Data estimation apparatus, method, and computer-readable storage medium

Abstract

The present invention provides a data inference device, method and computer-readable storage medium, which can improve the accuracy of inferring a target variable from an explanatory variable. In the data inference device (100) of the present invention, a learning unit (20) uses training data containing explanatory variables and target variables to generate a machine learning model for inferring the target variable from the explanatory variables. The learning unit (20) generates an explanatory variable group (E) containing one or more explanatory variables. i ) Estimated target variable (O i )’s machine learning model (M i ), which will be represented by the machine learning model (M i ) The estimated target variable O i ) is added to the explanatory variable group (E i ) and set a new explanatory variable group (E i+1 ), generate the explanatory variable group (E i+1 ) Estimated target variable (O i+1 )’s machine learning model (M i+1 )(wherein, i=1). The learning unit (20) repeatedly generates a machine learning model corresponding to i=2 to (n-1) (n is a natural number greater than or equal to 2).

Description

TECHNICAL FIELD

[0001] The present application relates to a data estimation technique of estimating a target variable from an explanatory variable. BACKGROUND

[0002] As a method of machine learning, decision tree learning is used, which is a classifier that generates a tree structure from training data including an explanatory variable and a target variable. The learned decision tree can be used to predict a classification result for unknown input data. Furthermore, a random forest is used, which is a technique of learning a plurality of decision trees by randomly changing the training data and making a majority decision to predict, thereby improving generalization ability.

[0003] The learning device described in Patent Literature 1 generates a plurality of decision trees that respectively estimate a target variable from a combination of explanatory variables and the truth of the explanatory variables, respectively, from training data including an explanatory variable and a target variable, generates a linear model that is equivalent to the plurality of decision trees and lists items including a combination of explanatory variables without omission, and thereby outputs a stable prediction result using the linear model from input data.

[0004] PRIOR ART DOCUMENTS

[0005] PATENT LITERATURE

[0006] Patent Literature 1: Japanese Patent Laid-Open No. 2020-46891 SUMMARY

[0007] PROBLEMS TO BE SOLVED BY THE INVENTION

[0008] Machine learning that predicts a target variable from an explanatory variable has a problem that the estimation accuracy of the target variable reaches a limit.

[0009] The present application has been achieved in view of the above-described problems, and an object of the present application is to provide a data estimation technique capable of improving the accuracy of estimating a target variable from an explanatory variable.

[0010] MEANS OF SOLVING THE PROBLEMS

[0011] To solve the above-described problems, a data estimation device according to an embodiment of the present application includes a learning unit that generates a machine learning model for estimating a target variable from an explanatory variable using training data including an explanatory variable and a target variable. The learning unit generates a machine learning model for estimating a target variable from an explanatory variable group E i estimating a target variable O i from a machine learning model M i , which estimates a target variable O i from a machine learning model M iAdd to explanatory variable group E i and set a new explanatory variable group E i+1 , generate the explanatory variable group E i+1 Estimated target variable O i+1 Machine learning model M i+1 (where i=1).

[0012] Another embodiment of the present invention is a data inference method. The method includes a learning step, wherein the learning step uses training data containing explanatory variables and target variables to generate a machine learning model that infers the target variable based on the explanatory variables. The learning step generates an explanatory variable group E containing one or more explanatory variables. i Estimated target variable O i Machine learning model M i , will be represented by the machine learning model M i The estimated target variable O i Add to explanatory variable group E i and set a new explanatory variable group E i+1 , generate the explanatory variable group E i+1 Estimated target variable O i+1 Machine learning model M i+1 (where i=1).

[0013] Furthermore, any combination of the above-described constituent elements or embodiments obtained by converting the expression of the present invention into methods, apparatuses, systems, computer programs, data structures, recording media, etc. may also be effective as embodiments of the present invention.

[0014] Effects of the Invention

[0015] According to the present invention, it is possible to improve the accuracy of estimating a target variable from an explanatory variable. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a diagram showing the configuration of a data estimation device according to this embodiment.

