JP2021005191A - Generation device, generation method, and generation program - Google Patents

Abstract

To efficiently achieve highly accurate data analysis.SOLUTION: A generation device including a processor for executing a program and a storage device for storing the program can access time-lapse feature information showing a time-lapse feature acquired from time-lapse data, and grouping information in which a plurality of groups with the time-lapse data having to be belonged thereto is regulated. The processor executes: generation processing for generating a plurality of pieces of time-lapse feature data showing the time-lapse features from time-lapse data of an analysis object on the basis of the time-lapse feature information; division processing for dividing the plurality of pieces of time-lapse feature data generated by the generation processing into the plurality of groups based on the grouping information; and dimensional compression processing for performing dimensional compression of each of the plurality of groups divided by the division processing.SELECTED DRAWING: Figure 2

Description

The present invention relates to a generator, a generation method, and a generation program for generating data.

In the life insurance enrollment examination, future onset and hospitalization risk are predicted based on the health condition of the applicant. The health condition is expressed by multivariate data such as health diagnosis results and notification information. Furthermore, when considering changes in health status, the number of dimensions of multivariate data becomes even larger because risk prediction is performed in consideration of health status for multiple years.

As a method for time-series data analysis, there are a vector self-regression model and an LSTM (Long short-term memory), and there is a method of analyzing using a regression model or a neural network as an independent variable at each time point.

Further, Patent Document 1 discloses a data analyzer that analyzes a health condition for a plurality of years. This data analyzer stores quantitative data and qualitative data having ID and time information, respectively, generates time-series quantitative event data from the quantitative data, and generates time-series qualitative event data from the qualitative data. A characteristic part with a change is extracted from one of the series quantitative and qualitative event data, a time series event pattern is generated from the set of event data corresponding to the feature part, and the ID included in the time series event pattern and the time mentioned above. The ID included in the other of the series quantitative and qualitative event data is associated with the ID, and the associated time-series event pattern and the other of the associated time-series quantitative and qualitative event data are displayed.

Japanese Unexamined Patent Publication No. 2006-285672

However, when the features required for the analysis are extracted in a full search, the features unnecessary for the analysis are also generated, and the calculation cost increases. In addition, features that are not necessary for analysis may adversely affect the target analysis.

An object of the present invention is to realize efficient and highly accurate data analysis.

The generator, which is one aspect of the invention disclosed in the present application, is a generator having a processor for executing a program and a storage device for storing the program, and exhibits temporal characteristics obtained from temporal data. It is possible to access the time-lapse feature information and the grouping information in which a plurality of groups to which the time-lapse data belongs are defined, and the processor uses the time-time data to be analyzed based on the time-time feature information. A generation process for generating a plurality of time-dependent feature data showing characteristics over time, and a division process for dividing the plurality of time-related feature data generated by the generation process into the plurality of groups based on the grouping information. It is characterized by executing a dimensional compression process of dimensionally compressing each of a plurality of groups divided by the division process.

According to a typical embodiment of the present invention, efficient and highly accurate data analysis can be realized. Issues, configurations and effects other than those described above will be clarified by the description of the following examples.

FIG. 1 is a block diagram showing a hardware configuration example of the generator. FIG. 2 is a block diagram showing a functional configuration example of the generator according to the first embodiment. FIG. 3 is an explanatory diagram showing an example of stored contents of notification information. FIG. 4 is an explanatory diagram showing an example of stored contents of domain knowledge. FIG. 5 is an explanatory diagram showing an example of the characteristics over time. FIG. 6 is an explanatory diagram showing an example of the characteristics over time of division. FIG. 7 is an explanatory diagram showing an example of higher-order feature data. FIG. 8 is an explanatory diagram showing an input / output screen example 1. FIG. 9 is an explanatory diagram showing an example of a notification information input screen. FIG. 10 is an explanatory diagram showing an input / output screen example 2. FIG. 11 is a flowchart showing an example of a generation processing procedure by the generation device. FIG. 12 is a block diagram showing a functional configuration example of the generator according to the second embodiment. FIG. 13 is an explanatory diagram showing an example of a multimodal neural network. FIG. 14 is an explanatory diagram showing an example of an input / output screen showing an analysis result by a multimodal neural network.

Hereinafter, the generator according to the present invention will be described with reference to the accompanying drawings. This specification shows an example of insurance payment risk prediction in the underwriting assessment of life insurance. In the underwriting assessment, future insurance payment risk is assessed based on the information notified by the contract applicant (hereinafter referred to as notification information), and approval or abandonment of insurance coverage is decided. The notification information includes the test result of the medical examination, the interview, the medical history, and the like.

<Hardware configuration example of generator>
FIG. 1 is a block diagram showing a hardware configuration example of the generator. The generation device 100 includes a processor 101, a storage device 102, an input device 103, an output device 104, and a communication interface (communication IF) 105. The processor 101, the storage device 102, the input device 103, the output device 104, and the communication IF 105 are connected by the bus 106. The processor 101 controls the generator 100. The storage device 102 serves as a work area for the processor 101. Further, the storage device 102 is a non-temporary or temporary recording medium for storing various programs and data. Examples of the storage device 102 include a ROM (Read Only Memory), a RAM (Random Access Memory), an HDD (Hard Disk Drive), and a flash memory. The input device 103 inputs data. The input device 103 includes, for example, a keyboard, a mouse, a touch panel, a numeric keypad, and a scanner. The output device 104 outputs data. The output device 104 includes, for example, a display and a printer. The communication IF 105 connects to the network and transmits / receives data.

