How to learn model parameters

Abstract

To provide a learning method of a model parameter capable of accurately learning the model parameter of a prediction model for predicting a feeling of a formulation.SOLUTION: A learning apparatus 1 acquires an actual value of a feeling as teacher data (STEP 11), acquires a feature quantity including a physical property value of a formulation (STEP 13, 15, and 16), creates learning data using the feature quantity and the teacher data (STEP 17), and performs learning of a model parameter of a prediction model according to a type of the feeling by using a learning algorithm to which a machine learning algorithm with a teacher is applied and the learning data (STEP 19, 21, and 22).SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present invention relates to a method for learning model parameters of a predictive model for predicting the feel of a formulation. [Background technology]

[0002] The present applicant has already proposed a method for learning model parameters of a prediction model, as described in Patent Document 1. In this learning method, model parameters of a prediction model in which the physical property values ​​of a formulation are used as objective variables are learned using a predetermined learning algorithm. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Application No. 2020-126447 (confidential) Summary of the Invention [Problem to be solved by the invention]

[0004] When predicting the properties of a formulation, there is a demand for predicting properties other than physical values. For example, when the formulation is a cosmetic, the feel of the cosmetic (feel on the skin) is an important factor that is directly linked to the commercial value, so a method is used in which a prototype of the cosmetic is actually produced and its feel is actually confirmed. However, this method leads to an increase in the number of steps and manufacturing costs, so it is desirable to use a predictive model to accurately predict the feel of formulations such as cosmetics.

[0005] The present invention has been made to solve the above-mentioned problems, and has an object to provide a model parameter learning method that can accurately learn model parameters of a predictive model for predicting the feel of a formulation. [Means for solving the problem]

[0006] In order to achieve the above object, the invention according to claim 1 provides a model parameter learning method for learning model parameters of a prediction model that predicts the texture of a preparation produced by blending a plurality of raw materials as a response variable by a learning device, the learning device performing a teacher data acquisition step of acquiring actual measured values ​​of the response variable as teacher data, a feature amount acquisition step of acquiring feature amounts including physical property values ​​of the preparation, a learning data creation step of creating learning data using the feature amounts and the teacher data, and a learning step of executing learning of the model parameters of the prediction model by using a learning algorithm to which a supervised machine learning algorithm is applied and the learning data. The formulation is a monophasic formulation, and in the feature acquisition step, one of a first acquisition method of acquiring only the physical property values ​​of the formulation as features, a second acquisition method of acquiring first selected data selected from the blending ratios of multiple raw materials and the physical property values ​​of the formulation as features, and a third acquisition method of acquiring second selected data selected from the blending ratios of multiple raw materials and the physical property values ​​of the formulation as features is selected according to the type of feel, and in the learning step, one of a regression model and a classification model is selected as the prediction model according to the type of feel, and one of an algorithm in which a multivariate prediction method is applied to a supervised machine learning algorithm and a supervised machine learning algorithm is selected according to the type of feel. It is characterized by:

[0007] According to this model parameter learning method, learning data is created using features including the physical properties of the formulation and teacher data that is the actual measured value of the objective variable, and model parameters of the prediction model are learned using a learning algorithm to which a supervised machine learning algorithm is applied and the learning data. In contrast, as described below, the applicant's experiments have confirmed that when learning model parameters of a prediction model that predicts the texture of a formulation as an objective variable, in the case of a texture such as "thick", the learning accuracy of the model parameters is improved by using only the physical properties of the formulation as features. Therefore, according to this model parameter learning method, the learning accuracy can be improved when learning model parameters of a prediction model that predicts the texture of a formulation such as "thick" as an objective variable.

[0009] As described below, the applicant's experiments have confirmed that when learning model parameters of a prediction model that predicts the feel of a single-phase formulation as a dependent variable, the accuracy of learning the model parameters can be improved by selecting one of the above-mentioned first to third acquisition methods depending on the type of feel to acquire features, selecting either a regression model or a classification model as the prediction model, and selecting either an algorithm in which a multivariate prediction method is applied to a supervised machine learning algorithm, or a supervised machine learning algorithm as the learning algorithm.

[0010] Therefore, according to this model parameter learning method, by selecting one of the first to third acquisition methods depending on the type of feel, and selecting the prediction model and learning algorithm as described above, it is possible to improve the learning accuracy of the model parameters of a prediction model that predicts the feel of a single-phase formulation as a dependent variable.

[0011] Claim 2 The invention according to claim 1 In the model parameter learning method described in the above, the second acquisition technique includes a first selection step in which, using the feature values, which are the blending ratios of a plurality of ingredients, and teacher data as learning data, learning of model parameters of the machine learning model is performed using a supervised machine learning algorithm, and the learned machine learning model is used to calculate the importance of the feature values ​​and select the feature values ​​in order of highest importance; a second selection step in which, by applying a first filter process to the feature values, which are the blending ratios of the plurality of ingredients, a correlation between the target variable and the feature values ​​is obtained and the feature values ​​are selected in order of highest correlation with the target variable; a third selection step in which, by applying a second filter process different from the first filter process to the feature values, which are the blending ratios of the plurality of ingredients, a contribution of the feature values ​​to the target variable is obtained and the feature values ​​are selected in order of highest contribution; and a fourth selection step in which, when the feature value with the highest average value is ranked first, the feature values ​​from first rank to a predetermined rank are selected as first selection data.

[0012] According to this model parameter learning method, the second acquisition method executes first to fourth selection steps. In this first selection step, the model parameters of the machine learning model are learned by a supervised machine learning algorithm using the feature values, which are the blending ratios of a plurality of ingredients, and the teacher data as learning data, and the importance of the feature values ​​is acquired using the machine learning model after learning, and the feature values ​​are selected in order of the highest importance. Therefore, the feature values ​​can be selected in order of the most important in the learning of the supervised machine learning model.

[0013] In the second selection step, a first filter process is applied to the feature values, which are the blending ratios of the ingredients, to obtain correlations between the objective variable and the feature values, and the feature values ​​are selected in descending order of correlation with the objective variable. Therefore, the feature values ​​can be selected in descending order of correlation with the objective variable.

[0014] Furthermore, in the third selection step, a second filter process different from the first filter process is applied to the feature values, which are the blending ratios of the multiple ingredients, to obtain the contributions of the feature values ​​to the objective variable, and the feature values ​​are selected in descending order of contribution. Therefore, the feature values ​​can be selected in descending order of contribution to the objective variable.

[0015] In the fourth selection step, the average value of the ranks of the feature amounts selected in the first to third selection steps is calculated, and the feature amounts from the first place to a predetermined rank, with the feature amount with the highest average value being ranked first, are selected as the first selection data, so that the first selection data can be appropriately selected while suppressing the effects of variance and noise in the selection results of the feature amounts in the first to third selection steps, thereby improving the suitability and selection accuracy of the first selection data.

