JP7325598B1 - Recommendation server, recommendation method and program - Google Patents

Abstract

A recommendation server, a recommendation method, and a program for associating users and contents with predetermined factors and recommending contents based on the factors. SOLUTION: In an information processing system having a plurality of user terminals, a recommendation server 801 uses factor information associated with content and factor score information associated with a user as explanatory variables, and reaction probability as an objective variable. a first model acquisition unit that acquires a first trained estimation model generated by learning the first estimation model as, for each of a plurality of target contents, factor information associated with the target content; A first estimation unit that inputs the factor score information associated with the target user into the first trained estimation model to estimate the reaction probability, and based on the estimated reaction probability for each of the plurality of target contents, A first extraction unit extracting one or more target contents from a plurality of target contents is provided. [Selection drawing] Fig. 8

Description

The present invention relates to a recommendation server, a recommendation method, and a program.

JP-A-2011-175362 (Patent Document 1) is a document that discloses the background art of this technical field. This publication describes "recommendation engine: a system module that recommends content based on user's preferences or user's actions on content" (see paragraph 0028).

More specifically, in this publication, "First, the analysis unit 120 acquires the basic attributes of the content stored in the attribute table of the storage unit 110 (step S102). Next, the analysis unit 120 , PLSA, or LDA, based on the attribute value of the basic attribute, the attribute value of the extended attribute is calculated (step S104). stored in.” (see paragraph 0102). Also, in this publication, "First, the recommendation unit 140 selects content to be recommended to the user using the attribute values stored in the attribute table held by the storage unit 110 (step S202). .” (see paragraph 0106). That is, this publication describes that attribute values are calculated by PLSA and stored in an attribute table, and content to be recommended to the user is selected using the attribute values stored in the attribute table. .

JP 2011-175362 A

PLSA described in Patent Document 1 is known as one of recommendation algorithms. This PLSA is an algorithm for a posteriori (implicit) estimation of latent factors behind users and contents. In contrast to the recommendation method using such an algorithm, the present invention provides a mechanism for associating users and contents with predetermined factors and recommending contents based on the factors.

In order to solve the above problems, for example, the configurations described in the claims are adopted .

According to the present invention, it is possible to provide a mechanism for associating users and contents with predetermined factors and recommending contents based on the factors.
Problems, configurations, and effects other than those described above will be clarified by the following description of the embodiments.

FIG. 1 shows an example of a recommendation method using Matrix Factorization. FIG. 2 shows an example of a conceptual model of a global domain and individual domains. FIG. 3 shows another example of a conceptual model of global domains and individual domains. FIG. 4 shows examples of WB factors. FIG. 5 shows an example of a factor score table. FIG. 6 shows a comparative example of the recommendation method. FIG. 7 shows an example of a topic model. FIG. 8 shows an example of the configuration of an information processing system 800. As shown in FIG. FIG. 9 shows an example of the configuration of the recommendation server 801. As shown in FIG. FIG. 10 shows an example of user attribute information 1000. As shown in FIG. FIG. 11 shows an example of factor score information 1100. FIG. FIG. 12 shows an example of a content action log 1200. As shown in FIG. FIG. 13 shows an example of content action score information 1300. As shown in FIG. FIG. 14 shows an example of a service behavior log 1400. As shown in FIG. FIG. 15 shows an example of service behavior score information 1500 . FIG. 16 shows an example of content information 1600. As shown in FIG. FIG. 17 shows an example of service information 1700 . FIG. 18 shows an example activity log 1800 . FIG. 19 shows an example of the flow of recommendation processing. FIG. 20 shows an example of this scoring process. FIG. 21 shows an example of this scoring process. FIG. 22 shows an example of input data to be learned by the global optimum model. FIG. 23 shows an example of input data to be input to the trained global optimum model. FIG. 24 shows an example of input data to be learned by the individual optimum model. FIG. 25 shows an example of input data to be input to the trained individual optimum model. FIG. 26 shows an example of a registration screen. FIG. 27 shows an example of a registration screen. FIG. 28 shows an example of a registration screen. FIG. 29 shows an example of a recommendation screen. FIG. 30 shows an example of a recommendation screen. FIG. 31 shows an example of the detail screen. FIG. 32 shows an example of My Page. FIG. 33 shows an example of the detail screen. FIG. 34 shows an example of an edit screen.

1. Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings.
1-1. Overview Conventionally, many recommendation services have made recommendations within individual domains. For example, FIG. 1 shows an example of a recommendation method using Matrix Factorization.

In the recommendation method shown in the figure, the evaluation matrix 101 is decomposed into a user feature amount matrix 102 and an item feature amount matrix 103 . Then, the two decomposed matrices are multiplied to generate a new evaluation matrix 104 . Finally, items with high values in the generated new evaluation matrix 104 are recommended.
The upper part of the figure shows a method for recommending products based on product purchase data, and the lower part shows a method for recommending movies based on movie viewing data.

In such a recommendation method, user preferences are defined from behavior data of individual domains. Therefore, it is difficult to make recommendations across domains.

On the other hand, the recommendation method according to the present embodiment uses a factor that expresses the potential psychological state of the user, specifically, a well-being factor that expresses the value of human happiness. Specifically, in this recommendation method, well-being factors predefined on the business side are associated with each of users, well-being contents, and well-being services. On top of that, from the viewpoint of maximizing the well-being of users, we aim for overall domain optimization using well-being content, and then aim for individual domain optimization using well-being services. Therefore, according to this recommendation method, recommendations can be made across domains.

