JP2016110289A - Abnormality extraction device, abnormality extraction method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an abnormality extraction device, an abnormality extraction method, and a program that extract abnormality by taking into account periodicity of survival time and periodicity of a time series.SOLUTION: A survival function estimating part 24 estimates a cycle/term cancellation rate and a period cancellation rate of cancelled contract terms and contracts, for each combination of cycle units of a time series, on the basis of contract history information inputted for each of a plurality of users. A cancellation number estimating part 26 estimates a cancellation number reached by users of a specified attribute in each period, on the basis of a contract number reached by users of the specified attribute in each contract period, that is acquired from the contract history information, and on the basis of the estimated cycle/term cancellation rate and the period cancellation rate. An abnormality extracting part 28 extracts a period in which a cancellation number is abnormal, on the basis of the cancellation number reached by users of the specified attribute in each period, that is acquired from the contract history information, and on the basis of the estimated cancellation number for each period.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an abnormality extraction device, method, and program, and more particularly, to an abnormality extraction device, method, and program for grasping a contract cancellation record.

  As a conventional technique, there is anomaly extraction by a time series analysis technique such as an ARIMA model or a Kalman filter.

  In addition, there are abnormal extractions using survival time analysis techniques such as the mixed Weibull model and the Kaplan-Meier method.

JP 2014-093042 A

  In the prior art, there is a problem that an increase in churn risk due to seasonality in time series, periodicity in survival time, and local churn risk increase in a specific attribute group cannot be separated.

  In addition, when applying the survival analysis technique to solve the above problem, it is difficult to distinguish whether the extracted one is a local anomaly or the assumed distribution is wrong. There is a problem that the placing method cannot be used.

  In the present invention, an object of the present invention is to provide an anomaly extraction apparatus, method, and program for extracting an anomaly in consideration of the periodicity of the survival time and the periodicity of the time series. And

  In order to achieve the above object, the abnormality extraction device according to the first aspect of the present invention is the input information on the contract and cancellation of the user for each of the plurality of users, and the user's information for each of the plurality of users. Based on the contract history information including attribute information, the contract period of the contract and the unit of the period for each combination of the units of the time-series period with the predetermined period as a unit, which has been canceled. A survival function estimation unit for estimating a period / period churn rate at which the contract of the contract period is terminated and a period churn rate at which the contract at the contract period is terminated for each of the contract periods of the contract; and The number of contracts that the user with the specified attribute acquired from the contract history information contracts with each of the contract periods of the contract, and the period / period churn rate estimated by the survival function estimation unit and A churn number estimation unit that estimates the number of churns that the user of the designated attribute churns at each of the times based on the period churn rate, and the user of the designated attribute that is acquired from the contract history information An abnormal extraction unit that extracts a period when the churn number becomes abnormal based on the churn number canceled at each of the periods and the churn number of each of the periods estimated by the churn number estimation unit; , Including.

  The abnormality extraction method of the second invention is an abnormality extraction method in an abnormality extraction device including a survival function estimation unit, a churn number estimation unit, and an abnormality extraction unit, wherein the survival function estimation unit is input A contract period of the contract based on contract history information including information on the contract and cancellation of the user for each of a plurality of users, and information on the attributes of the user for each of the plurality of users; and The period / period cancellation rate at which the contract of the contract period is canceled in the period unit, and the contract of the contract, for each combination of time-series period units with a predetermined period as a unit after the contract is canceled For each of the periods, the contract cancellation rate of the contract period is estimated, and the cancellation number estimation unit is obtained from the contract history information. Based on the number of contracts contracted for each contract period of the contract and the period / period churn rate and period churn rate estimated by the survival function estimation unit, the user of the specified attribute is assigned to each period. The number of cancellations to be canceled is estimated, and the abnormality extraction unit is acquired from the contract history information, and is estimated by the number of cancellations that the user with the specified attribute has canceled at each of the times, and the cancellation number estimation unit. In addition, the time when the number of cancellations becomes abnormal is extracted based on the number of cancellations at each time period.

  According to the first and second inventions, the contract period of the contract and the time-series cycle in which the contract is canceled based on the contract history information for each of the plurality of users input by the survival function estimation unit For each combination of units, the period / period churn rate and the period churn rate are estimated, and the churn number estimation unit obtains the contract with the specified attribute acquired from the contract history information. Based on the number of contracts contracted for each period, the estimated period / period churn rate and the period churn rate, the number of churns that the user with the specified attribute cancels for each period is estimated, and an anomaly extraction unit When the number of churns becomes abnormal on the basis of the number of churns that the user with the specified attribute cancels at each time and the estimated number of churns for each time obtained from the contract history information To extract.

