Data estimation device, data estimation method, and data estimation program

Abstract

A data estimation device is provided that can suppress overlearning in estimating state transition probabilities between stratified data sets. [Solution] The data estimation device includes an acquisition unit 11, a stratification unit 12, an estimation unit 13, and an output unit 15. The acquisition unit 11 acquires datasets including data exhibiting different probability distributions and attributes used for stratifying the datasets. The stratification unit 12 stratifies the datasets based on the attributes. The estimation unit 13 estimates state transition probabilities between stratified datasets based on the difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified for each attribute. The output unit 15 outputs the state transition probabilities between the stratified datasets. By using the state transition probabilities estimated in this manner, the data estimation device 10 can support decision-making based on the estimation results of the data transition destination.

Description

[Technical Field]

[0001] The present disclosure relates to a data estimation device and the like. [Background technology]

[0002] In data analysis, etc., data trends are sometimes estimated based on state transition probabilities between multiple data sets having different distributions. Furthermore, when using a data set for analysis, a data set stratified according to attributes may be used to perform analysis according to the analysis target. For example, if the attribute is gender, data stratified based on gender is used to estimate the state transition probabilities between data of different genders. Then, based on the estimated state transition probabilities, the transition destination of the data to be analyzed is estimated.

[0003] The predictive model building device of Patent Document 1 clusters medical data in the form of word frequencies by age group to generate clusters by age group.The predictive model building device of Patent Document 1 then estimates the transition probability in the direction of age progression between clusters of adjacent age groups based on the similarity between the clusters. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Laid-Open No. 2016-95684 Summary of the Invention [Problem to be solved by the invention]

[0005] In the technique described in Patent Document 1, when there is little data for each attribute, overlearning may occur in estimating state transition probabilities between data sets stratified based on attributes.

[0006] In order to solve the above-mentioned problems, the present disclosure aims to provide a data estimation device and the like that can suppress overlearning in estimating state transition probabilities between stratified data sets. [Means for solving the problem]

[0007] In order to solve the above problems, the data estimation device of the present disclosure includes: datasets including data showing different probability distributions; an acquisition means for acquiring attributes used for stratifying each of the datasets; a stratification means for stratifying each of the datasets based on the acquired attributes; an estimation means for estimating state transition probabilities between the datasets after stratification based on the difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified for each attribute; and an output means for outputting the state transition probabilities between the datasets after stratification.

[0008] The data estimation method disclosed herein acquires datasets including data exhibiting different probability distributions and attributes used for stratifying each of the datasets, stratifies each of the datasets based on the acquired attributes, estimates state transition probabilities between the stratified datasets based on the difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified by attribute, and outputs the state transition probabilities between the stratified datasets.

[0009] The data estimation program disclosed herein causes a computer to execute the following processes: a process of acquiring datasets including data showing different probability distributions and attributes to be used for stratifying each of the datasets; a process of stratifying each of the datasets based on the acquired attributes; a process of estimating state transition probabilities between the stratified datasets based on the difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified by the attributes; and a process of outputting the state transition probabilities between the stratified datasets. [Effects of the Invention]

[0010] According to the present disclosure, it is possible to suppress overfitting in estimating state transition probabilities between stratified data sets. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating an example of a configuration of a data estimation system according to the present disclosure. [Figure 2] FIG. 1 is a diagram illustrating an example of data transition in the present disclosure. [Figure 3] FIG. 1 is a diagram illustrating an example of a configuration of a data estimation device according to the present disclosure. [Figure 4] FIG. 10 is a diagram illustrating an example of an operation flow of a data estimation device according to the present disclosure. [Figure 5] FIG. 10 is a diagram illustrating an example of an operation flow of a data estimation device according to the present disclosure. [Figure 6] FIG. 1 is a diagram illustrating an example of a hardware configuration of a data estimation device according to the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0012] An embodiment of the present disclosure will be described in detail with reference to the drawings. FIG. 1 is a diagram illustrating an example of the configuration of a data estimation system. The data estimation system includes a data estimation device 10, a terminal device 20, and a data management device 30. The data estimation device 10 is connected to the terminal device 20, for example, via a network. The data estimation device 10 is connected to the data management device 30, for example, via a network. There may be a plurality of terminal devices 20 and a plurality of data management devices 30. The number of terminal devices 20 and the number of data management devices 30 may be set as appropriate.

[0013] The data estimation system, for example, estimates state transition probabilities between data sets stratified based on attributes. The data estimation system, for example, stratifies two data sets exhibiting different probability distributions based on attributes. The data estimation system then estimates state transition probabilities between the two data sets stratified based on attributes, for example, using an algorithm related to optimal transportation. The algorithm related to optimal transportation is an algorithm for solving a target problem as an optimal transportation problem. The data estimation system also estimates, for example, a transition destination on the probability distribution of one of the two stratified data sets, based on the estimated state transition probabilities.

