirement at multiple points in the future, it is possible to understand why the current employee being predicted is considering retiring, and to make appropriate decisions. Make it possible to take measures to prevent retirement.

Next, the operation of the retirement prediction processing device 10 will be explained. FIG. 15 is a flowchart showing the flow of learning processing by the retirement prediction processing device 10. The learning process is performed by the CPU 11 reading out the retirement prediction learning program from the ROM 12 or the storage 14, loading it onto the RAM 13, and executing it. Note that the learning process is an example of the retirement prediction learning method of the present invention.

In step S<b>11 , the CPU 11 , as the filter unit 21 , acquires the personnel-related information of the incumbent from the incumbent DB 51 and acquires the personnel-related information of the retiree from the retiree DB 52 .

Next, in step S12, the CPU 11, as the filter unit 21, extracts, for the retiree, personnel-related information up to the first period before the time of retirement. In addition, for the current employee, the filter unit 21 extracts personnel-related information from a point in time that is a first period before the present time and a point in time that is a second period before the present time.

Next, in step S13, the CPU 11, as the generation unit 22, performs various processing such as statistical processing and conversion into categorical variables for each item of personnel-related information of the current employees and retirees extracted in step S12. to generate a feature vector containing multiple feature quantities.

Next, in step S14, the CPU 11, as the learning unit 23, learns the prediction model 31 using the retiree's feature vector generated in step S13 as a positive example and the incumbent's feature vector as a negative example. Next, in step S15, the CPU 11, as the learning unit 23, uses the interpretation model 32 to calculate the degree of contribution of each feature included in the feature vector to the prediction result of the prediction model 31, based on the learned prediction model 31. is generated, and the learning process ends.

FIG. 16 is a flowchart showing the flow of prediction processing by the retirement prediction processing device 10. The prediction process is performed by the CPU 11 reading out the retirement prediction program from the ROM 12 or the storage 14, loading it into the RAM 13, and executing it. Note that the prediction process is an example of the retirement prediction method of the present invention.

In step S21, the CPU 11, as the generation unit 41, acquires the 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, as the generation unit 41, generates a feature vector from the personnel-related information of the incumbent who is the prediction target.

Next, in step S23, the CPU 11, as the prediction unit 42, predicts the retirement probability of the incumbent to be predicted at multiple points in the future by inputting the feature vector of the incumbent to be predicted to the prediction model 31. .

Next, in step S24, the CPU 11, as the interpretation unit 43, uses the prediction model 31, the feature vectors of each of the current and retired persons to be predicted, and the interpretation model 32 to calculate each feature included in the feature vector. Calculate the degree of contribution to the predicted amount. Next, in step S25, the CPU 11, as the interpreter 43, refers to the correspondence table 33 and selects a combination of a feature with a high degree of contribution and a value of that feature among the features of the incumbent to be predicted. Obtain the corresponding interpretation natural language and retirement prediction reasons.

Next, in step S26, the CPU 11, as the interpreter 43, determines the degree of similarity between the degree of contribution of each feature of the retiree and the degree of contribution of each feature of the incumbent to be predicted, from among the plurality of retirees. is greater than a predetermined value, or a predetermined number of top retirees are extracted in descending order of similarity. Next, in step S27, the CPU 11, as the interpreter 43, acquires the reason for retirement stored in the retiree DB 52 for the extracted retiree.

Next, in step S28, the CPU 11, as the interpreter 43, creates a prediction result list including the retirement probabilities of incumbents whose retirement probabilities after the first period are greater than or equal to a predetermined value in step S23. Further, the CPU 11, as the interpreter 43, uses the natural language indicating the basis for prediction and the retirement prediction reason acquired in step S25, the personal code of the retiree extracted in step S26, and the retirement of the retiree acquired in step S27. Include the reason in the prediction results list. The CPU 11, as the interpreter 43, outputs the created prediction result list, and the prediction process ends.

As explained above, the retirement prediction unit of the retirement prediction processing device according to the present embodiment generates a feature vector including a plurality of feature quantities from the personnel-related information for a predetermined period of each of the current employee and the retiree, and This is a predictive model that is trained using the feature vectors generated for each of the retirees as positive examples and the feature vectors generated for each of the multiple incumbents as negative examples, and the feature vectors for the incumbents are input. In this case, a prediction model that predicts the retirement probability indicating the likelihood of an incumbent retiring at multiple points in the future, and a feature vector about the incumbent to be predicted, are used to Predict the probability of retirement at the point in time. This makes it possible to predict retirement in order to support taking measures to prevent retirement at an appropriate time.

