JP2015060237A - Prediction model learning device, prediction model learning method, and computer program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide technology for generating a model capable of highly accurate prediction even when some of explanatory variables included in training data is lost.SOLUTION: In the case of machine learning of a model conducted on the basis of training data in which samples composed of a pair of objective variable and explanatory variable are collected, the machine learning using a plurality of prediction models set for each group of samples of the training data that is divided into a plurality of groups, a use rate calculation unit 14 calculates, using an estimated parameter, a use rate of each of the prediction models constituting a model to be outputted for a lost pattern indicating a loss of component in an explanatory variable vector. An estimation unit 13 estimates a parameter of each prediction model by using the use rate of each prediction model for the lost pattern. A process performed by the use rate calculation unit 14 and a process performed by the estimation unit 13 using the use rate of each prediction model calculated by the process are alternately repeated by an instruction unit 15.

Description

  The present invention relates to a technique for predicting prediction target data based on available data.

  Forecasting the future based on available data is useful for business improvement. For example, in a store, if the sales of a product can be predicted based on sales data for the last two weeks, the store can appropriately manage the inventory of the product. In addition, if a sales office can predict the sales method that will increase the possibility of receiving an order by analyzing the relationship between the sales method based on business records such as daily business reports, etc. Can be improved.

  Here, data that serves as a clue to prediction (for example, actual sales data or an executed sales technique) is referred to as an explanatory variable. In addition, data to be predicted (for example, sales of a product to be predicted or order status to be predicted) is referred to as an objective variable. Furthermore, a function that can obtain an objective variable (predicted value) by substituting (inputting) an explanatory variable (data) is called a model or a predicting function. Furthermore, a set of combinations of explanatory variables and objective variables that are past data (samples) is referred to as training data. Machine learning is used as a technique for creating a model (a function for outputting an objective variable using explanatory variables) based on the training data.

  By the way, in the machine learning, some explanatory variables in the training data may be missing. Specifically, for example, if the product A has not been put out in the store during a certain time zone, the sales of the product A in that time zone will be lost. In addition, when there is a day forgetting to record in the business daily report, the data for the forgotten day is lost. Thus, when machine learning a model (prediction function) based on training data in which some of the explanatory variables are missing, for example, a method of using the average value of the explanatory variables as the missing explanatory variables May be adopted. There is also a method of machine learning of a model (prediction function) by using a value predicted based on another explanatory variable as a missing explanatory variable.

  However, with such a method, an assumed value (substitute value) used as a missing explanatory variable may be greatly deviated from the original value, so that there is a possibility that a highly accurate model cannot be created. If an inaccurate model is used, there is a problem that the accuracy of prediction is lowered.

  Non-Patent Document 1 discloses a method of machine learning a model (prediction function) when some of explanatory variables in training data are missing. In the technique shown in Non-Patent Document 1, a machine (computer) that performs machine learning detects which explanatory variable is missing in training data, and samples that have the same missing explanatory variable ( The same label is given to past data that is a combination of explanatory variables and objective variables. Then, the apparatus outputs (generates) a model by performing machine learning using only a set of samples having the same label as training data.

Maytal Saar-Tsechansky and Foster Provost, “Handling Missing Values when Applying Classification Models” Journal Of Machine Learning Research, 8, (2007), 1625-1657

  However, if an explanatory variable having a small contribution to the objective variable is missing, if the model is generated using the method of Non-Patent Document 1, the accuracy of the model may be deteriorated. This is because in the method of Non-Patent Document 1, the training data is divided based on the missing state of the explanatory variable, and the degree of involvement of the explanatory variable with respect to the objective variable is not considered.

  The present invention has been made to solve the above problems. That is, the main object of the present invention is to provide a machine that can generate a model (prediction function) that enables highly accurate prediction even if some of the explanatory variables included in the training data (past data) are missing. It is to provide technology related to learning.

In order to achieve the above object, the predictive model learning device of the present invention provides:
An output target model that uses a plurality of prediction models respectively set for each group of the samples grouped into a plurality of groups in the training data in which samples that are pairs of objective variables and explanatory variable vectors are collected is a machine When learning, the usage rate of each prediction model constituting the output target model with respect to the missing pattern indicating the missing state of the component in the explanatory variable vector is calculated using the estimated parameter of the prediction model. Usage rate calculator,
An estimation unit that estimates a parameter of each prediction model using a use ratio of each prediction model with respect to the missing pattern;
The estimation unit estimates a parameter of each prediction model using the use rate of each prediction model for the missing pattern calculated by the use rate calculation unit, and each prediction estimated by the estimation unit A command unit that controls a process of alternately repeating the process of calculating the usage ratio of each prediction model with respect to the missing pattern using the model parameters.