[0017] Figure 2 It is based on Figure 1 Flowchart of a procedure for generating a machine learning model of a data inference device.

[0018] Figure 3 This is a graph illustrating secondary data related to running performance.

[0019] Figure 4 This is a flowchart illustrating the data estimation procedure in the embodiment.

[0020] Figure 5 (a) Figure 5 (b) is based on Figure 1Fig. 1 is a diagram showing an example of display of evaluation items of the evaluation item display section.

[0021] Figure 6 Fig. 2 is a diagram showing the estimation accuracy of the target variables in the embodiment.

[0022] Figure 7 Fig. 3 is a diagram showing the correlation coefficient between the four target variables A to D.

[0023] Figure 8 (a), Figure 8 (b) is a diagram showing the correlation coefficient between the selected target variable and the remaining target variables.

[0024] Explanation of symbols

[0025] 10: explanatory variable input section

[0026] 20: learning section

[0027] 30: target variable output section

[0028] 40: explanatory variable addition section

[0029] 50: evaluation item display section

[0030] 60: explanatory variable storage section

[0031] 70: learning model storage section

[0032] 80: target variable storage section

[0033] 100: data estimation device DETAILED DESCRIPTION

[0034] Figure 1 Fig. 1 is a diagram showing an example of display of evaluation items of the evaluation item display section. Figure 1 The functional blocks depicted in the block diagrams are implemented in various forms, by hardware, software, or a combination thereof.

[0035] In the learning phase, the data estimation device 100 generates a machine learning model for estimating the target variables from the explanatory variables, using training data including the explanatory variables and the target variables. In the prediction phase, the data estimation device 100 inputs unknown explanatory variables into the generated machine learning model and predicts the target variables.

[0036] First, the structure and operation of the data estimation device 100 in the learning stage will be described. The learning unit 20 is provided with values of the explanatory variables and values of the target variables as training data. The learning unit 20 generates a machine learning model that estimates the target variables from the explanatory variables using the provided training data, and stores the model in the learning model storage unit 70. As a method of machine learning, a regression model, a decision tree, a random forest, Bayesian inference, a neural network, or the like can be used.

[0037] The target variable output unit 30 outputs the values of the target variables estimated from the values of the explanatory variables based on the learned machine learning model, and stores the values in the target variable storage unit 80. The explanatory variable addition unit 40 newly adds the target variables estimated based on the machine learning model to the explanatory variables, and stores the values in the explanatory variable storage unit 60.

[0038] The explanatory variable input unit 10 reads out the newly set explanatory variables from the explanatory variable storage unit 60, and supplies the variables to the learning unit 20. The learning unit 20 generates a machine learning model that estimates the target variables from the newly set explanatory variables, and stores the model in the learning model storage unit 70. Thereafter, the machine learning model is repeatedly generated by newly adding the estimated target variables to the explanatory variables.

[0039] Next, the structure and operation of the data estimation device 100 in the prediction stage will be described. In the prediction stage, the learning unit 20 functions as a prediction unit.

[0040] The explanatory variable input unit 10 reads out the values of the explanatory variables stored in the explanatory variable storage unit 60 as unknown data, and supplies the values to the learning unit 20.

[0041] The learning unit 20 reads out the machine learning model stored in the learning model storage unit 70, and estimates the target variables from the explanatory variables based on the machine learning model. The target variable output unit 30 stores the values of the estimated target variables in the target variable storage unit 80.

[0042] The explanatory variable addition unit 40 newly adds the target variables estimated based on the machine learning model to the explanatory variables, and stores the values in the explanatory variable storage unit 60.

[0043] The explanatory variable input unit 10 reads out the newly set explanatory variables from the explanatory variable storage unit 60, and supplies the variables to the learning unit 20. The learning unit 20 estimates the target variables from the newly set explanatory variables based on the machine learning model stored in the learning model storage unit 70. Thereafter, the machine learning model is repeatedly used to estimate the target variables by newly adding the estimated target variables to the explanatory variables.

