JP2020181423A - Cosmetic article recommendation discrimination method and system - Google Patents

Abstract

PURPOSE: To provide a cosmetic article recommendation discrimination method and a related system that reflect the beauty consciousness of a user, without assuming the discrimination of a skin type, by using machine learning, based on inquiry data of the user.SOLUTION: It is a method for determining whether or not a cosmetic article g should be recommended to a user v by executing a computer control program. The computer control program has a cosmetic article recommendation discrimination program using machine learning. User feature quantities of a large number of users prepared in advance and numerical values indicating whether or not these users use cosmetic articles are used as training data for the machine learning. It is a cosmetic article recommendation determination method and a related system for determining whether or not the cosmetic article g should be recommended to the user v, by executing the recommendation discrimination program, using a user feature quantity related to an item other than the cosmetic article g of the user v as a variable. The user feature quantity includes answers to cosmetic-related questions and numerical values indicating whether to use cosmetic articles.SELECTED DRAWING: Figure 9

Description

The present invention relates to a method and a system for determining whether or not a beauty product should be recommended to a user by using machine learning.

As a technique for determining a skin condition for evaluating a user's skin condition based on interview data, for example, a technique disclosed in Patent Document 1 (Patent No. 541004) by the present inventor is known. This technology collects basic data by interviewing a large number of subjects by age in advance, extracts sebum excess factors and skin dryness factors from the basic data by factor analysis, and interviews users of a specific age. This is a technique for classifying a user's skin type two-dimensionally based on the interview data. The results of skin type classification obtained by this technique can be used to select and provide appropriate beauty products, that is, cosmetics, beauty services, and the like to users.

In the process of implementing this patented invention, the present inventor has newly obtained three notices that are disclosed for the first time in the present specification.
The first thing to notice is that in order to select and provide the appropriate beauty products to each user, it is necessary to review the question set so that the user's beauty consciousness can be grasped in addition to the skin type. .. The above-mentioned two-dimensional classification of skin types based on sebum excess factor and skin dryness factor has a large correlation with basic cosmetics such as face washing, cleansing, lotion, milky lotion, and milk in the cosmetics classification. In addition to the basic cosmetics group, there are other types of cosmetics such as esthetic cosmetics such as beauty essences and creams, base makeup cosmetics such as foundations, and point makeup cosmetics such as lipsticks. In order to understand the customer's psychology in purchasing cosmetics of types other than these basic cosmetics, the above two-dimensional classification of skin types is not perfect. It is desirable to devise a question set so that in addition to the above-mentioned skin type of the user, the beauty consciousness of the user related to base makeup and aesthetics can be accurately grasped.

The second awareness is, of course, that the user's answers to the interview questions can be used to understand the user's beauty consciousness, but in addition to that, what is the beauty product that the user usually uses? The information is very useful.

The third thing to notice is that if you use a machine learning method using big data, you do not necessarily have to go through the steps of grasping the user's skin type and beauty consciousness, but directly from the user's answer to the question, a certain beauty product Is to be able to determine whether is appropriate for the user. At that time, in order to avoid that the basis of discrimination is confined in a black box and becomes unclear as is often seen in application cases of machine learning, the user can use or use the sigmoid collaborative filtering method described later. It is possible to determine whether or not a beauty product is suitable for a user while clarifying the skin type and beauty consciousness of the product as a latent feature amount.

Patent Document 2 (Japanese Patent Laid-Open No. 2012-523947) includes questions such as age, gender, sensation, skin condition and color gloss, sleep pattern, eating habits, energy level, stress level, physical health condition and emotional health. By answering the questionnaire, the technology of the computer system that classifies the user's skin composition based on the principle of traditional Chinese medicine is disclosed, and based on the determined skin composition, the user recommends a product that is effective for the skin. It is stated that it should be done. The document discloses an example in which a distinctive analysis is used in the classification process of the skin composition, and although there is no description of the example, a logistic regression, a support vector machine, a neural network, or the like can be used in the classification process. It is stated that. However, the document does not describe the importance of knowing the beauty consciousness of the user, nor does it describe the use of information about what the beauty product the user usually uses. In addition, the document does not mention a specific statistical processing method that associates skin composition with effective products, and classifies skin composition by machine learning methods to determine whether a product is appropriate for the user. It is not described that it is directly determined from the answer of the user's questionnaire without going through the step.

Patent Document 3 (Japanese Unexamined Patent Publication No. 2002-157469) provides a checklist regarding age, gender, living environment, skin moisture content, sebum content, skin troubles, skin sensitivity, and skin feel preference. Discloses the invention of a cosmetics purchasing system that has a means of automatically determining a user's skin quality. Here, the skin type of the user is classified into sensitive skin type, all skin type, mixed skin type, normal skin type, oily skin type, and dry skin type. In addition, the system stores cosmetic data including a cosmetic name associated with information about the user's skin, and this cosmetic data can be transmitted to the user terminal. However, the document does not mention the importance of knowing the beauty consciousness of the user, nor does it mention the use of information about what beauty products the user usually uses, skin type and cosmetics. There is no description about the specific statistical processing method to relate.

In Patent Document 4 (Japanese Patent No. 6404677), data processing is applied to interview data on skin and image data of keratin and / or sebum to obtain diagnostic data, and these diagnostic data are calculated to output skin diagnostic results. A system for determining the condition of the skin to be treated is disclosed. However, this technique is limited to determining the skin condition. There is no description in this document about a specific statistical processing method for associating the skin condition with the cosmetics suitable for the skin condition.

Patent Document 5 (Japanese Unexamined Patent Publication No. 2017-120595) discloses an invention of a method for improving the response accuracy when evaluating the application state of cosmetics by a questionnaire. In the present invention, the evaluation score is calculated based on the evaluation value which is the answer to the evaluation term (corresponding to the above question item), and the evaluation score is totaled for each classification of the generic evaluation term (corresponding to the above skin type). Then, the evaluation result including the aggregation result is output. However, the disclosure of the invention is limited to the evaluation method of the application state of cosmetics, and what kind of evaluation term or generic evaluation term should be adopted when the user wants to evaluate the skin type and beauty consciousness, and , It is completely unclear how the evaluation score should be set based on the evaluation terms in the above case. In addition, the document states that questions other than the application state of cosmetics, such as the type of cosmetics that are usually used, may be added to the questionnaire, but the user's answer to such questions is There is no mention of how the cosmetics are specifically used to evaluate the user's skin type or beauty consciousness, or to determine whether the cosmetics are suitable for the user.

Japanese Patent No. 541004 Special table 2012-523947 JP-A-2002-157469 Japanese Patent No. 6404677 JP-A-2017-120595

As described above, in the conventional technique of evaluating the user's skin type and beauty consciousness by a computer based on the interview data and determining whether the beauty product is suitable for the user, first of all, the user Not recognizing the importance of grasping not only the skin type but also the beauty consciousness, secondly, the beauty that users usually use to evaluate the skin type and beauty consciousness and to determine the appropriate beauty product. It is implicit that it does not utilize information about what the product is, and thirdly, it is necessary to evaluate the user's skin type and beauty consciousness in advance in order to determine whether the beauty product is suitable for the user. Fourth, it is unclear how to relate the user's skin type and beauty consciousness to the beauty product suitable for the user by using statistical processing by computer. was there.

In view of the above problems, an object of the present invention is to provide a beauty product to a user by a machine learning method using big data based on interview data including information on what beauty product the user usually uses. The present invention is to provide a recommended determination method and system for beauty products that can determine whether or not they are suitable. A further object of the present invention is to provide a recommended discrimination method and system for beauty products that can evaluate a user's beauty consciousness based on interview data and the like including the above information by the above machine learning method. is there.

The present invention has been made to solve the above problems, and the first aspect of the present invention is a method of determining whether or not the beauty product g should be recommended to the user v by executing a computer control program. , The computer control program has a recommendation determination program for beauty products using machine learning, and prepares in advance user feature quantities of a large number of users and numerical values indicating the degree of recommendation of beauty products to those users. Should the beauty product g be recommended to the user v by executing the recommended determination program using the training data for machine learning and the user feature amount related to the item other than the beauty product g of the user v as a variable? This is a recommended determination method for beauty products, which is characterized by determining whether or not it is present.

The second aspect of the present invention is a method for determining the recommendation of a beauty product, in which the numerical value indicating the degree of recommendation of the beauty product to the user is a numerical value indicating whether or not the user uses the beauty product.

A third aspect of the present invention is a recommended method for determining a beauty product, wherein the item includes a beauty product and a question about beauty.

In the fourth aspect of the present invention, the recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function, the predictive function is defined by a parameter, and the beauty product g of the user. This is a recommended determination method for beauty products, in which a user feature amount related to an item other than the above is used as an input value, and a predicted value of a numerical value indicating the degree of recommendation of the beauty product g for the user is used as an output value.

In the fifth embodiment of the present invention, the recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function, and the predictive function is a user latent feature amount p = p of the user u. (U) and item i's item latent feature amount q = q (i) is used as an input value, and the predicted value of user feature amount y (u, i) related to user u's item i is used as an output value. It is a discrimination method.

A sixth aspect of the present invention is a recommended discrimination method for a beauty product in which the recommended discrimination program has a user latent feature amount display means for displaying the user latent feature amount of the user v.

A seventh aspect of the present invention is a recommended discrimination method for a beauty product, wherein the recommended discrimination program has a discrimination accuracy display means for displaying a discrimination accuracy rate and / or a recall rate.