[0014] An attribute is, for example, information indicating the characteristics of the entity to which the data belongs. An entity to which the data belongs is, for example, information indicating what the data relates to. For example, if the data is a person, the entity to which the data belongs is each person. For example, if the data is an object, the entity to which the data belongs is each object. If the data is data about a person, the attribute is, for example, information indicating the characteristics of the person. In this case, the attribute is one or more pieces of information from the nationality, race, occupation, work history, educational background, family structure, place of residence, height, weight, hobbies, medical history, and friendships of the person to whom the data relates. Furthermore, if the data is data about an object, the attribute is, for example, information indicating the characteristics of the object. Furthermore, the attribute is not limited to the above.

[0015] Furthermore, each dataset is constructed, for example, by classifying cross-sectional data. Cross-sectional data is, for example, data collected at a single point in time, and data having different positions in the time series direction with respect to the data's destination. A single point in time may have a period of time. For example, a single point in time may have a period of time, such as within the same fiscal year. For example, if the data is data about a person, the position in the time series direction with respect to the data's destination is the age of the person to which the data corresponds. For example, if the data is data about an object, the position in the time series direction with respect to the data's destination is an elapsed period, such as the period from the start of use, manufacture, or installation of the object to which the data corresponds. The position in the time series direction is not limited to the above. For example, if datasets are constructed from data about health checkup results, each dataset is constructed, for example, from the results of health checkups conducted by age group within the same fiscal year. In this case, the results of the health checkups are an example of health-related data.

[0016] The data estimation system, for example, estimates state transition probabilities between data sets obtained by classifying cross-sectional data in a time-series direction. When data sets are composed of data about people, each data set is composed of data about people in different age groups. For example, when there are two data sets, if the first data set is composed of data about people in their 50s and the second data set is a data set about people in their 70s, the data estimation system estimates the state transition probability from data included in the data set for people in their 50s to data included in the data set for people in their 70s. Then, the data estimation system estimates data for a target person in their 70s, whose time-series changes are to be estimated, based on data for that person in their 50s and the state transition probabilities. In this way, by estimating the transition destination of data based on the state transition probabilities estimated based on cross-sectional data, it is possible to estimate time-series changes for a target person without requiring time-series data about the same person as training data.

[0017] Furthermore, when a dataset is composed of data about objects, each dataset is composed of data about objects that have been in operation for a different number of years. For example, when there are two datasets, each dataset is about a machine, each dataset is composed of data about machines that have been in operation for a different number of years. For example, when a first dataset is composed of data about machines that have been in operation for less than five years, the second dataset is a dataset about machines that have been in operation for five years or more. In this case, the data estimation system estimates a state transition probability from data included in the data about machines that have been in operation for less than five years to data included in the dataset about machines that have been in operation for five years or more. Then, the data estimation system estimates data for when the machine whose time series change is to be estimated has been in operation for five years or more based on data from the machine that has been in operation for less than five years and the state transition probability.

[0018] When a dataset is composed of health-related data, the data estimation system estimates, for example, state transition probabilities between health-related data stratified by age group based on attributes. FIG. 2 is a diagram schematically illustrating an example of data transition when health checkup results are used as a dataset. The example in FIG. 2 shows data transition from the probability distribution of health checkup results at age 55 to the probability distribution of health checkup results at age 75. The example in FIG. 2 also shows data transition when not stratified by attributes and when stratified by gender. In the example in FIG. 2, when stratification is not performed, the health checkup results of a person belonging to data group A1 at age 55 transition to data group C1 when the person's age increases to 75. On the other hand, in the example in FIG. 2, when stratification is not performed, the health checkup results of a person belonging to data group A1 at age 55 transition to data group C2 when the person's age increases to 75.

[0019] Stratification in this manner may change the transition destination of data. Therefore, estimating the transition destination of data in a data set stratified based on attributes may improve the accuracy of estimating the transition destination of data. On the other hand, stratification reduces the number of data, which may lead to a state of overfitting that is strongly influenced by some data, and the accuracy of estimating the transition destination of data may decrease. For example, to suppress such a state of overfitting, the data estimation device 10 estimates the state transition probabilities between stratified data sets by regularizing them so that they do not deviate from the trend of the state transition probabilities between non-stratified data sets.

[0020] Here, an example of the configuration of the data estimation device 10 will be described. Fig. 3 is a diagram showing an example of the configuration of the data estimation device 10. The data estimation device 10 basically includes an acquisition unit 11, a stratification unit 12, an estimation unit 13, and an output unit 15. The data estimation device 10 further includes, for example, an onset probability estimation unit 14 and a storage unit 16.

[0021] The acquisition unit 11 acquires datasets including data showing different probability distributions and attributes used for stratifying each of the datasets. For example, the acquisition unit 11 acquires two datasets including data showing different probability distributions and attributes used for stratifying the datasets.

[0022] The acquiring unit 11 acquires, for example, a dataset configured by data at different positions in the time series direction. The dataset configured by data at different positions in the time series direction is, for example, a dataset in which cross-sectional data is classified in the time series direction. For example, the datasets at different positions in the time series direction are datasets in which the age groups or the number of years elapsed differ for each dataset.

[0023] A data set including data showing different probability distributions is, for example, a data set obtained by classifying cross-sectional data in a time series direction. Also, a data set including data showing different probability distributions is, for example, a data set in which the probability distributions of the respective data sets are different from each other when the probability distributions of the respective data sets are generated.

