JP6899805B2 - Characteristic estimation device, characteristic estimation method, characteristic estimation program, etc. - Google Patents

Description

The present invention relates to technical fields such as methods for estimating user characteristics.

Conventionally, in a communication network such as the Internet, many users are classified (segmented) into any of a plurality of types of characteristics, and information is provided according to the characteristics of the classified users. For example, in the field of marketing, it is known that it is effective to appeal to users who have characteristics called innovators as opinion leaders. As a method of estimating the characteristics of a user, there are a method of estimating by conducting a questionnaire to the user, a method of estimating by analyzing the access history of information by the user, and the like. Patent Document 1 discloses a method for estimating user characteristics. Specifically, the method of Patent Document 1 uses the characteristics of the teacher user as the objective variable, which is specified based on the result of the questionnaire to the teacher user, and the behavior of the teacher user generated based on the access history of the teacher user. A characteristic estimation model using data as an explanatory variable is generated, and the characteristics of the estimated user are estimated from this characteristic estimation model and the behavior data of the estimated user.

Japanese Unexamined Patent Publication No. 2009-289068

However, in the technique of Patent Document 1, the relationship with the hypothesis that is the basis of the question of the questionnaire used for the characteristic discrimination is unclear, and there is no idea of evaluating the validity of the hypothesis for discriminating the characteristic.

Therefore, the present invention has been made in view of the above points and the like, and provides a characteristic estimation device, a characteristic estimation method, a characteristic estimation program, etc. that enable empirical characteristic estimation based on the results of a questionnaire of a subject. The task is to do.

In order to solve the above problem, the invention according to claim 1 is one of a plurality of characteristics set in advance for the first user based on the answers of the plurality of first users who responded to the questionnaire. A classification means for classifying into the characteristics, a determination means for determining whether or not the behavior history of the first user meets the criteria set in relation to the characteristics based on the hypothesis common to the questionnaire, and the plurality. the first user of the action history of the first user is determined to prior meet Kimoto quasi as a teacher user, the action information extracted from the behavior history of the teacher user as explanatory variables, the characteristics of the teacher user The second model uses the generation means for generating the trained model learned from the training data as the objective variable, the behavior information extracted from the behavior history of the second user different from the first user, and the trained model. It is characterized by including an estimation means for estimating user characteristics.

According to the present invention, empirical characteristic estimation can be performed based on the results of a questionnaire of the subject.

In the invention according to claim 2, when the action history of the second user is updated in the characteristic estimation device according to claim 1, the estimation means uses the action information extracted from the updated action history. It is characterized in that it is updated by re-estimating the characteristics of the second user using the trained model and the trained model.

According to the present invention, the characteristics estimated for the second user can be made to correspond to the change in the behavior history of the second user.

The invention according to claim 3 is the characteristic estimated according to a predetermined classification criterion in the trained model when the behavior history of the second user is updated in the characteristic estimation device according to claim 1 or 2. Change to change the predetermined classification criteria so as to satisfy the condition that the appearance rate of the second user belonging to the above is equal and the condition that the second user belonging to the estimated characteristic has a predetermined degree of agreement. It is characterized by further providing means.

According to the present invention, it is possible to improve the estimation accuracy by the trained model by following the change in the behavior history of the second user whose characteristics have been estimated.

In the invention according to claim 4, in the characteristic estimation device according to any one of claims 1 to 3, the action history of the teacher user is updated by updating the action history of the first user. In this case, the behavior information extracted from the updated behavior history is used as an explanatory variable, and the learning data using the characteristics of the teacher user as the objective variable is further provided as an update means for updating the trained model. To do.

According to the present invention, it is possible to improve the estimation accuracy by the trained model by following the change in the behavior history of the teacher user.

According to the fifth aspect of the present invention, in the characteristic estimation device according to the fourth aspect, when the action history of the teacher user is updated by updating the action history of the first user, the determination means may be used. It is characterized in that it is determined whether or not the updated behavior history of the teacher user meets the criteria set in relation to the characteristic.

According to the present invention, it is possible to improve the estimation accuracy by the trained model including the empirical judgment by following the change in the behavior history of the teacher user.

The invention described in claim 6 is the characteristic estimation apparatus according to any one of claims 1 to 5, before SL generating means, first to the action history is judged that the not satisfied Kimoto quasi It is characterized in that the user is not adopted as a teacher user.

According to the present invention, the demonstrability and objectivity can be ensured and maintained from the stage of the teacher data which is the base point of the extension set specified by estimation.

The invention according to claim 7 is the updated behavior when the behavior history of the teacher user is updated by updating the behavior history of the first user in the characteristic estimation device according to claim 2. An update means for updating the trained model using the training data with the behavior information extracted from the history as the explanatory variable and the characteristics of the teacher user as the objective variable is further provided, and the update interval of the trained model is the above. It is characterized in that it is longer than the update interval of the characteristics of the second user.

According to the present invention, it is possible to minimize the update of the trained model and reduce the load on the device.

The invention according to claim 11 is a characteristic estimation method executed by a computer, in which the first user is preset with a plurality of characteristics based on the answers of the plurality of first users who have answered the questionnaire. The step of classifying into one of the characteristics, the step of determining whether or not the behavior history of the first user meets the criteria set in relation to the characteristics based on the hypothesis common to the questionnaire, and the above. Among the plurality of first users, the first user whose behavior history is determined to satisfy the criteria is used as the teacher user, the behavior information extracted from the behavior history of the teacher user is used as an explanatory variable, and the characteristics of the teacher user are set. The second user uses the step of generating a trained model learned from the training data used as the objective variable, the behavior information extracted from the behavior history of the second user different from the first user, and the trained model. It is characterized by including a step of estimating the characteristics of.

The invention according to claim 12 is a step of classifying the first user into one of a plurality of preset characteristics based on the answers of the plurality of first users who responded to the questionnaire. The step of determining whether or not the behavior history of the first user satisfies the criteria set in relation to the characteristics based on the hypothesis common to the questionnaire, and the behavior history of the plurality of first users is said. A trained model learned by learning data with the first user determined to meet the criteria as the teacher user, the behavior information extracted from the behavior history of the teacher user as the explanatory variable, and the characteristics of the teacher user as the objective variable. And a step of estimating the characteristics of the second user using the behavior information extracted from the behavior history of the second user different from the first user and the learned model are executed on the computer. It is characterized by letting it.

According to the present invention, empirical characteristic estimation can be performed based on the results of a questionnaire of a subject.

It is a flow chart which shows the flow of the basic operation of a characteristic estimation apparatus. It is a conceptual diagram of the basic operation of a characteristic estimation device. It is a figure which shows an example of the outline structure of the communication system S. (A) is a diagram showing a hardware configuration example of the characteristic estimation server 3, and (B) is a diagram showing a functional block example of the control unit 31 of the characteristic estimation server 3. It is a figure which shows an example of the contents of the user information database 301. As an example of the trained model, it is a conceptual diagram showing an example of conditional branching in a decision tree. It is a flowchart which shows an example of the trained model generation processing executed by the system control unit 33. It is a flowchart which shows an example of the characteristic estimation processing executed by the system control unit 33. It is a flowchart which shows an example of the characteristic update process executed by the system control unit 33. It is a flowchart which shows an example of the trained model update process executed by the system control unit 33. It is a conceptual diagram which shows the improvement content when the questionnaire option is specified as the improvement item of the questionnaire.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following embodiment, the first user is referred to as a target user, and the second user is referred to as an estimated user. The target user is a target person of the questionnaire, and includes a teacher user who meets the criteria described later. The estimated user is a user different from the target user, and the characteristic is estimated from the trained model generated by using the teacher data of the teacher user. The target user and the estimated user are identified by the user ID (identification information) given to each user. In the following description, the target user and the estimated user may be collectively referred to as a user.