The eighth embodiment of the present invention comprises an independent computer or a server and a terminal capable of bidirectional communication with each other, and determines whether or not the beauty product g should be recommended to the user v by executing a computer control program. It is a product recommendation discrimination system, and the computer control program has a beauty product recommendation discrimination program using machine learning, and a large number of user feature quantities prepared in advance and recommendation of beauty products to those users. A training data storage means for storing a numerical value representing a degree as training data for machine learning, and a user feature amount related to an item other than the beauty product g of the user v as a variable. Beauty characterized by having a variable storage means and executing the recommended determination program to determine whether or not the beauty product g should be recommended to the user v based on the training data and the variables. This is a recommended product discrimination system.

In the ninth aspect of the present invention, the recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function, the predictive function is defined by a parameter, and the beauty product g of the user. This is a beauty product recommendation determination system in which a user feature amount related to an item other than the above is used as an input value, and a predicted value of a numerical value indicating the degree of recommendation of the beauty product g to the user is used as an output value.

In the tenth aspect of the present invention, the recommended discrimination program has a prediction function output value calculation means for calculating the output value of the prediction function, and the prediction function is a user latent feature amount p = p of the user u. (U) and item i's item latent feature amount q = q (i) is used as an input value, and the predicted value of user feature amount y (u, i) related to user u's item i is used as an output value. It is a discrimination system.

The eleventh aspect of the present invention is the computer control program.

A twelfth aspect of the present invention is a computer-readable storage medium characterized in that the computer control program and / or the parameters and / or the item latent features are recorded.

A thirteenth aspect of the present invention is an e-learning system in which a student can learn whether or not to provide a beauty product g to a customer, and is an independent computer or a server and a terminal capable of bidirectional communication with each other. It is composed of the training data storage means, is configured to be able to execute the recommended determination program, and is based on a user feature amount other than the beauty product g of a real or virtual user and the training data. It is characterized by having a discrimination result output means for displaying the discrimination result obtained by executing the recommended discrimination program, and / or a discrimination result comparison means for comparing the discrimination result with the discrimination result by the student himself / herself. It is a learning system.

According to the first aspect of the present invention, it is a method of determining whether or not the beauty product g should be recommended to the user v by executing a computer control program, and the computer control program is a method of determining the beauty product using machine learning. The user feature amount of a large number of users prepared in advance having a recommendation determination program and a numerical value indicating the degree of recommendation of beauty products to those users are used as training data for machine learning, and the user v said. The recommendation determination of a beauty product, which is characterized in that it is determined whether or not the beauty product g should be recommended to the user v by executing the recommendation determination program with the user feature amount related to the item other than the beauty product g as a variable. A method can be provided.

Here, the "beauty product" means a group consisting of at least one cosmetic product and / or beauty service. Of course, one cosmetic product and one cosmetological service can be called a beauty product, but a group of cosmetics and cosmetological services may be collectively called one cosmetological product. When a user uses a beauty product, it means that he / she buys or uses at least one cosmetic product in a group of cosmetics or beauty services, or receives at least one beauty service.

Further, "machine learning" is a function of the user feature amount x and the weight w, and the weight w in the function f (x, w) for use in the recommended determination of the beauty product is optimized based on the training data. A learning method to do. The training data is not limited, but it is desirable that the number of samples is preferably 1,000 or more, and more preferably 10,000 or more big data. The above-mentioned "machine learning" includes, but is not limited to, forward propagation neural networks, support vector machines, collaborative filtering, and the like.

The "user feature amount" refers to a quantified answer to a question, whether or not a beauty product is used, an image, a measured value or other information, or an aggregate amount calculated using a part or all of these. .. Here, the "aggregate amount" includes, but is not limited to, an average value, a weighted average value, a median value, each principal component in the principal component analysis, each factor score in the factor analysis, and the like. The term "item" refers to questions, beauty products, images, measured values, and other information. The "user feature amount related to the item" refers to the answer to the quantified question when the item is a question, and whether or not the quantified beauty product is used when the item is a beauty product. , When the item is an image, measured value or other information, it means a quantified image, measured value or other information.

Further, the "numerical value indicating the degree of recommendation of the beauty product to the user" may be any numerical value indicating how suitable the beauty product is to the user. In one embodiment of the present invention, the numerical value representing the degree of recommendation is a numerical value (for example, 1 or 0) corresponding to whether or not the user is using the beauty product. In another embodiment of the present invention, the numerical value indicating the degree of recommendation is an evaluation value (for example, 0, 1, 2, 3) evaluated by an expert familiar with cosmetology to what extent the beauty product is suitable for the user. Either).

According to the first aspect of the present invention, should the beauty product g be recommended to the user v based on the user feature amount by performing machine learning based on the training data, without going through the diagnosis result of the skin type? It is possible to provide a recommended discrimination method for beauty products, which can directly determine whether or not it is present.

According to the second aspect of the present invention, it is possible to provide a method for determining the recommendation of a beauty product, in which the numerical value indicating the degree of recommendation of the beauty product to the user is a numerical value indicating whether or not the user uses the beauty product. According to this embodiment, it is not necessary to ask a beauty expert to evaluate the user and the beauty product in order to obtain the numerical value indicating the above recommendation level, so that the user feature amount can be obtained without using the skin type diagnosis result. Based on this, it is possible to provide a method for determining the recommendation of a beauty product, which can directly determine whether or not the beauty product g should be recommended to the user v, simply and at low cost. In addition, this discrimination method is reliable and persuasive because it utilizes an objective user feature amount of whether or not the user uses a beauty product.

According to the third aspect of the present invention, the item can provide a recommended method for determining a cosmetology product, including a cosmetology product and a question about cosmetology. An invention that uses information on whether or not a beauty product g is used by a user v of a beauty product other than the beauty product g in order to determine whether or not the beauty product g should be recommended to the user v cannot be found in the prior art. According to this embodiment, the beauty consciousness of the user v is implicitly reflected more than the conventional technique by directly using the above-mentioned information on the presence or absence of use without going through the diagnosis result of the skin type. , It is possible to provide a recommended discrimination method for beauty products, which enables excellent recommended discrimination.

According to the fourth aspect of the present invention, the recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function, the predictive function is defined by a parameter, and the beauty of the user. It is possible to provide a recommended determination method for a beauty product using a user feature amount related to an item other than the product g as an input value and a predicted value of a numerical value indicating the degree of recommendation of the beauty product g for the user as an output value. The prediction function of this embodiment can be expressed preferably by a forward propagation neural network, and more preferably by a deep forward propagation neural network. In the present embodiment, the parameter corresponds to the weight w of the prediction function, and by optimizing the parameter based on the training data, it is possible to provide a recommended discrimination method for a beauty product with high discrimination accuracy.

According to the fifth embodiment of the present invention, the recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function, and the predictive function is a user latent feature amount p of the user u. = P (u) and the item latent feature amount q = q (i) of the item i are input values, and the predicted value of the user feature amount y (u, i) related to the item i of the user u is the output value. Can provide a recommended determination method for. The prediction function of this embodiment can be preferably constructed by a method of Sigmoid Collaborative Filtering (SCF) described later. In the present embodiment, the item latent feature amount corresponds to the weight w of the prediction function, and by optimizing this weight w based on the training data, it is possible to provide a recommended discrimination method for a beauty product with high discrimination accuracy. .. Further, in this embodiment, or by using this embodiment in combination with another embodiment of the present invention, the user's latent feature amount can be used to objectively identify the user's beauty consciousness, and the item latent feature amount can be used to objectively identify features such as beauty products and questions. Since it is possible to grasp the recommended judgment of the beauty product, the basis for the recommended judgment of the beauty product is clarified, and the judgment result is convincing.

According to the sixth aspect of the present invention, the recommended discrimination program can provide a recommended discrimination method for a beauty product having a user latent feature amount display means for displaying the user latent feature amount of the user v. In this embodiment, since the user's beauty consciousness can be objectively grasped by the user latent feature amount displayed by the user latent feature amount display means, the basis for the recommended discrimination of the beauty product is clarified, and the discrimination result. Is convincing.

According to the seventh aspect of the present invention, the recommended discrimination program can provide a recommended discrimination method for a beauty product having a discrimination accuracy display means for displaying a discrimination accuracy rate and / or a recall rate. In this embodiment, it is possible to quantitatively grasp the degree to which it is appropriate to recommend a certain beauty product to the user from the conformity rate and / or the recall rate of the discrimination displayed by the discrimination accuracy display means. ..

According to the eighth aspect of the present invention, it is composed of an independent computer or a server and a terminal capable of bidirectional communication with each other, and it is determined by executing a computer control program whether or not the beauty product g should be recommended to the user v. It is a recommended discrimination system for beauty products, and the computer control program has a recommended discrimination program for beauty products using machine learning, and a large number of user feature quantities prepared in advance and beauty products for those users. A numerical value indicating the degree of recommendation of the above, a training data storage means for storing as training data for machine learning, and a user feature amount related to an item other than the beauty product g of the user v are stored as variables. It is characterized in that it has a variable storage means for the purpose, and by executing the recommended determination program, it is determined whether or not the beauty product g should be recommended to the user v based on the training data and the variable. It is possible to provide a recommended discrimination system for beauty products.

