JP7354844B2 - Impact determination program, device, and method - Google Patents

Description

The disclosed technology relates to an impact determination technology.

Time-series data is input into a model learned by supervised machine learning, and the state etc. at a time later than the input time-series data is estimated (which can also be called inference). Estimation in the human resources and financial fields may require interpretability of estimation results. For example, when inputting past attendance data and estimating the possibility of future leave of absence, it is important to know which of the inputted attendance data was most affected by the possibility of leave of absence or not. You may be required to provide the reason for your assumption.

A technique called LIME (Local Interpretable Model-agnostic Explanations) has been proposed as a technique for interpreting estimation results. In LIME, a multiple regression model is generated by locally approximating a learned model regarding time series data using learning data around the data to be evaluated. Then, based on the magnitude of the partial regression coefficient corresponding to each explanatory variable of the regression equation representing the multiple regression model, explanatory variables that have more influence on the estimation are identified.

Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin, ""Why Should I Trust You?" Explaining the Predictions of Any Classifier", arXiv:1602.04938v3 [cs.LG] 9 Aug 2016.

However, in the conventional technology, the time-series characteristics of the data cannot be captured, so there is a problem that the interpretability of the estimation results deteriorates.

One aspect of the disclosed technique is to determine, in estimation using time-series data, a portion of the time-series data that has more influence on the estimation result.

As one aspect, the disclosed technology associates data at each time point of the time series data in chronological order with each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data. input. At this time, the disclosed technology sets the partial regression coefficient of the term of the multiple regression model corresponding to data at a time later than each of the data at each time to 0, and Calculate the regression value. Further, the disclosed technique identifies a period that satisfies a specific condition among the periods indicated by the time series data, according to changes in the multiple regression values calculated for each of the data at each point in time. Then, the disclosed technology outputs information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result.

One aspect is that in estimation using time-series data, it is possible to determine which part of the time-series data has more influence on the estimation result.

FIG. 1 is a block diagram showing a functional schematic configuration of an estimation system according to the present embodiment. FIG. 3 is a diagram for explaining learning time series data. FIG. 3 is a diagram for explaining estimation using time series data. FIG. 2 is a diagram schematically showing a learning model. It is a figure which shows an example of the list output as an estimation result. FIG. 3 is a diagram for explaining generation of a multiple regression model. FIG. 3 is a diagram for explaining problems in interpreting estimation results using existing methods. FIG. 3 is a diagram for explaining calculation of influence degree in this embodiment. FIG. 3 is a diagram for explaining identification of an important period. FIG. 3 is a diagram for explaining identification of an important period. It is a figure which shows the example of an output of an estimation reason. It is a figure which shows another output example of an estimation reason. FIG. 1 is a block diagram showing a schematic configuration of a computer that functions as an influence determination device according to the present embodiment. 7 is a flowchart illustrating an example of influence determination processing in the present embodiment.

Hereinafter, an example of an embodiment according to the disclosed technology will be described with reference to the drawings. In the following embodiment, a case will be described in which the impact determination device of the disclosed technology is applied to an estimation system that estimates the occurrence of medical absences that will lead to absence from work due to mental illness several months in the future based on employee attendance data. .

As shown in FIG. 1, an estimation system 100 according to the present embodiment includes an influence determination device 10 and a learning estimation device 30.

Functionally, the learning estimation device 30 includes a learning section 31 and an estimation section 32, as shown in FIG. Further, a learning model 40 is stored in a predetermined storage area of the learning estimation device 30.

The learning unit 31 receives learning time series data. As shown in Figure 2, the learning time-series data is based on the attendance data of each employee for a predetermined period of time, and features such as overtime work, early leaving, tardiness, etc., and whether they are late or on vacation are extracted for each date. This is the data. In the example of FIG. 2, each feature amount is represented by a block for each date for each item of overtime, leaving early, being late, vacation, and attendance, and the shaded block represents that it corresponds to each item. Note that the characteristics used as the learning data are not limited to the above example, and other characteristics such as whether or not the person has a business trip or the length of overtime work may be used. Further, each of the learning time series data is associated with a correct estimation result.

