JP2021103339A - Device, method, and program for analyzing customer attribute information - Google Patents

Abstract

To improve the accuracy and efficiency of data analysis of customer attribute information.SOLUTION: A device includes: an attribute database connection unit connected to an attribute database for storing a plurality of attribute values corresponding to a plurality of attributes for each of a plurality of target people; an attribute prediction model generation unit that generates a plurality of first attribute prediction models that each predict an attribute value of a first prediction target attribute, which is a prediction target, based on an attribute value of at least one attribute other than the first prediction target attribute of the plurality of attributes, by using the attribute database; and an attribute prediction model selection unit that selects a first attribute prediction model used for prediction of an attribute value of the first prediction target attribute based on a prediction error of each of the plurality of first attribute prediction models.SELECTED DRAWING: Figure 1

Description

The present invention relates to devices, methods, and programs for analyzing customer attribute information.

Conventionally, a system for collecting and analyzing customer data is known (see, for example, Patent Document 1). The data analyzed by such a system is used for customer recommendations, market research, and the like.
Patent Document 1 Japanese Unexamined Patent Publication No. 2015-76076

In recent years, expectations for analyzing and utilizing data related to customers have been increasing more and more, and it is desired to further improve the accuracy and efficiency of data analysis.

In order to solve the above problem, in the first aspect of the present invention, for each of a plurality of target persons, an attribute database connection connected to an attribute database for storing a plurality of attribute values corresponding to the plurality of attributes. First, the attribute value of the first prediction target attribute, which is the prediction target, is predicted based on the attribute value of at least one attribute other than the first prediction target attribute among the plurality of attributes, using the unit and the attribute database. The first attribute prediction model used for predicting the attribute value of the first prediction target attribute based on the prediction error of each of the attribute prediction model generation unit that generates a plurality of attribute prediction models and the first plurality of attribute prediction models. Provided is an apparatus including an attribute prediction model selection unit for selecting.

In the second aspect of the present invention, the computer uses an attribute database for storing a plurality of attribute values corresponding to a plurality of attributes for each of the plurality of subjects, and is a first prediction target to be predicted. The attribute prediction model generation stage of generating the first plurality of attribute prediction models that predict the attribute value of the attribute based on the attribute value of at least one attribute other than the first prediction target attribute among the plurality of attributes, and the computer. Provided a method including an attribute prediction model selection step of selecting a first attribute prediction model to be used for predicting the attribute value of the first prediction target attribute based on the prediction error of each of the first plurality of attribute prediction models. To do.

In a third aspect of the invention, an attribute database connection executed by a computer that connects the computer to an attribute database for storing a plurality of attribute values corresponding to the plurality of attributes for each of the plurality of subjects. First, the attribute value of the first prediction target attribute, which is the prediction target, is predicted based on the attribute value of at least one attribute other than the first prediction target attribute among the plurality of attributes, using the unit and the attribute database. The first attribute prediction model used for predicting the attribute value of the first prediction target attribute based on the prediction error of each of the attribute prediction model generation unit that generates a plurality of attribute prediction models and the first plurality of attribute prediction models. Provides a program that functions as an attribute prediction model selection unit for selecting.

The outline of the above invention does not list all the necessary features of the present invention. Sub-combinations of these feature groups can also be inventions.

The configuration of the system 100 according to this embodiment is shown together with the terminal 112, the terminal 152, and the terminal 192. An example of the data structure stored in the attribute DB 122 and the attribute DB 167 according to the present embodiment is shown. The operation flow of the prediction model generation device 150 according to this embodiment is shown. An example of the evaluation result of the attribute prediction model generated by the attribute prediction model generation unit 180 according to the present embodiment is shown. An example of the selection result of the attribute prediction model generated by the attribute prediction model selection unit 185 according to the present embodiment is shown. The attribute information acquisition flow in the attribute prediction apparatus 110 according to this embodiment is shown. The attribute prediction flow in the attribute prediction device 110 according to this embodiment is shown. An example of the dependency of the predicted attribute is shown. The attribute addition flow in the system 100 which concerns on this embodiment is shown. An example of the configuration of the computer 1900 according to the present embodiment is shown.

Hereinafter, the present invention will be described through embodiments of the invention, but the following embodiments do not limit the inventions that fall within the scope of the claims. Also, not all combinations of features described in the embodiments are essential to the means of solving the invention.

FIG. 1 shows the configuration of the system 100 according to the present embodiment together with the terminal 112, the terminal 152, and the terminal 192. The system 100 registers and manages attribute data (also referred to as "attribute information") of each target person who is a subscriber or member of a point service or the like in an attribute database (attribute DB). The system 100 predicts the attribute value by generating an attribute prediction model that mainly predicts the attribute in the attribute DB for which the attribute data has not been acquired. Then, the system 100 makes it possible to utilize the attribute DB including the predicted attribute value for the recommendation of the product or service (hereinafter collectively referred to as “product or the like”), market research, and the like.

The system 100 includes an attribute prediction device 110, a prediction model generation device 150, and a recommendation processing device 190. The system 100 may be realized by a computer such as a server computer, may be a system realized by a plurality of computers, or may be a distributed system realized by a plurality of locally distributed computers. Good. System 100 manages the attribute information of the target person such as each member in the point system provided by one business operator, the common common point system across a plurality of business operators, credit card, electronic money, and any other member service. ..

The attribute prediction device 110 inputs and registers the attribute information of the target person, and mainly predicts the unknown attribute information based on the registered attribute information. The attribute prediction device 110 according to the present embodiment is for a target person whose attribute information is known so that the attribute value between a plurality of target persons can be compared regardless of whether the attribute information is known or unknown. Also predicts its attribute information. The attribute prediction device 110 includes an attribute information acquisition unit 115, an attribute DB connection unit 120, an attribute DB 122, a dimension reduction unit 125, a reduction DB connection unit 130, a reduction DB 132, an attribute prediction unit 135, and the like. It has an attribute value updating unit 140 and an attribute predicted value updating unit 142.

The attribute information acquisition unit 115 acquires the attribute information of the target person and stores it in the attribute DB 122 via the attribute DB connection unit 120. As an example, the source of the attribute information of the target person is the information that the target person must or optionally fills in or inputs when registering a new member, the response of the target person to the member questionnaire, the target person purchased the product, etc. at the store, etc. Payment information according to the situation, payment information according to the purchase of products by the target person on the electronic commerce site, information on the website accessed by the target person, Internet advertisement clicked on the website by the target person (web advertisement) ), And at least one of the information of the TV program viewed by the target person, etc., and is provided with the consent of the target person. The attribute information acquisition unit 115 accesses the attribute DB 122 via the attribute DB connection unit 120 at an arbitrary timing such as periodically, searches for a target person whose attribute value for at least one attribute is unknown, and visits a website or the like. Unknown attribute values may be actively collected by conducting a member questionnaire or the like.

The attribute DB connection unit 120 is connected to the attribute DB 122 and processes access to the attribute DB 122 from each unit in the system 100. The attribute DB 122 stores attribute information including a plurality of attribute values corresponding to the plurality of attributes for each of the plurality of target persons. The attribute DB 122 may be realized by a storage area of at least a part of an external storage device such as a hard disk drive connected to a computer that processes the system 100, and is external to the system 100 provided by, for example, a cloud storage service. It may be realized by a storage device.

The dimension reduction unit 125 reduces the dimensions of a plurality of attributes based on the plurality of attribute values of each of the plurality of subjects stored in the attribute DB 122. As an example, the dimension reduction unit 125 uses a topic model to calculate a plurality of values representing the degree or probability that the attribute information of each target person included in the attribute DB 122 corresponds to each of the plurality of topics. Then, the dimension reduction unit 125 adds the calculated plurality of values to the attribute information of each target person as a plurality of attributes for dimension reduction. As a result, the dimension reduction unit 125 can reduce a large number of attributes constituting the attribute information of the target person to a smaller number of attributes for each topic. Along with this, the dimension reduction unit 125 may remove at least a part or all of the attributes other than the plurality of attributes added for the dimension reduction from the explanatory variables of the attribute prediction model. The dimension reduction unit 125 applies the above-mentioned dimension reduction processing to the attribute DB 122 to convert it into the reduction DB 132.

The contraction DB connection unit 130 is connected to the contraction DB 132 and processes a write access from the dimension contraction unit 125 to the contraction DB 132 and a read access of the contraction DB 132 from the attribute prediction unit 135. The contraction DB 132 stores a plurality of attribute values corresponding to a plurality of attributes as attribute information, including at least a part of the attributes contracted by the dimension contraction unit 125 for each of the plurality of subjects. Similar to the attribute DB 122, the contraction DB 132 may be realized by an external storage device connected to a computer that processes the system 100, or is realized by a storage device provided by a cloud storage service or the like outside the system 100. You may.

The attribute prediction unit 135 receives the attribute prediction model selected for each of one or more attributes (referred to as “prediction target attribute”) of the prediction target in the attribute DB 122 from the prediction model generation device 150. Then, the attribute prediction unit 135 predicts the attribute value of the prediction target attribute for each of the plurality of target persons by using the received attribute prediction model. More specifically, the attribute prediction unit 135 executes a selected attribute prediction model for each prediction target attribute for all or at least a part of the target persons, and stores each in the reduction DB 132. The predicted value of the predicted target attribute is calculated based on the attribute information of the target person. The attribute prediction unit 135 stores the calculated predicted value in the attribute DB 122 via the attribute DB connection unit 120.

