Information processing apparatus, program, and information processing method

Abstract

The estimation device (100) is characterized by having: an input unit (110) that generates supervisory data containing causal variables, treatment types and outcome variables for each of the multiple treatments; and a learning unit (120) that uses the supervisory data to learn outcome variables based on causal variables and treatment types for each of the multiple treatments, thereby generating a learning model.

Description

Technical Field

[0001] This disclosure relates to information processing apparatus, programs, and methods. Background Technology

[0002] Various techniques for estimating the effect of a treatment, i.e., the effect of a single treatment, are known. Among existing techniques for estimating the effect of a single treatment, for example, there is a method, as shown in Non-Patent Document 1, that uses a multi-task Gaussian process to estimate the effect of a single treatment at a single time point.

[0003] Existing technical documents

[0004] Non-patent literature

[0005] Non-patent literature 1: Alaa, AM & van der Schaar, M. "Bayesian inference of individualized treatment effects using multi-task gaussian processes", InAdvances in Neural Information Processing Systems, pages 3424-3432, 2017 Summary of the Invention

[0006] The problem that the invention aims to solve

[0007] In non-patent literature 1, only the effects of individual treatments at a single point in time and in the absence of such a treatment are known, and the degree of treatment amount cannot be considered. Therefore, when there are multiple types of treatments, it is unknown at what level of amount each treatment would yield the highest individual treatment effect.

[0008] Therefore, the purpose of one or more methods disclosed herein is to enable the learning of the processing effects of multiple processes.

[0009] Methods for solving problems

[0010] The information processing apparatus of the first aspect of this disclosure is characterized in that it comprises: an input unit that generates supervisory data including causal variables, processing types, and result variables for each of a plurality of processes; and a learning unit that uses the supervisory data to learn the result variables based on the causal variables and the processing types for each of the plurality of processes, thereby generating a learning model.

[0011] The information processing apparatus of the second aspect of this disclosure is characterized in that it comprises: an input unit that generates, for each of a plurality of processes, supervisory 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 representing the history of the time change of the first causal variable; and a learning unit that, using the supervisory data, learns, for each of the plurality of processes, according to the historical information, the time change of the first causal variable at a certain time based on the first causal variable, the second causal variable, and the processing type at a certain time, thereby generating a learning model.

[0012] The information processing apparatus of the third aspect of this disclosure is characterized in that it comprises: an estimation unit that estimates an outcome variable by inputting a causal variable and a processing type to a learning model for each of a plurality of processes, wherein the learning model is generated by learning the outcome variable based on the causal variable and the processing type for each of the plurality of processes using supervisory data containing the causal variable, the processing type, and the outcome variable; and an output unit that outputs the estimation result.

[0013] The information processing apparatus of the fourth aspect of this disclosure is characterized in that it comprises: an estimation unit that, for each of a plurality of processes, uses a learning model to estimate the time variation of a first causal variable that changes over time, obtained by combining two or more processes selected from the plurality of processes within a certain period, wherein the learning model is generated in such a manner as follows: using supervisory data containing the first causal variable, a second causal variable that does not change over time, a processing type, and historical information representing the history of the time variation of the first causal variable, for each of the plurality of processes, learning the time variation of the first causal variable for the next time period based on the historical information, according to the first causal variable, the second causal variable, and the processing type at a certain time; and an output unit that outputs the result of the estimation.

[0014] The program of the first aspect of this disclosure is characterized in that the program enables a computer to function as: an input unit that generates supervisory data, including causal variables, processing types, and outcome variables, for each of a plurality of processes; and a learning unit that uses the supervisory data to learn the outcome variables based on the causal variables and the processing types for each of the plurality of processes, thereby generating a learning model.

[0015] The program of the second aspect of this disclosure is characterized in that the program enables a computer to function as: an input unit that generates, for each of a plurality of processes, supervisory data comprising a first causal variable that changes over time, a second causal variable that does not change over time, a processing type, and historical information representing the history of the time change of the first causal variable; and a learning unit that, using the supervisory data, for each of the plurality of processes, learns, according to the historical information, the time change of the first causal variable at a certain time based on the first causal variable, the second causal variable, and the processing type at a certain time, thereby generating a learning model.

[0016] The program of the third aspect of this disclosure is characterized in that the program enables a computer to function as: an estimation unit that estimates an outcome variable for each of a plurality of processes by inputting a causal variable and a processing type to a learning model, wherein the learning model is generated by learning the outcome variable based on the causal variable and the processing type for each of the plurality of processes using supervised data containing the causal variable, the processing type, and the outcome variable; and an output unit that outputs the estimation result.

[0017] The program of the fourth aspect of this disclosure is characterized in that the program enables a computer to function as: an estimation unit, which, for each of a plurality of processes, uses a learning model to estimate the time variation of a first cause variable that varies over time, obtained by combining two or more processes selected from the plurality of processes within a certain period, the learning model being generated in such a way that: using supervisory data containing the first cause variable, a second cause variable that does not change over time, a processing type, and historical information representing the history of the time variation of the first cause variable, for each of the plurality of processes, learning the time variation of the first cause variable for the next time period based on the historical information, according to the first cause variable, the second cause variable, and the processing type at a certain time; and an output unit that outputs the result of the estimation.

[0018] The information processing method of the first aspect of this disclosure is characterized in that, for each of the plurality of processes, supervisory data including causal variables, processing types and outcome variables are generated, and using the supervisory data, for each of the plurality of processes, the outcome variables are learned according to the causal variables and the processing types, thereby generating a learning model.

[0019] The information processing method of the second aspect of this disclosure is characterized in that, for each of the plurality of processes, supervisory data is generated, which includes a first causal variable that changes over time, a second causal variable that does not change over time, a processing type, and historical information representing the history of the time change of the first causal variable. Using the supervisory data, for each of the plurality of processes, according to the historical information, the time change of the first causal variable at a certain time is learned for the next time based on the first causal variable, the second causal variable, and the processing type at a certain time, thereby generating a learning model.

[0020] The information processing method of the third aspect of this disclosure is characterized in that, for each of the plurality of processes, a causal variable and a processing type are input into a learning model to estimate an outcome variable, wherein the learning model is generated in the following manner: using supervised data containing the causal variable, the processing type, and the outcome variable, the outcome variable is learned for each of the plurality of processes based on the causal variable and the processing type, and the estimated result is output.

[0021] The information processing method of the fourth aspect of this disclosure is characterized in that, for each of the plurality of processes, a learning model is used to estimate the time variation of a first causal variable that changes over time, obtained by combining two or more processes selected from the plurality of processes within a certain period. The learning model is generated in such a way that, using supervised data containing the first causal variable, a second causal variable that does not change over time, a processing type, and historical information representing the history of the time variation of the first causal variable, for each of the plurality of processes, the time variation of the first causal variable at the next time period is learned based on the historical information, according to the first causal variable, the second causal variable, and the processing type at a certain time, and the estimated result is output.

[0022] Invention Effects

[0023] According to one or more methods disclosed herein, it is possible to learn the processing effects of multiple processes. Attached Figure Description

[0024] Figure 1 This is a graph showing the difference in infection rates between vaccinated and unvaccinated cases, as an example of the treatment effect.

[0025] Figure 2 This is a block diagram that schematically illustrates the structure of the estimation device according to Embodiment 1.

[0026] Figure 3 This is a table showing examples of treatments for categorical variables.

[0027] Figure 4 This is a schematic diagram used to illustrate the outline of the learning model in Implementation 1.

[0028] Figure 5 It is a schematic diagram used to illustrate the outline of a multi-task Gaussian process.

[0029] Figure 6 This is a table showing an example of a list of parameters, initial values, and number of parameters in Implementation 1.

[0030] Figure 7 This is a block diagram illustrating an example of a hardware structure.

[0031] Figure 8 This is a flowchart illustrating the processing operation of the estimation device in Embodiment 1.

[0032] Figure 9 This is a graph showing the difference in infection rates between those who received multiple vaccinations and those who did not, as an example of the effects of treatment at multiple points in time.

[0033] Figure 10 This is a block diagram that schematically illustrates the structure of the estimation device in Embodiment 2.

[0034] Figure 11 (A) to (D) are schematic diagrams showing the features of the estimation device of Embodiment 2.

[0035] Figure 12 (A) and (B) are schematic diagrams used to illustrate the outline of the learning model in Implementation 2.

[0036] Figure 13 This is a schematic diagram used to illustrate the general outline of a deep multitasking Gaussian process.

[0037] Figure 14 This is a schematic diagram used to illustrate the inputs and outputs of a deep multitasking Gaussian process.

[0038] Figure 15 This is a flowchart illustrating the processing actions of the encoder unit for estimating the individual processing effects at multiple time points in Embodiment 2.

