Information processing device, program, and information processing method

Abstract

The estimation device (100) is characterized by comprising: an input unit (110) that generates teacher data including a cause variable, a process type, and a result variable for each of a plurality of processes; and a learning unit (120) that uses the teacher data to generate a learning model by learning the result variable from the cause variable and the process type for each of the plurality of processes.

Description

[Technical field]

[0001] The present disclosure relates to an information processing device, a program, and an information processing method. [Background technology]

[0002] There are various known techniques for estimating individual processing effects, which are effects of a certain process. For example, a conventional technique for estimating individual processing effects is a method for estimating individual processing effects at a single time point using a multitask Gaussian process, as shown in Non-Patent Document 1. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Alaa, AM & van der Schaar, M. “Bayesian inference of individualized treatment effects using multi-task gaussian processes”, In Advances in Neural Information Processing Systems, pages 3424-3432, 2017. Summary of the Invention [Problem to be solved by the invention]

[0004] In Non-Patent Document 1, only the effect of individual processing can be known when there is one type of processing at a single time point and when there is no such processing, and the degree of processing amount cannot be taken into consideration. Therefore, when there are multiple types of processing, it is not known which processing and how much of it should be performed to maximize the effect of each individual processing.

[0005] Therefore, one or more aspects of the present disclosure aim to make it possible to learn the processing effects of multiple processes. [Means for solving the problem]

[0006] An information processing device according to a first aspect of the present disclosure includes: an input unit that generates teacher data including a cause variable, a process type, and a result variable for each of a plurality of processes; and a learning unit that generates a learning model by learning the result variable from the cause variable and the process type for each of the plurality of processes using the teacher data. The causal variables are at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use, the process types are fare increase, fare reduction, campaign for promoting railway use, and distribution of coupons for promoting railway use, and the result variable is the sales amount of high-priced railway services. It is characterized by:

[0007] An information processing device according to a second aspect of the present disclosure is characterized in that it comprises an input unit that generates teacher data including, for each of a plurality of processes, a first cause variable that changes over time, a second cause variable that does not change over time, a process type, and historical information indicating a history of the change over time of the first cause variable, and a learning unit that uses the teacher data to generate a learning model for each of the plurality of processes by learning the change over time of the first cause variable at a next time from the first cause variable, the second cause variable, and the process type at a certain time according to the historical information.

[0008] An information processing device according to a third aspect of the present disclosure includes an estimation unit that estimates a result variable by inputting a cause variable and a process type into a learning model generated by learning the result variable from the cause variable and the process type for each of a plurality of processes using teacher data including the cause variable, the process type, and a result variable for the plurality of processes, and an output unit that outputs a result of the estimation. The causal variables are at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use, the process types are fare increase, fare reduction, campaign for promoting railway use, and distribution of coupons for promoting railway use, and the result variable is the sales amount of high-priced railway services. It is characterized by:

[0009] An information processing device according to a fourth aspect of the present disclosure is characterized in that it comprises: an estimation unit that uses teacher data including, for each of a plurality of processes, a first causal variable that changes over time, a second causal variable that does not change over time, a process type, and historical information indicating a history of the time change of the first causal variable, and uses a learning model generated by learning the time change of the first causal variable at a next time from the first causal variable, the second causal variable, and the process type at a certain time according to the historical information for each of the plurality of processes, to estimate the time change of the first causal variable obtained by a combination of two or more processes selected from the plurality of processes during a certain period; and an output unit that outputs the result of the estimation.

[0010] A program according to a first aspect of the present disclosure causes a computer to function as an input unit that generates teacher data including a cause variable, a process type, and a result variable for each of a plurality of processes, and a learning unit that generates a learning model by learning the result variable from the cause variable and the process type for each of the plurality of processes using the teacher data. The causal variables are at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use, the process types are fare increase, fare reduction, campaign for promoting railway use, and distribution of coupons for promoting railway use, and the result variable is the sales amount of high-priced railway services. It is characterized by:

[0011] A program according to a second aspect of the present disclosure causes a computer to function as an input unit that generates teacher data including, for each of a plurality of processes, a first causal variable that changes over time, a second causal variable that does not change over time, a process type, and historical information indicating a history of the change over time of the first causal variable, and a learning unit that uses the teacher data to generate a learning model by learning the change over time of the first causal variable at a next time from the first causal variable, the second causal variable, and the process type at a certain time according to the historical information for each of the plurality of processes.

[0012] A program according to a third aspect of the present disclosure causes a computer to function as an estimation unit that estimates a result variable by inputting a cause variable and a process type into a learning model generated by learning the result variable from the cause variable and the process type for each of a plurality of processes using teacher data including the cause variable, the process type, and a result variable for the each of the plurality of processes, and an output unit that outputs a result of the estimation. The causal variables are at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use, the process types are fare increase, fare reduction, campaign for promoting railway use, and distribution of coupons for promoting railway use, and the result variable is the sales amount of high-priced railway services. A program characterized by.

[0013] A program according to a fourth aspect of the present disclosure causes a computer to function as an estimation unit that estimates the time change of the first causal variable obtained by a combination of two or more processes selected from the multiple processes during a certain period using a learning model generated by learning the time change of the first causal variable at a next time from the first causal variable, the second causal variable, and the processing type at a certain time according to the historical information for each of the multiple processes, using teacher data including a first causal variable that changes over time, a second causal variable that does not change over time, a processing type, and historical information indicating a history of the time change of the first causal variable for each of the multiple processes, and an output unit that outputs the result of the estimation.

[0014] An information processing method according to a first aspect of the present disclosure generates training data including a cause variable, a process type, and a result variable for each of a plurality of processes, and generates a learning model by learning the result variable from the cause variable and the process type for each of the plurality of processes using the training data. An information processing method, wherein the causal variables are at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use, the processing types are fare increase, fare reduction, campaign for promoting railway use, and distribution of coupons for promoting railway use, and the result variable is the sales amount of high-priced railway services. It is characterized by:

[0015] An information processing method according to a second aspect of the present disclosure is characterized in that it generates, for each of a plurality of processes, teacher data including a first cause variable that changes over time, a second cause variable that does not change over time, a process type, and historical information indicating a history of the change in the first cause variable over time, and generates a learning model by using the teacher data to learn the change in the first cause variable over time at a next time from the first cause variable, the second cause variable, and the process type at a certain time according to the historical information for each of the plurality of processes.

[0016] An information processing method according to a third aspect of the present disclosure includes: using teacher data including a cause variable, a process type, and a result variable for each of a plurality of processes, inputting the cause variable and the process type into a learning model generated by learning the result variable from the cause variable and the process type for each of the plurality of processes, thereby estimating the result variable, and outputting a result of the estimation. An information processing method, wherein the causal variables are at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use, the processing types are fare increase, fare reduction, campaign for promoting railway use, and distribution of coupons for promoting railway use, and the result variable is the sales amount of high-priced railway services. It is characterized by:

[0017] An information processing method according to a fourth aspect of the present disclosure is characterized in that, for each of a plurality of processes, using teacher data including a first causal variable that changes over time, a second causal variable that does not change over time, a process type, and historical information indicating a history of the time change of the first causal variable, and using a learning model generated by learning the time change of the first causal variable at a next time from the first causal variable, the second causal variable, and the process type at a certain time according to the historical information, for each of the plurality of processes, an estimation of the time change of the first causal variable obtained by a combination of two or more processes selected from the plurality of processes during a certain period of time is performed, and the estimation result is output. Effect of the Invention

[0018] According to one or more aspects of the present disclosure, the processing effects of multiple processes can be learned. [Brief description of the drawings]

