Model learning device, method and program

Abstract

One embodiment of the present invention acquires training data that includes grouping uncouple data and grouping magnitude comparison data that are acquired from a plurality of groups being examined. First, a process for updating a hyperparameter using a first optimization method is executed on the acquired grouping uncouple data, and an optimization hyperparameter with which a first objective function becomes minimized is estimated. Next, a process for updating a parameter using a second optimization method is executed on the basis of the estimated optimization hyperparameter and the acquired grouping uncouple data and grouping magnitude comparison data, and an optimization parameter with which a second objective function becomes minimized is estimated. Finally, the estimated optimization parameter is outputted.

Description

[Technical field]

[0001] One aspect of the present invention relates to a model learning device, a method, and a program for performing model learning using uncoupled data. [Background technology]

[0002] Learning a model that expresses input-output relationships from data is one of the typical problems in the fields of machine learning and artificial intelligence. In this problem, data that is usually given as a set of input-output value pairs that show what the output value was for a certain input value, that is, data with input-output correspondence, is used to learn the model.

[0003] However, in recent years, for example, due to the influence of data collection methods or privacy protection processing, there are an increasing number of situations where it is necessary to estimate parameters and learn a model that represents the input-output relationship based on so-called uncoupled data, where the input-output relationship is not known.One example can be considered as a case where information indicating a user's basic attributes (gender, age, etc.) and life patterns (average wake-up time, average weekly exercise time, etc.) is input, and the user's annual income is estimated based on this input information.

[0004] In normal model training using input-output correspondence data, i is an index representing a user, x i Let be the input value (attributes and lifestyle patterns) of user i, and y i is the output value (annual income) of user i, the data represented as a set of input and output pairs {x i , y i} n i=1 We estimate the model parameters using,where,n,represents the total number of users.

[0005] In contrast, in model training using uncoupled data, the set of inputs {xm} n′X m=1 and the set of outputs {ym′} n′Y m′=1and are provided separately without being associated with each other. Here, n' X and n′ Y represents the number of data for each data item, but because there are users who do not answer the output value, n′ X and n′ Y In general, these data are not equivalent to input and output. For example, if the input value is x m The output value of the user who answered is { y 1 , y 2 , … , y n′Y Uncoupled data is created when sensitive data such as "annual income" is collected in a way that does not link the output value to the user, for example, from the viewpoint of privacy protection.

[0006] As existing techniques for model learning using uncoupled data, for example, the method described in Non-Patent Document 1 and the method described in Non-Patent Document 2 are known. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] A. Carpentier and T. Schlu¨ter. “Learning relationships between data obtained independently.” In Artificial Intelligence and Statistics, pp. 658 - 666, 2016. [Non-Patent Document 2] Liyuan Xu, Gang Niu, Junya Honda, and Masashi Sugiyama. “Uncoupled regression from pairwise comparison data.” In Proceedings of the 33rd International Conference on Neural Information Processing Systems, pp. 3992 - 4002, 2019. Summary of the Invention [Problem to be solved by the invention]

[0008] However, the technique described in Non-Patent Document 1 requires satisfying conditions that are difficult to use in practice, and is therefore unsuitable for practical use.

[0009] On the other hand, the method described in Non-Patent Document 2 is a model learning method under practical conditions that use size comparison data in addition to uncoupled data, and is expected to be used. + m ,x- m )} n′c m=1 The data is given in the form: n' C represents the number of data. This data is input value x + m The output value of the user who answered "yes" is x - m This indicates that the output value of a user who answered "yes" is greater than the output value of a user who answered "yes," and does not mean that the output value itself was observed. This type of data can be obtained by having a user answer whether their output value (e.g., annual income) is greater than the output value of another user. This type of data places a smaller psychological burden on the user than answering the actual annual income, such as 3 million yen or 5 million yen, and is often easier to collect.

[0010] However, Non-Patent Document 2 does not consider the case where uncoupled data are grouped. That is, in the field of data analysis, data is generally collected multiple times (n K In this case, the available data is grouped and uncoupled data.