[0039] Figure 16 This is a flowchart illustrating the processing actions of the decoder unit for estimating the individual processing effects at multiple time points in Implementation 2. Detailed Implementation

[0040] Implementation Method 1

[0041] In Implementation 1, an example of using a multi-task Gaussian process to estimate the individual processing effect at a single point in time will be described.

[0042] Here, as an example of the individual treatment effect at a single point in time, we estimate the individual treatment effect (ITE) at a single point in time.

[0043] Below, "disposal" is an example of "processing," and "dosage" is an example of "processing quantity."

[0044] The effect of individual treatment is defined as the difference between the outcome when an individual is treated and the outcome when no treatment is given.

[0045] The effectiveness of the treatment also varies depending on the type and quantity of treatment, as well as the individual attributes of the recipients.

[0046] For example, such as Figure 1 As shown, the difference in infection rates between vaccinated and unvaccinated individuals represents the effectiveness of the treatment. This effectiveness can be considered to vary depending on the attributes of the individuals to be vaccinated (e.g., gender, age, and pre-existing medical conditions), the type of vaccine, the dosage, and the environment at the time of vaccination.

[0047] Furthermore, this disclosure is not limited to individual treatment effects; by setting the attributes of individual objects as group attributes, it is also possible to calculate the average treatment effect (ATE) for each group.

[0048] Similarly, this disclosure can also be applied to Conditional Average Treatment Effect (CATE) or Local Average Treatment Effect (LATE).

[0049] In addition, the "single" in the effect of individual treatment can be referred to as individual or separate, and "treatment" can be referred to as intervention, treatment, measures, or exposure.

[0050] Furthermore, the effect of the treatment can also be called the causal effect.

[0051] Figure 2 This is a block diagram that schematically shows the structure of the estimation device 100, which is an information processing apparatus according to Embodiment 1.

[0052] The estimation device 100 includes an input unit 110, a learning unit 120, an estimation unit 140, and an output unit 150.

[0053] Input unit 110 inputs supervisory data to learning unit 120.

[0054] For example, the input unit 110 generates supervisory data containing cause variables, process types, and result variables for each of the multiple processes.

[0055] The input unit 110 includes a DB (Data Base) 111, which serves as a data storage unit, an input execution unit 112, and a preprocessing unit 113.

[0056] DB111 stores the data required for processing in the estimation device 100. Here, DB111 stores at least the cause variable X#, the treatment type W#, and the result variable Y#.

[0057] The input execution unit 112 obtains the cause variable X#, the treatment type W#, and the result variable Y# from DB111 and provides them to the preprocessing unit 113.

[0058] The causal variable X#, treatment type W#, and outcome variable Y# are represented by the following equation (1) based on the data observed for each individual i.

[0059]

Mathematical Formula 1

[0060]

[0061] As shown in equation (2) below, X i Let J1 represent the set of causal variables. Here, Q represents the number of causal variables.

[0062]

Mathematical Formula 2

[0063]

[0064] Furthermore, as shown in equation (3) below, each element of the causal variable can also be multidimensional information, and the dimensions of each element can also be different.

[0065]

Mathematical Expression 3

[0066]

[0067] Cause variable x q It can be expressed by the following equations (4) to (7).

[0068]

Mathematical Expression 4

[0069]

[0070]

Mathematical Expression 5

[0071]

[0072]

Mathematical Expression 6

[0073]

[0074]

Mathematical Expression 7

[0075]

[0076] Here, the characteristics of the object being processed include, for example, gender, age, and weight.

[0077] The amount of data processed is, for example, the amount of vaccines distributed.

[0078] Historical information on vaccine administration includes details such as the types of vaccines administered in the past, the number of administrations, and their effectiveness at the time.

[0079] Environmental information at the time of disposal is a factor that affects the disposal effect. Environmental information at the time of disposal includes, for example, weather, location, and market conditions.

[0080] As shown in equation (8) below, the treatment type W# is a categorical variable representing the type of treatment. For example, the categorical variable is the type of vaccine.

[0081]

Mathematical Expression 8

[0082]

[0083] The outcome variable Y# represents the treatment effect. For example, the treatment effect is the reduction in infection rate achieved based on vaccination.

[0084] The preprocessing unit 113 performs preprocessing on the cause variable X#, the treatment type W#, and the result variable Y# from the input execution unit 112, as described later, and provides the preprocessed cause variable X, treatment type W, and result variable Y to the learning unit 120.

[0085] If at least one of the processing type W# and the result variable Y# has a missing value, the preprocessing unit 113 removes the missing data {X}. i W i Y i}

[0086] Pretreatment Department 113 regarding treatment type W#, such as Figure 3 As shown, in the presence of multiple combined patterns observed with respect to the treatment, integer values ​​(1, ..., D) are assigned to each pattern. Here, D is the number of patterns.

[0087] Preprocessing unit 113 in element x of causal variable X# q In the case of missing values, use x q The missing value is filled by the average value of each dimension.

[0088] Preprocessing unit 113 processes the elements x of the causal variable X# qNormalize it so that it has a mean of 0 and a variance of 1 for each of its dimensions.

[0089] The learning unit 120 uses the supervised data provided by the input unit 110 to learn the learning model.

[0090] For example, the learning unit 120 uses supervised data from the input unit 110 to learn the outcome variables for each of the multiple processes, based on the causal variables and the type of process, thereby generating a learning model. Furthermore, the generated learning model is stored in a storage unit (not shown).

[0091] The learning unit 120 includes a calculation unit 121 and an optimization unit 122.

[0092] The calculation unit 121 initializes the parameters θ of the Gaussian process, inputs the cause variable X, calculates the variance-covariance matrix K(X, X') of the Gaussian process, and outputs the variance-covariance matrix K to the optimization unit 122.

[0093] Figure 4 This is a schematic diagram used to illustrate the outline of the learning model performed by the learning unit 120.

[0094] The correlation between the sequences of observed variables in the learning model can be represented by the linear correlation of the LMC (Linear Model of Coregionalization) kernel of the well-known MTGP (Multitask Gaussian Process).

[0095] In addition, MTGP is described in detail in the following document 1.

[0096] The following describes the method for calculating the variance-covariance matrix of the LMC kernel in this embodiment.

[0097] The treatment effect Y shown in equation (9) below uses LMC and is expressed by equations (10) and (11) below.

[0098]

Mathematical Expression 9

[0099]

[0100]

Mathematical Formula 10

[0101]

[0102]

Mathematical Expression 11

[0103]

[0104] Among them, a d,q From the state function gq To the observation function f d The linear correlation coefficient, σ d 2 It is the variance of the observation error.

[0105] (10) g q (x q ) indicates that based on the element x of the cause variable q The generated Gaussian process gp(μ(x) q ), k q (x q x q The function of ')).

[0106] Where, μ(x) q ) represents x q The average matrix, however, in the preprocessing unit 113, x q Normalized, therefore, becomes μ(x) q ) = 0.

[0107] k q This indicates a positive definite kernel. Among the positive definite kernels to be used, a kernel suitable for the data, such as the RBF (Radial Basis Function) kernel, can be selected. For example, the RBF kernel is represented by the following equation (12).

[0108] Among them, l 1 q and l 2 q These are the parameters of the RBF core.

[0109]

Mathematical Expression 12

[0110]

[0111] When the function f shown in equation (13) below is based on a Gaussian process, the function f is represented by a multidimensional Gaussian distribution as shown in equation (14) below.

[0112]

Mathematical Expression 13

[0113]

[0114]

Mathematical Expression 14

[0115]

[0116] Where K(X, X') and (K(X, X')) d,d’ It can be expressed by the following equations (15) and (16).

[0117]

Mathematical Expression 15

[0118]

[0119]

Mathematical Expression 16

[0120] (K(X,X′)) d,d′ =cov[f d (X),f d′ (X′)] (16)

[0121] In equation (14), K(X, X') is the variance-covariance matrix or Gram matrix, which is a matrix representing the similarity between causal variables such as X and X'.

[0122] μ(X) represents the average matrix of the causal variable X. However, in the preprocessing unit 113, the causal variable X is normalized, and therefore becomes μ(X) = 0.

[0123] The components of the variance-covariance matrix K(X, X') (K(X, X')) d,d’ f can be used in equation (10) d (X) is calculated as shown in equation (17) below.

[0124] Furthermore, in the transformation from the second row to the third row in equation (17), the different causal variables (q≠q') are independent. Therefore, the content shown in equation (18) below can be used.

[0125]

Mathematical Expression 17

[0126]

[0127]

Mathematical Expression 18

[0128] cov[g q (x q ),g q′ (x′ q′ )]=0 (18)

[0129] K(X, X') is expressed as shown in equations (19) and (20) below.

[0130]

Mathematical Expression 19

[0131]

[0132]

Mathematical Expression 20

[0133]

[0134] Among them, B q It satisfies equation (21) below. Furthermore, B qIt is the coregionalization matrix, representing the state function g. q To the observation function f d Linear transformation.

