DATA ESTIMATION DEVICE, METHOD, AND PROGRAM

Abstract

In a data estimation device, a learning unit creates, using training data including an explanatory variable and an objective variable, a machine learning model that estimates an objective variable from an explanatory variable. The learning unit creates a machine learning model Mi that estimates an objective variable Oi from an explanatory variable group Ei including one or more explanatory variables, sets a new explanatory variable group Ei+1 by adding the objective variable Oi estimated by the machine learning model Mi to the explanatory variable group Ei, and creates a machine learning model Mi+1 that estimates an objective variable Oi+1 from the explanatory variable group Ei+1 (where i=1). The learning unit repeatedly creates a machine learning model while i is in a range of from 2 to (n−1) (n is a natural number greater than or equal to 2).

Description

TECHNICAL FIELD

The present invention relates to a data estimation technique for estimating an objective variable from an explanatory variable.

BACKGROUND ART

As a machine learning method, decision tree learning for creating a classifier having a tree structure from training data including an explanatory variable and an objective variable is used. A classification result can be predicted for unknown input data using a learned decision tree. Further, random forest in which a plurality of decision trees are learned with training data randomly changed, and a prediction is made by taking a majority decision to enhancing generalization ability is used.

The learning device disclosed in Patent Literature 1 creates a plurality of decision trees, using pieces of training data each including an explanatory variable and an objective variable, which are configured by a combination of the explanatory variables and each estimate the objective variable on the basis of true or false of the explanatory variables. The learning device creates a linear model that is equivalent to the plurality of decision trees and lists all terms configured by a combination of the explanatory variables without omission to output a stable prediction result by using the linear model from input data.

CITATION LIST

Patent Literature

[Patent Literature 1] JP 2020-46891 A

SUMMARY OF INVENTION

Technical Problem

Machine learning for predicting an objective variable from an explanatory variable has a problem that accuracy in objective variable estimation reaches a plateau.

The present invention has been made in view of such a problem, and it is therefore an object of the present invention to provide a data estimation technique capable of improving accuracy in estimation of an objective variable from an explanatory variable.

Solution to Problem

In order to solve the above-described problem, a data estimation device according to one aspect of the present invention includes a learning unit that creates, using training data including an explanatory variable and an objective variable, a machine learning model that estimates an objective variable from an explanatory variable. The learning unit creates a machine learning model M_i that estimates an objective variable O_i from an explanatory variable group E_i including one or more explanatory variables, sets a new explanatory variable group E_i+1 by adding the objective variable O_i estimated by the machine learning model M_i to the explanatory variable group E_i, and creates a machine learning model M_i+1 that estimates an objective variable O_i+1 from the explanatory variable group E_i+1 (where i=1).

Another aspect of the present invention is a data estimation method. The method includes a learning process of creating, using training data including an explanatory variable and an objective variable, a machine learning model that estimates an objective variable from an explanatory variable. The learning process includes creating a machine learning model M_i that estimates an objective variable O_i from an explanatory variable group E_i including one or more explanatory variables, setting a new explanatory variable group E_i+1 by adding the objective variable O_i estimated by the machine learning model M_i to the explanatory variable group E_i, and creating a machine learning model M_i+1 that estimates an objective variable O_i+1 from the explanatory variable group E_i+1 (where i=1).

Note that any combination of the above-described components, or an entity that results from replacing expressions of the present invention among a method, a device, a system, a computer program, a data structure, a recording medium, and the like is also valid as an aspect of the present invention.

Advantageous Effects of Invention

According to the present invention, it is possible to increase accuracy in estimation of an objective variable from an explanatory variable.

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 is a configuration diagram of a data estimation device according to the present embodiment.

FIG. 2 is a flowchart for describing a procedure of creating a machine learning model by the data estimation device in FIG. 1.

FIG. 3 is a diagram for describing secondary data on running performance.

FIG. 4 is a flowchart for describing a data estimation procedure in an example.

FIGS. 5(a) and 5(b) are diagrams for describing display examples of evaluation items output by an evaluation item display unit in FIG. 1.

FIG. 6 is a diagram for describing accuracy in objective variable estimation in the example.

FIG. 7 is a diagram for describing correlation coefficients between four objective variables A to D.

FIGS. 8(a) and 8(b) are diagrams for describing correlation coefficients between a selected objective variable and the remaining objective variables.