The prediction model learning method of the present invention includes:
An output target model that uses a plurality of prediction models respectively set for each group of the samples grouped into a plurality of groups in the training data in which samples that are pairs of objective variables and explanatory variable vectors are collected is a machine When learning, the computer uses the estimated parameters of the prediction model to determine the usage ratio of each prediction model that constitutes the output target model with respect to the missing pattern indicating the missing state of the component in the explanatory variable vector. Calculate
Utilizing the usage rate of each prediction model for the missing pattern, the computer estimates the parameters of each prediction model,
The computer alternately repeats the process of estimating the parameters of each prediction model and the process of calculating the usage rate of each prediction model with respect to the missing pattern using the estimated parameters of each prediction model.

Furthermore, the computer program of the present invention is
An output target model that uses a plurality of prediction models respectively set for each group of the samples grouped into a plurality of groups in the training data in which samples that are pairs of objective variables and explanatory variable vectors are collected is a machine When learning, the usage rate of each prediction model constituting the output target model with respect to the missing pattern indicating the missing state of the component in the explanatory variable vector is calculated using the estimated parameter of the prediction model. Processing,
A processing procedure for causing a computer to execute a process of estimating a parameter of each prediction model using a use ratio of each prediction model with respect to the missing pattern is shown.
Furthermore, a process of alternately repeating the process of estimating the parameters of each prediction model and the process of calculating the usage ratio of each prediction model with respect to the missing pattern using the estimated parameters of each prediction model A processing procedure to be executed by the computer is shown.

  The object of the present invention is also achieved by the prediction model learning method of the present invention corresponding to the prediction model learning apparatus of the present invention having the above-described configuration. The object of the present invention is also achieved by a computer program for realizing the prediction model learning device and the prediction model learning method of the present invention by a computer and a computer program storage medium for storing the computer program.

  According to the present invention, it is possible to generate a model (prediction function) that enables highly accurate prediction even if some of the explanatory variables included in the training data (past data) are missing.

It is a block diagram which simplifies and represents the structure of the prediction model learning apparatus of 1st Embodiment which concerns on this invention. It is a block diagram which simplifies and represents the structure of the prediction model learning apparatus of 2nd Embodiment which concerns on this invention. It is a table | surface showing the specific example of the defect pattern in a training pattern. It is a table | surface utilized for description of the process which clusters a training pattern. It is a flowchart showing the operation example of the machine learning in the prediction model learning apparatus of 2nd Embodiment.

  Embodiments according to the present invention will be described below with reference to the drawings.

(First embodiment)
FIG. 1 is a block diagram showing a simplified configuration of the prediction model learning apparatus according to the first embodiment of the present invention. The prediction model learning device 10 of the first embodiment is a device that machine-learns the following model based on training data in which samples that are pairs of objective variables and explanatory variable vectors are collected. The model is a model (prediction function) configured by using a plurality of prediction models respectively set for each group of the samples grouped into a plurality of groups in the training data.

  The prediction model learning device 10 according to the first embodiment includes a control device 11 and a storage device 12. The storage device 12 stores a computer program (hereinafter also referred to as a program) 16 in which a control procedure for controlling the operation of the control device 11 is represented.

  The control device 11 has, for example, a CPU (Central Processing Unit), and the control device (computer) 11 can have the following functions by executing the program 16 read from the storage device 12. That is, the control device 11 includes an estimation unit 13, a usage rate calculation unit 14, and a command unit 15 as functional units.

  The usage rate calculation unit 14 uses the estimated parameters of the prediction model to calculate the usage rate of each prediction model that constitutes the output target model with respect to the missing pattern indicating the missing state of the component in the explanatory variable vector. Has the function to calculate.

  The estimation unit 13 has a function of estimating a parameter of each prediction model by using a use ratio of each prediction model with respect to the missing pattern.