[0044] The evaluation item display section 50 displays the value of each evaluation item calculated based on the estimated value of the target variable. The evaluation item display section 50 can also calculate the value of each evaluation item based on the value of the explanatory variable and the estimated value of the target variable.

[0045] Figure 2 is a flowchart illustrating a procedure of generating a machine learning model based on the data estimation device 100.

[0046] As the training data, one or more explanatory variables and a plurality of target variables O i (i = 1 ~ n).

[0047] The explanatory variable input section 10 sets one or more explanatory variables as an explanatory variable group E1 (S10). The learning section 20 sets 1 as the variable i (S20).

[0048] The learning section 20 generates a learning model M i estimating the target variable O i from the explanatory variable group E i The target variable output section 30 outputs the estimated target variable O i to the target variable storage section 80 (S30).

[0049] The explanatory variable addition section 40 adds the target variable O i estimated by the learning model M i to the explanatory variable group E i and sets a new explanatory variable group E i+1 , and stores it in the explanatory variable storage section 60 (S40).

[0050] The learning section 20 increments the variable i by 1 (S50). In the case where the variable i is greater than n (Y of S60), the generation process of the machine learning model is ended. In the case where the variable i is n or less (N of S60), it returns to step S30 and repeats the procedure thereafter.

[0051] As described above, in the case where the explanatory variables are the same and there are a plurality of target variables, the estimation accuracy of the target variable can be improved by the following procedures (1) to (3).

[0052] (1) A machine learning model that can estimate an arbitrary target variable is constructed using an explanatory variable prepared in advance.

[0053] (2) A new machine learning model is constructed by newly adding the estimated target variable to the explanatory variable.

[0054] (3) The above (2) is repeated.

[0055] Next, as an example, the following example is described, in which, using the data estimation device 100, secondary data related to running performance is learned from primary data related to running of a runner and physical characteristic data of the runner, and prediction is performed.

[0056] As the primary data related to running that can be measured using a sensor provided on a shoe of a runner, there are running pace, stride, pitch, ground contact time, and flight time. As the physical characteristic data of the runner, there are height and weight. These measurable primary data and physical characteristic data are explanatory variables.

[0057] As the secondary data related to running performance to be estimated, there are a second peak value of a z component of a ground reaction force (expressed as "Fz") (expressed as "Fz 2nd max"), propulsive impulse, braking impulse, and a rate of increase of the z component of the ground reaction force (expressed as "Fz Loading Rate"). These secondary data are target variables.

[0058] Figure 3 is a graph that illustrates secondary data related to running performance. The horizontal axis of the graph is time, and the vertical axis is a ground reaction force. The ground reaction force has a component in the vertical direction (z direction) and a component in the front-rear direction (y direction), and the solid line indicates the z component Fz of the ground reaction force, and the broken line indicates the y component Fy of the ground reaction force.

[0059] The second peak value of the value of the z component Fz of the ground reaction force is Fz 2nd max, and the slope of the increase of the z component Fz of the ground reaction force is Fz Loading Rate. The area of the region in which the y component of the ground reaction force has a positive value is the propulsive impulse, and the area of the region in which the y component of the ground reaction force has a negative value is the braking impulse.

[0060] Hereinafter, for convenience of explanation, Fz 2nd max is referred to as secondary data A, the propulsive impulse is referred to as secondary data B, the braking impulse is referred to as secondary data C, and Fz Loading Rate is referred to as secondary data D. These secondary data A to D are also referred to as target variables A to D.

[0061] As an example, by the machine learning model, secondary data A, secondary data B, secondary data C, secondary data D as target variables are estimated from the primary data and the body characteristic data as explanatory variables in the order, and the estimated secondary data are added to the explanatory variables step by step in the order. The order of estimating the target variables, in other words, the order of inputting the estimated target variables to the explanatory variables can also be an order other than the above, and a method of determining the input order of the target variables in which the estimation accuracy of the target variables is higher will be described below.

[0062] Figure 4 is a flowchart illustrating a data estimation procedure in the embodiment.