Further, according to the retirement prediction unit of the retirement prediction processing device according to the present embodiment, as shown in A of FIG. It may be predicted that Furthermore, as shown in B of FIG. 17, predictions of retirement probabilities and reasons for retirement are also performed for incumbents who were not employed at the time of learning.

Additionally, the retirement prediction unit may output the retirement probabilities at multiple points in the future as a graph showing time-series changes in the retirement probabilities. This makes it possible to appropriately judge when to take measures to prevent retirement.

In addition, the retirement prediction unit uses a prediction model, a feature vector of the incumbent to be predicted, and an interpretation model that calculates the degree of contribution of each feature to the prediction result of the prediction model. The feature amounts or the predetermined top feature amounts in descending order of contribution may be presented as the basis for the prediction model predicting the prediction result. This makes it possible to understand the basis for predicting the probability of retirement, and to consider appropriate measures to prevent retirement.

In addition, the retirement prediction unit selects from among the plurality of retirees, the degree of similarity between the feature vector of the retiree and the feature vector of the incumbent to be predicted is a predetermined value or more, or retirees, or the degree of similarity between the degree of contribution of the feature amount of the retiree and the degree of contribution of the feature amount of the incumbent person to be predicted is greater than or equal to a predetermined value, or a predetermined number of retirees in the top order of similarity. may be extracted and presented. This makes it possible to consider appropriate measures to prevent retirement by referring to information on retirees who have similar trends to the current employee who is the target of prediction. In particular, when using the similarity of the contribution of features, it is assumed that the features that were the reason for resignation are similar, so it is assumed that the reason for retirement is appropriately assumed and more appropriate retirement is achieved. Prevention measures can be considered.

Additionally, the retirement prediction unit may present the reason for retirement stored in advance for the extracted retiree. This makes it possible to appropriately assume the reasons for retirement and consider more appropriate retirement prevention measures.

In addition, the retirement prediction unit refers to a correspondence table in which combinations of feature quantities and reference values of feature quantity values are associated with and stored text in natural language that interprets the combinations, and the feature quantities to be presented are determined in a natural manner. It may be converted into a language and presented. As a result, even if personnel such as human resources personnel or managers are not familiar with machine learning, they can easily understand the basis of predictions made by the predictive model.

In addition, the retirement prediction unit corresponds to the presented feature by referring to a correspondence table in which combinations of feature values and reference values of feature values and retirement prediction reasons represented by the combinations are stored in association with each other. You may also present reasons for predicting retirement. Thereby, it is possible to easily understand the reason for the target prediction according to the feature quantity that is the basis of the prediction.

In addition, among personnel-related information that includes multiple items, for items whose values do not change over a predetermined period or whose values change irregularly, the retirement prediction department will leave the value of the feature as is or categorize it. For items of personnel-related information that are converted into variables and whose values are numerical and change regularly, feature vectors may be generated by statistically processing the values for the most recent predetermined period. As a result, from static items such as qualifications and departments, features are generated that can learn past experience, etc., and from statistical items such as attendance and salary, it is possible to learn by taking into account the most recent situation. Features are generated.

Furthermore, the retirement prediction learning unit of the retirement prediction processing device according to the present embodiment generates a feature vector including a plurality of feature quantities from the personnel-related information for a predetermined period of time for each of the current employees and retirees, and The feature vectors generated for each of the incumbents are taken as positive examples, and the feature vectors generated for each of the multiple incumbents are taken as negative examples.If the feature vectors for the incumbents are input, the incumbents at multiple future points in time are Learn a predictive model that predicts the probability of retirement. This makes it possible to generate a prediction model that can predict retirement to support taking measures to prevent retirement at an appropriate time.

In addition, the retirement prediction learning unit uses personnel-related information to be used for generating feature quantities for learning a prediction model that predicts the probability of retirement after a first period from the present time. Extract personnel-related information prior to a time period, and for the current employee, go back from the current time and extract personnel-related information prior to a second period prior to the first period. A feature vector may be generated by setting a plurality of periods as the second period and using the extracted personnel-related information as personnel-related information for a predetermined period. Thereby, the same feature amount is generated based on the values of items common to all incumbents, and it is possible to prevent this from becoming a leak feature amount.