In this embodiment, the training data storage means is an internal storage means such as a memory such as RAM or ROM in a computer, a server, or a terminal, or a hard disk drive (HDD) or a solid state drive (SSD), a USB memory, or an SD card. It is composed of a removable storage device such as, or an external storage means such as CD-R or DVD. The terminal includes any fixed terminal such as a desktop computer or a mobile terminal such as a smartphone or a laptop computer. The server and the terminal are configured to be able to communicate with each other in both directions via a communication network such as a LAN or the Internet. The variable storage means is composed of, for example, an area reserved for the internal storage means in a computer, a server, or a terminal, or a register of a CPU. According to this embodiment, by performing machine learning based on the training data, it is directly determined whether or not the beauty product g should be recommended to the user v based on the user feature amount, without going through the diagnosis result of the skin type. It is possible to provide a recommended cosmetological product discrimination system.

According to the ninth aspect of the present invention, the recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function, the predictive function is defined by a parameter, and the beauty of the user. It is possible to provide a beauty product recommendation determination system in which a user feature amount related to an item other than the product g is used as an input value and a predicted value of a numerical value indicating the degree of recommendation of the beauty product g to the user is used as an output value. In the present embodiment, the parameter corresponds to the weight w of the prediction function, and by optimizing the parameter based on the training data, it is possible to provide a recommended discrimination system for beauty products with high discrimination accuracy.

According to the tenth aspect of the present invention, the recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function, and the predictive function is a user latent feature amount p of the user u. = P (u) and the item latent feature amount q = q (i) of the item i are input values, and the predicted value of the user feature amount y (u, i) related to the item i of the user u is the output value. Recommended discrimination system can be provided. In the present embodiment, the item latent feature amount corresponds to the weight w of the prediction function, and by optimizing this weight w based on the training data, it is possible to provide a recommended discrimination system for beauty products with high discrimination accuracy. Further, in this embodiment, the user's latent feature amount can be used to objectively grasp the user's beauty consciousness, and the item latent feature amount can be used to objectively grasp the characteristics of beauty products and questions. The grounds for this are clarified, and the judgment result is convincing.

According to the eleventh aspect of the present invention, it is possible to provide a program characterized by being the computer control program.

According to a twelfth aspect of the present invention, it is possible to provide a computer-readable storage medium characterized by recording the computer control program and / or the parameter and / or the item latent feature amount.

According to the thirteenth aspect of the present invention, it is an e-learning system capable of learning whether or not to provide beauty products g to customers, from an independent computer or a server and a terminal capable of bidirectional communication with each other. The training data storage means is configured so that the recommended discrimination program can be executed, and the training data is based on the user feature amount other than the beauty product g of the real or virtual user and the training data. E-learning characterized by having a discrimination result output means for displaying the discrimination result obtained by executing the recommended discrimination program, and / or a discrimination result comparison means for comparing the discrimination result with the discrimination result by the student himself / herself. Can provide a system.

It is a table diagram explaining the content of the question about cosmetology included in the questionnaire used in one Embodiment of this invention. It is a table diagram explaining the classification of cosmetics which are beauty products in one Embodiment of this invention. It is explanatory drawing which shows the division of the data set used in one Embodiment of this invention. It is a table diagram explaining the characteristic of the statistical distribution of each data set obtained by dividing the data set. It is explanatory drawing which shows the general structure of the forward propagation type neural network used in one Embodiment of this invention. It is explanatory drawing which shows three kinds of concrete configurations of the forward propagation type neural network in embodiment of this invention, and the configuration of a network used for an SCF model. It is explanatory drawing which shows the general relationship between the threshold value used for recommended discrimination and discrimination accuracy. It is a graph which shows the relationship between the utilization rate of the beauty product and the f1 score in one Embodiment of this invention, and is the explanatory diagram explaining the concept of the modified f1 score. It is a table figure which shows the result of the recommended discrimination of the beauty product for the user and the conformity rate of the discrimination in one Embodiment of this invention. It is a flow chart which shows the execution flow of the computer control program in one Embodiment of this invention. It is a chart which shows eight components of the item latent feature amount of each item in one Embodiment which concerns on SCF model of this invention. It is explanatory drawing which shows how the user latent feature amount is displayed on the output device by the user latent feature amount display means in embodiment which concerns on the SCF model of this invention. It is a flow chart which shows the execution flow of the computer control program in one Embodiment which concerns on the SCF model of this invention. It is a block diagram which shows the system structure in one Embodiment of this invention. It is a block diagram explaining a part of the system structure in detail. It is a block diagram which shows the system structure in another embodiment of this invention. It is a block diagram which shows the system structure in still another Embodiment of this invention. In one embodiment of the invention, there are two types of discriminant surfaces for plotting the score of each factor by factor analysis by the user.

The recommended determination method and system embodiment of the beauty product according to the present invention will be described in detail below with reference to the drawings.

<1. Data set ＞
(Dataset Data Structures) Table 1 shows the dataset data structures used in some embodiments of the present invention. We interviewed 20000 users and obtained valid responses for 19468 people. Each user's answer includes a 2-choice (0,1) or 5-choice (-2,-1,0,1,2) answer to the beauty question (Q1 to Q21) and an age question (Q22). ), An integer value indicating the age, and a two-choice (0,1) answer indicating whether or not the beauty products (G1 to G12) are used are included. The age of the user ranges from 20 to 69 years. All genders are female, but in another embodiment of the invention, genders may span men and women, in which case a two-choice (0,1) answer to the gender question is added to the dataset. You may.

(Contents of Questions) FIG. 1 is a table diagram for explaining the contents of questions about cosmetology included in the questionnaire related to the data set. The questions (Q1 to Q13) are two-choice questions, and the questions (Q14 to Q21) are five-choice questions. The former is a question mainly related to the user's skin type, and the latter is a question related to the user's beauty consciousness. In addition, z may be a natural number of 2 or more, and these questions may be questions of z selection, or a slide bar may be placed on the line segment displayed on the screen with a mouse. It may be a question to answer the real value included in the closed section such as 0 or more and 1 or less. Question Q22 is a question that asks the age of the user as an integer value.

(Selection of Question Items) In Japanese Patent No. 54190004 (Patent Document 1) by the present inventor, sebum excess factor and skin dryness factor are determined by factor analysis method from interview data by age group regarding beauty and skin condition of women. Extracted. These two factors can be explained by an inversely correlated "one-dimensional model" in the younger age group and an almost uncorrelated "two-dimensional model" in the middle-aged and older age groups. As shown in FIG. (18A), the subjects are distributed on a two-dimensional plane having an X-axis related to excess sebum and a Y-axis related to dry skin, and dry skin, normal skin, and oily skin (or excess sebum). It is classified into four quadrants: skin) and oily dry skin (or hidden dry skin). Assuming that the axes obtained by rotating the X-axis and Y-axis around the origin by 45 ° counterclockwise are the A-axis and the B-axis in order, the B-axis is a one-dimensional model for the younger age group, and the A-axis is the middle-aged and older group. Corresponds to the two-dimensional model of, and the A-axis coordinates indicate skin dissatisfaction related to sebum and dry skin.
As a result of reanalyzing the data accumulated by carrying out the above patented invention, in addition to the sebum excess factor and the skin dryness factor, an independent third factor (No. 3) was found. It was found that this factor is highly related to the question items related to makeup dissatisfaction (skin quality dissatisfaction) such as "makeup collapse", "bad makeup" and "rough texture". Therefore, by finding the fourth factor that is paired (= orthogonal) with this factor and obtaining a new discriminating surface, the possibility of a new beauty determination method that includes the findings of the above patented invention is considered and examined. Stacked. Specifically, the question item set (13 items) constructed in Patent Document 1 is combined with a new set of question items, and the third factor and the fourth factor to be paired with the skin dryness and oiliness are set. The extraction was repeated. As a result, the above purpose was achieved by adding eight new question items.
That is, 5 factors were extracted by the factor analysis method of the answers of the subjects to the questions of 13 + 8 = 21 items. The five factors are called EST factor, O1 factor, O2 factor, MB factor, and D factor in order. The EST factor is the most influential factor in cosmetology interviews (with the largest ∵ eigenvalue) and is the fourth factor of interest. The O1 factor is a factor (sebum excess factor) substantially corresponding to the X-axis in the above patented invention (∵ factor score correlation coefficient R = 0.968). The O2 factor is also a factor close to the X-axis in the above patented invention, and is presumed to be involved in the skin satisfaction caused by oiliness. The MB factor is the above-mentioned No3 factor. The EST factor is the fourth factor that was the purpose of the above search. Factor D is almost the Y-axis factor (skin dryness factor) of the above-mentioned patented invention. The MB factor is called a make-up base factor (MB factor) because it is strongly related to the question items related to base make-up (makeup collapse, bad make-up, and rough texture) when summarized in the cosmetics category. The MB factor can be interpreted as a factor that extracts and manifests latent dissatisfaction with base makeup, and is closely related to the products and services (foundation, makeup base, etc.) that prepare the foundation for aesthetics of the skin. The EST factor is called the aesthetic factor (EST factor) because it is strongly related to the question items related to aesthetics (tension elasticity, brightness and transparency, skin texture, youthfulness) when summarized in the cosmetics category. To do. The EST factor can be interpreted as a factor that extracts and manifests latent aesthetic needs, and is closely related to products and services (beauty essence, cream, etc.) that improve skin beauty. Since the MB factor and the EST factor are orthogonal to each other, the score of each factor is obtained from the results of the questionnaire, and for example, as shown in FIG. (18B), the two-dimensional plane (discrimination plane) developed on the EST axis and the MB axis. The factor score of the person concerned can be plotted and expressed above. From the plot on the above identification surface, the user's visit / stop-by needs, specifically, (1) strength of needs for beauty products (cosmetics and beauty services), (2) center of gravity (where of interest), (3) ) It is possible to quickly classify the psychological tolerance for purchase / enjoyment and realize effective customer service. Therefore, the above set of 13 + 8 = 21 questions (Q1 to Q21) includes the knowledge of the above patented invention, minimizes the burden on the customer, and provides prompt, accurate, and comprehensive beauty counseling. A set of questions that can be done, increase sales opportunities, and promote royal customerization.