  As described above, based on the entered contract history information for each of the plurality of users, the contract period of the contract, and the period / period cancellation for each of the combinations of the units of the time-series period that canceled the contract. Estimate the rate and the period churn rate, and the number of contracts that the user with the specified attribute that is acquired from the contract history information contracts at each of the contract timings of the contract, the estimated period / period churn rate and Based on the churn rate, the number of churns that users with the specified attribute churn at each time period is estimated, and the churns that the user with the specified attribute churn at each time period are obtained from the contract history information. By extracting the period when the churn number becomes abnormal based on the number and the estimated churn number of each period, abnormal extraction considering the periodicity of the survival time and the periodicity of the time series is performed. be able to.

  In the first and second aspects of the invention, the abnormality extraction unit estimates a confidence interval of the estimated value of the churn number based on the estimated churn number at each of the estimated times, and determines from the contract history information Based on the obtained churn number that the user of the specified attribute churns at each of the times and the confidence interval of the estimated value of the estimated churn number, the churn number belongs to the confidence interval. The period when there is no cancellation may be extracted as the period when the number of cancellations becomes abnormal.

  In the first and second aspects of the invention, the survival function estimation unit estimates the contract in the unit of the period estimated for each of the contract period of the contract and the combination of the units of the period in which the contract is canceled. The period / period churn rate may be estimated by performing a smoothing process for smoothing in the direction of the contract period with respect to the period / period churn rate at which the contract of the period is canceled.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said abnormality extraction apparatus.

  As described above, according to the abnormality extraction device, method, and program of the present invention, the contract period of the contract and the contract are canceled based on the input contract history information for each of the plurality of users. Estimate the period / period churn rate and period churn rate for each combination of time-series cycle units, and the user with the specified attribute obtained from the contract history information Based on the number of contracts contracted with each, the estimated period / period churn rate and period churn rate, the number of churns that the user with the specified attribute cancels at each period is estimated and obtained from the contract history information By extracting the period when the churn number becomes abnormal based on the number of churns that the user with the specified attribute quits at each period and the estimated number of churns at each period Time periodicity and time It can be an abnormality extracted in consideration of the periodicity of the column.

It is an example of the figure which demonstrated simply the flow of the process of the abnormality extraction apparatus which concerns on this Embodiment. It is a block diagram which shows the functional structure of the abnormality extraction apparatus which concerns on this Embodiment. It is a figure which shows the example of the information of the observation start date contained in contract history information. It is a figure which shows the example of the information of the observation end date contained in contract history information. It is a figure which shows the example of the contract information contained in contract history information. It is a figure which shows the example of the survival function of the macro for every period. It is a figure which shows the example of the churn rate at the time of the end of observation for every contract time. It is a figure which shows the example of calculation of the churn number estimated value. It is a figure which shows the example of the table output. It is a figure which shows the example of the graph output. It is a flowchart figure of the abnormality extraction process routine in the abnormality extraction device which concerns on this Embodiment. It is a figure which shows the example of the surrender increase which is not assumed.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Principle of this embodiment>

  The principle of this embodiment will be described. As shown in FIG. 1, first, a macro survival function is obtained for each unit of a time-series period from nationwide contract / cancellation information. Next, the churn person estimated value is obtained based on the transition of the number of contracts for each region and the macro survival function. Then, the churn person transition of the region a and the churn person estimated value are compared, and the time when the divergence occurs is extracted.

<Configuration of Abnormality Extraction Device According to Embodiment of the Present Invention>
Next, the configuration of the abnormality extraction device according to the embodiment of the present invention will be described. As shown in FIG. 2, the abnormality extraction device 100 according to the embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing an abnormality extraction processing routine described later. It can consist of computers. As shown in FIG. 2, this abnormality extraction device is functionally configured to include an input unit 10, a calculation unit 20, and an output unit 90.