[0135]

Mathematical Expression 21

[0136]

[0137] The outcome variable Y is defined using the observation function f, as shown in equations (22), (23) and (24) below.

[0138]

Mathematical Expression 22

[0139]

[0140]

Mathematical Expression 23

[0141]

[0142]

Mathematical Expression 24

[0143]

[0144] K'(X, X') can be calculated using the causal variable X and the parameter θ shown in equation (25) below. The initial value of the parameter θ is as follows: Figure 6 As shown. These parameters are optimized by the optimization unit 122 of the learning unit 120.

[0145]

Mathematical Expression 25

[0146]

[0147] 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 to minimize the marginal likelihood.

[0148] The marginal likelihood is obtained by calculation as follows.

[0149] The probability p(Y|X, θ) of observing the outcome variable Y can be obtained by the following equation (26).

[0150]

Mathematical Expression 26

[0151]

[0152] Among them, K' θ (X, X') represents K'(X, X') calculated using the parameter θ.

[0153] When taking the logarithm of both sides of equation (26), the marginal likelihood log can be calculated using the following equation (27). p (Y|X、θ).

[0154]

Mathematical Expression 27

[0155]

[0156] Where N is the length of vector X, and D is the output dimension of Y as the number of tasks.

[0157] Regarding the optimization of parameter θ, the marginal likelihood log p (Y|X、θ) can be maximized. For example, to make the form consistent with general optimization problems, the optimization function E can be minimized, as shown in equation (28) below, which is obtained by multiplying both sides by a negative sign.

[0158]

Mathematical Expression 28

[0159]

[0160] The optimization unit 122 updates θ so that the optimization function E of equation (28), i.e., the marginal likelihood - log p Minimize (Y|X, θ). During optimization, K also needs to be updated when updating θ. θ Therefore, the calculation unit 121 calculates K' (X, X'). θ (X, X').

[0161] In the optimization of the optimization function E, known techniques such as stochastic gradient descent can be used.

[0162] As described above, the optimization unit 122 is capable of optimizing the parameter θ.

[0163] The estimation unit 140 uses the variance-covariance matrix K' calculated using the parameters θ optimized by the calculation unit 121. θ (X, X'), calculate the new input X ＊ The effectiveness of the treatment.

[0164] The new input X shown in equation (29) below ＊ The output y in ＊ It can be expressed by the following equation (30).

[0165]

Mathematical Expression 29

[0166]

[0167]

Mathematical Expression 30

[0168]

[0169] The output y ＊ Able to determine the variance-covariance matrix K' θ (X, X') and observation points {X, Y} are calculated, for example, using Gaussian process regression with known techniques, by the following equation (31).

[0170]

Mathematical Expression 31

[0171]

[0172] Among them, K ＊ B can be expressed by the following equation (32). q [:,d] is B q The vector in the d-th column.

[0173]

Mathematical Expression 32

[0174]

[0175] Here, the treatment effect ITE in the case of no treatment (W=1) and the case of treatment (W=t) can be calculated by the following equation (33).

[0176]

Mathematical Expression 33

[0177]

[0178] The estimation unit 140 inputs the causal variable and treatment type into the learning model, thereby estimating the outcome variable. Based on the estimated outcome variable, the estimation unit 140 can determine the optimal combination of the causal variable and treatment type.

[0179] For example, the estimation unit 140 calculates the treatment effect ITE by using all combinations of treatments and treatment quantities, and determines the treatment T with the highest treatment effect and the combination of treatment quantities.

[0180] The output unit 150 outputs the estimation results performed by the estimation unit 140.

[0181] For example, output unit 150 is for object X ＊ The output displays the processing quantity (T) with the highest processing effect, its processing efficiency (ITE), and its processing effect (ITE). For example, the output unit 150 displays them.

[0182] The estimation device 100 described above, for example, can be used by... Figure 7 This is achieved using a PC (Personal Computer) 10 as shown.

[0183] PC10 has auxiliary storage devices such as HDD (Hard Disk Drive) and SSD (Solid State Drive) 11, memory 12, processors such as CPU (Central Processing Unit) 13, communication I / F (Interface) such as NIC (Network Interface Card) 14, input I / F such as keyboard and mouse 15, and a display 16.

[0184] For example, DB111 can be implemented through auxiliary storage device 11.

[0185] The program stored in the auxiliary storage device 11 is loaded into the memory 12, and the processor 13 executes the program, thereby enabling the input execution unit 112, the preprocessing unit 113, the calculation unit 121, the optimization unit 122, and the estimation unit 140 to be implemented.

[0186] The output unit 150 can be implemented through the display 16.

[0187] The program can also be downloaded to the auxiliary storage device 11 via a reader / writer (not shown) from a recording medium (not shown), or downloaded to the auxiliary storage device 11 from a network via communication I / F 14, and then loaded onto the memory 12 for execution by the processor 13. Alternatively, it can be loaded directly onto the memory 12 from the recording medium via a reader / writer, or directly onto the memory 12 from a network via communication I / F 14, and then executed by the processor 13.

[0188] In other words, programs can also be provided through program products such as recording media.

[0189] In addition, the storage unit (not shown) of the storage learning model can also be implemented by the auxiliary storage device 11.

[0190] Figure 8 This is a flowchart illustrating the operation of the estimation device 100 in Embodiment 1.

[0191] First, the input execution unit 112 obtains the cause variable X#, the treatment type W#, and the result variable Y# from DB111 (S10).

[0192] Next, the preprocessing unit 113 removes any missing data from the processing type W# and the result variable Y# (S11).

[0193] Next, the pretreatment unit 113 performs pretreatment (S12) of the treatment type W#.

[0194] Next, the preprocessing unit 113 processes the missing values ​​of the cause variable X# (S13).

[0195] Next, the preprocessing unit 113 normalizes the causal variable X# (S14).

[0196] Through the above steps S11 to S14, the causal variable X#, treatment type W#, and outcome variable Y# are preprocessed to become the causal variable X, treatment type W, and outcome variable Y.

[0197] Next, the calculation unit 121 sets the parameter θ shown in equation (25) above to an initial value (S15).

[0198] Next, the calculation unit 121 uses X and θ to calculate K using the above equation (19). θ (X, X')(S16).

[0199] Next, the optimization unit 122 calculates the optimization function E(S17) using the above equation (28).

[0200] Next, the optimization unit 122 optimizes θ to minimize the optimization function E (S18).

[0201] Next, the computation unit 121 determines whether the optimization function E converges (S19). The condition for the optimization function E to converge is the marginal likelihood - log p The value of (Y|X、θ) becomes the minimum, or the processing in steps S16 to S18 reaches the predetermined number of repetitions.

[0202] If the optimization function E does not converge (S19: No), the process returns to step S16; if the optimization function E converges (S19: Yes), the process proceeds to step S20.

[0203] In step S20, the estimation unit 140 calculates the new input X. ＊ The effect of ITE in the treatment.

[0204] Next, the estimation unit 140 calculates the combination of treatments with the highest treatment effect ITE and their treatment quantities (S21).

[0205] In the method described in Non-Patent Document 1, only one treatment is known to exist and the treatment effect of no treatment is known. However, as described above, in the estimation device 100 of Embodiment 1, the treatment effects of multiple treatments can be estimated.

[0206] Furthermore, in the method of Non-Patent Document 1, only the characteristic quantities of the object to be disposed of are considered. However, in the estimation apparatus 100 of Embodiment 1, in addition to the characteristic quantities of the object to be disposed of, disposal history information and environmental information at the time of disposal are set as separate independent causal variables (latent variables). In this way, by inputting separate independent causal variables, for example, compared to merging the characteristic quantities of the object and the disposal, which should be independent, into a single characteristic quantity to calculate the variance-covariance matrix of the Gaussian process, when calculating the variance-covariance matrix as separate latent variables, no apparent correlation is generated, and the estimation accuracy can be improved. Here, apparent correlation means that even when calculating the correlation coefficient between variables that are not actually correlated, it will not be "0".

[0207] By estimating the individual treatment effect at a single point in time using this implementation method, and considering the characteristics of the target population, it is possible to determine the appropriate level and extent of each treatment (measure). For example, by considering the target population, it can be determined which vaccine, and in what quantity, will result in the lowest infection rate.

[0208] Furthermore, in Implementation 1, the determination of vaccine treatment effect was described as an example of treatment effect, but it is not limited to this. For example, as described below, if variables are set as follows, this implementation can be applied to cancer treatment, promotion of high-priced services such as reserved seats on railways or luxury rooms in hotels, or increase in store sales.

[0209] The variables in the context of cancer treatment are as follows.

[0210] The characteristics of the object being processed can be set as at least one of the attributes of the individual receiving cancer treatment (such as gender, age, presence or absence of disease) and physical condition information.

[0211] The environmental information during processing can be set as factors that influence the effectiveness of cancer treatment. For example, factors influencing the effectiveness of cancer treatment can be set as information on at least one of the factors: the region where treatment is received and the hospital.