[1. Basic operation of the characteristic estimation device]
First, the basic operation of the characteristic estimation device according to the present invention will be described with reference to FIG. FIG. 1 is a conceptual diagram showing the basic operation of the characteristic estimation device, and FIG. 2 is a conceptual diagram of the basic operation of the characteristic estimation device.

As shown in FIGS. 1 and 2, first, the characteristic estimation device (classification means) is based on the respective answers (questionnaire results) of the plurality of target users who answered the questionnaire (questionnaire survey), and the plurality of target users. Is classified (segmented) into one of a plurality of preset characteristics (Phase F1). User characteristics include hypotheses based on behavioral characteristics and market appearance distribution, and characteristic classification based on questionnaire results can be said to be hypothesis verification based on the subjective and conscious judgment of the target user. Examples of user characteristics include, but are not limited to, multiple characteristics in line with innovator theory (innovators, early adopters, early majority, late majority, and lagard), for example, in the user's life. It is possible to adopt characteristics that take values into consideration.

Next, the characteristic estimation device (judgment means) determines for each target user whether or not the behavior history of the target user who responded to the questionnaire meets a predetermined standard (objective standard), and acts among the target users. The target user whose history is determined to satisfy the above criteria is specified as a teacher user (Phase F2). Here, the action history represents an objective fact (event) based on the actual action of the target user. The set of the behavior information extracted from the behavior history of the teacher user and the characteristics of the teacher user is used as teacher data in generating the trained model described later. The above criteria are, for example, set in relation to characteristics based on a hypothesis common to the questionnaire, and include criteria related to the questionnaire. Examples of the above criteria include whether a new product is pre-ordered, whether a trending product is purchased or searched, and the like. In the process of creating teacher data, by using the behavior history of the target user who answered the questionnaire (that is, data based on objective facts), the incompleteness and uncertainty that can occur by relying on the consciousness of the target user can be determined. Can be complemented. For example, the behavior history of the target user can be used to describe the uncertainty based on the phenomenon that the accuracy of detailed data that relies on the memory of the target user, such as "what month and day the website A was visited," remains uncertain. Can be complemented. As a result, the amount of information that can be used while the accuracy is guaranteed increases, and the accuracy of teacher data can be expected to improve. In addition, the standard is a standard that can classify the characteristics by verifying the behavior history only by the standard.

Whether or not the behavior history meets the above criteria is determined, for example, by setting an inductive hypothesis verification item group (verification item group based on a hypothesis model common to the questionnaire) for each characteristic (a group of verification items based on a hypothesis model common to the questionnaire) that derives highly probable characteristics from individual events. Hypothesis verification items may overlap between characteristics), and it may be performed by determining whether or not the corresponding rate for this item group exceeds a predetermined ratio. Here, the corresponding rate for the hypothesis verification item group is calculated based on the behavior information (actual data such as purchase data) extracted from the behavior history. Such a judgment can be said to be a hypothesis verification based on the objective facts of the target user. Then, the target user whose corresponding rate exceeds a predetermined rate is specified as a teacher user. On the other hand, the target user whose corresponding rate does not exceed the predetermined rate is not adopted as the teacher user. If the applicable rate does not exceed the specified rate, in addition to the case where the subjective evaluation by the questionnaire and the objective evaluation are different (that is, when the hypothesis is not met), the purchase data of the target user, etc. This applies when the actual data is not sufficient and cannot be proved. That is, even when it is difficult to prove, the target user is not adopted as the teacher user, and the pair of the behavior information and the characteristic of the target user is excluded from the teacher data. By these processes, the demonstrability and objectivity can be secured and maintained from the stage of the teacher data which is the base point of the extended set specified by the estimation, and the estimation process proceeds in a state based on the demonstrability and objectivity at the time of the later estimation. be able to.

Next, the characteristic estimation device (generation means) uses the target user determined to meet the criteria as the teacher user, the behavior information extracted from the behavior history of the teacher user as the explanatory variable, and the characteristic of the teacher user as the objective variable. A trained model trained from the trained data is generated (Phase F3). That is, the characteristic estimation device generates a trained model in which the behavior information and characteristics of the teacher user are learned as teacher data. This enables hypothesis verification based on the subjective judgment of the target user and empirical characteristic estimation based on hypothesis verification based on the objective facts of the target user.

Next, the characteristic estimation device (estimating means) estimates (predicts) the characteristics of the estimated user using the behavior information extracted from the behavior history of the estimated user and the learned model (Phase F4). That is, the behavior information extracted from the behavior history of the estimated user is input to the trained model, and the characteristics of the estimated user are output. Since the trained model is modeled using, for example, each of a plurality of characteristics as an objective variable and one behavioral information as an explanatory variable, the degree of fit of the plurality of objective variables to each segment is calculated for each estimated user. Among them, the segment for which the maximum fit is calculated is adopted to determine the characteristics of the estimated user. For example, in FIG. 6, which will be described later, the characteristic 01 of the segment Se1 for which the maximum fit degree is calculated is determined as the characteristic of the estimated user.

[2. Overview of communication system S Configuration and functions]
Next, with reference to FIG. 3, the outline configuration of the communication system S to which the present invention is applied will be described. FIG. 3 is a diagram showing an example of an outline configuration of the communication system S.

As shown in FIG. 3, the communication system S includes a questionnaire management server (research server) 1, an action history management server 2, a characteristic estimation server 3, and the like. The questionnaire management server 1 and the action history management server 2 can communicate with the characteristic estimation server 3, for example, via an intranet, respectively. The questionnaire management server 1, the action history management server 2, and the characteristic estimation server 3 may be configured by one server.

(2-1. Questionnaire management server 1)
The questionnaire management server 1 is a server that manages the results of questionnaires conducted on the target users. For example, the questionnaire management server 1 transmits question information related to the questionnaire to each user terminal UTm (m = 1, 2, 3 ...) Of a plurality of target users via a communication network NW such as the Internet, and conducts a questionnaire. The answer information according to the above is received from each user terminal UTm, and the answer indicated by each answer information is stored as a questionnaire result in the questionnaire database (DB) 101 in association with the user ID of each target user. The user terminal UTm includes one used by the target user and one used by the estimated user. Examples of the user terminal UTm include personal computers, smartphones, tablets, game machines, and the like.

The question information and the answer information related to the questionnaire may be sent to the user terminal UTm of the target user by e-mail, or may be sent to the user terminal UTm in response to a request from the browser or application of the user terminal UTm of the target user. May be good. In the question information related to the questionnaire, a plurality of questions (questionnaire questions (questions)) for determining the characteristics of the target user and answer options (questionnaire options, for example, five) for each question are described. In the answer information related to the questionnaire, the answer selected for each question is described. The questionnaire management server 1 may have a function of a classification means of the characteristic estimation device according to the present invention.