[0019] [Figure 1]1 is a graph showing the difference in infection rate between vaccinated and unvaccinated cases as an example of treatment efficacy. [Diagram 2] 1 is a block diagram illustrating a schematic configuration of an estimation device according to a first embodiment. [Diagram 3] 13 is a table showing an example of categorical variable processing of treatment variables. [Figure 4] FIG. 2 is a schematic diagram for explaining an overview of a learning model in the first embodiment. [Diagram 5] FIG. 1 is a schematic diagram for explaining an overview of a multitask Gaussian process. [Figure 6] 4 is a table showing an example of a list of parameters, their initial values, and the number of parameters according to the first embodiment. [Figure 7] FIG. 2 is a block diagram showing an example of a hardware configuration. [Figure 8] 4 is a flowchart showing a processing operation of the estimation device according to the first embodiment. [Figure 9] 1 is a graph showing the difference in infection rates with and without multiple vaccinations as an example of treatment effect over multiple time points. [Figure 10] FIG. 11 is a block diagram illustrating a schematic configuration of an estimation device according to a second embodiment. [Figure 11] 11(A) to 11(D) are schematic diagrams illustrating features of an estimation device according to a second embodiment. [Figure 12] 13(A) and 13(B) are schematic diagrams for explaining an overview of a learning model in the second embodiment. [Figure 13] FIG. 1 is a schematic diagram for explaining an overview of a deep multitask Gaussian process. [Figure 14] FIG. 13 is a schematic diagram for explaining input and output of a deep multitasking Gaussian process unit. [Figure 15] 13 is a flowchart showing a processing operation of an encoder for estimating individual treatment effects at multiple time points in the second embodiment. [Figure 16] 13 is a flowchart showing the processing operation of a decoder unit for estimating individual treatment effects at multiple time points in the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] Embodiment 1 In the first embodiment, an example will be described in which the effect of individual processing at a single time point is estimated using a multitask Gaussian process. Here, we estimate the single-point individual treatment effect (ITE) as an example of a single-point individual treatment effect. In the following, treatment is an example of a process, and dosage is an example of a quantity.

[0021] Individual treatment effectiveness is defined as the difference between the outcome for an individual when they receive treatment and when they do not receive treatment. The magnitude of the treatment effect varies depending on the type and amount of treatment, as well as the attributes of the individual being treated.

[0022] For example, the difference in infection rate between those who receive a vaccine and those who do not receive a vaccine represents the treatment effect, as shown in Figure 1. This treatment effect is thought to vary depending on the attributes of the vaccinated individual (e.g., gender, age, and the presence or absence of chronic diseases), the type of vaccine, the amount of vaccine, and the environment at the time of vaccination.

[0023] Note that the present disclosure is not limited to individual treatment effects, and by making the attributes of individual subjects into attributes of a group, it is also possible to obtain the average treatment effect (ATE) for each group. Similarly, the present disclosure may be applied to the Conditional Average Treatment Effect (CATE) or the Local Average Treatment Effect (LATE). In addition, the "individual" in "individual treatment effect" can be referred to as an individual or individual, and the "treatment" can be referred to as intervention, treatment, policy, or exposure. Furthermore, the treatment effect may be called a causal effect.

[0024] FIG. 2 is a block diagram illustrating a schematic configuration of an estimation device 100 serving as an information processing device according to the first embodiment. The estimation device 100 includes an input unit 110 , a learning unit 120 , an estimation unit 140 , and an output unit 150 .

[0025] The input unit 110 inputs teacher data to the learning unit 120 . For example, the input unit 110 generates training data including a cause variable, a process type, and a result variable for each of a plurality of processes. The input unit 110 includes a DB (Date Base) 111 as a data storage unit, an input execution unit 112, and a preprocessing unit 113.

[0026] The DB 111 stores data necessary for processing in the estimation device 100. Here, the DB 111 stores at least a cause variable X#, a treatment type W#, and a result variable Y#.

[0027] The input execution unit 112 acquires the cause variable X#, the action type W#, and the result variable Y# from the DB 111 and provides these to the preprocessing unit 113 .

[0028] The cause variable X#, treatment type W#, and result variable Y# are data observed for each individual i, and are expressed by the following equation (1).

number

[0029] X i represents a set J1 of causal variables as shown in the following equation (2), where Q indicates the number of causal variables.

number

[0030] Furthermore, each element of the causal variable may be multidimensional information as shown in the following formula (3), and the number of dimensions of each element may be different.

number

[0031] Cause variable x q is expressed by the following equations (4) to (7).

number

number

number

number

[0032] Here, the feature amount of the treatment target is, for example, sex, age, weight, and the like. A treatment amount is, for example, a vaccine dose. The treatment history information includes, for example, the type of vaccine administered in the past, the number of times, and the effect at that time. The environmental information at the time of treatment is a factor that affects the effect of the treatment, such as the weather, the region, and the economic situation.

[0033] Treatment type W# is a categorical variable that represents the type of treatment, as shown in the following formula (8): For example, the categorical variable is the type of vaccine.

number

[0034] The outcome variable Y# represents the treatment effect, e.g., the reduction in infection rate due to vaccination.

[0035] The preprocessing unit 113 performs preprocessing, described below, on the cause variable X#, action type W#, and result variable Y# from the input execution unit 112, and provides the preprocessed cause variable X, action type W, and result variable Y to the learning unit 120.

[0036] When there is a missing value in at least one of the treatment type W# and the result variable Y#, the preprocessing unit 113 extracts the missing data {X i ,W i ,Y i} is removed.

[0037] For the treatment type W#, when there are multiple combination patterns observed for the treatment as shown in Fig. 3, the preprocessing unit 113 assigns an integer value (1,...,D) to each pattern, where D is the number of patterns.

[0038] The preprocessing unit 113 processes the element x of the cause variable X#. q If there are missing values ​​in x q The mean value of each dimension is used for interpolation.

[0039] The preprocessing unit 113 processes the element x of the cause variable X#. q is normalized so that each dimension has a mean of 0 and a variance of 1.

[0040] The learning unit 120 uses the teacher data provided from the input unit 110 to learn a learning model. For example, the learning unit 120 generates a learning model by learning result variables from cause variables and process types for each of a plurality of processes using teacher data from the input unit 110. The generated learning model is stored in a storage unit (not shown). The learning unit 120 includes a calculation unit 121 and an optimization unit 122 .

[0041] The calculation unit 121 initializes a parameter θ of the Gaussian process, inputs a causal variable X, calculates a variance-covariance matrix K(X, X′) of the Gaussian process, and outputs the variance-covariance matrix K to the optimization unit 122.

[0042] FIG. 4 is a schematic diagram for explaining an overview of the learning model learned by learning unit 120. As shown in FIG. The correlation between series of observed variables in the learning model can be expressed by the linear correlation of the LMC (Linear Model of Coregionalization) kernel of the well-known MTGP (Multitask Gaussian Process). For more information on MTGP, see Reference 1 below.

[0043] A method for calculating the variance-covariance matrix of the LMC kernel according to this embodiment will be described below. The treatment effect Y shown in equation (9) below can be expressed using LMC as equations (10) and (11) below.

number

number

number

[0044] g in Eq. (10) q (x q ) is the element x of the causal variable q Gaussian process gp(μ(x q ),k q (x q , x q ')) function. However, μ(x q ) is x q represents the average matrix of x q Since is normalized, μ(x q )=0.

[0045] k qrepresents a positive definite kernel. The positive definite kernel to be used can be selected from those suitable for the data, such as the RBF (Radial Basis Function) kernel. For example, the RBF kernel is expressed by the following equation (12). However, 1 q and l 2 q are the parameters of the RBF kernel.

number

[0046] If the function f shown in the following equation (13) follows a Gaussian process, the function f is expressed as a multidimensional Gaussian distribution as shown in the following equation (14).

number

number

number

number

[0047] In equation (14), K(X,X') is a variance-covariance matrix or Gram matrix, which represents the similarity between causal variables X and X'.