[0063] Primary data related to running is acquired from a sensor of a shoe of a runner, and the primary data is stored to an explanatory variable storage section 60 together with body characteristic data of the runner (S100).

[0064] The learning section 20 inputs the primary data and the body characteristic data to the regression model as explanatory variables, and estimates secondary data A related to the running performance as a target variable (S110).

[0065] The explanatory variable addition section 40 newly adds the estimated secondary data A to the explanatory variables, the learning section 20 inputs the primary data, the body characteristic data, and the secondary data A to the regression model as explanatory variables, and estimates secondary data B as a target variable (S120).

[0066] The explanatory variable addition section 40 newly adds the estimated secondary data B to the explanatory variables, the learning section 20 inputs the primary data, the body characteristic data, the secondary data A, and the secondary data B to the regression model as explanatory variables, and estimates secondary data C as a target variable (S130).

[0067] The explanatory variable addition section 40 newly adds the estimated secondary data C to the explanatory variables, the learning section 20 inputs the primary data, the body characteristic data, the secondary data A, the secondary data B, and the secondary data C to the regression model as explanatory variables, and estimates secondary data D as a target variable (S140).

[0068] The evaluation item display section 50 displays the values of the evaluation items calculated on the basis of the primary data and the secondary data A to D (S150).

[0069] Figure 5 (a), Figure 5 (b) is a diagram illustrating an example of display of the evaluation items based on the evaluation item display section 50. On the basis of the primary data and the estimated secondary data A to D, evaluation items V1 to V4 of the left foot, the right foot, the rebound force, the degree of reduction in the burden of impact, the braking effect, the rebound efficiency, and the like are displayed.M For example, the display is performed in 5-level evaluation. The display mode can be as follows Figure 5 (a) Figure 5 The radar chart shown in (b) can also be a bar chart.

[0070] Figure 6 This graph illustrates the estimation accuracy of the target variable in the example. The coefficient of determination is calculated when the primary data and physical characteristic data from the example are input into a regression model as explanatory variables, and secondary data A to D are estimated as the target variable. Figure 6 For comparison, the coefficient of determination for estimating target variables A through D using only explanatory variables, as in the conventional method, is shown, as well as the coefficient of determination for estimating target variables using this method by adding the target variables estimated in the order of secondary data A through D to the explanatory variables. The coefficient of determination is also called the contribution rate; the closer it is to 1, the higher the target variable estimation accuracy.

[0071] The coefficient of determination of the target variable A is 0.84 when predicted using only the explanatory variables. Since this is the initial estimation of the target variable, it is also 0.84 in this method.

[0072] The coefficient of determination of the target variable B is 0.67 when the prediction is made using only the explanatory variables, and it increases to 0.69 when the prediction is made by adding the estimated target variable A to the explanatory variables.

[0073] The coefficient of determination of the target variable C is 0.36 when the prediction is performed using only the explanatory variables, and it increases to 0.52 when the estimated target variable B is added to the explanatory variables for prediction.

[0074] The coefficient of determination of the target variable D is 0.69 when the prediction is performed using only the explanatory variables, and it increases to 0.74 when the estimated target variable C is added to the explanatory variables for prediction.

[0075] In the embodiment, the target variable estimated by the regression model is sequentially added to the explanatory variables and the next target variable is estimated by the regression model, thereby improving the estimation accuracy of the target variable.

[0076] Next, a method of further improving the estimation accuracy of the target variables by changing the order in which a plurality of target variables are input into the explanatory variables will be described.

[0077] In order to determine the order in which n target variables are input as explanatory variables, a machine learning model is generated according to each input order and the prediction accuracy of the n target variables is calculated. Finally, the input order with the largest average value of the prediction accuracy of the n target variables or the smallest standard deviation of the prediction accuracy of the n target variables is selected as the optimal input order.

