Data analysis device, data analysis system, and data analysis method

Abstract

To analyze a relationship between variables uniquely included in two datasets, respectively.SOLUTION: A data analysis apparatus includes a model integration unit 36 which constructs an integrated model which configured to predict a fourth data item serving as an objective variable, using a first data item serving as an explanatory variable, by inputting a result predicted by a first prediction model as an explanatory variable of a second prediction model, the first prediction model predicting a second data item serving as an objective variable using a first data item serving as an explanatory variable, the second prediction model using the second data item serving as an explanatory variable and the fourth data item serving as an objective variable. The model integration unit executes training for adjusting a parameter of the first prediction model included in the integrated model so that a loss function may be small.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a data analysis apparatus, a data analysis system, and a data analysis method.

Background Art

[0002] In recent years, the utilization of data for the planning and implementation of policies and measures has been active, and companies, public institutions, and organizations have accumulated various datasets on their own. If it becomes possible to combine these datasets distributed among organizations in some way, an extended dataset with a large amount of information can be obtained without newly collecting data. And if such an extended dataset can be obtained, the relationship between variables uniquely existing in each of the two distributed datasets can be analyzed.

[0003] Generally, a technique for combining such multiple datasets is called data linkage. In particular, a method of calculating the similarity between records based on information that commonly exists in each dataset when there is no collation key for identifying each record in multiple datasets and combining similar records is called statistical matching.

[0004] Statistical matching is described in, for example, Patent Document 1.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] When analyzing the relationship between variables in each of two datasets, in order to apply the technique disclosed in Patent Document 1 (hereinafter referred to as the prior art), it is necessary that each is obtained as a dataset.

[0007] For example, consider a case where we have a first dataset consisting of variables X and Y, and a second dataset consisting of variables Y and Z, and we want to analyze the relationship between variable X and variable Z. In this case, if we cannot obtain the second dataset itself, and instead are provided with a statistical model that predicts variable Z from variable Y as information about the second dataset, then the prior art cannot be applied. Furthermore, even if we possess both the first and second datasets, if there are many missing data points for variable Y, which is common to both datasets, the accuracy of the analysis of the relationship between variable X and variable Z using the prior art will decrease.

[0008] Therefore, the challenge is to accurately analyze the relationship between variable X and variable Z, which are unique to each of the two datasets, even when information about one of the datasets is available in the form of a statistical model, or when there are many missing data points in the variables common to both datasets. [Means for solving the problem]

[0009] A data analysis device, which is one of the representative embodiments of the present invention for solving the above problems, includes a storage device that stores: a first information database that stores first data of a first data item collected from a first population; a second information database that stores second data of a second data item and a third data item collected from the first population; and a prediction model database that stores parameters of a second prediction model that is trained using third data collected from a second population different from the first population, and that predicts a fourth data item with the second data item as an explanatory variable and the second data item as an objective variable; The system includes a model integration unit that constructs an integrated model that predicts a fourth data item, which uses the first data item as an explanatory variable and is the dependent variable, by inputting the prediction results of a first prediction model, which predicts a second data item, which uses the first data item as an explanatory variable and is the dependent variable, as explanatory variables of the second prediction model. The model integration unit performs training using the first data and the second data of the second data item to adjust the parameters of the first prediction model included in the integrated model to minimize a predetermined loss function. The predetermined loss function is defined as a function calculated based on the probability distribution of the prediction results of the first prediction model and the probability distribution of the prediction results of the integrated model. [Effects of the Invention]

[0010] According to the present invention, it is possible to analyze the relationships between variables that are unique to each of two datasets. Other issues, configurations, and effects will be clarified by the following description of embodiments. [Brief explanation of the drawing]

[0011] [Figure 1] This figure shows an example configuration of a data analysis system. [Figure 2] This diagram shows an example of the configuration of a user management information database. [Figure 3] This diagram shows an example of the configuration of an behavioral information database. [Figure 4] This figure shows an example of the structure of a survey information database. [Figure 5] This figure shows an example of the structure of a survey distribution information database. [Figure 6] This figure shows an example of the structure of an observation data database. [Figure 7A] This figure shows an example of the configuration of an observation data distribution information database. [Figure 7B] This figure shows an example of the configuration of an observation data distribution information database. [Figure 7C] This figure shows an example of the configuration of an observation data distribution information database. [Figure 8] This figure shows an example of the configuration of a prediction probability information database. [Figure 9] This diagram shows an example of the configuration of the correction amount database. [Figure 10] This is a sequence diagram illustrating an example of data analysis. [Figure 11] This is a flowchart illustrating an example of model integration processing. [Figure 12] This is an example of a screen displaying analysis results. [Figure 13A] This is an example of a screen displaying analysis results. [Figure 13B] This is an example of a screen displaying analysis results. [Figure 14] This is a functional block diagram of the analysis server. [Figure 15] This flowchart illustrates another example of the model integration process. [Modes for carrying out the invention]

[0012] Hereinafter, specific embodiments of the present invention will be described with reference to the drawings.

[0013] Figure 1 shows an example configuration of a data analysis system, which is an embodiment of the system. The data analysis system comprises a user terminal 101, an administrator terminal 151, and an analysis server 201. The user terminal 101, the administrator terminal 151, and the analysis server 201 are connected to each other via a network 10 so that they can communicate with one another. First, an overview of an example of processing in the data analysis system will be described.

[0014] The user terminal 101 collects sensor data and questionnaire data from the user and transmits this sensor data and questionnaire data to the analysis server 201. The questionnaire data asks about the user's daily activities. The analysis server 201 identifies behavioral information indicating the user's daily activities from the sensor data acquired by the user terminal 101. The analysis server 201 performs analysis on the behavioral information and questionnaire data and transmits the analysis results to other devices. In this analysis, the analysis server 201 combines similar questionnaire data from a different population with a risk prediction model generated based on, for example, the results of a follow-up survey on whether or not the user received long-term care certification. This allows the analysis server 201 to predict the user's risk in the near future based on objective sensor data that reflects the user's behavior. The analysis results are transmitted from the analysis server 201 to, for example, the user terminal 101 or the administrator terminal 151 and output. Output includes displaying the results on a screen. In other words, the analysis server 201 provides a behavioral information analysis service to the user terminal 101 or the administrator terminal 151. Furthermore, the connection method for the network 10 used to realize the behavioral information analysis service can be either wired or wireless, and the network 10 can be implemented via the internet, etc.

[0015] Next, the configuration of each device in the data analysis system will be described. The user terminal 101 can be a terminal (such as a smartphone or wearable device) that performs some kind of measurement related to the user's behavior and is used by the user. For this reason, it is desirable that the user terminal 101 be a portable device. The data measured by the user terminal includes sensor data that reflects the user's behavior, and questionnaire data which is the user's response to some kind of question.

[0016] The user terminal 101 can be implemented as a computer that primarily comprises an input / output device 102, an arithmetic unit 103, a sensor 104, a communication interface 105, a main memory device 106, and an auxiliary memory device 109.

[0017] The input / output device 102 is, for example, a display device equipped with a touch panel, which accepts user input and outputs the execution results of a program corresponding to the input in a format that the user can see. For example, it displays the risk prediction analysis results received from the analysis server 201 in response to user input. The input / output device 102 may be configured separately as an input device such as a keyboard and an output device such as a display. The arithmetic unit 103 is implemented as a processor such as a CPU and loads programs (applications) stored in the auxiliary storage device 109 into the main memory device 106 and executes them. The sensor 104 is composed of various sensors that measure the user's actions. Specifically, it can be composed of a location information sensor that identifies the user's current location and an acceleration sensor that measures the user's activity status. The communication interface 105 is a communication interface device that controls communication with the analysis server 201 via the network 10 according to a predetermined protocol and performs data transmission and reception with the analysis server 201.