[0048] μ(X) represents the average matrix of the cause variable X, but since the cause variable X has been normalized in the preprocessing unit 113, μ(X)=0.

[0049] Each component of the variance-covariance matrix K(X,X') (K(X,X')) d,d’ is f in Eq. (10). d Using (X), the calculation can be made as shown in the following equation (17). In the transformation from the second line to the third line in equation (17), the different causal variables (q ≠ q') are independent, so the transformation is as shown in equation (18) below.

number

number

[0050] K(X,X') is expressed as in the following equations (19) and (20).

number

number

number

[0051] The result variable Y is defined using the observation function f as shown in the following equations (22), (23), and (24).

number

number

number

[0052] K'(X,X') can be calculated using the cause variable X and the parameter θ shown in the following equation (25). The initial value of the parameter θ is shown in FIG. 6. These parameters are optimized by the optimization unit 122 of the learning unit 120.

number

[0053] The optimization unit 122 receives the variance-covariance matrix K'(X,X') calculated by the calculation unit 121, calculates the marginal likelihood, and optimizes the parameters so as to minimize the marginal likelihood.

[0054] The marginal likelihood is calculated as follows: The probability p(Y|X, θ) of observing the outcome variable Y can be obtained by the following equation (26).

number

[0055] Taking the logarithm of both sides of equation (26), we obtain the marginal likelihood log p (Y|X, θ) can be calculated.

number

[0056] To optimize the parameter θ, we use the marginal likelihood log p We can maximize (Y|X, θ). For example, to match the form of a general optimization problem, we can minimize the optimization function E, as shown in the following equation (28) with both sides multiplied by minus.

number

[0057] The optimization unit 122 calculates the optimization function E in equation (28), that is, the marginal likelihood −log p Update θ so that (Y|X, θ) is minimized. During optimization, updating θ results in K θ Since it is also necessary to update '(X,X'), the calculation unit 121 calculates K' θ Calculate (X,X').

[0058] The optimization function E can be optimized using a known technique such as stochastic gradient descent. As described above, the optimization unit 122 can optimize the parameter θ.

[0059] The estimation unit 140 estimates the variance-covariance matrix K′ calculated by the calculation unit 121 using the optimized parameter θ. θ Using (X,X'), new input X * Calculate the treatment effect of

[0060] The new input X shown in the following equation (29) * Output at y * is expressed by the following equation (30).

number

number

[0061] The output y * is the variance-covariance matrix K' θ From (X, X') and the observation point {X, Y}, for example, it can be calculated using the known technique of Gaussian process regression, using the following equation (31).

number

number

[0062] Here, the treatment effect ITE in the case of no treatment (W=1) and the case of treatment (W=t) can be calculated using the following equation (33).

number

[0063] The estimation unit 140 estimates the result variable by inputting the cause variable and the type of processing to the learning model. The estimated result variable allows the estimation unit 140 to identify an optimal combination of the cause variable and the type of processing. For example, the estimation unit 140 calculates the treatment effect ITE for all combinations of treatments and treatment doses, and identifies the combination of the treatment T with the highest treatment effect and its treatment dose.

[0064] The output unit 150 outputs the result of the estimation performed by the estimation unit 140 . For example, the output unit 150 outputs the object X * For each of the treatments, the output unit 150 outputs the treatment T having the highest treatment effect, the treatment amount, and the treatment effect ITE. For example, the output unit 150 displays them.

[0065] The above-described estimation device 100 can be realized, for example, by a PC (Personal Computer) 10 as shown in FIG. The PC 10 includes an auxiliary storage device 11 such as an HDD (Hard Disk Drive) and an SSD (Solid State Drive), a memory 12, a processor 13 such as a CPU (Central Processing Unit), a communication I / F (Interface) 14 such as a NIC (Network Interface Card), an input I / F 15 such as a keyboard and a mouse, and a display 16.

[0066] For example, the DB 111 can be realized by the auxiliary storage device 11. The input execution unit 112, the preprocessing unit 113, the calculation unit 121, the optimization unit 122 and the estimation unit 140 can be realized by loading a program stored in the auxiliary storage device 11 into the memory 12 and having the processor 13 execute the program. The output unit 150 can be realized by the display 16 .

[0067] The program may be downloaded to the auxiliary storage device 11 from a recording medium (not shown) via a reader / writer (not shown) or from a network via the communication I / F 14, and then loaded onto the memory 12 and executed by the processor 13. Alternatively, the program may be directly loaded onto the memory 12 from a recording medium via the reader / writer or from a network via the communication I / F 14, and executed by the processor 13. In other words, the program may be provided by a program product such as a recording medium. In addition, a storage unit (not shown) that stores the learning model can also be realized by the auxiliary storage device 11.

[0068] FIG. 8 is a flowchart showing the operation of the estimation device 100 according to the first embodiment. First, the input execution unit 112 acquires a cause variable X#, a treatment type W#, and a result variable Y# from the DB 111 (S10).

[0069] Next, the preprocessing unit 113 deletes missing data from the treatment type W# and the result variable Y# (S11). Next, the preprocessing unit 113 performs preprocessing for the procedure type W# (S12). Next, the preprocessing unit 113 processes missing values ​​of the cause variable X# (S13). Next, the preprocessing unit 113 normalizes the cause variable X# (S14). By the above steps S11 to S14, pre-processing is performed on the cause variable X#, the action type W#, and the result variable Y# to obtain the cause variable X, the action type W, and the result variable Y.

[0070] Next, the calculation unit 121 sets the parameter θ shown in the above equation (25) to an initial value (S15). Next, the calculation unit 121 calculates K by using X and θ according to the above formula (19). θ Calculate (X,X') (S16).

[0071] Next, the optimization unit 122 calculates the optimization function E by the above equation (28) (S17). Next, the optimization unit 122 optimizes θ so that the optimization function E is minimized (S18).

[0072] Next, the calculation unit 121 judges whether the optimization function E has converged (S19). The condition for the optimization function E to converge is that the marginal likelihood -log p The value of (Y|X, θ) is minimized, or the processes in steps S16 to S18 are repeated a predetermined number of times. If the optimization function E has not converged (NO in S19), the process returns to step S16, and if the optimization function E has converged (YES in S19), the process proceeds to step S20.

[0073] In step S20, the estimation unit 140 estimates the new input X * Determine the treatment effect ITE in Next, the estimation unit 140 determines the combination of the treatment with the highest treatment effect ITE and the treatment amount (S21).

[0074] The method described in Non-Patent Document 1 was only able to determine the treatment effects of one type of treatment and no treatment, but as described above, the estimation device 100 of embodiment 1 can estimate the treatment effects of multiple types of treatment.

[0075] In addition, in the method of Non-Patent Document 1, only the feature of the treatment target was taken into consideration, but in the estimation device 100 of the first embodiment, in addition to the feature of the treatment target, the treatment history information and the environmental information at the time of the treatment are treated as separate independent causal variables (latent variables). By inputting separate independent causal variables in this way, for example, rather than combining the feature of the target and the feature of the treatment, which should be originally independent, into one feature and calculating the variance-covariance matrix of a Gaussian process, calculating the variance-covariance matrix as separate latent variables can prevent spurious correlation and improve the accuracy of estimation. Here, spurious correlation means that the correlation coefficient between variables that are not originally correlated does not become "0".

[0076] The estimation of individual treatment effects at a single time point in this embodiment makes it possible to know which treatment (measure) should be taken and to what extent in accordance with the characteristics of the subject. For example, it is possible to know which vaccine should be administered in what amount in accordance with the subject to be administered in order to most reduce the infection rate.