(Classification of Cosmetological Products) FIG. 2 is a table diagram illustrating an example of classification of cosmetics as beauty products in one embodiment of the present invention. The rightmost column shows the utilization rate of each cosmetic product among the users of the data set. Focus on 12 cosmetics with a utilization rate of 7% or more and 70% or less, and each of them is represented by a beauty product symbol (G1 to G12). In the present specification, the discrimination accuracy of the recommended discrimination method for the beauty product according to the present invention is verified by taking the above 12 beauty products as an example, but the present invention is of course applicable to other beauty products. It is also applicable to the classified beauty products when classified by other classification methods.

(Dataset division) In machine learning applications, the entire data set is often divided into three parts: training data, cross-validation data, and test data. FIG. 3 shows an example of such a trisection of the dataset according to the embodiment of the present invention. The number of samples in the divided data set is 11868, 3800, and 3800, respectively. Samples in the entire dataset were randomly sorted and then divided into three datasets by two dividers. Generally, training data is used for determining parameters of prediction functions, cross-validation data is used for determining hyperparameters such as regularization parameters, and test data is used for verification of discrimination accuracy and prediction accuracy.

(Statistics of each data set) FIG. 4 shows statistics for each of the above divided data sets and the entire data set. The numbers shown in each column of the questions (Q1 to Q21) and the beauty products (G1 to G12) indicate the frequency of answers as a percentage. For example, in the training data, 16.7% of users answered 1 (YES) to the two-choice question Q7. In addition, in the cross-validation data, 7.2% of users answered that the beauty product G4 is used.

<2. Forward Propagation Neural Network>
<2-1. General forward propagation neural network configuration>
FIG. 5 is an explanatory diagram showing a general configuration of a forward propagation neural network representing a prediction function used in some embodiments of the present invention. In these embodiments, the prediction function f takes the user feature amount x = x (u) of any user u as an input value, and the value y = y (u) relating to the item (question, beauty product, etc.) of the user u. The predicted value a of is used as the output value, and it depends on the parameter Θ. The parameter Θ is a collective representation of the weights of all neurons present in the neural network. Written in the formula (1) a = f (x, Θ)
Is.

(Input layer) The first layer (input layer) consists of n ₁ neurons, and each neuron is composed of each component of the user feature amount x = x (u) of the user u, and each neuron in the second layer (excluding bias neurons). ) Is output. It is desirable to standardize the user features before performing machine learning. The standardization is, for example, normalizing the average value of each component of the user feature amount to zero and the standard deviation to one.

(Second layer) The second layer (one of the middle layers) consists of n ₂ neurons, and the k-th neuron in the second layer has the output value x of the neuron in the input layer and its own weight θ ⁽¹⁾ _k. The inner product of and is calculated, and the value of the inner product is z, and the value a ⁽²⁾ _k = g (z) obtained by applying the activation function g to z is set for each neuron in the third layer (excluding bias neurons). ) Is output. When written in the formula, (2) a ⁽²⁾ _k = g (z), z = x · θ ⁽¹⁾ _k
Is.
(Intermediate layer and the output layer) j-th layer (j = 3, ··, L ) consists n _j neurons, k-th neuron in the j-th layer is the output value of the neuron of the j-1 Layer a ⁽ Calculate the inner product of ^j-1) and its own weight θ ^(j-1) _k, and add its own bias values θ ^(j-1) _{k, 0} to z, and let z be the activation function g. The value a ^(j) _k = g (z) obtained by applying the above is output toward each neuron (excluding the bias neuron) in the j + 1 layer. When written in the formula, (3) a ^(j) _k = g (z), z = a ^(j-1) · θ ^(j-1) _k + θ ^(j-1) _{k, 0}
Is.
(Output layer) The Lth layer (output layer) is composed of n _L neurons, and outputs an output value a = a ^(L) when j = L in the above equation (3). The equation (3) in this case is a concrete expression of the contents of the equation (1). The output value a is a predicted value of the value related to the item of the user u, and is used for the recommended determination of the beauty product. For example, when the k-th component a _{k of} the output value a is a predicted value of a value related to a beauty product, if a _{k is} approximately 1 for the user u, the beauty product should be recommended to the user u. It is determined that if a _{k is} almost 0, it should not be recommended.

(Activation function) In this embodiment, a sigmoid function is used as the activation function g of each neuron. However, it is also possible to use a hyperbolic tangent function or a rectified linear function as the activation function, and an activation function different depending on the neuron may be used. However, when the presence or absence of the use of beauty products is expressed by a numerical value of 0 or 1, the activation function is used for neurons in the output layer that output predicted values of values related to beauty product items. It is desirable to use the sigmoid function as. It is also possible to include neurons that do not use a bias value. Furthermore, it is possible to configure some neurons belonging to the neural network as pass-through neurons. A pass-through neuron is an output value of any neuron belonging to any layer behind the layer to which the neuron belongs, as its own output value, and any or a part or all of the layer one before. A neuron that outputs to a neuron.

<2-2. Cost function>
(Cost function) The parameter Θ representing the weight of each neuron is optimized by machine learning based on training data. The optimization is performed by sequentially updating the parameter Θ with the aim of minimizing the cost function J represented by the following equation (4).

In equation (4), the sum for user u covers all m = 11868 training data. The sum for the item k covers all of the items (question or beauty product) k that output the predicted value a _k (u) of the value y _k (u) related to the item k in the output layer. Here, the predicted value a (u) = f (x (u), Θ) is a function of the user feature amount x (u) and the parameter Θ. The neurons in the input layer output the user feature amount x (u), but the user feature amount related to the item that outputs the predicted value in the output layer is excluded from this x (u). The value y _k (u) related to the item k in the output layer has a maximum value of 1 and a minimum value of 0 in advance before machine learning.
It is desirable to standardize so that For example, the answer values of 5 choices -2, -1, 0, 1, 2 become 0, 0.25 0.50, 0.75, 1.0 in order by standardization.

(Regularization parameter) λ is a regularization parameter, and the square of the norm of Θ || Θ || ² is the sum of the squares of the weights (excluding the bias value) of all neurons existing in the neural network. is there. In machine learning, the regularization parameter λ is fixed, and the parameter Θ is repeatedly updated with the aim of minimizing the cost function J for the training data. As a result, a optimized Θ = Θ (λ) is obtained. This is done for some preselected λ values (eg 0.1, 0.032, 0.001, 0.00032, 0.0001, ...). Next, for the cross-validation data, λ is selected so as to maximize the f1 score or the modified f1 score described later when the parameter Θ (λ) is used.

(Initial value of parameter Θ) The initial value of parameter Θ was given by a random number. For example, the weight θ ^(j-1) _k of the k-th neuron in the ^j- th layer is n _j-1 components θ ^(j-1) _{k, r} (r = 1, 2, ..., n _{j- Although} it has ₁ ), each component is given by a uniform random number whose value range is the interval [−ε, ε] (where ε = 0.5 × n _j-1 ^{-1 / 2} ). In addition, we mainly used a forward-propagation neural network with three intermediate layers, but when using a forward-propagation neural network with a larger number of layers, we perform pre-learning with an autoencoder. , The value obtained by pre-learning may be used as the initial value of the parameter Θ.
It is also possible to use a forward propagation neural network having one or two intermediate layers. In that case, in order to favorably optimize the parameter Θ, the number of neurons in the second layer (first intermediate layer) is preferably increased, for example, the number of neurons in the first layer (input layer). The initial value of the parameter Θ is given by a random number by the above method as twice or more of.

(Update of parameter Θ) In this embodiment, the parameter Θ is updated by batch learning, and the ADADELTA method (see M.D.Zeiler, arXiv: 1212.5701v1) that can automatically adjust the learning rate is used. However, it is also possible to update the parameter Θ by mini-batch learning or online learning. In the case of mini-batch learning or online learning, when additional data is added to the training data, it is possible to train only the additional data and update the parameter Θ. Further, the parameter Θ may be updated not only by the ADADELTA method but also by a method such as a stochastic gradient descent method, a conjugate gradient method, a momentum method, an ADAGRAD method, or a combination of two or more of them. The calculation of the gradient ∂J / ∂Θ can be performed efficiently by the error back propagation method. Further, in the present embodiment, when it is determined that the cost function J has reached the minimum value (stationary point), the update of the parameter Θ is completed, but even if the cost function does not stop, the cross-validation data is used. When it is confirmed that the cost function calculated by the above or the f1 score or the modified f1 score described later reaches the stationary value and the sign of the difference is inverted, it is judged that overfitting has started, and the update of the parameter Θ is stopped. May be good. In the following three examples of this embodiment, the number of iterations of batch learning of training data was 2000 to 4000 times. For the calculations in the examples of this specification, GNU Octave, version 4.4.1 is used as the program language, and a commercially available desktop computer (Dell Vostro 230, Pentium (R) Dual-Core CPU E5800 @ 3.20 GHz) is used. I went.