[0212] Processing historical information can be set as the treatment history of past cancer treatments. For example, the treatment history of past cancer treatments includes the types of cancer treatments received, the number of times, and the effects at the time.

[0213] The treatment can be set to either radiation therapy or chemotherapy.

[0214] The treatment capacity can be set to the radiation intensity and the dosage of anticancer agents.

[0215] The treatment effect can be set as the amount of reduction in the size of the cancer.

[0216] The variables for an increase in store sales are as follows.

[0217] The characteristics of the processing object can be set as at least one of the attributes of the purchaser of the product (e.g., gender, age, annual income, hobbies, etc.) and purchase history.

[0218] The environmental information processed can be set as factors that affect the store's sales. For example, factors affecting the purchase of goods can be set as at least one of the following: the popularity of the goods, the weather, and market conditions.

[0219] The processing history information can be set to the purchase history at the time of the processing. For example, the purchase history at the time of the processing includes the types of goods purchased, the number of times, and the purchase date and time.

[0220] The processing categories can be set as discount coupons, internet ads, television ads, and travel booking vouchers.

[0221] The processing volume can be set to the discount amount, advertising frequency, and winning probability.

[0222] The processing result can be set as the store's sales revenue.

[0223] The variables for promoting high-priced (high-value) services are as follows.

[0224] Here, high-priced services include reserved seats on trains or buses, premium seats on airplanes, premium cabins on ships, reserved seats for sporting or entertainment events, priority tickets for sightseeing facilities, or premium rooms in hotels.

[0225] In addition, with or without a ticket, priority includes guaranteed seating, same-day no-queue access to rides, and time-reserved privileges. Specifically, higher-priced railway services include not only the usual reserved seats, but also reserved seats with good views, luxury sleeper cars, and reserved seats on special trains.

[0226] The characteristics of the processing object can be set to at least one of the attributes of the service user (e.g., gender, age, annual income, work style, place of residence, place of work, hobbies, family structure, etc.) and purchase history.

[0227] Environmental information during processing can be defined as factors affecting service sales revenue. For example, factors affecting service sales revenue can include at least one of the following: reservation rate, congestion rate, weather, accidents, sales plans, and the spread of infectious diseases. Specifically, in railway services, weather includes disasters such as typhoons, earthquakes, and heavy rain. Furthermore, accidents include vehicle malfunctions, overhead line failures, personal injury accidents, and fires along the line. Additionally, sales plans include planned service cancellations and temporary increases in service due to power outages, construction, strikes, extended holidays, etc.

[0228] Processing history information can be set as the action history at the time of processing in the past. Action history is the history of actions taken by service users. For example, action history also includes the history of users purchasing tickets (in other words, purchase history), the history of users registering for activities, the history of users exercising rights, the history of users using coupons, etc.

[0229] The processing categories can be set to price increase, price decrease, promotion, and coupon issuance.

[0230] The processing volume can be set to the price increase amount, price decrease amount, or coupon amount. Additionally, the price increase amount and price decrease amount can also be the difference between the price increases (price increase rate) and the difference between the price decreases (price decrease rate), respectively.

[0231] The processing result is the sales revenue of high-priced (high-value) services.

[0232] As described above, when applying Implementation 1 to railway services, it is as follows.

[0233] The causal variables are at least one of the attributes and purchase history of railway users, factors affecting the sales amount of railway services, the behavior history of railway users, the amount of freight rate increases, the amount of freight rate decreases, and the amount of coupons used to promote railway use.

[0234] The types of transactions handled include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use.

[0235] The outcome variable became the sales revenue of high-priced railway services.

[0236] Furthermore, depending on the processing results obtained, operations such as the first to third operations can be performed.

[0237] Operation 1: In order to increase the reservation rate of reserved seats on the railway, the price of reserved seats can be reduced according to the travel time, so that long-term users can be seated when crowded, and conversely, when idle, the price of reserved seats can be reduced for short-term users.

[0238] The second action: It can conduct activities to upgrade the seat class of railway and airplane trains for free, guiding people to make reservations for higher-class seats next time.

[0239] The third operation: It can recommend services within a price range that matches the user's characteristics.

[0240] In addition, it is possible to guide the purchase of higher-priced services by issuing coupons and other means.

[0241] Implementation Method 2

[0242] In Implementation 2, a deep multi-task Gaussian process is used to estimate the individual processing effects at multiple time points. Specifically, in Implementation 2, the case of estimating the individual processing effects when processing is performed at multiple time points with varying times is described.

[0243] In implementation method 2, for example, as Figure 9 The following describes a scenario where four vaccinations are planned to reduce the infection rate of the virus. Specifically, in Implementation Method 2, if two doses of the vaccine have already been administered and two more doses are planned, it is estimated which vaccine, when administered twice, would result in the lowest infection rate. Alternatively, the effectiveness of the subsequent four vaccinations can be estimated even if no vaccinations have been administered at all.

[0244] Figure 10 This is a block diagram that schematically shows the structure of the estimation device 200, which is an information processing apparatus according to Embodiment 2.

[0245] The estimation device 200 includes an input unit 210, an encoder learning unit 220, a decoder learning unit 230, an estimation unit 240, and an output unit 250.

[0246] The input unit 210 inputs supervisory data to the encoder learning unit 220.

[0247] For example, the input unit 210 generates monitoring data for each of the multiple processes, including a first cause variable that changes over time, a second cause variable that does not change over time, the type of process, and historical information representing the history of the time change of the first cause variable.

[0248] The input unit 110 includes a DB 211 which serves as a data storage unit, an input execution unit 212, and a preprocessing unit 213.

[0249] DB211 stores the data required for processing in the estimation device 200. Here, DB211 stores at least the feature value X#, the feature value V#, and the processing type W#.

[0250] Here, feature X# is the feature of the object being disposed of that changes due to the disposal, while feature V# is the feature of the object being disposed of that does not change due to the disposal.

[0251] The input execution unit 212 obtains the feature quantity X#, feature quantity V# and processing type W# from DB211 and provides them to the preprocessing unit 213.

[0252] The characteristic quantity X#, characteristic quantity V# and treatment type W# are represented by the following equation (34) based on the data observed for each individual i.

[0253]

Mathematical Expression 34

[0254]

[0255] Where N is the number of observed data points, s max This indicates the duration of the observed time series data.

[0256] The preprocessing unit 213 performs preprocessing on the feature quantity X#, feature quantity V# and processing type W# from the input execution unit 212, as described later, and provides the preprocessed feature quantity X, feature quantity V and processing type W to the encoder learning unit 220.

[0257] If there are missing values ​​in the feature quantity X#, feature quantity V# and treatment type W# of individual i, the preprocessing unit 213 removes the data J2 of individual i as shown in the following formula (35).

[0258]

Mathematical Expression 35

[0259]

[0260] In addition, the pretreatment unit 213 pretreatments the treatment type W# using the same steps as the pretreatment unit 113 in Embodiment 1.

[0261] Furthermore, the preprocessing unit 213 normalizes the feature quantity X# and the feature quantity V# to obtain a mean of 0 and a variance of 1.

[0262] The encoder learning unit 220 and the decoder learning unit 230 function as learning units for learning the learning model in Embodiment 2.

[0263] For example, in Implementation 2, the learning unit uses supervisory data from the input unit 210 to learn the temporal changes of the first cause variable for the next time period based on historical information, according to the first cause variable, the second cause variable, and the type of processing at a certain time, thereby generating a learning model. The learned learning model is stored in a storage unit (not shown).

[0264] Figure 11 (A) to (D) are schematic diagrams illustrating the action concepts of the learning model in Implementation 2.

[0265] exist Figure 11 The diagram shows the causal variable X that changes over time, the causal variable V that does not change over time, the treatment type W, and historical information H.

[0266] Here, X is a factor, such as the infection rate, that changes due to treatment; it is both a causal variable and a target variable. For example, treatment W is performed in step s. s In X s Transform into X s+1 In the case of Xs It is the causal variable, X s+1 It becomes the target variable.

[0267] Will Figure 11 The single-dotted line indicates the treatment W s and the results after treatment X s+1 The impact of X on both sides s It is called the interleaving factor.

[0268] In the presence of an interleaving factor, due to X s To W s The increased number of observations on treatments with high probabilities leads to a bias when the model is optimized for high-probability treatments during learning. Consequently, when estimating the effects of treatments with low probabilities, the prediction accuracy deteriorates, making it difficult to accurately predict the treatment outcomes.

[0269] Therefore, in Implementation 2, performance learning, as shown in the fourth feature below, is used to mitigate the influence of the interleaving factor.

[0270] First, let's explain the learning model.

[0271] Figure 12 (A) and (B) are schematic diagrams used to illustrate the learning model in Implementation Method 2.