The input unit 10 receives the contract history information and stores it in the contract history storage unit 22. Here, the contract history information is, for example, information on the observation start date when the contract history information shown in FIG. 3 is acquired, and information on the observation end date when the acquisition of the contract history information shown in FIG. 5, for each contract shown in FIG. 5, the contract information consisting of a combination of the contractor's customer ID, the contractor's address code, the contract date, the cancellation date, and the contract status, and the correspondence relationship between the date and the cycle unit ( (Not shown). The unit of the time-series period is a predetermined period. In this embodiment, the unit of the period is a calendar month, and the possible values are January to December. Also, the time corresponding to the observation start date is set to t _str and the time corresponding to the observation end date is set to t _end .

  The calculation unit 20 includes a contract history storage unit 22, a survival function estimation unit 24, a churn number estimation unit 26, and an abnormality extraction unit 28.

  The contract history storage unit 22 stores the contract history information received by the input unit 10.

  Based on the contract history information stored in the contract history storage unit 22, the survival function estimation unit 24 performs the period / period churn rate, which is a macro survival function for each cycle s, and the observation end time for each contract time t. The term churn rate H (t), which is the churn rate of the contract, is acquired. An example of the period / period churn rate is shown in FIG. 6, and an example of the timing churn rate H (t) is shown in FIG. The unit of time is one month.

  First, the acquisition of the period / period churn rate will be described. Specifically, first, based on the contract state information included in the contract history information, the total number of cancellations N on the observation end date is acquired.

  Next, for each of the contracts corresponding to the contract information included in the contract history information, it is determined from the contract status information included in the contract information of the contract whether or not the contract has been canceled on the observation end date. . Here, when it is determined that the contract has been canceled on the observation end date, the contract period x of the contract and the contract period x and the contract date are included in the contract information of the contract. The period s corresponding to the cancellation date is acquired. Note that when determining whether or not the contract has been canceled on the observation end date, the number of contracts that have been canceled is counted, and the total number of cancellations N.

Next, based on the contract period x of the contract and the period s at which the contract is canceled, acquired for each contract canceled on the observation end date, all contract periods x and periods s For each combination, the number _nxs canceled in the period s in the contract period x is acquired. Here, x takes each of the values 1 to X, and s takes each of the values 1 to 12.

Next, based on the acquired total number N of cancellations and each of the numbers _nxs canceled in the period s in the contract period x acquired for all combinations of the contract period x and the period s, the following (1) The period / period churn rate P (x, s) is acquired according to the equation. Then, for each of the combinations of all contract periods x and periods s, a smoothing process (moving average of the average period / period churn rate P (x, s)) is smoothed in the x-axis direction as the contract period. Any other method may be used to obtain the final period / period churn rate P (x, s) from which noise has been removed. In this way, a macro survival function can be created by non-parametric estimation for each period.

Next, acquisition of the time lapse rate H (t) will be described. First, based on the contract information included in the contract history information, for each time _{t (t = t str ~t end} ), get to the time t, contract date corresponds, subscriptions _{m t} contracted at a time t To do. Next, based on the contract information of each contract determined to be canceled on the observation end date, the contract date corresponds to the time t for each time t (t = t _{str to} t _end ), The number n _t that has already been canceled is acquired. Next, the number of contracts m _t contracted in t time acquired, contract the time t, already on the basis of the number n _t are canceled, according to the following equation (2), timing churn H a (t) get.

The churn number estimation unit 26, the contract history information stored in the contract history storage unit 22, the period / period churn rate P (x, s) and the timing churn rate H (t) acquired by the survival function estimation unit. Based on the above, the churn number estimated value ^ N _a (t) at each time t for each of the areas a designated as the estimation target is calculated. A symbol with “^” in front of the symbol indicates that the value of the symbol is an estimated value.

Specifically, first, each of the contracts belonging to each region a which is an attribute is classified based on the contract information of the contracts included in the contract history information stored in the contract history storage unit 22. In the present embodiment, it is assumed that each contract having the same address code belongs to the same area. Next, the area a to be processed is determined from the designated areas. Next, based on each information of the contract date included in the contract information of the contract of each contract belonging to the region a to be processed, and acquires the contract number m _at every time t in the region a. Next, based on the number of contracts mat _at each time t in the acquired area a, the period / period churn rate P (x, s), and the time churn rate H (t) acquired by the survival function estimation unit. Then, according to the following equation (3), the churn number estimated value ^ N _a (t) for each time t in the region a is calculated. Note that t takes a value of t _{str to} t _end . Here, the estimated churn number estimated value ^ N _a (t) for each period t in the region a will be described. For example, among the contractors in December 2012, the number of churners to be canceled in January 2013 is (contract period). Cycle / period churn rate P (1 month, January) for 1 month contractor cancels in January x (contractor churn rate H in December 2012 (December 2012)) x (2012 12 It is obtained by integrating the number of monthly contracts m _t ). The sum of the period from the observation start month to December 2012 (one period before January 2013) is the estimated churn number estimate for January 2013. Thus, the churn number estimation can be obtained by applying the macro survival function to the number of contracts in the region. FIG. 8 shows an example of calculating the churn number estimated value.