(2-2. Behavior history management server 2)
The action history management server 2 is a server that manages a history of actions (behavior history) performed by a user online or offline. For example, the action history management server 2 is a server SAo (o = 1,2,3 ...) that can be accessed from the user terminal UTm or the processing terminal STn (n = 1,2,3 ...) via the communication network NW. The action history of each user (target user and estimated user) is collected from (), and each action history is stored in the action history database 201 in association with the user ID of each user. As an example of the processing terminal STn, for example, a payment card installed in a store and possessed by the user (for example, an IC card capable of electronic money payment or credit card payment) or a user terminal UTm (for example, electronic money payment or the like) A payment processing terminal capable of communicating with a payment processing application (such as a smartphone equipped with a payment application capable of credit card payment) can be mentioned. Further, as an example of the processing terminal STn, there is an entrance processing terminal installed at a venue such as an exhibition or an event of a product or the like.

Examples of server SAo include shopping server, free market server, auction server, travel reservation server, TV viewing analysis server, payment processing server, video / music streaming server, SNS (Social Networking Service) server, ticket issuance management server, and actual server. An action tracking server and the like can be mentioned. Here, if the user's behavior history can be widely collected from more servers SAo, it is possible to improve the estimation accuracy using the trained model. Therefore, it is desirable that the user ID that identifies the user is configured to be common or associated with more servers SAo.

The action history includes actions performed online by the user via the communication network NW (for example, searching for a product or service, browsing a product or service, paying (purchasing) a product or service, posting a product or service, etc.). Etc.), as well as the history of actions that the user went to the actual store and performed offline (for example, payment for products and services), and the user went to an exhibition or event of the product etc. to enter. It is good to include the history of actions. Examples of behavior history are search history of goods or services, browsing history of goods or services, bookmark registration history of goods or services (favorite registration), reservation history of goods or services (purchase reservation history in the case of goods), goods or Payment history for services (purchase history in the case of products), posting history of comments (word-of-mouth) or tweets about products or services, playback history of videos or songs as products, and to venues such as product exhibitions and events The admission history and the like can be mentioned, and which action history is used by the characteristic estimation server 3 is arbitrarily set by, for example, an administrator or the like.

The search history and browsing history are stored in, for example, a shopping server, a flea market server, an auction server, and a travel reservation server. The search history includes the date and time when the information of the product or service was published (that is, the date and time when it became searchable by the user), the search date and time of the product or service, the search condition (keyword used for the search), and the product or the product to be searched. Service information (eg, ID, name, category, price) and the like are included. The browsing history is stored in, for example, a shopping server, a flea market server, an auction server, and a travel reservation server. The browsing history includes the publication date and time of the product or service information (that is, the date and time when the product or service became available for viewing), the browsing date and time of the product or service, and the information of the product or service to be browsed (for example, ID, name, category). , Price) etc. are included.

The browsing history may be the browsing history of the product or service displayed on the TV screen by receiving the radio wave of the TV broadcast. In this case, for example, the audio data acquired by the microphone mounted on the user terminal UTm (for example, a smartphone) is acquired by the television viewing analysis server. Then, the TV viewing analysis server collates the acquired audio data with the original (master) audio data to determine at what time the user was watching the video, thereby determining the viewing date and time of the product or service, and the viewing date and time of the service. Identify the browsed product or service information and store it as a browsing history. Alternatively, when information prompting the user to select a selection button is displayed on the TV screen and the selection button displayed on the user terminal UTm (for example, a smartphone) is pressed by the user, the pressing signal is sent to the television viewing analysis server. Obtained by. Then, the TV viewing analysis server determines the viewing date and time of the product or service and the information of the viewed product or service by determining which time the user was watching the video from the acquisition date and time of the acquired pressing signal. Identify and store as browsing history.

The bookmark registration history is stored in, for example, a shopping server, a flea market server, an auction server, and a travel reservation server. The bookmark registration history includes the publication date and time of the product or service information, the bookmark registration date and time of the product or service, the information of the product or service to be bookmarked, and the like. The reservation history is stored in, for example, a shopping server, a flea market server, an auction server, and a travel reservation server. The reservation history includes the publication date and time of the product or service information, the reservation date and time of the product or service, the information of the product or service to be reserved, and the like. The payment history is stored in, for example, a shopping server, a flea market server, an auction server, a travel reservation server, and a payment processing server. The payment history includes the publication date and time of the information on the product or service, the payment date and time for the product or service, the information on the product or service to be paid, and the like.

The posting history is stored in, for example, a shopping server, a flea market server, an auction server, a travel reservation server, a video / music streaming server, and an SNS server. The posting history includes the publication date and time of information on the product or service, the posting date and time of comments or tweets on the product or service, the content of the posting, information on the product or service to be posted, and the like. The posting history should include the number of followers of the user who posted. The playback history is stored in, for example, a video / music streaming server. The playback history includes the release date and time of the product (video or music) information, the playback date and time of the product, the information of the product to be played, and the like.

The admission history is stored in, for example, the ticket issuance management server. The admission history includes information on the date and time of admission to the venue such as an exhibition or event of the product, and information on the product to be exhibited. For example, when the ticket issued by the ticket issuance management server is detected by the processing terminal STn installed at the venue, the issued ticket number and the like are acquired by the ticket issuance management server. Then, the ticket issuance management server determines the admission date and time from the acquisition time of the acquired ticket number or the like, identifies the information of the product to be exhibited from the ticket number or the like, and stores it as the admission history.

The admission history may be stored in the execution behavior tracking server. Position information detected by the user terminal UTm by the GPS (Global Positioning System) function and the GPS function (
When an application that acquires (latitude and longitude) and sends it to the execution motion tracking server is installed, the acquired position information is transmitted to the execution motion tracking server, for example, at predetermined time intervals. Then, the execution motion tracking server collates the location information and the reception date and time received from the user terminal UTm with the location information and the implementation date and time of the venue such as an exhibition or event described in the list prepared in advance. When it is determined that the user has actually visited the venue, the admission history including the date and time of admission to the venue and information on the products to be exhibited is stored.

(2-3. Characteristic estimation server 3)
The characteristic estimation server 3 is a server that functions as a characteristic estimation device according to the present invention alone or together with the questionnaire management server 1. FIG. 4A is a diagram showing a hardware configuration example of the characteristic estimation server 3, and FIG. 4B is a diagram showing a functional block example of the system control unit 33 of the characteristic estimation server 3. Although not shown, an operator's terminal can be connected to the characteristic estimation server 3.

As shown in FIG. 4A, the characteristic estimation server 3 includes a communication unit 31, a storage unit 32, a system control unit 33, and the like, and these components are connected to the system bus 34. The communication unit 31 is connected to, for example, an intranet, and controls the communication state between the questionnaire management server 1 and the action history management server 2.

The storage unit 32 is composed of, for example, a hard disk drive or the like, and stores various programs such as the OS and the characteristic estimation program of the present invention. The characteristic estimation program may be downloaded from a predetermined server to the characteristic estimation server 3, or recorded on a recording medium such as a CD, DVD, or BD (recorded readable by a computer) from the recording medium. It may be read and stored in the storage unit 32.

Further, a user information database 301 and the like are constructed in the storage unit 32. FIG. 5 is a diagram showing an example of the contents of the user information database 301. In the user information database 301 shown in FIG. 5, for example, the user ID, user name, user type, characteristics, and the like of each user whose account has been issued as a user member of a predetermined site are stored in association with each other. Here, the user type indicates whether or not the user is a teacher user (in the figure, "1" indicates a teacher user). As the user's characteristic, any one of the five characteristics 01 to 05 is given. The user whose characteristic is blank in the user information database 301 is a user whose characteristic has not been estimated yet.