<Example of functional configuration of the generator 100>
FIG. 2 is a block diagram showing a functional configuration example of the generator 100 according to the first embodiment. The generation device 100 includes a determination unit 201, a generation unit 202, a division unit 203, a dimension compression unit 204, a coupling unit 205, and an analysis unit 206. Specifically, the determination unit 201, the generation unit 202, the division unit 203, the dimension compression unit 204, the coupling unit 205, and the analysis unit 206 specifically, for example, process the program stored in the storage device 102 shown in FIG. 1 in the processor 101. It is realized by letting it execute.

Further, the generation device 100 only needs to have at least a generation unit 202, a division unit 203, and a dimensional compression unit 204, and the determination unit 201, the coupling unit 205, and the analysis unit 206 can communicate with the generation device 100. It may be realized by another computer.

Further, the generation device 100 stores the notification information 300 and the domain knowledge 400 in the storage device 102. The notification information 300 and the domain knowledge 400 may be stored in the generator 100 in advance, or may be acquired from another computer capable of communicating with the generator 100. First, the notification information 300 will be described in detail.

FIG. 3 is an explanatory diagram showing an example of stored contents of the notification information 300. The notification information 300 is information necessary for the insurance contract notified by the contract applicant, and is the data to be analyzed. The notification information 300 includes basic notification information 310, a medical examination result 320, an interview result 330, and a medical history 340. The notification basic information 310 is basic information regarding notification of a contract applicant. The notification basic information 310 includes a name ID 311 and a date of birth 312 and an age 313.

The name ID 311 is identification information that uniquely identifies the contract applicant. Date of birth 312 is the date of birth of the contract applicant. The three entries of the contract applicant whose name ID 311 is "0001" in FIG. 3 indicate the data to be analyzed for the past three years of the contract applicant. Age 313 is the number of years elapsed from the date of birth 312 of the contract applicant. In the example described later, for the three entries of the contract applicant whose name ID 311 is "0001", the age 313 is "47" in the first year of the time series, "48" in the second year of the time series, and "49". This is the third year of the time series.

The medical examination result 320 is the result of the medical examination received by the contract applicant. The medical examination result 320 includes a body weight 321, a BMI (Body Mass Index) 322, a systolic blood pressure 323, a diastolic blood pressure 324, and a fasting blood glucose 325. The weight 321 is the body weight of the contract applicant. BMI322 is a physique index representing the degree of obesity in humans, and is calculated by body weight / (height ² ). BMI322 indicates that the smaller the value, the thinner the body, and the larger the value, the thicker the body.

The systolic blood pressure 323 is the pressure exerted on the blood vessel wall of the aorta by the blood pushed out by the contraction of the heart in the state of pumping blood from the heart to the aorta. Diastolic blood pressure 324 is the pressure exerted on the aortic vessel wall, which is reduced by the influx of blood from the aorta into the heart and the decrease in blood volume in the aorta when the blood returns to the heart. Fasting blood glucose 325 is a blood glucose level measured in a fasting state.

The interview result 330 is the result of the interview received by the contract applicant. The interview result 330 includes smoking habit 331, drinking habit 332, and exercise habit 333. The smoking habit 331 is the presence / absence, frequency, and amount of smoking of the contract applicant. The drunken habit 332 is the presence / absence, frequency, and amount of drunkenness of the contract applicant. The exercise habit 333 is the presence / absence, frequency, and amount of exercise of the contract applicant.

The medical history 340 is a history that the contract applicant has already received or been hospitalized. The medical history 340 includes a history of hypertension consultation 341, a history of hypertension hospitalization 342, and a history of diabetes consultation 343. The hypertension consultation history 341 is a history of a contract applicant having a medical examination regarding hypertension. The history of hospitalization for hypertension 342 is a history of hospitalization for hypertension by a contract applicant. Diabetes consultation history 343 is a history of a contract applicant having a consultation regarding diabetes.

FIG. 4 is an explanatory diagram showing an example of stored contents of the domain knowledge 400. The domain knowledge 400 is qualitative information for various information such as a medical examination result 320, an interview result 330, and a medical history 340 included in the notification information 300, and corresponds to medical knowledge. Specifically, for example, the domain knowledge 400 includes time-time data determination knowledge 410, time-time feature knowledge 420, and time-time feature division knowledge 430.

The time-lapse data determination knowledge 410 is determination information for determining whether the notification information 300 of the contract applicant, which is the analysis target data, corresponds to the non-time-lapse data 231 or the time-lapse data 232. Specifically, for example, the time-dependent data determination knowledge 410 defines the items of the notification information 300 corresponding to the non-time-lapse data 231 and the time-lapse data 232.

The non-time data item 411 includes an item corresponding to the non-time data 231. The non-periodic data 231 is data in which there is no time-series change in the data, or even if there is such a change, it is meaningless. The non-time-lapse data 231 corresponds to, for example, smoking habits 331, drinking habits 332, various consultation histories, and various hospitalization histories. The time-lapse data item 412 includes an item corresponding to the time-lapse data 232. The time-series data 232 is time-series data in which the time-series changes in the data are meaningful. The time-dependent data 232 includes, for example, body weight 321 and BMI 322, systolic blood pressure 323, diastolic blood pressure 324, and fasting blood glucose 325.