[0077] In addition, in the first embodiment, the measurement of the vaccine treatment effect has been described as an example of the treatment effect, but the present invention is not limited to this. For example, if the following variables are set for each variable as shown below, the present embodiment can be applied to the sales promotion of expensive services such as cancer treatment, reserved seats on trains or high-grade hotel rooms, or sales promotion of stores.

[0078] For cancer treatment, the variables are: The feature to be processed may be at least one of the attributes of the individual undergoing cancer treatment (e.g., gender, age, presence or absence of chronic illness, etc.) and physical condition information. The environmental information at the time of treatment may be a factor that affects the effect of cancer treatment. For example, the factor that affects the effect of cancer treatment may be at least one of information on the region and hospital where the treatment is performed. The processing history information may be a history of past cancer treatments, such as the types, number, and effects of cancer treatments received in the past. The types of treatment can be radiation therapy and chemotherapy. The treatment amount can be the radiation intensity and the dose of the anticancer drug. The effect of treatment can be the amount of reduction in cancer size.

[0079] The variables for store sales promotion are as follows: The feature quantity to be processed may be at least one of the attributes of the purchaser of the product (for example, gender, age, annual income, preferences, etc.) and the purchase history. The environmental information at the time of processing may be factors that affect sales at a store. For example, factors that affect the purchase of a product may be at least one of product trends, weather, and economic conditions. The transaction history information may be a purchase history from a past transaction, for example, the purchase history from a past transaction may include the type of product purchased in the past, the number of purchases, and the purchase date and time. The types of transactions can be discount coupons, internet advertisements, television advertisements and travel vouchers. The transaction amount can be the discount amount, the advertisement frequency and the probability of winning. The effect of the treatment may be sales at the store.

[0080] The variables for the promotion of high-priced (high-value) services are: Here, high-priced services refer to reserved seats on trains or buses, high-class seats on airplanes, high-class cabins on ships, reserved seats at sports or entertainment events, priority tickets to tourist facilities, or high-class guest rooms in hotels.

[0081] In the case of seats or tickets, this also includes priority rights such as guaranteed seating, being able to ride attractions without lining up on the day, and time reservation. In particular, in the case of railway services, expensive services include not only regular reserved seats, but also reserved seats with good views, luxurious sleeper cars, and reserved seats on special trains.

[0082] The feature quantity to be processed may be at least one of the attributes of the user of the service (for example, gender, age, annual income, employment status, place of residence, place of work, preferences, family structure, etc.) and purchase history. The environmental information at the time of processing can be a factor that affects the sales amount of the service. For example, the factor that affects the sales amount of the service can be at least one of the reservation rate, the congestion rate, the weather, an accident, a sales plan, and the prevalence of an infectious disease. In particular, in the case of railroad services, the weather includes disasters such as typhoons, earthquakes, and heavy rain. Accidents also include vehicle breakdowns, overhead line failures, personal injury accidents, and fires along the tracks. Sales plans also include planned service suspensions and increased number of special services due to power outages, construction work, strikes, and consecutive holidays. The processing history information can be a behavior history during past processing. The behavior history is a history of actions taken by a user of a service. For example, the behavior history includes a history of a user purchasing a ticket (in other words, a purchase history), a history of a user applying for a campaign, a history of a user exercising a right, a history of a user using a coupon, and the like. The types of the process can be a price increase, a price decrease, a campaign, and a coupon distribution. The processing amount can be the price increase amount, the price reduction amount, and the coupon amount. Note that the price increase amount and the price reduction amount may be the difference between the price increase (price increase rate) and the price reduction amount (price reduction rate), respectively. The processing effect is the sales amount of high-priced (high-value) services.

[0083] From the above, when the first embodiment is applied to railway services, the following occurs. The causal variables are at least one of the attributes and purchasing history of railway users, factors influencing the sales amount of railway services, the behavioral history of railway users, the amount of fare increase, the amount of fare decrease, and the amount of coupons to promote railway use. The types of processing include a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways. The outcome variable is the amount of sales of the higher-priced railroad's services.

[0084] Then, based on the obtained processing effect, for example, the following first to third operations can be performed. First operation: In order to increase the reservation rate of reserved seats on trains, the price of reserved seats can be reduced according to the length of travel when the train is crowded so that long-term users can stay in their seats, and conversely, the price of reserved seats can be reduced for short-term users when the train is empty. Second action: You can run a promotion offering a free upgrade to train or airplane seats to encourage customers to make a higher-grade reservation next time. Third operation: Services in a price range that matches the user's characteristics can be recommended. In addition, coupons can be distributed, etc., to induce purchases of higher-priced services.

[0085] Embodiment 2 In the second embodiment, the deep multitasking Gaussian process is used to estimate the individual processing effects at multiple points in time. Specifically, in the second embodiment, a case will be described in which the individual processing effects are estimated when processing is performed at multiple points in time with different times. In the second embodiment, for example, as shown in FIG. 9, a case where four vaccinations are planned to reduce the infection rate of a virus will be described. Specifically, in the second embodiment, when two vaccinations have already been administered and two more vaccinations are planned, it is estimated which vaccine should be administered two times to reduce the infection rate the most. Note that the treatment effect of four vaccinations to be administered in the future may be estimated in a state where no vaccinations have been administered even once.

[0086] FIG. 10 is a block diagram illustrating a schematic configuration of an estimation device 200 serving as an information processing device according to the second embodiment. The estimation device 200 includes an input unit 210 , an encoder training unit 220 , a decoder training unit 230 , an estimation unit 240 , and an output unit 250 .

[0087] The input unit 210 inputs training data to the encoder learning unit 220 . For example, the input unit 210 generates teacher data for each of a plurality of processes, the teacher data including a first cause variable that changes over time, a second cause variable that does not change over time, a process type, and historical information indicating the history of the change over time of the first cause variable. The input unit 110 includes a DB 211 as a data storage unit, an input execution unit 212, and a preprocessing unit 213.

[0088] The DB 211 stores data necessary for processing in the estimation device 200. Here, the DB 211 stores at least a feature amount X#, a feature amount V#, and a treatment type W#. Here, the feature amount X# is a feature amount of the treatment target that varies depending on the treatment, and the feature amount V# is a feature amount of the treatment target that does not vary depending on the treatment.

[0089] The input execution unit 212 acquires the feature amount X#, the feature amount V#, and the treatment type W# from the DB 211 , and provides these to the pre-processing unit 213 .

[0090] The feature amount X#, the feature amount V#, and the treatment type W# are data observed for each individual i, and are expressed by the following equation (34).

number

[0091] The pre-processing unit 213 performs pre-processing, described below, on the feature X#, feature V#, and action type W# from the input execution unit 212, and provides the feature X, feature V, and action type W after pre-processing to the encoder learning unit 220.

[0092] When there are missing values ​​in the feature amount X#, the feature amount V#, and the treatment type W# of the individual i, the preprocessing unit 213 removes the data J2 of the individual i, which is shown in the following equation (35).

number

[0093] Furthermore, the pre-processing unit 213 pre-processes the procedure type W# in the same manner as the pre-processing unit 113 of the first embodiment.

[0094] Furthermore, the preprocessing unit 213 normalizes the feature amount X# and the feature amount V# so that the mean and variance are zero and unity, respectively.

[0095] The encoder learning unit 220 and the decoder learning unit 230 function as a learning unit that learns the learning model in the second embodiment. For example, the learning unit in the second embodiment generates a learning model by learning, for each of a plurality of processes, a time change of the first cause variable at the next time from the first cause variable, the second cause variable, and the process type at a certain time according to the history information, using the teacher data from the input unit 210. The learned learning model is stored in a storage unit (not shown).