The system control unit 33 (an example of a computer in the present invention) includes a CPU (Central Processing Unit) 33a, a ROM (Read Only Memory) 33b, and a RAM (Random Access).
Memory) 33c and the like are provided, and a characteristic estimation program and the like are executed on the OS. By executing the characteristic estimation program, the system control unit 33 executes the characteristic estimation program, and as shown in FIG. 4B, the target user classification unit 331, the teacher user identification unit 332, the model generation unit 333, the characteristic estimation unit 334, and the model update unit. It functions as 335, classification standard changing unit 336, and the like. The target user classification unit 331 is an example of the classification means in the present invention. The teacher user identification unit 332 is an example of the determination means in the present invention. The model generation unit 333 is an example of the generation means in the present invention. The characteristic estimation unit 334 is an example of the estimation means in the present invention. The model update unit 335 is an example of the update means in the present invention. The classification standard changing unit 336 is an example of the changing means in the present invention.

(2-3-1. Target user classification unit 331)
First, the target user classification unit 331 obtains the results of a questionnaire for each of a plurality of target users (that is, some users among all users) determined from all the users to whom the account has been issued. Based on the answers of the plurality of target users, the plurality of target users are classified into one of a plurality of preset characteristics. Various classification methods can be considered. For example, the answer score table data that describes the correspondence between the answer options for the question and the score (the score that is the basis of the score), and the total score (score) and the characteristics. Characteristic discrimination condition data (in other words, segment definition data) that describes the correspondence may be used.

As an example, in the answer score table data, option 1 (applicable) has +2 points, option 2 (somewhat applicable) has +1 point, option 3 (neither can be said) has 0 points, and option 4 (not very applicable). ) Is associated with -1 point, option 5 (not applicable) is associated with -2 points, and so on. Further, as an example, in the characteristic discrimination condition data, the characteristic C01 (for example, an innovator) has a total score of 12 points or more, the characteristic C02 (for example, an early adopter) has a total score of 5 to 11, and the total score of 2 to 4 has a characteristic C02 (for example, an early adopter). The characteristics C03 (for example, early majority), ... Are associated with each other.

For example, the target user classification unit 331 determines the score corresponding to the answer (any option) to each question for each question by referring to the above answer score table data, and totals each determined score for all the questions. By doing so, the total score is calculated (calculated for each target user). Then, the target user classification unit 331 refers to the characteristic discrimination condition data and assigns the characteristic corresponding to the calculated total score to the target user, thereby assigning the plurality of target users to any one of the plurality of characteristics. (Classify by target user). In this way, the characteristics given to the target user are stored in the user information database 301 in association with the user ID of the target user.

The questionnaire management server 1 may classify the plurality of target users into one of a plurality of preset characteristics based on the responses of the plurality of target users. In this case, the target user classification unit 331 acquires information indicating the correspondence between the target user (user ID) and the characteristics from the questionnaire management server 1.

(2-3-2. Teacher user identification unit 332)
Next, the teacher user identification unit 332 acquires the behavior history of each of the plurality of target users from the behavior history database 201 of the behavior history management server 2, and the acquired behavior history is related to the characteristics based on the hypothesis common to the questionnaire. It is determined for each target user whether or not the criteria set in the above are satisfied, and among the target users, the target user whose action history is determined to meet the above criteria is specified as a teacher user. Here, the "criteria set in relation to the characteristics" include "criteria set for each characteristic", "criteria having continuity across the characteristics", and "behavior distribution based on the behavior history of the entire user". At least one of the "criteria showing the correspondence between the characteristics and the characteristics" is included. To determine whether or not the criteria set for each characteristic are met, for example, are you searching for information on attributes (product name, service name, etc.) that are common to the terms used in questionnaire questions and answer options? , Are you purchasing the product in the information, or are using the service in the information, or search including not only the product name and service name but also the common brand and store, and the behavior related to this (purchasing behavior and use) It should include determining whether or not you are doing (acting, etc.). Furthermore, in order to judge whether or not this criterion is met, it is necessary to search for the area and time related to the questions in the questionnaire, and to judge whether or not the person is taking action (purchasing behavior, usage behavior, etc.) in the area or time. May be included.

In addition, as an example of determining whether or not a criterion having continuity across characteristics is satisfied, the time from when a specific product (for example, a trend product) is opened to the public until the target user accesses the web page ( Based on the access timing), did you access the above web page within 3 days from the publication (in this case, it corresponds to characteristic 01), or accessed the above web page after 4 days and within 14 days. It is possible to judge whether or not it was (in this case, it corresponds to the characteristic 02). In this case, for example, if the target user classified into the characteristic 01 accesses the web page after 4 days and within 14 days, the target user is excluded from the teacher user because it does not satisfy the criteria. In addition, as an example of a standard showing the correspondence between the behavior distribution (in other words, the distribution of the execution dynamic data set) based on the behavior history of the entire user and the characteristics, the access log of social media in a certain field (for example, the automobile field) Among the characteristics 01 to 05 in the overall distribution (behavior distribution), there is a criterion that the characteristic 01 and the characteristic 02 have a larger amount of information transmission and information reception in the social media in the field than the characteristic 04 and the characteristic 05. In this case, for example, when the amount of information transmitted by the target user classified into the characteristic 01 by the questionnaire is included in the lower level (for example, the distribution layer below the average) of the overall distribution of the access log of the social media, the criterion is used. Is excluded from the teacher user as not satisfying.

As a preferred example for identifying a teacher user, the teacher user identification unit 332 sets a group of inductive hypothesis verification items for each characteristic to derive highly probable characteristics from individual events (hypothesis verification items overlap between characteristics). Then, based on the behavior information extracted from the behavior history of the target user, it is determined whether or not the corresponding rate for the hypothesis verification item group exceeds the predetermined ratio, and the corresponding ratio is the predetermined ratio. It is advisable to identify the target user exceeding the above (that is, the target user determined to meet the criteria) as the teacher user. Here, the hit rate for the hypothesis verification item group is based on, for example, the result of calculating whether or not each of the plurality of hypothesis verification items is applicable (whether or not it is applicable) based on the behavior history of the target user (= number of applicable items /). Total number of hypothesis verification items). The hypothesis verification item group is set for each characteristic based on the hypothesis common to the questionnaire. For example, it is assumed that the item group corresponds to the question and answer options of the questionnaire and complements the question and answer options of the questionnaire. Good. As a result, for example, when the target user who purchased a certain trend product or new product is determined to have the characteristic 01 based on the questionnaire result, the certainty of the characteristic 01 is determined from the corresponding rate for the hypothesis verification item group for the target user. Judgment (in other words, demonstrating that it is characteristic 01) can be made. For example, whether or not a specific product (for example, a new product) or service is pre-ordered, whether or not a specific product (for example, a trend product) or service is searched or browsed, and whether or not a specific product (for example, a trend product) is purchased. , Whether or not a specific service is used, when the user accesses public information (for example, information on a specific product or service), and information about the user regarding public information (for example, information on a specific product or service). A hypothesis verification item group is set based on the amount of transmission and the user's ability to act on the disclosed information (for example, whether or not he / she went to a venue such as an exhibition or event of a specific product).