[0011] Specifically, the total number of users (depending on which data collection session they participated in) is n K The data is divided into n groups, and like the uncoupled data mentioned above, there is no correspondence between the input and the output, but it is possible to know which group the user's answer belongs to. In other words, data collection is K After the first run, the available training data is Set of input values ​​D X = {D Xk} nK k=1 = {{x km}n Xk m=1}n K k=1 Output value set D Y = {D Yk} nK k=1 = {{y km}n Yk m=1}n K k=1 is given as:

[0012] However, x km is the input value of one of the users belonging to the kth group, y km represents the output value of any user belonging to the kth group, and n Xk is the number of input values ​​in the kth group, n Yk represents the number of output values ​​in the kth group. X ,n Y Let n denote the total number of data. X ,n Y are respectively n X= Σ nk k=1 n Xk n Y = Σ nk k=1 n Yk It is defined as follows:

[0013] Considering such a practical situation, a technique is needed that allows model learning even when uncoupled data are grouped.

[0014] The present invention has been made in light of the above-mentioned circumstances, and aims to provide a technique that enables highly accurate model learning even when grouped uncoupled data is used. [Means for solving the problem]

[0015] In order to solve the above problems, one aspect of the model learning device or model learning method according to the present invention is to provide a method for learning a plurality of groups to be surveyed. each Obtained 、 Learning data including grouped uncoupled data and grouped magnitude comparison data is acquired. Then, first, a process of updating hyperparameters is executed for the acquired grouped uncoupled data using a first optimization method, and optimized hyperparameters that minimize a first objective function are estimated. Next, All of the groups Grouping Uncoupled Data and All of the groups Executing a process of updating parameters using a second optimization method based on the grouped magnitude comparison data and the estimated optimized hyperparameters; The grouped uncoupled data and the grouped magnitude comparison data of all groups are included. The optimization parameters that minimize the second objective function are estimated, and finally, the estimated optimization parameters are output. Effect of the Invention

[0016] According to one aspect of the present invention, a technology can be provided that enables highly accurate model learning while satisfying practical conditions by utilizing grouped magnitude comparison data in addition to uncoupled data that has been grouped. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 is a block diagram showing an example of a hardware configuration of a model learning device according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a block diagram showing an example of a software configuration of the model learning device according to an embodiment of the present invention. [Diagram 3] FIG. 3 is a flowchart showing an example of a processing procedure and processing contents of the parameter estimation processing executed by the model learning device shown in FIG. [Figure 4] FIG. 4 is a flowchart showing an example of the processing procedure and processing contents of the hyper-parameter estimation processing from the processing procedures shown in FIG. [Diagram 5] FIG. 5 is a flowchart showing an example of the procedure and contents of the parameter estimation process from the procedure shown in FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0019] [One embodiment] In one embodiment of the present invention, a method is used to learn a model using grouped uncoupled data and magnitude comparison data, and this method will hereinafter be referred to as Grouped Uncoupled Regression (GUR).

[0020] (Configuration example) 1 and 2 are block diagrams showing an example of a hardware configuration and a software configuration, respectively, of a model learning device according to an embodiment of the present invention.

[0021] The model learning device ML is, for example, a server computer or a personal computer. The model learning device ML includes a control unit 1 using a hardware processor such as a central processing unit (CPU), and a storage unit having a program storage unit 2 and a data storage unit 3, and an input / output interface (hereinafter, the interface will be referred to as I / F) unit 4 are connected to the control unit 1 via a bus 5. The model learning device ML may also include a communication I / F unit, etc.

[0022] An external device EX that performs data analysis processing and the like is connected to the input / output I / F unit 4 via a signal cable or a network. The input / output I / F unit 4 is used to receive learning data used for model learning from the external device EX and to output parameters estimated by model learning to the external device EX.

[0023] The program memory unit 2 is configured, for example, by combining a non-volatile memory such as an HDD (Hard Disk Drive) or SSD (Solid State Drive) as a storage medium that can be written to and read at any time, and a non-volatile memory such as a ROM (Read Only Memory), and stores various programs necessary for executing various control processes according to one embodiment of the present invention, in addition to middleware such as an OS (Operating System).