  The command unit 15 has a function of controlling the estimation unit 13 and the usage rate calculation unit 14. For example, when the usage rate calculation unit 14 calculates the usage rate of each prediction model for the missing pattern, the command unit 15 outputs the calculation result to the estimation unit 13. Thereby, the estimation part 13 estimates the parameter of each said prediction model using the usage rate of each said prediction model with respect to the said missing pattern which is the calculation result. The command unit 15 outputs the parameters of each prediction model estimated by the estimation unit 13 to the usage rate calculation unit 14. Thereby, the usage rate calculation unit 14 calculates the usage rate of each prediction model with respect to the missing pattern in the same manner as described above, using the estimated parameters of each prediction model. In this way, the command unit 15 has a function of controlling processing that alternately repeats processing by the estimation unit 13 and processing by the usage rate calculation unit 14.

  Since the prediction model learning apparatus 10 of the first embodiment has a configuration for machine learning of a model in consideration of a missing pattern, some components of explanatory variable vectors included in training data (past data) Even if is missing, it is possible to generate a model that enables highly accurate prediction.

(Second Embodiment)
The second embodiment according to the present invention will be described below.

  FIG. 2 is a block diagram illustrating a simplified configuration of the prediction model learning apparatus according to the second embodiment. The prediction model learning device 20 includes a control device 21 and a storage device 22 when roughly classified. The storage device 22 has a storage medium (not shown), and a computer program (program) 30 and various data are stored in the storage medium. The program 30 represents a processing procedure for controlling the operation of the prediction model learning device 20.

  The control device (computer) 21 includes, for example, a CPU (Central Processing Unit). The control device 21 (CPU) can have the following functions by operating according to the program 30 read from the storage device 22. That is, in the second embodiment, the control device 21 includes a clustering unit 23, a complementing unit 24, a command unit 25, a usage rate calculating unit 26, an estimating unit 27, and a setting unit 28 as functional units. Have.

The clustering unit 23 has a function of performing clustering by analyzing the given training data (past data). Training data is a data group in which samples that are combinations of objective variables and explanatory variable vectors based on past data (actual data) are collected. For example, the training data is given (inputted) to the prediction model learning device 20 from the outside. Here, it is assumed that the explanatory variable vector is represented as x and the objective variable is represented as y. A sample (a combination of an objective variable and an explanatory variable vector) is represented as (x _i , y _i ). Note that i is a positive integer from 1 to N. Thereby, the training data D is

It can be expressed as.

In the second embodiment, the clustering unit 23 has a function of detecting whether a part of the component in the explanatory variable vector x is missing in each sample (x _i , y _i ) of the given training data D. It has. In addition, the clustering unit 23 has a function of detecting (identifying) a missing pattern representing a missing state when a part of the component in the explanatory variable vector x is missing. Further, the clustering unit 23 has a function of giving the same label to samples having the same or similar missing pattern of the explanatory variable vector x based on the detected missing pattern. Here, the clustering represents the process from classifying the samples as described above until labeling.

  There are various clustering methods. Here, when a part of the component of the explanatory variable vector x is missing, any method may be adopted as long as the sample can be classified based on the missing pattern. One specific example is described below.

  In this specific example, it is assumed that the number of samples included in the training data is 40, and the explanatory variable vectors x in these samples S1 to S40 have 10 components X1 to X10, respectively. FIG. 3 is a table showing a missing state in the components X1-X10 of the explanatory variable vector x in the samples S1-S40. In FIG. 3, “NA” is shown at the position corresponding to the missing component, and the numerical values of the other components are omitted. According to FIG. 3, in the samples S1-S5, all the components X1-X10 of the explanatory variable vector x are not missing. In samples S6-S10, the components X1-X5 of the explanatory variable vector x are missing. Further, in the samples S11 to S20, the components X1 to X6 of the explanatory variable vector x are missing, and in the samples S21 to S40, the components X7 to X10 of the explanatory variable vector x are missing.

  Regarding the training data including such an explanatory variable vector x, the clustering unit 23 compares the explanatory variable vector x of each sample S1 to S40 with the explanatory variable vector x of the other samples S1 to S40, and resembles the explanatory variable vector x. Calculate the degree. Here, in the explanatory variable vector x in the two samples being compared, the number of components missing in common is M, and the number of missing components in the two explanatory variable vectors x (number of missing portions). ) Let N be the number of defects included in the sample with the larger number of). For example, the clustering unit 23 calculates the similarity R according to an equation of M / N. The similarity R is 1 when M is zero and N is zero.