For example, as shown in Figure 3, based on attendance data for a reference period (e.g. 180 days), medical absences that lead to leave due to mental problems occur within an estimated period (e.g. 90 days) after the reference period. Let us estimate whether or not. In this case, the learning time series data is time series data for the reference period, and the correct estimation result is the presence or absence of medical absence during the estimation period.

Of the received learning time series data, the learning unit 31 uses data associated with the correct answer ``with medical absence'' as positive example learning data, and data associated with the correct answer ``no medical absence'' as negative example learning data. As learning data, a learning model 40 is trained using an existing method. FIG. 4 schematically shows the learning model 40. In FIG. 4, "1" indicates learning data for positive examples, "0" indicates learning data for negative examples, and broken lines indicate decision boundaries of the learning model 40. The learning model 40 outputs the probability that medical absence will occur during the estimated period (90 days in the above example).

The estimation unit 32 receives estimation time series data. The time series data for estimation has a data structure similar to the time series data for learning conceptually shown in FIG. 2, and the correct answer of the estimation result is unknown. The estimation unit 32 inputs the estimation time series data into the learning model 40 learned by the learning unit 31, thereby obtaining the probability that medical absence will occur as an estimation result. For example, as shown in FIG. 3, the estimation unit 32 inputs the last 180 days of attendance data of each employee into the learning model 40 as estimation time series data once a month (for example, on the 1st of every month). By doing so, the probability that medical absence will occur within the immediately following 90 days is estimated.

In the example of FIG. 3, at the estimation stage 1, the probability that medical absence will occur is low, and at the estimation stages 2 to 4, if the probability that medical absence will occur is high, the estimation is successful.

For example, as shown in FIG. 5, the estimating unit 32 outputs an estimation result in which employees are listed in descending order of probability of occurrence of medical absence. In the example in Figure 5, the list showing the estimation results includes the "employee number" of the relevant employee, the "reference period" corresponding to the period of time series data for estimation, the "estimated period", and the occurrence of medical absences. Items such as "accuracy" and "reason presentation" are included. The "reason presentation" column displays a reason presentation button for instructing interpretation of the estimation results for each employee, that is, presentation of the reason for the estimation.

The influence determination device 10 functionally includes a multiple regression model learning section 11, a calculation section 12, a specifying section 13, and an output section 14, as shown in FIG. Further, a multiple regression model 20 is stored in a predetermined storage area of the influence determination device 10.

As shown in FIG. 6, the multiple regression model learning unit 11 determines the learning model 40 using learning data surrounding the estimation time series data for the relevant employee in the feature space corresponding to the learning model 40. A multiple regression model 20 that locally approximates the boundary is generated. In FIG. 6, "1" is the learning data of positive examples, "0" is the learning data of negative examples, "a" is the target time series data for estimation, the broken line is the decision boundary of the learning model 40, and the dashed line is multiple regression. The decision boundary of model 20 is shown. This multiple regression model 20 is expressed by the following equation (1).
y= _α1x1 + _α2x2 + _{... αnxn} +β... _{( 1} )

In equation (1), y is the probability that medical absence will occur, x _i is the i-th explanatory variable, which corresponds to the i-th date from the beginning of the time-series data, and n is the number of data included in the time-series data. , that is, the number of days. Further, α _i is a partial regression coefficient for the explanatory variable x _i , and β is an intercept.

Here, in the interpretation of the estimation results using _the above-mentioned LIME technology, as shown in FIG. 7, the partial regression coefficient _α The degree of influence indicates the degree of influence. Then, the attendance data of dates whose degree of influence is equal to or higher than the threshold value is presented as attendance data that has more influenced the estimation result. In this case, the same degree of influence is calculated for dates (explanatory variables) with the same feature amount, and only the degree of influence of the date alone is known. For example, a person in charge of the human resources department, health management department, etc. takes measures such as interviews with the employees listed in the estimation results as shown in FIG. At this time, the person in charge may want to understand the characteristics of the time-series data that influenced the estimation results, such as how the degree of influence changes over time and which period has the greatest degree of influence.