DESCRIPTION OF EMBODIMENTS

FIG. 1 is a configuration diagram of a data estimation device 100 according to the present embodiment. The data estimation device 100 includes an explanatory variable input unit 10, a learning unit 20, an objective variable output unit 30, an explanatory variable adding unit 40, an evaluation item display unit 50, an explanatory variable storage unit 60, a learning model storage unit 70, and an objective variable storage unit 80. This drawing illustrates a block diagram focused on functions, and such functional blocks may be implemented in various forms such as hardware, software, or a combination of hardware and software.

In a training phase, the data estimation device 100 creates, using training data including explanatory variables and objective variables, a machine learning model that estimates an objective variable from an explanatory variable. In a prediction phase, the data estimation device 100 inputs an unknown explanatory variable to the created machine learning model to predict an objective variable.

First, a configuration and operation of the data estimation device 100 in the training phase will be described. A value of an explanatory variable and a value of an objective variable are given to the learning unit 20 as training data. The learning unit 20 creates, using the given training data, a machine learning model that estimates an objective variable from an explanatory variable, and stores the machine learning model into the learning model storage unit 70. As a machine learning method, a regression model, a decision tree, a random forest, Bayesian estimation, a neural network, or the like may be used.

The objective variable output unit 30 outputs the value of the objective variable estimated from the value of the explanatory variable on the basis of the learned machine learning model, and stores the value of the objective variable into the objective variable storage unit 80. The explanatory variable adding unit 40 newly adds the objective variable estimated on the basis of the machine learning model to the explanatory variable and stores the explanatory variable into the explanatory variable storage unit 60.

The explanatory variable input unit 10 reads the newly set explanatory variable from the explanatory variable storage unit 60 and supplies the newly set explanatory variable to the learning unit 20. The learning unit 20 creates a machine learning model that estimates an objective variable from the newly set explanatory variable, and stores the machine learning model into the learning model storage unit 70. Subsequently, a machine learning model is repeatedly created with the estimated objective variable newly added to the explanatory variable.

Next, a configuration and operation of the data estimation device 100 in the prediction phase will be described. In the prediction phase, the learning unit 20 functions as a prediction unit.

The explanatory variable input unit 10 reads the value of the explanatory variable stored in the explanatory variable storage unit 60 as unknown data, and gives the explanatory variable to the learning unit 20.

The learning unit 20 reads the machine learning model stored in the learning model storage unit 70 and estimates an objective variable from the explanatory variable on the basis of the machine learning model. The objective variable output unit 30 stores the estimated value of the objective variable into the objective variable storage unit 80.

The explanatory variable adding unit 40 newly adds the objective variable estimated on the basis of the machine learning model to the explanatory variable and stores the explanatory variable into the explanatory variable storage unit 60.

The explanatory variable input unit 10 reads the newly set explanatory variable from the explanatory variable storage unit 60 and supplies the newly set explanatory variable to the learning unit 20. The learning unit 20 estimates an objective variable from the newly set explanatory variable on the basis of the machine learning model stored in the learning model storage unit 70. Subsequently, an objective variable is repeatedly estimated using the machine learning model with the estimated objective variable newly added to the explanatory variable.

The evaluation item display unit 50 calculates and displays a value of each evaluation item on the basis of the estimated value of the objective variable. The evaluation item display unit 50 may calculate the value of each evaluation item on the basis of the value of the explanatory variable and the estimated value of the objective variable.

FIG. 2 is a flowchart for describing a procedure of creating a machine learning model by the data estimation device 100.

It is assumed that one or more explanatory variables and a plurality of objective variables O_i (i=1 to n) are given as training data.

The explanatory variable input unit 10 sets the one or more explanatory variables to an explanatory variable group E₁ (S10). The learning unit 20 sets a variable i to 1 (S20).

The learning unit 20 creates a learning model M_i that estimates an objective variable O_i from an explanatory variable group E_i, and the objective variable output unit 30 outputs the estimated objective variable O_i to the objective variable storage unit 80 (S30).

The explanatory variable adding unit 40 adds the objective variable O_i estimated by the learning model M_i to the explanatory variable group E_i to set a new explanatory variable group E_i+1, and stores the explanatory variable group E_i+1 into the explanatory variable storage unit 60 (S40).