[0024] The data storage unit 3 is configured, for example, by combining a non-volatile memory such as an HDD or SSD, which can be written to and read from at any time, as a storage medium, with a volatile memory such as a RAM (Random Access Memory), and is provided with an input data storage unit 31, a hyper parameter storage unit 32, and a parameter storage unit 33 as storage areas necessary for implementing one embodiment of the present invention.

[0025] The input data storage unit 31 is used to store the learning data received from the external device EX.

[0026] The hyper-parameter storage unit 32 is used to temporarily store hyper-parameters estimated by a hyper-parameter estimation process by the control unit 1, which will be described later, for the parameter estimation process.

[0027] The parameter storage unit 33 is used to temporarily store parameters estimated by a parameter estimation process performed by the control unit 1, which will be described later, until the parameters are output to the external device EX.

[0028] The control unit 1 includes, as processing functions according to an embodiment of the present invention, a data acquisition processing unit 11, a hyper parameter estimation processing unit 12, a parameter estimation processing unit 13, and a parameter output processing unit 14. Each of these processing units 11 to 14 is realized by causing a hardware processor of the control unit 1 to execute an application program stored in the program storage unit 2. Note that the application program does not need to be stored in advance in the program storage unit 2, and may be downloaded, for example, from an external device EX or another server device when necessary.

[0029] The data acquisition processing unit 11 performs processing to acquire learning data used for model learning sent from the external device EX via the input / output I / F unit 4, and store the acquired learning data in the input data storage unit 31. The learning data includes grouped uncoupled data and size comparison data. Among these, the size comparison data is data acquired by, for example, asking surveyed users in a questionnaire whether their output value is larger or smaller than the output value of another user.

[0030] The hyper-parameter estimation processing unit 12 reads the grouped uncoupling data from the input data storage unit 31, and performs a hyper-parameter update process on the read uncoupling data using, for example, a subgradient method. Then, when the number of repetitions of this update process exceeds a predetermined number of times or the change range before and after the update becomes smaller than a threshold, the hyper-parameters after the update are stored in the hyper-parameter storage unit 32.

[0031] The parameter estimation processing unit 13 reads the input uncoupled data and magnitude comparison data from the input data storage unit 31, and also reads the updated hyper-parameters from the hyper-parameter storage unit 32. Then, the parameter estimation processing unit 13 performs processing to estimate parameters that minimize the objective function using, for example, a gradient method, and stores the estimated parameters in the parameter storage unit 33.

[0032] After the model learning process is completed, the parameter output processing unit 14 reads out the estimated parameters from the parameter storage unit 33 and transmits the read parameters from the input / output I / F unit 4 to the external device EX.

[0033] (Example of operation) Next, an example of the operation of the model learning device ML configured as above will be described. (1) Overview of operation (1-1) Formulation First, we explain the problem setting of GUR. GUR uses grouped uncoupled data D X ,D Y In addition, we use the grouped size comparison data D C D C = {D Ck} nK k=1 = {{ (x + km , x - km )} nCk m=1} nK k=1 Use the following.

[0034] However, x + km ,x - km Both represent the input value of any user belonging to the k-th group, and x + km The output value of the user who answered is x - km This indicates that the output value is greater than the output value of the user who answered "n Ck represents the number of data for size comparison in the k-th group. Also, let the total number of size comparison data be n C Let n C =P nK k=1 n Ck It is defined as:

[0035] As will be described later, uncoupled data D Y Even if the probability distribution of the output values ​​of each group is not available, it is possible to train a model using GUR if information about the probability distribution of the output values ​​of each group is available.