[0016] Claim 3 The invention according to claim 1 or 2In the model parameter learning method described in the above, the third acquisition technique includes a fifth selection step in which feature values, which are the blending ratios and physical property values ​​of a plurality of raw materials, and teacher data are used as learning data to learn model parameters of a machine learning model using a supervised machine learning algorithm, and the learned machine learning model is used to calculate the importance of the feature values ​​and select the feature values ​​in order of highest importance; a sixth selection step in which a third filter process is applied to the feature values, which are the blending ratios and physical property values ​​of a plurality of raw materials, to obtain a correlation between the target variable and the feature values ​​and select the feature values ​​in order of highest correlation with the target variable; a seventh selection step in which a fourth filter process different from the third filter process is applied to the feature values, which are the blending ratios and physical property values ​​of a plurality of raw materials, to obtain a contribution of the feature values ​​to the target variable and select the feature values ​​in order of highest contribution; and an eighth selection step in which an average value of the ranks of the feature values ​​selected in the fifth to seventh selection steps is calculated, and the feature values ​​from the first place to a predetermined rank when the feature value with the highest average value is set as the second selection data are selected.

[0017] According to this model parameter learning method, the third acquisition method executes the fifth to eighth steps. In this fifth selection step, the model parameters of the machine learning model are learned by a supervised machine learning algorithm using the feature amounts, which are the blending ratios and physical property values ​​of a plurality of raw materials, and the teacher data as learning data, and the importance of the feature amounts is acquired using the machine learning model after learning, and the feature amounts are selected in order of the highest importance. Therefore, the feature amounts can be selected in order of the most important in the learning of the supervised machine learning model.

[0018] In the sixth selection step, a third filter process is applied to the feature quantities, which are the blending ratios and physical property values ​​of a plurality of raw materials, to obtain correlations between the objective variables and the feature quantities, and the feature quantities are selected in descending order of correlation with the objective variables. Therefore, the feature quantities can be selected in descending order of correlation with the objective variables.

[0019] Furthermore, in the seventh selection step, a fourth filter process different from the third filter process is applied to the feature quantities, which are the blending ratios and physical property values ​​of a plurality of raw materials, to obtain the contributions of the feature quantities to the objective variable, and the feature quantities are selected in descending order of contribution. Therefore, the feature quantities can be selected in descending order of contribution to the objective variable, and the appropriateness and selection accuracy of the feature quantity selection results can be improved.

[0020] In the eighth selection step, the average value of the ranks of the feature amounts selected in the fifth to seventh selection steps is calculated, and the feature amounts from the first place to a predetermined rank, with the feature amount with the highest average value being ranked first, are selected as second selection data, so that the second selection data can be appropriately selected while suppressing the effects of variance and noise in the selection results of the feature amounts in the fifth to seventh selection steps, thereby improving the suitability and selection accuracy of the second selection data.

[0021] In order to achieve the above-mentioned object, the invention according to claim 4 provides a model parameter learning method for learning model parameters of a prediction model that predicts the texture of a preparation produced by blending a plurality of raw materials as a response variable by a learning device, the learning device performing a teacher data acquisition step of acquiring actual measured values ​​of the response variable as teacher data, a feature acquisition step of acquiring feature amounts including physical property values ​​of the preparation, a learning data creation step of creating learning data by using the feature amounts and the teacher data, and a learning step of executing learning of the model parameters of the prediction model by using a learning algorithm to which a supervised machine learning algorithm is applied and the learning data, The formulation is an emulsion-based formulation, and in the feature acquisition step, one of a first acquisition method of acquiring only the physical property values ​​of the formulation as features, a second acquisition method of acquiring first selected data selected from a plurality of raw material blending ratios and formulation manufacturing methods and the physical property values ​​of the formulation as features, and a third acquisition method of acquiring second selected data selected from a plurality of raw material blending ratios, the physical property values ​​of the formulation, and the manufacturing method of the formulation as features is executed according to the type of feel, and in the learning step, one of a regression model and a classification model is selected as the prediction model according to the type of feel, and an algorithm in which a multivariate prediction method is applied to a supervised machine learning algorithm, and one of the supervised machine learning algorithms is selected according to the type of feel.

[0022] As described below, the applicant's experiments have confirmed that when learning model parameters of a prediction model that predicts the feel of an emulsion-based formulation as a target variable, the accuracy of learning the model parameters can be improved by selecting one of the above-mentioned first to third acquisition methods depending on the type of feel to acquire features, selecting either a regression model or a classification model as the prediction model, and selecting either an algorithm in which a multivariate prediction method is applied to a supervised machine learning algorithm, or a supervised machine learning algorithm as the learning algorithm.

[0023] Therefore, according to this model parameter learning method, by selecting one of the first to third acquisition methods depending on the type of feel, and selecting the prediction model and learning algorithm as described above, it is possible to improve the learning accuracy of the model parameters of a prediction model that predicts the feel of an emulsion-based formulation as a target variable.

[0024] Claim 5 The invention according to claim 4 In the model parameter learning method described in the above, the second acquisition method includes a first selection step in which feature amounts, which are the blending ratios of multiple raw materials and the manufacturing method of the formulation, and teacher data are used as learning data to learn a machine learning model using a supervised machine learning algorithm, and the learned machine learning model is used to calculate the importance of the feature amounts and select the feature amounts in order of highest importance; a second selection step in which a first filter process is applied to the feature amounts, which are the blending ratios of multiple raw materials and the manufacturing method of the formulation, to select the feature amounts in order of highest correlation with the target variable; a third selection step in which a second filter process different from the first filter process is applied to the feature amounts, which are the blending ratios of multiple raw materials and the manufacturing method of the formulation, to obtain the contribution of the feature amounts to the target variable and select the feature amounts in order of highest contribution; and a fourth selection step in which an average value of the ranks of the feature amounts selected in the first to third selection steps is calculated, and the feature amounts from the first rank to a predetermined rank when the feature amount with the highest average value is set as the first selection data are executed.

[0025] According to this model parameter learning method, the second acquisition method executes steps 1 to 4. In this first selection step, the model parameters of the machine learning model are learned by a supervised machine learning algorithm using the feature values ​​and teacher data, which are the blending ratios of multiple raw materials and the manufacturing method of the formulation, as learning data, and the importance of the feature values ​​is acquired using the machine learning model after learning, and the feature values ​​are selected in order of the highest importance. Therefore, the feature values ​​can be selected in order of the most important in the learning of the supervised machine learning model.

[0026] In the second selection step, a first filter process is applied to the feature quantities, which are the blending ratios of the raw materials and the manufacturing method of the formulation, to obtain correlations between the objective variables and the feature quantities, and the feature quantities are selected in descending order of correlation with the objective variables. Therefore, the feature quantities can be selected in descending order of correlation with the objective variables.