The temporal feature knowledge 420 is temporal feature information indicating the temporal characteristics (temporal features) obtained from the temporal data 232. For example, the temporal feature knowledge 420 defines the basic statistic item 421, the change amount item 422, the change rate item 423, and so on. The basic statistic item 421 includes an item corresponding to the basic statistic obtained from the time-dependent data 232. The basic statistic is, for example, the maximum value, the minimum value, and the average value of the time-dependent data 232.

The change amount item 422 is an item corresponding to the change amount indicating the change of two consecutive values in the time-dependent data 232. For example, when the time-dependent data 232 is the annual weight 321 for 1 to 3 years, the amount of change (1st and 2nd years) is the difference between the 1st year weight 321 and the 2nd year weight 321. The amount of change (2nd and 3rd years) is the difference between the body weight 321 in the 2nd year and the body weight 321 in the 3rd year.

The change rate item 423 is an item corresponding to a value indicating the rate of change of two consecutive values in the time-dependent data 232. For example, when the time-dependent data 232 is the annual body weight 321 for 1 to 3 years, the change rate (1st and 2nd years) is the difference between the 1st year body weight 321 and the 2nd year body weight 321. It is the value divided by the weight 321 of the year, and the amount of change (2nd and 3rd years) is the value obtained by dividing the difference between the weight 321 of the 2nd year and the weight 321 of the 3rd year by the weight 321 of the 2nd year. Is.

The temporal feature division knowledge 430 is grouping information in which the group to which the temporal data 232 should belong is defined. The temporal feature division knowledge 430 is specifically defined based on, for example, statistical knowledge or medical knowledge. Specifically, for example, the time-dependent characteristic division knowledge groups the body shape basic information item 431, the blood pressure system test value item 432, the blood glucose system test value item 433, the liver function system test value item 434, and so on, respectively. Include as.

The body shape basic information item 431 includes an item corresponding to the body shape basic information. The basic body shape information is basic information about the body shape of the contract applicant. The body shape basic information item 431 includes, for example, age 313, body weight 321 and BMI 322 as items.

The blood pressure system test value item 432 includes an item corresponding to the blood pressure system test value. The blood pressure system test value is a test value related to the blood pressure of the contract applicant. The blood pressure system test value item 432 includes, for example, systolic blood pressure 323 and diastolic blood pressure 324 as items.

The blood glucose test value item 433 includes an item corresponding to the blood glucose test value. The blood glucose system test value is a test value related to blood glucose of a contract applicant. The blood glucose system test value item 433 includes, for example, fasting blood glucose 325 and HbA1c as items.

The liver function test value item 434 includes an item corresponding to the liver function test value. Liver function test values are test values related to liver function of contract applicants. Liver function test value item 434 includes, for example, GOT (glutamate oxaloacetate transaminase), GPT (glutamate pyruvate transaminase), and γ-GTP (γ-glutamyl transpeptidase).

Returning to FIG. 2, the determination unit 201 determines whether the notification information 300 corresponds to the temporal data 232 or the non-chronological data 231 based on the temporal data determination knowledge 410. An entry in which the name ID 311 in the notification information 300 of FIG. 3 is "0001" is taken as an example. Age 313, body weight 321 and BMI 322, systolic blood pressure 323, diastolic blood pressure 324 and fasting blood glucose 325 are included in the temporal data item 412. Therefore, the determination unit 201 has age 313 "47", "48", "49", weight 321 "83.4", "86.6", "92.0", and BMI 322 "22.8". , "24.3", "26.0", systolic blood pressure 323 "124.9", "128.5", "133.8", diastolic blood pressure 324 "80.7", "86. "1", "90.0", "104.5", "107.2", "110.0" of fasting blood pressure 325 are time-lapse data 232 of the contract applicant whose name ID 311 is "0001". , Is determined. The determination unit 201 outputs the determined time-dependent data 232 as a determination result.

In addition, smoking habits 331, drinking habits 332, exercise habits 333, hypertension consultation history 341, hypertension hospitalization history 342 and diabetes consultation history 343 are included in the non-time-lapse data item 411. Therefore, the determination unit 201 has "none", "none", "none" of smoking habit 331, "week 1", "week 1", "week 2" of drinking habit 332, and "week 1" of exercise habit 333. , "Week 1", "None", Hypertension consultation history 341 "None", "None", "None", Hypertension hospitalization history 342 "None", "None", "None", Diabetes consultation history 343 "None", "None", and "None" are determined to be non-time data 231 of the contract applicant whose name ID 311 is "0001". The determination unit 201 outputs the determined non-time-lapse data 231 as a determination result.

The generation unit 202 generates the temporal feature data showing the temporal characteristics from the temporal data 232 based on the temporal characteristic knowledge 420. The generation unit 202 calculates the maximum value 511, the minimum value 512, the average value, ... Included in the basic statistic items for the time-dependent data 232. The generation unit 202 calculates the amount of change (1st and 2nd year) 521, the amount of change (2nd and 3rd year) 522, and the like included in the change amount item for the time-dependent data 232. The generation unit 202 calculates the change rate (1st and 2nd year) 531 and the change rate (2nd and 3rd year) 532, etc. included in the change rate item for the time-dependent data 232.

An entry in which the name ID 311 in the notification information 300 of FIG. 3 is "0001" is taken as an example. For example, in the case of "83.4", "86.6", and "92.0" of the body weight 321 included in the time data 232, the maximum value 511 is "92.0", which is the body weight 321 of the third year. The minimum value 512 is "83.4", which is the weight of 321 in the first year, and the average value is "87.3", which is the average of "83.4", "86.6", and "92.0". is there.