The churn number estimation unit 26 calculates the churn number estimated value ^ N _a (t) for each time period t for each of the areas a by performing the above processing for all the specified areas a.

Based on the churn number estimation value ^ N _a (t) for each time period t for each region a acquired by the churn number estimation unit 26, the abnormality extraction unit 28 calculates the churn number estimation value for each time period t of the region a. Confidence intervals (^ N _{a, min} (t) to ^ N _{a, max} (t)) are calculated. Specifically, the obtained confidence interval (the upper limit _{^ N a} cancellation by the estimated value for each time t of the region _{a ^ N a (t),} max (t), the lower limit _{^ N a,} and _min (t) Is calculated according to the following formula (4) and the following formula (5) based on the estimated churn number estimated value ^ N _a (t) for each time t in the area a.

Incidentally, abnormality extraction unit 28, the above process, by performing the all regions a specified, for each of the areas a, for each time t, churn estimate ^ N a _(t) of the confidence interval (^ N _{a, min} (t) to ^ N _{a, max} (t)) are calculated.

In addition, the abnormality extraction unit 28 includes the contract history information stored in the contract history storage unit 22 and the churn number estimated value ^ N _a (t) for each time t acquired for each area a by the churn number estimation unit 26. And the confidence interval of the estimated churn number estimated value ^ N _a (t) for each period t acquired for each region a, there is a discrepancy between the actual value and the estimated value for each designated region a. Ano _a (t) is calculated for each time t indicating the occurrence. Moreover, the abnormality extraction part 28 produces the graph and table showing the deviation of an actual value and an estimated value based on the value of Ano _a (t) for every time t acquired about each area a, and outputs it. Output to the unit 90.

When the value of the abnormal value Ano _a (t) is 0, the value is not treated as an abnormal value. Also, output examples are shown in FIGS. FIG. 9 shows an example in which the abnormal value Ano _a (t) is output in a table, and the square color is changed according to the degree of abnormality. FIG. 10 shows an example in which the actual churn number (actual), the upper limit value (max_line) of the confidence interval of the churn number estimated value, and the lower limit value (min_line) of the confidence interval of the churn number estimated value are output as a graph. Further, the possible values of t are t _{str to} t _end .

Specifically, first, a region a to be processed is determined from the designated regions. Next, the number of cancellations N _a (t) for each period t in the region a is acquired based on each of the cancellation date information included in the contract information of each contract belonging to the region a to be processed. . Next, the churn number N _a (t) for each time t in the acquired region a, the churn number estimated value ^ N _a (t) for each time t of the region a acquired by the churn number estimation unit 26, and the acquired Based on the confidence interval of the estimated churn value ^ N _a (t) for each time t, according to the following equation (6), all the possible values of t (t _{str to} t _end ) are abnormal values Ano _a (t ) Is calculated. In this way, the deviation between the actual number of cancellations and the estimated value is extracted as a regional local abnormality.

<Operation of Abnormality Extraction Device According to Embodiment of the Present Invention>
Next, the operation of the abnormality extraction device 100 according to the present embodiment will be described. First, the contract history information is received by the input unit 10 and stored in the contract history storage unit 22. Next, in the abnormality extraction device 100, when the contract history information stored in the contract history storage unit 22 is read, the abnormality extraction processing routine shown in FIG.

  First, in step S104, a contract to be processed is determined from each contract included in the read contract history information.

  Next, in step S106, it is determined whether or not the contract is canceled based on the contract status information of the contract information corresponding to the contract among the read contract history information for the contract to be processed. . If it is determined that the contract has been canceled, the process proceeds to step S108. On the other hand, if it is determined that the contract is in contract, the process proceeds to step S114. In addition, in the process of step S106, the number of times determined to have been canceled is counted, and the counted value is set as the total number of cancellations N.

  Next, in step S108, for the contract to be processed, the contract period x of the contract is calculated based on the contract date and cancellation date information of the contract information corresponding to the contract in the read contract history information. .