[0027] Furthermore, in the third selection step, a second filter process different from the first filter process is applied to the feature quantities, which are the blending ratios of the multiple raw materials and the manufacturing method of the formulation, to obtain the contributions of the feature quantities to the objective variable, and the feature quantities are selected in descending order of contribution. Therefore, the feature quantities can be selected in descending order of contribution to the objective variable.

[0028] In the fourth selection step, the average value of the ranks of the feature amounts selected in the first to third selection steps is calculated, and the feature amounts from the first place to a predetermined rank, with the feature amount with the highest average value being ranked first, are selected as the first selection data, so that the first selection data can be appropriately selected while suppressing the effects of variance and noise in the selection results of the feature amounts in the first to third selection steps, thereby improving the suitability and selection accuracy of the first selection data.

[0029] Claim 6 The invention according to claim 4or 5 In the model parameter learning method described in the above, the third acquisition method includes a fifth selection step of using the feature values, which are the blending ratios of multiple raw materials, physical properties, and the manufacturing method of the formulation, and training data as training data to learn a machine learning model using a supervised machine learning algorithm, calculating the importance of the feature values ​​using the trained machine learning model, and selecting the feature values ​​in order of the highest importance, and a fifth selection step of applying a third filter process to the feature values, which are the blending ratios of multiple raw materials, physical properties, and the manufacturing method of the formulation, to obtain correlations between the objective variable and the feature values, and selecting the feature values ​​in order of the highest correlation with the objective variable. a sixth selection step of selecting a feature quantity that is a blending ratio of a plurality of raw materials, a physical property value, and a manufacturing method of the formulation, by applying a fourth filter process different from the third filter process to the feature quantities, thereby obtaining the contribution of the feature quantity to the target variable, and selecting the feature quantities below the first place in order of contribution when the feature quantity with the highest contribution is ranked first; and an eighth selection step of calculating the average rank of the feature quantities selected in the fifth to seventh selection steps, and selecting the feature quantities from the first place to a predetermined rank when the feature quantity with the highest average value is ranked first as second selection data.

[0030] According to this model parameter learning method, the third acquisition method executes steps 5 to 8. In this fifth selection step, the model parameters of the machine learning model are learned by a supervised machine learning algorithm using the feature values ​​and teacher data, which are the blending ratios of a plurality of raw materials, the physical properties, and the manufacturing method of the formulation, as learning data, and the importance of the feature values ​​is acquired using the machine learning model after learning, and the feature values ​​are selected in order of the highest importance. Therefore, the feature values ​​can be selected in order of the most important in the learning of the supervised machine learning model.

[0031] In the sixth selection step, a third filter process is applied to the feature quantities, which are the blending ratios of the raw materials, the physical properties, and the manufacturing method of the formulation, to obtain correlations between the objective variables and the feature quantities, and the feature quantities are selected in descending order of correlation with the objective variables. Thus, the feature quantities can be selected in descending order of correlation with the objective variables.

[0032] Furthermore, in the seventh selection step, a fourth filter process different from the third filter process is applied to the features, which are the blending ratios of multiple raw materials, physical property values, and the manufacturing method of the formulation, to obtain the contributions of the features to the objective variable, and the features are selected in descending order of contribution. Therefore, the features can be selected in descending order of contribution to the objective variable, and the appropriateness and selection accuracy of the feature selection results can be improved.

[0033] In the eighth selection step, the average value of the ranks of the feature amounts selected in the fifth to seventh selection steps is calculated, and the feature amounts from the first place to a predetermined rank, with the feature amount with the highest average value being ranked first, are selected as second selection data, so that the second selection data can be appropriately selected while suppressing the effects of variance and noise in the selection results of the feature amounts in the fifth to seventh selection steps, thereby improving the suitability and selection accuracy of the second selection data. [Brief description of the drawings]

[0034] [Figure 1] FIG. 1 is a diagram showing a learning device for executing a model parameter learning method according to an embodiment of the present invention. [Diagram 2] FIG. 13 is a diagram illustrating an example of features of a database used for learning. [Diagram 3] FIG. 13 is a diagram illustrating an example of a target variable of a database used for learning. [Figure 4] 4 is a flowchart showing a learning process performed by the learning device. [Diagram 5] 13 is a flowchart showing a learning process for a one-phase system. [Figure 6] FIG. 13 is a diagram showing an example of a sensation for which good results were obtained when learning of a predictive model was performed. [Figure 7] FIG. 13 is a diagram showing the results of learning a prediction model for the sensations of “stretchy” and “fresh” using learning patterns 1 to 3. [Figure 8] FIG. 13 is a diagram showing the results when a regression model is used as a prediction model for the sensations of “rough” and “warm.” [Figure 9] FIG. 13 shows the results when a classification model is used as a predictive model for the sensations of “rough” and “warm.” [Figure 10] FIG. 13 is a diagram showing an example of a feel for which good results were obtained when a predictive model was trained using an algorithm that applies a multivariate prediction technique. [Figure 11] FIG. 13 is a diagram showing the results of a prediction model for the sensations of “coating” and “comfort” trained using a multivariate prediction method and trained without using the multivariate prediction method. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0035] Hereinafter, a method for learning model parameters according to one embodiment of the present invention will be described with reference to the drawings. In this embodiment, when a mixture of various raw materials is stirred to produce a plurality of cosmetic products as a plurality of formulations, the learning method described below is used to learn model parameters of a predictive model that predicts the texture of these formulations as a response variable. In the following description, learning the model parameters of the predictive model will be referred to as "learning the predictive model" as appropriate.

[0036] The learning method of this embodiment is specifically executed by a learning device 1 shown in Fig. 1. This learning device 1 is a personal computer type, and includes a display 1a, a device main body 1b, and an input interface 1c. The device main body 1b includes a storage device such as an HDD, a processor, and a memory (RAM, ROM, etc.) (none of which are shown).

[0037] A database is stored in the memory of the device main body 1b, which includes various formulation names, formulation feature values ​​(physical properties, etc.), and actual measured values ​​of objective variables (feel) as shown in Figures 2 and 3.

[0038] In this database, the various formulations include one-phase formulations such as lotions, creams, and serums, as well as two-phase formulations. Here, one-phase formulations are formulations such as lotions that are composed of an aqueous solution or oil, and two-phase formulations are formulations with an emulsion formula, i.e., formulations such as creams that are composed of a mixture of water and oil.

[0039] The characteristic quantities include physical property values ​​of various preparations. Specifically, in addition to the viscosity, dry residual rate, specific gravity, and thermal conductivity of various preparations, specific gravity, dry residual rate, shear stress, contact angle, thermal conductivity, peeling force, and friction coefficient (not shown) are included. Here, the dry residual rate represents the weight ratio of the preparation before and after drying, and the peeling force represents the peeling force per unit area of ​​the preparation.