[0096] 11(A) to 11(D) are schematic diagrams showing the operation concept of the learning model in the second embodiment. In FIG. 11, a time-varying cause variable X, a time-invariant cause variable V, a type of action W, and history information H are shown. Here, X is a factor that changes depending on the treatment, such as the infection rate, and is both a causal variable and a response variable. For example, at step s, treatment W s is performed, and X s X s+1 If it changes to X s are causal variables, and X s+1 is the objective variable.

[0097] As shown by the dashed line in FIG. s and the result after treatment X s+1 Affects both X s These are called confounding factors. If there are confounding factors, X s From W sDue to the effect of the effect of the treatment with a low probability of treatment, the observed data for the treatment with a high treatment probability increases. Therefore, if the model is trained using the data as is, a bias occurs in which the model is optimized for the treatment with a high probability, and when predicting the effect of a treatment with a low treatment probability, the prediction accuracy decreases, making it difficult to estimate the correct treatment effect. Therefore, in the second embodiment, the influence of confounding factors is reduced by using expression learning as shown in the fourth feature below.

[0098] First, the learning model will be described. 12(A) and (B) are schematic diagrams for explaining a learning model in the second embodiment. The learning model in the second embodiment is a combination of a known technique, Encoder-Decoder (or Seq2Seq), and a known technique, Deep Multitask Gaussian Process.

[0099] First, the features of the learning model in the second embodiment will be described. First feature: When considering measures at multiple points in time, the question arises as to what extent past influences should be considered. As shown in FIG. 11(A), the more past information is included, the more complex the causal graph becomes, and the more factors must be considered. Therefore, in the second embodiment, for example, by introducing the Markov property as shown in FIG. 11(B), the information up to the previous time is treated as history information H s-1 The learning model can be simplified as follows:

[0100] Second feature: As shown in FIG. 11(B) and FIG. 13, the history H s For the calculation of, for example, an intermediate representation (Hidden Variable) calculated by a deep multitasking Gaussian process, which is a known technique, can be used.

[0101] Third feature: As shown in Fig. 11(C) and Fig. 13, the second layer of the deep multi-task Gaussian process is branched, and one of them is a multi-task Gaussian process H sFrom Hs, the cause variable J3 shown in the following equation (36) is predicted, and the treatment type J4 shown in the following equation (37) is predicted from Hs using a Gaussian process classification model, which is a known technique.

number

number

[0102] Fourth feature: Applying representation learning techniques based on known technologies, s While improving the prediction accuracy of the causal variable J3 shown in the above equation (36), H s Therefore, it is difficult to predict the treatment type J4 shown in the above equation (37). s Calculate f h By learning H s From W s This can reduce the influence of confounding factors on

[0103] Fifth feature: As shown in Fig. 11(D) and Fig. 12, the intermediate representation of the deep multitask Gaussian process H s By recursively transmitting the next step (s+1), long-term prediction is possible, and any multiple treatment plans W t:t+τ This allows us to estimate the treatment effect in

[0104] Sixth feature: As described above, the estimation device 200 according to the second embodiment estimates the processing effect of multiple processes using a deep multitasking Gaussian process. Therefore, in addition to the effects described in the first embodiment, the reliability of the estimated processing effect can also be known.

[0105] The deep multi-task Gaussian process and the learning model of the Encoder-Decoder according to the second embodiment will be described below.

[0106] First, we explain the deep multitask Gaussian process. The following equations (38) to (43) show relational expressions of the deep multitasking Gaussian process. Equation (38) is the Gaussian process relation in the first layer MTGP, and equation (39) is the likelihood function in the first layer MTGP. Equation (40) is the Gaussian process relation in the second layer MTGP, and equation (41) is the likelihood function in the second layer MTGP. Equation (42) is a relational expression of the Gaussian process in the GP Classification layer, and equation (43) is a likelihood function in the GP Classification layer.

number

number

number

number

number

number

[0107] Here, X' s is expressed by the following equation (44).

number

[0108] Furthermore, because of the Gaussian process, in the above equations (40) and (41), or (42) and (43), in addition to the average values ​​J5 and J6 of the estimated values ​​shown in the following equations (45) and (46), the respective standard deviations J7 and J8 shown in the following equations (47) and (48) can be obtained.

number

number

number

number

[0109] In the above formulas (38) to (43), f is a function of the Gaussian process, and μ, sm, and N are mean functions, and represent a softmax function and a Gaussian distribution, respectively. θ is a parameter of the Gaussian process, and σ is a variance value of the Gaussian distribution. Note that none of the parameters have time dependency and are the same at all times.

[0110] X' s Each of the causal variables constituting the above equations may be treated as an independent causal variable as in the first embodiment, or each causal variable (latent variable) may be divided into a time-invariant element and a time-variant element, and may be, for example, the element Xs and the element V shown in the following equations (49) and (50).

number

number

[0111] The prediction target J9 shown in the following formula (51), which is included in the above formula (41), may be set to predict only factors that change due to treatment (e.g., infection rate, etc.). For example, only the factor J10 shown in the following formula (52) may be set as the prediction target.

number

number

[0112] Next, the Gaussian process will be described. The matrix K1 shown in the following equation (53) represents the variance-covariance matrix of a normal one-output Gaussian process. The matrices K2 and K3 shown in the following equations (54) and (55) represent the variance-covariance matrices of the multitask Gaussian process.

number

number

number

[0113] Here, as shown in FIG. 14, the output dimension number D of the variance-covariance matrix K2 of the multitask Gaussian process shown in the above equation (54) is H The number of dimensions is set to 1 or more, and the output dimension number N of the variance-covariance matrix K3 shown in equation (55) is task is set to the number of categories in W.

[0114] The kernel for calculating the variance-covariance matrices K2 and K3 shown in the above formulas (54) and (55) may be an Independent Multi-task Kernel that has no correlation between output dimensions, or a Dependent Multi-task Kernel that has correlation between output dimensions. In the Dependent Multi-task Kernel, for example, an LMC kernel that provides linear correlation between output dimensions may be used.

[0115] Calculating a Gaussian process as is takes O(N 3 ), the amount of calculation may be reduced by using Stochastic Variational Inference for Gauss Process (SVI-GP) using inducing points, which is a known technique.

[0116] Next, the likelihood function will be described. Prediction distribution of the output of the first-layer MTGP, f h (Xs Since the second-layer Gaussian process cannot be directly used as the input, the predictive distribution f h (X t ') to calculate the intermediate representation J11 shown in the following equation (56), or the predictive distribution f h (X s ') to obtain the intermediate representation J11 shown in equation (56) below.

number

[0117] The second layer MTGP generates one task (e.g., processing W) for each input data. s ) is observed, as shown in Figure 14, the following equation (57) is used for the output of MTGP J12, W s The index matrix I in one-hot format tasks (W s ) so that only the treated types are output.

number

[0118] The second layer, the Classification Layer, uses the softmax function as the likelihood function, as shown in equation (48) above, to determine the treatment of categorical variables W s Make it predictable.

[0119] Next, the Encoder-Decoder model will be described. Using the known Encoder-Decoder technique, we obtain an intermediate representation H of the deep multitask Gaussian process. sis transmitted to the next step. As shown in FIG. 12, the Encoder unit can perform one-step ahead prediction, and the Decoder unit can perform multiple-step ahead prediction. The Encoder-Decoder model of the second embodiment is not limited to the example shown in FIG. 12, and for example, the Encoder-Decoder may be provided with an Attention mechanism or may be implemented using a Transformer.

[0120] The structure of the deep multi-task Gaussian process in the encoder and decoder is the same, but the parameters are refined separately.

[0121] Below, the operation of each unit will be described in detail based on the above-mentioned deep multi-task Gaussian process and Encoder-Decoder model.