  Next, in step S110, for the contract to be processed, information on the correspondence between the date and the cycle unit of the read contract history information, information on the date of cancellation of the contract information corresponding to the contract, Based on the above, the contract cancellation period s of the contract is acquired.

  Next, in step S112, for the contract to be processed, the contract time t of the contract is acquired based on the contract date information of the contract information corresponding to the contract in the read contract history information.

  Next, in step S114, it is determined whether or not the processing in step S106 has been completed for all contracts included in the read contract history information. If the process of step S106 has not been completed for all contracts, the process proceeds to step S104, the contract to be processed is changed, and the processes after step S106 are repeated. On the other hand, when the process of step S106 has been completed for all contracts, the process proceeds to step S116.

Next, in step S116, all possible values x can be taken based on the contract period x of each contract that has been canceled acquired in step S108 and the contract period s of each contract that has been canceled acquired in step S110. and the value (1 to x), for all combinations of all possible values of s (1 to 12), and acquires the number _{n xs} you cancel the cycle s contract period x.

Next, in step S117, for all combinations of all values (1 to X) that x can take and all values (1 to 12) that s can take, the total churn number N acquired in step S106 is Based on the number _nxs canceled in the period s in the contract period x acquired in step S116, the period / period churn rate P (x, s) is calculated according to the above equation (1).

  Next, in step S118, the period / period churn rate acquired in step S117 for all combinations of all possible values of x (1 to X) and all possible values of s (1 to 12). A smoothing process for smoothing P (x, s) in the x-axis direction, which is the contract period, is performed and acquired as a final period / period churn rate P (x, s).

Next, in step S120, based on the contract date information and the observation start date information of each contract information included in the read contract history information, each time t (t = t _{str to} t _end ). to calculate the number of contracts _{m t} in.

Next, in step S122, the number n _t contracted at each time t (t = t _{str to} t _end ) and already canceled based on each contract time t of the contracts canceled in step S112. calculate.

Next, in step S124, the based and subscriptions _{m t} at each time t obtained in step S120, on the number _{n t} under contract already cancel each time t obtained in step S122, the (2) According to the equation, the time lapse rate H (t) at each time t (t = t _{str to} t _end ) is calculated.

  Next, in step S125, each of the contracts is classified for each region a (address code) based on the address code information of the contract information of each contract included in the read contract history information.

  Next, in step S126, the area a to be processed is determined from the designated areas.

Next, in step S128, based on the contract date information of the contract information included in the read contract history information of each contract belonging to the area a to be processed acquired in step S125, in the area a. The contract number mat _at each time t (t _{str to} t _end ) is acquired.

Next, in step S130, the period / period churn rate P (x, s) acquired in step S118, the time lapse rate H (t) acquired in step S124, and the time t in the region a acquired in step S128. based on the contract number _{m at} every, according to the above (3) to calculate the churn estimate _{^ N} a (t) for each time _{t (t} str _{~t end)} in the region a.

Next, in step S132, based on the churn number estimated value ^ N _a (t) for each time t in the region a acquired in step S130, the churn number estimation for each time t in the region a is performed according to the above equation (4). the upper limit of the confidence interval of values _{^ N a (t) ^ N} a, and calculates the _max (t).

Next, in step S134, based on the churn number estimated value ^ N _a (t) for each time t in the region a acquired in step S130, the churn number estimation for each time t in the region a is performed according to the above equation (5). the lower limit of the confidence interval for the value _{^ N a (t) ^ N} a, and calculates the _min (t).

  Next, in step S136, it is determined whether or not the processing in steps S128 to S134 has been completed for all designated areas a. When the processing of step S128 to step S134 is completed for all designated areas a, the process proceeds to step S138. On the other hand, if the processing of step S128 to step S134 has not been completed for all designated areas a, the process proceeds to step S126, the area a to be processed is changed, and the processes of steps S128 to S136 are performed. repeat.

  Next, in step S138, the designated area a to be processed is determined.

Next, in step S140, in the region a based on the contract date cancellation date information of each contract belonging to the region a to be processed acquired in step S125 included in the read contract history information. The number of cancellations N _a (t) at each time t is calculated.