Here, FIG. 18 shows an example of the verification results of the prediction accuracy of the prediction model in this embodiment. FIG. 18 shows the recall rate and precision rate for each employment category of incumbents as verification results for a prediction model that predicts the probability of retirement six months later. Note that A to H in FIG. 18 mask data representing employment categories such as part-time workers and full-time employees. Although there are differences by employment category, high recall rates are obtained for specific employment categories. It can be seen that even in employment category G, which has the lowest recall rate, approximately one in four incumbents can be found showing signs of retiring.

<Modified example>
In the above embodiment, the case where N is an integer larger than 0 has been described as an example of predicting the probability of retirement after N months, but it is possible to set an arbitrary number such as N=1.5.

Furthermore, in the above embodiment, when extracting retirees who are similar to the current employee who is the prediction target, a case has been described in which the similarity of the contribution of the feature amount is used, but as shown in FIG. Similarity may also be used.

Furthermore, in the above embodiment, a case has been described in which a prediction model and an interpretation model are provided separately, but a machine learning model such as LightGBM that can calculate the importance of a feature amount may be used as a prediction model.

Further, in the above embodiment, a case has been described in which the retirement probability and information on the basis of the prediction and the interpretation of the prediction result, such as the reason for predicting retirement, are output together, but only the retirement probability may be output.

Furthermore, in the above embodiment, the case where the retirement prediction learning section and the retirement prediction section are realized by one computer (retirement prediction processing device) has been described, but they may be realized by different computers.

Furthermore, various processors other than the CPU may execute the retirement prediction learning process and the retirement prediction process that the CPU reads and executes the software (program) in the above embodiments. The processor in this case is a PLD (Programmable Logic Device) whose circuit configuration can be changed after manufacturing, such as an FPGA (Field-Programmable Gate Array), and an ASIC (Application Specific Intel). In order to execute specific processing such as egrated circuit) An example is a dedicated electric circuit that is a processor having a specially designed circuit configuration. Further, the retirement prediction learning process and the retirement prediction process may be executed by one of these various processors, or by a combination of two or more processors of the same type or different types (for example, multiple FPGAs and CPUs). and FPGA). Further, the hardware structure of these various processors is, more specifically, an electric circuit that is a combination of circuit elements such as semiconductor elements.

Furthermore, in the above embodiment, a mode has been described in which the retirement prediction learning program and the retirement prediction program are stored (installed) in the ROM 12 or the storage 14 in advance, but the present invention is not limited to this. The program can be installed on CD-ROM (Compact Disk Read Only Memory), DVD-ROM (Digital Versatile Disk Read Only Memory), and USB (Universal Serial Bus) stored in a non-transitory storage medium such as memory It may be provided in the form of Further, the program may be downloaded from an external device via a network.

Regarding the above embodiments, the following additional notes are further disclosed.

(Additional note 1)
memory and
at least one processor connected to the memory;
including;
The processor includes:
Generate a feature vector including multiple feature quantities from personnel-related information for a predetermined period of time for each of current and retired employees,
A predictive model that is trained using feature vectors generated for each of a plurality of retirees as a positive example and feature vectors generated for each of a plurality of incumbents as a negative example, and the feature vector for the incumbent is input. In the case that A retirement prediction device configured to predict retirement probabilities at the plurality of future points in time.

(Additional note 2)
memory and
at least one processor connected to the memory;
including;
The processor includes:
Generate a feature vector including multiple feature quantities from personnel-related information for a predetermined period of time for each of current and retired employees,
The feature vectors generated for each of multiple retirees are taken as positive examples, and the feature vectors generated for each of multiple incumbents are taken as negative examples. When the feature vectors for incumbents are input, future multiple A retirement prediction learning device configured to learn a prediction model that predicts a retirement probability indicating the possibility that the current employee will retire at a point in time.

(Additional note 3)
A non-temporary recording medium storing a program executable by a computer to execute retirement prediction processing,
The retirement prediction process includes:
Generate a feature vector including multiple feature quantities from personnel-related information for a predetermined period of time for each of current and retired employees,
A predictive model that is trained using feature vectors generated for each of a plurality of retirees as a positive example and feature vectors generated for each of a plurality of incumbents as a negative example, and the feature vector for the incumbent is input. In this case, the prediction model predicts the retirement probability indicating the possibility that the incumbent will retire at multiple points in the future, and the feature vector for the prediction target incumbent is used to predict the prediction target incumbent. predicting retirement probabilities at said plurality of future points in time.