[0078] In the embodiment, the prediction accuracy of the four target variables A to D was evaluated for all 24 different orderings of their introduction. When the four target variables were introduced in the order of A, D, B, and C, the coefficients of determination for target variables A, B, C, and D were 0.84, 0.70, 0.55, and 0.71, respectively. The average value of the coefficients of determination for the four target variables A to D was 0.7000, and the standard deviation of the coefficients of determination for the four target variables A to D was 0.1186. Among the 24 orderings, the average value of the coefficients of determination for the four target variables A to D was the largest and the standard deviation was the smallest when introduced in the order of A, D, B, and C. The order of introduction of A, D, B, and C was selected as the most appropriate order of introduction.

[0079] The optimal order of injection of target variables can be derived based on the correlation coefficients between target variables without trying all injection orders. Next, a method for determining the optimal order of injection of target variables will be described.

[0080] Figure 7 This is a diagram illustrating the correlation coefficients among the four target variables A to D. Referring to the correlation coefficients, the optimal order of inputting the target variables is determined according to the following procedure. Figure 8 (a) Figure 8 (b) is from Figure 7 The correlation coefficient between the selected target variable and the remaining target variables is selected and referred to in the following procedure description.

[0081] (Step 1) Select the target variable with the highest prediction accuracy when building a regression model using only explanatory variables, add it to the explanatory variables, and build a regression model.

[0082] In the case of the embodiment, if reference is made to Figure 6 , then when only explanatory variables are used for prediction, the prediction accuracy of target variable A is the highest, so the target variable A is initially added to the explanatory variables to build a regression model.

[0083] (Step 2) Calculate the correlation coefficient between the selected target variable and all remaining target variables. Select the target variable with the largest absolute value of the correlation coefficient and add it to the explanatory variables to construct the regression model. If multiple target variables are selected, the average of the absolute values ​​of the correlation coefficients is used.

[0084] In the case of the embodiment, as shown in Figure 8 (a), a target variable D having the largest absolute value of the correlation coefficient with the selected target variable A is newly added to the explanatory variables, and a regression model is constructed.

[0085] (Step 3) Step 2 is repeated.

[0086] In the case of the embodiment, as shown in Figure 8 (b), a target variable B having the largest average of the absolute values of the correlation coefficients with the selected target variables A and B is newly added to the explanatory variables, and a regression model is constructed.

[0087] (Step 4) Further, in the case of a target variable remaining, Step 2 is repeated.

[0088] In the case of the embodiment, a final target variable C is newly added to the explanatory variables, and a regression model is constructed.

[0089] As described above, according to the data estimation device 100 of the present embodiment, in the case where there are a plurality of target variables to be predicted from the explanatory variables, first, a machine learning model capable of estimating an arbitrary target variable is constructed using the explanatory variables, next, a newly estimated target variable is added to the explanatory variables, a machine learning model capable of estimating a next target variable is constructed, and the operation is repeated, whereby the estimation accuracy of the target variables can be improved. Furthermore, the present application is not limited to the embodiment in which the target variables are added one by one to the explanatory variables, but can be applied to an embodiment in which a plurality of target variables are added to the explanatory variables.

[0090] The present application has been described above based on the embodiments. It will be understood by those skilled in the art that the embodiments are illustrative, and that various modifications of the respective components or combinations of the respective processes are possible, and that such modifications are also included in the scope of the present application.

[0091] An embodiment in which secondary data related to running performance is estimated from primary data related to running and physical characteristic data of a runner has been described, but as long as a target variable is estimated from an explanatory variable, the present application can be applied to any embodiment.

[0092] Industrial applicability

[0093] The present application can be applied to data estimation technology.