[0122] The encoder learning unit 220 in FIG. 10 performs learning on the encoder side in an encoder-decoder model. The encoder learning unit 220 in FIG. 10 includes an initialization unit 221, a prediction unit 222, a classification optimization unit 223, and a prediction optimization unit 224.

[0123] The initialization unit 221 initializes the number of dimensions D of H shown in FIG. H Set the number of dimensions to 1 or more. Moreover, the initialization unit 221 initializes a parameter J13 of the deep multitasking Gaussian process of Encorder, which is shown in the following equation (58). The initialized value may be, for example, 1, or may be a sampling from an arbitrary Gaussian distribution.

number

[0124] 12A, first, in the first step (e.g., s=1) in Encoder, the prediction unit 222 inputs H0, which is a zero vector, and {X1, V, W1}, and calculates data J14 shown in the following formula (59). After that, the prediction unit 222 recursively inputs data J15 shown in the following formula (60) and data J16 shown in the following formula (61) up to step t (t is an integer equal to or greater than 2), thereby calculating data J17 shown in the following formula (62).

number

number

number

number

[0125] Then, the prediction unit 222 calculates data J18 shown in the following equation (63) for all individuals (i) and provides the data J18 to the classification optimization unit 223.

number

[0126] The classification optimization unit 223 receives data J19 calculated by the prediction unit 222 and shown in the following equation (64), and calculates data J20 shown in the following equation (65) from the above equations (42) and (43).

number

number

[0127] Then, the classification optimization unit 223 optimizes a parameter J21 shown in the following equation (67) by minimizing the loss function shown in the following equation (66) using SVI-GP, a known technique.

number

number

[0128] The prediction optimization unit 224 receives as input the data J22 calculated by the prediction unit 222 shown in (68) below and the observed data J23 shown in (69) below, and calculates the data J24 shown in the following equation (70) using the above equations (38) and (39).

number

number

number

[0129] The prediction optimization unit 224 receives data J25 shown in the following equation (71) as an input, and calculates data J26 shown in the following equation (72) using the above equations (40) to (43).

number

number

[0130] In order to reduce the influence of confounding factors, the prediction optimization unit 224 optimizes a parameter J30 shown in the following formula (77) using the known technology DSVI-DGP with a loss function shown in the following formula (76) in order to increase the prediction accuracy of a treatment effect J29 shown in the following formula (75) from information J27 while preventing a treatment J28 shown in the following formula (74) from being predicted from information J27 shown in the following formula (73). However, a parameter J31 shown in the following formula (78) is fixed.

number

number

number

number

number

number

number

[0131] The prediction optimization unit 224 returns to the processing in the prediction unit 222 and repeats the processing in the classification optimization unit 223 and the prediction optimization unit 224 until the value J33 shown in the following equation (80) and the value J34 shown in the following equation (81) converge, thereby optimizing the parameter J35 shown in the following equation (82).

number

number

number

[0132] When the value J33 shown in the above equation (80) and the value J34 shown in the above equation (81) converge, the prediction optimization unit 224 provides the decoder learning unit 230 with the data J18 calculated in the above equation (63) using the optimized parameters J35 shown in the above equation (82) and the parameters optimized in the prediction unit 222.

[0133] The decoder learning unit 230 in FIG. 10 performs learning on the decoder side in an encoder-decoder model. The decoder learning unit 230 in FIG. 10 includes an initialization unit 231, a prediction unit 232, a classification optimization unit 233, and a prediction optimization unit 234.

[0134] The initialization unit 231 initializes the parameters J36 of the decoder unit, shown in the following equation (83), with the parameters J35 optimized in the encoder unit, shown in the above equation (82).

number

[0135] Prediction unit 232 performs the following process of predicting variable J37 τ ahead of step t (τ-step ahead prediction), which is shown in the following equation (84).

number

[0136] As shown in FIG. 12B, the prediction unit 232 compares the data J38 last predicted by the Encoder and expressed by the following equation (85), with the observed data {V, W (t+1)} is input, and the data J39 shown in the following equation (86) is calculated.

number

number

[0137] Next, the prediction unit 232 recursively inputs data J40 shown in the following equation (87) and predicted data J41 shown in the following equation (88) up to step t+τ−1 to calculate data J42 shown in the following equation (89) and obtains data J43 shown in the following equation (90).

number

number

number

number

[0138] The prediction unit 232 calculates data J44 represented by the following equation (91) for all individuals (i) and provides the data J44 to the classification optimization unit 233. max represents the maximum number of prediction steps.

number

[0139] The classification optimization unit 233 receives data J45 calculated by the prediction unit 232 and expressed by the following equation (92), and calculates data J46 expressed by the following equation (93) from the above equations (42) and (43).

number

number

[0140] The classification optimization unit 233 optimizes the parameter J47 shown in the following equation (95) by minimizing the loss function shown in the following equation (94) using the known technique SVI-GP. Note that, during optimization, τ=1 to τ max Optimization is performed by calculating the sum of the losses and performing back propagation.

number

number

[0141] The prediction optimization unit 234 uses data J48 shown in the following equation (96) as an initial input, and calculates data J50 shown in the following equation (98) by recursively inputting data J49 shown in the following equation (97) up to s=t+τ−1, as shown in FIG. 12(B).

number

number

number

[0142] Then, the prediction optimization unit 234 optimizes the parameter J51 shown in the following formula (100) with the loss function shown in the following formula (99) using the DSVI-DGP, which is a known technique. Note that the parameter J52 shown in the following formula (101) is fixed.

number

number

number

number

[0143] The prediction optimization unit 234 returns to the processing in the prediction unit 232 and repeats the processing in the classification optimization unit 233 and the prediction optimization unit 234 until the value J54 shown in the following equation (103) and the value J55 shown in the following equation (104) converge, thereby optimizing the parameter J36 shown in the above equation (83).

number

number

[0144] When the value J54 shown in the above formula (103) and the value J55 shown in the above formula (104) converge, the prediction optimization unit 234 provides the optimized parameters J36 shown in the above formula (83) and the optimized parameters J35 of the Encoder shown in the above formula (82) to the estimation unit 240.

[0145] The optimization of parameter J35 of encoder learning unit 220 shown in equation (82) and the optimization of parameter J36 of decoder learning unit 230 shown in equation (83) above may use a method other than the method shown above. For example, training may be performed using the well-known technique of adversarial training. In that case, the Discriminator (f a ) learns to predict the action J57 shown in the following equation (106) from the intermediate representation J56 shown in the following equation (105), and the Generator (f h ) improves the prediction accuracy of the treatment effect J58 shown in the following equation (107) from the intermediate representation J56 shown in the following equation (105), while a) may learn to be unable to predict action J57 shown in equation (106) below from intermediate representation J56 shown in equation (105) below.

number

number

number

[0146] The estimation unit 240 uses the learning model to estimate a time change of a first causal variable obtained by a combination of two or more processes selected from a plurality of processes during a certain period of time. This allows the estimation unit 240 to identify an optimal combination of two or more processes based on the estimated time change.

[0147] For example, the estimation unit 240 calculates a treatment effect J60 of a treatment plan J59 from an arbitrary time t to a time τ ahead, which is shown in the following equations (108) and (109), and a variance value J61, which is shown in the following equation (110), from the optimized Encoder parameters J35 shown in the above equation (82) and the optimized Decoder parameters J36 shown in the above equation (83).

number

number

number

[0148] The estimation unit 240 selects the treatment plan J59 shown in the above formula (108) that maximizes the final treatment effect J62 shown in the following formula (111).

number

[0149] Then, the estimation unit 240 provides the optimal treatment plan J59 shown in the above equations (108), (109), and (110), its treatment effect J62, and its variance value J63 shown in the following equation (112) to the output unit 250.

number

[0150] When there are many combinations of the treatment plan J59 shown in the above equation (108) or when τ is a long period, the treatment plan may be searched for by Monte Carlo sampling or by reinforcement learning that learns the combination that has the highest final treatment effect.