[0272] The learning model in Implementation 2 is a combination of the known Encoder-Decoder (or Seq2Seq) and the known Deep Multitask Gaussian Process.

[0273] First, the features of the learning model in Implementation Method 2 will be explained.

[0274] The first characteristic: When considering the handling of situations at multiple points in time, the question arises as to what extent the impact of the past should be taken into account. For example... Figure 11 As shown in (A), the more past information is included, the more complex the causal graph becomes, and the more factors need to be considered. Therefore, in implementation 2, for example, by importing... Figure 11 The Markov property shown in (B) allows information from previous times to be used as historical information H. s-1 To simplify the learning model.

[0275] Second feature: such as Figure 11 (B) and Figure 13 As shown, in history H s In the calculation, for example, intermediate representations (Hidden Variables) can be computed using deep multi-task Gaussian processes with known techniques.

[0276] The third feature: such as Figure 11 (C) and Figure 13 As shown, the second-level branch of the deep multi-task Gaussian process enables one side to utilize the multi-task Gaussian process according to H. s To predict the causal variable J3 shown in equation (36) below, so that the other party uses the Gaussian process classification model of known technology to predict the treatment type J4 shown in equation (37) below based on Hs.

[0277]

Mathematical Expression 36

[0278]

[0279]

Mathematical Expression 37

[0280]

[0281] Fourth Feature: Utilizing performance learning methods based on known technologies, calculations are performed according to H... s Improve the prediction accuracy of the causal variable J3 shown in equation (36) above and make it difficult to predict based on H. s To anticipate the treatment type J4 shown in equation (37) above, H s f h Learning from H, thereby reducing the risk of infection. s To W s The influence of intertwining factors.

[0282] Fifth feature: such as Figure 11 (D) and Figure 12 As shown, by using the intermediate representation H of the deep multi-task Gaussian process... s Recursively transferring to the next step (s+1), it enables long-term prediction and can estimate the disposal plan W for any number of iterations. t：t+τ The effectiveness of the treatment.

[0283] Sixth feature: The estimation device 200 of Embodiment 2, as described above, uses a deep multitasking Gaussian process to estimate the processing effect of multiple processes. Therefore, based on the effect described in Embodiment 1, the reliability of the estimated processing effect can also be determined.

[0284] The following describes the deep multi-task Gaussian process and the learning model of the Encoder-Decoder in Implementation Method 2.

[0285] First, the deep multi-task Gaussian process is explained.

[0286] Equations (38) to (43) below show the relationships of deep multitasking Gaussian processes.

[0287] Equation (38) is the relation of the Gaussian process in the first layer MTGP, and equation (39) is the likelihood function in the first layer MTGP.

[0288] Equation (40) is the relation of the Gaussian process in the second layer MTGP, and equation (41) is the likelihood function in the second layer MTGP.

[0289] Equation (42) is the relation of the Gaussian process in the GP classification layer, and equation (43) is the likelihood function in the GP classification layer.

[0290]

Mathematical Expression 38

[0291]

[0292]

Mathematical Expression 39

[0293]

[0294]

Mathematical Expression 40

[0295]

[0296]

Mathematical Expression 41

[0297]

[0298]

Mathematical Expression 42

[0299]

[0300]

Mathematical Expression 43

[0301]

[0302] Here, X' s It can be expressed by the following equation (44).

[0303]

Mathematical Expression 44

[0304] X′ s ={X s ,V,H s-1} (44)

[0305] Furthermore, since it is a Gaussian process, in addition to the average values ​​J5 and J6 of the estimated values ​​shown in equations (40) and (41) or (42) and (43) below, we also obtain the standard deviations J7 and J8 shown in equations (47) and (48) below.

[0306]

Mathematical Expression 45

[0307]

[0308]

Mathematical Expression 46

[0309]

[0310]

Mathematical Expression 47

[0311]

[0312]

Mathematical Expression 48

[0313]

[0314] Furthermore, in equations (38) to (43) above, f is a function of the Gaussian process, μ, sm, and N are the average functions, representing the flexible maximum function and the Gaussian distribution, respectively. θ represents the parameters of the Gaussian process, and σ represents the variance of the Gaussian distribution. In addition, none of the parameters are time-dependent, and are assumed to be the same at all times.

[0315] Composition of X' s Each causal variable can be set as an independent causal variable in the same way as in Implementation 1, or each causal variable (potential variable) can be divided into elements that do not change with time and elements that change with time, for example, elements Xs and elements V as shown in equations (49) and (50) below.

[0316]

Mathematical Expression 49

[0317]

[0318]

Mathematical Expression 50

[0319] V = {v} covariate ,v dosage ,v history ,v environment} (50)

[0320] The prediction object J9 shown in equation (51) below, which is included in equation (41) above, can also only predict factors that change due to treatment (e.g., infection rate). For example, only factor J10 shown in equation (52) below can be set as the prediction object.

[0321]

Mathematical Expression 51

[0322]

[0323]

Mathematical Expression 52

[0324]

[0325] Next, the Gaussian process will be explained.

[0326] The matrix K1 shown in equation (53) below represents the variance-covariance matrix of a typical single-output Gaussian process.

[0327] The matrices K2 and K3 shown in equations (54) and (55) below represent the variance-covariance matrices of the multitasking Gaussian process.

[0328]

Mathematical Expression 53

[0329]

[0330]

Mathematical Expression 54

[0331]

[0332]

Mathematical Expression 55

[0333]

[0334] Here, as Figure 14 As shown, the output dimension D of the variance-covariance matrix K2 of the multi-task Gaussian process represented by equation (54) above is... H The output dimension N of the variance-covariance matrix K3 shown in equation (55) is set to a dimension of 1 or higher. task The number of categories that are set for W.

[0335] The kernels for calculating the variance-covariance matrices K2 and K3 shown in equations (54) and (55) above can be either Independent Multi-task Kernel, which is uncorrelated between output dimensions, or Dependent Multi-task Kernel, which is correlated between output dimensions. In Dependent Multi-task Kernel, for example, an LMC kernel, which is linearly correlated between output dimensions, can also be used.

[0336] Directly calculating a Gaussian process requires O(N) 3 The computational cost of this method can be reduced by using known techniques such as Stochastic Variational Inference for Gauss Process (SVI-GP), which utilizes inducing points.

[0337] Next, the likelihood function will be explained.

[0338] The predicted distribution f of the output of the first layer MTGP h (X sSince it cannot be directly used as the input to the Gaussian process of the second layer, the predicted distribution f of equation (39) mentioned above, which is used in the known technique of Double Stochastic Variational Inference for Deep Gaussian Processes (DSVI-DGP), can also be taken. h (X t The average of ') is used to calculate the intermediate performance J11 shown in equation (56) below, or taken from the predicted distribution f. h (X s The sampling of ') yields the intermediate representation J11 shown in the following equation (56).

[0339]

Mathematical Expression 56

[0340]

[0341] The second layer of MTGP observes only one task for each input data (e.g., processing W). s Therefore, as Figure 14 As shown, only the output J12 of the MTGP shown in equation (57) below is multiplied by W. s Becoming a one-hot index matrix I tasks (W s And the type of treatment.

[0342]

Mathematical Expression 57

[0343]

[0344] The second layer, the Classification Layer, is configured as a softmax function, as shown in equation (48) above, to predict the treatment W of the categorical variable. s .

[0345] Next, the encoder-decoder model will be explained.

[0346] Using known encoder-decoder techniques, the intermediate representation H of a deep multi-task Gaussian process is obtained. s Proceed to the next step. For example... Figure 12 As shown, the encoder unit is capable of one-step ahead prediction, and the decoder unit is capable of multiple-step ahead prediction. Furthermore, the encoder-decoder model in Embodiment 2 is not limited to... Figure 12As shown in the example, an attention mechanism can also be set up in the encoder-decoder and a Transformer can be used.

[0347] The deep multitasking Gaussian processes in the encoder and decoder sections have the same structure, but the parameters are refined separately.

[0348] Below, we will explain the details of the actions of each part based on the deep multi-task Gaussian process and encoder-decoder model described above.

[0349] Figure 10 The encoder learning unit 220 performs encoder-side learning in the encoder-decoder model.

[0350] Figure 10 The encoder learning unit 220 has an initialization unit 221, a prediction unit 222, a classification optimization unit 223, and a prediction optimization unit 224.

[0351] Initialization section 221 Figure 14 The dimension D of H shown H Set the dimension to 1 or higher.

[0352] Furthermore, the initialization unit 221 initializes the parameter J13 of the depth multitasking Gaussian process of the encoder shown in equation (58) below. The initialization value can be, for example, 1, or it can be a sample from an arbitrary Gaussian distribution.

[0353]

Mathematical Expression 58

[0354]

[0355] like Figure 12 As shown in (A), firstly, the prediction unit 222 uses the encoder to input H0 and {X1, V, W1} as zero vectors in the initial step (e.g., s=1) and calculates the data shown in the following equation (59).