[0018] The main memory 106 temporarily stores the program executed by the arithmetic unit 103 and the data used during program execution. In this embodiment, the sensor data collection program 107 and the questionnaire data collection program 108 are loaded. By executing the sensor data collection program 107, the arithmetic unit 103 functions as a sensor data collection unit, and by executing the questionnaire data collection program 108, the arithmetic unit 103 functions as a questionnaire data collection unit. The sensor data collection unit controls the sensor 104 to collect sensor data. The questionnaire data collection unit displays the questionnaire text and answer choices on the input / output device 102 and collects the questionnaire data.

[0019] The auxiliary storage device 109 can be implemented by a high-capacity, non-volatile recording device such as a magnetic storage device (HDD), a semiconductor storage device (SSD) including flash memory, etc., and stores programs executed by the arithmetic unit 103 and various data. In this embodiment, the auxiliary storage device 109 stores the sensor data collection program 107 and the questionnaire data collection program 108 that are loaded into the main memory 106, as well as the sensor information database 110 and the questionnaire information database 111.

[0020] The sensor information database 110 is a database that stores sensor data collected according to the sensor data collection program 107. Specifically, it stores time-series data of latitude and longitude collected by the location information sensor and time-series data of 3D acceleration collected by the acceleration sensor. The questionnaire information database 111 stores questionnaire data collected according to the questionnaire data collection program 108. In this embodiment, it stores behavioral questionnaire data, which is questionnaire data about the user's usual behavior, such as the average number of steps and the average number of days spent outside, and non-behavioral questionnaire data, which is questionnaire data about things other than usual behavior, such as the user's attributes and status (date of birth, gender, health status, etc.).

[0021] The administrator terminal 151 can be implemented as a terminal (PC or tablet) that displays the analysis results from the analysis server 201, and is used by users of the behavioral information analysis service. Users of the behavioral information analysis service are expected to include not only the user themselves, but also the local government where the user resides and the insurance company to which the user subscribes. The administrator terminal 151's main purpose is to output the analysis results from the analysis server 201, and it may be a portable terminal or a permanently installed terminal. The administrator terminal 151 is a computer that, like the user terminal 101, has as its main components an input / output device 152, an arithmetic unit 153, a communication interface 154, a main memory 155, and an auxiliary memory 156. The hardware configuration of the administrator terminal 151 is the same as that of the user terminal 101, so redundant explanations will be omitted.

[0022] The analysis server 201 can be implemented using a computer having an input / output device 202, an arithmetic unit 203, a communication interface 204, a main memory 210, and an auxiliary memory 220. The analysis server 201 is sometimes referred to as a data analysis device. Since the hardware configuration of the analysis server 201 is the same as that of the user terminal 101 and the administrator terminal 151, redundant explanations will be omitted, and the explanation will focus on the software configuration. Note that the analysis server 201 may be accessed from a terminal device in a separate enclosure by omitting the input / output device 202. Furthermore, the functions of the analysis server described below may be distributed across multiple computers, or implemented as a cloud application.

[0023] The main memory 210 is loaded with the following programs: the behavioral information identification program 211, the risk prediction model generation program 212, the correction amount calculation program 213, the model integration program 214, and the risk prediction analysis program 215. Each program is stored in the auxiliary memory 220 and is loaded into the main memory 210 when each program is executed by the arithmetic unit 203.

[0024] Figure 14 shows the analysis server 201 of this embodiment represented as a functional block. The analysis server 201 has an input unit 31, an output unit 32, an action information identification unit 33, a risk prediction model generation unit 34, a correction amount calculation unit 35, a model integration unit 36, and a risk prediction analysis unit 37. The input unit 31 and the output unit 32 are functional units that perform data input and output to the analysis server 201 via a communication I / F 204 or an input / output device 202. The action information identification unit 33, the risk prediction model generation unit 34, the correction amount calculation unit 35, the model integration unit 36, and the risk prediction analysis unit 37 are functional units realized by the arithmetic unit 203 executing an action information identification program 211, a risk prediction model generation program 212, a correction amount calculation program 213, a model integration program 214, and a risk prediction analysis program 215, respectively. Although this example shows the functional units being realized by executing software, it is also possible to realize them with hardware.

[0025] The behavior information identification unit 33 identifies behavior information from sensor data from the user terminal 101. Here, behavior information refers to information that indicates the user's behavior in daily life. The method by which the behavior information identification unit 33 identifies behavior information is arbitrary; it may extract behavior information from sensor data, or it may calculate or generate behavior information based on sensor data.

[0026] The risk prediction model generation unit 34 generates a long-term care risk prediction model in this embodiment, based on observational data stored in the observational data database 225, which is stored in the auxiliary storage device 220, with questionnaire data as the explanatory variable and long-term care certification as the dependent variable.

[0027] The correction amount calculation unit 35 calculates the probability distribution of the questionnaire data received from the user terminal 101 and stores the ratio of the probability density in the first population from which the questionnaire data was collected to the probability density in the second population from which the observation data was collected as a correction amount. As will be described later, the correction amount is used to correct the data of the data analysis system user (user) with the aim of aligning the data distribution of the data analysis system user with the distribution of the observation data.

[0028] The model integration unit 36 ​​generates a questionnaire prediction model using the behavioral information identified by the behavioral information identification unit 33 as explanatory variables and the questionnaire data as the dependent variable. Furthermore, it generates an integrated model that uses the output of this questionnaire prediction model as input to the risk prediction model generated by the risk prediction model generation unit 34. In other words, the integrated model becomes a risk prediction model for calculating the probability of becoming certified as requiring long-term care in the future based on the behavioral information identified by the behavioral information identification unit 33.

[0029] The risk prediction and analysis unit 37 performs risk prediction and analysis based on behavioral information, based on analysis requests from the user terminal 101 or the administrator terminal 151.

[0030] Furthermore, the auxiliary storage device 220 stores the following: user management information database 221, behavioral information database 222, questionnaire information database 223, questionnaire distribution information database 224, observation data database 225, observation data distribution information database 226, prediction probability information database 227, correction amount database 228, risk prediction model database 229, and integrated model database 230.

[0031] Figures 2 to 9 illustrate examples of data structures for each database. While these examples show a table-based structure, other data structures such as lists or queues are also acceptable. Furthermore, the contents of each database are merely examples; some information may be removed or new information added.

[0032] Figure 2 shows an example of the configuration of the user management information database 221. The user management information database 221 stores management data of users whose data is collected by the user terminal 101 in the data analysis system, and includes user ID 301, date of birth 302, gender 303, address 304, and email address 305. Each user is uniquely assigned an identification number as their user ID 301.

[0033] Figure 3 shows an example of the configuration of the behavioral information database 222. The behavioral information database 222 records user behavioral information identified by the behavioral information identification unit 33 from sensor data and includes basic information 401 and behavioral information 402. Basic information 401 includes user ID 403. Behavioral information 402 includes average number of days out 404. Here, only one data point (average number of days out per week) is shown as an example, but it includes all behavioral information identified from sensor data for use by the data analysis system for analysis. For example, it may include any behavioral information that can be identified from sensor data, such as the number of steps, frequency of going out, distance traveled, travel time, walking time, continuous walking time, decrease in going out, number of places stayed, status of stay at specific spots (number of times and time, etc.), car usage, motorcycle usage, train usage, bus usage, number of trips, high-intensity activity time, participation in group activities (number of times and time, etc.), employment status (number of times and time, etc.), number of falls, etc.