As an example of a special case of the general embodiment of the present invention shown in FIG. 5, three examples (Deep model, DeepPa model, and DeepGo model) will be described in order.
<2-3. Deep model>
FIG. (6A) is an explanatory diagram showing a specific configuration of a forward propagation neural network representing a prediction function used in an embodiment of the present invention called a Deep model. In this embodiment, the user feature amount used for the input layer is only the answer to the question (Q1 to Q22) and does not include the presence or absence of the use of beauty products (G1 to G12), and the predicted value of the output layer is one. It is characterized in that it is only a predicted value of whether or not a beauty product (G1 in the example of the figure) is used. In this embodiment, a neural network having 3 intermediate layers and 5 layers as a whole is used, and the number of neurons is 22 for the input layer, 16 for the 2nd layer, 8 for the 3rd layer, and 8 for the 4th layer. 4. The output layer is 1. The preferred value of the regularization parameter λ was λ = 1.0 × 10 ^-4 .

<2-4. DeepPa model>
FIG. (6B) is an explanatory diagram showing a specific configuration of a forward propagation neural network representing a prediction function used in an embodiment of the present invention called a DeepPa model (Pa means parallel). In this embodiment, the user feature amount used for the input layer is only the answer to the question (Q1 to Q22) and does not include the presence or absence of the beauty products (G1 to G12), and the predicted value of the output layer is 12 beauty products. It is a predicted value of the use or non-use of products (G1 to 12), and is characterized in that the use or non-use of all beauty products is predicted at the same time. In this embodiment, a neural network having 3 intermediate layers and 5 layers as a whole is used, and the number of neurons is 22 for the input layer, 24 for the second layer, 12 for the third layer, and 12 for the fourth layer. 12, the output layer is 12. The preferred value of the regularization parameter λ was λ = 1.0 × 10 ^-5 .

<2-5. DeepGo model>
FIG. (6C) is an explanatory diagram showing a specific configuration of a forward propagation neural network representing a prediction function used in an embodiment of the present invention called a DeepGo model (Go means goods). In this embodiment, the user feature amount used for the input layer includes both the answer to the question (Q1 to Q22) and the presence or absence of the use of beauty products (G2 to G12 in the example of the figure), and the predicted value of the output layer is 1. It is characterized in that it is a predicted value of whether or not one beauty product (G1 in the example of the figure) is used. It should be noted that the presence or absence of the beauty product (G1 in the example of the figure) related to the predicted value of the output layer is excluded from the user feature amount used for the input layer. In this embodiment, a neural network having 3 intermediate layers and 5 layers as a whole is used, and the number of neurons is 33 for the input layer, 24 for the second layer, 12 for the third layer, and 12 for the fourth layer. 6. The output layer is 1. The preferred value of the regularization parameter λ was λ = 1.0 × 10 ^-4 .

<3. Discrimination accuracy>
<3-1. Threshold and discrimination accuracy>
(Threshold value) In the present invention, the threshold value h is used to determine whether or not the beauty product g should be recommended to the user u. For example, when the sigmoid function is used as the activation function of the output layer, the threshold value h is a real number having a value of 0 or more and 1 or less, and is generally selected as a different value for each beauty product g. When the user feature amount of the user u related to the item other than the beauty product g is input to the input layer of the neural network, the output value (predicted value) of the neuron related to the item which is the beauty product g of the output layer is set to a. .. If a ≧ h, it is determined that the beauty product g should be recommended to the user u, and if a <h, it is determined that the beauty product g should not be recommended.

(Compliance rate, recall rate, f1 score) The precision rate, recall rate, and f1 score, which are concepts representing discrimination accuracy, will be described. The discrimination value based on the above predicted value a is set to 1 when it is determined that it should be recommended, and 0 when it is not. Further, the observed value is set to 1 when the user u is using the beauty product g, and 0 when the user u is not using the beauty product g. All users contained in a certain data set, for example, test data, are classified into one of the four cases shown in Table 2, True Positive, False Positive, False Negative, and True Negative. Let TP, FP, FN, and TN be the percentages of users who fit each case in the dataset. At this time, the precision rate and the recall rate are defined by the following equations, respectively.
Conformity rate = TP / (TP + FP)
Recall rate = TP / (TP + FN)
The f1 score is a harmonic mean of the precision rate and the recall rate, and is defined by the following equation.
f1 score = ((Compliance rate ^-1 + recall rate ^-1 ) / 2) ^-1

FIG. 7 is an explanatory diagram showing a general relationship between the threshold value used for the recommended discrimination and the discrimination accuracy. There is a trade-off between recall and precision. Generally, as the threshold h is increased, the precision rate increases but the recall rate decreases. In FIG. 7, the recall rate is larger than the precision rate for the threshold value h1, but the recall rate is smaller than the precision rate for the threshold value h2. What value the threshold should be set to depends on the nature and purpose of the recommended determination. If you want to reduce the mistake of determining that a beauty product is not actually used even though it is determined to be recommended, you should emphasize the precision rate, and even though you did not determine that it should be recommended, in reality If you want to reduce the omission of discrimination that beauty products are used, you should emphasize the recall rate. In the example of FIG. 7, the minimum value of the conformance rate is 0.2, which is equal to the utilization rate of the beauty product g in the data set. In the embodiment of the present invention, the proportion of users whose discrimination value based on the predicted value a in the cross-validation data or training data is 1, in other words, the proportion of users whose predicted value a is equal to or higher than the threshold value h is beauty. The threshold value h was set so as to be equal to the utilization rate of the product g.

<3-2. f1 score and modified f1 score>
FIG. (8A) is a graph showing the relationship between the utilization rate of beauty products and the f1 score in the user included in the test data of the DeepGo model according to the embodiment of the present invention. The 12 plotted dots represent each of the G1 to G12 beauty products. It can be seen that the f1 score tends to increase as the usage rate of beauty products increases. In order to measure the discrimination ability of the model, it is desirable to use the discrimination accuracy that corrects the difference due to the utilization rate of beauty products. Therefore, consider the amount called "corrected f1 score". FIG. (8B) is an explanatory diagram illustrating the concept of the modified f1 score. Roughly speaking, the modified f1 score is a quantity indicating how much the f1 score would be if the utilization rate of beauty products was 0.20. 0.20 is an approximate average value of the utilization rates of the 12 types of beauty products G1 to G12 taken up in this example. We use the sigmoid function σ and its inverse function σ ^-1 shown in the curve of the graph in Fig. (8B) for the above correction. Generally, in the discrimination by a certain model, when the f1 score of the beauty product with the utilization rate r is s, the modified f1 score is the modified f1 score = σ (x + Δx).
Defined in. Here, σ (x) = 0.20, Δx = σ ^-1 (s) −σ ^-1 (r). For example, in the discrimination by a certain model, it is assumed that the f1 score of the beauty product having a utilization rate of 0.90 is 0.99 (point A and point A'in the figure). Since 0.90 = σ (2.2) and 0.99 = σ (4.2), the difference between the x-coordinates (horizontal axis coordinates) of the point A and the point A'is Δx = 2.0. Here, the coordinates of the point S'on the graph advanced by Δx to the right from the point S (? 1.4, 0.20) where σ (x) = 0.20 are (0.6, 0.65). Therefore, the modified f1 score of the beauty product in the discrimination by the model is defined as 0.65. Similarly, when the f1 score of a cosmetological product with a utilization rate of 0.05 is 0.29 (point B and point B'in the figure), the modified f1 score is 0.65.

<3-3. Discrimination accuracy of each model>
Table 3 is a table showing the f1 score and the modified f1 score, which are the discrimination accuracy of the discrimination by the Deep model, the DeepPa model, the DeepGo model, and the SCF model described later, which are examples of the present invention, for 12 types of beauty products. .. The discrimination accuracy of the Deep model and the DeepPa model is almost the same on average, but the latter has higher discrimination accuracy than the former depending on the beauty product. It is considered that this is because the information on whether or not a beauty product other than the beauty product related to the discrimination is used is reflected by error back propagation in the process of optimizing the weight of the neural network. Therefore, in order to positively use the information on whether or not the beauty product is used for the recommendation determination, the DeepGo model uses the information on whether or not the beauty product is used in the input layer. As expected, the DeepGo models had the highest f1 and modified f1 scores of these models. The discrimination accuracy of the discrimination by the DeepGo model is about 0.17 measured by the f1 score and about 0. 17 also improved.

<3-4. Discrimination accuracy display means>
FIG. 9 is a table showing the result of recommended discrimination of a beauty product for a certain user and the conformity rate of the discrimination by the DeepGo model according to the embodiment of the present invention. Table 4 shows the user feature amount before standardization (however, the information on whether or not the beauty product is used for discrimination is not input to the input layer) of the user input to the neurons of the input layer of the neural network. This user sample belongs to the test data (data set 3). The recommended discrimination program compares the predicted value a by the forward propagation neural network with the threshold value to determine whether or not the beauty product should be recommended to the user, and the discrimination result is determined by the discrimination result output means of the display, printer, etc. It is output to the output device. The discrimination result output means preferably has a discrimination accuracy display means. The discrimination accuracy display means displays on the output device whether or not each beauty product is used by the user, the discrimination accuracy rate, and the discrimination recall rate. The precision of discrimination is preferably displayed in a visually easy-to-understand form such as a bar graph in addition to the numerical value. Further, it is preferable that the displayed beauty products are sorted according to the magnitude relation of the matching rate of discrimination.