[0151] The treatment plan J59 shown in the above equations (108) and (111) may be determined to be optimal when it has the largest final treatment effect J62 or the largest cumulative treatment effect J64 shown in the following equation (113).

number

[0152] In addition, the estimation unit 240 may create a treatment plan that will achieve the target treatment effect most quickly, a treatment plan that will improve the situation to the best on the last day of the period, or a sustainable treatment plan that will maintain the current situation so as not to exceed the indicator condition.

[0153] The output unit 250 outputs the result of the estimation performed by the estimation unit 240 . For example, the output unit 250 outputs the treatment plan, the treatment effect, and the variance value from the estimation unit 240. Specifically, the output unit 250 displays the treatment plan, the treatment effect, and the variance value.

[0154] The above-described estimation device 200 can be realized, for example, by a PC 10 as shown in FIG. For example, the DB 211 can be realized by the auxiliary storage device 11. The input execution unit 112, the pre-processing unit 113, the encoder learning unit 220, the decoder learning unit 230 and the estimation unit 140 can be realized by loading a program stored in the auxiliary storage device 11 into the memory 12 and having the processor 13 execute the program. The output unit 250 can be realized by the display 16 .

[0155] 15 and 16 are flowcharts showing the operations of the estimation device 200 according to the second embodiment described above. First, the input execution unit 212 obtains {X#, V#, W#} from the DB 211 and provides them to the preprocessing unit 213 (S30).

[0156] Next, the preprocessing unit 213 deletes data having missing values ​​in X#, V#, and W# (S31).

[0157] Next, the preprocessing unit 213 performs preprocessing of W# (S32).

[0158] Next, the preprocessing unit 213 normalizes X# (S33).

[0159] Next, the pre-processing unit 213 provides the {X, V, W} that has been subjected to the above processing to the encoder learning unit 220 (S34).

[0160] Next, the initialization unit 221 of the encoder learning unit 220 initializes the number of dimensions D H Set (S35).

[0161] Next, the initialization unit 221 initializes the parameter J13 shown in the above equation (58) (S36).

[0162] Next, the prediction unit 222 calculates data J65 shown in the following equation (114) (S37).

number

[0163] Next, the classification optimization unit 223 performs calculation of data J66 shown in the following equation (115) (S38).

number

[0164] Next, the classification optimization unit 223 optimizes the parameter J67 shown in the following equation (116) (S39).

number

[0165] Next, the prediction optimization unit 224 calculates data J68 shown in the following equation (117) (S40).

number

[0166] Next, the prediction optimization unit 224 calculates data J69 and data J70 shown in the following equations (118) and (119) (S41).

number

number

[0167] Next, the prediction optimization unit 234 optimizes the parameter J30 shown in the above equation (77) (S42). Here, the prediction optimization unit 224 returns the process to step S37 and repeats each process until the loss functions, value J33 shown in the above equation (80) and value J34 shown in the above equation (81), are minimized or a predetermined number of repetitions is reached.

[0168] Next, the prediction optimization unit 234 provides the optimized parameter J35 shown in the above equation (82) to the decoder learning unit 230 (S43). Then, the process proceeds to step S44 in FIG.

[0169] In step S44 of FIG. 16, the initialization unit 231 of the decoder learning unit 230 initializes the parameter J36 of the Decoder unit shown in the above equation (83) with the parameter J35 optimized in the Encoder unit shown in the above equation (82).

[0170] Next, the prediction unit 232 calculates data J71 shown in the following equation (120) (S45).

number

[0171] Next, the classification optimization unit 233 calculates data J72 shown in the following equation (121) (S46).

number

[0172] Next, the classification optimization unit 233 optimizes a parameter J73 shown in the following equation (122) (S47).

number

[0173] Next, the prediction optimization unit 234 calculates data J74 and data J75 shown in the following equations (123) and (124) (S48).

number

number

[0174] Next, the prediction optimization unit 234 optimizes the parameter J51 shown in the above equation (100) (S49). Here, the prediction optimization unit 234 returns the process to step S45 and repeats each process until the loss functions, value J54 shown in the above equation (103) and value J55 shown in the above equation (104), are minimized or a predetermined number of repetitions is reached.

[0175] Next, the prediction optimization unit 234 provides the parameters J35 optimized by the encoder learning unit 220, shown in the above equation (82), and the parameters J36 optimized by the decoder learning unit 230, shown in the above equation (83), to the estimation unit 240 (S50).

[0176] Next, the estimation unit 240 calculates the treatment effect J60 of the treatment plan J59 from an arbitrary time t to a time τ ahead, which are shown in the above equations (108) and (109), and the variance value J61, which is shown in the above equation (110), from the optimized Encoder parameters J35 shown in the above equation (82) and the optimized Decoder parameters J36 shown in the above equation (83) (S51).

[0177] Next, the estimation unit 240 selects the treatment plan J59 shown in the above formula (108) that maximizes the final treatment effect J62 shown in the above formula (111) (S52).

[0178] Next, the estimation unit 240 provides the optimal treatment plan J59, its treatment effect J62, and its variance value J61, which are shown in the above equations (108), (109), and (110), to the output unit 250.

[0179] Next, the output unit 250 displays the treatment plan, treatment effect, and variance value from the estimation unit 240 (S54).

[0180] Here, when the second embodiment is applied to railway services, the following occurs. The first causal variable is the amount of the fare increase, the amount of the fare decrease, and the amount of coupons to encourage rail travel. The second causal variable is at least one of the attributes and purchase history of railway users, factors that affect the sales amount of railway services, and the behavioral history of railway users. The types of processing include a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways. The history information includes a history of fare increases, a history of fare decreases, and a history of coupon distribution to promote rail travel.

[0181] As described above, according to the second embodiment, it becomes possible to know when and what policy plan should be made in accordance with the characteristics of the target. In addition, it becomes possible to know the reliability of the estimated value of the treatment effect estimated in the policy plan. For example, as shown in Figure 9, when administering a vaccine four times, it becomes possible to create an administration plan including the combination of vaccines to be administered (multiple combinations of treatments), the order of treatments, the timing of treatment, and when to stop treatment.

[0182] In addition, for example, in dynamic pricing of reserved train seats or hotel rooms, it becomes possible to know what timing and from what period to increase or decrease the price of seats or rooms, the number or combination of price increases and reductions, and the order in which price increases and reductions should be performed in order to increase final sales.

[0183] Reference 1: Bica, I., Alaa, AM, Jordon, J., and van der Schaar, M., “Estimating counterfactual treatment outcomes over time through adversarially balanced representations”, International Conference on Learning Representationsm 2020. [Explanation of symbols]

[0184] 100,200 estimation device, 110,210 input unit, 111,211 DB, 112,212 input execution unit, 113,213 preprocessing unit, 120 learning unit, 121 calculation unit, 122 optimization unit, 220 encoder learning unit, 221 initialization unit, 222 prediction unit, 223 classification optimization unit, 224 prediction optimization unit, 230 decoder learning unit, 231 initialization unit, 232 prediction unit, 233 classification optimization unit, 234 prediction optimization unit, 140,240 estimation unit, 150,250 output unit.

Claims

1. an input unit that generates training data including a cause variable, a process type, and a result variable for each of a plurality of processes; a learning unit that uses the teacher data to learn the result variable from the cause variable and the process type for each of the plurality of processes, thereby generating a learning model; The causal variables are at least one of attributes and purchase history of railway users, factors that affect the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The outcome variable is the sales amount of high-priced rail services. An information processing device comprising:

2. The method further includes an estimation unit that estimates the result variable by inputting the cause variable and the type of processing into the learning model.