[0024] The acquisition unit 11 acquires, for example, a dataset composed of health-related data. When the dataset is composed of health-related data, the dataset is, for example, health-related data for a first age group and a second age group that is an age group older than the first age group. For example, when the dataset is composed of health-related data, each dataset is health-related data for a group in each age group. For example, because health-related data shows different trends depending on the age group, health-related data corresponding to each age group may have different probability distributions. The acquisition unit 11 acquires, for example, a dataset used to estimate state transition probabilities from the data management device 30.

[0025] Health-related data is, for example, data indicating a health condition. Health-related data is, for example, data on one or more items of health checkup results, hospital test results, vital signs, whether or not a disease has occurred, the probability of disease onset, a doctor's findings, motor function, whether or not care is required, and the level of care required. Health checkup results are, for example, data on one or more items of height, weight, eyesight, blood pressure, abdominal circumference, hearing, blood test measurements, image diagnostic results, and doctor's interview results measured during the health checkup. Health-related data may also include expenses necessary for maintaining a healthy state or for daily life. Health-related data is not limited to the above.

[0026] When the health-related data is the results of a health checkup, the acquisition unit 11 acquires, for example, the results of the health checkup within a predetermined period as the health-related data. The predetermined period is set, for example, as a period during which a sufficient amount of data can be collected and during which the trend of the data does not change. The trend of the data does not change, for example, when the standards for conducting health checkups do not change and data can be acquired using the same standards. The predetermined period is, for example, within the same fiscal year. The predetermined period may be multiple years. The predetermined period may be one month or multiple months. The predetermined period is not limited to the above.

[0027] The acquisition unit 11 acquires, for example, one or more attributes as attributes to be used for stratification. The acquisition unit 11 may acquire attributes that are candidates for attributes to be used for stratification. When the dataset is composed of health-related data, the attributes are, for example, information on one or more items of gender, place of residence, nationality, occupation, medical history, and medical history of family members. The attributes are not limited to the above. Furthermore, when the dataset is classified by an attribute other than age group, the acquisition unit 11 may acquire information indicating age as an attribute. The acquisition unit 11 acquires information indicating the attributes to be used for stratification from, for example, the terminal device 20.

[0028] Furthermore, the acquisition unit 11 acquires data and attributes related to a person or object whose time-series change is to be estimated. When the dataset is composed of health-related data, the acquisition unit 11 acquires, for example, the health-related data of the target person. For example, the acquisition unit 11 acquires the most recent health-related data from the health-related data of the target person. For example, when the health-related data is the results of a medical checkup, the acquisition unit 11 acquires the results of the most recent medical checkup the target person underwent. Furthermore, the acquisition unit 11 acquires, for example, the age of the target person. The age of the target person may be, for example, the age at the time the health-related data of the target person was measured. The acquisition unit 11 acquires data and attributes related to the person or object whose time-series change is to be estimated from, for example, the data management device 30. The acquisition unit 11 may also acquire data related to the person or object whose time-series change is to be estimated from the terminal device 20.

[0029] The stratification unit 12 stratifies each of the datasets based on the attributes acquired by the acquisition unit 11. For example, the stratification unit 12 extracts data that matches the attributes acquired by the acquisition unit 11 from each of the datasets. Then, the stratification unit 12 generates a dataset constituted by the extracted data as a stratified dataset. The stratified dataset is also referred to as a sub-dataset, for example. Matching may also include similarity.

[0030] The stratification unit 12 may stratify each of the data sets using an attribute related to the characteristics of a cluster when at least one of the data sets is clustered. For example, the stratification may be performed using candidate attributes S1, ..., S j When a cluster classifier is trained using some of these, the set of attribute information used as input to the cluster classifier with the highest accuracy rate is set as the attribute to be used for stratification.

[0031] The estimation unit 13 estimates the state transition probability between the stratified datasets based on the difference in distribution between the state transition probability between the datasets before stratification and the state transition probability between the datasets stratified by attribute. The estimation unit 13 estimates the state transition probability between the stratified datasets using an algorithm related to optimal transportation. In the process of estimating the optimal transportation between the datasets, the estimation unit 13 estimates the state transition probability between the stratified datasets using, for example, a first loss related to the transportation cost between the datasets stratified by attribute and a second loss based on the difference in distribution between the state transition probability between the datasets before stratification and the state transition probability between the stratified datasets. In addition, the first loss is calculated based on, for example, the sum of the transportation costs in each of the stratified datasets.

[0032] The loss functions of the first loss and the second loss are weighted using, for example, a weighting function based on attributes. The weights are set, for example, so that the weight is heavier for attributes that are less likely to change. For example, when a dataset is composed of health-related data and the attribute is gender, the weight is set to be heavier than for attributes that may change, because biological sex does not change. For example, when the attribute is a lifestyle habit such as whether or not smoking, the weight is set to be smaller than for attributes that do not change, because changes may occur. The weights may be set in stages based on the likelihood of change. Furthermore, the weights may be determined so as to improve the accuracy of estimating the transition destination between datasets by using actual data for which the transition destination between datasets is known as ground truth data.