Here, the amount of information transmission is, for example, at least one of the number of comments (or tweets) posted and the amount of data (or the number of characters) per comment (or tweet). The access timing is, for example, access within 3 days from the date when the information (for example, product or service information) is published, access from the date when the information is published to 4 to 14 days, ... It can be divided into multiple stages. The amount of information sent is 100 or more comments posted for published information (or 300 characters or more per comment), and 60 to 99 comments posted for published information (or per comment). The number of characters is 100 to 299 characters), and so on, and it can be divided into a plurality of stages. The teacher user identification unit 332 may set a hypothesis verification item group based on the number of followers of the user. By using the above-mentioned criteria, the characteristic estimation server 3 determines which stage of the plurality of stages in the hypothesis verification item (for example, access timing) having continuity across the characteristics corresponds to the behavior of the target user. Judgment based on history, the stage of judgment (for example, access within 3 days from the date of publication) and the characteristics of the target user classified based on the questionnaire result are related (or matched). In some cases, the target user may be specified as a teacher user. Further, the characteristic estimation server 3 determines which level in the behavior distribution corresponds to the hypothesis verification item showing the correspondence between the behavior distribution and the characteristic based on the behavior history of the entire user based on the behavior history of the target user. Even if it is judged and the judged level (for example, lower level) and the characteristics of the target user classified based on the questionnaire result are related (or matched), the target user is specified as a teacher user. Good.

(2-3-3. Model generation unit 333)
Next, the model generation unit 333 has learned by using the behavior information extracted from the behavior history of the teacher user specified by the teacher user identification unit 332 as an explanatory variable and learning data using the characteristics of the teacher user as an objective variable. Generate a model. Here, the action information can be obtained, for example, by setting the data extracted from the action history in a data format in which data of a plurality of items can be set. Examples of this item are whether or not you have a pre-order for a particular product or service, whether or not you have searched or browsed for a particular product or service, whether or not you have purchased a particular product, whether or not you have used a particular service, and how many searches (or searches) Frequency), search conditions, browsing frequency (or browsing frequency), payment frequency (or payment frequency), posting frequency (posting frequency), access timing, information transmission amount, and the like.

For the trained model, for example, a decision tree is adopted. The decision tree is a method of dividing and classifying segments by conditional branching according to a predetermined branching rule (an example of a classification standard). In such a decision tree, the degree of fit of a plurality of objective variables for each segment is calculated. FIG. 6 is a conceptual diagram showing an example of conditional branching in a decision tree. In the segment shown in FIG. 6, each characteristic and the ratio (appearance rate) of the specified number of people are shown. “X” and “xx” in the segments Se1 and Se2 shown in FIG. 6 indicate numerical values, respectively, and indicate that “xx” is a larger numerical value than “x”.

For example, the teacher data classified by the questionnaire is classified into characteristic 01 to characteristic 05, and the ratio of the number of people (appearance rate) is specified for each characteristic. This set is classified according to the characteristics that are the objective variables using various explanatory variables. Here, for example, a branch rule has been set in which items including items including purchase of product ABC (X1), purchase of product XYZ (X2), and search for a keyword expressing a predetermined trend (X3) are set as explanatory variables. Become. When these are classified using a decision tree as an explanatory variable, for example, the segment Se1 shows that the appearance rate xx of one characteristic 01 is equal to or higher than the specified value yy. The explanatory variables in this case are X1 and X2 when referring to the branch rule. In this way, the branching rule is set so that the degree of fit of the characteristic 01, which is one objective variable, is equal to or higher than a predetermined reference, and the classification is specified as the classification using the explanatory variables X1 and X2. Similarly, the segment Se2 is specified as a set of characteristics 02 with explanatory variables X1, X2, and X3 by a branching rule such that the appearance rate xx of the characteristic 02 is equal to or higher than the specified value.

In the generation of the trained model, of the teacher data of each of the plurality of teacher users, a trained model is constructed using some teacher data as training data, and the remaining part is used as test data to verify the trained model. It is good to do. Also, the supervised machine learning algorithms for generating trained models are not particularly limited, but for example, Gradient Boosting, or Gradient Boosting and Random Forests. The combined XG Boost may be used. Gradient boosting is a method of creating multiple decision trees and training them in a group. It constructs multiple decision trees as a trained model and uses the information of the previous decision tree to create a new decision tree. It is designed to be boosted to build. Specifically, a new decision tree is constructed with the gradient of the loss function, which could not be predicted by the previous decision tree, as the objective variable. In each decision tree, the explanatory variables are classified according to the conditions on the way from the root to the branch, and when the terminal leaf is reached, the value given to the terminal leaf is returned as a predicted value. XGBoost is basically the same as the gradient boosting method, but because it uses a random forest, not all explanatory variables are used, but the number of explanatory variables at a randomly determined rate. Is selected and a decision tree is constructed.

(2-3-4. Characteristic estimation unit 334)
Next, the characteristic estimation unit 334 uses the behavior history management server to collect the behavior history of the estimated user (user other than the teacher user) determined as the estimation target from all the users whose accounts have been issued as members of the predetermined site, for example. The characteristics of the estimated user are estimated using the behavior information acquired from the behavior history database 201 of 2 and extracted from the acquired behavior history and the learned model generated by the model generation unit 333. That is, the characteristic estimation unit 334 obtains the characteristics (predicted value) for the behavior information by inputting the behavior information extracted from the behavior history into the trained model. The characteristics estimated for the estimated user in this way are associated with the user ID of the estimated user and stored in the user information database 301. When the behavior history of the estimated user is updated, the characteristic estimation unit 334 re-estimates the characteristics of the estimated user using the behavior information extracted from the updated behavior history and the learned model. It is recommended to update by. As a result, the characteristics estimated for the estimated user can be made to correspond to the change in the behavior history of the estimated user.

(2-3-5. Model update unit 335)
When the behavior history of the teacher user is updated by updating the behavior history of the target user, the model update unit 335 uses the behavior information extracted from the updated behavior history as an explanatory variable and sets the characteristics of the teacher user. The trained model (factor (parameter) in the trained model) is updated using the training data used as the objective variable. As a result, the trained model can be updated according to, for example, a trend (trend) that changes at predetermined intervals. In this case, the teacher user identification unit 332 determines whether or not the updated behavior history of the teacher user satisfies the above criteria, and the model update unit 335 updates the teacher user determined to meet the criteria. It is advisable to update the trained model using the training data with the behavior information extracted from the behavior history as the explanatory variable and the characteristics of the teacher user as the objective variable. As a result, it is possible to improve the estimation accuracy by the trained model by following the change in the behavior history of the teacher user.

Further, the update interval of the trained model due to the update of the behavior history of the teacher user may be longer than the update (re-estimation) interval of the characteristics of the estimated user. As a result, the update of the trained model due to the update of the behavior history of the teacher user can be minimized, and the server load can be reduced. When the characteristics of the estimated user are estimated, the model update unit 335 uses the behavior information extracted from the behavior history of the estimated user as an explanatory variable, and the learning data using the estimated characteristics of the estimated user as an objective variable. The trained model may be updated using.

(2-3-6. Classification Criteria Change Unit 336)
When the behavior history of the estimated user is updated, the classification standard changing unit 336 sets the condition that the appearance rate of the estimated user belonging to the estimated characteristic is the same in the predetermined branching rule in the trained model, and is estimated. The branching rule is modified so that the presumed user belonging to the above characteristic satisfies the condition that the estimated user has a predetermined degree of matching. The reason for this will be explained below.

As the behavior history of the estimated user changes, for example, when the number of people who purchase the "product XYZ" shown in FIG. 6 increases, the appearance rate of the characteristic 01 (for example, 10%) increases in the branch rule. Therefore, the branching rule is changed so that the appearance rate of the characteristic 01 is maintained at the same level, for example, at 10% (that is, the appearance rate of the estimated users belonging to the estimated characteristics is satisfied). Equivalent to changing the trained model). As an example of changing the branching rule, change the weighting (coefficient) of the explanatory variable "product XYZ" purchase (X2), and replace the explanatory variable "product XYZ" purchase (X2) with "product ABC" purchase (X1). That, adding a new branch for purchasing "Product ABC" (X1), and changing the branch configuration. In other words, when the behavior history of the estimated user is changed, the variable configuration and weighting of the trained model change, so the change is taken into consideration to recreate the overall optimal model.