[0040] Furthermore, when the preparation is an emulsion-based preparation such as a cream, i.e., a two-phase preparation, the characteristic quantities include the blending ratios of the raw materials (pure water, 1,3-BG, petrolatum, etc.) and the manufacturing method (mixing method, mixing time, mixing rotation speed, and emulsification temperature). These blending ratios are set to values ​​normalized based on the maximum blending values ​​of the various raw materials, and these maximum blending values ​​are calculated from the applicant's past performance values.

[0041] On the other hand, the measured values ​​of the feel, which is the objective variable, include the following 18 items of data (A1) to (A18) (only some of which are shown in Figure 3). These 18 measured values ​​of the feel were obtained using the CATA (Check-All-That-Apply) method.

[0042] (A1) soft, (A2) hard, (A3) smooth, (A4) moist, (A5) silky, (A6) rough, (A7) warm, (A8) cold, (A9) sticky, (A10) thick, (A11) firm, (A12) stretchy, (A13) absorbed, (A14) penetrating, (A15) coating, (A16) oily, (A17) moist, (A18) plump.

[0043] The above items (A1) to (A10) are general touch items based on the tactile dimension, and the items (A11) to (A18) are touch items that are frequently used in the field of cosmetic preparation production. In the case of this embodiment, the database is configured as above, and the reason for this will be described later.

[0044] On the other hand, application software for executing the learning process described below is installed in the storage of the device main body 1b. The input interface 1c is composed of a keyboard, a mouse, and the like for operating the learning device 1.

[0045] Next, the contents of the learning process executed by the learning device 1 of this embodiment will be described with reference to Fig. 4. In this learning process, as described below, learning of a prediction model is executed using the above-mentioned database. This prediction model is for predicting the above-mentioned 18 texture items of a preparation.

[0046] As shown in the figure, in this learning process, first, it is determined whether the formulation for which learning is to be performed this time is a monophasic formulation (FIG. 4 / STEP 1). If this determination is positive (FIG. 4 / STEP 1...YES) and the formulation is a monophasic formulation, learning processing for monophasic systems is executed (FIG. 4 / STEP 2). After that, this processing ends. The contents of this learning processing for monophasic systems will be described later.

[0047] On the other hand, if the above-mentioned determination is negative (FIG. 4 / STEP 1...NO) and the preparation is a two-phase system, a learning process for a two-phase system is executed (FIG. 4 / STEP 3). After that, this process ends. The contents of this learning process for a two-phase system will be described later.

[0048] Next, the contents of the learning process for the one-phase system described above will be described with reference to Fig. 5. As shown in the figure, first, a prediction model determination process is executed (Fig. 5 / STEP 10). In this determination process, the prediction model to be the target of the current learning process is determined to be one of the 18 prediction models that predict the 18 tactile sensations described above (A1) to (A18) as objective variables.

[0049] Specifically, this determination is made based on the result of an operation of the input interface 1c by an operator (not shown). In the following description, the feel of the objective variable of the prediction model that is the subject of the current learning process is referred to as the "feel of the learning target."

[0050] Next, a process of reading out the teacher data is executed (FIG. 5 / STEP 11). Specifically, the actual measurement value of the feel of the learning object determined in the above determination process is read out as the teacher data from the above database.

[0051] Next, it is determined whether the texture to be learned is a texture of Pattern 1 (Fig. 5 / STEP 12). This texture of Pattern 1 is a type of texture for which the prediction accuracy of the prediction model is most improved by implementing the learning method of Pattern 1 described below. Specifically, a texture such as "thick and creamy" corresponds to the texture of Pattern 1.

[0052] Here, the learning method of Pattern 1 is a method in which only the physical property values ​​of the preparation are acquired from the database as the feature quantities of Pattern 1 (first acquisition method), and the feature quantities of Pattern 1 are used to learn the prediction model. In the case of the feel of Pattern 1, it has been confirmed through an experiment by the present applicant, which will be described later, that the prediction accuracy of the prediction model is most improved by using the learning method of Pattern 1.

[0053] If this determination is positive (FIG. 5 / STEP 12...YES), and the learning target texture is the texture of pattern 1, the feature amount of pattern 1 is read out from the database (FIG. 5 / STEP 13).

[0054] On the other hand, when the above-mentioned determination is negative (FIG. 5 / STEP 12...NO), that is, the feel to be learned is not the feel of pattern 1, it is determined whether or not the feel to be learned is the feel of pattern 2 (FIG. 5 / STEP 14).

[0055] This pattern 2 feel is a type of feel for which the prediction accuracy of the prediction model is most improved by implementing the learning method of pattern 2 described below. Specifically, the feel of pattern 2 includes "stretchy" and "smooth."

[0056] Here, the learning method of pattern 2 is a method in which first selection data selected from the blending ratios of the raw materials of the formulation by the methods of steps (B2) to (B5) described below and the physical property values ​​of the formulation are acquired as feature quantities of pattern 2 (second acquisition method), and a prediction model is learned using the feature quantities of pattern 2. In the case of the feel of pattern 2, it has been confirmed through experiments by the applicant that the prediction accuracy of the prediction model is most improved by using the learning method of pattern 2.

[0057] If this determination is positive (FIG. 5 / STEP 14...YES), and the learning target texture is the texture of pattern 2, the feature amount of pattern 2 is read from the database (FIG. 5 / STEP 15). Next, the process proceeds to the creation of learning data (FIG. 5 / STEP 17), which will be described later.

[0058] On the other hand, if the above-mentioned judgment is negative (FIG. 5 / STEP 14...NO), and the feel of the learning subject is not the feel of pattern 2, the feel of the learning subject is determined to be the feel of pattern 3, and the feature values ​​of pattern 3 are read out from the database (FIG. 5 / STEP 16).

[0059] The pattern 3 feel is a type of feel for which the prediction accuracy of the prediction model is most improved by implementing the learning method of pattern 3 described below. Specifically, the pattern 3 feel includes "fresh," "moist," and "sticky."

[0060] Here, the learning method of Pattern 3 is a method in which second selection data selected from the blending ratios of the raw materials of the formulation and the physical properties of the formulation by the methods of steps (B2) to (B5) described below is acquired as feature quantities of Pattern 3 (third acquisition method), and a prediction model is learned using the feature quantities of Pattern 3. In the case of the feel of Pattern 3, it has been confirmed by the applicant's experiments described below that the prediction accuracy of the prediction model is most improved by using the learning method of Pattern 3.

[0061] After the feature quantity readout process of any one of the above patterns 1 to 3 is executed, a learning data creation process is executed (FIG. 5 / STEP 17). In this creation process, learning data is created so as to include any one of the feature quantities of patterns 1 to 3 read out as described above and the teacher data (actually measured values ​​of the feel).