[0033] The estimation unit 13 estimates, for example, a state transition probability from data on the probability distribution of an earlier data set in the time series to data on the probability distribution of a later data set in the time series. For example, suppose that two data sets are each composed of health-related data, and the two data sets are one composed of health-related data for a first age group and the other composed of health-related data for a second age group that is an age group older than the first age group. In this case, the estimation unit 13 estimates, for example, a state transition probability between a data set on the health of people in a first age group stratified based on attributes and a data set on the health of people in a second age group stratified based on attributes.

[0034] The process of estimating the state transition probability between stratified datasets using an algorithm related to optimal transportation will be described in more detail. Assume that two datasets for which state transition probabilities are to be estimated have distributions μ and ν. In this case, the distributions in a dataset stratified only to data with attribute S among the data contained in the two datasets are represented by distributions μ(·|S) and ν(·|S), respectively. In this case, the estimation unit 13 estimates the optimal transportation between μ(·|S) and ν(·|S) as shown in the following equation 1, for example.

[0035]

number

[0036] The estimation unit 13 estimates the optimal transportation between data sets independent of attributes, which is expressed by the following equation 2.

[0037]

number

[0038]

number

[0039]

number

[0040] The estimation unit 13 simultaneously optimizes π(·|S) for all attributes S by performing multi-task optimization for the first loss and the second loss. That is, the estimation unit 13 estimates the optimal transfer between (·|S) and ν(·|S), for example, by estimating π(·|S) that simultaneously minimizes the first loss and the second loss for all attributes S. Furthermore, such multi-task optimization corresponds to, for example, regularization of the second loss by the first loss.

[0041] In the above process, the estimation unit 13 solves the optimal transportation problem using the dataset before stratification, estimates the overall trend of state transition probabilities, and estimates the state transition probabilities between each dataset stratified by attribute information so that they do not deviate from the overall trend. By making estimations in this manner, overlearning can be suppressed when solving the optimal transportation problem, even if the number of data included in the dataset stratified by attribute is small.

[0042] Also, if attribute S is S1,...,S j and attribute S in the dataset j The number of j In this case, the estimation unit 13 may calculate the optimal transportation π independent of the attribute using the following formula 5.

[0043]

number

[0044] The estimation unit 13 estimates a transition destination on the probability distribution of one dataset for data that is within the range of the probability distribution of the other dataset, based on the state transition probability between stratified datasets. For example, the estimation unit 13 estimates a transition destination on the probability distribution of a subsequent dataset in the time series for data that is within the range of the probability distribution of an earlier dataset in the time series, based on the state transition probability between stratified datasets.

[0045] The estimation unit 13 estimates a transition destination when data on a probability distribution based on an earlier dataset in the time series transitions to a probability distribution based on a later dataset in the time series, for example, based on the state transition probabilities between stratified datasets. For example, when a dataset is composed of health-related data, the estimation unit 13 estimates health-related data when a person in a first age group changes age to a second age group, based on health-related data of the person, attributes of the person, and state transition probabilities according to the attributes. For example, when the attribute is male, the estimation unit 13 estimates health-related data when a man in the first age group changes age to the second age group, using state transition probabilities estimated based on age-group datasets of men's health-related data.

[0046] The estimation unit 13 may estimate a state transition probability between data sets based on the disease onset probability. For example, the estimation unit 13 estimates a state transition probability between data sets based on the disease onset probability for each attribute using a data set configured of data on the disease onset probability by age group estimated by the onset probability estimation unit 14. Then, for example, the estimation unit 13 estimates a disease onset probability when a person in a first age group changes age to a second age group based on the disease onset probability of a person in a first age group, the attributes of the person, and the state transition probability according to the attributes.

[0047] When the dataset is composed of health-related data, the onset probability estimation unit 14 estimates the disease onset probability for each dataset of health-related data, for example, using an estimation model that estimates the disease onset probability from the health-related data. The onset probability estimation unit 14 also estimates the disease onset probability from the health-related data stratified based on attributes, for example, using the estimation model. The estimation model is, for example, a machine learning model that uses the health-related data as input and estimates the disease onset probability.

[0048] The estimation model is generated, for example, based on the difference between the disease incidence probability estimated using stratified health data and the disease incidence probability based on unstratified health data. The estimation model is generated, for example, by machine learning using a third loss based on the disease incidence probability estimated using the stratified health data and ground truth data, and a fourth loss based on the difference between the disease incidence probability estimated using the stratified health data and the disease incidence probability based on unstratified health data. The loss function representing the third loss and the fourth loss each include, for example, a weight function based on an attribute. The weight is set, for example, so that the greater the impact on the disease, the greater the weight. For example, when estimating the probability of developing diabetes, the weight of the attribute regarding whether or not the person consumes alcohol is set greater than the other weights.

[0049] The estimation model is generated, for example, as follows: A learning device that generates the estimation model generates an estimation model f(x) that is not dependent on attributes. The learning device generates the estimation model f(x) by performing machine learning using, for example, training data that is not stratified by attributes.

[0050] The learning device defines the first loss as the third loss for any weight function w1(S) as shown in the following equation (6).