The amount of change (1st and 2nd year) 521 is "3.2", which is obtained by subtracting "83.4" from "86.6", and the amount of change (2nd and 3rd year) 522 is "92.0". It is "5.4" which is obtained by subtracting "86.6" from "". The rate of change (1st and 2nd year) 531 is calculated by dividing "3.2", which is the amount of change (1st and 2nd year) 521, by "83.4" of the weight 321 in the first year. "04", the rate of change (2nd and 3rd years) 532 is "0" obtained by dividing "5.4", which is the amount of change (2nd and 3rd years) 522, by "86.6" of the weight 321 in the second year. It is ".06". The same applies to BMI 322 other than body weight 321, systolic blood pressure 323, diastolic blood pressure 324, fasting blood glucose 325 and the like. The generation unit 202 outputs the generation result as the time-dependent feature data 500.

FIG. 5 is an explanatory diagram showing an example of time-dependent feature data 500. The temporal feature data 500 includes a basic statistic 510, a change amount 520, and a change rate 530. The basic statistic 510 includes a maximum value 511, a minimum value 512, an average value (not shown), ... Calculated according to the basic statistic item 421. The change amount 520 includes the change amount (1st and 2nd year) 521, the change amount (2nd and 3rd year) 522, ... Calculated according to the change amount item 422. The change rate 530 includes the change rate (1st and 2nd year) 531 calculated according to the change amount item 422, the change rate (2nd and 3rd year) 532, and so on.

The basic statistic 510, the change amount 520, and the change rate 530 are generated by the generation unit 202 for each type of time-dependent data 232, respectively. The entries 501-1, 501-2, ..., 502-1, 502-2, ..., 501-3, 503-2, ..., 504-1, 504-2, ... In the temporal features shown in FIG. , It is time-dependent characteristic data about a certain contract applicant (for example, a contract applicant whose name ID 311 is "0001"). The generation unit 202 has entries 501-1, 501-2, ..., 502-1, 502-2, ..., 501-3, 503-2, ..., 504-1, 504-2, ..., For each contract applicant. To generate.

Entry 501 is time-dependent feature data according to body shape basic information item 431. Specifically, for example, entry 501-1 is temporal feature data showing basic statistics, amount of change and rate of change for body weight 321. Entry 501-2 is temporal feature data showing basic statistics, amount of change and rate of change for BMI322.

Entry 502 is temporal feature data according to blood pressure system test value item 432. Specifically, for example, entry 502-1 is temporal feature data showing basic statistics, amount of change and rate of change for systolic blood pressure 323. Entry 502-2 is temporal feature data showing basic statistics, amount of change and rate of change for diastolic blood pressure 324.

Entry 503 is temporal feature data according to blood glucose test value item 433. Specifically, for example, entry 503-1 is temporal feature data showing basic statistics, changes, and rates of change for fasting blood glucose 325. Entry 503-2 is temporal feature data showing the basic statistics, amount of change and rate of change for HbA1c.

Entry 504 is temporal feature data according to liver function test value item 434. Specifically, for example, entry 504-1 is temporal feature data showing basic statistics, amount of change, and rate of change for GOT. Entry 504-2 is temporal feature data showing the basic statistics, amount of change and rate of change for GPT.

Returning to FIG. 2, the division unit 203 divides the plurality of time-dependent feature data generated by the generation unit 202 into a plurality of groups based on the time-dependent feature division knowledge 430, and outputs the divided time-dependent feature data 600. The plurality of temporal feature data are, for example, the respective entries 501-1, 501-2, ..., 502-1, 502-2, ..., 501-3, 503-2, ..., 504-1 shown in FIG. , 504-2, ..., Each of the values constituting. The plurality of groups are divided time-dependent feature data 600-1, 600-2, ..., 600-n (n is an integer of 1 or more). Each of the divided temporal feature data 600-1, 600-2, ..., 600-n is the body shape basic information item 431, the blood pressure system test value item 432, and the blood glucose system test value item of the temporal feature division knowledge 430 shown in FIG. It is defined as 433, liver function test value item 434, ...

The temporal feature division knowledge 430 is defined based on statistical or medical findings. (Example 1) BMI 322 is an index calculated from height and weight 321. If it is the age of the insurance contract applicant, the change in height is not large. Therefore, the change in BMI 322 has a very strong correlation with the change in body weight 321. If there are a plurality of highly correlated items in the notification information 300, they become redundant information and cause inefficiency in data analysis.

On the other hand, by applying the temporal feature division knowledge 430, the temporal feature data of a plurality of items (age 313, weight 321, BMI 322, ...) Including redundancy can be divided into the temporal feature data 600 of the body shape basic information item 431. It is grouped into a group of -1 (hereinafter referred to as body type group 600-1). As a result, the generation device 100 can extract the high-dimensional features (hereinafter referred to as high-order features) by collectively compressing the time-dependent feature data generated using the values of the plurality of items as a body shape group. This makes it possible to improve the efficiency of data analysis.

(Example 2) It is known that both systolic blood pressure 323 and diastolic blood pressure 324 gradually deteriorate with aging. Further, when the blood pressure decreases or increases due to improvement or deterioration of lifestyle, the systolic blood pressure 323 and the diastolic blood pressure 324 decrease and increase while maintaining a balance. However, when the balance changes, it is said to be a sign of hypertensive diseases such as arteriosclerosis.