[0036] (1-2) Loss function The loss function is defined as the Bregman Divergence (BD) d, which is defined as φ Use d φ is expressed as follows: d φ (x, y) = φ(x) - φ(y) - (x -y ) φ(y) Note that φ is a convex function, and ψ is its first derivative ψ = ∇φ. By changing the function φ, BD can express various functions. For example, φ(x) =x 2 When φ(x) = x log(x) + (1-x) log(1-x), it corresponds to the logistic loss; φ(x) = x log(x) corresponds to the I-divergence (also known as the generalized KL divergence); and φ(x) = -log(x) corresponds to the Itakura-Saito pseudodistance.

[0037] Determining the loss function to be used is equivalent to making an assumption about the probability distribution of the generated data. Specifically, using squared error, I-divergence, and Itakura-Saito pseudodistance correspond to the assumption that the data is generated according to normal distribution, Poisson distribution, and exponential distribution, respectively.

[0038] To define the loss function, we define symbols. The random variable whose realization value corresponds to the group index is written as K, the random variable corresponding to the input value is written as X, and the random variable corresponding to the output value is written as Y. We write all possible values ​​of the input value and the output value as Xall and Yall, respectively. The probability distribution that these random variables follow is P K ,X ,Y This probability distribution P K ,X ,Y The distribution of P with respect to K is X ,Y , and the conditional probability distribution and probability density function when K=k are respectively P X,Y|k ,f X,Y|k Let us write this conditional distribution as P Y|k ,P X|k We write it as follows. Probability distribution P Y|k The probability density function of f Y|k , the cumulative density function F Y|k With the above definitions, the cumulative density function F Y|k teeth, F Y|k (y) = ∫ y -∞ f Y|k (y′)dy′ It is.

[0039] A learning model belonging to the hypothesis space H is written as h: Xall →Yall. In the model learning according to the present invention, the learning model to be used is not limited. For example, the present invention uses a linear model (hypothesis space H={ h(x) =θ T x|θ∈R d}) and nonlinear models including deep learning and kernel methods. The loss function used to train the model using the random variables above is the expected value of the Bregman divergence,

number

[0040] However, E K,X,Y [·] is the probability distribution P X,Y This is the expected value by R k teeth,

number

[0041] In equation (2), E Y|k ,E X|k ,E X,Y|k are probability distributions P Y|k ,P X|k ,P X,Y|k The difficulty in evaluating this loss function is that R k The final term of E X,Y|k [Y ψ(h(X))]. This is because this term is defined by the joint distribution of random variables X and Y that represent the input and output values, but the problem setting in this invention uses uncoupled data in which the input and output values ​​are not observed simultaneously, so it cannot be calculated even if sample approximation is performed. Therefore, from here on, we consider approximately evaluating this term.

[0042] That is, a new set of random variables (X + ,X - ) is introduced. This is because, with a certain group k fixed, + The output value of input X - is greater than the output value of

number

[0043] However, (X,Y) and (X′,Y′) are both probability distributions P X,Y|k From this definition, the above-mentioned size comparison data D Ck We will later use the fact that we can consider X to be the realization of this random variable. + and X - Let f be the probability density function that follows. X+|k and f X-|k , and X + and X - The expected value operation regarding the outcome of E X+|k [·] ,E X-|k Let us write it as [·]. Using this random variable,

number

[0044] Below, we will prove the above equations (3) and (4). That is, X + From the definition of f X+|k teeth,

number

[0045] Here, Z is a normalization constant, and by partial integration, we can derive that Z = 1 / 2. Using this,

number

[0046] Using equations (3) and (4), the probability density function f Yk is uniformly distributed on [0,1], i.e., F Y|k When (y) = y, the last term of equation (2) becomes E X,Y|k [Yψ(h(X))] = EX+|k [ψ(h(X + ))] / 2 It can be transformed as follows.

[0047] Considering this fact, a certain hyperparameter w k1 ,w k2 Using ∈ R,

number

[0048] As above, Y|k If is a uniform distribution on [0,1], then (w k1 ,w k2 This approximation is accurate if we set θ = (1 / 2,0), which is the case for a uniform distribution on [a,b]. F Y|k (y) = (ya) / (ba) for all y∈[a, b] This can also be generalized to (w k1 ,w k2 ) = (b / 2,a / 2). If we consider a more general distribution than a uniform distribution, we set the hyperparameters (w k1 ,w k2 ), which will be discussed later.