[0034] Figure 4 shows an example of the configuration of the questionnaire information database 223. The questionnaire information database 223 is a database that stores questionnaire data received from the user terminal 101 and includes basic information 501, behavioral questionnaire information 502, and non-behavioral questionnaire information 503. Basic information 501 includes the user ID 504 and the response date 505, behavioral questionnaire information 502 includes the average number of days spent out 506, and non-behavioral questionnaire information includes the date of birth 507 and gender 508. The analysis server 201 in this embodiment constructs a questionnaire prediction model that predicts questionnaire data from sensor data. For this reason, only one data point (average number of days spent out) is shown as an example of behavioral questionnaire information 502 here, but it is not limited to this, and it is sufficient if behavioral questionnaire data corresponding to the behavioral information recorded in the behavioral information 402 shown in Figure 3 is registered. Also, the non-behavioral questionnaire information 503 is not limited to the example data. For example, data items such as marital status, equivalent income, medication information, medical history, and subjective health perception may be included as non-behavioral questionnaire data. The average number of days out, 506, is the user's response to the question "Average number of days out per week," with four answer options: "5 or more times / week," "4 times / week," "2-3 times / week," and "1 time / week or less."

[0035] Figure 5 shows an example of the configuration of the survey distribution information database 224. The survey distribution information database 224 is a database that stores distribution information of users (the first population), and here an example of a joint probability distribution of sex and age 601 is shown. The survey distribution information database 224 may include the distributions of any variables contained in the survey information database 223. An example of a multivariate distribution is shown as the distribution information, but a univariate distribution may also be used.

[0036] Figure 6 shows an example of the configuration of the observation data database 225. The observation data is data used by the risk prediction model generation unit 34 to construct a risk prediction model, and is collected from a population different from the population of users who collect data using the user terminal 101. The risk prediction model in this embodiment is a model that estimates the likelihood of receiving long-term care certification in the future from the subject's behavioral questionnaire information. The observation data database 225 records observation data including basic information 701, behavioral questionnaire information 702, non-behavioral questionnaire information 703, and outcome information 704. The basic information 701 includes the subject ID 705, which is an identification number uniquely assigned to each subject of the observation data survey. The behavioral questionnaire information 702 includes the average number of days spent outside 706, the non-behavioral questionnaire information 703 includes age 707 and sex 708, and the outcome information 704 includes long-term care certification 709.

[0037] Here, only one data point (average number of days spent outside) is used as an example for behavioral questionnaire information 702, but it is not limited to this; any behavioral questionnaire data identical to that of behavioral questionnaire information 502 shown in Figure 4 is acceptable. Similarly, two data points (age, sex) are used as an example for non-behavioral questionnaire information 703, but it is not limited to these; any non-behavioral questionnaire data identical to that of non-behavioral questionnaire information 503 shown in Figure 4 is acceptable. Outcome information 704 records data corresponding to the risk information predicted by the model. Here, the care needs assessment 709 refers to whether or not a care needs assessment was conducted within a three-year follow-up period, with "not certified" or "certified" as elements. Alternatively, care needs assessment 709 can be changed to care needs assessment categories, with the elements being "not certified," "care level 1," "care level 2," "care level 3," "care level 4," and "care level 5." In this case, "support level 1" and "support level 2" may also be added as elements.

[0038] Figures 7A-7C show examples of the configuration of the observation data distribution information database 226. The observation data distribution information database 226 is a database that stores distribution information of observation data stored in the observation data database 225, and includes the sex and age joint probability distribution 801 (Figure 7A), the average number of days spent outside distribution 802 (Figure 7B), and the long-term care certification distribution 803 (Figure 7C).

[0039] The sex-age joint probability distribution 801 shown in Figure 7A is the joint probability distribution of age 707 and sex 708 of subjects (second population) stored in the observational data database 225. The average number of days spent out distribution 802 shown in Figure 7B is the distribution of the average number of days spent out 706 of subjects stored in the observational data database 225. The long-term care certification distribution 803 shown in Figure 7C is the distribution of long-term care certification 709 of subjects stored in the observational data database 225.

[0040] Figure 8 shows an example of the configuration of the prediction probability information database 227. The prediction probability information database 227 is a database that stores output information when observation data stored in the observation data database 225 is input into the risk prediction model stored in the risk prediction model database 229, and includes subject ID 901 and prediction probability 902. The prediction probability 902 is the output information in this embodiment and represents the probability of being certified as requiring long-term care within 3 years.

[0041] Figure 9 shows an example of the configuration of the correction amount database 228. The correction amount database 228 is a database that stores the correction amounts calculated by the correction amount calculation unit 35, and includes user ID 1001 and correction amount 1002. Correction amount 1002 is used for analysis in the model integration unit 36.

[0042] In the above explanation, User ID 301, User ID 403, User ID 504, and User ID 1001 are the same user assigned the same identification number. Also, Subject ID 705 and Subject ID 901 are the same subject assigned the same identification number.

[0043] Figure 10 shows a sequence diagram illustrating an example of data analysis in a data analysis system. In the following explanation, the functional units shown in Figure 14 will be used as the processing units in the analysis server 201.

[0044] (Step S1) The user terminal 101 receives user registration application input from the user via the input / output device 102. Subsequently, the user terminal 101 transmits user registration application data corresponding to the received user registration application to the analysis server 201 via the communication interface 105.

[0045] (Step S2) The input unit 31 receives user registration application data from the user terminal 101. The behavioral information identification unit 33, etc., registers or updates user management data in the user management information database 221 based on the received user registration application data. After that, the output unit 32 notifies the user terminal 101 that registration is complete. Specifically, it assigns a user ID 301 to the user registration application data, adds an entry to the user management information database 221, and registers the date of birth 302, gender 303, address 304, and email address 305.

[0046] (Step S3) The user terminal 101 receives a registration completion notification via the communication interface 105. Upon receiving the registration completion notification or starting the sensor data collection program 107, the computing unit 103 performs measurements using the sensor 104 according to the sensor data collection program 107 and stores the sensor data in the sensor information database 110. Similarly, upon receiving the registration completion notification or starting the questionnaire data collection program 108, the computing unit 103 presents the questionnaire to the user, collects responses from the user, and stores the collected questionnaire data in the questionnaire information database 111 according to the questionnaire data collection program 108.

[0047] (Step S4) The user terminal 101 transmits sensor data stored in the sensor information database 110 to the analysis server 201 at any time. Specifically, the computing unit 103 extracts sensor data according to the sensor data collection program 107 and transmits it periodically to the analysis server 201 via the communication interface 105. For example, the data may be transmitted at the same time every day. Similarly, the user terminal 101 transmits questionnaire data stored in the questionnaire information database 111 to the analysis server 201 at any time. Specifically, the computing unit 103 collects questionnaire responses from the user according to the questionnaire data collection program 108 and transmits them to the analysis server 201 via the communication interface 105. For example, the data may be transmitted when the user has completed answering the questionnaire.

[0048] (Step S5) When the input unit 31 receives the survey data, the behavioral information identification unit 33 updates the survey information database 223. Specifically, the behavioral information identification unit 33 adds an entry to the survey information database 223 and registers basic information 501, behavioral survey information 502, and non-behavioral survey information 503.

[0049] When the input unit 31 receives sensor data, the behavior information identification unit 33 identifies the user's behavior information from the received sensor data and updates the behavior information database 222 with the identified behavior information. In this example, the behavior information identification unit 33 estimates the average number of days spent outside from the acceleration information and location information included in the sensor data, adds an entry to the behavior information database 222, and registers basic information 401 and behavior information 402. Here, the average number of days spent outside 404 is registered as behavior information 402.

[0050] Furthermore, if the user terminal 101 can identify user behavior information based on sensor data, the user terminal 101 may send the user behavior information it has identified itself to the analysis server 201 instead of the sensor data.

[0051] (Step S6) The risk prediction model generation unit 34 generates a risk prediction model based on the observation data stored in the observation data database 225 and registers it in the risk prediction model database 229 (see Figure 1). Specifically, the risk prediction model database 229 stores information necessary for restoring the risk prediction model, such as the model type, hyperparameters, and weight parameter values. The specific method for generating the risk prediction model is described below. Here, we will generate a model using the long-term care certification 709 from the observation data (see Figure 6) as the dependent variable, and the average number of days spent outside 706 and age 707 as independent variables.