Claims

1. A data estimation device characterized by comprising a computer, and a learning section executed by the computer, the learning section generating a machine learning model for estimating a target variable from an explanatory variable using training data containing the explanatory variable and the target variable, the explanatory variable including primary data measured by a sensor in relation to a runner's running and physical characteristic data of the runner, and the target variable including secondary data in relation to running performance, the learning section repeatedly generating the machine learning model in correspondence with i = 2 to (n-1), n being a natural number of 2 or more, The learning unit generates an explanatory variable group E including one or more explanatory variables. i Estimated target variable O i Machine learning model M i , will be represented by the machine learning model M i The estimated target variable O i Add to explanatory variable group E i and set a new explanatory variable group E i＋1 , generate the explanatory variable group E i＋1 Estimated target variable O i＋1 Machine learning model M i＋1 , where i=1, the learning section generating the machine learning model in accordance with each input order and calculating the prediction accuracy of the n target variables in order to determine the input order of the n target variables as explanatory variables, and finally selecting the input order in which the average of the prediction accuracy of the n target variables is the largest or the standard deviation of the prediction accuracy of the n target variables is the smallest.

2. A data estimation device characterized by comprising a computer, and a learning section executed by the computer, the learning section generating a machine learning model for estimating a target variable from an explanatory variable using training data containing the explanatory variable and the target variable, the explanatory variable including primary data measured by a sensor in relation to a runner's running and physical characteristic data of the runner, and the target variable including secondary data in relation to running performance, the learning section selecting a target variable having the highest prediction accuracy in which the target variable is predicted from the explanatory variable using the machine learning model, among n target variables, as the target variable O1, 3. A data estimation method characterized by comprising a learning step executed by a computer, the learning step being for generating a machine learning model for estimating a target variable from an explanatory variable using training data containing the explanatory variable and the target variable, the explanatory variable including primary data measured by a sensor in relation to a runner's running and physical characteristic data of the runner, and the target variable including secondary data in relation to running performance, The learning unit generates a machine learning model M i that estimates a target variable O i based on an explanatory variable group E i that includes one or more explanatory variables i . The machine learning model M i estimates a target variable O i , adds the target variable O i＋1 estimated by the machine learning model M i＋1 to the explanatory variable group E i＋1 , and sets a new explanatory variable group E i＋1 . The machine learning model M i = 1, the learning step repeatedly generating the machine learning model in correspondence with i = 2 to (n-1), n being a natural number of 2 or more, The learning unit is based on the selected target variables O1 to O i The target variable O is selected in descending order of correlation. i＋1 , i=1~(n-1). the learning step generating the machine learning model in accordance with each input order and calculating the prediction accuracy of the n target variables in order to determine the input order of the n target variables as explanatory variables, and finally selecting the input order in which the average of the prediction accuracy of the n target variables is the largest or the standard deviation of the prediction accuracy of the n target variables is the smallest. the data estimation program being executed by a computer to: The learning step is to generate an explanatory variable group E containing one or more explanatory variables. i Estimated target variable O i Machine learning model M i , will be represented by the machine learning model M i The estimated target variable O i Add to explanatory variable group E i and set a new explanatory variable group E i＋1 , generate the explanatory variable group E i＋1 Estimated target variable O i＋1 Machine learning model M i＋1 , where i = 1, cause the computer to execute a learning step, the learning step being for generating a machine learning model for estimating a target variable from an explanatory variable using training data containing the explanatory variable and the target variable, the explanatory variable including primary data measured by a sensor in relation to a runner's running and physical characteristic data of the runner, and the target variable including secondary data in relation to running performance, the learning step repeatedly generating the machine learning model in correspondence with i = 2 to (n-1), n being a natural number of 2 or more, 4. A computer-readable storage medium storing a data estimation program, characterized by, ​ ​ The learning step is to generate an explanatory variable group E containing one or more explanatory variables. i Estimated target variable O i Machine learning model M i , will be represented by the machine learning model M i The estimated target variable O i Add to explanatory variable group E i and set a new explanatory variable group E i＋1 , generate the explanatory variable group E i＋1 Estimated target variable O i＋1 Machine learning model M i＋1 , where i = 1, ​ The learning step generates machine learning models according to each input order in order to determine an input order of the n target variables as explanatory variables, and calculates prediction accuracies of the n target variables, and finally selects an input order in which an average of the prediction accuracies of the n target variables is maximum or a standard deviation of the prediction accuracies of the n target variables is minimum.

Images