The recall rate is calculated by determining the threshold value by a method determined based on the utilization rate. In principle, the precision rate is cross-validated by considering ε1, ε2 as a small positive number and considering the interval [a-ε1, a + ε2] on the number line including the predicted value a of the user concerned with the beauty product. It can be obtained by calculating the ratio (adaptation rate) of the samples whose actual observed value is 1 for the samples whose predicted value a belongs to the above interval in one data set such as data for use. However, in order to speed up the calculation, for example, ε1 = ε2 = 0.05 is set in advance, and the predicted values are 0.0, 0.1, 0.2, ..., 0.9, 1.0, etc. The matching rate in the above data set when set to a discrete value may be obtained, and the matching rate may be calculated by interpolation using a linear or polynomial or spline curve according to the value of the predicted value a.
Similarly, in the SCF model described later, which is another embodiment of the present invention, the discrimination result output means and / or the discrimination accuracy display means is used in the output device to determine the result and / or discrimination of the recommended discrimination of the beauty product. Outputs the precision rate and recall rate.

In the example of FIG. 9, the output of the discrimination result output means and the discrimination accuracy display means is the result of recommended discrimination for the beauty products G5 (milky lotion / milk) and G7 (beauty essence (whitening)) not used by the user. It is 1 (recommended), and the accuracy rate of these discriminations is also quite high, 55% and 40%, respectively. Therefore, it is considered appropriate to first recommend these beauty products to the user.

<3-5. Flow diagram>
FIG. 10 is a flow chart showing a processing flow of a computer control program in one embodiment of the present invention using a forward propagation neural network such as a DeepGo model. First, in step S11, it is determined whether or not the parameter Θ can be read from a storage means such as a memory, HD, or removable memory, or a server connected via a network. If it can be read, the parameter Θ is read in step S14, and the process proceeds to step S15. If it is not readable, the process proceeds to step S12. In step S12, a suitable value of the parameter Θ is estimated by machine learning. Next, in step S13, the estimated parameter Θ is stored in the above storage means and / or transmitted to a server, a terminal, or the like connected via the network. Next, in step S15, the user feature amount of the user is input to the input layer of the forward propagation neural network, and the predicted value is obtained from the output layer. In step S16, the recommended determination of the beauty product is performed based on the predicted value and the predetermined threshold value. Next, in step S17, the precision and recall are calculated based on the predicted values. Finally, in step S18, the discrimination result output means and the discrimination accuracy display means output the discrimination result and / or the matching rate and / or the recall rate to the output device.

<4. SCF model>
(Structure of SCF model) FIG. (6D) is an explanatory diagram showing a configuration of a neural network that expresses a prediction function used in one embodiment of the present invention called an SCF model. In the present embodiment, the prediction function f is a user latent feature amount p = p (u) of any user u and an item latent feature amount q = q (k) of any item (question, beauty product, etc.) k. , Is used as the input value, and the predicted value a = a (p, q) of the value y = y (u, k) related to the item k of the user u is used as the output value. Both the user latent feature amount p and the item latent feature amount q are vectors belonging to the b-dimensional number vector space, and are estimated based on the training data. Here, the natural number b is a hyperparameter of this model, and is set to a suitable value by using the cross-validation data together with the regularization parameter λ described later. For the data set used this time, b = 8.

(Input layer) The input layer consists of (b + b) neurons, and each neuron is each component of the user latent feature p = p (u) of the user u, or the item latent feature q = q (of item k). Each component of k) is output toward the neurons in the output layer.

(Output layer) The output layer consists of one neuron, which calculates the standard inner product p · q of the user latent feature amount p and the item latent feature amount q, and is active in z with the value of the inner product as z. The value g (z) obtained by applying the conversion function g is output as the predicted value a = a (p, q) of the value y (u, k) related to the item k of the user u. When written in the formula, (5) a = g (z), z = p · q = p (u) · q (k)
Is. The output value a is a predicted value of the value related to the item k of the user u, and is used for the recommended determination of the beauty product. For example, when the item k is a beauty product, if a is close to 1, or more accurately, a predetermined threshold value is set to h, and if a ≧ h, the beauty product should be recommended to the user u. It is determined that if a is close to 0, or more accurately a <h, it should not be recommended.

(Activation function) In the present embodiment, it is desirable to use the sigmoid function as the activation function g of the neurons in the output layer. It is also possible to use a hyperbolic tangent function or a rectified linear function as the activation function, but in that case, it is necessary to convert the user features according to the range of the activation function in advance before machine learning. is there. Further, although the bias value is not used in the present embodiment, it is also possible to introduce the bias value as a parameter and use the value obtained by adding the bias value to the inner product as the z.

(Cost function) The user latent feature amount p (u) of each user u and the item latent feature amount q (k) of each item k are optimized by machine learning based on training data. The optimization aims at minimizing the cost function J represented by the following equation (6), that is, the latent feature amount, that is, the user latent feature amount p (u) of each user u and the item of each item k. This is performed by sequentially updating the latent feature amount q (k).

In equation (6), the sum for user u covers all m = 11868 users of the training data. The sum for item k spans all n = 34 items (questions or beauty products) in the training data. Here, it is desirable that the value y (u, k) related to the item k of the user u is standardized in advance so that the maximum value is 1 and the minimum value is 0 before machine learning.

(Regularization parameter) λ is a regularization parameter, which is a standard square norm of p (u) and q (k), which are vectors in a b-dimensional number vector space | p (u) | and | q (k). It works to impose a limit on the size of |. In machine learning, the regularization parameter λ is fixed, and the latent features p (u) and q (k) are repeatedly updated with the aim of minimizing the cost function J for the training data. As a result, the optimized item latent feature amount q (k) = q (k, λ) is obtained. This is done for some preselected λ values (eg 0.1, 0.032, 0.001, 0.00032, 0.0001, ...). Next, the item latent features are fixed to the above q (k, λ), and cross-validation data (m = 3800) is used instead of the training data, and each user v and each item that is a beauty product. For k, the user latent feature amount p (v, λ; k) is defined so as to minimize the modified cost function described later, which is a data version for cross-validation of the above cost function. Then, λ is selected so as to maximize the f1 score or the modified f1 score for the cross-validation data. The preferred value of the regularization parameter λ was λ = 0.032.

(Initial value of latent feature amount) The initial value of each component of the user latent feature amount and each component of the item latent feature amount is a uniform random number with a small absolute value, that is, the interval [-ε, ε] (however, ε = It was given as a uniform random number with 0.05) as the range.

(Update of latent feature amount) In this embodiment, the latent feature amount is updated by batch learning, and the ADADELTA method that can automatically adjust the learning rate is used. However, it is also possible to update the latent features by mini-batch learning or online learning. In the case of mini-batch learning or online learning, when additional data is added to the training data, it is also possible to train only the additional data and update the latent features. Further, the latent feature amount may be updated not only by the ADADELTA method but also by a method such as a stochastic gradient descent method, a conjugate gradient method, a momentum method, an ADAGRAD method, or a combination of two or more of these. The gradients ∂J / ∂p and ∂J / ∂q are easy to calculate. Further, in the present embodiment, when it is determined that the cost function J has reached the minimum value (stationary point), the update of the latent feature amount is completed, but even if the cost function does not stay, the cross-validation data is used. It may be determined that overfitting has started when the calculated f1 score or the modified f1 score reaches the stationary value and it is confirmed that the sign of the difference is inverted, and the update of the latent feature amount may be stopped. ..

(Item latent feature amount) FIG. 11 shows 34 items (questions or beauty products) k estimated to be suitable values by machine learning using training data in one embodiment according to the SCF model of the present invention. , Is a chart showing eight components of each item latent feature amount q (k) which is an eight-dimensional number vector. The present inventor has an item latent based on the correlation coefficient between each component of the user latent feature amount and the factor score derived by applying the factor analysis method to the answer to the question, the numerical value in the table of FIG. The beauty consciousness expressed by each component of the feature amount is q ₁ (beauty interest), q ₂ (greasy), q ₃ (dry), q ₄ (aggressive beauty) q ₅ (natural age), q ₆ (intermediate), q _It was identified as ₇ (natural beauty) and q ₈ (cover). Please refer to Table 6 below for a more detailed explanation of the beauty consciousness expressed by each component of the item latent feature amount and the user latent feature amount.

(User latent features) Table 5 shows each user latent feature, which is an 8-dimensional number vector for each user belonging to the training data, estimated at the same time as the item feature amount by machine learning using the training data. It is a chart which showed 8 components of the quantity p (k) for some users. Each component p _j of the user latent feature amount expresses the same beauty consciousness as the corresponding component q _j of the item feature amount.