The learning unit 20 increments the variable i by 1 (S50). In a case where the variable i is greater than n (Y in S60), the processing of creating a machine learning model is brought to an end. In a case where the variable i is less than or equal to n (N in S60), the processing returns to step S30, and the subsequent procedure is repeated.

As described above, in a case where explanatory variables are unchanged and there are a plurality of objective variables, it is possible to increase the accuracy in objective variable estimation by the following procedures (1) to (3):

• • (1) Build a machine learning model capable of estimating any objective variable using an explanatory variable prepared in advance;
  • (2) Build a new machine learning model with an estimated objective variable newly added to the explanatory variable; and
  • (3) Repeat the above (2).

Next, a case where secondary data on running performance is learned and predicted from primary data on running of a runner and physical characteristic data on the runner using the data estimation device 100 will be described as an example.

The primary data on running that is measurable using a sensor attached to shoes of the runner includes a running pace, a stride, a pitch, a grounding time, and a hang time. The physical characteristic data on the runner includes height and weight. Such measurable primary data and physical characteristic data serve as an explanatory variable.

The secondary data on running performance to be estimated includes a second peak value (denoted as “Fz 2nd max”) of a z component of a ground reaction force (denoted as “Fz”), a propulsion force product, a braking force product, and a rising rate of the z component of the ground reaction force (denoted as “Fz Loading Rate”). Such secondary data serves as an objective variable.

FIG. 3 is a diagram for describing the secondary data on running performance. The horizontal axis of the graph represents a time, and the vertical axis represents a ground reaction force. The ground reaction force has a component in a vertical direction (z direction) and a component in a front-back direction (y direction), a solid line represents the z component Fz of the ground reaction force, and a dashed line represents a y component Fy of the ground reaction force.

A second peak value of the z component Fz of the ground reaction force is Fz 2nd max, and a slope of the rise of the z component Fz of the ground reaction force is Fz Loading Rate. An area of a region where the y component of the ground reaction force has a positive value is the propulsion force product, and an area of a region where the y component of the ground reaction force has a negative value is the braking force product.

Hereinafter, for convenience of description, Fz 2nd max is referred to as secondary data A, the propulsion force product is referred to as secondary data B, the braking force product is referred to as secondary data C, and Fz Loading Rate is referred to as secondary data D. Such pieces of secondary data A to D are also referred to as objective variables A to D.

As an example, the pieces of secondary data A, B, C, D that are objective variables are estimated in this order from the primary data and the physical characteristic data that are explanatory variables using the machine learning model, and the estimated pieces of secondary data A, B, C, D are added to the explanatory variables in this order. The order in which the objective variables are estimated, in other words, the order in which the estimated objective variables are input to the explanatory variables, may be different from the above-described order, and how to determine an input order that makes the accuracy in objective variable estimation higher will be described later.

FIG. 4 is a flowchart for describing a data estimation procedure in the example.

The primary data on running is acquired from the sensor attached to the shoes of the runner and is stored into the explanatory variable storage unit 60 together with the physical characteristic data on the runner (S100).

The learning unit 20 inputs, to a regression model, the primary data and the physical characteristic data as explanatory variables to estimate the secondary data A on running performance as an objective variable (S110).

The explanatory variable adding unit 40 newly adds the estimated secondary data A to the explanatory variables, and the learning unit 20 inputs, to the regression model, the primary data, the physical characteristic data, and the secondary data A as explanatory variables to estimate the secondary data B as an objective variable (S120).

The explanatory variable adding unit 40 newly adds the estimated secondary data B to the explanatory variables, and the learning unit 20 inputs, to the regression model, the primary data, the physical characteristic data, the secondary data A, and the secondary data B as explanatory variables to estimate the secondary data C as an objective variable (S130).

The explanatory variable adding unit 40 newly adds the estimated secondary data C to the explanatory variables, and the learning unit 20 inputs, to the regression model, the primary data, the physical characteristic data, the secondary data A, the secondary data B, and the secondary data C as explanatory variables to estimate the secondary data D as an objective variable (S140).

The evaluation item display unit 50 calculates and displays the value of each evaluation item on the basis of the primary data and the secondary data A to D (S150).

FIGS. 5(a) and 5(b) are diagrams illustrating display examples of the evaluation items output by the evaluation item display unit 50. Evaluation items V₁ to V_M such as a kicking force of a left foot, a kicking force of a right foot, a degree of reduction in impact load, a braking effect, and a kicking efficiency are displayed on, for example, a scale of 0 to 5 on the basis of the secondary data A to D estimated as the primary data. A display mode may be a radar chart as illustrated in FIGS. 5(a) and 5(b), or may be a bar graph.