[0049] Adding together equations (3) and (4), we get

number

[0050] Using this, a certain constant λ k Using equation (5),

number

number

[0051] Therefore, the random variables K, X, and X + ,X - By replacing the expectation with the sample mean,

number

[0052] Here, C is a constant that does not depend on the model h. Since this quantity can be calculated from the data excluding the constant C, it can be used as an objective function for parameter estimation. Therefore, the model can be trained by optimizing the following objective function L, which is obtained by removing the constant C from the empirical loss R^.

number

[0053] Any method can be used for optimization, such as gradient methods, (pseudo)Newton methods, stochastic gradient methods, Adam, etc. For example, when training a model with a parameter θ, the following optimization process is performed using gradient methods:

number

[0054] In addition, the following objective function L^, which is an approximation of the objective function L, can be used to train the model.

number

[0055] The above objective function L^ is the output value y ~ km Input value x km The input and output values ​​are considered to correspond to the pseudo-values ​​{x km ,y ~ km}n Xk m=1}n K k=1 This is equivalent to the objective function used when learning a model using L^. However, the constant term is excluded. Therefore, to optimize the objective function L^, the model learning method used when using data with a correspondence between input and output values ​​can be applied as is.

[0056] In other words, when estimating the optimization parameters, the above hyperparameters w k1 ,w k2 The value calculated based on the above may be regarded as a pseudo-output value corresponding to the input value, and a process of updating the parameters may be performed using an objective function L^ that approximates the objective function L, thereby estimating the optimization parameter θ that minimizes the objective function L^.

[0057] (1-3) Hyperparameter Estimation Finally, the hyperparameters {w k1 ,w k2 We will now describe the estimation method for}. This hyperparameter is

number

[0058] However, F Y|k ^ is the cumulative density function F Y|k Empirical approximation of

number

number

[0059] Function E rrk The probability density function f Y|k and its cumulative density function F Y|k Since is generally unknown, the function E rrk However, the function E rrk ^ is data D Yk and optimization can be performed.

[0060] Any optimization method can be used for the optimization process. For example, since equation (9) is defined as the sum of absolute values, it is desirable to use a method that can handle even non-differentiable points in the objective function, such as the subgradient method, or a linear programming method. When using the subgradient method, the function E rrk ^w k =(w k1 ,w k2 ) the set of subgradients in ∂E rrk ^(w k1 ,w k2 ) using any vector g belonging to

number

[0061] Also, as is clear from the above discussion, data D Y Even if the probability density function {f Y|k} nK k=1 If prior knowledge about E rrk We can estimate the hyperparameters by directly minimizing E rrk contains integrals, so for example, we can discretely divide all possible values ​​of y into {y L} nsplit L=1 For example, the 0.01 quantile of y is y and the 0.99 quantile is y. y L =y+(L-1) / n split (y ￣ -y) Then, by minimizing the following equation (11), E rrk ^ It can be estimated using any optimization method, just like .

number

[0062] (2) Operation of the model learning device ML FIG. 3 is a flowchart showing the procedure and contents of the model learning process executed by the control unit 1 of the model learning device ML.

[0063] (2-1) Acquiring learning data In step S1, the control unit 1 of the model learning device ML monitors the input of learning data from the external device EX. In this state, when learning data is sent from the external device EX, the control unit 1 of the model learning device ML, under the control of the data acquisition processing unit 11, receives the learning data sent from the external device EX via the input / output I / F unit 4 in step S2, and stores the received learning data in the input data storage unit 31.

[0064] The input learning data is the grouped uncoupled data D X ,D YAnd grouped size comparison data D C Among these, the size comparison data D C is the data obtained by asking the surveyed users whether their output value is greater or smaller than the output value of another user. As shown in the formulation of (1-1) above, D C = {D Ck} nK k=1 = {{ (x + km , x - km )} nCk m=1} nK k=1 This is expressed as:

[0065] However, x + km ,x - km Both represent the input value of any user belonging to the k-th group, and x + km The output value of the user who answered is x - km This indicates that the output value is greater than the output value of the user who answered "n Ck represents the number of data points for comparison in the kth group.