Therefore, by applying the temporal feature division knowledge 430, the temporal characteristic data of the systolic blood pressure 323 and the diastolic blood pressure 324 are the group of the divided temporal characteristic data 600-2 called the blood pressure system test value item 432 (hereinafter, the blood pressure system group). It is summarized in 600-2). As a result, the generator 100 collects the temporal feature data generated using the value of systolic blood pressure 323 and the value of diastolic blood pressure 324 as a blood pressure system group and dimensionally compresses them to extract higher-order features. Therefore, the complex features required for data analysis can be obtained.

For the same reason, the temporal characteristic data of fasting blood glucose 325 and HbA1c are summarized in the group of divided temporal characteristic data 600-3 (hereinafter referred to as blood glucose group 600-3) called blood glucose test value item 433, and GOT, GPT. , Γ-GTP temporal feature data is summarized in a group of divided temporal feature data 600-4 (hereinafter, liver function system group 600-4) called liver function test value item 434.

FIG. 6 is an explanatory diagram showing an example of the divided time-dependent feature data 600. The divisional time-dependent features include body type group 600-1, blood pressure group 600-2, blood glucose group 600-3, and liver function group 600-4. The body shape group 600-1 is a data set including time-dependent feature data corresponding to the body shape basic information item 431. The blood pressure system group 600-2 is a data set including time-dependent characteristic data corresponding to the blood pressure system test value item 432. The blood glucose system group 600-3 is a data set including time-dependent characteristic data corresponding to the blood glucose system test value item 433. The liver function system group 600-4 is a data set including time-dependent feature data corresponding to the liver function system test value item 434.

The body type group 600-1 includes time-related characteristic data 601-1 of body weight 321 and time-related characteristic data 601-2 of BMI 322. Let the data string of the temporal feature data 601-1 of the body weight 321 be the vector Ub1-1. Let the data sequence of the time-dependent feature data 601-2 of BMI322 be the vector Ub1-2.

Blood pressure group 600-2 includes temporal characteristic data 602-1 for systolic blood pressure 323 and temporal characteristic data 602-2 for diastolic blood pressure 324. The data sequence of the temporal characteristic data 602-1 of the systolic blood pressure 323 is defined as the vector Ub2-1. Let the data string of the time-dependent characteristic data 602-2 of the diastolic blood pressure 324 be the vector Ub2-2.

The blood glucose system group 600-3 includes time-related characteristic data 603-1 of fasting blood glucose 325 and time-related characteristic data 602-2 of HbA1c. Let the data sequence of the time-dependent characteristic data 603-1 of the fasting blood glucose 325 be the vector Ub3-1. Let the data string of the time-dependent feature data 602-2 of HbA1c be the vector Ub3-2.

Liver function group 600-4 includes GOT temporal feature data 604-1 and GPT temporal characteristic data 604-2. Let the data string of the time-dependent feature data 604-1 of GOT be the vector Ub4-1. Let the data string of the time-dependent feature data 604-2 of GPT be the vector Ub4-2.

Returning to FIG. 2, the dimensional compression unit 204 dimensionally compresses the input data to generate higher-order feature data 700. Specifically, for example, the dimensional compression unit 204 dimensionally compresses each of the plurality of groups divided by the division unit 203. Specifically, for example, the dimensional compression unit 204 performs dimensional compression on each of the divided chronological feature data 600-1 to 600-n to generate higher-order feature data 702 to 702-n related to the chronological data 232. ..

Further, the dimensional compression unit 204 performs dimensional compression on the non-time-lapse data 231 and generates higher-order feature data 701 regarding the non-time-lapse data 231. Extraction of higher-order feature data by dimensional compression of data is realized by a known dimensional compression method such as Principal components analysis (PCA) or Stacked autoencoder. Further, the dimension compression unit 204 may generate high-order feature data by expanding the number of dimensions of the data once by using a neural network or the like, and then perform dimension compression.

FIG. 7 is an explanatory diagram showing an example of higher-order feature data. The higher-order feature data includes higher-order feature data 701 related to non-time-lapse data 231, body-type higher-order feature data 702-1 related to time-lapse data 232, blood pressure-based higher-order feature data 702-2 related to time-lapse data 232, and time-lapse data. Includes glycemic system higher-order feature data 702-3 for 232 and hepatic function system higher-order feature data 702-4 for time-dependent data 232.

Higher-order feature data 701 for non-time data 231, body-type higher-order feature data 702-1 for time-lapse data 232, blood-blood-type higher-order feature data 702-2 for time-lapse data 232, and blood-type higher-order feature data 702 for time-lapse data 232 -3, and hepatic functional system higher-order feature data 702-4 relating to time-dependent data 232, respectively, include feature amount 1, feature amount 2, ... Obtained by dimensional compression.

The set of values in the column of the feature amount 1 of the higher-order feature data 701 regarding the non-periodic data 231 is the vector Va1, and the set of the values in the column of the feature amount 2 is the vector Va2. The set of values in the column of feature amount 1 of the body type system higher-order feature data 702-1 regarding the time-dependent data 232 is the vector Vba1-1, and the set of values in the column of feature amount 2 is the vector Vb1-2.

The set of values in the column of feature amount 1 of the blood pressure system higher-order feature data 702-2 with respect to the time-dependent data 232 is the vector Vb2-1, and the set of values in the column of feature amount 2 is the vector Vb2-2. The set of values in the column of feature amount 1 of the blood glucose system higher-order feature data 702-3 with respect to the time-dependent data 232 is the vector Vb3-1, and the set of values in the column of feature amount 2 is the vector Vb3-2. The set of values in the column of feature amount 1 of the liver function system higher-order feature data 702-4 with respect to the time-dependent data 232 is the vector Vb4-1, and the set of values in the column of feature amount 2 is the vector Vb4-2.