The attribute DB 122 according to the present embodiment corresponds to each of a plurality of attributes, and has a known attribute value based on the attribute information acquired by the attribute information acquisition unit 115 and a predicted value predicted by the attribute prediction unit 135. Store both as attribute values in a broad sense. In the present embodiment, for convenience of explanation, the attribute DB 122 stores known attribute values and predicted values using the same expression format. In this case, the attribute information acquisition unit 115 converts the representation format of the raw data of the acquired attribute information, and the attribute DB 122 stores the attribute value whose representation format has been converted as a known attribute value. For example, the attribute information acquisition unit 115 may convert the attribute value of the absolute evaluation acquired as the raw data of the attribute information into the attribute value of the relative evaluation, and the attribute DB 122 may store the attribute value converted into the relative evaluation. As such an example, the attribute DB 122 has a relative value (for example, deviation value or percentile ranking) among all subjects whose ages are known, instead of the age of the subject represented by an absolute evaluation such as 35 years old. Etc.), and the converted attribute value may be stored. As another example, there is a case where the attribute DB 122 stores the attribute value converted into the probability of being male, the deviation value of the possibility of being male, or the like with respect to the gender of the subject. Here, the attribute DB 122 stores the predicted value of the attribute in an expression format such as a deviation value, a percentile ranking, or a probability in the same manner as the known attribute value after conversion, so that the predicted value is between the known attribute value and the predicted value. You can get compatibility with. As a result, the system 100 can use a known attribute value as an explanatory variable when generating the attribute prediction model, and can also use the predicted value as an explanatory variable when using the attribute prediction model. As a result, the attribute prediction device 110 uses the attribute prediction model to obtain the attribute value of the prediction target attribute even for the target person for which the known attribute value has not been obtained for at least some of the attributes that are the explanatory variables of the attribute prediction model. Can be predicted.

The attribute DB 122 can be realized in various other forms such as a form in which a known attribute value is stored in the expression format before conversion and a predicted value is stored in the expression format after conversion. Even in such a case, the system 100 is known by converting the representation format of the predicted value to the representation format of the known attribute value each time the predicted value is referred to instead of the known attribute value. It is possible to realize the same operation as when the attribute value and the predicted value are stored in the same expression format. Further, the attribute DB 122 may separately store the raw data of the attribute information acquired by the attribute information acquisition unit 115 in addition to the known attribute value whose expression format has been converted.

The attribute value update unit 140 updates the attribute value of a certain attribute stored in the attribute DB 122 on the condition that the certainty of the predicted value is equal to or higher than the threshold value. As a result, when the accuracy of the predicted value is sufficiently high, the unacquired attribute value can be set based on the predicted value.

When the attribute value of a certain attribute stored in the attribute DB 122 is known, the attribute predicted value update unit 142 sets the predicted value of the attribute according to the deviation from the known attribute value by a reference or more. Update the predicted value of the attribute based on the known attribute value. As a result, the predicted value can be rewritten to a more accurate value when the predicted value greatly deviates from the known attribute value for some target persons.

Further, the attribute prediction device 110 may have an attribute data acquisition unit 144 and an attribute addition unit 146. The attribute data acquisition unit 144 acquires known information indicating the known attribute values of the additional attributes to be added to the attribute database for at least a part of the plurality of target persons. For example, the attribute data acquisition unit 144 receives known information from the user of the system 100 or the attribute prediction device 110 via the terminal 112.

The attribute addition unit 146 adds additional attributes to a plurality of attributes of the attribute DB 122. The attribute addition unit 146 sets the known attribute value of the additional attribute for each target person included in the known information based on the attribute value indicated in the known information acquired by the attribute data acquisition unit 144. For the target person whose known attribute value is not shown in the known information, the attribute addition unit 146 does not set the known attribute value.

When the additional attribute is added to the attribute DB 122, the system 100 may process the additional attribute as a prediction target attribute. That is, the prediction model generation device 150 may generate one or a plurality of attribute prediction models for predicting the attribute value of the prediction target attribute by the attribute prediction model generation unit 180, using the additional attribute as the prediction target attribute. Further, when a plurality of attribute prediction models are generated, the prediction model generation device 150 selects an attribute prediction model to be used for predicting the attribute value of the prediction target attribute by the attribute prediction model selection unit 185 based on each prediction error. You can do it.

The prediction model generation device 150 generates an attribute prediction model that predicts each of one or more attributes to be predicted in the attribute DB 122. The prediction model generation device 150 includes a model update instruction unit 155, a sampling unit 160, an attribute DB connection unit 165, an attribute DB 167, a dimension contraction unit 170, a contraction DB connection unit 175, and a contraction DB 177. It has an attribute prediction model generation unit 180 and an attribute prediction model selection unit 185.

The model update instruction unit 155 instructs the update of the attribute prediction model used by the attribute prediction device 110 according to the elapse of a predetermined period. The model update instruction unit 155 triggers the sampling unit 160 to start processing according to the elapse of a period such as one day, one week, or one month since the last time the attribute prediction model was updated. Instruct the update of the attribute prediction model by such means. Instead, the model update instruction unit 155 transmits the attribute prediction model already prepared in the prediction model generation device 150 during the period to the attribute prediction device 110 according to the elapse of the period. This may be instructed to the attribute prediction model selection unit 185.

The sampling unit 160 samples a part of the target persons in which the attribute DB 122 stores the attribute information, and sets the attribute information stored in the attribute DB 122 in association with the sampled target person as the attribute DB. It is stored or copied in the attribute DB 167 via the connection unit 165. Here, the sampling unit 160 is among the target persons registered in the attribute DB 122, among the target persons whose attribute values of the predicted target attribute are known, the target persons for the number of samples used to generate the attribute prediction model. May be extracted. Instead of this, the sampling unit 160 may randomly extract as many target persons as the number of samples used for generating the attribute prediction model from the target persons registered in the attribute DB 122.

The attribute DB connection unit 165 is connected to the attribute DB 167 and processes access to the attribute DB 167 from each unit in the prediction model generation device 150. The attribute DB 167 stores attribute information including a plurality of attribute values corresponding to the plurality of attributes for each of the plurality of target persons sampled by the sampling unit 160 among the target persons registered in the attribute DB 122. Like the attribute DB 122, the attribute DB 167 may be realized by an external storage device connected to a computer that processes the system 100, or is realized by a storage device provided by a cloud storage service or the like outside the system 100. May be good.

The dimension reduction unit 170 reduces the dimensions of a plurality of attributes based on the plurality of attribute values of each of the plurality of subjects stored in the attribute DB 167. The dimensional contraction unit 170 may apply the same dimensional contraction process as the dimensional contraction unit 125 to the attribute DB 167 to convert the attribute DB 167 into a contraction DB 177.

The contraction DB connection unit 175 is connected to the contraction DB 177 and processes a write access from the dimension contraction unit 170 to the contraction DB 177 and a read access of the contraction DB 177 from the attribute prediction model generation unit 180. The contraction DB 177 has a plurality of attribute values corresponding to a plurality of attributes including at least a part of the attributes contracted by the dimension contraction unit 170 for each of the plurality of subjects sampled by the sampling unit 160. Memorize information. Similar to the attribute DB 122, the contraction DB 177 may be realized by an external storage device connected to a computer that processes the system 100, or is realized by a storage device provided by a cloud storage service or the like outside the system 100. You may.

The attribute prediction model generation unit 180 uses the attribute DB 167 to predict the attribute value of the prediction target attribute based on the attribute value of at least one attribute other than the prediction target attribute among the plurality of attributes. To generate. The attribute prediction model generation unit 180 according to the present embodiment generates a plurality of attribute prediction models by referring to the attribute DB 167 and using the attribute values associated with some of the target persons sampled from the attribute DB 122. To do. As a result, the attribute prediction model generation unit 180 reduces the amount of calculation and the calculation time required when each of the plurality of attribute prediction models is generated using the attribute information of all the target persons stored in the attribute DB 122. be able to. If sufficient computing power can be secured, the attribute prediction model generation unit 180 may generate a plurality of attribute prediction models using the attribute information of each target person stored in the attribute DB 122 without using sampling. Good.

Further, in the present embodiment, the attribute prediction model generation unit 180 may indirectly use the attribute DB 167 by referring to the contraction DB 177 converted from the attribute DB 167. As a result, the attribute prediction model generation unit 180 can predict the attribute value of the prediction target attribute from the attribute value of at least one of the plurality of dimension-reduced attributes. Here, by excluding all attributes or at least some attributes other than the plurality of attributes added by the dimension reduction unit 170 for dimension reduction from the explanatory variables of the attribute prediction model, the attribute prediction model generation unit 180 can generate the dimension. By effectively using the result of reduction, it is possible to generate an attribute prediction model using only some of the attributes that can affect the prediction of the prediction target attribute as explanatory variables. As a result, the attribute prediction model generation unit 180 can reduce the amount of calculation and the calculation time required to generate a plurality of attribute prediction models.

In the above, the attribute prediction model generation unit 180 can generate a plurality of attribute prediction models for predicting the prediction target attribute for each of the plurality of prediction target attributes. Here, the attribute prediction model generation unit 180 is different as a first attribute prediction model used for predicting the attribute value of the first prediction target attribute and a second attribute prediction model used for predicting the attribute value of the second prediction target attribute. Attribute prediction model can be selected. Each of these plurality of attribute prediction models may differ from other attribute prediction models in at least one of the hyperparameters and prediction algorithms that are not updated by training. The attribute prediction model generation unit 180 trains the learnable parameters in the attribute prediction model for each of the plurality of attribute prediction models, and each of the plurality of attribute prediction models predicts the attribute value of the prediction target attribute. Is performed to minimize.