[0066] (2-2) Hyperparameter Estimation When the learning data is acquired, the control unit 1 of the model learning device ML, under the control of the hyper-parameter estimation processing unit 12, first reads the uncoupled data D from the input data storage unit 31 in step S3. Y Then, the uncoupled data D Y For each group k=1,...,nk, an update process using the subgradient method described below is performed to find the hyperparameter w by minimizing the objective function shown in equation (9) above.

[0067] FIG. 4 is a flowchart showing an example of the processing procedure and processing contents of the hyper-parameter update processing using the subgradient method.

[0068] That is, the hyper-parameter estimation processing unit 12 first calculates the hyper-parameter w k1 ,w k2 When this initialization process is completed, the hyper-parameter estimation unit 12 then initializes a variable δ in step S42. This variable δ is a variable used as a termination condition, and indicates the maximum change width of the update amount. Additionally, in step S42, the hyper-parameter estimation unit 12 sets a threshold ε and a maximum number of iterations C as termination conditions. Values ​​indicating these termination conditions are stored in advance in a variable storage area of ​​the data storage unit 3.

[0069] Next, in step S43, the hyper-parameter estimation processing unit 12 updates the hyper-parameter w according to the above-mentioned formula (10). k1 ,w k2 Each time the update process is performed, the hyperparameters w before and after the update are k The maximum absolute value of the difference max(|w old k1 -w new k1 |,|w old k2 -w new k2 |) is set as the variable δ. Note that the hyperparameter w k The elements of w old k1 ,w old k2 , the updated element is w new k1 ,w new k2 Each is described as follows.

[0070] Next, in step S44, the hyper-parameter estimation processor 12 updates the number of update repetitions C.

[0071] The hyper-parameter estimation processing unit 12 estimates the hyper-parameter w k1 ,w k2 Each time the update process is performed once, it is determined in step S45 whether or not the termination condition is satisfied. In this example, it is determined whether the update repetition count C has exceeded a preset maximum value Cmax or the variable δ has become smaller than a threshold value ε. If the result of this determination is that the repetition count C has not exceeded the maximum value Cmax and the variable δ has not become smaller than the threshold value ε, the hyperparameter estimation processor 12 returns to step S42 to initialize the variable δ to 0, and then performs the update process in steps S43 to S45 again. This update process is repeatedly performed until the termination condition is satisfied.

[0072] On the other hand, if the number of update iterations C exceeds the maximum value C max or the variable δ becomes smaller than the threshold ε, the hyper-parameter estimation unit 12 terminates the update process and estimates the finally obtained hyper-parameter w k1 ,w k2 is stored in the hyper-parameter storage unit 32.

[0073] (2-3) Parameter Estimation When the hyper-parameter estimation process is completed, the control unit 1 of the model learning device ML then, under the control of the parameter estimation processing unit 13, first reads the uncoupled data D from the input data storage unit 31 in step S5. X and size comparison data D C At the same time, the estimated hyper-parameter w k1 ,w k2 Then, the parameter estimation processor 13 executes an update process using the gradient method described below to minimize the objective function of the above-mentioned equation (7), thereby finding the optimal parameter θ.

[0074] FIG. 5 is a flowchart showing an example of the procedure and contents of the parameter update process using the gradient method.

[0075] That is, the parameter estimation processing unit 13 first initializes a parameter θ in step S51. Then, in step S52, a variable δ indicating the maximum change width of the update amount, which is one of the variables used as a termination condition, is similarly initialized, and further a threshold ε and a maximum number of iterations C indicating the termination condition are set. These values ​​indicating the termination condition are stored in advance in a variable storage area of ​​the data storage unit 3.