[0356] J14. Thereafter, the prediction unit 222 recursively inputs the data J15 shown in equation (60) and the data J16 shown in equation (61) below until step t (t is an integer greater than or equal to 2), thereby calculating the data J17 shown in equation (62) below.

[0357]

Mathematical Expression 59

[0358]

[0359]

Mathematical Expression 60

[0360]

[0361]

Mathematical Expression 61

[0362]

[0363]

Mathematical Expression 62

[0364]

[0365] Then, the prediction unit 222 calculates the data J18 shown in equation (63) below for all individuals (i), and provides the data J18 to the classification optimization unit 223.

[0366]

Mathematical Expression 63

[0367]

[0368] The classification optimization unit 223 takes the data J19 calculated by the prediction unit 222 as shown in equation (64) below as input, and calculates the data J20 shown in equation (65) below according to the above equations (42) and (43).

[0369]

Mathematical Expression 64

[0370]

[0371]

Mathematical Expression 65

[0372]

[0373] Then, the classification optimization unit 223 uses the known technique SVI-GP to minimize the loss function shown in equation (66) below, thereby optimizing the parameter J21 shown in equation (67) below.

[0374]

Mathematical Expression 66

[0375]

[0376]

Mathematical Expression 67

[0377]

[0378] Among them, L ELBO The Evidence Lower Bond (ELBO) represents the loss function of SVI-GP.

[0379] The prediction optimization unit 224 takes the data J22 calculated by the prediction unit 222 as shown in (68) below and the observation data J23 shown in (69) below as input, and uses the above equations (38) and (39) to calculate the data J24 shown in equation (70) below.

[0380]

Mathematical Expression 68

[0381]

[0382]

Mathematical Expression 69

[0383]

[0384]

Mathematical Expression 70

[0385]

[0386] The prediction optimization unit 224 takes the data J25 shown in equation (71) below as input and uses equations (40) to (43) above to calculate the data J26 shown in equation (72) below.

[0387]

Mathematical Expression 71

[0388]

[0389]

Mathematical Expression 72

[0390]

[0391] To mitigate the influence of the interleaving factor, the prediction optimization unit 224 uses the loss function shown in equation (76) below and the known technique DSVI-DGP to optimize the parameter J30 shown in equation (77) below, so that the treatment J28 shown in equation (74) below cannot be predicted based on the information J27 shown in equation (73) below, and the prediction accuracy of the treatment effect J29 shown in equation (75) below is improved based on the information J27. However, the parameter J31 shown in equation (78) below is fixed.

[0392]

Mathematical Expression 73

[0393]

[0394]

Mathematical Expression 74

[0395]

[0396]

Mathematical Expression 75

[0397]

[0398]

Mathematical Expression 76

[0399]

[0400]

Mathematical Expression 77

[0401]

[0402]

Mathematical Expression 78

[0403]

[0404] Where, α e Consider the proportion of term J32 shown in equation (79) below.

[0405]

Mathematical Expression 79

[0406]

[0407] The prediction optimization unit 224 returns to the processing in the prediction unit 222, and repeatedly performs the processing of the classification optimization unit 223 and the prediction optimization unit 224, and optimizes the parameter J35 shown in the following equation (82) until the value J33 shown in the following equation (80) and the value J34 shown in the following equation (81) converge.

[0408]

Mathematical Expression 80

[0409]

[0410]

Mathematical Formula 81

[0411]

[0412]

Mathematical Expression 82

[0413]

[0414] When the values ​​J33 shown in equation (80) and J34 shown in equation (81) converge, the prediction optimization unit 224 uses the optimized parameter J35 shown in equation (82) and the optimized parameter in the prediction unit 222 to provide the data J18 calculated using equation (63) to the decoder learning unit 230.

[0415] Figure 10 The decoder learning unit 230 performs decoder-side learning in the encoder-decoder model.

[0416] Figure 10 The decoder learning unit 230 has an initialization unit 231, a prediction unit 232, a classification optimization unit 233, and a prediction optimization unit 234.

[0417] The initialization unit 231 initializes the parameters J36 of the decoder unit shown in equation (83) below using the optimized parameters J35 of the encoder unit as shown in equation (82) above.

[0418]

Mathematical Formula 83

[0419]

[0420] The prediction unit 232 performs the following processing of the variable J37 (τ-step ahead prediction) from prediction step t to τ as shown in equation (84).

[0421]

Mathematical Expression 84

[0422]

[0423] like Figure 12 As shown in (B), the prediction unit 232 will use the data J38 and the observation data {V,W}, which are finally predicted by the encoder in the decoder, as shown in the following equation (85). (t+1) As input, calculate the data J39 shown in equation (86) below.

[0424]

Mathematical Expression 85

[0425]

[0426]

Mathematical Expression 86

[0427]

[0428] Next, the prediction unit 232 recursively inputs the data J40 shown in equation (87) and the predicted data J41 shown in equation (88) until step t+τ-1, thereby calculating the data J42 shown in equation (89) and obtaining the data J43 shown in equation (90).

[0429]

Mathematical Expression 87

[0430]

[0431]

Mathematical Formula 88

[0432]

[0433]

Mathematical Expression 89

[0434]

[0435]

Mathematical Expression 90

[0436]

[0437] The prediction unit 232 calculates the data J44 shown in equation (91) for all individuals (i), and provides the data J44 to the classification optimization unit 233. max This represents the maximum number of prediction steps.

[0438]

Mathematical Expression 91

[0439]

[0440] The classification optimization unit 233 takes the data J45 shown in equation (92) calculated by the prediction unit 232 as input, and calculates the data J46 shown in equation (93) according to the above equations (42) and (43).

[0441]

Mathematical Expression 92

[0442]

[0443]

Mathematical Expression 93

[0444]

[0445] The classification optimization unit 233 uses the known technique SVI-GP to minimize the loss function shown in equation (94) below, thereby optimizing the parameter J47 shown in equation (95) below. Furthermore, during optimization, τ = 1 to τ... max The total loss is calculated, and Back Propagation is performed to optimize the process.

[0446]

Mathematical Expression 94

[0447]

[0448]

Mathematical Expression 95

[0449]

[0450] The prediction optimization unit 234 takes the data J48 shown in equation (96) below as the initial input, such as Figure 12 As shown in (B), the data J49 shown in the following equation (97) is recursively input until s = t + τ - 1, thereby calculating the data J50 shown in the following equation (98).

[0451]

Mathematical Expression 96

[0452]

[0453]

Mathematical Expression 97

[0454]

[0455]

Mathematical Formula 98

[0456]

[0457] Then, the prediction optimization unit 234 uses the known technique DSVI-DGP to optimize the parameter J51 shown in equation (100) using the loss function shown in equation (99) below. In addition, it is assumed that the parameter J52 shown in equation (101) below is fixed.

[0458]

Mathematical Expression 99

[0459]

[0460]

Mathematical Expression 100

[0461]

[0462]

Mathematical Expression 101

[0463]

[0464] Where, α d Consider the proportion of term J53 shown in equation (102) below.

[0465]

Mathematical Expression 102

[0466]

[0467] The prediction optimization unit 234 returns to the processing in the prediction unit 232, and repeatedly performs the processing of the classification optimization unit 233 and the prediction optimization unit 234 to optimize the parameter J36 shown in the above equation (83) until the value J54 shown in the following equation (103) and the value J55 shown in the following equation (104) converge.

[0468]

Mathematical Expression 103

[0469]

[0470]

Mathematical Expression 104

[0471]

[0472] When the values ​​J54 shown in equation (103) and J55 shown in equation (104) converge, the prediction optimization unit 234 provides the optimized parameter J36 shown in equation (83) and the optimized encoder parameter J35 shown in equation (82) to the estimation unit 240.

[0473] 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 can also be achieved using methods other than those shown above.

[0474] For example, Adversarial Training, a known technique, can also be used for learning. In this case, it could also be Discriminator(f) a ) learns in a way that allows the generator to predict the action J57 shown in equation (106) based on the intermediate performance J56 shown in equation (105) below. h The prediction accuracy of the treatment effect J58 shown in equation (107) is improved by using the intermediate performance J56 shown in equation (105) below, and the Discriminator(f a It is impossible to learn the manner of the treatment J57 shown in equation (106) based on the intermediate performance J56 shown in equation (105) below.

[0475]

Mathematical Expression 105

[0476]

[0477] [Mathematical Expression 106]

[0478]

[0479]

Mathematical Expression 107

[0480]

[0481] The estimation unit 240 uses a learning model to estimate the time variation of the first cause variable by combining two or more treatments selected from multiple treatments within a certain period.

[0482] Therefore, the estimation unit 240 is able to determine the optimal combination of two or more processes based on the estimated time changes.