[0052] The risk prediction model generation unit 34 uses the observation data read from the observation data database 225 as training data to adjust the parameters of the risk prediction model. Here, since the average number of days out 706 is a categorical variable, these variables are converted into dummy variables and input into the model. Specifically, since the average number of days out 706 can take four values: "5 times / week or more", "4 times / week", "2-3 times / week", and "1 time / week or less", four dummy variables are created that take the value 1 if the value corresponds to each value, and 0 if it does not. A neural network can be used in the risk prediction model, and the number of hidden layers of the neural network and the number of units in each hidden layer can be set to any number.

[0053] Furthermore, other survey information stored in the observational data database 225 may be used as explanatory variables for the risk prediction model. In addition, although this risk prediction model predicts the risk of requiring long-term care within three years, a mortality risk prediction model or a disease incidence risk prediction model can be constructed by using, for example, mortality information or disease incidence information as the dependent variables.

[0054] (Step S7) The correction amount calculation unit 35 calculates the distribution of survey data based on the survey data stored in the survey information database 223 and updates the survey distribution information database 224. Specifically, the correction amount calculation unit 35 reads the survey data from the survey information database 223 and calculates the age of each user based on the date of birth 507. Next, the correction amount calculation unit 35 calculates the joint probability distribution of the calculated age and sex 508 and updates the sex-age joint probability distribution 601 in the survey distribution information database 224.

[0055] Next, the correction amount calculation unit 35 reads the sex-age joint probability distribution 601 from the questionnaire distribution information database 224 and the sex-age joint probability distribution 801 from the observation data distribution information database 226, and calculates the probability density ratio of the subject's sex-age joint probability distribution 801 to the user's sex-age joint probability distribution 601. That is, for the user's sex-age (attribute), if the probability density of the user's (first population) sex-age joint probability distribution 601 is H1 and the probability density of the subject's (second population) sex-age joint probability distribution 801 is H2, then the probability density ratio R can be expressed as H2 / H1.

[0056] Next, the correction amount calculation unit 35 calculates a correction amount for each user based on the calculated probability density ratio R and updates the correction amount database 228. In the example in Figure 9, the probability density ratio R corresponding to the user's age and sex combination is used as the correction amount for the user. By defining the correction amount in this way, a correction amount less than 1 is given when the probability of the user's attribute existing in the survey population is higher than the probability of the attribute existing in the observed data population, and a correction amount greater than 1 is given when the probability of the user's attribute existing in the survey population is lower than the probability of the attribute existing in the observed data population.

[0057] (Step S8) The model integration unit 36 ​​constructs a survey prediction model using the behavioral information identified by the behavioral information identification unit 33 as explanatory variables and the survey data as the dependent variable. Furthermore, it generates an integrated model using the output of this survey prediction model as input to a risk prediction model generated by the risk prediction model generation unit 34, which uses the survey data as explanatory variables and risk prediction as the dependent variable. The model integration unit 36 ​​then updates the integrated model database 230 based on the generated integrated model. Specifically, it stores information necessary for restoring the integrated model, such as the model type, hyperparameters, and weight parameter values, in the integrated model database 230. The details of this integrated model generation will be explained below using Figure 11.

[0058] (Step S20) The model integration unit 36 ​​reads the risk prediction model from the risk prediction model database 229. Specifically, the model integration unit 36 ​​reads the information necessary to restore the risk prediction model from the risk prediction model database 229 and restores the risk prediction model. Here, if the risk prediction model is g, the input / output relationship of the risk prediction model g is expressed by (Equation 1) shown below.

[0059]

number

[0060] Here, y is the input to the risk prediction model g, specifically representing the average number of days spent outside (706). Also, a is the input to the risk prediction model g, specifically representing age (707). Also, p g θg is the output of the risk prediction model g when y and a are input, and represents the probability of being certified as requiring long-term care within 3 years. θg represents the parameters of the risk prediction model g.

[0061] (Step S21) Next, the model integration unit 36 ​​sets the structure of the questionnaire prediction model that predicts questionnaire information from behavioral information. Specifically, the model integration unit 36 ​​sets the number of layers and the number of units in each layer of the neural network that predicts the average number of days out 506, based on the average number of days out 404. The number of units in the input and output layers of the model is automatically determined by the input and output variables, but the number of intermediate layers and the number of units in the intermediate layers of the model can be set to arbitrary values. If the questionnaire prediction model is denoted as f, the input-output relationship of the questionnaire prediction model f is expressed by (Equation 2) shown below.

[0062]

number

[0063] Note that x is the input to the survey prediction model f, and specifically represents the average number of days spent out, which is 404. Also, p f This is the output when x is input into the survey prediction model f, and represents the classification probability for each class of an average number of days out of the house of 506. In other words, it represents the classification probability for each class: "5 times / week or more", "4 times / week", "2-3 times / week", and "1 time / week or less". Also, θ f This represents the parameters of the survey prediction model f.

[0064] Here, the average number of days spent out, 404, which is an explanatory variable in the survey prediction model, is the number of days spent out identified by the behavior information identification unit 33 based on sensor information measured by the user terminal 101. In contrast, the average number of days spent out, 506, which is the dependent variable in the survey prediction model, is the number of days spent out based on the survey measured by the user terminal 101. In other words, the survey prediction model f in this embodiment can be described as a model that converts the number of days spent out calculated from sensor information into the number of days spent out based on the survey.

[0065] In this step, the survey prediction model is in the untrained stage, and the weight parameter θ f The initial values ​​are set to arbitrary values. The weight parameters θ of the survey prediction model f. fThe update method will be described later in step S27.

[0066] Note that, as explanatory variables of the questionnaire prediction model, a plurality of variables included in the behavior information 402 may be added. Further, as explanatory variables of the questionnaire prediction model, variables of the non-behavior questionnaire information 503 may be added. Further, in this step, the structures of a plurality of questionnaire prediction models for predicting different items of questionnaire data may be set.

[0067] (Step S22) Subsequently, the model integration unit 36 integrates the risk prediction model read in step S20 and the questionnaire prediction model whose structure was determined in step S21, and generates one integrated model. Specifically, as expressed by the following (Equation 3), the model integration unit 36 substitutes the output p f of the questionnaire prediction model f into the input y of the risk prediction model g, thereby integrating the risk prediction model g and the questionnaire prediction model f, and constructing one integrated model.

[0068]

Equation

[0069] Note that p fg is the output when p f and a are input to the risk prediction model g, and represents the probability of being certified as needing care within 3 years.

[0070] In this way, the integrated model replaces a part of the input of the risk prediction model with the output of the questionnaire prediction model, and enables risk prediction of needing care based on the behavior information specified by the behavior information specifying unit 33. On the other hand, in order for the integrated model to exhibit high prediction performance, two parameters θ g and θ f among them, the weight parameter θ f of the questionnaire prediction model that has not yet been learnedTo optimize this, it is necessary to train (learn) the integrated model. In the following steps S23 to S28, the two parameters θ of the integrated model g and θ f Of these, the parameter θ g With the parameter θ fixed, f This section describes how to optimize p. f The intermediate output of the integrated model, p fg We will refer to this as the final output of the integrated model.

[0071] Here, if we let N1 be the number of input data to the integrated model (Equation 3), that is, the number of rows of input a and x, then the intermediate output p f is row N1 C f It is a matrix of columns, and the final output is p fg is row N1 C fg It is a matrix of columns. Here, C f The intermediate output p of the integrated model is f This is the number of possible events for the variable being predicted. Specifically, it is the intermediate output p of the integrated model. f The average number of days spent out, 506, which is the target of this prediction, can take one of four forms: "5 times / week or more", "4 times / week", "2-3 times / week", or "1 time / week or less". fg The final output p of the integrated model is fg This is the number of possible events for the variable being predicted. Specifically, it is the final output p of the integrated model. fg The possible outcomes for the long-term care certification level 709, which is the target of this prediction, are "not certified" and "certified."