The attribute prediction model selection unit 185 predicts the attribute value of the prediction target attribute based on the prediction error of each of the plurality of attribute prediction models generated by the attribute prediction model generation unit 180 for each of the plurality of prediction target attributes. Select the attribute prediction model used for. For example, the attribute prediction model selection unit 185 selects an attribute prediction model having the smallest prediction error from a plurality of attribute prediction models for each prediction target attribute. At this time, the attribute prediction model selection unit 185 can select an attribute prediction model in which at least one of the hyperparameters and the prediction algorithm is different for each prediction target attribute. The attribute prediction model selection unit 185 provides the attribute prediction model selected for each prediction target attribute to the attribute prediction unit 135 in the attribute prediction device 110. As an example, the attribute prediction model selection unit 185 provides the attribute prediction unit 135 with a set of identification information and hyperparameter values that specify a prediction algorithm used for the attribute prediction model.

The recommendation processing device 190 selects whether or not to recommend a product or the like associated with the prediction target attribute to the target person based on the attribute value of the prediction target attribute set as the prediction target in each of the plurality of target persons. Here, the recommendation processing device 190 indicates that, for each of a plurality of target persons, when the attribute value of the predictive target attribute indicating the preference for a specific product or the like shows the preference of the target person or more, the product or the like is given to the target person. May be decided to recommend. In addition, the recommendation processing device 190 is a prediction target that indicates one or more basic attributes, living attributes, and at least a part of orientation (for example, place of residence, whether or not a car is owned, luxury-oriented or frugality-oriented). It may be decided to recommend a product or the like based on the attribute value of the attribute.

Then, the recommendation processing device 190 recommends a product or the like to a target person who has decided to make a recommendation among a plurality of target persons. As an example, the recommendation processing device 190 may provide an e-mail including an advertisement for the product or the like, a direct mail, an Internet advertisement or the like to the target person, and a TV commercial containing the advertisement for the product or the like may be provided to the target person. It may be provided to viewers including those, and services such as coupons, discounts, and point awarding that give preferential treatment to the purchase of the product or the like may be provided.

The terminal 112 is directly or indirectly connected to the attribute prediction device 110 via a network, and statistically processes the attribute information of a plurality of target persons stored in the attribute DB 122 or the contraction DB 132 of the attribute prediction device 110, and the attributes of the attributes. To perform database processing such as addition / deletion, setting of attribute values, narrowing down of target persons who satisfy the extraction conditions set by at least one set of attributes, and evaluation of the attribute prediction model for each prediction target attribute of the prediction target. Provides a user interface for.

The terminal 152 is directly or indirectly connected to the prediction model generation device 150 via a network, and the attribute prediction model generation unit 180 adds, changes, or deletes a prediction algorithm that can be generated by the attribute prediction model generation unit 180. Specifying the range and value of hyperparameters given to the attribute prediction model, checking and evaluating the trained parameters and prediction results of each of the plurality of attribute prediction models generated by the attribute prediction model generation unit 180, comparing between attribute models, and attributes. Manual selection specification or selection assistance of the attribute prediction model for the prediction model selection unit 185, addition / deletion of attributes for the attribute DB167 which is a subset of the attribute DB122, setting of the attribute value, and extraction conditions set by at least one set of attributes. Provided is a user interface for performing database processing such as narrowing down of target persons satisfying the conditions, and statistical processing of attribute information of a plurality of target persons stored in the attribute DB 167 or the contraction DB 177 of the prediction model generation device 150.

The terminal 192 is directly or indirectly connected to the terminal 192 via a network to provide a user interface for managing the recommendation processing in the recommendation processing device 190. As an example, the terminal 192 sets and adjusts conditions for narrowing down the recommendation target people according to the advertising budget and the discount or point granting budget for products, etc., confirms the narrowing down result of the recommendation target people, and makes a recommendation method. It provides a user interface for setting and instructing recommendation execution.

The terminal 112, the terminal 152, and the terminal 192 may be desktop computers or mobile terminals such as tablets and smartphones.

According to the system 100 shown above, a plurality of attributes that predict the attribute value for each prediction target attribute based on the attribute information of the plurality of target persons stored in the attribute DB 122 or the attribute DB 167 that is a subset of the attribute DB 122. A prediction model can be generated and a suitable attribute prediction model can be selected for each prediction target attribute. Therefore, according to the system 100, it is possible to further improve the prediction accuracy for each prediction target attribute.

FIG. 2 shows an example of the data structure stored in the attribute DB 122 and the attribute DB 167 according to the present embodiment. The attribute DB 122 and the attribute DB 167 store personal identification information (personal ID) for identifying an individual and attribute information about a plurality of attributes possessed by the individual for each of the plurality of target persons.

The "individual ID" is an identifier for identifying an individual target person in the system 100, and is, for example, a membership number or a login ID of a service provided by the system 100. Instead, the attribute DB 122 and the attribute DB 167 can be used as the "personal ID" in the target person's name, e-mail address, address, telephone number, identification information of the target person's mobile terminal, or at least one combination thereof. Information generated based on this may be used.

"Attribute information" is an attribute value for various attributes possessed by the target person. Each attribute stored in the attribute DB 122 and the attribute DB 167 according to the present embodiment is roughly classified into general-purpose attribute data, purchasing potential data, and recommended potential data.

The "general-purpose attribute data" is a set of attributes showing the characteristics of each target person, and particularly shows the characteristics of the target person itself in a general-purpose manner. "General-purpose attribute data" includes at least one of one or more attributes classified as basic attributes, one or more attributes classified as living attributes, and one or more attributes classified as intentionality. May include.

The "basic attribute" is the basic information of each target person, and one or more attributes classified into the basic attributes include, for example, at least one such as name, date of birth, age or age, gender, address, and telephone number. Including one. The "basic attribute" is an attribute that is mainly input when the target person is newly registered or when the registered contents are changed, but at least some of the attributes may be voluntary registration and may be a prediction target.

The "living attribute" is information on the lifestyle of the target person, and one or more attributes classified into the "living attribute" include, for example, married / unmarried, housing form, household annual income, personal annual income, occupation, and ownership of a car. It may include at least one such as presence / absence and possession of a residence. The attribute value related to the "living attribute" may be collected at the time of new registration or the like, may be collected by various methods such as a questionnaire, or may be a prediction target.

"Intentionality" is information indicating the intention, tendency and / or preference of the subject, and one or more attributes classified as "intentionality" include, for example, quality-oriented / challenge-oriented / solid-oriented / brand for clothing. It may include at least one of health-oriented, career-oriented, global-oriented, etc., such as high-class-oriented / frugal-oriented / discount-oriented, convenience-oriented / urban-oriented / community-oriented, etc. for housing. Alternatively, intentionality is an attribute related to the taste of the subject, for example, whether or not there is a preference for various hobbies such as driving, gourmet, travel, and sports, or the degree of preference, whether or not there is a preference for various products, or the degree of preference, and It may include at least one such as presence / absence of preference or degree of preference for various websites and the like. Attributes related to "intentionality" can be added in various surveys depending on the purpose of the survey. The attribute value related to "intentionality" may be collected at the time of new registration or the like, may be collected by various methods such as a questionnaire, or may be a prediction target.

The "purchasing potential data" is a set of attributes indicating the purchaseability of each target person for each of a plurality of products, etc., or a plurality of product groups or service groups. "Purchase potential data" may include, for example, attributes associated with each product or service corresponding to the genre, type, or classification of products such as entertainment, food, and daily necessities. .. Each attribute of the "purchasing potential data" may be a preference attribute indicating the preference of the target person for the product or the like associated with the attribute.

As an example, the "purchasing potential data" includes attributes for each of a large number of products and the like subject to sales management by the member service provided by the system 100. The "purchase potential data" may include each attribute corresponding to each code value of the JAN code that identifies each product or the like. In this case, 1 for each product or the like to which the JAN code is assigned. Or two or more attributes are assigned. When the member service provided by the system 100 provides a common point system across a plurality of businesses, the purchasing potential data may include hundreds of thousands to millions or more attributes. Further, the "purchasing potential data" may have an attribute corresponding to a product group or a service group. For example, the "purchasing potential data" may have attributes corresponding to product groups such as beer, alcoholic beverages, and / or beverages.

"Purchase potential data" is the degree of preference that quantifies the purchase record (presence or absence of purchase, purchase amount, purchase time, purchase location, etc.) and / or the reaction of the target person to the product questionnaire or advertisement for each target person. Etc. may be stored as known attribute values. Further, in the "purchasing potential data", the predicted value predicted by using the attribute prediction model may be stored as at least a part of the attribute value.

"Recommend potential data" is a set of attributes indicating the characteristics of each subject regarding recommendations. The "recommended potential data" may include, for example, at least one or more attributes of the media reaction, one or more attributes of the incentive reaction, and one or more attributes of the possibility of separation.

"Media reaction" is, for example, the effectiveness or evaluation of the recommendation using the media for the target person for each media used for the recommendation such as direct mail, e-mail, advertisement to be printed on the receipt, Internet advertisement, and TV advertisement. Includes one or more attributes that indicate values and the like. For example, if a target person does not respond to direct mail, the recommendation processing device 190 may reduce the attribute value of the attribute related to the effectiveness or evaluation value of direct mail, or set a lower attribute value. Good. In addition, when a certain target person clicks on the Internet advertisement more frequently than the standard frequency, or when the product is purchased via the Internet advertisement, the recommendation processing device 190 sets the effectiveness or evaluation value of the Internet advertisement, etc. You may increase the attribute value of the attribute related to, or set a higher attribute value.