FIG. 6 is a diagram for describing the accuracy in objective variable estimation in the example. The primary data and the physical characteristic data in the above-described example were input to the regression model as explanatory variables, and coefficients of determination when the pieces of secondary data A to D were estimated as objective variables were obtained. For comparison, FIG. 6 shows coefficients of determination when the objective variables A to D were estimated only with the explanatory variables under a known method, and coefficients of determination when the objective variables estimated in the order of the secondary data A to D were added to the explanatory variables and estimated under the present method. The coefficient of determination is also referred to as a contribution ratio, and the closer to 1, the higher the accuracy in objective variable estimation.

The coefficient of determination of the objective variable A is 0.84 in a case where a prediction is made only with the explanatory variables, which is the first objective variable estimation, so that the coefficient of determination of the objective variable A is also 0.84 under the present technique.

The coefficient of determination of the objective variable B is 0.67 in a case where a prediction is made only with the explanatory variables, and is improved to 0.69 in a case where a prediction is made with the estimated objective variable A added to the explanatory variables.

The coefficient of determination of the objective variable C is 0.36 in a case where a prediction is made only with the explanatory variables, and is improved to 0.52 in a case where a prediction is made with the estimated objective variable B further added to the explanatory variables.

The coefficient of determination of the objective variable D is 0.69 in a case where a prediction is made only with the explanatory variables, and is improved to 0.74 in a case where a prediction is made with the estimated objective variable C further added to the explanatory variables.

In the example, sequentially adding each objective variable estimated by the regression model to the explanatory variables and estimating the next objective variable by the regression model makes it possible to improve the accuracy in objective variable estimation.

Next, a method for further improving the accuracy in objective variable estimation by changing the order in which the plurality of objective variables are input to the explanatory variables will be described.

In order to determine an order in which n objective variables are input as explanatory variables, a machine learning model is created in each input order to calculate accuracy in prediction about the n objective variables, and an input order in which the mean value of the accuracy in prediction about the n objective variables becomes the largest or the standard deviation of the accuracy in prediction about the n objective variables becomes the smallest is finally selected as an optimum input order.

In the above-described example, the accuracy in prediction about the four objective variables A to D was evaluated for all 24 input orders of the four objective variables A to D. In a case where the four objective variables were input in the order of A, D, B, C, the coefficients of determination of the objective variables A, B, C, D were 0.84, 0.70, 0.55, and 0.71, respectively, the mean value of the coefficients of determination of the four objective variables A to D was 0.7000, and the standard deviation of the coefficients of determination of the four objective variables A to D was 0.1186. In a case where input was made in an order of A, D, B, C among the 24 input orders, the mean value of the coefficients of determination of the four objective variables A to D was the largest, and the standard deviation was the smallest. The input order of A, D, B, C is selected as an optimum input order.

The optimum input order of the objective variables can be derived on the basis of correlation coefficients between the objective variables without testing all the input orders. Next, a method for determining the optimum input order of the objective variables will be described.

FIG. 7 is a diagram illustrating correlation coefficients between the four objective variables A to D. With reference to the correlation coefficients, the optimum input order of the objective variables is determined by the following procedure. FIGS. 8(a) and 8(b) are obtained by extracting correlation coefficients between a selected objective variable and the remaining objective variables from FIG. 7, and the following procedure will be described with reference to FIGS. 8(a) and 8(b).

(Step 1) Select an objective variable that is the highest in prediction accuracy when a regression model is built only with explanatory variables, and build a regression model with the objective variable added to the explanatory variables.

In the example, with reference to FIG. 6, in a case where a prediction is made only with the explanatory variables, the objective variable A is the highest in prediction accuracy, so that the objective variable A is first added to the explanatory variables, and the regression model is built.

(Step 2) Obtain correlation coefficients between the selected objective variable and all the remaining objective variables, select an objective variable having the largest absolute value of the correlation coefficient, and build the regression model with the objective variable newly added to the explanatory variables. Here, in a case where a plurality of objective variables are selected, the mean of the absolute values of the correlation coefficients is obtained.

In the example, as shown in FIG. 8(a), the objective variable D having the largest absolute value of the correlation coefficient with the selected objective variable A is newly added to the explanatory variables, and the regression model is built.