  FIG. 4 is a table showing the similarity R calculated based on the above calculation method. For example, since all the components of the explanatory variable vector x are not missing in the samples S1-S5, the similarity R of the explanatory variable vectors x as a result of comparing each sample S1-S5 with the samples S1-S5 is R = M ÷ N = 0 ÷ 0 = 1. Further, the similarity R of the explanatory variable vector x based on the result of comparison of each sample S1-S5 with samples S6-S10, S21-S40 is R = M ÷ N = 0 ÷ 5 = 0. Further, the similarity R of the explanatory variable vector x resulting from the comparison of each sample S1-S5 with the samples S11-S20 is R = M ÷ N = 0 ÷ 6 = 0.

  The clustering unit 23 sets (applies) the same label to a set of samples whose similarity R calculated in this way is 0.8 or more. For example, based on the similarity R shown in FIG. 4, the clustering unit 23 sets a label C1 for each of the samples S1-S5, sets a label C2 for each of the samples S6-S20, and samples S21-S40. Is set with a label C3.

  As described above, the clustering unit 23 has a function of clustering a plurality of samples by paying attention to the missing pattern of the explanatory variable vector x.

  The complement unit 24 has a function of complementing data (numerical values) in place of the missing component in the explanatory variable vector x. For example, in each sample S6-S40, the complement unit 24 substitutes (complements) the average value of the missing components in the explanatory variable vector x as the missing component. More specifically, in sample S6-S10, complement part 24 substitutes (complements) the average value of component X6-X10 into missing component X1-X5. In samples S11 to S20, the complementing unit 25 substitutes (complements) the average value of the components X7 to X10 into the missing components X1 to X6. Further, in the samples S21 to S40, the complementing unit 24 substitutes (complements) the average value of the components X1 to X6 into the missing components X7 to X10.

  The setting unit 28 has a function of setting a prediction model based on training data. For example, the setting unit 28 classifies the training data samples S1 to S40 into the following four groups based on combinations (patterns) of missing components of the explanatory variable vector x. That is, when the training data is a set of samples having a missing pattern as shown in the table of FIG. 3, the samples S1-S5 are not missing the components X1-X10 of all the explanatory variable vectors x. It is a group (referred to as group G1). Samples S6-S10 are groups in which the components X6-X10 of the explanatory variable vector x are not missing (referred to as group G2). Samples S11 to S20 are groups in which the components X7 to X10 of the explanatory variable vector x are not missing (referred to as group G3). Samples S21 to S40 are groups in which the components X1 to X6 of the explanatory variable vector x are not missing (referred to as group G4). The setting unit 28 sets a prediction model corresponding to each group of samples grouped in this way.

Here, it is assumed that the prediction model (function) associated with each group G1-G4 is represented by Expression (1).

  In addition, x represented by Formula (1) is an explanatory variable vector, and y is an objective variable. K is a code for identifying the prediction model, and is an integer of 1 or more (k = 1, 2,..., K). Here, k of each prediction model is set to a numerical value corresponding to each of the groups G1-G4 grouped as described above. That is, 1 is set to k of the prediction model corresponding to the sample group G1, and 2 is set to k of the prediction model corresponding to the group of samples G2. In addition, 3 is set to k of the prediction model corresponding to the group of samples G3, and 4 is set to k of the prediction model corresponding to the group of samples G4. That is, in this case, K = 4.

Θ ^(k) represents a parameter in the prediction model f _k .

Here, when the label given to the sample by the clustering process of the clustering unit 23 is c _(xi) , the usage ratio (model allocation latent variable) of the prediction model for each label is Z _{c (xi), k} Let it be represented. In this case, the prediction model considering the usage rate is expressed by Equation (2).

More specifically, it is assumed that the probability density function family represented by Expression (3) is set (defined) as the prediction model.

In Equation (3), θ: = (β, σ ² ) (β represents an average value (also referred to as a weight when represented by a linear function of explanatory variables), and σ represents variance). Further, τε {1, 2,...}.

Based on Expression (3), the prediction model corresponding to each group G1-G4 is expressed (defined) as Expression (4) -Expression (7).