However, it is assumed that the classification will be completely different if the appearance rates are kept the same, so as a constraint condition, the rate at which each estimated user shifts to another characteristic before and after the change of the trained model is set. Set low (for example, set the degree of matching before and after the change of the trained model to 90% or more) (that is, set so that the estimated user belonging to the estimated characteristic satisfies the condition having a predetermined degree of matching). As a result, it is possible to improve the estimation accuracy by the trained model by following the change in the behavior history of the estimated user whose characteristics have been estimated.

The change interval of the trained model due to the update of the behavior history of the estimated user (according to the classification standard change unit 336) is the trained model due to the update of the behavior history of the teacher user (according to the model update unit 335). It is better to make it shorter than the update interval of. As a result, the update of the trained model due to the update of the behavior history of the teacher user can be minimized, and the server load can be reduced.

[3. Various processes of the characteristic estimation server 3]
Next, various processes executed by the system control unit 33 of the characteristic estimation server 3 will be described.

(3-1. Trained model generation process)
First, the trained model generation process executed by the system control unit 33 will be described with reference to FIG. 7. FIG. 7 is a flowchart showing an example of the trained model generation process executed by the system control unit 33.

The trained model generation process shown in FIG. 7 is started in response to an instruction from an operator, for example. When the process shown in FIG. 7 is started, the system control unit 33 sends the questionnaire results of the plurality of target users (plurality of predetermined target users) that can be identified from the user information database 301 to the questionnaire management server 1. Obtained from the questionnaire database 101 (step S1). For example, the system control unit 33 questionnaires the questionnaire results of each target user by transmitting a request including a target user list in which the user ID (or user name) of each target user is described to the questionnaire management server 1. Obtained from management server 1. The questionnaire result acquired in this way is stored in the memory in association with the user ID (or user name) of each target user, for example.

Next, the system control unit 33 acquires the action history of each of the plurality of target users from the action history database 201 of the action history management server 2 (step S2). For example, the system control unit 33 acquires the action history of each target user from the action history management server 2 by transmitting the request including the target user list to the action history management server 2. The action history acquired in this way is stored in the memory in association with the user ID (or user name) of each target user, for example.

Next, the system control unit 33 selects one target user (for example, by user ID or user name) from the target users for which the questionnaire result and the action history have been acquired (step S3).

Next, the system control unit 33 (target user classification unit 331) has one of a plurality of characteristics preset by the target user based on the questionnaire result (answer) of the target user selected in step S3. It is specified whether it is classified into a characteristic (step S4). The characteristics identified by this are stored in the memory in association with, for example, the user ID (or user name) of the target user.

Next, the system control unit 33 (teacher user identification unit 332) determines whether or not the behavior history of the target user selected in step S3 satisfies the criteria set based on the hypothesis common to the questionnaire (step S5). .. When the system control unit 33 (teacher user identification unit 332) determines that the behavior history of the target user satisfies the criterion (step S5: YES), the system control unit 33 identifies the target user as a teacher user (step S6), and steps. Proceed to S7. The user type indicating the teacher user is stored in the memory in association with the user ID (or user name) of the target user identified as the teacher user in this way. On the other hand, when the system control unit 33 (teacher user identification unit 332) determines that the behavior history of the target user does not satisfy the criterion (step S5: NO), the step is performed without specifying the target user as the teacher user. Proceed to S7.

In step S7, the system control unit 33 determines whether or not there is a target user who has not yet been selected among the target users whose questionnaire results and action history have been acquired. When the system control unit 33 determines that there is a target user who has not been selected yet (step S7: YES), the system control unit 33 returns to step S3 and performs the same process as described above. On the other hand, when the system control unit 33 determines that there is no target user who has not been selected yet (step S7: NO), the system control unit 33 proceeds to step S8.

In step S8, the system control unit 33 (model generation unit 333) uses the action information extracted from each action history of the teacher user specified in step S6 as an explanatory variable, and each of the teacher users specified in step S6. A trained model trained by training data with the characteristics of is used as the objective variable is generated. The trained model generated in this way is used in the characteristic estimation process described later.

Next, the system control unit 33 stores the characteristics identified in step S4 for each target user in the user information database 301 in association with the user ID of each target user, and further identifies the target user as a teacher user in step S6. For the target user, the user type indicating the teacher user is stored in the user information database 301 in association with the user ID (or user name) of the target user (step S9), and the process shown in FIG. 7 is terminated. ..

The trained model generation process will be executed again when a new questionnaire is conducted for the target user.

(3-2. Characteristic estimation processing)
Next, the characteristic estimation process executed by the system control unit 33 will be described with reference to FIG. FIG. 8 is a flowchart showing an example of the characteristic estimation process executed by the system control unit 33.

The characteristic estimation process shown in FIG. 8 is started in response to an instruction from an operator, for example. When the process shown in FIG. 8 is started, the system control unit 33 displays the behavior history of one or more estimated users (for example, one or more estimated users whose characteristics have not been estimated yet) that can be specified from the user information database 301. It is acquired from the action history database 201 of the action history management server 2 (step S11). For example, the system control unit 33 transmits the action history of the estimated user to the action history management server 2 by transmitting a request including the estimated user list in which the user ID (or user name) of the estimated user is described to the action history management server 2. Get from 2. The action history acquired in this way is stored in the memory in association with, for example, the user ID (or user name) of the estimated user.

Next, the system control unit 33 selects one estimated user (for example, by user ID or user name) from the estimated users whose action history has been acquired (step S12).

Next, the system control unit 33 (characteristic estimation unit 334) estimates the characteristics of the behavior information by inputting the behavior information extracted from the behavior history of the estimated user selected in step S12 into the learned model (step). S13). The characteristics estimated in this way are stored in the memory in association with, for example, the user ID (or user name) of the estimated user. In this way, it is possible to extend the characteristics to the estimated users who are not the subjects of the questionnaire.

The system control unit 33 (model update unit 335) uses the behavior information extracted from the behavior history of the estimated user whose characteristics are estimated in step S13 as an explanatory variable, and the estimated characteristics of the estimated user as an objective variable. The trained model may be updated using the trained data.

Next, the system control unit 33 determines whether or not there is an estimated user who has not yet been selected among the estimated users whose action history has been acquired (step S14). When the system control unit 33 determines that there is an estimated user who has not been selected yet (step S14: YES), the system control unit 33 returns to step S12 and performs the same process as described above. On the other hand, when the system control unit 33 determines that there is no estimated user who has not been selected yet (step S14: NO), the system control unit 33 proceeds to step S15.

In step S15, the system control unit 33 stores the characteristics estimated in step S13 for the estimated user in the user information database 301 in association with the user ID of the estimated user, and ends the process shown in FIG.

The characteristic estimation process will be executed for the user to whom the account is newly issued.

(3-3. Characteristic update processing)
Next, the characteristic update process executed by the system control unit 33 will be described with reference to FIG. FIG. 9 is a flowchart showing an example of the characteristic update process executed by the system control unit 33.

The characteristic update process shown in FIG. 9 is executed, for example, at predetermined time intervals (for example, every 24 hours or every week). When the process shown in FIG. 9 is started, the system control unit 33 acquires a list of users whose action history has been updated from the action history management server 2 (step S21). In this list, for example, the user ID (or user name) of the user whose action history has been updated is described.