The "incentive reaction" is, for example, for each incentive such as discount, coupon grant, point grant, point increase, and provision of sales promotion products, the effectiveness or evaluation value of the recommendation using the incentive to the target person, etc. Includes one or more attributes to indicate. For example, if a certain target person does not purchase a product or the like even if the discount is provided, the recommendation processing device 190 reduces the attribute value of the attribute related to the effectiveness of the discount or the evaluation value or the like. Alternatively, a lower attribute value may be set. In addition, when a target person purchases a product or the like for which points have been granted, the recommendation processing device 190 increases the attribute value of the attribute related to the effectiveness of point granting or the evaluation value, or increases the attribute value. It may be set.

"Possibility of separation" includes, for example, one or more attributes indicating the possibility that each seller and / or each product maker or service provider will not use the products of the business operator or the like. .. For example, when a certain target person has passed a reference period or more since the last purchase of a product or the like from a certain sales company, the attribute prediction device 110 considers that the target person has separated from the sales company and indicates that the target person has separated. The attribute value indicating is stored in the attribute DB 122. Instead of this, the attribute prediction device 110 may store the separation evaluation value calculated according to the elapsed period since the target person last used the seller as the attribute value in the attribute DB 122.

FIG. 3 shows an operation flow of the prediction model generation device 150 according to the present embodiment. In step S310, the model update instruction unit 155 instructs to update the attribute prediction model used by the attribute prediction device 110 according to the elapse of a predetermined period. Alternatively or additionally, the model update instruction unit 155 may instruct the update of the attribute prediction model in response to an instruction from a user or administrator of the system 100.

Here, the prediction model generation device 150 may use all the attributes of the attribute DB 122 or the attribute DB 167 as prediction targets, or each attribute excluding some attributes such as the name of the target person as the prediction target, and general-purpose attribute data. Each attribute excluding the basic attribute included in may be the prediction target, and the purchase potential data and / or the recommendation potential data may be the prediction target. Further, the prediction model generation device 150 may exclude the attribute for dimension reduction added by the dimension reduction unit 170 from the prediction target. Further, the prediction model generation device 150 may target only the attributes specified by the user or the administrator of the system 100 among all the attributes possessed by the attribute DB 122 or the attribute DB 167.

The prediction model generation device 150 may instruct the update of the attribute prediction model at a timing that may differ depending on the prediction target attribute. For example, the prediction model generation device 150 periodically instructs the update of the attribute prediction model for the attribute belonging to the general-purpose attribute data, and purchases a product or the like corresponding to the attribute for the attribute belonging to the purchasing potential data as an example. Instruct the update of the attribute prediction model every time the data is received in a predetermined quantity, and for the attributes belonging to the recommendation potential data, instruct the update of the attribute prediction model every time the recommendation is performed a predetermined number of times as an example. May be good. Further, the prediction model generation device 150 may update the attribute prediction model that predicts the attributes belonging to the purchasing potential data lower than the update frequency of the attribute prediction model that predicts the attributes belonging to the general-purpose attribute data. Since the purchasing potential data includes attributes corresponding to individual products and the like, the number of attributes is usually larger than that of general-purpose attribute data. Therefore, the prediction model generator 150 predicts the attributes belonging to the purchasing potential data. By lowering the update frequency of, the calculation amount of the prediction model generation device 150 can be significantly reduced.

In S320, the sampling unit 160 samples a part of the target person from the attribute DB 122, and the attribute information stored in the attribute DB 122 in association with the sampled target person is transmitted to the attribute DB 167 via the attribute DB connection unit 165. Store in.

In S330, the dimension contraction unit 170 uses a topic model such as Latent Dirichlet Allocation (LDA) or Stochastic Latent Semantics Analysis (LDA) as an example to determine the dimension of the attribute DB 167. Reduce. When the topic model is used, the dimension reduction unit 170 models the number of topics as each of a plurality of types such as 10, 20, 30, ..., 100, etc., and a plurality of attributes generated for each number of topics are used for dimension reduction. It may be stored in the contraction DB 177 as an attribute of. As a result, the attribute prediction model generation unit 180 can use the number of topics to be used as a hyperparameter of the attribute prediction model, and the attribute prediction model selection unit 185 can select the attribute prediction model with the optimum number of topics. ..

Here, the dimension reduction unit 170 may use all the attributes of the attribute DB 122 or the attribute DB 167 as evaluation targets for topic generation, and instead, a part of the attributes of the attribute DB 122 or the attribute DB 167 may be evaluated for topic generation. May be. For example, in the dimension reduction unit 170, the attributes included in the general-purpose attribute data may be evaluated for topic generation, and the attributes included in the purchasing potential data and the recommendation potential data may not be used for topic generation. Further, for example, the dimension reduction unit 170 may evaluate only the basic attributes, the living attributes, and the attributes included in only a part of the orientations included in the general-purpose attribute data as the evaluation target of the topic generation. As an example, the dimension reduction unit 170 targets only the attributes included in the basic attribute and the living attribute as the evaluation target of the topic generation, and the attribute included in the intentionality does not have to be used for the topic generation. When all attributes are evaluated for topic generation, even small differences between the target people (for example, whether or not a specific product was purchased) can be included in the evaluation and reflected in the attribute value after dimension reduction. The amount of calculation required for the reduction process increases. When only some attributes are to be evaluated for topic generation, the amount of calculation can be suppressed by using only some attributes that are emphasized by the user or administrator of the system 100.

Further, the dimension reduction unit 170 may reduce the dimension of the attribute DB 167 by another method instead of the topic model or in addition to the topic model. For example, the dimension reduction unit 170 may exclude an attribute in which the ratio of the target person in which the known attribute value is stored is less than the reference ratio from the explanatory variables of the attribute prediction model.

Next, the prediction model generation device 150 repeats the processing (S350 to S370) between S340 and S380 for each prediction target attribute.

In S350, the attribute prediction model generation unit 180 generates one or a plurality of attribute prediction models for predicting the attribute value of the prediction target attribute. First, the attribute prediction model generation unit 180 extracts the attribute information of a plurality of target persons for modeling used for generating the attribute prediction model from the attribute information of each target person stored in the attribute DB 167. The attribute prediction model generation unit 180 extracts the attribute information for which the attribute value of the prediction target attribute is known, that is, the attribute information for which the attribute value is actually obtained for the prediction target attribute, as the attribute information of the target person for modeling. .. Here, the attribute prediction model generation unit 180 knows the attribute value of each attribute that may be used as an explanatory variable of the attribute prediction model (that is, an attribute that inputs an attribute value to the attribute prediction model as a candidate for the explanatory variable). Attribute information with a larger proportion of those that are may be extracted with higher priority. When the number of subjects for modeling is equal to or less than a predetermined reference number, the attribute prediction model generation unit 180 may increase the number of samples by using a method such as boost sampling.

Next, the attribute prediction model generation unit 180 uses the extracted attribute information of the target person as learning attribute information to generate a plurality of attribute prediction models. The attribute prediction model generation unit 180 generates a plurality of attribute prediction models in which at least one of the hyperparameters that are not updated by the prediction algorithm and learning is different from each other. The attribute prediction model generation unit 180 may select a prediction algorithm from various machine learning algorithms including, for example, random forest, gradient boosting, logistic regression, neural network, and support vector machine (SVM). In addition, the attribute prediction model generation unit 180 includes the number of dimensional reduction topics in the dimensional reduction unit 170, the depth of the decision tree in the random forest, the depth of the tree in gradient boosting, the normalization parameters in logistic regression, and the neural. By independently selecting each of one or more hyperparameters that can be set in a prediction algorithm such as the number of neurons and the number of layers in a network from a set of multiple settings that can be set as hyperparameters. You may obtain a plurality of hyperparameter sets that are different from each other for each selected prediction algorithm.

The attribute prediction model generation unit 180 optimizes each of a plurality of attribute prediction models in which at least one of the prediction algorithm and the hyperparameter set obtained as described above is different from each other by learning using the learning attribute information. .. That is, when the attribute prediction model generation unit 180 inputs an attribute value other than the prediction target attribute among the learning attribute information for each learning attribute information as the input of the attribute prediction model, the prediction target in the learning attribute information. The trainable parameters of the attribute prediction model are updated so that the predicted values closer to the known attribute values of the attributes are output. As a result, the attribute prediction model generation unit 180 learns and optimizes the learnable parameters for each of the attribute prediction models. Here, "optimization" does not necessarily mean minimizing the prediction error of the attribute prediction model (that is, setting all the learnable parameters to the optimum values), and the prediction error of the attribute prediction model is used. It means a state in which the attribute prediction model generation unit 180 has completed the learning process to be further reduced (that is, a state in which the practically feasible learning process is completed).

Here, the attribute prediction model generation unit 180 may limit the explanatory variables of the attribute prediction model to only known attribute values. In this case, the attribute prediction model generation unit 180 is limited in the number of samples that can be used as the learning attribute information, but increases the possibility that a more accurate attribute prediction model can be generated based on the actual attribute information. Can be done. Instead, the attribute prediction model generation unit 180 may use the predicted value of the attribute as a part or all of the explanatory variables. That is, the attribute prediction model generation unit 180 predicts the attribute value of the second prediction target attribute by using the prediction value of the first prediction target attribute predicted by the first attribute prediction model. An attribute prediction model may be generated. In the generation of the second plurality of attribute prediction models, the attribute prediction model generation unit 180 by making the predicted value of the first prediction target attribute predicted by the already generated first attribute prediction model available as an explanatory variable. Can also use training attribute information that lacks some known attribute values, and can improve the accuracy of the attribute prediction model by using more training attribute information.