(Step 3) Repeat step 2.

In the example, as shown in FIG. 8(b), the objective variable B having the largest mean of the absolute values of the correlation coefficients with the selected objective variables A, B is newly added to the explanatory variables, and the regression model is built.

(Step 4) Repeat step 2 in a case where an objective variable still remains.

In the example, the regression model is built with the last objective variable C newly added to the explanatory variables.

As described above, the data estimation device 100 of the present embodiment can improve the accuracy in objective variable estimation by first building, in a case where there are a plurality of objective variables to be predicted from explanatory variables, a machine learning model capable of estimating any objective variable using the explanatory variables, and then repeatedly building, with an estimated objective variable added to the explanatory variables, a machine learning model capable of estimating the next objective variable. Note that the present invention is applicable to not only an example where an objective variable is added to explanatory variables on a one-by-one basis but also an example where a plurality of objective variables are added to the explanatory variables.

The present invention has been described on the basis of the embodiment. It is to be understood by those skilled in the art that the embodiment is illustrative and that various modifications are possible for a combination of components or processes, and that such modifications are also within the scope of the present invention.

Although the example where the secondary data on running performance is estimated from the primary data on running of the runner and the physical characteristic data on the runner has been described, the present invention is applicable to any example as long as an objective variable is estimated from an explanatory variable.

INDUSTRIAL APPLICABILITY

The present invention is applicable to a data estimation technique.

REFERENCE SIGNS LIST

10 explanatory variable input unit, 20 learning unit, 30 objective variable output unit, 40 explanatory variable adding unit, 50 evaluation item display unit, 60 explanatory variable storage unit, 70 learning model storage unit, 80 objective variable storage unit, 100 data estimation device

Claims

1. A data estimation device comprising a learning unit structured to create, using training data including an explanatory variable and an objective variable, a machine learning model that estimates an objective variable from an explanatory variable, wherein the learning unit creates a machine learning model Mi that estimates an objective variable Oi from an explanatory variable group Ei including one or more explanatory variables, sets a new explanatory variable group Ei+1 by adding the objective variable Oi estimated by the machine learning model Mi to the explanatory variable group Ei, and creates a machine learning model Mi+1 that estimates an objective variable Oi+1 from the explanatory variable group Ei+1 (where i=1).

2. The data estimation device according to claim 1, wherein the learning unit repeatedly creates a machine learning model while i is in a range of from 2 to (n−1) (n is a natural number greater than or equal to 2).

3. The data estimation device according to claim 2, wherein in order to determine an order in which n objective variables are input as explanatory variables, the learning unit creates a machine learning model in each input order to calculate accuracy in prediction about the n objective variables, and finally selects an input order in which a mean value of the accuracy in prediction about the n objective variables becomes largest or a standard deviation of the accuracy in prediction about the n objective variables becomes smallest.

4. The data estimation device according to claim 2, wherein the learning unit selects, as an objective variable O1, an objective variable highest in accuracy in prediction using the explanatory variables by the machine learning model from among n objective variables.

5. The data estimation device according to claim 4, wherein the learning unit selects, in descending order of a correlation with already selected objective variables O1 to Oi, subsequent objective variables Oi+1 (i=1 to (n−1)).

6. A data estimation method comprising a learning process of creating, using training data including an explanatory variable and an objective variable, a machine learning model that estimates an objective variable from an explanatory variable, wherein the learning process includes creating a machine learning model Mi that estimates an objective variable Oi from an explanatory variable group Ei including one or more explanatory variables, setting a new explanatory variable group Ei+1 by adding the objective variable Oi estimated by the machine learning model Mi to the explanatory variable group Ei, and creating a machine learning model Mi+1 that estimates an objective variable Oi+1 from the explanatory variable group Ei+1 (where i=1).

7. A data estimation program that causes a computer to execute a learning process of creating, using training data including an explanatory variable and an objective variable, a machine learning model that estimates an objective variable from an explanatory variable, wherein the learning process includes creating a machine learning model Mi that estimates an objective variable Oi from an explanatory variable group Ei including one or more explanatory variables, setting a new explanatory variable group Ei+1 by adding the objective variable Oi estimated by the machine learning model Mi to the explanatory variable group Ei, and creating a machine learning model Mi+1 that estimates an objective variable Oi+1 from the explanatory variable group Ei+1 (where i=1).