Therefore, the influence determination device 10 according to the present embodiment determines the portion of the time series data that influenced the estimation result based on an index that captures the characteristics of the time series data. Hereinafter, each of the calculation section 12, the identification section 13, and the output section 14 will be explained in detail.

In the multiple regression model 20 generated by the multiple regression model learning unit 11, the calculation unit 12 sets the partial regression coefficient of the term corresponding to the date after the date corresponding to each data included in the estimation time series data to 0. In this case, the multiple regression value is calculated as the degree of influence for each date (explanatory variable).

Specifically, the calculation unit 12 calculates the degree of influence y _i for the i-th date (explanatory variable) as shown below.
y ₀ =0+0+...0+β
y ₁ =α ₁ x ₁ +0+...0+β
y ₂ =α ₁ x ₁ +α ₂ x ₂ +0+...0+β
...
y _n = α ₁ x ₁ + α ₂ x ₂ +... α _n x _n + β

As a result, as shown in FIG. 8, the degree of influence is calculated by integrating the partial regression coefficients of each date in chronological order. In this way, by integrating the partial regression coefficients of each date in chronological order to calculate the degree of influence, as shown in the upper diagram of Figure 9, the degree of influence be able to understand periods when the value is high.

The calculation unit 12 passes the calculated degree of influence y _i for each date (explanatory variable x _i ) to the identification unit 13 .

The specifying unit 13 specifies an important period, which is a period that has a greater influence on the estimation result, among the periods indicated by the estimation time series data, according to the change in the degree of influence y _i in the time series. The identifying unit 13 can identify a period in which the degree of influence y _i continuously increases in chronological order as an important period. More specifically, the identifying unit 13 calculates the slope of the influence degree y _i for each date (explanatory variable x _i ), and determines that a predetermined number of consecutive dates (explanatory variable x _i ) have a slope equal to or greater than a predetermined threshold. Periods can be identified as critical periods.

For example, the specifying unit 13 determines the degree of influence of the explanatory variable x _i using the degree of influence of surrounding explanatory variables, for example, x _i-2 , x _i-1 , x _i+1 , x _{i+2 .} Calculate the slope. The specifying unit 13 can calculate the slope b _i for the explanatory variable x _i using the following equation (2), for example.

In formula (2), x ^- (in the formula, a "bar" above x) is the average of the explanatory variable x _i and its surrounding explanatory variables, and y ^- (in the formula, a "bar" above y '') is the average degree of influence of the explanatory variable x _i and each of its surrounding explanatory variables. The middle part of FIG. 9 shows the waveform of the slope b _i in chronological order.

As shown in the lower diagram of FIG. 9 and FIG. 10, the identification unit 13 defines a period in which the slope b _i of the degree of influence is equal to or higher than a predetermined threshold value th (for example, 1.5) for a predetermined number or more consecutive periods as an important period. Specify as. If there are a plurality of periods that correspond to important periods within the period indicated by the estimation time series data, the specifying unit 13 specifies all of the periods as important periods. In this case, the identifying unit 13 may rank the important periods in descending order of the average value of the slope b _i within each important period. The identifying unit 13 notifies the output unit 14 of the identified important period.

The output unit 14 outputs the important period specified by the identification unit 13 as the estimation reason for the employee corresponding to the reason presentation button selected in the estimation result as shown in FIG. 5, for example, output from the learning estimation device 30. Output. In addition, the output unit 14 includes at least one of the data included in the important period of the estimation time series data and the aggregation result of the data included in the important period as a factor that influenced the estimation result in the estimation reason. You can also output it. FIG. 11 shows an output example of the estimated reason. In the example of FIG. 11, one important period is specified, and an example is shown in which aggregation results of data included in the important period are output together with the important period.

FIG. 12 shows another output example of the presumed reason. In the example of FIG. 12, the output unit 14 outputs the result of comparing the data of the important period and the data of the period other than the important period among the periods indicated by the estimation time-series data, as well as the important period. It is output as a factor that gave.