The attribute prediction model generation unit 180 determines the difference between the predicted value and the expected value of the predicted target attribute predicted by the attribute prediction model from each attribute information as the prediction error for all the attribute information to be evaluated for the prediction error. Cumulative values may be used. As an example of such an error calculation method, various known methods can be applied, including a method of summing the squares of the difference between the predicted value and the expected value. Instead, the attribute prediction model generation unit 180 may use a value (1-AUC value as an example) that decreases as the AUC (Area Under the Curve), which will be described later, increases, as an index indicating a prediction error.

Here, the "prediction error" includes an index that changes according to the "prediction error". That is, for example, an index indicating prediction accuracy also corresponds to an index indicating prediction error, and a higher prediction accuracy corresponds to a lower prediction error, and a lower prediction accuracy means a higher prediction error. Corresponds to high. Therefore, the attribute prediction model generation unit 180 may use the AUC, which increases as the prediction accuracy increases, as an index indicating a prediction error (an index that decreases as the prediction error increases). In this case, a high prediction error means that the index value indicating the prediction error is small.

In S360, the attribute prediction model generation unit 180 evaluates each of the generated plurality of trained attribute prediction models. The attribute prediction model generation unit 180 may evaluate the attribute prediction model based only on the prediction error, and not only the prediction error but also, for example, a smaller amount of calculation and / or a smaller number of referenced attributes. Other conditions such as may be included as evaluation parameters. Further, the attribute prediction model generation unit 180 may evaluate the attribute prediction model by using the value of the objective function including the term related to the prediction error, which is used for optimizing the prediction algorithm. Here, the attribute prediction model generation unit 180 may calculate the prediction error of each of the plurality of attribute prediction models by using the cross test. In this case, the attribute prediction model generation unit 180 trains a plurality of attribute prediction models using a part of the extracted attribute information for modeling, and uses the attribute information for modeling that is not used for training. Calculate the prediction error of each of the multiple attribute prediction models.

In S370, the attribute prediction model selection unit 185 selects an attribute prediction model to be used for predicting the attribute value of the prediction target attribute based on the evaluation results such as the prediction errors of each of the plurality of attribute prediction models. As an example, the attribute prediction model selection unit 185 selects the attribute prediction model having the smallest prediction error (that is, the maximum prediction accuracy) from the plurality of attribute prediction models. Instead, the attribute prediction model selection unit 185 may select the attribute prediction model having the highest evaluation from the plurality of attribute prediction models by using the evaluation result including conditions other than the prediction error as evaluation parameters. .. Here, the attribute prediction model selection unit 185 is based on the evaluation result of each prediction error of the second one or a plurality of attribute prediction models using the prediction value by another attribute prediction model as an explanatory variable as described above. , The second attribute prediction model used for predicting the attribute value of the second prediction target attribute may be further selected.

The prediction model generation device 150 can select an attribute prediction model individually optimized for each prediction target attribute by performing the processes S350 to S370 for each of the prediction target attributes of the prediction target. it can. As a result, the prediction model generation device 150 can further improve the prediction accuracy of each prediction target attribute.

FIG. 4 shows an example of the evaluation result of the attribute prediction model generated by the attribute prediction model generation unit 180 according to the present embodiment. The attribute prediction model generation unit 180 has attribute identification information (attribute ID), dimension reduction parameters, and prediction algorithm identification information (prediction algorithm ID) for each of one or a plurality of attribute prediction models generated for each prediction target attribute. ), Hyperparameters, prediction errors, and evaluation results including model evaluation.

The "attribute ID" is identification information of the prediction target attribute of the prediction target. The "dimensional contraction parameter" is a parameter such as the number of topics used by the dimensional contraction unit 125 and the dimensional contraction unit 170 for dimensional contraction. When there are a plurality of dimensional reduction parameters, the attribute prediction model generation unit 180 records the set value for each parameter. Dimensional contraction parameters can also be a kind of hyperparameters.

The "prediction algorithm ID" is identification information that identifies the prediction algorithm selected in the entry. "Hyperparameter" indicates the setting value for each hyperparameter in the selected prediction algorithm. “Forecast error” indicates the prediction error of the attribute prediction model generated using the dimension reduction parameters, prediction algorithms, and hyperparameters specified in the entry. "Model evaluation" indicates the evaluation result of the attribute prediction model generated corresponding to the entry. Here, when the attribute prediction model generation unit 180 evaluates the attribute prediction model based only on the prediction error, it is not necessary to separately provide a column for "model evaluation".

FIG. 5 shows an example of the selection result of the attribute prediction model generated by the attribute prediction model selection unit 185 according to the present embodiment. The attribute prediction model selection unit 185 selects an attribute prediction model to be used for predicting the attribute value for each prediction target attribute by using the evaluation result of each attribute prediction model by the attribute prediction model generation unit 180. The attribute prediction model selection unit 185 generates a selection result including an attribute ID, a dimension reduction parameter selection value, a selection prediction algorithm ID, and a hyperparameter selection value for each prediction target attribute of the prediction target.

The "attribute ID" is the identification information of the prediction target attribute corresponding to the entry. The "dimensional contraction parameter selection value" includes each set value of one or a plurality of dimension contraction parameters used for predicting the prediction target attribute corresponding to the entry. The “dimensional reduction parameter value” corresponds to the dimension reduction parameter of the selected attribute prediction model in FIG. The "selection prediction algorithm ID" is identification information that identifies the prediction algorithm used for predicting the prediction target attribute corresponding to the entry. The “selection prediction algorithm ID” corresponds to the prediction algorithm ID of the selected attribute prediction model in FIG. The "hyperparameter selection value" is a set value for each hyperparameter given to the prediction algorithm used for predicting the prediction target attribute corresponding to the entry. The "hyperparameter selection value" corresponds to the hyperparameter of the selected attribute prediction model in FIG.

The attribute prediction model selection unit 185 supplies the selection result of the attribute prediction model for each prediction target attribute to the attribute prediction unit 135, thereby supplying the dimension reduction parameters, prediction algorithms, and hyper parameters used for prediction for each prediction target attribute. It can be set for the attribute prediction unit 135. The attribute prediction model selection unit 185 may supply the prediction error and / or the evaluation result of the selected attribute prediction model to the attribute prediction unit 135.

FIG. 6 shows an attribute information acquisition flow in the attribute prediction device 110 according to the present embodiment. In S610, the attribute information acquisition unit 115 acquires the attribute information of the target person. In S620, the attribute information acquisition unit 115 writes the acquired attribute information to the attribute DB 122 via the attribute DB connection unit 120.

When the attribute DB 122 separately stores the known attribute value and the predicted value for the attribute to be written, the attribute information acquisition unit 115 may write the acquired attribute information to the attribute DB 122 as a known attribute value. Further, when the attribute DB 122 and the attribute DB 167 store a known attribute value in the same expression format as the predicted value for the attribute to be written, the attribute information acquisition unit 115 predicts the expression format of the raw data of the acquired attribute information. It is converted into the same expression format as the value and stored in the attribute DB 122.

FIG. 7 shows an attribute prediction flow in the attribute prediction device 110 according to the present embodiment. In step S710, the dimension contraction unit 125 contracts the dimension of the attribute DB 122 by using a topic model such as a latent Dirichlet allocation method or a stochastic latent semantic analysis method as an example, similarly to the sampling unit 160. Here, the dimension reduction unit 125 refers to the dimension reduction parameter selection value corresponding to the attribute prediction model selected for each prediction target attribute, and the dimension reduction parameter specified for at least one prediction target attribute. The attribute for dimension reduction according to the selected value is stored in the reduction DB 132. Further, the dimension reduction unit 125 copies an attribute used as an explanatory variable in at least one attribute prediction model selected by the attribute prediction model selection unit 185 among the plurality of attributes in the attribute DB 122 as an attribute of the reduction DB 132. To do.

The attribute prediction device 110 repeats the processing (S730 to S780) between S730 and S790 for each prediction target attribute (S720, S790).

In S730, the attribute prediction unit 135 predicts the attribute value of the prediction target attribute by using the attribute prediction model selected for the prediction target attribute by the attribute prediction model selection unit 185. The attribute prediction unit 135 acquires the attribute values of the attributes corresponding to the explanatory variables of the attribute prediction model included in the attribute information stored in the reduction DB 132 for each target person via the reduction DB connection unit 130. The attribute value of is input to the attribute prediction model to calculate the predicted value of the predicted target attribute. Here, the attribute prediction unit 135 describes the second attribute prediction model that predicts the prediction value of the second prediction target attribute, which is the prediction target, using the prediction value of the first prediction target attribute by the first attribute prediction model. As the attribute value of the first prediction target attribute that becomes a variable, the prediction value of the first prediction target attribute is input to the second attribute prediction model, and the second attribute prediction model is used for each of the plurality of target persons. Further predict the prediction target attribute.

Further, when the known attribute value and the predicted value for the first predicted target attribute are represented in the same expression format, or when one can be converted to the other, the attribute prediction unit 135 can be used for each of the plurality of target persons. The predicted value of the second prediction target attribute of the target person is predicted using the known attribute value on condition that the attribute value of the first prediction target attribute is known, and the attribute value of the first prediction target attribute is unknown. The predicted value of the second prediction target attribute of the target person may be predicted by using the attribute value of the first prediction target attribute on the condition that. As a result, the attribute prediction unit 135 predicts the second prediction target attribute with higher accuracy by using the known attribute value for the target person whose first prediction target attribute is known, which is the prediction target by the first attribute prediction model. At the same time, it is possible to make a prediction even for a target person whose first prediction target attribute is unknown.