[0062] Next, it is determined whether the texture to be learned is a texture of a classification model (FIG. 5 / STEP 18). Here, a texture of a classification model is a texture for which the prediction accuracy can be improved when a classification model is used as a prediction model, compared to when a regression model is used as a prediction model. Specifically, the two textures "rough" and "warm" correspond to the textures of the classification model.

[0063] If this determination is positive (FIG. 5 / STEP 18...YES), and the feel to be learned is a feel of the classification model, a classification model learning process is executed (FIG. 5 / STEP 19). Specifically, the model parameters of the classification model are learned using the learning data described above.

[0064] In this case, the classification model is created by applying the XGBoost algorithm. In this embodiment, the XGBoost algorithm corresponds to a supervised machine learning algorithm. After the classification model learning process is performed as described above, this process ends.

[0065] On the other hand, when the above-mentioned judgment is negative (FIG. 5 / STEP 18...NO), and the learning target feel is not a classification model feel, it is judged whether the learning target feel is a regression model feel or not (FIG. 5 / STEP 20). Here, the regression model feel is a feel that can ensure high prediction accuracy when the regression model is used as a prediction model. Specifically, the regression model feel includes "stretchy," "fresh," and "moist."

[0066] If the determination is positive (FIG. 5 / STEP 20...YES), and the feel to be learned is the feel of the regression model, the regression model learning process is executed (FIG. 5 / STEP 21). Specifically, the model parameters of the regression model are learned using the learning data described above.

[0067] In this case, the regression model is created by applying the XGBoost algorithm. In this embodiment, the XGBoost algorithm corresponds to a supervised machine learning algorithm. After the regression model learning process is performed as described above, this process ends.

[0068] On the other hand, when the above-mentioned judgment is negative (FIG. 5 / STEP 20...NO), and the learning target feel is not the feel of the regression model, the MP learning process is executed (FIG. 5 / STEP 22). In this MP learning process, as described below, learning of a prediction model is executed using a learning algorithm that applies a multivariate prediction method.

[0069] The following describes an example in which learning of a prediction model for the "coating" sensation is performed. First, as learning step 1, the following learning processes (1A) to (1F) are performed.

[0070] (1A) First, using the above-mentioned learning data, a prediction model of the feeling of "hard" is learned by a random forest algorithm. Note that in this embodiment, the random forest algorithm corresponds to a supervised machine learning algorithm.

[0071] (1B) Next, learning data 1B is created by adding the score (predicted value) of the “hard” feeling in the predictive model after the learning process in (1A) above as a feature to the aforementioned learning data, and this learning data 1B is used to learn a predictive model for the “firm” feeling using a random forest algorithm.

[0072] (1C) Next, learning data 1C is created by adding the score for the “firm” feel in the predictive model after the learning process in (1B) above as a feature to the above-mentioned learning data 1B, and this learning data 1C is used to learn a predictive model for the “fresh” feel using a random forest algorithm.

[0073] (1D) Furthermore, by adding the score for the “fresh” sensation in the predictive model after the learning process in (1C) above as a feature to the above-mentioned learning data 1C, learning data 1D is created, and using this learning data 1D, a predictive model for the “soft” sensation is trained by a random forest algorithm.

[0074] (1E) Next, learning data 1E is created by adding the score for the feeling of “penetrating” in the predictive model after the learning process of (1D) above as a feature to the above-mentioned learning data 1D, and using this learning data 1E, learning of a predictive model for the feeling of “cold” is performed using a random forest algorithm.

[0075] (1F) Next, the score of the "cold" feeling in the prediction model that has completed the learning in (1E) above is added as a feature to the learning data 1E described above to create learning data 1F, and using this learning data 1F, a prediction model of the "coating" feeling is trained by a random forest algorithm. Then, the predicted value of the "coating" feeling by the learned prediction model is set to the predicted value 1.

[0076] Following the above learning step 1, the following learning processes (2A) to (2G) are executed as learning step 2.

[0077] (2A) First, using the training data described above, a random forest algorithm is used to train a predictive model for the “smooth” texture.

[0078] (2B) Next, learning data 2B is created by adding the score for the “smooth” texture in the predictive model after the learning process in (2A) above as a feature to the aforementioned learning data. Using this learning data 2B, a predictive model for the “hard” texture is trained using a random forest algorithm.

[0079] (2C) Next, learning data 2C is created by adding the score for the “hard” texture in the predictive model after the learning process in (2B) has been completed as a feature to the above-mentioned learning data 2B, and this learning data 2C is used to learn a predictive model for the “fresh” texture using a random forest algorithm.

[0080] (2D) Furthermore, by adding the score for the “fresh” sensation in the predictive model after the learning process of (2C) above as a feature to the above-mentioned learning data 2C, learning data 2D is created, and using this learning data 2D, a predictive model for the “cold” sensation is trained by a random forest algorithm.

[0081] (2E) Next, learning data 2E is created by adding the score for the “cold” feeling in the predictive model after the learning process of (2D) above as a feature to the above-mentioned learning data 2D, and using this learning data 2E, a predictive model for the “fluffy” feeling is trained using a random forest algorithm.

[0082] (2F) Next, learning data 2F is created by adding the score for the “plump” feeling in the predictive model that has completed the training in (2E) as a feature to the training data 2E described above, and this training data 2F is used to train a predictive model for the “firm” feeling using a random forest algorithm.

[0083] (2G) Furthermore, by adding the score of the "firm" feeling in the prediction model that has completed the learning in (2F) above as a feature to the learning data 2F described above, learning data 2G is created, and using this learning data 2G, a prediction model of the "coated" feeling is learned by a random forest algorithm. Then, the predicted value of the "coated" feeling by the learned prediction model is set to predicted value 2.

[0084] Following the above learning step 2, a learning step similar to learning steps 1 and 2 is executed multiple times to calculate N predicted values ​​1 to N (N is a plural number). Then, finally, the average value of these N predicted values ​​1 to N is determined as the predicted value of the "coated" feeling according to the prediction model. After the MP learning process is executed as described above, this process ends.

[0085] In this embodiment, the learning process for the one-phase system (FIG. 4 / STEP 2) is executed as described above. On the other hand, although not shown, the learning process for the two-phase system (FIG. 4 / STEP 3) is executed in the same manner as the learning process for the one-phase system shown in FIG. 5, except for the points described below.

[0086] That is, in the learning process for a two-phase system, in the process corresponding to STEP 13 in FIG. 5, only the physical property values ​​of the formulation are acquired as feature quantities of pattern 1 (first acquisition method).

[0087] In addition, in the process corresponding to STEP 15 in FIG. 5, first selection data selected from the mixing ratio of the raw materials of the formulation and the manufacturing method of the formulation by a first selection method described below, and the physical property values ​​of the formulation are acquired as features of pattern 2 (second acquisition method).

[0088] Furthermore, in the process corresponding to STEP 16 in FIG. 5, second selection data selected from the mixing ratio of the raw materials of the formulation, the physical properties of the formulation, and the manufacturing method of the formulation by a second selection method described below is acquired as a feature of pattern 3 (third acquisition method).