Returning to FIG. 2, the coupling unit 205 combines a plurality of groups after the dimension compression by the dimension compression unit 204 to generate the combined higher-order feature data 710. The plurality of groups after the dimensional compression are higher-order feature data 702 to 702-n regarding the chronological data 232 when the divided chronological feature data 600-1 to 600-n are dimensionally compressed. When the non-time-lapse data 231 and the divided time-lapse feature data 600-1 to 600-n are dimensionally compressed, the higher-order feature data 701 for the non-time-lapse data 231 and the higher-order feature data 702 to 702-n for the time-lapse data 232 is there.

FIG. 7 shows an example of coupling by the coupling portion 205. The coupling portion 205 connects Va1, Va2, ..., Vb1-1, Vb1-2, ..., Vb3-1, Vb3-2, ..., Vb4-1, Vb4-2, ..., And the higher-order feature vector Val. Is generated as the combined high-order feature data 710. The number of dimensions of the higher-order feature vector Val is the number of dimensions of each element of Va1, vA2, ..., Vb1-1, Vb1-2, ..., Vb3-1, Vb3, 2, ..., Vb4-1, Vb4-2, ... It is the sum.

Returning to FIG. 2, the analysis unit 206 uses the combination result (combination higher-order feature data 710) of the connection unit 205 as an explanatory variable and outputs the corresponding objective variable. For example, the analysis unit 206 performs insurance payment risk analysis, and specifically outputs future occurrence probabilities such as death, hospitalization, surgery, and outpatient treatment as objective variables. Specifically, for example, the analysis unit 206 executes data analysis using a known technique such as a statistical method such as multiple regression analysis or a machine learning method such as a neural network. Specifically, for example, the analysis unit 206 generates a learning model using the combination of the higher-order feature vector Val obtained from the existing notification information 300 and the analysis result as training data, and obtains it from the new notification information 300. By inputting the higher-order feature vector Val into the training model, a new analysis result corresponding to the new notification information 300 is obtained.

<Screen example>
FIG. 8 is an explanatory diagram showing an input / output screen example 1. The input / output screen 800 is displayed on a display that is an example of the output device 104 of the generation device 100 or a display of another computer that can communicate with the generation device 100.

The input / output screen 800 includes a notification information read button 801, a domain knowledge read button 802, a feature extraction method selection pull-down 803, an analysis method selection pull-down 804, an analysis execution button 805, and an execution result display area 806. .. The notification information reading button 801 is a button pressed by the input device 103. When the notification information reading button 801 is pressed, the notification information 300 of the contract applicant stored in the storage device 102 is read.

In addition to the method of pressing the notification information reading button 801, the notification information 300 can display the notification information input screen by pressing the notification information input button 807 and can be input by the input device 103. FIG. 9 is an explanatory diagram showing an example of a notification information input screen. The notification information input screen 900 includes a medical examination result input area 901 and a medical inquiry result input area 902. In the medical examination result input area 901, values such as systolic blood pressure 323, diastolic blood pressure 324, and fasting blood glucose 325 can be set by the input device 103. In the interview result input area 902, the presence / absence of smoking habits 331, drinking habits 332, exercise habits 333, etc. can be set by the input device 103.

Returning to FIG. 8, the domain knowledge reading button 802 is a button pressed by the input device 103. When the domain knowledge read button 802 is pressed, the domain knowledge 400 stored in the storage device 102 is read. Alternatively, by pressing the domain knowledge input button 808, a setting screen (not shown) for inputting domain knowledge can be displayed. On the setting screen, the input device 103 enables addition, change, and deletion of the contents of the time-lapse data determination knowledge 410, the time-lapse feature knowledge 420, and the time-lapse feature division knowledge 430.

The feature extraction method selection pull-down 803 is a button for displaying a plurality of feature extraction methods in a pull-down display on the input device 103 and selecting one of them. The plurality of feature extraction methods include known dimensional compression methods such as the above-mentioned PCA and Stacked autoencoder. For example, when PCA is selected, the generator 100 will perform dimensional compression on the PCA.

The analysis method selection pull-down 804 is a button for displaying a plurality of analysis methods in a pull-down display on the input device 103 and selecting one of them. The plurality of analysis methods include known methods such as the above-mentioned multiple regression analysis and other statistical methods and machine learning methods such as neural networks. For example, if multiple regression analysis is selected, the generator 100 will perform data analysis in multiple regression analysis.

The analysis execution button 805 is a button pressed by the input device 103. When the analysis execution button is pressed, the generation device 100 loads the notification information 300 and the domain knowledge 400 from the storage device 102, performs dimension compression by the method selected by the feature extraction method selection pull-down 803, and performs the analysis method selection pull-down. Data analysis is performed by the method selected in 804. The execution result display area 806 is an area in which the execution result of the data analysis executed by pressing the analysis execution button 805 is displayed.

FIG. 10 is an explanatory diagram showing an input / output screen example 2. In FIG. 10, the analysis result 1000 for each name ID 311 is displayed in the execution result display area 806.

<Example of generation processing procedure>
FIG. 11 is a flowchart showing an example of a generation processing procedure by the generation device 100. By pressing the analysis execution button 805, the generation device 100 reads the notification information 300 and the domain knowledge 400 (step S1101), and for each of the analysis target data of the contract applicant in the notification information 300, it is time-lapse data 232 or not. The determination unit 201 determines whether the data is the time-lapse data 231 (step S1102). Regarding the data determined to be the time-lapse data 232 (step S1102: Yes), the generation device 100 generates the time-lapse feature data 500 by the generation unit 202 (step S1103), and the division unit 203 generates a plurality of time-lapse feature data. (Step S1104).