In S740, the attribute prediction unit 135 stores the predicted value of the prediction target attribute in the attribute DB 122 via the attribute DB connection unit 120. Here, when the predicted value of the predicted target attribute can be calculated, the attribute prediction unit 135 stores the predicted value in the attribute DB 122 even if the known attribute value of the predicted target attribute is stored in the attribute DB 122. You may.

In S750, the attribute prediction value update unit 142 sets the prediction value of the prediction target attribute when the attribute value of the prediction target attribute is known for one or more of the plurality of target persons registered in the attribute DB 122. Judge whether or not the value deviates from the known attribute value by more than the standard. This criterion may be predetermined by the designer, system builder, user, and / or administrator of the system 100, and the attribute prediction value update unit 142 determines the variance or standard deviation of the known attribute values of the prediction target attribute. Etc. may be calculated by multiplying the above by a constant.

According to the fact that the predicted value of the predicted target attribute deviates from the known attribute value by more than the standard (Yes in S750), the attribute predicted value update unit 142 sets the predicted value of the predicted target attribute to the known attribute value in S760. May be updated based on. That is, when the known attribute value and the predicted value of the predicted target attribute have the same expression format, the attribute predicted value update unit 142 may rewrite the predicted value of the predicted target attribute to the known attribute value. When the known attribute value and the predicted value of the predicted target attribute have different expression formats, the attribute predicted value update unit 142 converts the expressed format of the known attribute value and rewrites the predicted value of the predicted target attribute. May be good. Here, the criteria used for determining whether or not there is a divergence may be preset by the designer, system builder, user, and / or administrator of the system 100.

As a result, the attribute prediction value update unit 142 uses the predicted value that is significantly wrong in other attribute prediction models that use the predicted target attribute as the explanatory variable for the target person whose attribute value of the predicted target attribute is known. It is possible to prevent the prediction of other prediction target attributes. For example, the attribute prediction value update unit 142 sets the prediction value to "unmarried" if it is known that the target person is "unmarried" even if the target person is mistakenly predicted to be "married". It can be updated and the prediction accuracy of other attributes predicted by using this attribute as an explanatory variable can be improved.

Further, the attribute value update unit 140 sets the attribute value of the prediction target attribute based on the prediction value of the prediction target attribute, provided that the prediction accuracy of the prediction value of the prediction target attribute is equal to or higher than the threshold value (Yes in S770). May be updated (S780). Here, when the attribute value update unit 140 takes the same expression format as the known attribute value and the predicted value of the predicted target attribute, the attribute predicted value updating unit 142 predicts the known attribute value of the predicted target attribute. It may be rewritten as. When the known attribute value of the predicted target attribute and the predicted value take different representation formats, the attribute value update unit 140 may convert the representation format of the predicted value and rewrite the attribute value of the predicted target attribute. Here, the threshold value used for determining the accuracy of the prediction may be preset by the designer, system builder, user, and / or administrator of the system 100.

As a result, the attribute value update unit 140 adapts to the situation such as when the acquisition time of the attribute value is old or when the situation of the target person changes even for the target person whose attribute value of the predicted target attribute is known. , The known attribute value can be replaced based on the predicted value as the predicted value of the predicted attribute becomes sufficiently accurate. As a result, as an example, the target person who started to purchase diapers was kept as the "childless" attribute, and the target person who started to purchase gasoline frequently was set as the "carless" attribute. It is possible to prevent the maintenance as it is.

Here, the attribute value update unit 140 is included in all the target persons whose predicted target attributes are known, all the target persons included in the learning attribute information, or the learning attribute information for the cross test, as the certainty of the prediction. AUC (Area Under the Curve) based on the predicted values of all subjects may be used. As an example, the attribute value update unit 140 arranges each target person whose predicted target attribute is known on the horizontal axis in descending order of the predicted value, and for each target person on the horizontal axis, all the target persons having a predicted value equal to or higher than that target person. The area below the ROC curve in the ROC curve (Receiver Operating Characteristic Curve) plotting the ratio of the subjects satisfying the predicted target attribute among the subjects is calculated as AUC. At this time, the attribute value updating unit 140 may normalize the maximum value that the AUC can take to 1 by normalizing the horizontal axis of the ROC curve to 1. Instead, the attribute value update unit 140 may use a parameter value such as (constant-prediction error) that increases as the prediction error decreases, as the certainty of the prediction.

FIG. 8 shows an example of the dependency of the predicted target attribute. The attribute prediction model A810 predicts the attribute value of the prediction target attribute c based on the attribute value of the attribute a and the attribute value of the attribute b. The attribute prediction model B820 predicts the attribute value of the prediction target attribute e based on the attribute value of the attribute c and the attribute value of the attribute d. The attribute prediction model C830 predicts the attribute value of the prediction target attribute b based on the attribute value of the attribute e and the attribute value of the attribute f.

In this example, the attribute prediction model A810 inputs the attribute value of the attribute b and outputs the attribute value of the attribute c, and the attribute prediction model B820 inputs the attribute value of the attribute c and outputs the attribute value of the attribute e. The attribute prediction model C830 inputs the attribute value of the attribute e and outputs the attribute value of the attribute b. Therefore, in the chain of explanatory variables and objective variables in a plurality of attribute prediction models, there is a cyclic dependency between two or more prediction target attributes such as attribute b → attribute c → attribute e → attribute b.

Here, in the attribute prediction model generation unit 180, the attribute prediction model selection unit 185, and / or the attribute prediction unit 135, for example, the attribute prediction model B820 uses a known attribute value of the attribute c as an explanatory variable, and the predicted value of the attribute c is If it is not used as an explanatory variable, the attribute prediction model B820 does not input the predicted value of the attribute c output by the attribute prediction model A810, so that the cyclic dependence between the predicted target attributes can be eliminated. That is, when the attribute prediction model generation unit 180, the attribute prediction model selection unit 185, and / or the attribute prediction unit 135 have a circular dependence of the input / output relationship to the attribute prediction model between two or more prediction target attributes. May generate or select an attribute prediction model to enter a known attribute value as the attribute value of the attribute included in the circular dependency in at least one attribute prediction model constituting the circular dependency.

However, when a known attribute value is input, the attribute prediction unit 135 cannot predict the attribute value of the prediction target attribute for a target person who does not have a known attribute value, or the prediction accuracy is lowered. .. Therefore, the attribute prediction unit 135 may have a configuration that allows a circular dependency between two or more prediction target attributes.

In this case, the attribute prediction unit 135 determines the certainty of prediction of other prediction target attributes used for each prediction of the two or more prediction target attributes according to the existence of a cyclic dependency between the two or more prediction target attributes. , And the prediction order of the two or more prediction target attributes may be determined based on at least one of the contributions of the other prediction target attributes to each of the two or more prediction target attributes. For example, in the example of this figure, if the order is attribute c, attribute b, and attribute e in descending order of prediction probability, the attribute prediction unit 135 inputs the attribute c having the highest prediction probability. The prediction order may be determined so that B820 is executed first. After that, the attribute prediction unit 135 may execute the attribute prediction model in which the attributes are input in descending order of the probability of prediction among the attributes that depend on the circulation. Instead, when the attribute prediction model B820 is executed, the attribute prediction unit 135 inputs the prediction values output by the attribute prediction model B820 and the attribute prediction model C830, and inputs the prediction values output by the attribute prediction model C830. After the attribute prediction model B820 such as the attribute prediction model A810, the attribute prediction model may be executed along the dependency direction between the prediction target attributes. As a result, the attribute prediction unit 135 predicts the predicted value of the predicted attribute of the dependent destination by preferentially using the predicted value that is expected to be closer to the original attribute value due to the high accuracy of the prediction. It is possible to improve the accuracy of the predicted value more quickly.

Further, for example, in the example of this figure, the attribute prediction unit 135 describes the contribution of the attribute c in the prediction of the prediction target attribute e, the contribution of the attribute b in the prediction of the prediction target attribute c, and the attribute e in the prediction of the prediction target attribute b. If the contribution degree decreases in this order, the attribute prediction unit 135 may determine the prediction order so that the attribute prediction model C830 having the attribute e having the lowest contribution degree as an input is executed first with the highest priority. Good. After that, the attribute prediction unit 135 may execute the attribute prediction model in which the attribute is input in ascending order of contribution in the prediction of the next prediction target attribute by the attribute prediction model among the attributes that depend on the circulation. Instead, when the attribute prediction model C830 is executed, the attribute prediction unit 135 inputs the prediction values output by the attribute prediction model C830 and the attribute prediction model A810 and the prediction values output by the attribute prediction model A810. After the attribute prediction model C830, such as the attribute prediction model B820, the attribute prediction model may be executed along the dependency direction between the prediction target attributes. As a result, the attribute prediction unit 135 preferentially uses the prediction value that has a low contribution to the prediction of the attribute prediction model in the subsequent stage and is unlikely to be affected by the circular dependency, and determines the prediction value of the prediction target attribute of the dependency. Predictions can be made, and the effects of circular dependencies of predicted values can be attenuated more quickly.

The attribute prediction unit 135 determines both the certainty of the prediction of the other prediction target attributes used for each prediction of the two or more prediction target attributes and the contribution of the other prediction target attributes to each of the two or more prediction target attributes. Based on this, the prediction order of two or more prediction target attributes may be determined, or the prediction order may be determined in consideration of other conditions. Here, the attribute prediction unit 135 inputs the attribute for the attribute included in the circular dependency when the prediction accuracy of the attribute used for the prediction is higher or the contribution of the attribute in the prediction is lower. Prioritize model execution. As an example, the attribute prediction unit 135 may determine the priority by the weighted sum of the certainty of the prediction (or the uncertainty of the prediction) of the attribute used for the prediction and the contribution of the attribute in the prediction.