[0089] As described above, in the learning process for two-phase systems, the manufacturing method of the formulation is used as the feature amount for patterns 2 and 3. This is because two-phase formulations are emulsion-based formulations, and the manufacturing method has a large effect on the feel of the formulation.

[0090] Next, the reason and principle of using the above learning method in this embodiment will be described. First, the applicant conducted an experiment to learn a prediction model of the feel of a formulation using a method similar to that of Japanese Patent Application No. 2020-126447, as shown in the following steps (B1) to (B6).

[0091] (B1) That is, a database of features was created that includes the blending ratio of various ingredients, the preparation method, the mixing method, the mixing time, the mixing rotation speed, and the phase state, etc. In this database, the blending ratio of various ingredients was normalized, ingredients that have little effect on the texture were omitted, and features that cause multicollinearity were omitted.

[0092] (B2) Furthermore, a prediction model (supervised machine learning model) was created by learning all the features using the random forest algorithm, and the importance of the features was calculated using this prediction model, and the features were selected in order of importance. In this case, the importance of the features is calculated according to the frequency of use of the features at the branching points of the decision tree, and the more frequently a feature is used, the higher its importance is calculated to be.

[0093] (B3) In addition, the association between the feature amount and the objective variable (touch score) was calculated by network analysis, and the feature amount was selected in descending order of association with the objective variable.

[0094] (B4) In addition, the contribution of the feature to the actual measured value of the objective variable was calculated using the Relief algorithm, and the feature was selected in descending order of contribution.

[0095] (B5) Next, the average value of the ranks of the three types of features selected as above was calculated, and the feature with the highest average rank was ranked first. Then, the features from this rank first to a specified rank were selected.

[0096] (B6) Then, using the XGBoost algorithm, we created a predictive model that learned the features from the first place to the specified rank.

[0097] Experimental results have revealed that the prediction accuracy of the prediction model created by the above learning method is low. Therefore, the applicant presumed that the learning accuracy of the prediction model for the texture of a formulation would be improved if the physical property values ​​of the formulation were used as feature values ​​instead of the blending ratios of various raw materials and the manufacturing method, and conducted an experiment to train the prediction model using the method described below. In this case, the prediction model was created as a regression model.

[0098] That is, the applicant created a database as shown in the above-mentioned FIGS. 2 and 3, and used this database to train a prediction model with the XGBoost algorithm.

[0099] In this case, in the case of a monophasic formulation, learning was performed using the feature quantities of the above-mentioned patterns 1 to 3. Hereinafter, learning of a prediction model using the feature quantities of patterns 1 to 3, respectively, will be referred to as "learning patterns 1 to 3").

[0100] That is, in learning pattern 1, the prediction model was trained using only the physical properties of the formulation as features.

[0101] In addition, in learning pattern 2, first selection data was selected from the raw material blend ratios using the above-mentioned steps (B2) to (B5), and the prediction model was trained using this first selection data and the physical properties of the formulation as features.

[0102] In this case, steps (B2) to (B5) correspond to the first to fourth selection steps, the network analysis corresponds to the first filtering process, and the Relief algorithm corresponds to the second filtering process.

[0103] Furthermore, in learning pattern 3, second selection data was selected from the physical properties of the formulation and the mixing ratios of the raw materials by the above-mentioned steps (B2) to (B5), and this second selection data was used as features to learn the predictive model.

[0104] In this case, steps (B2) to (B5) correspond to the fifth to eighth selection steps, the network analysis corresponds to the third filtering process, and the Relief algorithm corresponds to the fourth filtering process.

[0105] On the other hand, in the case of a two-phase formulation, learning was performed using the following learning patterns 1 to 3. That is, in learning pattern 1, the prediction model was trained using only the physical property values ​​of the formulation as feature quantities.

[0106] In addition, in learning pattern 2, first selection data was selected from the formulation's manufacturing method and the raw material mixing ratios by the above-mentioned steps (B2) to (B5), and a predictive model was trained using this first selection data and the formulation's physical property values ​​as features.

[0107] In this case, steps (B2) to (B5) correspond to the first to fourth selection steps, the network analysis corresponds to the first filtering process, and the Relief algorithm corresponds to the second filtering process.

[0108] Furthermore, in learning pattern 3, second selection data was selected from the physical properties of the formulation and the mixing ratios of the raw materials by the above-mentioned steps (B2) to (B5), and this second selection data was used as features to learn the predictive model.

[0109] In this case, steps (B2) to (B5) correspond to the fifth to eighth selection steps, the network analysis corresponds to the third filtering process, and the Relief algorithm corresponds to the fourth filtering process.

[0110] As described above, by training a model to predict the texture of one-phase and two-phase formulations, we were able to ensure a certain level of prediction accuracy (prediction error within ±10% of the actual measured value) for the multiple texture scores shown in Figure 6.

[0111] In the experimental results shown in the figure, the phase state "AQ" indicates that the formulation is a one-phase system, and the phase state "EM" indicates that the formulation is a two-phase system. As shown in the figure, the LOO (Leave One Out) method and the 5-fold cross-validation method (5-fold CV) were used as the accuracy evaluation methods, and the NRMSE (Normalized Root Mean Square Error) and the coefficient of determination R 2 was used.

[0112] Here, in the case of the sensations of "stretching" and "freshness" in a single-phase formulation, when learning patterns 1 to 3 were implemented, the learning results shown in Figure 7 were obtained. As shown in the figure, in the case of the sensation of "stretching," as is clear from comparing the learning results of learning patterns 1 to 3, it can be seen that the NRMSE does not change much even when the learning pattern is changed.

[0113] In contrast, the coefficient of determination R 2 It can be seen that the learning results of learning pattern 2 are clearly better than those of learning patterns 1 and 3, and the prediction accuracy is improved. In other words, when training a prediction model (regression model) to predict the "stretching" sensation of a monophasic formulation, it was found that the learning accuracy of the prediction model is most improved by training with learning pattern 2.

[0114] In addition, as shown in Figure 7, when comparing the learning results of learning patterns 1 to 3 for the sensation of "fresh," it can be seen that the NRMSE does not change much even when the learning pattern is changed. In contrast, the coefficient of determination R 2 It can be seen that the learning results of learning pattern 3 are clearly improved compared to learning patterns 1 and 2, and the prediction accuracy is improved.

[0115] In other words, when training a predictive model (regression model) to predict the "fresh" sensation in a monophasic formulation, it was found that training using training pattern 3 maximizes the learning accuracy of the predictive model.

[0116] Based on the above experimental results, it was found that in the case of a prediction model that predicts the sensations of "stretchable," "fresh," "moist," "smooth," "sticky," "thick," "firm," "penetrating," and "cool" for one-phase formulations, and the sensations of "hard," "smooth," "bleeds in," "oily," "moist," and "stretchable" for two-phase formulations, a high prediction accuracy can be ensured by using a regression model and performing learning with the learning pattern shown in Figure 6.