Next, in step S142, the upper limit value ^ N _{a, max} (t) of the confidence interval of the churn number estimated value ^ N _a (t) at each time t in the region a obtained in step S132, and obtained in step S134. The lower limit value ^ N _{a, min} (t) of the confidence interval of the estimated number of cancellations ^ N _a (t) at each time t in the region a, and the number of cancellations N _a (for each time t in the region a acquired in step S140. Based on t), the abnormal value At _{a a} (t) at each time t (t = t _{str to} t _end ) of the region _a is acquired according to the above equation (6).

  Next, in step S144, it is determined whether or not the processing in steps S140 to S142 has been completed for all designated areas a. If the processing of step S140 to step S142 has been completed for all designated areas a, the process proceeds to step S146. On the other hand, if the processing of step S140 to step S142 has not been completed for all designated areas a, the process proceeds to step S138, the area a to be processed is changed, and the processes of steps S140 to S142 are performed. repeat.

Next, in step S146, a table or a graph as shown in FIG. 9 and FIG. 10 is created based on the abnormal value At _a (t) at each time t for each region a acquired in step S142.

  Next, in step S148, the table or graph acquired in step S146 is output to the output unit 90, and the abnormality extraction processing routine is terminated.

  As described above, according to the abnormality extraction device according to the present embodiment of the present invention, the contract period of the contract and the contract are canceled based on the input contract history information for each of the plurality of users. Estimate the period / period churn rate and period churn rate for each combination of cycle units, and users with specified attributes obtained from the contract history information Based on the number of contracts contracted and the estimated period / period churn rate and period churn rate, the number of churns that the user with the specified attribute churns at each period is estimated and obtained from the contract history information. Based on the number of surrenders that the user with the specified attribute cancels at each time period and the estimated number of surrenders at each time period, the period of survivability is extracted by extracting the time when the surrender number is abnormal Periodicity and periodicity of time series It can be a consideration anomaly extracted.

  Further, by obtaining and superimposing the survival function for each period, it is possible to realize abnormality extraction considering not only the periodicity of the survival time but also the periodicity of the time series.

  In addition, since the survival function obtained using a nonparametric counting method is used, it is not necessary to confirm the validity of the distribution assumption.

  Further, it is possible to realize abnormality extraction by dividing the whole data and the local data, obtaining a distribution function from the reliable whole data, and applying it to the local data.

  Further, when compared with the conventional time series analysis technique, the analysis is performed in consideration of not only the information on cancellation but also the information on the contract, so that the abnormality can be extracted with higher accuracy.

  In addition, by analyzing historical data with the start / end time of service contract / cancellation, survival / death, equipment failure, etc., it occurred locally only in a certain group caused by a certain event in time series An increase in churn risk can be extracted. Thereby, it can utilize for the retention analysis in marketing, equipment failure prediction, etc.

  Also, as shown in FIG. 12, among the surrender increases caused by some event on the time series, the surrender can be explained by the periodicity or the milestones of the contract period enclosed by the solid line or the dashed-dotted circle in FIG. The increase and the other unexpected increase in surrender surrounded by a dotted circle in FIG. 12 can be distinguished to enable discovery and extraction.

  By dealing with two different time axes (time series and survival time), the churn risk increase caused by events on the time series, the period on the time series (periodicity, etc.), the periodicity on the survival time (contracts) It is possible to distinguish and extract from the increase in the number of cancellations at milestones).

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, in the present embodiment, the case where the abnormality extraction device is used for application to sales data has been described, but the present invention is not limited to this, such as product life analysis, company management analysis, disaster impact analysis, etc. As long as the start and end times can be acquired, the abnormality extraction device may be applied.

  In the present embodiment, the case where the local (attribute) is a specific area (residence area) corresponding to the whole country is described, but the present invention is not limited to this. For example, any classification may be used, such as a specific customer segment (customer attribute) or service category.

  In the present embodiment, an example of handling time series in units of one month has been described, but the present invention is not limited to this. For example, an arbitrarily set unit such as a year unit, a day unit, or a time unit may be used.

  In the present embodiment, the case where the unit of the period is a calendar month has been described. However, the present invention is not limited to this. For example, a unit set arbitrarily such as a quarter unit, a week unit, a day unit, or a time unit may be used as a cycle unit.

  Further, in the present embodiment, a cycle / period churn rate P (() obtained by performing a smoothing process for smoothing the acquired cycle / period churn rate P (x, s) in the x-axis direction which is a contract period. Although the case of acquiring as x, s) has been described, the present invention is not limited to this. For example, the acquired cycle / period churn rate P (x, s) may be used as it is as the cycle / period churn rate P (x, s).