2. The information processing apparatus according to claim 1,

3. The estimation unit specifies an optimal combination of the cause variable and the processing type based on the estimated result variable.

3. The information processing apparatus according to claim 2,

4. an input unit that generates teacher data including, for each of a plurality of processes, a first cause variable that changes with time, a second cause variable that does not change with time, a process type, and history information indicating a history of a time change of the first cause variable; a learning unit that uses the teacher data to learn a time change of the first cause variable at a next time from the first cause variable, the second cause variable, and the process type at a certain time according to the history information for each of the plurality of processes, thereby generating a learning model. An information processing device comprising:

5. The method further includes an estimation unit that estimates a time change of the first causal variable obtained by a combination of two or more processes selected from the plurality of processes during a certain period of time, using the learning model.

5. The information processing apparatus according to claim 4,

6. The estimation unit specifies an optimal combination of the two or more processes based on the estimated time change.

6. The information processing device according to claim 5,

7. The first causal variable is the amount of the fare increase, the amount of the fare decrease, and the amount of the coupon to promote the use of the railway; The second causal variable is at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, and behavioral history of railway users; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The history information includes a history of fare increases, a history of fare reductions, and a history of coupon distribution to promote railroad use.

7. The information processing device according to claim 4, wherein:

8. an estimation unit that estimates the result variable by inputting the cause variable and the process type into a learning model generated by learning the result variable from the cause variable and the process type for each of a plurality of processes using teacher data including the cause variable, the process type, and the result variable for each of the plurality of processes; an output unit that outputs a result of the estimation, The causal variables are at least one of attributes and purchase history of railway users, factors that affect the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The outcome variable is the sales amount of high-priced rail services. An information processing device comprising:

9. an estimation unit that uses teacher data including, for each of a plurality of processes, a first cause variable that changes with time, a second cause variable that does not change with time, a process type, and history information indicating a history of a time change of the first cause variable, and estimates a time change of the first cause variable obtained by a combination of two or more processes selected from the plurality of processes during a certain period of time using a learning model generated by learning a time change of the first cause variable at a next time from the first cause variable, the second cause variable, and the process type at a certain time in accordance with the history information; An output unit that outputs the result of the estimation. An information processing device comprising:

10. The first causal variable is the amount of the fare increase, the amount of the fare decrease, and the amount of the coupon to promote the use of the railway; The second causal variable is at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, and behavioral history of railway users; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The history information includes a history of fare increases, a history of fare reductions, and a history of coupon distribution to promote railroad use. The information processing device according to claim 9 .

11. Computer, an input unit that generates training data including a cause variable, a process type, and a result variable for each of a plurality of processes; a learning unit that generates a learning model by learning the result variable from the cause variable and the process type for each of the plurality of processes using the teacher data; The causal variables are at least one of attributes and purchase history of railway users, factors that affect the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The outcome variable is the sales amount of high-priced rail services. A program characterized by.

12. Computer, an input unit that generates teacher data including, for each of a plurality of processes, a first cause variable that changes over time, a second cause variable that does not change over time, a process type, and history information indicating a history of a time change of the first cause variable; and a learning unit that generates a learning model by using the teacher data and learning a time change of the first cause variable at a next time from the first cause variable, the second cause variable, and the process type at a certain time according to the history information for each of the plurality of processes. A program characterized by.

13. The first causal variable is the amount of the fare increase, the amount of the fare decrease, and the amount of the coupon to promote the use of the railway; The second causal variable is at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, and behavioral history of railway users; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The history information includes a history of fare increases, a history of fare reductions, and a history of coupon distribution to promote railroad use. The program according to claim 12 .

14. Computer, an estimation unit that estimates the result variable by inputting the cause variable and the process type into a learning model generated by learning the result variable from the cause variable and the process type for each of a plurality of processes using teacher data including the cause variable, the process type, and the result variable for each of the plurality of processes; and an output unit that outputs a result of the estimation; The causal variables are at least one of attributes and purchase history of railway users, factors that affect the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The outcome variable is the sales amount of high-priced rail services. A program characterized by.

15. Computer, an estimation unit that uses teacher data including, for each of a plurality of processes, a first cause variable that changes with time, a second cause variable that does not change with time, a process type, and history information indicating a history of a change in the first cause variable with time, and uses a learning model generated by learning a change in the first cause variable with time at a next time from the first cause variable, the second cause variable, and the process type at a certain time in accordance with the history information for each of the plurality of processes, to estimate a change in the first cause variable with time obtained by a combination of two or more processes selected from the plurality of processes during a certain period; and functioning as an output unit that outputs the result of the estimation. A program characterized by.

16. The first causal variable is the amount of the fare increase, the amount of the fare decrease, and the amount of the coupon to promote the use of the railway; The second causal variable is at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, and behavioral history of railway users; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The history information includes a history of fare increases, a history of fare reductions, and a history of coupon distribution to promote railroad use. The program according to claim 15 .

17. generating training data including a cause variable, a process type, and an effect variable for each of a plurality of processes; An information processing method for generating a learning model by learning the result variable from the cause variable and the process type for each of the plurality of processes using the teacher data, comprising: The causal variables are at least one of attributes and purchase history of railway users, factors that affect the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The outcome variable is the sales amount of high-priced rail services. An information processing method comprising:

18. generating teacher data including, for each of a plurality of processes, a first cause variable that changes with time, a second cause variable that does not change with time, a process type, and history information indicating a history of a change in the first cause variable over time; generating a learning model by learning a time change of the first cause variable at a certain time from the first cause variable, the second cause variable, and the process type at a next time according to the history information, for each of the plurality of processes, using the teaching data; An information processing method comprising:

19. The first causal variable is the amount of the fare increase, the amount of the fare decrease, and the amount of the coupon to promote the use of the railway; The second causal variable is at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, and behavioral history of railway users; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The history information includes a history of fare increases, a history of fare reductions, and a history of coupon distribution to promote railroad use.

20. The information processing method according to claim 18, further comprising:

20. using teacher data including a cause variable, a process type, and a result variable for each of a plurality of processes, the cause variable and the process type are input into a learning model generated by learning the result variable from the cause variable and the process type for each of the plurality of processes, thereby estimating the result variable; An information processing method for outputting a result of the estimation, The causal variables are at least one of attributes and purchase history of railway users, factors that affect the sales amount of railway services, behavioral history of railway users, fare increase amount, fare reduction amount, and coupon amount for promoting railway use; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The outcome variable is the sales amount of high-priced rail services. An information processing method comprising:

21. using teacher data including, for each of a plurality of processes, a first cause variable that changes with time, a second cause variable that does not change with time, a process type, and history information indicating a history of the time change of the first cause variable, and using a learning model generated by learning the time change of the first cause variable at a next time from the first cause variable, the second cause variable, and the process type at a certain time in accordance with the history information, for each of the plurality of processes, estimating the time change of the first cause variable obtained by a combination of two or more processes selected from the plurality of processes during a certain period; outputting a result of said estimation; An information processing method comprising:

22. The first causal variable is the amount of the fare increase, the amount of the fare decrease, and the amount of the coupon to promote the use of the railway; The second causal variable is at least one of attributes and purchase history of railway users, factors influencing the sales amount of railway services, and behavioral history of railway users; the processing types are a fare increase, a fare decrease, a campaign to promote the use of railways, and distribution of coupons to promote the use of railways; The history information includes a history of fare increases, a history of fare reductions, and a history of coupon distribution to promote railroad use.

22. The information processing method according to claim 21, further comprising:

Images