In the second embodiment, the machine learning of the model is machine learning of the parameter θ ^(k) and the usage ratio Z _{c (xi), k} . The command unit 25 has a function of controlling the operations of the usage rate calculation unit 26 and the estimation unit 27 for the machine learning. For example, when receiving the training data, the command unit 25 determines whether or not the information on the use ratios _{Zc (xi), k} of the prediction model is stored in the storage unit 33 provided in the control device 21, for example. If it is determined that it is not stored _, the initial values of the use ratios _{Zc (xi), k} are set (generated). As a specific example, when the prediction models f ₁ -f ₄ for the groups G ₁ -G _{4 as described above} are set, the command unit 25 uses the usage ratio Z _c in all the prediction models f ₁ -f ₄ . _{, k} , set the same constant. That is, the usage ratio Z _{c, k} is set to 0.25. In this case, c = 1, 2, 3 and k = 1, 2, 3, 4.

The command unit 25 outputs the usage rate _{c, k} and training data to the estimation unit 27 when the usage rate Z _{c, k} information can be acquired. Thereby, the estimation unit 27 starts to function, and estimates the parameter θ ^(k) of each prediction model as described later. When the command unit 25 receives the parameter θ ^(k) estimated (calculated) by the estimation unit 27 from the estimation unit 27, the command unit 25 outputs the parameter θ ^(k) and training data to the use ratio calculation unit 28. As a result, the usage rate calculation unit 28 starts to function and calculates the usage rate Z _{c, k} as described later. When the command unit 25 receives the use rate Z _{c, k} calculated by the use rate calculation unit 28 from the use rate calculation unit 28, the command unit 25 outputs the use rate Z _{c, k} and training data to the estimation unit 27.

In this way, the command unit 25 performs machine learning of the parameter θ ^(k) and the usage rate Z _{c (xi), k} by controlling the estimation unit 27 and the usage rate calculation unit 28 to function alternately and repeatedly. Proceed. The command unit 25 continues such machine learning until a predetermined stop condition is satisfied. The stop condition, for example, a parameter which is newly calculated theta ^(k), the parameter theta ^(k) of each component of the parameter calculated by the calculation before one calculated theta ^(k) of the difference There is a condition that the sum of squares is 10 ⁻⁵ or less.

In the above example, the command unit 25 controls the repetitive operations of the estimation unit 27 and the usage rate calculation unit 28 after setting the initial values of the usage rate _{Zc (xi), k} . Instead, the command unit 25 sets (generates ⁾ an initial value of the parameter θ ^(k) , and outputs the set initial value and the training data to the use ratio calculation unit 28, thereby estimating as described above. The start of the repetitive operation of the unit 27 and the usage rate calculation unit 28 may be controlled.

The estimation unit 27 appropriately uses information on the prediction model set by the setting unit 28 based on the training data and the usage ratio Z _{c (xi), k} of the prediction model f ₁ -f ₄ for each label. Thus, a function for estimating the parameter θ is provided. For example, the estimation unit 27 uses the training data output from the command unit 25 and the usage rate information _{Zc (xi), k} to make predictions so that the log likelihood represented by Expression (8) increases. The parameter θ ⁽¹⁾ −θ ⁽⁴⁾ of the model f ₁ −f ₄ is calculated. In addition, when a part of the component of the explanatory variable vector x in the training data is missing, the data (numerical value) supplemented by the complementing unit 24 is used.

There are various methods for calculating the parameter θ so as to increase the log likelihood, and the estimation unit 27 may employ an appropriate method from among these methods. For example, the estimation unit 27 may use a regularization method in order to prevent the calculation from becoming complicated. Further, when the simultaneous equations represented by the equation (9) can be solved analytically, the estimation unit 27 can calculate (estimate) the parameter θ by substituting the calculation result into the equation (8). it can. Further, when the simultaneous equations of Equation (9) cannot be solved analytically, the estimation unit 27 may calculate (estimate) the parameter θ using numerical calculation such as Newton's method.

  Note that the operation symbol に お け る in equation (9) represents a nabla which is a vector differential operator.

The result of the estimation unit 27 estimating the parameter θ (θ = (β, σ ² )) using the equation (9) is as follows.

X ^(k) and Y ^(k) are defined as follows.

Note that d (k) represented in the column vector represents the total number of components constituting the column vector corresponding to the prediction model f _k .