(Varimax rotation) The item latent feature q shown in FIG. 11 and the user latent feature p shown in Table 5 are not actually suitable values estimated by repeatedly updating by machine learning, but are factor analysis methods. It is a "varimax rotation" of them, following the method in. This "varimax rotation" will be described. The predicted value a = g (p · q) in the SCF model is the rotation R in the b-dimensional number vector space, where p → p'= pR, q → q'= qR.
Is invariant to. Here, R is an arbitrary b-dimensional square matrix acting on a b-dimensional number vector, and p and q are expressed as row vectors. By utilizing this arbitrariness of R, it is possible to select R well and facilitate the interpretation of the meaning of each component of the rotated item latent feature q'. That is, in the same item latent feature amount shown in the table of FIG. 11 (however, unlike FIG. 11, the preferable value itself estimated by machine learning), the item latent feature amount of each item is q (k). ) (K = 1,2, ···, n) (n = 34), first, the correlation coefficient r = q (k) · q (l) / (| q) between the row vectors representing each item. (K) || q (l) |) is calculated, and the "commonality" of the item k is determined by the sum of the squares of the correlation coefficients for the item k and the other item l. Next, the item latent feature amount qs (k) (k = 1, 2, ..., N) after the scale obtained by dividing the item latent feature amount of each item k by the commonality of the item is obtained. Next, the Quarteymax rotations R ₁ , R ₂ , ..., R _c are repeatedly applied to the scaled item latent features until they converge. Assuming that the product of these rotation matrices is R = R ₁ R ₂ ... R _c , the user latent feature amount p'and the item latent feature amount q'are obtained by "varimax rotation" of p and q by the matrix R. The above "Quality Max rotation" is a b-dimensional number that maximizes the sum of the squares of the squares of each component of the post-scaled item latent features for each item. Rotation in vector space. It should be noted that each of the above R ₁ , R ₂ , ..., R _c is usually derived from a rotation matrix that transforms only two components of the b-dimensional number vector and retains the remaining (b-2) components. Be selected.

(Accuracy of Discrimination of Each Model) Table 3 above shows the f1 score and the modified f1 score calculated from the samples belonging to the test data in one embodiment according to the SCF model of the present invention. The discrimination accuracy of the SCF model is higher than that of the Deep model and the DeepPa model and lower than that of the DeepGo model in both the f1 score and the modified f1 score. It is a pity that the discrimination accuracy of the SCF model is lower than that of the DeepGo model, but the SCF model has an advantage that the DeepGo model does not have. In the DeepGo model, the basis of the recommended discrimination tends to be confined in the black box and it tends to be unclear, but in the SCF model, the item characteristics and the user's beauty consciousness are expressed in the form of item latent features and user latent features. Since it becomes clear, the basis of the recommendation judgment becomes clear, and the result of the recommendation judgment can be persuasive.

(Display of User Latent Feature) In FIG. 12, in one embodiment according to the SCF model of the present invention, eight components of the user latent feature are displayed on the output device by the user latent feature display means of the recommended discrimination program. It is explanatory drawing which shows the state of this. Previously, it was described that the result of recommended determination by the DeepGo model or the like, the conformance rate, the recall rate, etc. for a certain user are output to the output device in the form of the table shown in FIG. When this output device is a display or a touch panel display, for example, when the mouse pointer is placed on the beauty product G7 (beauty essence (whitening)) and clicked, each component of the user latent feature amount as shown in FIG. 12 is displayed. The sub-window to be displayed opens, or the screen display changes to the content displaying each component of the user latent feature amount as shown in FIG. 12, and the user's beauty consciousness can be read. The display is performed by the user latent feature amount display means. Preferably, it is desirable to display each component of the item latent feature amount of the beauty product (G7 in this example) together with each component of the user latent feature amount on the output device, and more preferably, it is directly related to the recommended determination. It is desirable to display the product of each component of the user latent feature amount and the component of the corresponding item latent feature amount on the output device. The display is preferably performed in a visually recognizable manner such as a bar graph. From the display contents of FIG. 12, it was determined that G7, which is a beauty product having an item latent feature amount positively related to the oiliness consciousness, is recommended because the user is highly interested in oiliness. I can guess. Since the user or the beauty advisor can objectively recognize the user's beauty consciousness by observing the display content by the user latent feature amount display means as shown in FIG. 12, the present embodiment can be applied to the user's beauty consciousness. You can select and recommend suitable beauty products.

(Display of Factor Score on Discriminating Surface) In another embodiment of the present invention, the user latent feature amount display means may be used together with the display of the user latent feature amount in the embodiment illustrated in FIG. Instead of displaying the user latent feature amount in the mode illustrated in FIG. 12, the user's factor score calculated from the answer to the question consisting of 21 items can be displayed on the identification surface by the factor analysis method. .. In this case, the user or the beauty advisor can obtain the user or beauty advisor from the factor scores such as the user's O1 factor, D factor, EST factor, MB factor, etc. plotted on the identification plane shown in FIGS. (18A) and / or (18B). Can objectively recognize the user's skin type and beauty consciousness. Therefore, in the present embodiment, the beauty advisor selects and recommends a beauty product suitable for the user's skin type and beauty consciousness, and the user or the beauty advisor understands the determination result of the recommendation determination of the beauty product by the computer control program. Above, it helps a lot.

(Flow Diagram of SCF Model) FIG. 13 is a flow diagram showing a processing flow of a computer control program in one embodiment according to the SCF model of the present invention. First, in step S21, it is determined whether or not the item latent feature amount can be read from a storage means such as a memory, HD, or removable memory, or a server connected via a network. If it can be read, hyperparameters such as the item latent feature amount and the regularization parameter and the dimension b of the latent feature amount are read in step S24, and the process proceeds to step S25. If it is not readable, the process proceeds to step S22. In step S22, preferable values of the item latent feature amount and the user latent feature amount are estimated with respect to the training data by machine learning. In addition, the preferred value of the regularization parameter λ is estimated using the cross-validation data. Next, in step S23, the item latent features and hyperparameters are stored in the above storage means and / or transmitted to a server, terminal, or the like connected via a network. Next, in step S25, for the purpose of minimizing the correction cost function described later for the user v whose item latent feature amount is fixed and the target of the recommended determination of the beauty product and each item k which is the beauty product, the user v The user latent feature amount of the user v is repeatedly updated, and the preferable value p (v, λ; k) of the user latent feature amount of the user v is estimated. Next, in step S26, the predicted value a of the user feature amount related to the beauty product k of the user v is calculated based on the item latent feature amount and the estimated user latent feature amount of the user v. Next, in step S27, the recommended determination of the beauty product is performed based on the predicted value and the predetermined threshold value. Next, in step S28, the precision and recall are calculated based on the predicted values. Finally, in step S29, the discrimination result output means and the discrimination accuracy display means output the discrimination result and / or the conformity rate and / or the recall rate to the output device, for example, in the manner shown in FIGS. 9 and / or FIG.

(Modified cost function) The modified cost function J is a function of the user latent feature amount p = p (v) of the user v who is the target of the recommended determination of the beauty product, and is defined by almost the same equation as the equation (6). .. However, unlike the equation (6), the sum of the first and second terms for the user u consists of only one term in the case of u = v. Further, since the item latent feature amount q (k) is fixed when estimating the user latent feature amount, the third term may be ignored. As the regularization parameter λ of the second term, the value estimated to be a suitable value in the previous step S22 and saved in the previous step S23, or the value read in the previous step S24 is used.

<5. Recommended discrimination system and e-learning system>
<5-1. Independent system>
FIG. 14 is a configuration diagram showing a system configuration according to an embodiment of the present invention. The recommended discrimination system of the present embodiment includes an independent computer 1 such as a personal computer, an input device 2 composed of a keyboard, a mouse, a touch panel display, and the like, and an output device 3 composed of a display, a printer, and the like. This recommended discrimination system may be operated by the user, or may be operated by a cosmetologist such as a cosmetology store while asking the user. The input device 2 is used to input the user feature amount of each user. Here, the user feature amount of each user includes answers to questions related to beauty, whether or not beauty products are used, images, measured values, and other information. The computer 1 executes a computer control program to perform recommended discrimination of beauty products and calculate the discrimination accuracy. The output device 3 displays the result of the recommended discrimination of the beauty product, and preferably displays the discrimination accuracy such as the conformity rate and the recall rate of the discrimination.

FIG. 15 is a configuration diagram for explaining a part of the configuration of the recommended determination system in detail. The computer 1 has a basic configuration 10. The basic configuration 10 includes a user feature amount input means 4 for inputting a user feature amount, a discrimination result output means 5 for displaying the result of recommended discrimination of a beauty product, and training data for storing training data. Data storage means 6, variable storage means 7 for storing user features of user v, which is the target of recommended determination, computer control program 8, computer control program 8, and / or parameter Θ, and / or item. It has a latent feature quantity and / or a computer-readable storage medium 9 for recording the regularization parameter λ. Here, the computer control program 8 has a recommended discriminant program 80, and the recommended discriminant program 80 has a predictive function output value calculation means 81 for calculating the output value of the predictive function.

Further, when the recommended discrimination system of the present embodiment includes the function of the beauty product e-learning system that allows the student to learn whether or not the beauty product g should be provided to the customer, the basic configuration 10 further includes the user. It has a user discrimination result input means 101 for inputting the discrimination result by the user, and a discrimination result comparison means 102 for comparing the discrimination result by the user with the discrimination result by the recommended discrimination system.

<5-2. System that determines by server>
FIG. 16 is a configuration diagram showing a system configuration according to another embodiment of the present invention. In this embodiment, as a general rule, the server 12 determines the recommendation of the beauty product, and the terminal 11 close to the user performs input and output. The recommended discrimination system of the present embodiment includes a terminal 11 and a server 12 capable of bidirectional communication via a network 13 or the like such as the Internet. The terminal 11 is composed of a computer such as a personal computer, a smartphone, and a tablet, an input device 2 composed of a keyboard, a mouse, a touch panel display, and the like, and an output device 3 composed of a display, a printer, and the like. The server 12 is a computer such as a personal computer or a server machine, and includes the basic configuration 10 and optionally an input device including a keyboard, a mouse, a touch panel display, and an output device including a display and a printer. May have. The terminal 11 in this embodiment may be operated by the user, or may be operated by a beauty advisor such as a beauty goods store while asking the user. The input device 2 is used to input the user feature amount of each user. The terminal 11 has a user feature amount transmitting means 41 for transmitting the user feature amount input by the input device 2 to the server 12. The server 12 receives the user feature amount transmitted by the user feature amount transmitting means 41 of the terminal 11 by the user feature amount receiving means 42. The server 12 executes a computer control program to perform recommended discrimination of beauty products, calculates discrimination accuracy, and transmits discrimination result transmission means 51 for transmitting the discrimination result and / or discrimination accuracy to the terminal 11. Have. The terminal 11 has a discrimination result receiving means 52 for receiving the discrimination result and / or the discrimination accuracy transmitted by the discrimination result transmitting means 51 of the server 12. The terminal 11 displays the discrimination result and / or the discrimination accuracy received by the discrimination result receiving means 52 on the output device 3.