The influence determination device 10 can be realized, for example, by a computer 50 shown in FIG. 13. The computer 50 includes a CPU (Central Processing Unit) 51, a memory 52 as a temporary storage area, and a nonvolatile storage section 53. The computer 50 also includes an input/output device 54 such as an input section and a display section, and an R/W (Read/Write) section 55 that controls reading and writing of data to and from a storage medium 59. The computer 50 also includes a communication I/F (Interface) 56 connected to a network such as the Internet. The CPU 51, memory 52, storage section 53, input/output device 54, R/W section 55, and communication I/F 56 are connected to each other via a bus 57.

The storage unit 53 can be realized by an HDD (Hard Disk Drive), an SSD (Solid State Drive), a flash memory, or the like. An impact determination program 60 for causing the computer 50 to function as the impact determination device 10 is stored in the storage unit 53 as a storage medium. The influence determination program 60 includes a multiple regression model learning process 61 , a calculation process 62 , a specific process 63 , and an output process 64 .

The CPU 51 reads the impact determination program 60 from the storage unit 53, expands it into the memory 52, and sequentially executes the processes included in the impact determination program 60. The CPU 51 operates as the multiple regression model learning section 11 shown in FIG. 1 by executing the multiple regression model learning process 61. Further, the CPU 51 operates as the calculation unit 12 shown in FIG. 1 by executing the calculation process 62. Further, the CPU 51 operates as the specifying unit 13 shown in FIG. 1 by executing the specifying process 63. Further, the CPU 51 operates as the output unit 14 shown in FIG. 1 by executing the output process 64. Further, the CPU 51 develops the generated multiple regression model 20 in the memory 52. Thereby, the computer 50 that has executed the influence determination program 60 functions as the influence determination apparatus 10. Note that the CPU 51 that executes the program is hardware.

Note that the functions realized by the influence determination program 60 can also be realized, for example, by a semiconductor integrated circuit, more specifically, an ASIC (Application Specific Integrated Circuit).

Like the influence determination device 10, the learning estimation device 30 can also be realized by a computer including a CPU, memory, storage section, input/output device, R/W section, communication I/F, etc., so a detailed explanation will be omitted. Omitted.

Next, the operation of the estimation system 100 according to this embodiment will be explained. First, when learning time series data is input to the learning estimation device 30, the learning section 31 receives the input learning time series data. Then, the learning unit 31 performs learning to estimate the probability that medical absence will occur within a subsequent estimation period (for example, 90 days) based on the time series data for the reference period (for example, 180 days). A model 40 is generated. The learning unit 31 stores the generated learning model 40 in a predetermined storage area.

Then, when the estimation time series data is input to the learning estimation device 30, the estimation unit 32 inputs the estimation time series data to the learning model 40, thereby obtaining the probability that medical absence will occur as an estimation result. . For example, as shown in FIG. 5, the estimating unit 32 outputs an estimation result in which employees are listed in descending order of probability of occurrence of medical absence.

When a list indicating the output estimation results is displayed on the display section of an information processing device used by a person in charge of the human resources department, health management department, etc., the impact determination device 10 performs the impact determination shown in FIG. 14. Processing is executed. Note that the impact determination process is an example of the impact determination method of the disclosed technology.

In step S11, the multiple regression model learning unit 11 determines whether presentation of a reason has been instructed by determining whether any reason presentation button included in the list indicating the estimation results has been selected. If the reason presentation is instructed, the process moves to step S12, and if the reason presentation is not instructed, the process moves to step S18.

In step S12, the multiple regression model learning unit 11 searches the feature space corresponding to the learning model 40 for learning data around the estimation time series data for the employee corresponding to the selected reason presentation button. For example, the multiple regression model learning unit 11 searches for learning data in which the Euclidean distance between the vector representing the estimation time series data and the vector representing the learning data is equal to or less than a predetermined value, as peripheral learning data.