[0051]

number

[0052]

number

[0053] The learner simultaneously optimizes f(x|S) for all attributes S, for example, by performing multi-task optimization for the third loss and the fourth loss. For example, the learner generates an estimation model by estimating f(x|S) that simultaneously minimizes the third loss and the fourth loss for all attributes S. Furthermore, such multi-task optimization corresponds to, for example, regularization of the fourth loss by the third loss.

[0054] In the above process, the learning device generates an estimation model so that the estimation results obtained using the stratified dataset do not deviate from the overall trend. By generating an estimation model in this way, overfitting in the generation of the estimation model can be suppressed even when the number of data points included in the attribute-stratified dataset is small.

[0055] The output unit 15 outputs, for example, the state transition probabilities between the stratified datasets. The output unit 15 stores, for example, the state transition probabilities between the stratified datasets as data for each attribute. The output unit 15 may output, as data for each attribute, a map of the state transition probabilities between the stratified datasets. The output unit 15 may also output the state transition probabilities for each attribute and the state transition probabilities between non-stratified datasets. For example, the output unit 15 outputs a map of the state transition probabilities for each attribute and a map of the state transition probabilities between non-stratified datasets. The output unit 15 stores the state transition probabilities between the stratified datasets for each attribute by outputting the state transition probabilities between the stratified datasets for each attribute to, for example, the storage unit 16. The output unit 15 also outputs the state transition probabilities between the stratified datasets for each attribute to, for example, the terminal device 20. The output unit 15 may also output the state transition probabilities between the stratified datasets for each attribute to, for example, the data management device 30.

[0056] The output unit 15 outputs, for example, an estimation result of the target person's future health data. When the dataset is composed of health-related data, the output unit 15 outputs, for example, an estimation result of the target person's future health data. The output unit 15 outputs health data for when the target person's age increases based on the estimation result of the target person's future health data. The output unit 15 outputs, for example, health data for when the target person's age increases to a second age group. For example, when the first age group is 50 years old and the second age group is 70 years old, the output unit 15 outputs health data for when the target person is 50 years old and an estimation result of the health data for when the target person is 70 years old. The output unit 15 may also output an estimation result of the target person's future disease onset probability. The output unit 15 outputs the estimation result of the target person's future health data to, for example, the terminal device 20.

[0057] The storage unit 16 stores, for example, information related to the process of estimating a state transition function between data sets. The storage unit 16 stores, for example, a data set used in the process of estimating a state transition function. The storage unit 16 stores, for example, a state transition probability for each attribute. When a data set is composed of health-related data, the storage unit 16 estimates, for example, a state transition probability between health-related data for each attribute. The storage unit 16 stores, for example, an estimation model. The estimation model may be stored in a storage means external to the data estimation device 10.

[0058] The terminal device 20 is, for example, a terminal device that accesses the data estimation device 10 and is used for processing to estimate state transition probabilities between data sets. The terminal device 20 acquires, for example, an estimation result of the state transition probabilities between data sets from the output unit 15 of the data estimation device 10. When the data sets are composed of health-related data, the terminal device 20 acquires, for example, a state transition probability between health-related data for different age groups. Furthermore, the terminal device 20 outputs, for example, an estimation result of the state transition probabilities between data sets to a display device (not shown).

[0059] The terminal device 20 acquires the estimation result of the data transition destination from, for example, the output unit 15 of the data estimation device 10. When the dataset is composed of health-related data, the terminal device 20 acquires, for example, the estimation result of the health-related data when the target person's age increases. The terminal device 20 also outputs the estimation result of the data transition destination to, for example, a display device (not shown).

[0060] The terminal device 20 acquires, for example, attributes of data for which a transition destination is to be estimated, which are input by an operator's operation. Then, the terminal device 20 outputs, for example, the attributes of the data for which a transition destination is to be estimated to the acquisition unit 11 of the data estimation device 10. Furthermore, when data for which a transition destination is to be estimated is input to the terminal device 20, the terminal device 20 outputs, for example, the data for which a transition destination is to be estimated to the acquisition unit 11 of the data estimation device 10.

[0061] When the data set is composed of health-related data, the worker using the terminal device 20 may be, for example, a person who provides advice to the target person for making decisions regarding health or assets, or the target person. A person who provides advice to the target person may be, for example, a medical professional, an insurance officer, a human resources officer, a financial planner, or a financial institution employee. A medical professional may be a doctor, nurse, physical therapist, pharmacist, laboratory technician, or counselor. Medical professionals are not limited to the above. Furthermore, a person who provides advice to the target person is not limited to the above.

[0062] For example, a personal computer, a tablet computer, a smartphone, or a smartwatch can be used as the terminal device 20. The information processing device used as the terminal device 20 is not limited to the above.

[0063] The data management device 30 stores, for example, a dataset used to estimate the state transition probability. For example, the data management device 30 stores a dataset in association with attributes of the data included in the dataset. The data management device 30 stores, for example, health-related data as a dataset used to estimate the state transition probability. For example, the data management device 30 may store the health-related data in association with the measurement date of the health-related data and the age of the person corresponding to the health-related data. Furthermore, the data management device 30 outputs the dataset and attributes to, for example, the acquisition unit 11 of the data estimation device 10.