  The estimation unit 27 has a function of outputting information on the estimated parameter θ to the command unit 25 and a function of registering information on the parameter θ in, for example, the storage unit 33 provided in the control device 21.

The usage rate calculation unit 26 uses the prediction model usage rate Z based on the training data and the parameter θ information output from the command unit 25 and appropriately uses the prediction model information set by the setting unit 28. _It has a function to calculate _{c (xi), k} . For example, the usage rate calculation unit 26 uses the prediction model usage rate so that the usage rate of the prediction model of the label increases with respect to the prediction model having a large probability for the sample of each label by the clustering process of the clustering unit 23. Is calculated. For example, the usage rate calculation unit 26 calculates the usage rate Z _{c (xi), k} of the prediction model based on the likelihood ratio of the prediction model represented by Expression (10).

Note that p _k (c) and τ (c) are defined as follows.

The ratio calculation unit 26, the proportion Z _c of the calculated prediction model _(xi), and a function of outputting information about _k in the command unit 25, the use of the prediction model ratio Z _{c (xi),} the information about the _k example And a function of registering in the storage unit 33 provided in the control device 21.

  Hereinafter, an operation example related to prediction model learning in the prediction model learning device 20 of the second embodiment will be described with reference to the flowchart of FIG. FIG. 5 is a flowchart of operations related to prediction model learning executed by the prediction model learning device 20 of the second embodiment. The flowchart is a computer executed by the control device 21 (CPU) of the prediction model learning device 20. The processing procedure of the program 30 is shown.

  For example, when the control device 21 (clustering unit 23) receives the training data from the outside of the control device 21, the control device 21 (clustering unit 23) specifies a missing pattern that represents the missing state of the explanatory variable vector x in each sample in the training data (step S101). ). And the clustering part 23 classify | categorizes the sample of training data based on the defect | deletion pattern, and provides the same label to the sample of the same classification | category. In other words, the control device 21 (clustering unit 23) clusters the training data based on the missing pattern (step S102).

Thereafter, the control device 21 (command unit 25) sets (generates) an initial value of the use rate Z _{c (xi), k} of the prediction model (step S103). The prediction model is a model set (defined) by the setting unit 28 based on the training data given to the control device 21 as described above.

  Thereafter, the control device 21 determines whether or not some components of the explanatory variable vector x in the training data are missing (step S104). Accordingly, when it is determined that the component is missing, the control device 21 (complement unit 24) supplements the missing component (step S105).

After the supplement processing or when the component of the explanatory variable vector x in the training data is not missing, the control device 21 (command unit 25) determines whether or not the stop condition is satisfied (step S106). ). When the command unit 25 determines that the stop condition is not satisfied, the command unit 25 outputs an initial value of the use ratio Z _{c (xi), k} and training data (complemented data) to the estimation unit 27. Thereby, the estimation part 27 starts a function and estimates parameter (theta) of a prediction model (step S107). Information on the estimated parameter θ is output from the estimation unit 27 to the command unit 25 and registered in the storage unit 33.

When the command unit 25 receives the parameter θ information and the training data from the estimation unit 27, the command unit 25 outputs the information to the use ratio calculation unit 26. Thereby, the usage rate calculation unit 26 calculates the usage rate Z _{c (xi), k} based on the received information (step S108). Information on the calculated use ratio Z _{c (xi), k} is output to the command unit 25 and registered in the storage unit 33.

  Thereafter, the command unit 25 determines whether or not the stop condition is satisfied (step S106). If it is determined that the stop condition is not satisfied, the operation after step S107 is repeated. If the command unit 25 determines that the stop condition is satisfied, it ends the machine learning of the model.

  With the above operation, the control device 21 performs machine learning on the model based on the training data.

  As described above, the prediction model learning device 20 of the second embodiment performs machine learning on the use ratio (model allocation latent variable) of the prediction model with respect to the missing pattern of the explanatory variable vector x included in the training data. Then, the prediction model learning device 20 performs machine learning on a model that takes into account the usage rate of the prediction model. That is, the prediction model learning device 20 can perform machine learning in which the degree of involvement of the missing component of the explanatory variable vector x with respect to the objective variable is considered. Thereby, the prediction model learning apparatus 20 is a model (prediction function) that enables highly accurate prediction even when a part of the components of the explanatory variable vector x included in the training data (past data) is missing. Can be generated.