[0072] (Step S23) Next, the model integration unit 36 ​​calculates the loss function loss1 based on the intermediate output of the integrated model, the correction amounts stored in the correction amount database 228, and the distribution information stored in the observation data distribution information database 226.

[0073] First, the model integration unit 36 ​​reads the correction amount from the correction amount database 228. Next, the model integration unit 36 ​​reads the distribution P of the intermediate output of the integrated model. fThis is calculated based on (Equation 4) shown below.

[0074]

number

[0075] Here, N1 is the number of rows in the intermediate output, i.e., the number of input data to the integrated model. Also, w is a correction amount for each data, and is a vector of length N1. Here, as explained in step S7, the correction amount w is determined based on the probability distribution of the user's attributes in the population, so attributes that have a higher probability of being present in the survey population than in the observed data population are given a smaller weight, and conversely, attributes that have a lower probability of being present in the survey population than in the observed data population are given a larger weight. This absorbs the differences in attributes between the survey population and the observed data population from which the model was based.

[0076] Next, the model integration unit 36 ​​retrieves the intermediate output p of the integrated model from the observation data distribution information database 226. f The distribution of the variable to be predicted is read out. In other words, the intermediate output p of the integrated model. f Since this represents the classification probability for each class of 506 average days out, we read out the corresponding average days out distribution 802. In the following explanation, we will refer to the read average days out distribution 802 as Q f Far away.

[0077] Finally, the model integration unit 36 ​​calculates the distribution P of the intermediate output of the obtained integrated model. f And the distribution Q of the variable to be predicted for the intermediate output f The loss function loss1 is calculated by substituting the following into (Equation 5).

[0078]

number

[0079] PSI(P,Q) is known as the Population Stability Index, which represents the dissimilarity between the probability distributions P and Q of discrete random variables, and is defined as shown below (Equation 6).

[0080]

number

[0081] Here, C is the number of possible events for the discrete random variables of probability distributions P and Q (in this case, 4). Note that in the calculation of loss1, Kullback-Leibler divergence, Jensen-Shannon divergence, or Maximum Mean Discrepancy may be used instead of the Population Stability Index.

[0082] Thus, the loss function loss1 is equal to the intermediate output p of the integrated model. f Distribution P f And the distribution of the observed data Q f It is calculated as the dissimilarity of the integrated model p. f Distribution P f In the calculation, a correction using a correction amount w is introduced. In this way, by adjusting the gender and age distribution of the analysis system users to match the gender and age distribution of the observed data, and then calculating the dissimilarity between the two distributions, the validity of the dissimilarity calculation results can be improved.

[0083] (Step S24) Next, the model integration unit 36 ​​calculates the loss function loss2 based on the intermediate output of the integrated model and the survey data stored in the survey information database 223.

[0084] First, the model integration unit 36 ​​retrieves the intermediate output p of the integrated model from the questionnaire information database 223. f This retrieves information about the variables to be predicted. In other words, it reads the intermediate output p of the integrated model. fSince this represents the classification probability for each class of the average number of days out of the house (506), we read out the corresponding average number of days out of the house (506). In the following explanation, we will refer to the read-out average number of days out of the house (506) as q f Let's assume that.

[0085] Subsequently, the model integration unit 36 ​​outputs the intermediate output p of the integrated model. f And the intermediate output p of the integrated model f Information about the variable to be predicted q f The loss function loss2 is calculated by substituting this into (Equation 7) shown below.

[0086]

number

[0087] Here, N1 is the number of intermediate output rows, i.e., the number of input data to the integrated model. Note that of the N1 data, q f Data with missing values ​​are not used in the calculation of the loss function loss2. CE(p,q) is an index known as cross-entropy, which calculates the classification error of the prediction probability p for the correct label q, and is defined as shown below (Equation 8).

[0088]

number

[0089] Here, C is the number of possible events for the correct label q.

[0090] The loss function loss2 is a commonly used loss function in training machine learning models to solve classification problems, and it is calculated as the average of the classification errors calculated for each sample. The loss function loss2 does not use information from observational data and is affected by missing data in survey data.

[0091] (Step S25) Next, the model integration unit 36 ​​calculates the loss function loss3 based on the final output of the integrated model, the correction amounts stored in the correction amount database 228, and the distribution information stored in the observation data distribution information database 226.

[0092] First, the model integration unit 36 ​​reads the correction amount from the correction amount database 228. Next, the model integration unit 36 ​​reads the distribution P of the final output of the integrated model. fg This is calculated based on (Equation 9) shown below.

[0093]

number

[0094] Here, N1 is the number of rows in the final output, i.e., the number of input data points to the integrated model. Also, w is the correction amount for each data point, and is a vector of length N1.

[0095] Next, the model integration unit 36 ​​retrieves the final output p of the integrated model from the observation data distribution information database 226. fg The distribution of the variable to be predicted is read out. In other words, the final output p of the integrated model. fg Since this represents the probability of being certified as requiring long-term care within three years, we read out the corresponding long-term care certification distribution 803. In the following explanation, we will refer to the read long-term care certification distribution 803 as Q fg Far away.

[0096] Finally, the model integration unit 36 ​​calculates the distribution P of the final output of the obtained integrated model. fg And the distribution Q of the variable that is the target of prediction for the final output. fg The loss function loss3 is calculated by substituting the following into (Equation 10).

[0097]

number

[0098] As with the loss function loss1, Kullback-Leibler divergence, Jensen-Shannon divergence, or Maximum Mean Discrepancy may be used instead of the Population Stability Index.

[0099] Thus, the loss function loss3 is calculated by multiplying the loss function loss1 calculated in step S23 by the final output p of the integrated model. fg This corresponds to the application to the above.

[0100] (Step S26) Next, the model integration unit 36 ​​calculates the loss function loss4 based on the final output of the integrated model, the predicted probability information stored in the predicted probability information database 227, and the correction amounts stored in the correction amount database 228.

[0101] First, the model integration unit 36 ​​reads the correction amount from the correction amount database 228. Next, the model integration unit 36 ​​outputs the final output p of the integrated model. fg Average entropy H fg This is calculated based on (Equation 11) shown below.

[0102]

number

[0103] Here, N1 is the number of rows in the final output, i.e., the number of input data to the integrated model. Also, w is the correction amount for each data point, and is a vector of length N1. Furthermore, H(p) is the entropy, which is an indicator of the uncertainty of the prediction, and is defined as shown below (Equation 12).

[0104]

number

[0105] Here, C is the number of possible outcomes for the random variable. p is a vector of length C, representing the probability of each outcome.

[0106] Next, the model integration unit 36 ​​reads the predicted probability 902 from the predicted probability information database 227. Then, the average entropy H is calculated based on the read predicted probability 902. ref This is calculated based on (Equation 13) shown below.

[0107]

number

[0108] Here, p ref N2 is the predicted probability of 902, and N2 is the number of data points (in this case, the number of subjects) for the predicted probability of 902.

[0109] Finally, the model integration unit 36 ​​uses these average entropy H fg and H ref The loss function loss4 is calculated by substituting the following into (Equation 14).