[0483] For example, the estimation unit 240 calculates the processing effect J60 of the processing plan J59 for any time t to τ as shown in equation (82) above and the variance value J61 as shown in equation (110) below, based on the parameters J35 of the optimized encoder shown in equation (82) above and the parameters J36 of the optimized decoder shown in equation (83) above.

[0484] [Mathematical Expression 108]

[0485]

[0486]

Mathematical Expression 109

[0487]

[0488]

Mathematical Expression 110

[0489]

[0490] The estimation unit 240 selects the treatment plan J59 shown in the above formula (108) that has the highest final treatment effect J62 as shown in the following formula (111).

[0491]

Mathematical Expression 111

[0492]

[0493] Then, the estimation unit 240 provides the output unit 250 with the optimal treatment plan J59 shown in equations (108), (109) and (110) above, its treatment effect J62, and its variance value J63 shown in equation (112) below.

[0494]

Mathematical Expression 112

[0495]

[0496] In cases where there are many combinations of the disposal plan J59 shown in equation (108) above, or when τ is a long period, disposal plans can also be searched by Monte Carlo sampling or reinforcement learning to learn the combination with the highest final disposal effect.

[0497] Alternatively, the treatment plan with the largest cumulative treatment effect J62 in the treatment plan J59 shown in equations (108) and (111) above, or the treatment plan with the largest cumulative treatment effect J64 shown in equation (113) below, can be set as the optimal treatment plan.

[0498]

Mathematical Expression 113

[0499]

[0500] In addition, the estimation department 240 can also formulate a disposal plan that achieves the target disposal effect as early as possible, a disposal plan that achieves the best situation on the final day of the period, or a continuous disposal plan that maintains the status quo in a way that does not exceed the target.

[0501] The output unit 250 outputs the estimation results performed by the estimation unit 240.

[0502] For example, the output unit 250 outputs the treatment plan, treatment effect, and variance value from the estimation unit 240. Specifically, the output unit 250 displays the treatment plan, treatment effect, and variance value.

[0503] The estimation device 200 described above, for example, can be used by... Figure 7 The PC10 shown is used to implement this.

[0504] For example, DB211 can be implemented through auxiliary storage device 11.

[0505] The program stored in the auxiliary storage device 11 is loaded into the memory 12, and the processor 13 executes the program, thereby enabling the input execution unit 112, the preprocessing unit 113, the encoder learning unit 220, the decoder learning unit 230, and the estimation unit 140 to be implemented.

[0506] The output unit 250 can be implemented through the display 16.

[0507] Figure 15 and Figure 16 This is a flowchart illustrating the operation of the estimation device 200 in Embodiment 2 described above.

[0508] First, the input execution unit 212 obtains {X#, V#, W#} from DB211 and provides them to the preprocessing unit 213 (S30).

[0509] Next, the preprocessing unit 213 deletes data with missing values ​​in X#, V# and W# (S31).

[0510] Next, the preprocessing unit 213 performs preprocessing of W# (S32).

[0511] Next, the preprocessing unit 213 performs normalization of X# (S33).

[0512] Next, the preprocessing unit 213 provides the {X, V, W} that has undergone the above processing to the encoder learning unit 220 (S34).

[0513] Next, the initialization unit 221 of the encoder learning unit 220 sets the dimension D of H. H (S35).

[0514] Next, the initialization unit 221 initializes the parameter J13 shown in equation (58) above (S36).

[0515] Next, the prediction unit 222 performs the calculation of the data J65 shown in the following formula (114) (S37).

[0516]

Mathematical Expression 114

[0517]

[0518] Next, the classification optimization unit 223 performs the calculation of the data J66 shown in the following formula (115) (S38).

[0519]

Mathematical Expression 115

[0520]

[0521] Next, the classification optimization unit 223 optimizes the parameter J67 shown in equation (116) below (S39).

[0522]

Mathematical Expression 116

[0523]

[0524] Next, the prediction optimization unit 224 performs the calculation of the data J68 shown in the following formula (117) (S40).

[0525]

Mathematical Expression 117

[0526]

[0527] Next, the prediction optimization unit 224 calculates the data J69 and data J70 shown in equation (118) and equation (119) below (S41).

[0528]

Mathematical Expression 118

[0529]

[0530]

Mathematical Expression 119

[0531]

[0532] Next, the prediction optimization unit 234 optimizes the parameter J30 shown in equation (77) above (S42).

[0533] Here, the prediction optimization unit 224 returns the process to step S37 and repeats each process until the values ​​J33 shown in equation (80) and J34 shown in equation (81) above, which are loss functions, are minimized, or until a predetermined number of repetitions is reached.

[0534] Next, the prediction optimization unit 234 provides the optimized parameter J35 shown in equation (82) above to the decoder learning unit 230 (S43). Then, processing begins. Figure 16 Step S44.

[0535] exist Figure 16 In step S44, the initialization unit 231 of the decoder learning unit 230 initializes the parameter J36 of the decoder unit shown in equation (83) using the parameter J35 optimized by the encoder unit as shown in equation (82).

[0536] Next, the prediction unit 232 performs the calculation of the data J71 shown in the following formula (120) (S45).

[0537]

Mathematical Expression 120

[0538]

[0539] Next, the classification optimization unit 233 performs the calculation of the data J72 shown in the following formula (121) (S46).

[0540]

Mathematical Expression 121

[0541]

[0542] Next, the classification optimization unit 233 optimizes the parameter J73 shown in equation (122) below (S47).

[0543]

Mathematical Expression 122

[0544]

[0545] Next, the prediction optimization unit 234 calculates the data J74 and data J75 shown in equations (123) and (124) below (S48).

[0546]

Mathematical Expression 123

[0547]

[0548]

Mathematical Expression 124

[0549]

[0550] Next, the prediction optimization unit 234 optimizes the parameter J51 shown in equation (100) above (S49).

[0551] Here, the prediction optimization unit 234 returns the process to step S45 and repeats each process until the value J54 shown in equation (103) and the value J55 shown in equation (104) above, which are loss functions, are minimized, or until a predetermined number of repetitions are reached.

[0552] Next, the prediction optimization unit 234 provides the parameter J35 optimized by the encoder learning unit 220 as shown in equation (82) above and the parameter J36 optimized by the decoder learning unit 230 as shown in equation (83) above to the estimation unit 240 (S50).

[0553] Next, the estimation unit 240 calculates the processing effect J60 of the processing plan J59 for any time t to τ shown in equation (82) above and the variance value J61 shown in equation (110) above (S51) based on the parameters J35 of the optimized encoder shown in equation (82) above and the parameters J36 of the optimized decoder shown in equation (83) above.

[0554] Next, the estimation unit 240 selects the treatment plan J59 (S52) shown in the above formula (108) that has the highest final treatment effect J62 shown in formula (111).

[0555] Next, the estimation unit 240 provides the output unit 250 with the optimal treatment plan J59, its treatment effect J62 and its variance value J61 shown in the above equations (108), (109) and (110).

[0556] Next, the output unit 250 displays the treatment plan, treatment effect, and variance value from the estimation unit 240 (S54).

[0557] Here, the application of Implementation Method 2 to railway services is as follows.

[0558] The primary variables are the amount of the freight rate increase, the amount of the freight rate decrease, and the amount of coupons used to promote railway use.

[0559] The second causal variable is at least one of the attributes and purchasing history of railway users, factors affecting the sales amount of railway services, and the behavioral history of railway users.

[0560] The types of transactions handled include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use.

[0561] Historical information includes the history of freight rate increases, freight rate decreases, and the distribution of coupons to promote railway use.

[0562] As described above, according to Implementation Method 2, the appropriate time and measure plan can be determined by considering the characteristics of the object. Furthermore, the reliability of the estimated treatment effect derived from the formulated measure plan can also be determined.

[0563] For example, such as Figure 9 As shown, in the case of four doses of vaccine, a distribution plan should be developed, including the combinations of treatments, the sequence of treatments, the timing of treatment, and when to stop treatment.

[0564] Furthermore, for example, it can be known at what time and from what period to increase or decrease the price of seats or rooms in dynamic pricing such as railway seats or hotel rooms, the number or combination of price increases and decreases, and the order in which price increases and decreases will increase the final sales volume.

[0565] Document 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

[0566] Label Explanation

[0567] 100, 200: Estimation unit; 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 information processing device, characterized in that, The information processing device has: The input section generates supervisory data containing causal variables, treatment types, and outcome variables for each of the multiple processes; and The learning unit uses the supervised data to learn the outcome variables for each of the plurality of processes, based on the causal variables and the type of processing, thereby generating a learning model.

2. The information processing device according to claim 1, characterized in that, The information processing device further includes an estimation unit that inputs the causal variable and the processing type into the learning model to estimate the outcome variable.

3. The information processing device according to claim 2, characterized in that, The estimation unit determines the optimal combination of the causal variable and the treatment type based on the estimated outcome variable.