As another example, the attribute prediction unit 135 has two or more prediction target attributes for each of the two or more prediction target attributes, based on the sum of the products of the prediction uncertainty of the other prediction target attributes and the contributions of the other prediction target attributes. The prediction order of the prediction values of the prediction target attributes may be determined. That is, for example, the attribute prediction unit 135 determines the uncertainty of each input attribute included in the circular dependency and each input for each of the plurality of attribute prediction models (attribute prediction model A810, attribute prediction model B820, attribute prediction model C830, etc.). The product of the contributions of the attributes is taken, and the attribute prediction model with a smaller sum of products value is executed with higher priority.

The attribute prediction unit 135 may use the value of a parameter whose value decreases as the certainty of the prediction increases, as the uncertainty of the prediction. For example, the attribute prediction unit 135 may use a value obtained by subtracting the certainty of the prediction based on the AUC normalized to a maximum of 1 from 1 as in (1-AUC).

As shown above, by making the prediction order between the prediction target attributes included in the circular dependence a more appropriate order, the attribute prediction unit 135 makes a prediction between two or more prediction target attributes calculated cyclically. The accuracy of the values can be increased faster and / or the predicted values can be converged faster. As a result, even when the execution cycle of each attribute prediction model becomes long due to the attribute DB 122 and the attribute DB 167 having many attributes, the predicted value can be optimized faster.

FIG. 9 shows an attribute addition flow in the system 100 according to the present embodiment. In S900, the attribute data acquisition unit 144 acquires known information indicating the known attribute value of the additional attribute to be added to the attribute database for at least a part of the plurality of target persons. For example, the terminal 112 can easily add an attribute to the attribute DB 122 so that an end user of the system 100 such as a marketer can easily predict an arbitrary characteristic of each target person. Provides a user interface for calculating predicted values of attributes. As an example, the terminal 112 adds an attribute according to the known information, sets a known attribute value, and responds to the drag and drop of the known information to the icon of the application providing the user interface. A user interface may be provided that performs at least part of the generation and selection of the attribute prediction model.

The known information acquired by the attribute data acquisition unit 144 may be information including a pair of personal identification information and known attribute values of one or a plurality of additional attributes for each of at least a part of a plurality of target persons. .. Further, when the additional attribute is a binary attribute indicating whether or not the target person has a specific characteristic (for example, "whether or not the product A is bought"), the known information is each target whose additional attribute is true. It may include at least one of a list of personally identifiable information of the person and a list of personally identifiable information of each subject whose additional attributes are false.

In S910, the attribute addition unit 146 adds an additional attribute to a plurality of attributes of the attribute DB 122. For example, the attribute addition unit 146 adds a column of additional attributes to the attribute DB 122 via the attribute DB connection unit 120, and adds the additional attribute in the attribute DB 122 based on the known attribute value of each target person included in the known information. Set the known attribute value to be stored. At this time, the attribute addition unit 146 may set the known attribute value indicated in the known information as the known attribute value of the additional attribute in the attribute DB 122, and the known attribute value indicated in the known information may be set as the attribute DB 122. It may be converted to the expression format in and set as a known attribute value of the additional attribute.

In S920, the prediction model generation device 150 in the system 100 generates one or a plurality of attribute prediction models with the additional attribute as the prediction target attribute. Here, the model update instruction unit 155 promptly responds to the addition of the additional attribute to the attribute DB 122 by the attribute data acquisition unit 144 without waiting for the elapse of a predetermined period, and promptly predicts the attribute of the additional attribute. May be instructed to generate. The process of generating one or more attribute prediction models with the additional attribute as the prediction target attribute may be the same as the process of S320 to S350 in the operation flow of FIG.

In S930, the prediction model generation device 150 in the system 100 selects an attribute prediction model used for predicting the attribute value of the prediction target attribute, which is an additional attribute, based on the prediction error of each of the plurality of attribute prediction models. The process of selecting the attribute prediction model used for predicting the attribute value of the additional attribute from the plurality of attribute prediction models may be the same as the process of S360 to S370 of FIG.

After the processing shown above, the attribute prediction device 110 in the system 100 may execute the attribute prediction flow shown in FIG. 7 to predict the attribute value of the additional attribute for each target person or all target persons. .. As a result, the system 100 learns an attribute prediction model that predicts the attribute value of the additional attribute by using the known attribute value of some target persons included in the known information as teacher data, and other target persons or all target persons. The predicted value of the additional attribute can be calculated for.

In the present embodiment, the attribute data acquisition unit 144 and the attribute addition unit 146 are provided in the attribute prediction device 110 and add additional attributes to the master attribute DB 122. Instead of this, the attribute data acquisition unit 144 and the attribute addition unit 146 may be provided in the prediction model generation device 150 and may adopt a form in which additional attributes are added to the attribute DB 167 which is a subset.

FIG. 10 shows an example of a computer 1900 in which a plurality of aspects of the present invention may be embodied in whole or in part. The program installed on the computer 1900 can cause the computer 1900 to function as an operation associated with the device according to an embodiment of the present invention or as one or more sections of the device, or the operation or the one or more. Sections can be run and / or computer 1900 can be run a process according to an embodiment of the invention or a stage of such process. Such a program may be run by the CPU 2000 to cause the computer 1900 to perform certain operations associated with some or all of the blocks of the flowcharts and block diagrams described herein.

The computer 1900 according to the present embodiment is connected to the host controller 2082 by the input / output controller 2084 and the CPU peripheral portion having the CPU 2000, the RAM 2020, the graphic controller 2075, and the display device 2080 which are interconnected by the host controller 2082. It includes an input / output unit having a communication interface 2030, a hard disk drive 2040, and a DVD drive 2060, and a legacy input / output unit having a ROM 2010, a flash memory drive 2050, and an input / output chip 2070 connected to the input / output controller 2084. ..

The host controller 2082 connects the RAM 2020 to the CPU 2000 and the graphic controller 2075 that access the RAM 2020 at a high transfer rate. The CPU 2000 operates based on the programs stored in the ROM 2010 and the RAM 2020, and controls each part. The graphic controller 2075 acquires image data generated on a frame buffer provided in the RAM 2020 by the CPU 2000 or the like, and displays the image data on the display device 2080. Instead, the graphic controller 2075 may internally include a frame buffer for storing image data generated by the CPU 2000 or the like.

The input / output controller 2084 connects the host controller 2082 to the communication interface 2030, the hard disk drive 2040, and the DVD drive 2060, which are relatively high-speed input / output devices. The communication interface 2030 communicates with other devices via a network by wire or wirelessly. In addition, the communication interface functions as hardware for communication. The hard disk drive 2040 stores programs and data used by the CPU 2000 in the computer 1900. The DVD drive 2060 reads a program or data from the DVD 2095 and provides it to the hard disk drive 2040 via the RAM 2020.

Further, the ROM 2010, the flash memory drive 2050, and the relatively low-speed input / output device of the input / output chip 2070 are connected to the input / output controller 2084. The ROM 2010 stores a boot program executed by the computer 1900 at startup and / or a program depending on the hardware of the computer 1900. The flash memory drive 2050 reads a program or data from the flash memory 2090 and provides it to the hard disk drive 2040 via the RAM 2020. The input / output chip 2070 connects the flash memory drive 2050 to the input / output controller 2084, and inputs / outputs various input / output devices via, for example, a parallel port, a serial port, a keyboard port, a mouse port, and the like. Connect to controller 2084.

The program provided to the hard disk drive 2040 via the RAM 2020 is stored in a recording medium such as a flash memory 2090, a DVD 2095, or an IC card and provided by the user. The program is read from the recording medium, installed on the hard disk drive 2040 in the computer 1900 via the RAM 2020, and executed in the CPU 2000. The information processing described in these programs is read by the computer 1900 and results in collaboration between the software and the various types of hardware resources described above. The device or method may be configured to perform manipulation or processing of information in accordance with the use of computer 1900.

As an example, when communicating between the computer 1900 and an external device or the like, the CPU 2000 executes a communication program loaded on the RAM 2020, and based on the processing content described in the communication program, a communication interface. Instruct 2030 to perform communication processing. Under the control of the CPU 2000, the communication interface 2030 reads the transmission data stored in the transmission buffer area or the like provided on the storage device such as the RAM 2020, the hard disk drive 2040, the flash memory 2090, or the DVD 2095, and transmits the transmission data to the network. Alternatively, the received data received from the network is written to a receive buffer area or the like provided on the storage device. As described above, the communication interface 2030 may transfer the transmitted / received data to / from the storage device by the DMA (direct memory access) method, and instead, the CPU 2000 may transfer the transfer source storage device or the communication interface 2030. The transmitted / received data may be transferred by reading the data from the data and writing the data to the communication interface 2030 or the storage device of the transfer destination.

Further, the CPU 2000 performs DMA all or necessary parts from files or databases stored in an external storage device such as a hard disk drive 2040, a DVD drive 2060 (DVD2095), and a flash memory drive 2050 (flash memory 2090). It is read into the RAM 2020 by transfer or the like, and various processes are performed on the data on the RAM 2020. Then, the CPU 2000 writes the processed data back to the external storage device by DMA transfer or the like. In such processing, the RAM 2020 can be regarded as temporarily holding the contents of the external storage device. Therefore, in the present embodiment, the RAM 2020 and the external storage device are collectively referred to as a memory, a storage unit, a storage device, or the like.