[0110]

number

[0111] Thus, the loss function loss4 is the average entropy of the predicted probability when the observational data stored in the observational data database 225 is input to the risk prediction model, and the final output p of the integrated model. fg It is calculated as the squared error of the mean entropy. Since entropy is an indicator of prediction uncertainty, the loss function loss4 can be interpreted as the difference between the uncertainty of the risk prediction based on observational data and the uncertainty of the risk prediction based on information measured by user terminal 101. Here, the final output p of the integrated model is fgIn calculating the average value of the entropy, a correction by a correction amount w is introduced. In this way, by adjusting the gender and age distribution of the analysis system users to match the gender and age distribution of the observed data, and then calculating the error between the two entropies, the validity of the calculation results can be improved.

[0112] (Step S27) Next, the model integration unit 36 ​​calculates the loss value using at least one combination of loss functions from loss1, loss2, loss3, and loss4 calculated in steps S23, S24, S25, and S26, and adjusts the parameters θ of the integrated model to minimize this loss. f Update the parameters of the integrated model θ, for example, to minimize the loss value calculated by the loss function defined in (Equation 15) shown below. f Update.

[0113]

number

[0114] In example (Equation 15), the loss function loss2 is weighted to match the order of the other loss functions. Model parameter optimization algorithms such as SGD and Adam can be used.

[0115] As mentioned above, the loss function `loss` does not necessarily include all of the loss functions `loss1` through `loss4`. Therefore, it is not necessary to calculate the loss functions `loss1` through `loss4` that are not used in the `loss` function.

[0116] When applying (Equation 15) as the loss function, the parameters θ of the integrated model should be minimized to minimize the four loss functions loss1, loss2, loss3, and loss4. f The following is updated. Here, the loss function loss2 is the intermediate output p of the integrated model. fThis is a general classification error calculated for each sample based on [the given formula]. In contrast, the loss functions loss1 and loss3 are calculated by converting the intermediate or final output of the integrated model into a distribution and determining the dissimilarity between that distribution and the ground truth distribution (the distribution of the observed data). Similarly, the loss function loss4 is calculated by converting the final output of the integrated model into entropy and determining the error between that entropy and the ground truth entropy (the entropy of the predicted probability when the observed data is input into the risk prediction model).

[0117] In other words, in calculating the loss functions loss1, loss3, and loss4, the ground truth data for each sample (average number of days out: 506) is not required; they can be calculated from the ground truth distribution or entropy information. Therefore, even if some of the ground truth data for each sample, specifically the survey data, is missing and the loss function loss2 cannot be calculated for all samples, the loss functions loss1, loss3, and loss4 can still be calculated. By minimizing these loss functions, it becomes possible to optimize the integrated model so that the distribution or entropy of the integrated model's output approaches the ground truth distribution and entropy, respectively, thereby improving the prediction accuracy of the integrated model.

[0118] (Step S28) Next, the model integration unit 36 ​​determines whether the termination condition for updating the parameters of the integrated model is met. For example, the termination condition for updating the parameters of the integrated model can be set to whether the number of parameter updates of the integrated model exceeds a pre-set threshold. If the termination condition is not met, the process returns to the calculation of the loss function (steps S23 to S26). On the other hand, if the termination condition is met, the process ends, and the model integration unit 36 ​​updates the integrated model database 230 based on the generated integrated model.

[0119] This concludes the explanation of Figure 11. Now, let's return to the explanation of Figure 10.

[0120] (Step S9) When the administrator terminal 151 receives an analysis request input from the administrator via the input / output device 152, it sends the analysis request information to the analysis server 201 via the communication interface 154. The analysis request information includes user IDs for the set of users that the administrator intends to analyze for risk prediction.

[0121] (Step S10) When the Risk Prediction Analysis Unit 37 receives an analysis request from the administrator terminal 151, it performs data analysis. The details of the data analysis are described below. First, the Risk Prediction Analysis Unit 37 reads user records corresponding to the user ID to be targeted for risk prediction analysis received from the administrator terminal 151 from the behavioral information stored in the behavioral information database 222. Specifically, the Risk Prediction Analysis Unit 37 reads user records corresponding to the user ID to be targeted for risk prediction analysis received from the administrator terminal 151 from the average number of days spent outside the home (404) stored in the behavioral information database 222. In addition, the Risk Prediction Analysis Unit 37 reads user records corresponding to the user ID to be targeted for risk prediction analysis received from the administrator terminal 151 from the questionnaire data stored in the questionnaire information database 223. Specifically, the Risk Prediction Analysis Unit 37 reads user records corresponding to the user ID to be targeted for risk prediction analysis received from the administrator terminal 151 from the date of birth (507) stored in the questionnaire information database 223.

[0122] Next, the risk prediction and analysis unit 37 reads the integrated model from the integrated model database 230. Subsequently, the risk prediction and analysis unit 37 calculates the user's age from the retrieved birth dates 507, substitutes the calculated age and average number of days spent outside 404 into the integrated model, and calculates the probability of being certified as requiring long-term care within three years. Furthermore, it calculates a histogram of the calculated long-term care certification probabilities for each user and the average long-term care certification probability for all users.

[0123] (Step S11) Once step S10 is completed, the output unit 32 sends the analysis results to the administrator terminal 151. Specifically, the output unit 32 sends to the administrator terminal 151 the histogram of the probability of requiring long-term care calculated in step S10, the average value of the probability of requiring long-term care, and other information to be displayed on the screen of the administrator terminal 151.

[0124] (Step S12) When the administrator terminal 151 receives the data analysis results, it displays the analysis results on the input / output device 152. Figure 12 shows an example of the display screen for the analysis results in this step, and shows the results of the risk prediction analysis. The administrator here is a local government.

[0125] The analysis results display screen shown in Figure 12 includes an analysis target selection area 1201, a prediction result display area 1203, and a prediction model display area 1205.

[0126] The analysis target selection area 1201 includes the analysis target file reference unit 1202. The analysis target file reference unit 1202 is used by the administrator to select a file containing user ID 401 of the user (user of user terminal 101) that they want to be included in the risk prediction analysis. Specifically, when the administrator selects the analysis target file reference unit 1202, a list of files stored on the administrator terminal 151 is displayed, and the administrator selects a file for the set of users they want to include in the risk prediction analysis. The file selected in the analysis target file reference unit 1202 is sent to the analysis server 201 along with the analysis request in step S9.

[0127] In this example, the prediction result display area 1203 displays the distribution of the probability of being certified as requiring long-term care within three years, and the average user probability of being certified as requiring long-term care within three years. This information is based on the information received from the analysis server 201 in step S11.

[0128] The prediction model display area 1205 includes the input display area 1206, the intermediate output display area 1207, and the final output display area 1208. This information is also based on the information received from the analysis server 201 in step S11. The input display area 1206 displays the input variables of the integrated model. Here, the variable names are displayed for the average number of days away from home (404) and age (calculated based on the date of birth (507)), which are inputs to the integrated model. The intermediate output display area 1207 displays the variable names of the variables that are predicted for the intermediate output of the integrated model. Here, the intermediate output p of the integrated model is displayed. f The variable names for the average number of days away from home (506), which is the target of the prediction, are displayed. The final output display area 1208 is the final output p of the integrated model. fg The variable names of the variables being predicted are displayed. Here, the final output p of the integrated model is shown. fg For the long-term care certification level 709, which is the target of the prediction, the variable name will be displayed.

[0129] By presenting the structure of the model used to calculate the probability of requiring long-term care, the basis for the prediction can be clarified, and the reliability of the administrator's analysis results can be improved.

[0130] The data analysis described above is an example performed at the request of an administrator, but it can also be performed at the request of individual users. The details are explained below.

[0131] (Step S13) When the user terminal 101 receives an analysis request input from the user via the input / output device 102, it sends the analysis request information to the analysis server 201 via the communication interface 105. The analysis request information includes the user's user ID.

[0132] (Step S14) When the risk prediction and analysis unit 37 receives analysis request information from the user terminal 101, it performs data analysis. The analysis in this step is equivalent to step S10 when analysis request information is received from the administrator terminal 151, so redundant explanations are omitted. The difference is that the analysis target in step S10 is a set of users, while in step S14 it is individual users.