  In the present embodiment, the description has been given of the case where the 95% confidence interval is calculated by normal distribution approximation assuming a Poisson distribution having a sufficiently large average and variance. However, the present invention is not limited to this. For example, other existing methods may be used. For example, the variance of the estimated churn number at each time t of the contract in the region a may be taken, and the time when the abnormal value may be extracted using the variance value.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do.

,
DESCRIPTION OF SYMBOLS 10 Input part 20 Calculation part 22 Contract history memory | storage part 24 Survival function estimation part 26 Cancellation number estimation part 28 Abnormality extraction part 90 Output part 100 Abnormality extraction apparatus

Claims (7)

The contract period of the contract based on the contract history information including the input information regarding the contract and cancellation of the user for each of the plurality of users and the information regarding the attribute of the user for each of the plurality of users. And the period / period cancellation rate at which the contract of the contract period is canceled in the unit of the period for each combination of the units of the time-series period with the predetermined period as the unit, and the contract A survival function estimator for estimating a contract cancellation rate at which the contract at the contract period is canceled for each of the contract periods;
Obtained from the contract history information, the number of contracts that the user of the specified attribute has contracted with each of the contract periods of the contract, the period / period churn rate and the period churn rate estimated by the survival function estimation unit, , A churn number estimation unit for estimating the number of churns that the user of the specified attribute churns at each time period,
Obtained from the contract history information, based on the number of churns canceled by the user of the designated attribute at each of the times and the number of churns at each of the times estimated by the churn number estimating unit. An anomaly extraction unit that extracts the time when the number of cancellations becomes abnormal,
An abnormal extraction device including The abnormality extraction unit estimates a confidence interval of an estimated value of the churn number based on the churn number of each of the estimated times,
Based on the number of churns that the user of the specified attribute churns at each of the times acquired from the contract history information, and the confidence interval of the estimated value of the churn number, the number of churns is: The abnormality extraction device according to claim 1, wherein a period that does not belong to the confidence interval is extracted as a period when the churn number becomes abnormal.   The survival function estimation unit estimates each of the contract period of the contract and the combination of the units of the cycle that has canceled the contract, and the period / period churn rate at which the contract of the contract period is canceled in the unit of the period The abnormality extraction device according to claim 1 or 2, wherein the period / period churn rate is estimated by performing a smoothing process for smoothing in the direction of the contract period. An abnormality extraction method in an abnormality extraction device including a survival function estimation unit, a churn number estimation unit, and an abnormality extraction unit,
The survival function estimation unit is input based on contract history information including information on the contract and cancellation of the user for each of the plurality of users and information on the attribute of the user for each of the plurality of users. For each of the combinations of the contract period of the contract and the unit of the unit of the time-series period with the predetermined period as the unit, the period in which the contract of the contract period is canceled in the unit of the period Estimating the period churn rate and the churn rate at which the contract at the contract date was canceled for each of the contract dates of the contract,
The churn number estimation unit is obtained from the contract history information, and the number of contracts that the user with the specified attribute has contracted with each contract period of the contract, and the period / period estimated by the survival function estimation unit. Based on the churn rate and the churn rate, the number of churns that the user of the specified attribute churns at each of the times is estimated,
The abnormality extraction unit is obtained from the contract history information, and the number of churns canceled by the user of the designated attribute at each of the times, and the churn number estimation unit estimated by the churn number estimation unit Based on the number of surrenders, extract the time when the surrender number is abnormal,
Abnormal extraction method. The extraction by the abnormality extraction unit estimates a confidence interval of the estimated value of the churn number based on the churn number of each of the estimated times,
Based on the number of churns that the user of the specified attribute churns at each of the times acquired from the contract history information, and the confidence interval of the estimated value of the churn number, the number of churns is: The abnormality extraction method according to claim 4, wherein a period that does not belong to the confidence interval is extracted as a period when the number of cancellations becomes abnormal.   Estimating by the survival function estimating unit estimates the contract period of the contract and the combination of the units of the period when the contract is canceled, and the period when the contract of the contract period is canceled in the unit of the period 6. The abnormality extraction method according to claim 4 or 5, wherein the period / period churn rate is estimated by performing a smoothing process to smooth the period churn rate in the direction of the contract period.   The program for functioning a computer as each part which comprises the abnormality extraction device of any one of Claims 1-3.