Next, in step S13, the multiple regression model learning unit 11 uses the searched surrounding learning data to locally adjust the decision boundary of the learning model 40 around the estimation time series data for the relevant employee. An approximate multiple regression model 20 is generated.

Next, in step S14, the calculation unit 12 calculates, in the multiple regression model 20, the partial regression coefficient of the term corresponding to the date after the date corresponding to each data included in the estimation time series data is set to 0. The multiple regression value is calculated as the degree of influence for each date (explanatory variable).

Next, in step S15, the specifying unit 13 calculates the slope of the degree of influence for each date (explanatory variable), for example, using equation (2).

Next, in step S16, the specifying unit 13 specifies as an important period a period in which the slope of the degree of influence is equal to or greater than a predetermined threshold value th for a predetermined number or more consecutively. The identifying unit 13 notifies the output unit 14 of the identified important period.

Next, in step S17, the output unit 14 presents the important period specified in step S16 as the reason for the estimation for the relevant employee. The output unit 14 outputs at least one of the data included in the important period of the estimation time series data and the aggregation result of the data included in the important period as factors that influenced the estimation result and included in the estimation reason. It's okay. Then, the process returns to step S11.

In step S18, it is determined whether an instruction has been given to end the display of the list showing the estimation results. If the end of list display is not instructed, the process returns to step S11, and if the end is instructed, the influence determination process ends.

As explained above, according to the estimation system according to the present embodiment, the influence determination device locally approximates the learning model that outputs the estimation result of the state in the subsequent estimation period based on the past time series data. generate a multiple regression model. Then, the influence determination device associates the data of each date of the time series data with each term of the regression equation indicating the multiple regression model in chronological order, and sets the partial regression coefficient of the term after the term corresponding to each date to 0. The multiple regression value is calculated as the degree of influence for each date. Further, the influence determination device identifies a period in which the slope of the degree of influence for each date is equal to or greater than a predetermined value for a predetermined number of consecutive periods as an important period, and outputs it as the reason for the estimation result. Thereby, in estimation using time-series data, it is possible to determine which part of the time-series data has more influence on the estimation result.

In the above embodiment, an example has been described in which the occurrence of medical absences that will lead to absence from work due to mental illness several months in the future is estimated based on the employee's attendance data, but the present invention is not limited to this. For example, it can be applied to estimation using time-series data, such as a system for predicting changes in stock prices.

Moreover, the unit of time series data is not limited to the date unit, but may also be an hourly unit, weekly unit, monthly unit, or the like. In either case, the disclosed technique can be applied by associating data at each point in time with each term of a regression equation representing a multiple regression model in chronological order.

Further, in the above embodiment, a case has been described in which the slope of the degree of influence is used as an index indicating a time-series change in the degree of influence, but the present invention is not limited to this. For example, an index that takes into account the degree of influence at surrounding times may be used, such as by taking a moving average of the degree of influence at each time point.

Further, in the above embodiment, the case where the influence degree determination device and the learning estimation device are realized by separate computers has been described, but the influence degree determination device and the learning estimation device may be realized by one computer.

Further, in the above embodiment, a mode has been described in which the influence determination program is stored (installed) in the storage unit in advance, but the present invention is not limited to this. The program according to the disclosed technology can also be provided in a form stored in a storage medium such as a CD-ROM, DVD-ROM, or USB memory.

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

(Additional note 1)
When inputting data at each point in time series data in chronological order to each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data, For each, calculate a multiple regression value by setting the partial regression coefficient of the term of the multiple regression model corresponding to data at a time after each of the data at each time as 0,
Identifying a period that satisfies a specific condition among the periods indicated by the time series data according to changes in the multiple regression values calculated for each of the data at each point in time;
outputting information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result;
An impact determination program characterized by causing a computer to execute processing.