[0064] When a data set is composed of health-related data, the health-related data may be, for example, the results of health checkups by age group. The data management device 30 stores, for example, the date of the health checkup, the attributes of the person who underwent the health checkup, and the results of the health checkup in association with each other. The attributes may be, for example, information on one or more of age, sex, place of residence, nationality, occupation, medical history, and medical history of family members. The attributes are not limited to the above. The date of the health checkup may be information indicated by the month, year, or fiscal year in which the health checkup was conducted. Furthermore, the data management device 30 may store the results of the health checkup as a database categorized based on at least one of the date of the health checkup and the attributes of the person who underwent the health checkup.

[0065] If the data set is the results of a health checkup, the data management device 30 may store the results of the health checkup as, for example, anonymously processed information or pseudonymized information. Anonymously processed information is, for example, information that has been processed so that an individual cannot be identified even when compared with other information. Pseudonymized information is, for example, information that cannot identify an individual by itself but has been processed so that an individual can be identified when compared with other information.

[0066] The data management device 30 may store the results of health checkups conducted for a predetermined group. The predetermined group may be, for example, a group for which a health checkup is conducted. The predetermined group may be, for example, a local government, a company, an organization, a cooperative, a school, or a health insurance association. The predetermined group is not limited to the above. The data management device 30 may also store the results of health checkups for multiple groups.

[0067] The following describes the process of estimating state transition probabilities between stratified data sets by the data estimation device 10. Fig. 4 shows an example of the operational flow of the process of estimating state transition probabilities between stratified data sets by the data estimation device 10.

[0068] The acquiring unit 11 acquires data sets including data showing different probability distributions and attributes used for stratifying each of the data sets (step S11). The acquiring unit 11 acquires the data sets from, for example, the data management device 30. In addition, the acquiring unit 11 acquires attributes used for stratifying each of the data sets from, for example, the terminal device 20.

[0069] When the data sets and the attributes used for stratifying each data set are acquired, the stratification unit 12 stratifies each data set based on the acquired attributes (step S12).

[0070] Once the datasets are stratified, the estimation unit 13 estimates the state transition probabilities between the datasets after stratification based on the difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified by attribute (step S13).

[0071] When the state transition probabilities between the stratified datasets are estimated, the output unit 15 outputs the state transition probabilities between the stratified datasets (step S14). The output unit 15 stores the state transition probabilities between the stratified datasets by outputting the state transition probabilities between the stratified datasets to, for example, the storage unit 16. The output unit 15 may output the state transition probabilities between the stratified datasets to, for example, the terminal device 20.

[0072] The following describes the operation of the data estimation device 10 to estimate health-related data for an age-group-specific data set based on state transition probabilities when the target person's age increases. Figure 5 shows an example of the operation flow of the process in which the data estimation device 10 estimates health-related data for an age-group-specific data set.

[0073] The acquisition unit 11 acquires, for example, data related to the health of the target person and attributes of the target person (step S21). The acquisition unit 11 acquires the data related to the health of the target person from, for example, the data management device 30. The acquisition unit 11 also acquires the age and attributes of the target person from, for example, the terminal device 20.

[0074] When the health data and attributes of the target person are acquired, the estimation unit 13 estimates the health data of the target person when the target person's age increases, based on the state transition probability data corresponding to the target person's age and attributes (step S22). The estimation unit 13 estimates the transition destination data of the health data of the target person as the health data of the target person when the target person increases in age, based on the state transition probability data corresponding to the target person's age and attributes.

[0075] When the data on the health of the target person as the target person ages is estimated, the output unit 15 outputs the estimated result of the data on the health of the target person as the target person ages (step S23). The output unit 15 outputs the data on the health of the target person as the target person ages to, for example, the terminal device 20.

[0076] The processes in the data estimation device 10 may be distributed and executed among a plurality of information processing devices connected via a network. For example, the process in the stratification unit 12 and the process in the estimation unit 13 may be executed in different information processing devices. It can be appropriately set which information processing device executes each process in the data estimation device 10.

[0077] The data estimation device 10 stratifies each of the datasets including data exhibiting different probability distributions based on attributes. The data estimation device 10 estimates the state transition probabilities between the stratified datasets based on the difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified by attributes. By making such estimations, the data estimation device 10 can suppress overlearning in estimating the state transition probabilities between the stratified datasets. As a result, the data estimation device 10 can improve the estimation accuracy of the state transition probabilities.

[0078] The data estimation device 10 solves an optimal transportation problem using, for example, a data set before stratification, and estimates an overall trend of state transition probabilities. Then, for example, the data estimation device 10 estimates the state transition probabilities between each stratified data set so that they do not deviate from the overall trend. By making such estimations, the data estimation device 10 can suppress overlearning when solving the optimal transportation problem, even when the number of data included in the data set stratified by attribute is small.

[0079] Furthermore, the data estimation device 10 weights the loss function used to solve the optimal transportation problem, for example, so that the weight is greater for attributes that are less likely to change. By weighting attributes that are less likely to change more, the influence of attributes that are more likely to cause fluctuations in the estimation results can be suppressed. Therefore, by suppressing the influence of data on attributes that are more likely to cause fluctuations in the estimation results, it is possible to suppress a situation in which a small amount of data has a large impact on the overall estimation results, thereby suppressing overlearning when solving the optimal transportation problem.