Next, the system control unit 33 identifies an estimated user from the list acquired in step S21 by referring to the user information database 301, and updates the action history of one or more identified estimated users in step S11. In the same manner as above, it is acquired from the action history database 201 of the action history management server 2 (step S22).

Next, the system control unit 33 selects one estimated user (for example, by user ID or user name) from the estimated users for whom the updated action history has been acquired (step S23).

Next, the system control unit 33 (characteristic estimation unit 334) inputs the behavior information extracted from the updated behavior history of the estimated user selected in step S23 into the trained model to input the characteristics for the behavior information. Is re-estimated (step S24). The characteristics re-estimated in this way are stored in the memory in association with, for example, the user ID (or user name) of the estimated user.

In addition, the system control unit 33 (classification standard change unit 336) is a condition that the appearance rate of the estimated users belonging to the estimated characteristics is the same in the predetermined branching rule in the trained model before or after the step S24. , And the condition that the estimated user belonging to the estimated characteristic has a predetermined degree of matching, the branching rule may be changed.

Next, the system control unit 33 determines whether or not there is an estimated user who has not yet been selected among the estimated users for whom the updated action history has been acquired (step S25). When the system control unit 33 determines that there is an estimated user who has not been selected yet (step S25: YES), the system control unit 33 returns to step S23 and performs the same processing as described above. On the other hand, when the system control unit 33 determines that there is no estimated user who has not been selected yet (step S25: NO), the system control unit 33 proceeds to step S26.

In step S26, the system control unit 33 stores (overwrites and updates) the characteristics re-estimated in step S24 for the estimated user in the user information database 301 in association with the user ID of the estimated user, and performs the process shown in FIG. finish.

(3-4. Trained model update process)
Next, the trained model update process executed by the system control unit 33 will be described with reference to FIG. FIG. 10 is a flowchart showing an example of the trained model update process executed by the system control unit 33.

The trained model update process shown in FIG. 10 is executed at a longer time interval (for example, every six months or every year) than the characteristic update process. When the process shown in FIG. 10 is started, the system control unit 33 acquires a list of users whose action history has been updated from the action history management server 2 (step S31).

Next, the system control unit 33 determines whether or not a predetermined number or more of teacher users can be specified from the list acquired in step S31 by referring to the user information database 301 (step S32). This predetermined number may be a number according to a ratio of, for example, 1% to 5% of the total number of teacher users that can be specified from the user information database 301. When the system control unit 33 determines that a predetermined number or more of teacher users have been identified from the acquired list (step S32: YES), the system control unit 33 proceeds to step S33. On the other hand, when it is determined that the teacher users cannot be specified by a predetermined number or more from the acquired list (step S32: NO), the system control unit 33 ends the process shown in FIG.

In step S33, the system control unit 33 acquires the characteristics of the teacher user whose action history has been updated from the user information database 301. Next, the system control unit 33 acquires the updated behavior history of the teacher user whose behavior history has been updated from the behavior history database 201 of the behavior history management server 2 (step S34).

Next, the system control unit 33 selects one teacher user (for example, by user ID or user name) from the teacher users whose action history has been updated (step S35).

Next, the system control unit 33 (teacher user identification unit 332) determines whether or not the updated behavior history of the teacher user selected in step S35 meets the criteria set based on the hypothesis common to the questionnaire ( Step S36). When the system control unit 33 (teacher user identification unit 332) determines that the updated action history of the teacher user satisfies the criterion (step S36: YES), the system control unit 33 identifies the teacher user (step S37), and steps. Proceed to S38. On the other hand, when the system control unit 33 (teacher user identification unit 332) determines that the updated behavior history of the teacher user does not satisfy the criterion (step S36: NO), the system control unit 33 does not specify the teacher user as a teacher user (adopt). ) Proceed to step S38.

As another example, the process of steps S35 to S38 may not be performed. That is, when the behavior history of the teacher user is updated, the system control unit 33 does not determine whether or not the updated behavior history of the teacher user satisfies the above criteria, but from the updated behavior history of the teacher user. The trained model may be updated using the training data using the extracted behavior information as the explanatory variable and the characteristics of the teacher user as the objective variable.

In step S38, the system control unit 33 determines whether or not there is a teacher user who has not yet been selected among the teacher users whose action history has been updated. When the system control unit 33 determines that there is a teacher user who has not been selected yet (step S38: YES), the system control unit 33 returns to step S35 and performs the same processing as described above. On the other hand, when the system control unit 33 determines that there is no teacher user who has not been selected yet (step S38: NO), the system control unit 33 proceeds to step S39.

In step S39, the system control unit 33 (model update unit 335) uses the action information extracted from the updated action history of the teacher user specified in step S37 (action history acquired in step S34) as an explanatory variable. , The trained model is updated using the training data with the characteristics of the teacher user identified in step 37 (characteristics acquired in step S33) as the objective variable, and the process shown in FIG. 10 is completed.

The trained model updated by the trained model update process is used when the characteristics of the estimated user are subsequently estimated.

As described above, according to the above-described embodiment, the characteristic estimation server 3 sets the plurality of target users among the plurality of characteristics preset based on the responses of the plurality of target users who responded to the questionnaire. It is classified into one of the following characteristics, and it is determined whether or not the behavior history of each target user meets a predetermined criterion. Then, the characteristic estimation server 3 uses learning data in which the target user determined to satisfy the criteria is the teacher user, the behavior information extracted from the behavior history of the teacher user is used as the explanatory variable, and the characteristic of the teacher user is used as the objective variable. A trained model was generated, and the behavior information extracted from the behavior history of the estimated user and the generated learned model were used to estimate the characteristics of the estimated user. Empirical characterization can be performed based on the results. That is, it is possible to perform empirical characteristic estimation based on hypothesis verification based on the subjective and conscious judgment of the target user and hypothesis verification based on the objective facts of the target user.

Further, in the above embodiment, the accuracy of the teacher data used for generating the trained model can be improved by supplementing the incompleteness and uncertainty that may occur in the questionnaire of the target user with the behavior history of the target user. However, in addition to this, it is possible to improve the teacher data (pre-teacher data) at the stage created based on the results of the characteristic classification by the questionnaire. That is, as described above, the characteristic estimation server 3 (an example of the classification means and the determination means of the characteristic classification device) sets the target users in advance based on the answers of the plurality of target users who answered the questionnaire. It is classified into one of the characteristics of the above, and it is determined whether or not the behavior history of the target user meets the criteria set in relation to the characteristics based on the hypothesis common to the questionnaire. Then, the characteristic estimation server 3 (an example of the identification means of the characteristic classification device) identifies the improvement items (improvement points) of the questionnaire from the judgment result of whether or not the criteria are satisfied. For example, what kind of question (questionnaire question) was useful or useless for characteristic classification judgment (segment judgment) based on the judgment result of whether or not the criteria were met, and what kind of questionnaire option (answer to the question) By determining whether the classification method (option) was useful or useless for the classification judgment of the characteristics, and what classification method (calculation method used for the classification of the characteristics) was suitable for the classification judgment of the characteristics, the improvement items of the questionnaire ( For example, at least one of questionnaire questions, questionnaire options, and classification methods) is specified. This can contribute to improving the quality of the questionnaire.