[0133] (Step S15) Once step S14 is completed, the output unit 32 sends the analysis results to the user terminal 101. Specifically, the output unit 32 sends to the user terminal 101 the user's long-term care certification probability calculated in step S14 and other information to be displayed on the screen of the user terminal 101.

[0134] (Step S16) When the user terminal 101 receives the data analysis results, it displays the analysis results on the input / output device 102. Figures 13A and 13B are examples of the display screen for the analysis results in this step, showing the results of the risk prediction analysis.

[0135] The display screen shown in Figure 13A includes a prediction result display area 1301. The prediction result display area 1301 also includes a prediction probability display area 1302, an input variable display area 1303, and a supplementary explanation display area 1304. The prediction probability display area 1302 displays the probability of being certified as requiring long-term care within three years. The input variable display area 1303 displays the values ​​of the input variables of the integrated model used to calculate the probability of requiring long-term care. This information is based on the information received from the analysis server 201 in step S15. The supplementary explanation display area 1304 is used to display the calculation method for the probability of requiring long-term care, and when the user selects this, the display screen switches to Figure 13B.

[0136] The display screen shown in Figure 13B includes an input display area 1305, an intermediate output display area 1306, and a final output display area 1307. The displayed content is equivalent to that displayed in the prediction model display area 1205 shown in Figure 12, and redundant explanations are omitted. This information is also based on the information received from the analysis server 201 in step S15.

[0137] (modified version) Below, an example of a model integration process different from the integrated model generation process (step S8) in the embodiment is described as a modification. In this modification, the model integration process is performed without using the distribution information stored in the observation data distribution information database 226 and the predicted probability information stored in the predicted probability information database 227. Figure 15 shows a flowchart of the integrated model generation process in the modification. Processes similar to the integrated model generation process in the embodiment are indicated with the same reference numerals, and redundant explanations are omitted. Here, the calculation of the loss function (steps S31 to S34), which is a characteristic process of the modification, will be explained.

[0138] (Step S33) The model integration unit 36 ​​calculates the loss function loss 5 based on the intermediate output of the integrated model and the survey data stored in the survey information database 223.

[0139] First, the model integration unit 36 ​​retrieves the intermediate output p of the integrated model from the questionnaire information database 223. f This retrieves information about the variables to be predicted. In other words, it reads the intermediate output p of the integrated model. f Since this represents the classification probability for each class of the average number of days out of the house (506), we read out the corresponding average number of days out of the house (506). In the following explanation, we will refer to the read-out average number of days out of the house (506) as q f Let's assume that.

[0140] Next, the model integration unit 36 ​​outputs the intermediate output p of the integrated model. f And the intermediate output p of the integrated model f Information about the variable to be predicted q fThe loss function loss5 is calculated by substituting this into (Equation 16) shown below.

[0141]

number

[0142] Here, MMD(p,q) is the Maximum mean discrepancy, which quantifies the dissimilarity between the distributions of variable p and variable q without estimating the distributions of each variable p and variable q. Since Maximum mean discrepancy is a comparison of the distributions of the population, rather than a comparison of predicted values ​​and ground truth data within the same sample, as calculated in step S24 or step S32 described later, the number of data points for variables p and q do not need to be the same. Therefore, q f Regardless of the missing status, all intermediate outputs p f The loss function loss5 can be calculated using this method.

[0143] (Step S32) Next, the model integration unit 36 ​​calculates the loss function loss6 based on the intermediate output of the integrated model and the survey data stored in the survey information database 223. The loss function loss6 calculated in this step is the same as the loss function loss2 (equation 7) calculated in step S24 of the flowchart in Figure 11. Therefore, an explanation is omitted here.

[0144] (Step S33) Next, the model integration unit 36 ​​calculates the loss function loss7 based on the final output of the integrated model and the survey data stored in the survey information database 223.

[0145] First, the model integration unit 36 ​​reads information from the survey information database 223 that corresponds to the input information for the risk prediction model g in the survey population. Specifically, it reads the average number of days spent away from home 506 and the date of birth 507, which correspond to the input for the risk prediction model g.

[0146] Next, the model integration unit 36 ​​calculates the age from the read date of birth 507, then inputs the calculated age and average number of days away from home 506 into the risk prediction model g, and outputs p g The following is calculated. Note that for data where either the average number of days out (506) or the date of birth (507) is missing, p g It is not used in the calculation.

[0147] Subsequently, the model integration unit 36 ​​outputs the final output p of the integrated model. fg The output p of the calculated risk prediction model g is... g The loss function loss7 is calculated by substituting this into (Equation 17) shown below.

[0148]

number

[0149] Similar to the calculation of the loss function loss5 in step S31, the loss function loss7 is calculated as output p g Regardless of the missing values, all final output p fg It is possible to calculate using this method.

[0150] (Step S34) Next, the model integration unit 36 ​​calculates the loss value using at least one combination of loss functions from loss5, loss6, and loss7 calculated in steps S31, S32, and S33, and adjusts the parameters θ of the integrated model to minimize this loss. f Update the parameters of the integrated model θ, for example, to minimize the loss value calculated by the loss function defined in (Equation 18) shown below. f Update.

[0151]

number

[0152] In example (Equation 18), the loss function loss6 is weighted to match the order of the other loss functions. Model parameter optimization algorithms such as SGD and Adam can be used.

[0153] As mentioned above, the loss function does not necessarily include all of the loss functions loss5 through loss7. Therefore, it is not necessary to calculate loss functions loss5 through loss7 that are not used in the loss function.

[0154] When applying (Equation 18) as the loss function, the parameters θ of the integrated model should be minimized to minimize the three loss functions loss5, loss6, and loss7. f The following is updated. Here, the loss function loss6 is the intermediate output p of the integrated model. f This is a general classification error calculated for each sample based on the following. In contrast, the loss functions loss5 and loss7 are calculated by converting the intermediate or final output of the integrated model into a distribution and determining the dissimilarity between that distribution and the ground truth distribution. However, in steps S23 and S25 of the example, the distribution information of the observation data distribution information database 226 was used as the ground truth distribution, whereas in this modified example, the distribution calculated based on the data of the questionnaire information database 223 is used as the ground truth distribution. In other words, by using the loss functions loss5 and loss7, even when distribution information regarding the observation data is not available, the parameters θ of the integrated model can be used. f Optimization becomes possible.

[0155] It should be noted that the present invention is not limited to the embodiments and modifications described above, but includes various other modifications. For example, the embodiments and modifications described above are explained in detail to make the present invention easier to understand, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace parts of the configuration of one embodiment or modification with the configuration of another embodiment or modification, and it is also possible to add the configuration of another embodiment or modification to the configuration of one embodiment or modification. In addition, it is possible to add, delete, or replace parts of the configuration of an embodiment or modification with other configurations.