[0076] Next, in step S53, the parameter estimation processing unit 13 updates the parameter θ according to the above-mentioned formula (8). In addition, every time the parameter θ is updated, the parameter θ ∈ R before and after the update is calculated. d The maximum absolute value of the difference max d |θ old d -θ new d | is set as the variable δ. Note that here, the elements of the parameter θ before the update are set as θ old d , the updated element is θ new d Each is described as follows.

[0077] The parameter estimation processor 13 then updates the update repetition count C in step S54.

[0078] The parameter estimation processor 13 judges in step S55 whether or not the termination condition of the update process is satisfied each time the update process of the parameter θ is performed once. In this example, it judges whether the update repetition count C exceeds a preset maximum value C max or the variable δ becomes smaller than a threshold value ε. If the result of this judgment is that the repetition count C does not exceed the maximum value C max and the variable δ is not yet smaller than the threshold value ε, the parameter estimation processor 13 returns to step S52 to initialize the variable δ to 0, and then executes the update process of steps S53 to S55 again. This update process is repeatedly executed until the termination condition is satisfied.

[0079] On the other hand, if the update repetition count C exceeds the maximum value C max or the variable δ becomes smaller than the threshold ε, the parameter estimation processing unit 13 ends the update process of the parameter θ and stores the finally obtained parameter θ in the parameter storage unit 33.

[0080] (2-4) Output of parameter θ When the series of model learning processes is completed, in step S6, the control unit 1 of the model learning device ML, under the control of the parameter output processing unit 14, reads out the estimated parameter θ from the parameter storage unit 33, and transmits the read parameter θ from the input / output I / F unit 4 to the external device EX.

[0081] The external device EX creates a learning model using the parameter θ received from the model learning device ML, and thereafter uses this learning model to execute, for example, data analysis processing relating to consumers.

[0082] (Action and effect) As described above, in the model learning device ML according to one embodiment, the grouped uncoupled data D X ,D Y And grouped size comparison data D C Then, the uncoupled data D Y For each group, the process of updating the hyperparameters w is repeated using the subgradient method, which is one of the optimization methods, to obtain the optimized hyperparameters w that minimize the objective function. k1 ,w k2 Next, the above uncoupled data D X and size comparison data D C and the optimized hyperparameter w k1 ,w k2 Based on this, the process of updating the parameter θ is repeatedly performed using the gradient method, which is one of the optimization methods, to find the optimization parameter θ that minimizes the objective function, and the found optimization parameter θ is output.

[0083] Therefore, by using grouped magnitude comparison data in addition to the uncoupled data, it is possible to perform highly accurate model learning while satisfying practical conditions, even for grouped uncoupled data.

[0084] [Other embodiments] (1) In the above embodiment, a subgradient method is used for estimating the hyper-parameter w, and a gradient method is used for estimating the parameter θ. However, the present invention is not limited to this. For example, a linear programming method may be used for estimating the hyper-parameter w, and a (quasi-)Newton method, a stochastic gradient method, an Adam method, or the like may be used for estimating the parameter θ. In short, any method may be used for estimating the hyper-parameter w and optimizing the parameter θ.

[0085] (2) In the embodiment, the hyper-parameter estimation process uses uncoupled data D Y Instead, we use the probability density function {f Y|k} nK k=1 For example, the information of the output value D may be used for the obtained grouped uncoupled data. Y If not, the probability density function {f Y|k} nK k=1 The information is obtained, and this probability density function is used as the output value data D Y Instead of and minimize the objective function using equation (11). In short, any method can be used to optimize the hyperparameters.

[0086] (3) In the above embodiment, the model learning device ML is provided as a device separate from the external device EX. However, the present invention is not limited to this. The functions of the model learning device ML may be provided in the external device EX, and the external device EX may be configured to execute the model learning process.

[0087] In addition, the functional configuration of the model learning device and the processing procedures and processing contents of the model learning process can be modified in various ways without departing from the spirit and scope of the present invention.

[0088] Although the embodiment of the present invention has been described in detail above, the above description is merely an example of the present invention in every respect. It goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. In other words, in implementing the present invention, specific configurations according to the embodiment may be appropriately adopted.