(Other embodiments)
The present invention is not limited to the first and second embodiments, and various embodiments can be adopted. For example, in the second embodiment, the setting unit 28 sets (defines) a prediction model corresponding to each group of samples grouped by paying attention to the pattern of the missing component of the explanatory variable vector x in the training data. doing. Instead of this, for example, the setting unit 28 may set (define) a prediction model corresponding to each group of samples grouped by focusing on the pattern (missing pattern) of the explanatory variable vector x in the training data. Good. Alternatively, the setting unit 28 may set (define) a prediction model corresponding to each group of samples divided based on the subject matter other than the missing pattern of each sample in the training data. As described above, there are various methods for setting (defining) the prediction model, and here, any method may be used to set (define) the prediction model.

10, 20 Prediction model learning device 13, 27 Estimation unit 14, 26 Usage ratio calculation unit 15, 25 Command unit 23 Clustering unit 24 Complement unit

Claims (7)

An output target model that uses a plurality of prediction models respectively set for each group of the samples grouped into a plurality of groups in the training data in which samples that are pairs of objective variables and explanatory variable vectors are collected is a machine When learning, the usage rate of each prediction model constituting the output target model with respect to the missing pattern indicating the missing state of the component in the explanatory variable vector is calculated using the estimated parameter of the prediction model. Usage rate calculator,
An estimation unit that estimates a parameter of each prediction model using a use ratio of each prediction model with respect to the missing pattern;
The estimation unit estimates a parameter of each prediction model using the use rate of each prediction model for the missing pattern calculated by the use rate calculation unit, and each prediction estimated by the estimation unit A prediction model learning apparatus comprising: a command unit that controls a process of alternately repeating a process in which the usage rate calculation unit calculates a usage rate of each prediction model with respect to the missing pattern using a model parameter.   The usage rate calculator is configured so that the usage rate of the prediction model having the highest likelihood for the sample having the missing pattern is the highest in the usage rate of each prediction model for the missing pattern. The prediction model learning apparatus according to claim 1, wherein the use ratio of each prediction model is calculated.   The said estimation part estimates the parameter of the prediction model so that machine learning advances in the direction where the log likelihood of the function which multiplied the said usage rate of the said prediction model to the said prediction model becomes large. 2. The prediction model learning device according to 2. And further comprising a complement that complements the missing component in the explanatory variable vector,
The prediction model according to claim 1, wherein the use ratio calculation unit and the estimation unit use the training data in which a missing component in the explanatory variable vector is supplemented by the complement unit. Learning device. Classifying the sample of the training data based on the missing pattern, further comprising a clustering unit that assigns a label to each classification,
The prediction model learning device according to any one of claims 1 to 4, wherein the usage rate calculation unit calculates a usage rate of each prediction model with respect to the missing pattern for each label. An output target model that uses a plurality of prediction models respectively set for each group of the samples grouped into a plurality of groups in the training data in which samples that are pairs of objective variables and explanatory variable vectors are collected is a machine When learning, the computer uses the estimated parameters of the prediction model to determine the usage ratio of each prediction model that constitutes the output target model with respect to the missing pattern indicating the missing state of the component in the explanatory variable vector. Calculate
Utilizing the usage rate of each prediction model for the missing pattern, the computer estimates the parameters of each prediction model,
A prediction model in which the computer alternately repeats the process of estimating the parameters of each prediction model and the process of calculating the usage rate of each prediction model with respect to the missing pattern using the estimated parameters of each prediction model Learning method. An output target model that uses a plurality of prediction models respectively set for each group of the samples grouped into a plurality of groups in the training data in which samples that are pairs of objective variables and explanatory variable vectors are collected is a machine When learning, the usage rate of each prediction model constituting the output target model with respect to the missing pattern indicating the missing state of the component in the explanatory variable vector is calculated using the estimated parameter of the prediction model. Processing,
A processing procedure for causing a computer to execute a process of estimating a parameter of each prediction model using a use ratio of each prediction model with respect to the missing pattern is shown.
Furthermore, a process of alternately repeating the process of estimating the parameters of each prediction model and the process of calculating the usage ratio of each prediction model with respect to the missing pattern using the estimated parameters of each prediction model A computer program showing the processing procedure to be executed by a computer.

Images