[0156] For example, in the above embodiment, we showed an example where the same analysis server creates the risk prediction model, but it is also possible to obtain a risk prediction model created by someone other than the service provider of the data analysis system. In this case, the risk prediction model generation program 212 (risk prediction model generation unit in Figure 14) becomes unnecessary from the example configuration of the data analysis system in Figure 1. Also, in this case, the data corresponding to the observation data stored in the observation data database 225 is usually unavailable, but even in that case, it is desirable to obtain the distribution information stored in the observation data distribution information database 226 and the prediction probability information stored in the prediction probability information database 227. By obtaining this information together with the risk prediction model, it becomes possible to calculate the loss function described in the embodiment and generate an integrated model. Even if distribution information for the observation data is not available, an integrated model can be constructed by calculating the loss function described in the modified example. [Explanation of symbols]

[0157] 10...Network, 31...Input Unit, 32...Output Unit, 33...Behavioral Information Identification Unit, 34...Risk Prediction Model Generation Unit, 35...Correction Amount Calculation Unit, 36...Model Integration Unit, 37...Risk Prediction Analysis Unit, 101...User Terminal, 102...Input / Output Device, 103...Calculation Unit, 104...Sensor, 105...Communication I / F, 106...Main Memory, 107...Sensor Data Collection Program, 108...Questionnaire Data Collection Program, 109...Auxiliary Storage Device, 110...Sensor Information Database, 111...Questionnaire Information Database, 151...Administrator Terminal, 152...Input / Output Device, 153...Calculation Unit, 154...Communication I / F, 155...Main Memory, 156...Auxiliary Storage Device, 201...Analysis Server, 202...Input / Output Device, 203... Calculation unit, 204... Communication I / F, 210... Main memory, 211... Behavioral information identification program, 212... Risk prediction model generation program, 213... Correction amount calculation program, 214... Model integration program, 215... Risk prediction analysis program, 220... Auxiliary memory, 221... User management information database, 222... Behavioral information database, 223... Questionnaire information database, 224... Questionnaire distribution information database, 225... Observation data database, 226... Observation data distribution information database, 227... Prediction probability information database, 228... Correction amount database, 229... Risk prediction model database, 230... Integrated model database.

Claims

1. A storage device that stores: a first information database that stores first data of a first data item collected from a first population; a second information database that stores second data of a second and a third data item collected from the first population; and a prediction model database that stores parameters of a second prediction model that is trained using third data collected from a second population different from the first population, and that predicts a fourth data item with the second data item as an explanatory variable and the second data item as an dependent variable. The system includes a model integration unit that constructs an integrated model that predicts a fourth data item, which uses the first data item as an explanatory variable and is the objective variable, by inputting the prediction results of a first prediction model that predicts a second data item, which uses the first data item as an explanatory variable and is the objective variable, as explanatory variables of the second prediction model. The model integration unit performs training to adjust the parameters of the first prediction model included in the integrated model so as to reduce a predetermined loss function using the first data and the second data of the second data item, and the predetermined loss function is defined as a function calculated based on the probability distribution of the prediction results of the first prediction model and the probability distribution of the prediction results of the integrated model.

2. In claim 1, The storage device stores a distribution information database that stores the probability distribution of the second data item, the probability distribution of the third data item, and the probability distribution of the fourth data item in the third data. A data analysis device in which the predetermined loss function is defined as either a first loss function that shows the dissimilarity between the probability distribution of the prediction result of the first prediction model and the probability distribution of the second data item in the third data, or a second loss function that shows the dissimilarity between the probability distribution of the prediction result of the integrated model and the probability distribution of the fourth data item in the third data, or a combination thereof.

3. In claim 1, A data analysis device in which the predetermined loss function is defined as either a fifth loss function that shows the dissimilarity between the probability distribution of the prediction results of the first prediction model and the probability distribution of the second data item in the second data, or a sixth loss function that shows the dissimilarity between the probability distribution of the prediction results of the integrated model and the probability distribution of the prediction results of the second prediction model inputting the second data, or a combination thereof.

4. In claim 2, The model integration unit has a correction amount calculation unit that calculates a correction amount to correct the first loss function and the second loss function used for training, The correction amount calculation unit is a data analysis device that calculates a correction amount for each record in the second information database so that the probability distribution of the third data item in the first population and the probability distribution of the third data item in the second population are consistent.

5. In claim 4, The third data item is a data item relating to the attributes of the population, The correction amount calculation unit is a data analysis device that calculates a probability density ratio for each record in the second information database, which is the ratio of the probability density in the second population to the probability density in the first population for the attributes of the record, as the correction amount for the record.

6. In claim 2 or claim 3, A data analysis device in which the predetermined loss function is defined as a combination of the predetermined loss function and a third loss function that indicates the classification error of the prediction result of the first prediction model.

7. In claim 2, The storage device stores a prediction probability information database that stores the prediction probabilities predicted for the fourth data item by the second prediction model for the second population. A data analysis device in which the predetermined loss function is defined as a combination of the average entropy of the prediction results predicted for the first population for the fourth data item by the integrated model and a fourth loss function that shows the error between the average entropy of the prediction probability stored in the prediction probability information database.

8. In claim 1, A data analysis device having a predictive analysis unit that uses the integrated model trained by the model integration unit to make predictions about the fourth data item for one or more records included in the first population.

9. In claim 1, The system includes a prediction model generation unit that performs training to adjust the parameters of the second prediction model using the third data of the second and fourth data items, The storage device is a data analysis device that stores an observation data database which stores the third data of the second data item, the third data item, and the fourth data item collected from the second population.

10. In claim 1, The first data is data that shows behavioral information estimated from sensor data detected by sensors regarding the behavior of people belonging to the first population, The data for the second data item of the second data is questionnaire data in which a person belonging to the first population responded about their behavior. A data analysis device in which the data of the second data item of the third data is questionnaire data in which a person belonging to the second population answers questions about their behavior.

11. In claim 10, The sensor is a data analysis device that includes at least one of a position information sensor and an acceleration sensor.

12. A data analysis system comprising a data analysis device and a user terminal that is communicably connected to the data analysis device via a network, The aforementioned data analysis device is A storage device that stores: a first information database that stores first data of a first data item collected from a first population; a second information database that stores second data of a second data item and a third data item collected from the first population; and a prediction model database that stores parameters of a second prediction model that is trained using data collected from a second population different from the first population, and that predicts a fourth data item with the second data item as an explanatory variable and the second data item as an dependent variable. The system includes a model integration unit that constructs an integrated model that predicts a fourth data item, which uses the first data item as an explanatory variable and is the objective variable, by inputting the prediction results of a first prediction model that predicts a second data item, which uses the first data item as an explanatory variable and is the objective variable, as explanatory variables of the second prediction model. The model integration unit performs training to adjust the parameters of the first prediction model included in the integrated model so as to reduce a predetermined loss function, using the first data and the second data of the second data item, and the predetermined loss function is defined as a function calculated based on the probability distribution of the prediction results of the first prediction model and the probability distribution of the prediction results of the integrated model. A data analysis system in which the user terminal is used by a person belonging to the first population, and the first data and the second data are transmitted from the user terminal to the data analysis device.

13. In claim 12, The system includes an administrator terminal that is connected to the aforementioned data analysis device via a network, The data analysis device includes a predictive analysis unit that uses the integrated model trained by the model integration unit to make predictions about the fourth data item for one or more records included in the first population. The prediction results from the predictive analysis unit are displayed on the administrator terminal or the user terminal in the data analysis system.

14. A data analysis method that performs data analysis using a data analysis device having a memory device and a model integration unit, The storage device stores: a first information database that stores first data of a first data item collected from a first population; a second information database that stores second data of a second data item and a third data item collected from the first population; and a prediction model database that stores parameters of a second prediction model that is trained using data collected from a second population different from the first population, and that predicts a fourth data item with the second data item as an explanatory variable and the second data item as an objective variable. The model integration unit constructs an integrated model that predicts the fourth data item, which uses the first data item as an explanatory variable and is the objective variable, by inputting the prediction results of the first prediction model, which predicts the second data item, which uses the first data item as an explanatory variable and is the objective variable, as explanatory variables of the second prediction model. A data analysis method in which the model integration unit performs training to adjust the parameters of the first prediction model included in the integrated model so as to reduce a predetermined loss function using the first data and the second data of the second data item, wherein the predetermined loss function is defined as a function calculated based on the probability distribution of the prediction results of the first prediction model and the probability distribution of the prediction results of the integrated model.

15. In claim 14, The data analysis device has a predictive analysis unit, The predictive analysis unit is a data analysis method that uses the integrated model trained by the model integration unit to predict the fourth data item for one or more records included in the first population.

Images