[0117] On the other hand, in the case of the sensations of “rough” and “warm,” regression models were created using the XGBoost algorithm for both the one-phase and two-phase formulations, and training of these regression models was performed. The training results are shown in Figure 8.

[0118] As can be seen from the learning results, in the case of the sensations of "rough" and "warm", the coefficient of determination R 2 It can be seen that the values ​​are low (0.01 to 0.36), and high prediction accuracy cannot be obtained.

[0119] Therefore, instead of the regression model, a classification model that classifies whether or not the texture feels “rough” and “warm” was created using the XGBoost algorithm, and training of this classification model was performed. The training results shown in Figure 9 were obtained.

[0120] As is clear from the learning results, in the case of the sensations of "rough" and "warm", the Kappa Score was 1 for both the one-phase and two-phase formulations, indicating that high prediction accuracy was obtained. In other words, it was found that when learning a prediction model for the sensations of "rough" and "warm" for both the one-phase and two-phase formulations, using a classification model can ensure higher prediction accuracy than using a regression model.

[0121] In addition, when the preparation is a cosmetic, it is well known that some textures of the preparation are correlated with each other. For example, there is a phenomenon that a preparation having a "thick" texture is likely to be associated with a "moist" texture.

[0122] Therefore, we used the multivariate prediction method described above to train a model for predicting the texture of one-phase and two-phase formulations. As a result, we were able to ensure a certain level of prediction accuracy (prediction error within ±10% of the actual measured value) for the multiple texture scores shown in Figure 10.

[0123] Here, FIG. 11 shows the learning results when a multivariate prediction method was implemented to train a predictive model (regression model) for the sensations of “coating” and “smoothness” in a monophasic formulation, and, for comparison, the learning results when the multivariate prediction method was not implemented (i.e., when only learning using the XGBoost algorithm was implemented).

[0124] As is clear from the figure, when the multivariate forecasting method was implemented, the coefficient of determination R 2 It can be seen that the prediction accuracy has improved significantly.

[0125] Based on the above experimental results, it was found that in the case of a prediction model predicting the sensations of "coating," "oily," "smooth," "plump," "hard," and "smooth" in a single-phase formulation, high prediction accuracy can be ensured by using a multivariate prediction method and performing learning with the learning pattern shown in Figure 10.

[0126] Similarly, in the case of a prediction model predicting sensations such as "coating-like," "fresh," "cold," "hard," "penetrating," "sticky," "thick," "plump," "firm," and "smooth" in two-phase formulations, it was found that high prediction accuracy could be ensured by using a multivariate prediction method and performing learning with the learning pattern shown in Figure 10.

[0127] In this embodiment, based on the above experimental results, in order to improve the prediction accuracy of various sensations, the learning process for the one-phase system is executed as shown in FIG. 5, and the learning process for the two-phase system is executed as described above.

[0128] At that time, one of the feature amounts from patterns 1 to 3 is selected according to the type of feel, and further, the feature amount thus selected is used to carry out a learning process for the prediction model. That is, as the learning process for the prediction model, one of the classification model learning process, regression model learning process, and MP learning process is carried out according to the type of feel so that the prediction accuracy after learning is maximized. As a result, a high prediction accuracy can be ensured in the prediction model that predicts the above-mentioned 18 types of feel.

[0129] In the embodiment, the actual measured values ​​of the feel of the formulation are obtained using the CATA method, but instead, the actual measured values ​​of the feel of the formulation may be obtained using a sensory evaluation method such as the QDA (Quantitative Descriptive Analysis) method.

[0130] In addition, the embodiment is an example in which data on 18 types of sensations (A1) to (A18) are used as the actual measured values ​​of the sensation of the formulation, but instead, data on 17 or less types of sensations or 19 or more types of sensations may be used as the actual measured values ​​of the sensation of the formulation.

[0131] Furthermore, the embodiment is an example in which the XGBoost algorithm is used as a supervised machine learning algorithm in the classification model learning process (FIG. 5 / STEP 19) and the regression model learning process (FIG. 5 / STEP 21). Alternatively, a neural network or a support vector machine may be used as the supervised machine learning algorithm.

[0132] On the other hand, the embodiment is an example in which a random forest algorithm is used as a supervised machine learning algorithm in the MP learning process (FIG. 5 / STEP 22), but instead of this, an XGBoost algorithm, a neural network, a support vector machine, or the like may be used.

[0133] Also, in the embodiment, the second acquisition method is an example in which the first selection data is selected by steps (B2) to (B5), but instead of this, the following method may be used: That is, any two of the processes (B2) to (B4) are executed, the average value of the ranks of the two types of feature amounts selected by these two processes is calculated, and the feature amount with the highest average rank is ranked first, and then the feature amounts from this rank to a predetermined rank may be selected as the first selection data.

[0134] Furthermore, in the embodiment, the third acquisition method is an example in which the second selection data is selected by steps (B2) to (B5), but instead of this, the following method may be used: That is, any two of the processes (B2) to (B4) are executed, the average value of the ranks of the two types of feature amounts selected by these two processes is calculated, and the feature amount with the highest average rank is ranked first, and then the feature amounts from this rank to a predetermined rank may be selected as the second selection data.

[0135] Furthermore, in the embodiment, a network analysis method is used as the first filter process and the third filter process, but the first filter process and the third filter process of the present invention are not limited to this, and may be any method capable of acquiring the correlation between the objective variable and the feature amount. For example, a correlation analysis algorithm, an analysis of variance algorithm, chi-square score, Fisher score, anova, or the like may be used as the first filter process and the third filter process. Also, different filter processes may be executed as the first filter process and the third filter process.

[0136] On the other hand, in the embodiment, the Relief algorithm is used as the second filter process and the fourth filter process, but the second filter process and the fourth filter process of the present invention are not limited to this, and may be any filter process that can obtain the contribution of the feature to the objective variable. For example, the second filter process and the fourth filter process may be a correlation analysis algorithm, a variance analysis algorithm, a chi-square score, a Fisher score, anova, or the like. Also, different filter processes may be executed as the second filter process and the fourth filter process.