[0089] In short, this invention is not limited to the above-mentioned embodiment as it is, and in the implementation stage, the components can be modified and embodied without departing from the gist of the invention. In addition, various inventions can be formed by appropriately combining multiple components disclosed in the above-mentioned embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, components from different embodiments may be appropriately combined. [Explanation of symbols]

[0090] ML: Model learning device EX…External device 1...Control unit 2...Program memory section 3. Data storage section 4... Input / Output I / F section 5. Bus 11...Data acquisition processing section 12...Hyperparameter estimation processing unit 13...Parameter estimation processing unit 14... Parameter output processing section 31... Input data storage unit 32...Hyperparameter storage unit 33...Parameter storage section

Claims

1. a data acquisition processing unit that acquires learning data including grouped uncoupled data and grouped magnitude comparison data acquired from each of a plurality of groups to be surveyed; A first estimation processing unit that executes a process of updating hyper-parameters using a first optimization method for the acquired grouped uncoupled data and estimates optimized hyper-parameters that minimize a first objective function; a second estimation processing unit that executes a process of updating parameters using a second optimization method based on the acquired grouped uncoupled data of all of the groups, the grouped magnitude comparison data of all of the groups, and the estimated optimization hyper-parameters, and estimates optimization parameters that minimize a second objective function including the grouped uncoupled data and the grouped magnitude comparison data of all of the groups; an output processing unit that outputs the estimated optimization parameters; A model learning device comprising:

2. 2. The model learning device according to claim 1, wherein the first estimation processing unit determines, each time an update process of the hyper-parameter is performed, whether an amount of change in the hyper-parameter before and after the update is less than a preset first threshold value or whether a number of times the update process has been performed exceeds a preset first number of times, and when the amount of change is less than the first threshold value or the number of times the update process has been performed exceeds the first number of times, terminates the update process and sets the hyper-parameter obtained at that time as the optimized hyper-parameter.

3. 2. The model learning device according to claim 1, wherein the second estimation processing unit determines, each time the parameter update process is performed, whether an amount of change in the parameter before and after the update becomes less than a predetermined second threshold value or whether the number of times the update process has been performed exceeds a predetermined second number of times, and when the amount of change becomes less than the second threshold value or the number of times the update process has been performed exceeds the second number of times, terminates the update process and sets the parameter obtained at that time as the optimized parameter.

4. 2. The model learning device according to claim 1, wherein the first estimation processing unit determines whether the acquired grouped uncoupled data includes output value data, and if the output value data is not included, obtains information regarding a probability distribution of the output value data of the multiple groups, and executes a process of estimating the optimized hyperparameters by using the information regarding the probability distribution instead of the output value data.

5. the first estimation processing unit uses a subgradient method or a linear programming method as the first optimization method; the second estimation processing unit uses a gradient method, a (pseudo) Newton method, a stochastic gradient method, or an Adam method as the second optimization method; The model learning device according to claim 1 .

6. 2. The model learning device according to claim 1, wherein the second estimation processing unit regards a value calculated based on the hyper-parameters estimated by the first estimation processing unit as a pseudo output value corresponding to an input value, and executes a process of updating the parameters using a third objective function equivalent to an objective function used when learning a model using data in which the input value and the output value correspond to each other, and estimates the optimization parameters that minimize the third objective function.

7. A model learning method executed by an information processing device, comprising: A step of acquiring learning data including grouped uncoupled data and grouped magnitude comparison data acquired from each of a plurality of groups to be surveyed; A process of updating hyperparameters using a first optimization method for the acquired grouped uncoupled data, and estimating optimized hyperparameters that minimize a first objective function; a step of executing a process of updating parameters using a second optimization method based on the acquired grouped uncoupled data of all of the groups and the grouped magnitude comparison data of all of the groups, and the estimated optimization hyperparameters, and estimating optimization parameters that minimize a second objective function including the grouped uncoupled data and the grouped magnitude comparison data of all of the groups; The model learning method comprises:

8. 7. A program for causing a processor included in the model learning device to execute processing by each of the processing units included in the model learning device according to claim 1.

Images