(Additional note 2)
The process of identifying a period that satisfies the specific condition includes determining a period in which the multiple regression values calculated for each of the data at each time point continuously increase in chronological order as a period that satisfies the specific condition. It is a process to identify as
The impact determination program according to Supplementary Note 1, characterized in that:

(Additional note 3)
The process of identifying the period in which the multiple regression value continuously increases includes determining the degree of change in the time series of the multiple regression value calculated for each of the data at each time point for each of the data at each time point. is a process of calculating a period in which a predetermined number of consecutive data indicating a degree of change equal to or greater than a threshold value is present as a period in which the multiple regression value continuously increases.
The impact determination program according to appendix 2, characterized in that:

(Additional note 4)
The information related to the period that satisfies the specific condition includes information indicating the period that satisfies the specific condition, information indicating data included in the period that satisfies the specific condition among the time series data, including at least one of the aggregation results of the data included in the period that satisfies the specific condition, and the comparison result of the data included in the period that satisfies the specific condition among the time series data and other data;
The impact determination program according to any one of Supplementary Notes 1 to 3, characterized in that:

(Appendix 5)
The multiple regression model is learning data used for learning the machine learning model, and is machine learning using the learning data located around the time series data in the feature space corresponding to the machine learning model. generated by,
The impact determination program according to any one of Supplementary Notes 1 to 4, characterized in that:

(Appendix 6)
When inputting data at each point in time series data in chronological order to each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data, a calculation unit that calculates a multiple regression value by setting the partial regression coefficient of the term of the multiple regression model corresponding to data at a time later than each of the data at each time to 0 for each;
a specifying unit that specifies a period that satisfies a specific condition among the periods indicated by the time-series data according to changes in the multiple regression values calculated for each of the data at each point in time;
an output unit that outputs information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result;
An influence determination device comprising:

(Appendix 7)
The identifying unit identifies a period in which the multiple regression value calculated for each of the data at each time point continuously increases in chronological order as a period that satisfies the specific condition.
The influence determination device according to appendix 6, characterized in that:

(Appendix 8)
The identification unit calculates, for each of the data at each time, the degree of change in the multiple regression value in a time series calculated for each of the data at each time, and determines the degree of change equal to or greater than a threshold value. identifying a period in which a predetermined number of consecutive data shown as a period in which the multiple regression value continuously increases;
The influence determination device according to appendix 7, characterized in that:

(Appendix 9)
The information related to the period that satisfies the specific condition includes information indicating the period that satisfies the specific condition, information indicating data included in the period that satisfies the specific condition among the time series data, including at least one of the aggregation results of the data included in the period that satisfies the specific condition, and the comparison result of the data included in the period that satisfies the specific condition among the time series data and other data;
The influence determination device according to any one of Supplementary Notes 6 to 8, characterized in that:

(Appendix 10)
The multiple regression model is learning data used for learning the machine learning model, and is machine learning using the learning data located around the time series data in the feature space corresponding to the machine learning model. generated by,
The influence determination device according to any one of Supplementary notes 6 to 9, characterized in that:

(Appendix 11)
When inputting data at each point in time series data in chronological order to each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data, For each, calculate a multiple regression value by setting the partial regression coefficient of the term of the multiple regression model corresponding to data at a time after each of the data at each time as 0,
Identifying a period that satisfies a specific condition among the periods indicated by the time series data according to changes in the multiple regression values calculated for each of the data at each point in time;
outputting information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result;
An impact determination method characterized in that processing is executed by a computer.

(Appendix 12)
The process of identifying a period that satisfies the specific condition includes determining a period in which the multiple regression values calculated for each of the data at each time point continuously increase in chronological order as a period that satisfies the specific condition. It is a process to identify as
The influence determination method according to appendix 11, characterized in that:

(Appendix 13)
The process of identifying the period in which the multiple regression value continuously increases includes determining the degree of change in the time series of the multiple regression value calculated for each of the data at each time point for each of the data at each time point. is a process of calculating a period in which a predetermined number of consecutive data indicating a degree of change equal to or greater than a threshold value is present as a period in which the multiple regression value continuously increases.
The influence determination method according to appendix 12, characterized in that:

(Appendix 14)
The information related to the period that satisfies the specific condition includes information indicating the period that satisfies the specific condition, information indicating data included in the period that satisfies the specific condition among the time series data, including at least one of the aggregation results of the data included in the period that satisfies the specific condition, and the comparison result of the data included in the period that satisfies the specific condition among the time series data and other data;
The influence determination method according to any one of Supplementary Notes 11 to 13, characterized in that:

(Appendix 15)
The multiple regression model is learning data used for learning the machine learning model, and is machine learning using the learning data located around the time series data in the feature space corresponding to the machine learning model. generated by,
The influence determination method according to any one of Supplementary Notes 11 to 14, characterized in that:

(Appendix 16)
When inputting data at each point in time series data in chronological order to each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data, For each, calculate a multiple regression value by setting the partial regression coefficient of the term of the multiple regression model corresponding to data at a time after each of the data at each time as 0,
Identifying a period that satisfies a specific condition among the periods indicated by the time series data according to changes in the multiple regression values calculated for each of the data at each point in time;
outputting information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result;
A storage medium storing an impact determination program that causes a computer to execute processing.

10 Impact determination device 11 Multiple regression model learning section 12 Calculation section 13 Specification section 14 Output section 20 Multiple regression model 30 Learning estimation device 31 Learning section 32 Estimation section 40 Learning model 50 Computer 51 CPU
52 Memory 53 Storage unit 59 Storage medium 60 Impact determination program 100 Estimation system

Claims (7)

When inputting data at each point in time series data in chronological order to each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data, For each, calculate a multiple regression value by setting the partial regression coefficient of the term of the multiple regression model corresponding to data at a time after each of the data at each time as 0,
Identifying a period that satisfies a specific condition among the periods indicated by the time series data according to changes in the multiple regression values calculated for each of the data at each point in time;
outputting information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result;
An impact determination program characterized by causing a computer to execute processing. The process of identifying a period that satisfies the specific condition includes determining a period in which the multiple regression values calculated for each of the data at each time point continuously increase in chronological order as a period that satisfies the specific condition. It is a process to identify as
The influence determination program according to claim 1, characterized in that: The process of identifying the period in which the multiple regression value continuously increases includes determining the degree of change in the time series of the multiple regression value calculated for each of the data at each time point for each of the data at each time point. is a process of calculating a period in which a predetermined number of consecutive data indicating a degree of change equal to or greater than a threshold value is present as a period in which the multiple regression value continuously increases.
The influence determination program according to claim 2, characterized in that: The information related to the period that satisfies the specific condition includes information indicating the period that satisfies the specific condition, information indicating data included in the period that satisfies the specific condition among the time series data, including at least one of the aggregation results of the data included in the period that satisfies the specific condition, and the comparison result of the data included in the period that satisfies the specific condition among the time series data and other data;
The influence determination program according to any one of claims 1 to 3, characterized in that: The multiple regression model is learning data used for learning the machine learning model, and is machine learning using the learning data located around the time series data in the feature space corresponding to the machine learning model. generated by,
The influence determination program according to any one of claims 1 to 4, characterized in that: When inputting data at each point in time series data in chronological order to each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data, a calculation unit that calculates a multiple regression value by setting the partial regression coefficient of the term of the multiple regression model corresponding to data at a time later than each of the data at each time to 0 for each;
a specifying unit that specifies a period that satisfies a specific condition among the periods indicated by the time-series data according to changes in the multiple regression values calculated for each of the data at each point in time;
an output unit that outputs information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result;
An influence determination device comprising: When inputting data at each point in time series data in chronological order to each term of a multiple regression model that approximates a machine learning model that outputs estimation results based on time series data, For each, calculate a multiple regression value by setting the partial regression coefficient of the term of the multiple regression model corresponding to data at a time after each of the data at each time as 0,
Identifying a period that satisfies a specific condition among the periods indicated by the time series data according to changes in the multiple regression values calculated for each of the data at each point in time;
outputting information regarding a period that satisfies the identified specific condition as a factor that influenced the estimation result;
An impact determination method characterized in that processing is executed by a computer.