Regarding the data determined to be the non-time-lapse data 231 (step S1102: No), the generation device 100 extracts the higher-order feature data 701 related to the non-time-lapse data 231 by the dimension compression unit 204 (step S1105). Similarly, the generation device 100 extracts the higher-order feature data 702 to 702-n regarding the time-dependent data 232 by the dimension compression unit 204 (steps S1106-1 to S1106-n).

Then, the generation device 100 combines the higher-order feature data 701 and 702 to 702-n by the coupling unit 205 to generate the combined higher-order feature data 710 (step S1107), and the analysis unit 206 performs data analysis. It is executed (step S1108), and the analysis result 1000 is displayed in the execution result display area 806 of the input / output screen 800. As a result, a series of generation processes is completed.

As described above, according to the first embodiment, since the temporal feature data 500 is grouped according to the temporal feature division knowledge 430, it is possible to suppress the generation of the feature amount unnecessary for the analysis. As a result, high-quality explanatory variables can be generated, and the accuracy of data analysis can be improved. In addition, by suppressing the generation of features unnecessary for branching, the calculation cost can be reduced and the calculation efficiency in data generation and data analysis can be improved.

Next, the second embodiment will be described. In the second embodiment, since the differences from the first embodiment will be mainly described, the same components as those in the first embodiment are designated by the same reference numerals, and the description thereof will be omitted.

FIG. 12 is a block diagram showing a functional configuration example of the generator 100 according to the second embodiment. The generator 100 according to the first embodiment executed the dimensional compression process by the dimensional compression unit 204, the coupling process by the coupling unit 205, and the analysis process by the analysis unit 206 as independent processes, but in the second embodiment, the dimensional compression process was performed. By applying the multimodal neural network 1200 instead of the unit 204, the coupling unit 205, and the analysis unit 206, the dimensional compression processing by the dimensional compression unit 204, the coupling processing by the coupling unit 205, and the analysis processing by the analysis unit 206 are continuously performed. Execute.

FIG. 13 is an explanatory diagram showing an example of the multimodal neural network 1200. FIG. 14 is an explanatory diagram showing an example of an input / output screen showing an analysis result 1400 by the multimodal neural network 1200.

First, the multimodal neural network 1200 performs feature extraction on the input vectors classified into a plurality of groups by the neural networks f1, f2, ..., Fn branched in each group. Next, the multimodal neural network 1200 combines the output vectors of the neural networks f1, f2, ..., Fn, performs feature extraction and analysis by the fully connected network g, and outputs the analysis result on the output layer h. The neural networks f1, f2, ..., Fn correspond to the dimensional compression processing, and the fully coupled network g and the output layer h correspond to the coupling processing and the analysis processing.

The learning of the multimodal neural network 1200 is supervised learning of the input vectors input to the neural networks f1, f2, ..., Fn and the analysis results which are the output values. The multimodal neural network 1200 can simultaneously learn three processes of feature extraction for each group, higher-order feature extraction by combining all groups, and analysis. Therefore, highly accurate analysis is possible depending on the network structure and parameter design. The operation of the multimodal neural network 1200 is expressed by, for example, the equation (1) shown in FIG.

h = g (f (Ua), ..., f (Ub1-1), f (Ub1-2), ..., f (Ub3-1), f (Ub3-2), ..., f (Ub4-1), f (Ub4-2), ...) ... (1)

In the equation (1), the vector Ua is a vector representation of the non-periodic data 231. By giving the vector Ua to the function f, the vectors Va1, Va2, ... Of the higher-order feature data 701 related to the non-periodic data 231 are generated. Further, by giving the vectors Ub1-1, vectors Ub1-2, ... To the function f, the vectors Vb1-1, Vba1-2, ... Of the higher-order feature data 702-1 regarding the time-dependent data 232 are generated.

Further, by giving the vectors Ub2-1, the vectors Ub2-2, ... To the function f, the vectors Vb2-1, Vba2-2, ... Of the higher-order feature data 702-2 relating to the time-dependent data 232 are generated. Further, by giving the vectors Ub3-1, the vectors Ub3-2, ... To the function f, the vectors Vb3-1, Vba3-2, ... Of the higher-order feature data 702-3 regarding the time-dependent data 232 are generated. Further, by giving the vectors Ub4-1, the vectors Ub4-2, ... To the function f, the vectors Vb4-1, Vba4-2, ... Of the higher-order feature data 702-4 relating to the time-dependent data 232 are generated.

As described above, according to the second embodiment, the dimensional compression process by the dimensional compression unit 204, the combination process by the coupling unit 205, and the analysis process by the analysis unit 206 are continuously executed, so that the generation process and the analysis process are high-speed. It is possible to improve the accuracy and accuracy.

Further, in the above-mentioned Examples 1 and 2, the insurance claim payment risk prediction in the underwriting assessment of life insurance has been described as an example, but it can also be applied to the financial analysis of a company. In this case, instead of the notification information 300, the data described in the securities report or the index data calculated from the data is used. Further, the temporal feature division knowledge 430 is, for example, information obtained by grouping the above data from the five viewpoints of profitability, safety, activity, productivity and growth potential.