4. The information processing apparatus according to any one of claims 1 to 3, characterized in that, The causal variables are at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, the railway user's behavior history, the amount of freight rate increases, the amount of freight rate decreases, and the amount of coupons used to promote the use of railway services. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The outcome variable is the sales revenue of high-priced railway services.

5. An information processing device, characterized in that, The information processing device has: The input unit, for each of the multiple processes, generates supervisory data containing a first causal variable that changes over time, a second causal variable that does not change over time, the type of process, and historical information representing the history of the time-varying nature of the first causal variable; and The learning unit uses the supervised data to learn the temporal changes of the first causal variable at a next time period based on the historical information, the first causal variable at a certain time, the second causal variable, and the type of the processing, thereby generating a learning model.

6. The information processing apparatus according to claim 5, characterized in that, The information processing device further includes an estimation unit that uses the learning model to estimate the time variation of the first causal variable by combining two or more processes selected from the plurality of processes within a certain period.

7. The information processing apparatus according to claim 6, characterized in that, The estimation unit determines the optimal combination of the two or more processes based on the estimated time changes.

8. The information processing apparatus according to any one of claims 5 to 7, characterized in that, The first causal variable is the amount of the freight rate increase, the amount of the freight rate decrease, and the amount of coupons used to promote railway use. The second causal variable is at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, and the railway user's behavioral history. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The historical information includes the history of freight rate increases, the history of freight rate decreases, and the history of the issuance of coupons to promote railway use.

9. An information processing device, characterized in that, The information processing device has: An estimation unit, for each of the multiple processes, inputs a cause variable and a processing type to a learning model, thereby estimating an outcome variable, wherein the learning model is generated in such a manner as follows: using supervised data containing the cause variable, the processing type, and the outcome variable, the outcome variable is learned for each of the multiple processes based on the cause variable and the processing type; as well as The output section outputs the estimated result.

10. The information processing apparatus according to claim 9, characterized in that, The causal variables are at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, the railway user's behavior history, the amount of freight rate increases, the amount of freight rate decreases, and the amount of coupons used to promote the use of railway services. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The outcome variable is the sales revenue of high-priced railway services.

11. An information processing device, characterized in that, The information processing device has: The estimation unit, for each of a plurality of treatments, uses a learning model to estimate the time variation of a first causal variable that changes over time, obtained by combining two or more treatments selected from the plurality of treatments within a certain period. The learning model is generated in such a manner as follows: using supervised data containing the first causal variable, a second causal variable that does not change over time, treatment type, and historical information representing the history of the time variation of the first causal variable, for each of the plurality of treatments, learning the time variation of the first causal variable for the next time period based on the historical information, according to the first causal variable, the second causal variable, and the treatment type at a certain time; and The output section outputs the estimated result.

12. The information processing apparatus according to claim 11, characterized in that, The first causal variable is the amount of the freight rate increase, the amount of the freight rate decrease, and the amount of coupons used to promote railway use. The second causal variable is at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, and the railway user's behavioral history. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The historical information includes the history of freight rate increases, the history of freight rate decreases, and the history of the issuance of coupons to promote railway use.

13. A program, characterized in that, This program enables the computer to function as the following parts: The input section generates supervisory data containing causal variables, treatment types, and outcome variables for each of the multiple processes; and The learning unit uses the supervised data to learn the outcome variables for each of the plurality of processes, based on the causal variables and the type of processing, thereby generating a learning model.

14. The procedure according to claim 13, characterized in that, The causal variables are at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, the railway user's behavior history, the amount of freight rate increases, the amount of freight rate decreases, and the amount of coupons used to promote the use of railway services. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The outcome variable is the sales revenue of high-priced railway services.

15. A program, characterized in that, This program enables the computer to function as the following parts: The input unit, for each of the multiple processes, generates supervisory data containing a first causal variable that changes over time, a second causal variable that does not change over time, the type of process, and historical information representing the history of the time-varying nature of the first causal variable; and The learning unit uses the supervised data to learn the temporal changes of the first causal variable at a next time period based on the historical information, the first causal variable at a certain time, the second causal variable, and the type of the processing, thereby generating a learning model.

16. The procedure according to claim 15, characterized in that, The first causal variable is the amount of the freight rate increase, the amount of the freight rate decrease, and the amount of coupons used to promote railway use. The second causal variable is at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, and the railway user's behavioral history. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The historical information includes the history of freight rate increases, the history of freight rate decreases, and the history of the issuance of coupons to promote railway use.

17. A program, characterized in that, This program enables the computer to function as the following parts: An estimation unit, for each of the multiple processes, inputs a cause variable and a processing type to a learning model, thereby estimating an outcome variable, wherein the learning model is generated in such a manner as follows: using supervised data containing the cause variable, the processing type, and the outcome variable, the outcome variable is learned for each of the multiple processes based on the cause variable and the processing type; as well as The output section outputs the estimated result.

18. The procedure according to claim 17, characterized in that, The causal variables are at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, the railway user's behavior history, the amount of freight rate increases, the amount of freight rate decreases, and the amount of coupons used to promote the use of railway services. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The outcome variable is the sales revenue of high-priced railway services.

19. A program, characterized in that, This program enables the computer to function as the following parts: The estimation unit, for each of a plurality of treatments, uses a learning model to estimate the time variation of a first causal variable that changes over time, obtained by combining two or more treatments selected from the plurality of treatments within a certain period. The learning model is generated in such a manner as follows: using supervised data containing the first causal variable, a second causal variable that does not change over time, treatment type, and historical information representing the history of the time variation of the first causal variable, for each of the plurality of treatments, learning the time variation of the first causal variable for the next time period based on the historical information, according to the first causal variable, the second causal variable, and the treatment type at a certain time; and The output section outputs the estimated result.

20. The procedure according to claim 19, characterized in that, The first causal variable is the amount of the freight rate increase, the amount of the freight rate decrease, and the amount of coupons used to promote railway use. The second causal variable is at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, and the railway user's behavioral history. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The historical information includes the history of freight rate increases, the history of freight rate decreases, and the history of the issuance of coupons to promote railway use.

21. An information processing method, characterized in that, For each of the multiple treatments, supervisory data is generated, including causal variables, treatment types, and outcome variables. Using the supervised data, for each of the plurality of processes, the outcome variable is learned based on the causal variable and the type of process, thereby generating a learning model.

22. The information processing method according to claim 21, characterized in that, The causal variables are at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, the railway user's behavior history, the amount of freight rate increases, the amount of freight rate decreases, and the amount of coupons used to promote the use of railway services. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The outcome variable is the sales revenue of high-priced railway services.

23. An information processing method, characterized in that, For each of the multiple processes, supervisory data is generated, which includes a first causal variable that changes over time, a second causal variable that does not change over time, the process type, and historical information representing the history of the time-varying nature of the first causal variable. Using the supervised data, for each of the plurality of processes, based on the historical information, the learning model is generated by learning the temporal changes of the first causal variable at the next time based on the first causal variable, the second causal variable, and the processing type at a certain time.

24. The information processing method according to claim 23, characterized in that, The first causal variable is the amount of the freight rate increase, the amount of the freight rate decrease, and the amount of coupons used to promote railway use. The second causal variable is at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, and the railway user's behavioral history. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The historical information includes the history of freight rate increases, the history of freight rate decreases, and the history of the issuance of coupons to promote railway use.

25. An information processing method, characterized in that, For each of the multiple processes, a cause variable and a process type are input into the learning model to estimate an outcome variable. The learning model is generated in the following manner: using supervised data containing the cause variable, the process type, and the outcome variable, the outcome variable is learned for each of the multiple processes based on the cause variable and the process type. Output the estimated result.

26. The information processing method according to claim 25, characterized in that, The causal variables are at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, the railway user's behavior history, the amount of freight rate increases, the amount of freight rate decreases, and the amount of coupons used to promote the use of railway services. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The outcome variable is the sales revenue of high-priced railway services.

27. An information processing method, characterized in that, For each of the multiple treatments, a learning model is used to estimate the time variation of a first causal variable that changes over time, obtained by combining two or more treatments selected from the multiple treatments within a certain period. The learning model is generated as follows: using supervised data containing the first causal variable, a second causal variable that does not change over time, treatment type, and historical information representing the history of the time variation of the first causal variable, for each of the multiple treatments, the time variation of the first causal variable for the next time period is learned based on the historical information, according to the first causal variable, the second causal variable, and the treatment type at a certain time. Output the estimated result.

28. The information processing method according to claim 27, characterized in that, The first causal variable is the amount of the freight rate increase, the amount of the freight rate decrease, and the amount of coupons used to promote railway use. The second causal variable is at least one of the railway user's attributes and purchase history, factors affecting the sales amount of railway services, and the railway user's behavioral history. The types of processing include increases in freight rates, decreases in freight rates, activities to promote railway use, and the issuance of coupons to promote railway use. The historical information includes the history of freight rate increases, the history of freight rate decreases, and the history of the issuance of coupons to promote railway use.

Images