[0080] By estimating the transition destination of data based on the state transition probability estimated as described above, the data estimation device 10 can estimate a future state without needing long-term data on the same person or object, for example. Therefore, the data estimation device 10 can support the decision-making of a person who uses the estimation result of the transition destination of data, for example.

[0081] When a dataset is composed of health-related data, the data estimation device 10 can estimate the health-related data of a target person as the target person ages by estimating the transition destination of the health-related data of the target person based on the state transition probability. Furthermore, the data estimation device 10 can estimate future health-related data without requiring long-term health-related data for the same person, for example. Therefore, the data estimation device 10 can easily estimate health-related data as the target person ages. Furthermore, by estimating health-related data as the target person ages, the data estimation device 10 can support health-related decision-making by the target person, for example.

[0082] Furthermore, by estimating the disease onset probability, the data estimation device 10 can provide information regarding the future disease onset risk of the target person. Furthermore, by estimating the medical expenses required for receiving medical treatment in the future, the data estimation device 10 can provide information for creating an asset plan for the target person.

[0083] Each process in the data estimation device 10 can be realized by executing a computer program on a computer. Fig. 6 shows an example of the configuration of a computer 100 that executes a computer program that performs each process in the data estimation device 10. The computer 100 includes a CPU (Central Processing Unit) 101, a memory 102, a storage device 103, an input / output I / F (Interface) 104, and a communication I / F 105.

[0084] The CPU 101 reads and executes computer programs for performing each process from the storage device 103. The CPU 101 may be configured by a combination of multiple CPUs. Furthermore, the CPU 101 may be configured by a combination of a CPU and another type of processor. For example, the CPU 101 may be configured by a combination of a CPU and a graphics processing unit (GPU). The memory 102 is configured by a dynamic random access memory (DRAM) or the like, and temporarily stores the computer programs executed by the CPU 101 and data being processed. The storage device 103 stores the computer programs executed by the CPU 101. The storage device 103 is configured by, for example, a non-volatile semiconductor storage device. Other storage devices such as a hard disk drive may also be used for the storage device 103. The input / output I / F 104 is an interface that receives input from an operator and outputs display data, etc. The communication I / F 105 is an interface that transmits and receives data between the terminal device 20, the data management device 30, and other information processing devices. Furthermore, the terminal device 20 and the data management device 30 may also have the same configuration as the computer 100.

[0085] The computer program used to execute each process can also be stored and distributed on a computer-readable recording medium that non-temporarily stores data. Examples of recording media that can be used include magnetic tapes for recording data and magnetic disks such as hard disks. Optical disks such as CD-ROMs (Compact Disc Read Only Memory) can also be used as recording media. Non-volatile semiconductor storage devices can also be used as recording media.

[0086] Some or all of the above-described embodiments can be described as, but are not limited to, the following supplementary notes.

[0087] [Appendix 1] a data set including data showing different probability distributions, and an acquisition means for acquiring attributes used for stratification of each of the data sets; a stratification means for stratifying each of the data sets based on the acquired attributes; an estimation means for estimating state transition probabilities between the stratified datasets based on a difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified for each attribute; an output means for outputting state transition probabilities between the stratified data sets; A data estimation device comprising:

[0088] [Appendix 2] the estimation means estimates the state transition probability between the stratified datasets by using a first loss related to transportation costs between the datasets stratified for each attribute, and a second loss based on a difference in distribution between the state transition probability between the datasets before stratification and the state transition probability between the stratified datasets; 2. The data estimation apparatus of claim 1.

[0089] [Appendix 3] the loss function of each of the first loss and the second loss includes a weight function based on an attribute; 3. The data estimation apparatus of claim 2.

[0090] [Appendix 4] The method further comprises an onset probability estimation means for estimating a disease onset probability for each data set of health-related data using an estimation model for estimating a disease onset probability from health-related data, the estimation means estimates a state transition probability between data sets based on the disease onset probability; 4. A data estimation device according to any one of appendixes 1 to 3.

[0091] [Appendix 5] The estimation model is generated by machine learning using a third loss based on the disease onset probability estimated using stratified health data and ground truth data, and a fourth loss based on the difference between the disease onset probability estimated using stratified health data and the onset probability based on unstratified health data. 5. The data estimation apparatus of claim 4.

[0092] [Appendix 6] the loss function of each of the third loss and the fourth loss includes a weight function based on an attribute; 6. The data estimation device according to claim 5.

[0093] [Appendix 7] The stratification means stratifies each of the data sets using an attribute related to a characteristic of a cluster when at least one of the data sets is clustered. 7. A data estimation device according to any one of appendixes 1 to 6.

[0094] [Appendix 8] each of the data sets is a data set configured by data at different positions in a time series direction, The estimation means estimates a state transition probability from data on the probability distribution of a data set that is earlier in the time series to data on the probability distribution of a data set that is later in the time series. 8. A data estimation device according to any one of appendices 1 to 7.

[0095] [Appendix 9] the estimation means estimates a transition destination when data on a probability distribution based on an earlier data set in the time series transitions to a probability distribution based on a later data set in the time series, based on the state transition probability. 9. The data estimation apparatus of claim 8.