In the present embodiment, the server 12 stores the user feature amount received by the user feature amount receiving means 42 in the variable storage means 7, and when the number of saved user feature amounts and the like satisfy a certain criterion. All or part of those user features are added to the training data and stored in the training data storage means 6, machine learning is performed based on the training data, and the trained parameters, that is, the parameters Θ and / or Hyper-parameters such as item features and / or regularization parameters may be updated and the updated learned parameters may be stored in variable storage means 7, storage medium 9, or other storage means.

Further, when the recommended discrimination system of the present embodiment includes the function of the e-learning system, the terminal 11 further includes a user discrimination result input means 101 for inputting a discrimination result by the user and a user discrimination result input means. It has a user discrimination result transmitting means 103 for transmitting the discrimination result by the user input by 101 to the server 12. Further, the server 12 further receives the user discrimination result receiving means 104 for receiving the discrimination result by the user transmitted by the user discrimination result transmitting means 103 of the terminal 11, the discrimination result by the user, and the discrimination result by the recommended discrimination system. It has a discrimination result comparison means 102 for comparison.

<5-3. System that determines by terminal>
FIG. 17 is a configuration diagram showing a system configuration according to still another embodiment of the present invention. Generally, machine learning requires high-load calculation, but if the trained parameters of the model are known, the calculation of predicted values and the calculation of recommended discrimination itself are generally light-load, and even on terminals such as mobile terminals. Can be done enough. Therefore, in the present embodiment, the terminal performs the calculation of the predicted value and the calculation of the recommended determination.
In the recommended discrimination system of the present embodiment, in principle, the terminal 11 performs the recommended discrimination of the beauty product in addition to the input of the user feature amount and the like and the output of the result of the recommended discrimination. The terminal 11 in the present embodiment includes an input device 2, an output device 3, and a user feature amount transmitting means 41 for transmitting the user feature amount input by the input device 2 to the server 12. The server 12 includes the basic configuration 10, has a user feature amount receiving means 42, and may optionally have an input device and an output device. The server 12 has learned parameter transmitting means 43 for transmitting learned parameters, that is, hyperparameters such as parameters Θ and / or item features and / or regularization parameters to the terminal 11. The terminal 11 has a learned parameter receiving means 44 for receiving the learned parameters transmitted by the learned parameter transmitting means 43 of the server 12. The terminal 11 executes a computer control program based on the learned parameters received by the learned parameter receiving means 44 and the user feature amount input by the input device 2, performs recommended discrimination of beauty products, and determines the discrimination accuracy. The calculation is performed, and the discrimination result and / or the discrimination accuracy is displayed on the output device 3.

Also in the present embodiment, the server 12 adds the user feature amount to the training data based on a certain standard, performs machine learning based on the additionally updated training data, updates the learned parameters, and is updated. The learned parameters may be configured to be stored in the variable storage means 7, the storage medium 9, or other storage means.

Further, when the recommended discrimination system of the present embodiment includes the function of the e-learning system, the terminal 11 further includes a user discrimination result input means 101 for inputting a discrimination result by the user and a user discrimination result input means. The user discrimination result transmission means 103 for transmitting the discrimination result by the user input by 101 to the server 12, and the discrimination result comparison means 102 for comparing the discrimination result by the user with the discrimination result by the recommended discrimination system. Have. In addition, the server 12 further has a user discrimination result receiving means 104 for receiving the discrimination result by the user transmitted by the user discrimination result transmitting means 103 of the terminal 11.

It goes without saying that the present invention is not limited to the above-described embodiment, and includes various modifications, design changes, and the like within the range that does not deviate from the technical idea of the present invention. ..

According to the present invention, a machine learning method using big data is used without premising the determination of skin type based on interview data including information on what beauty products the user usually uses. It is possible to provide a recommended determination method and system for beauty products that can directly determine whether or not the beauty product is suitable for the user. Further, according to the present invention, it is possible to provide a recommended determination method and system for a beauty product capable of evaluating the beauty consciousness of the user based on the interview data including the above information by the above machine learning method. .. According to the present invention, cosmetics and beauty services are recommended, beauty education, training of beauty advisors, etc. in online stores such as beauty salons, cosmetics stores, clinics, drug stores, esthetic salons, etc. or in real stores. be able to. The present invention has a wide range of industrial applicability.

1 Computer 2 Input device 3 Output device 4 User feature amount input means 5 Discrimination result output means 6 Training data storage means 7 Variable storage means 8 Computer control program 9 Storage medium 10 Basic configuration 11 Terminal 12 Server 13 Network 41 User feature amount transmission Means 42 User feature amount receiving means 43 Learned parameter transmitting means 44 Learned parameter receiving means 51 Discrimination result transmitting means 52 Discrimination result receiving means 80 Recommended discrimination program 81 Prediction function output value calculating means 101 User discrimination result input means 102 Discrimination result comparison Means 103 User discrimination result transmitting means 104 User discrimination result receiving means A, B, S Points on the graph of the sigmoid function G1 to G12 Beauty products Q1 to Q22 Questions related to beauty

Claims (13)

It is a method of determining whether or not the beauty product g should be recommended to the user v by executing a computer control program.
The computer control program has a recommended discrimination program for beauty products using machine learning.
The user features of a large number of users prepared in advance and the numerical values indicating the degree of recommendation of beauty products to those users are used as training data for machine learning.
It is characterized in that it is determined whether or not the beauty product g should be recommended to the user v by executing the recommendation determination program with the user feature amount related to the item other than the beauty product g of the user v as a variable. Recommended discrimination method for beauty products. The recommended determination method for a beauty product according to claim 1, wherein the numerical value indicating the degree of recommendation of the beauty product to the user is a numerical value indicating whether or not the beauty product is used by the user. The recommended determination method for a beauty product according to claim 1 or 2, wherein the item includes a beauty product and a question about beauty. The recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function.
The prediction function is defined by a parameter, and a request is made using a user feature amount related to an item other than the beauty product g of the user as an input value and a predicted value of a numerical value indicating the degree of recommendation of the beauty product g to the user as an output value. The recommended determination method for a beauty product according to any one of Items 1 to 3. The recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function.
The prediction function takes the user latent feature amount p = p (u) of the user u and the item latent feature amount q = q (i) of the item i as input values, and the user feature amount y (u) related to the item i of the user u. , I) The recommended determination method for beauty products according to any one of claims 1 to 3, wherein the predicted value is used as an output value. The recommended determination method for a beauty product according to claim 5, wherein the recommended determination program includes a user latent feature amount display means for displaying the user latent feature amount of the user v. The recommended discrimination method for a beauty product according to any one of claims 1 to 6, wherein the recommended discrimination program has a discrimination accuracy display means for displaying a discrimination accuracy rate and / or a recall rate. It consists of independent computers or servers and terminals that can communicate with each other in both directions.
It is a recommended determination system for beauty products that determines whether or not beauty product g should be recommended to user v by executing a computer control program.
The computer control program has a recommended discrimination program for beauty products using machine learning.
A training data storage means for storing user features of a large number of users and numerical values indicating the degree of recommendation of beauty products for those users as training data for machine learning, which are prepared in advance.
It has a variable storage means for storing a user feature amount related to an item other than the beauty product g of the user v as a variable.
A beauty product recommendation discrimination system, characterized in that it determines whether or not a beauty product g should be recommended to a user v based on the training data and the variables by executing the recommendation discrimination program. The recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function.
The prediction function is defined by a parameter, and a request is made using a user feature amount related to an item other than the beauty product g of the user as an input value and a predicted value of a numerical value indicating the degree of recommendation of the beauty product g to the user as an output value. Item 8. The recommended discrimination system for beauty products according to Item 8. The recommended discriminant program has a predictive function output value calculation means for calculating the output value of the predictive function.
The prediction function takes the user latent feature amount p = p (u) of the user u and the item latent feature amount q = q (i) of the item i as input values, and the user feature amount y (u) related to the item i of the user u. , I) The recommended discrimination system for beauty products according to claim 8, wherein the predicted value is used as an output value. The computer control program according to claim 1. A computer-readable memory comprising recording the computer control program according to claim 1 and / or the parameter according to claim 4 and / or the item latent feature amount according to claim 5. Medium. It is an e-learning system that allows students to learn whether or not to provide beauty products g to customers.
It comprises an independent computer or a server and a terminal capable of bidirectional communication with each other, has the training data storage means according to claim 8, and is configured to be able to execute the recommended determination program according to claim 8. ,
Discrimination result output means for displaying the discrimination result obtained by executing the recommended discrimination program based on the user feature amount other than the actual or virtual user's beauty product g and the training data, and / or the said An e-learning system characterized by having a discrimination result comparison means for comparing a discrimination result with a discrimination result by the student himself / herself.