(Additional note 4)
A non-temporary recording medium storing a program executable by a computer to execute a retirement prediction learning process,
The retirement prediction learning process includes:
Generate a feature vector including multiple feature quantities from personnel-related information for a predetermined period of time for each of current and retired employees,
The feature vectors generated for each of multiple retirees are taken as positive examples, and the feature vectors generated for each of multiple incumbents are taken as negative examples. When the feature vectors for incumbents are input, future multiple A non-temporary recording medium comprising learning a predictive model that predicts a retirement probability indicating the possibility that the incumbent will retire at a point in time.

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 section 21 Filter section 22 Generation section 23 Learning section 31 Prediction model 32 Interpretation model 33 Correspondence table 40 Retirement prediction section 41 Generation section 42 Prediction section 43 Interpretation section 51 Incumbent DB
52 Retirement DB

Claims

a generation unit that generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of time for each of the current and retired employees;
A predictive model that is trained using feature vectors generated for each of a plurality of retirees as a positive example and feature vectors generated for each of a plurality of incumbents as a negative example, and the feature vector for the incumbent is input. In the case that a prediction unit that predicts the probability of retirement at multiple points in the future;
Retirement prediction device including.
Using the prediction model, the feature vector of the incumbent to be predicted, and an interpretation model that calculates the degree of contribution of each feature to the prediction result of the prediction model, a feature whose degree of contribution is equal to or greater than a predetermined value; 2. The retirement prediction device according to claim 1, further comprising an interpretation unit that presents a predetermined number of top feature quantities in descending order of contribution as basis for the prediction model predicting the prediction result.
Using the prediction model, the feature vector of each of the plurality of retirees, the feature vector of the incumbent who is the prediction target, and an interpretation model that calculates the degree of contribution of each feature to the prediction result of the prediction model. , from among the plurality of retirees, the degree of similarity between the degree of contribution of the feature amount of the retiree and the degree of contribution of the feature amount of the incumbent person to be predicted is a predetermined value or more, or a top predetermined person is selected in descending order of similarity. The retirement prediction device according to claim 1, further comprising an interpreter that extracts and presents a number of retirees.
a generation unit that generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of time for each of the current and retired employees;
The feature vectors generated for each of multiple retirees are taken as positive examples, and the feature vectors generated for each of multiple incumbents are taken as negative examples. When the feature vectors for incumbents are input, future multiple a learning unit that learns a prediction model that predicts a retirement probability indicating the possibility that the incumbent will retire at a point in time;
Retirement prediction learning device including.
As personnel-related information used to generate features for learning a predictive model that predicts the probability of retirement after a first period from the current time, for the retiree, the information is collected before the first period, retroactive from the point of retirement. For the current employee, extract the personnel-related information from the time before the first period going back from the present time and extracting the personnel-related information from the time before the second period. includes a filter section that sets multiple periods as
The retirement prediction learning device according to claim 4, wherein the generation unit generates the feature vector using the personnel-related information extracted by the filter unit as the personnel-related information for the predetermined period.
a generation unit generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of each of the current employee and the retired employee;
A prediction model in which a prediction unit is trained using feature vectors generated for each of a plurality of retirees as positive examples and feature vectors generated for each of a plurality of incumbents as a negative example, the prediction unit learning features for the incumbent. When a vector is input, the prediction is made using the prediction model that predicts the retirement probability indicating the possibility that the incumbent will retire at multiple points in the future, and the feature vector for the incumbent to be predicted. A retirement prediction method that predicts the retirement probability of a target incumbent at multiple points in the future.
a generation unit generates a feature vector including a plurality of feature quantities from personnel-related information for a predetermined period of each of the current employee and the retired employee;
When the learning unit inputs the feature vectors for the incumbent, using the feature vectors generated for each of the multiple retirees as a positive example and the feature vectors generated for each of the multiple incumbents as a negative example, A retirement prediction learning method that learns a prediction model that predicts a retirement probability indicating the possibility that the incumbent will retire at multiple points in the future.
To cause a computer to function as each part constituting the retirement prediction device according to any one of claims 1 to 3, or each part constituting the retirement prediction learning device according to claim 4 or claim 5. program.