[0137] Furthermore, the embodiment is an example in which a personal computer type device is used as the learning device, but instead of this, a server or a cloud server may be used as the learning device. [Explanation of symbols]

[0138] 1 Learning device

Claims

1. A method for learning model parameters, in which model parameters of a prediction model that predicts the texture of a preparation produced by blending a plurality of raw materials as a response variable are learned by a learning device, comprising the steps of: The learning device includes: A teacher data acquisition step of acquiring actual measured values ​​of the objective variables as teacher data; A feature acquisition step of acquiring feature values ​​including physical property values ​​of the formulation; a learning data creation step of creating learning data using the feature amount and the teacher data; a learning step of executing learning of model parameters of the prediction model using a learning algorithm to which a supervised machine learning algorithm is applied and the learning data; Run The formulation is a monophasic formulation, In the feature acquisition step, one of a first acquisition method of acquiring only the physical property value of the formulation as the feature value, a second acquisition method of acquiring first selection data selected from the blending ratios of the plurality of ingredients and the physical property value of the formulation as the feature value, and a third acquisition method of acquiring second selection data selected from the blending ratios of the plurality of ingredients and the physical property value of the formulation as the feature value is selected according to the type of sensation; This method for learning model parameters is characterized in that, in the learning step, one of a regression model and a classification model is selected as the prediction model depending on the type of feel, and, as the learning algorithm, one of an algorithm in which a multivariate prediction technique is applied to the supervised machine learning algorithm and the supervised machine learning algorithm is selected depending on the type of feel.

2. 2. The method for learning model parameters according to claim 1, In the second acquisition method, a first selection step of executing learning of model parameters of a machine learning model by a supervised machine learning algorithm using the feature amounts, which are the blending ratios of the plurality of raw materials, and the teacher data as learning data, calculating the importance of the feature amounts using the machine learning model after the learning, and selecting the feature amounts in order of the highest importance; a second selection step of acquiring a correlation between the objective variable and the feature amount by applying a first filter process to the feature amount, which is the blending ratio of the plurality of raw materials, and selecting the feature amount in descending order of the highest correlation with the objective variable; a third selection step of acquiring contributions of the feature amounts to the objective variable by applying a second filter process different from the first filter process to the feature amounts, which are the blending ratios of the plurality of raw materials, and selecting the feature amounts in descending order of contributions; a fourth selection step of calculating an average value of the ranks of the feature quantities selected in the first to third selection steps, and selecting the feature quantities from the first rank to a predetermined rank when the feature quantity with the highest average value is ranked first as the first selection data; A method for learning model parameters, comprising:

3. 3. The method for learning model parameters according to claim 1, further comprising: In the third acquisition method, a fifth selection step of executing learning of model parameters of a machine learning model by a supervised machine learning algorithm using the blending ratios of the plurality of raw materials, the feature values ​​which are the physical property values, and the teacher data as learning data, calculating the importance of the feature values ​​using the machine learning model after the learning, and selecting the feature values ​​in order of the highest importance; a sixth selection step of acquiring a correlation between the objective variable and the feature amounts by applying a third filter process to the feature amounts, which are the blending ratios of the plurality of raw materials and the physical property values, and selecting the feature amounts in descending order of correlation with the objective variable; a seventh selection step of acquiring contributions of the feature amounts to the objective variable by applying a fourth filter process different from the third filter process to the feature amounts, which are the blending ratios of the plurality of raw materials and the physical property values, and selecting the feature amounts in descending order of contribution; an eighth selection step of calculating an average value of the ranks of the feature quantities selected in the fifth to seventh selection steps, and selecting, as the second selection data, the feature quantities from the first rank to a predetermined rank when the feature quantity with the highest average value is ranked first; A method for learning model parameters, comprising:

4. A method for learning model parameters, using a learning device, of a prediction model that predicts the texture of a preparation produced by blending a plurality of raw materials as a target variable, comprising: The learning device includes: A teacher data acquisition step of acquiring actual measured values ​​of the objective variables as teacher data; A feature acquisition step of acquiring feature values ​​including physical property values ​​of the formulation; a learning data creation step of creating learning data using the feature amount and the teacher data; a learning step of executing learning of model parameters of the prediction model using a learning algorithm to which a supervised machine learning algorithm is applied and the learning data; Run The preparation is an emulsion-based preparation, In the feature acquisition step, any one of a first acquisition method of acquiring only the physical property value of the formulation as the feature value, a second acquisition method of acquiring first selection data selected from the blending ratios of the plurality of ingredients and the manufacturing method of the formulation and the physical property value of the formulation as the feature value, and a third acquisition method of acquiring second selection data selected from the blending ratios of the plurality of ingredients, the physical property value of the formulation, and the manufacturing method of the formulation as the feature value is executed according to the type of sensation, This method for learning model parameters is characterized in that, in the learning step, one of a regression model and a classification model is selected as the prediction model depending on the type of feel, and, as the learning algorithm, one of an algorithm in which a multivariate prediction technique is applied to the supervised machine learning algorithm and the supervised machine learning algorithm is selected depending on the type of feel.

5. 5. The method for learning model parameters according to claim 4, In the second acquisition method, a first selection step of executing learning of a machine learning model by a supervised machine learning algorithm using the feature amounts, which are the blending ratios of the plurality of raw materials and the manufacturing method of the formulation, and the teacher data as learning data, calculating the importance of the feature amounts using the machine learning model after the learning, and selecting the feature amounts in order of the highest importance; a second selection step of applying a first filter process to the feature quantities, which are the blending ratios of the plurality of raw materials and the manufacturing method of the formulation, to select the feature quantities in descending order of correlation with the objective variable; a third selection step of acquiring contributions of the feature amounts to the objective variable by applying a second filter process different from the first filter process to the feature amounts, which are the blending ratios of the plurality of raw materials and the manufacturing method of the formulation, and selecting the feature amounts in descending order of contribution; a fourth selection step of calculating an average value of the ranks of the feature quantities selected in the first to third selection steps, and selecting the feature quantities from the first rank to a predetermined rank when the feature quantity with the highest average value is ranked first as the first selection data; A method for learning model parameters, comprising:

6. 6. The method for learning model parameters according to claim 4, further comprising: In the third acquisition method, a fifth selection step of executing learning of a machine learning model by a supervised machine learning algorithm using the feature amounts, which are the blending ratios of the plurality of raw materials, the physical property values, and the manufacturing method of the formulation, and the teacher data as learning data, calculating the importance of the feature amounts using the machine learning model after the learning, and selecting the feature amounts in order of the highest importance; a sixth selection step of acquiring a correlation between the objective variable and the feature quantities, which are the blending ratios of the plurality of raw materials, the physical property values, and the manufacturing method of the formulation, by applying a third filter to the feature quantities, and selecting the feature quantities in descending order of correlation with the objective variable; a seventh selection step of acquiring contributions of the feature amounts to the objective variable by applying a fourth filter different from the third filter to the feature amounts, which are the blending ratios of the plurality of raw materials, the physical property values, and the manufacturing method of the formulation, and selecting the feature amounts below the first rank in order of contribution when the feature amount with the highest contribution is ranked first; an eighth selection step of calculating an average value of the ranks of the feature quantities selected in the fifth to seventh selection steps, and selecting, as the second selection data, the feature quantities from the first rank to a predetermined rank when the feature quantity with the highest average value is ranked first; A method for learning model parameters, comprising:

Images