For example, the characteristic estimation server 3 determines whether or not each of the plurality of hypothesis verification items set for each characteristic is applicable based on the hypothesis common to the questionnaire based on the behavior history of the target user, and the determination result (that is, a plurality of). It is preferable to specify at least one item out of a plurality of items related to the questionnaire (for example, questionnaire questions, questionnaire options, and classification method) as improvement items based on whether or not each of the hypothesis verification items is applicable. Further, the characteristic estimation server 3 determines which stage corresponds to a plurality of stages in the hypothesis verification item (for example, access timing) having continuity across the characteristics based on the behavior history of the target user, and determines which stage corresponds to the stage. Based on the judgment result, at least one of a plurality of items related to the questionnaire (for example, questionnaire questions, questionnaire options, and classification method) may be specified as improvement items. For example, an item that influences the disagreement (or match) between the judged stage and the characteristics classified based on the questionnaire result is specified as an improvement item. Further, the characteristic estimation server 3 determines which level in the behavior distribution corresponds to the hypothesis verification item showing the correspondence between the behavior distribution and the characteristic based on the behavior history of the entire user based on the behavior history of the target user. Judgment may be made, and at least one of a plurality of items related to the questionnaire (for example, questionnaire questions, questionnaire options, and classification method) may be specified as improvement items based on the judgment result. For example, an item that influences the disagreement (or match) between the judged level and the characteristic classified based on the questionnaire result is specified as an improvement item.

FIG. 11 is a conceptual diagram showing the content of improvement when a questionnaire option is specified as an improvement item of the questionnaire. In FIG. 11A, the questionnaire option 2 is divided because the possibility of dividing the questionnaire option is found. In FIG. 11B, a part of the questionnaire option 2 is integrated into the questionnaire option 1 by finding different data trends in the questionnaire options. In addition, although not shown, when a classification method is specified as an improvement item in the questionnaire, a more suitable classification method is selected from a plurality of classification methods. For example, when classified into one of the characteristics based on the response of the target user, the answer to yes (applicable) in the question about trend product purchase in the questionnaire is scored as +5 points, and as a result, the characteristic of the target user. Is discriminated (classified) as the characteristic 01, the behavior history of the target user may not satisfy the criteria (that is, the target user does not correspond to the characteristic 01). In this case, it was not appropriate to give +5 points to the answer "Applicable" to the question about purchasing trend products, that is, the classification method was not appropriate, and the classification method was specified as an improvement item in the questionnaire. Will be done. As a result, the content of the question that affects the classification method is changed, the score of the question is changed (that is, the weighting of the options for the question is changed), or the question itself is deleted. In addition, when question 1 of the questionnaire sets yes to +2 points, no to 0 points, question 2 sets yes to +1 points and no to -1 points, for example, when a target user with a total of 3 points or more is determined as characteristic 01, The behavior history of the target user may not meet the criteria and may be inappropriate for the characteristic 01. In this case as well, the classification method is specified as an improvement item of the questionnaire, and the content of the question in the questionnaire is changed, the score of the question is changed, or the question itself is deleted. That is, the calculation logic used for characteristic discrimination is optimized.

Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment, and changes from the above embodiment to various configurations and the like are made without departing from the gist of the present invention. It may be added, and it is also included in the technical scope of the present invention.

1 Questionnaire management server 2 Action history management server 3 Characteristic estimation server 31 Communication unit 32 Storage unit 33 System control unit 34 System bus 101 Questionnaire database 201 Action history database 301 User information database UTm User terminal STn processing terminal SAo server NW communication network S communication system

Claims (12)

A classification means for classifying the first user into one of a plurality of preset characteristics based on the responses of the plurality of first users who responded to the questionnaire.
A means for determining whether or not the behavior history of the first user meets the criteria set in relation to the characteristics based on the hypothesis common to the questionnaire.
Among the plurality of first users, the first user whose behavior history is determined to satisfy the criteria is used as a teacher user, and the behavior information extracted from the behavior history of the teacher user is used as an explanatory variable, and the characteristics of the teacher user. A generation means for generating a trained model trained by training data with
An estimation means for estimating the characteristics of the second user using the behavior information extracted from the behavior history of the second user different from the first user and the trained model.
A characteristic estimation device comprising. When the behavior history of the second user is updated, the estimation means re-estimates the characteristics of the second user using the behavior information extracted from the updated behavior history and the learned model. The characteristic estimation device according to claim 1, wherein the characteristic estimation device is updated by When the behavior history of the second user is updated, the conditions under which the appearance rate of the second user belonging to the estimated characteristic is equivalent and the estimated characteristic are satisfied in the predetermined classification criteria in the trained model. The characteristic estimation device according to claim 1 or 2, further comprising a changing means for changing the predetermined classification standard so that the second user to which the second user belongs satisfies the condition having a predetermined degree of agreement. When the behavior history of the teacher user is updated by updating the behavior history of the first user, the behavior information extracted from the updated behavior history is used as an explanatory variable, and the characteristic of the teacher user is used as an objective variable. The characteristic estimation device according to any one of claims 1 to 3, further comprising an update means for updating the trained model using the training data. When the behavior history of the teacher user is updated by updating the behavior history of the first user, the determination means sets the updated behavior history of the teacher user in relation to the characteristic. The characteristic estimation device according to claim 4, wherein it is determined whether or not the above criteria are satisfied. The characteristic estimation device according to any one of claims 1 to 5, wherein the generation means does not employ a first user whose behavior history is determined not to satisfy the criteria as a teacher user. When the behavior history of the teacher user is updated by updating the behavior history of the first user, the behavior information extracted from the updated behavior history is used as an explanatory variable, and the characteristic of the teacher user is used as an objective variable. Further provided with an update means for updating the trained model using the training data of
The characteristic estimation device according to claim 2, wherein the update interval of the trained model is longer than the update interval of the characteristics of the second user. The characteristic estimation device according to any one of claims 1 to 7, wherein the standard set in relation to the characteristic is a standard set for each of the characteristics. The characteristic estimation device according to any one of claims 1 to 7, wherein the standard set in relation to the characteristic is a standard having continuity across the characteristic. The standard set in relation to the characteristic is described in any one of claims 1 to 7, wherein the standard is a standard showing a correspondence relationship between the behavior distribution based on the behavior history of the entire user and the characteristic. Characteristic estimation device. A method of character estimation performed by a computer
A step of classifying the first user into one of a plurality of preset characteristics based on the answers of each of the plurality of first users who responded to the questionnaire.
A step of determining whether or not the behavior history of the first user meets the criteria set in relation to the characteristics based on the hypothesis common to the questionnaire, and
Among the plurality of first users, the first user whose behavior history is determined to satisfy the criteria is used as a teacher user, and the behavior information extracted from the behavior history of the teacher user is used as an explanatory variable, and the characteristics of the teacher user. Steps to generate a trained model trained with training data with
A step of estimating the characteristics of the second user using the behavior information extracted from the behavior history of the second user different from the first user and the trained model.
A characteristic estimation method characterized by including. A step of classifying the first user into one of a plurality of preset characteristics based on the answers of each of the plurality of first users who responded to the questionnaire.
A step of determining whether or not the behavior history of the first user meets the criteria set in relation to the characteristics based on the hypothesis common to the questionnaire, and
Among the plurality of first users, the first user whose behavior history is determined to satisfy the criteria is used as a teacher user, and the behavior information extracted from the behavior history of the teacher user is used as an explanatory variable, and the characteristics of the teacher user. Steps to generate a trained model trained with training data with
A step of estimating the characteristics of the second user using the behavior information extracted from the behavior history of the second user different from the first user and the trained model.
A characteristic estimation program characterized by having a computer execute the above.

Images