[0096] [Appendix 10] the dataset is a dataset composed of health-related data in a first age group and a dataset composed of health-related data in a second age group that is an age group older than the first age group; the estimation means estimates a state transition probability between data on the health of people in a first age group stratified based on the attribute and data on the health of people in a second age group stratified based on the attribute; 4. A data estimation device according to any one of appendixes 1 to 3.

[0097] [Appendix 11] the estimation means estimates health-related data of the person in the first age group when the person's age changes to a second age group, based on the health-related data of the person in the first age group and the state transition probability; 11. The data estimation apparatus of claim 10.

[0098] [Appendix 12] The first loss is calculated based on a sum of transportation costs in each of the stratified data sets. 3. The data estimation apparatus of claim 2.

[0099] [Appendix 13] Obtaining data sets including data showing different probability distributions and attributes used for stratification of each of the data sets; stratifying each of the data sets based on the acquired attributes; Estimating state transition probabilities between the stratified datasets based on a difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified for each attribute; outputting state transition probabilities between the stratified datasets; Data estimation methods.

[0100] [Appendix 14] A process of acquiring data sets including data showing different probability distributions and attributes used for stratification of each of the data sets; stratifying each of the data sets based on the acquired attributes; A process of estimating state transition probabilities between the stratified datasets based on a difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified for each attribute; a process of outputting state transition probabilities between the stratified data sets; A data estimation program that causes a computer to execute the above.

[0101] Furthermore, some or all of the configurations described in Supplementary Notes 2 to 12, which are dependent on Supplementary Note 1, may also be dependent on Supplementary Notes 13 and 14 in the same dependent relationship as Supplementary Notes 2 to 12. Furthermore, not limited to Supplementary Notes 1, 13, and 14, some or all of the configurations described as Supplements may be made dependent on various hardware, software, various recording means for recording software, or systems, within the scope of each of the above-mentioned embodiments.

[0102] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate. [Explanation of symbols]

[0103] 10 Data Estimation Device 11 Acquisition Department 12 Stratification Department 13 Estimation part 14. Onset probability estimation section 15 Output section 16 Memory section 20 Terminal equipment 30 Data management device 100 computers 101 CPU 102 memory 103 Storage device 104 Input / Output Interface 105 Communication I / F

Claims

1. a data set including data showing different probability distributions, and an acquisition means for acquiring attributes used for stratification of each of the data sets; a stratification means for stratifying each of the data sets based on the acquired attributes; an estimation means for estimating state transition probabilities between stratified datasets based on a difference in distribution between state transition probabilities between datasets before stratification and state transition probabilities between datasets stratified for each attribute; an output means for outputting state transition probabilities between the stratified data sets; A data estimation device comprising:

2. the estimation means estimates a state transition probability between the stratified datasets by using a first loss related to transportation costs between the datasets stratified for each attribute, and a second loss based on a difference in distribution between a state transition probability between the datasets before stratification and a state transition probability between the stratified datasets; The data estimation device according to claim 1 .

3. the loss function of each of the first loss and the second loss includes a weight function based on an attribute; The data estimation device according to claim 2 .

4. The method further comprises an onset probability estimation means for estimating a disease onset probability for each of the data sets of health-related data using an estimation model for estimating a disease onset probability from health-related data, the estimation means estimates a state transition probability between data sets based on the disease onset probability; 4. The data estimation device according to claim 1.

5. The estimation model is generated by machine learning using a third loss based on the disease onset probability estimated using stratified health-related data and ground truth data, and a fourth loss based on the difference between the disease onset probability estimated using stratified health-related data and the onset probability based on non-stratified health-related data. The data estimation device according to claim 4.

6. The data sets are each composed of data at different positions in the time series direction. the estimation means estimates a state transition probability from data on the probability distribution of a data set earlier in the time series to data on the probability distribution of a data set later in the time series; 4. The data estimation device according to claim 1.

7. the dataset is a dataset composed of health-related data for a first age group and a dataset composed of health-related data for a second age group that is an age group older than the first age group; the estimation means estimates a state transition probability between data on the health of people in a first age group stratified based on the attribute and data on the health of people in a second age group stratified based on the attribute; 4. The data estimation device according to claim 1.

8. the estimation means estimates health-related data of the person in the first age group when the person's age changes to a second age group, based on health-related data of the person in the first age group and the state transition probability; The data estimation device according to claim 7.

9. Obtaining data sets including data showing different probability distributions and attributes used for stratification of each of the data sets; stratifying each of the data sets based on the acquired attributes; Estimating state transition probabilities between the stratified datasets based on a difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified for each attribute; outputting state transition probabilities between the stratified datasets; Data estimation methods.

10. A process of acquiring data sets including data showing different probability distributions and attributes used for stratification of each of the data sets; stratifying each of the data sets based on the acquired attributes; A process of estimating state transition probabilities between the stratified datasets based on a difference in distribution between the state transition probabilities between the datasets before stratification and the state transition probabilities between the datasets stratified for each attribute; a process of outputting state transition probabilities between the stratified data sets; A data estimation program that causes a computer to execute the above.

Images