Profitability is an item that indicates how much a company is making a profit. Total profit margin on sales, operating profit margin on sales, ordinary profit margin on total capital (ROA), return on equity (ROE) including. Safety is an item that indicates a company's ability to pay, such as the ability to repay loans from banks, and includes the current ratio, quick ratio, operating cash flow, investment cash flow, and so on. Activity is an item that indicates whether capital is used efficiently and a large amount of sales are being made, and includes total capital turnover, fixed asset turnover, inventory turnover, and the like.

Productivity is an item that indicates whether or not a company is efficiently utilizing employees and equipment, and includes sales value-added rate, labor share, labor productivity, and so on. Growth potential is an item that indicates the future growth potential of a company, and includes sales growth rate, ordinary profit growth rate, net income growth rate, and so on. As described above, the generator 100 according to the first and second embodiments can be applied to various data analyzes.

It should be noted that the present invention is not limited to the above-described examples, but includes various modifications and equivalent configurations within the scope of the attached claims. For example, the above-described examples have been described in detail in order to explain the present invention in an easy-to-understand manner, and the present invention is not necessarily limited to those having all the described configurations. In addition, a part of the configuration of one embodiment may be replaced with the configuration of another embodiment. In addition, the configuration of another embodiment may be added to the configuration of one embodiment. In addition, other configurations may be added, deleted, or replaced with respect to a part of the configurations of each embodiment.

Further, each of the above-described configurations, functions, processing units, processing means, etc. may be realized by hardware by designing a part or all of them by, for example, an integrated circuit, and the processor 101 performs each function. It may be realized by software by interpreting and executing the program to be realized.

Information such as programs, tables, and files that realize each function is recorded in a storage device such as a memory, a hard disk, an SSD (Solid State Drive), or an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versaille Disc). It can be stored in a medium.

In addition, the control lines and information lines indicate those that are considered necessary for explanation, and do not necessarily indicate all the control lines and information lines necessary for implementation. In practice, it can be considered that almost all configurations are interconnected.

100 Generator 101 Processor 102 Storage device 201 Judgment unit 202 Generation unit 203 Divided unit 204 Dimensional compression unit 205 Combined unit 206 Analytical unit 300 Notification information 400 Domain knowledge 500 Time-lapse feature data 600 Next feature data 1200 multimodal neural network

Claims (10)

A generator having a processor that executes a program and a storage device that stores the program.
It is possible to access the temporal feature information showing the temporal characteristics obtained from the temporal data and the grouping information in which a plurality of groups to which the temporal data should belong are defined.
The processor
A generation process for generating a plurality of temporal feature data showing the temporal features from the temporal data to be analyzed based on the temporal feature information.
A division process for dividing a plurality of time-dependent feature data generated by the generation process into the plurality of groups based on the grouping information,
A dimensional compression process that dimensionally compresses each of the plurality of groups divided by the division process,
A generator characterized by performing. The generator according to claim 1.
In the dimensional compression process, the processor is a generation device that dimensionally compresses non-time-lapse data to be analyzed. The generator according to claim 2.
It is possible to access the determination information for determining whether the data corresponds to the time-lapse data or the non-time-lapse data.
The processor
Based on the determination information, a determination process for determining whether the analysis target data corresponds to the time-lapse data or the non-time-lapse data is executed.
In the generation process, the processor uses the analysis target data determined to be the temporal data by the determination process as the temporal data of the analysis target, and based on the temporal feature information, from the temporal data of the analysis target, Generate the plurality of temporal feature data,
In the dimensional compression process, the processor dimensionally compresses the analysis target data determined to be the non-time data by the determination process as the analysis target non-time data. A generator characterized by. The generator according to claim 1.
The processor
A generator characterized by executing a joining process for joining a plurality of groups after the dimension compression by the dimension compression process. The generator according to claim 3.
A generator characterized by executing a combining process of combining a non-time-lapse data to be analyzed after dimension compression by the dimension compression process and a plurality of groups after dimension compression. The generator according to claim 4.
The processor
A generator characterized in that an analysis process for outputting a corresponding objective variable is executed using the combination result of the combination process as an explanatory variable. The generator according to claim 5.
The processor
A generator characterized in that an analysis process for outputting a corresponding objective variable is executed using the combination result of the combination process as an explanatory variable. The generator according to claim 7.
A generator characterized in that the dimension compression process, the combination process, and the analysis process are executed by a multimodal neural network. A generation method executed by a generator having a processor that executes a program and a storage device that stores the program.
The generator can access time-dependent feature information indicating the time-dependent characteristics of the time-lapse data and grouping information in which a plurality of groups to which the time-lapse data should belong are defined.
The processor
A generation process for generating a plurality of temporal feature data showing the temporal features from the temporal data to be analyzed based on the temporal feature information.
A division process for dividing a plurality of time-dependent feature data generated by the generation process into the plurality of groups based on the grouping information,
A dimensional compression process that dimensionally compresses each of the plurality of groups divided by the division process,
A generation method characterized by executing. A generator that lets a processor perform data generation
The processor can access the temporal feature information indicating the temporal characteristics of the temporal data and the grouping information in which a plurality of groups to which the temporal data should belong are defined.
To the processor
A generation process for generating a plurality of temporal feature data showing the temporal features from the temporal data to be analyzed based on the temporal feature information.
A division process for dividing a plurality of time-dependent feature data generated by the generation process into the plurality of groups based on the grouping information,
A dimensional compression process that dimensionally compresses each of the plurality of groups divided by the division process,
A generation program characterized by executing.

Images