Various information such as various programs, data, tables, and databases in the present embodiment are stored in such a storage device and are subject to information processing. The CPU 2000 can also hold a part of the RAM 2020 in the cache memory and read / write on the cache memory. Even in such a form, the cache memory plays a part of the function of the RAM 2020. Therefore, in the present embodiment, the cache memory is also included in the RAM 2020, the memory, and / or the storage device, unless otherwise indicated. To do.

In addition, the CPU 2000 includes various operations, information processing, condition determination, information retrieval / replacement, and the like specified in the instruction sequence of the program for the data read from the RAM 2020. Is performed, and the data is written back to the RAM 2020. For example, when the CPU 2000 determines a condition, whether or not various variables shown in the present embodiment satisfy conditions such as large, small, above, below, and equal to other variables or constants. If the condition is satisfied (or if it is not satisfied), it branches to a different instruction sequence or calls a subroutine.

In addition, the CPU 2000 can search for information stored in a file in the storage device, a database, or the like. For example, when a plurality of entries in which the attribute value of the second attribute is associated with the attribute value of the first attribute are stored in the storage device, the CPU 2000 describes the plurality of entries stored in the storage device. By searching for an entry in which the attribute value of the first attribute matches the specified condition and reading the attribute value of the second attribute stored in that entry, it is associated with the first attribute that satisfies the predetermined condition. The attribute value of the second attribute obtained can be obtained.

Further, when a plurality of elements are listed in the description of the embodiment, elements other than the listed elements may be used. For example, when it is described that "X executes Y using A, B and C", X may execute Y using D in addition to A, B and C.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments. It will be apparent to those skilled in the art that various changes or improvements can be made to the above embodiments. It is clear from the description of the claims that such modified or improved forms may also be included in the technical scope of the present invention.

The order of execution of operations, procedures, steps, steps, etc. in the devices, systems, programs, and methods shown in the claims, specification, and drawings is particularly "before" and "prior to". It should be noted that it can be realized in any order unless the output of the previous process is used in the subsequent process. Even if the scope of claims, the specification, and the operation flow in the drawings are explained using "first", "next", etc. for convenience, it means that it is essential to carry out in this order. It's not a thing.

100 System 110 Attribute prediction device 112 Terminal 115 Attribute information acquisition unit 120 Attribute DB Connection unit 122 Attribute DB
125 dimensional contraction 130 contraction DB connection 132 contraction DB
135 Attribute prediction unit 140 Attribute value update unit 142 Attribute prediction value update unit 144 Attribute data acquisition unit 146 Attribute addition unit 150 Prediction model generator 152 Terminal 155 Model update instruction unit 160 Sampling unit 165 Attribute DB connection unit 167 Attribute DB
170 dimensional contraction 175 contraction DB connection 177 contraction DB
180 Attribute prediction model generation unit 185 Attribute prediction model selection unit 190 Recommendation processing device 192 Terminal 810 Attribute prediction model A
820 Attribute Prediction Model B
830 Attribute Prediction Model C
1900 computer 2000 CPU
2010 ROM
2020 RAM
2030 Communication Interface 2040 Hard Disk Drive 2050 Flash Memory Drive 2060 DVD Drive 2070 I / O Chip 2075 Graphic Controller 2080 Display 2082 Host Controller 2084 I / O Controller 2090 Flash Memory 2095 DVD

Claims (20)

For each of multiple target persons, an attribute database connection part connected to an attribute database for storing multiple attribute values corresponding to multiple attributes, and
First, using the attribute database, the attribute value of the first prediction target attribute, which is the prediction target, is predicted based on the attribute value of at least one attribute other than the first prediction target attribute among the plurality of attributes. Attribute prediction model generator that generates multiple attribute prediction models of
A device including an attribute prediction model selection unit that selects a first attribute prediction model to be used for predicting the attribute value of the first prediction target attribute based on the prediction error of each of the first plurality of attribute prediction models. Further, a sampling unit for sampling a part of the plurality of target persons from the attribute database is provided.
The device according to claim 1, wherein the attribute prediction model generation unit generates the first plurality of attribute prediction models by using the attribute values associated with a part of the sampled target persons. A dimension reduction unit that reduces the dimensions of the plurality of attributes based on the plurality of attribute values of each of the plurality of target persons stored in the attribute database is further provided.
The device according to claim 1 or 2, wherein the attribute prediction model generation unit predicts the attribute value of the first prediction target attribute from the attribute value of at least one of the plurality of dimension-reduced attributes. The attribute prediction model selection unit has different attributes as the first attribute prediction model used for predicting the attribute value of the first prediction target attribute and the second attribute prediction model used for predicting the attribute value of the second prediction target attribute. The device according to any one of claims 1 to 3, wherein a prediction model can be selected. The device according to any one of claims 1 to 4, wherein the attribute prediction model generation unit learns learnable parameters in each of the first plurality of attribute prediction models. The device according to any one of claims 1 to 5, wherein each of the first plurality of attribute prediction models is different from other attribute prediction models in at least one of the hyperparameters and prediction algorithms that are not updated by learning. The apparatus according to any one of claims 1 to 6, further comprising an attribute prediction unit that predicts the attribute value of the first prediction target attribute for each of the plurality of target persons using the first attribute prediction model. .. When the attribute value of the first prediction target attribute is known for one of the plurality of target persons, the predicted value of the first prediction target attribute deviates from the known attribute value by a reference or more. The apparatus according to claim 7, further comprising an attribute predicted value updating unit that updates the predicted value of the first predicted target attribute based on the known attribute value. The attribute prediction model generation unit generates a second plurality of attribute prediction models that predict the attribute value of the second prediction target attribute, which is the prediction target, using the prediction value of the first prediction target attribute.
The attribute prediction model selection unit further selects a second attribute prediction model used for predicting the attribute value of the second prediction target attribute based on the prediction error of each of the second plurality of attribute prediction models.
The device according to claim 7 or 8, wherein the attribute prediction unit further predicts the second prediction target attribute for each of the plurality of target persons by using the second attribute prediction model.
The device according to claim 7 or 8. The attribute prediction unit predicts the second prediction target attribute of the target person using the known attribute value on condition that the attribute value of the first prediction target attribute is known for each of the plurality of target persons. A request for predicting a value and predicting the predicted value of the second prediction target attribute of the target person using the attribute value of the first prediction target attribute on condition that the attribute value of the first prediction target attribute is unknown. Item 9. The device according to item 9. Attribute value update that updates the attribute value of the first prediction target attribute based on the prediction value of the first prediction target attribute, provided that the prediction accuracy of the prediction value of the first prediction target attribute is equal to or greater than the threshold value. The device according to claim 7, further comprising a unit. Depending on the existence of a circular dependency between the two or more prediction target attributes, the attribute prediction unit determines the certainty of prediction of the other prediction target attributes used for each prediction of the two or more prediction target attributes, and According to any one of claims 7 to 11, which determines the prediction order of the two or more prediction target attributes based on at least one of the contributions of the other prediction target attributes to each of the two or more prediction target attributes. The device described. For each of the two or more prediction target attributes, the attribute prediction unit has the two or more prediction target attributes based on the sum of the uncertainties of prediction of the other prediction target attributes and the contribution of the other prediction target attributes. The apparatus according to claim 12, wherein the prediction order of the predicted values of the above is determined. The device according to any one of claims 7 to 13, wherein the first prediction target attribute is a preference attribute indicating a target person's preference for a product or service associated with the first prediction target attribute. 14. The claim 14 further includes a recommendation processing unit that selects whether or not to recommend a product or service associated with the first prediction target attribute to the target person based on the attribute value of the first prediction target attribute. Equipment. The apparatus according to any one of claims 1 to 15, further comprising a model update instruction unit for instructing the update of the first attribute prediction model according to the elapse of a predetermined period. A known information acquisition unit that acquires known information indicating known attribute values of additional attributes to be added to the attribute database for at least a part of the plurality of target persons.
An attribute addition part that adds the additional attribute to the plurality of attributes in the attribute database,
Further prepare
The attribute prediction model generation unit generates a plurality of attribute prediction models using the additional attribute as a prediction target attribute, and generates a plurality of attribute prediction models.
The attribute prediction model selection unit selects any one of claims 1 to 16 to select an attribute prediction model to be used for predicting the attribute value of the prediction target attribute based on the prediction error of each of the plurality of attribute prediction models. The device described in. The device according to claim 17, wherein the known information indicates the presence or absence of the additional attribute for each of at least a part of the plurality of subjects. The computer uses an attribute database for storing a plurality of attribute values corresponding to a plurality of attributes for each of the plurality of target persons, and sets the attribute value of the first prediction target attribute, which is the prediction target, to the plurality of attributes. Of these, an attribute prediction model generation stage that generates a first plurality of attribute prediction models that predict each based on the attribute values of at least one attribute other than the first prediction target attribute.
An attribute prediction model selection step in which the computer selects a first attribute prediction model to be used for predicting the attribute value of the first prediction target attribute based on the prediction error of each of the first plurality of attribute prediction models. How to prepare. Executed by a computer, said computer,
For each of multiple target persons, an attribute database connection part connected to an attribute database for storing multiple attribute values corresponding to multiple attributes, and
First, using the attribute database, the attribute value of the first prediction target attribute, which is the prediction target, is predicted based on the attribute value of at least one attribute other than the first prediction target attribute among the plurality of attributes. Attribute prediction model generator that generates multiple attribute prediction models of
A program that functions as an attribute prediction model selection unit that selects a first attribute prediction model to be used for predicting the attribute value of the first prediction target attribute based on the prediction error of each of the first plurality of attribute prediction models. ..

Images