The well-being mentioned above is a term that means a state of health and security, a satisfying living condition, happiness, and the like. In the following description, this well-being may be abbreviated as "WB".

FIG. 2 shows an example of a conceptual model of a global domain and individual domains. In the conceptual model shown in the figure, WB contents 202 (for example, "work") belonging to a general domain 201 are associated with one or more WB services 204 (for example, "facility environment") belonging to individual domains 203, respectively. ing.

To explain the recommendation method according to the present embodiment using this conceptual model, first, "work" is recommended to users who are highly interested in "work" as overall optimization. Then, as an individual optimum, a specific service related to "work" is recommended to the user.

FIG. 3 shows another example of a conceptual model of global domains and individual domains. In the conceptual model shown in the figure, WB contents 302 belonging to a general domain 301 are associated with one or more WB services 304 belonging to individual domains 303 respectively.

Using this conceptual model, the recommendation method according to the present embodiment will be described. First, WB content is recommended to the user as the overall optimum. At this time, an area of consciousness in which the user is highly interested is estimated. Then, as individual optimization, the WB service associated with the WB content is recommended to the user. In other words, further WB services are recommended within the user's area of interest. This improves the well-being of the user.

Next, the WB factor will be explained.
FIG. 4 shows examples of WB factors. A total of 21 WB factors shown in the figure are 7 WB factors related to activity or body belonging to the category "wellness", 7 WB factors related to element or spirit belonging to the category "newness", and WB factors belonging to the category "wellness". It consists of 7 WB factors related to the state or society belonging to the "community".

These WB factors are defined in advance as latent factors on the business side from the viewpoint of well-being. In other words, these WB factors are predefined by the operator as representing the user's happiness level or as representing the user's well-being values.

These WB factors are associated with users. As a specific method, first, a questionnaire about the WB factor is given to the user at the time of service registration. This questionnaire asks the user to answer the degree of importance of each WB factor. For details of this questionnaire, refer to the registration screen 2800 shown in FIG. 28, for example.

Then, the result of the questionnaire is stored in a table in association with the user. FIG. 5 shows an example of a factor score table. A factor score table 500 shown in the figure stores the score of each WB factor for each user.

Next, WB contents will be described.
The WB content is content for recommendation defined in advance by the provider, and is content for clarifying the user's current interest and area of concern. In other words, the WB content is pre-defined advice information that improves the user's happiness. As an example, 400 kinds of WB contents are prepared. For a specific example of this WB content, see the recommendation screen 2900 shown in FIG. 29, for example.

Each WB content is associated with a WB factor in a table. See, for example, content information 1600 shown in FIG. 16 for a table that associates WB content with WB factors.

Next, the WB service will be explained.
The WB service is a provider service for recommendation defined in advance on the provider side, and is a service within each user's current interest and area of concern. For a specific example of this WB service, refer to the recommendation screen 3000 shown in FIG. 30, for example.

Each WB service is associated with a WB factor in a table. For a table that associates WB services with WB factors, see service information 1700 shown in FIG. 17, for example.
Each WB service is associated with WB contents, as shown in FIGS.

Next, the flow of the recommendation method according to this embodiment will be described.
In the recommendation method according to the present embodiment, first, a user's interests and interest areas are estimated in the entire domain, and WB contents are recommended. Next, in the individual domain, the user's interests and WB services within the area of interest are recommended. In other words, in the individual domain, the WB service associated with the WB content that the user has reacted to is recommended.

Thus, the recommendation method according to the present embodiment has two stages. On the other hand, a recommendation method that directly recommends a WB service without recommending WB contents is conceivable. FIG. 6 shows a comparative example of the former recommendation method and the latter recommendation method.

All of the recommendation methods shown in the figure are methods of recommending books. However, the step-by-step recommendation 601 first recommends the genre of the book, and then recommends the book belonging to that genre. On the other hand, stepless recommendation 602 directly recommends a book without recommending the genre of the book.

A two-step recommendation method such as the step-by-step recommendation 601 gradually narrows down to a specific WB service from a vague area of interest or interest rather than suddenly recommending a WB service that may not be of interest or interest. It is based on the rule of thumb that users respond favorably to this method.

Next, the difference between the recommendation method according to the present embodiment and the topic model will be described.
First, the topic model is one of the recommendation algorithms, and is an algorithm for ex-post (implicitly) estimating latent factors behind users and recommended contents. PLSA (Probabilistic Latent Semantic Analysis), which is a representative example of this topic model, is illustrated as shown in FIG. 7 and expressed by the following equation.
... (1)

In this formula (1), P(c _i , g _j ) represents co-occurrence probability, P(z _k ), P(c _i |z _k ) and P(g _j |z _k ) represent random variables . Also, in this formula (1) and FIG. 7, c _i represents user information, g _j represents content information, and z _k represents latent factors.
This PLSA uses the EM algorithm or the like from user information and content information (in other words, user reaction information to content) to determine the latent factor z _k that maximizes the co-occurrence probability P(c _i , g _j ) ex post facto Ask (implicitly).

On the other hand, in the recommendation method according to the present embodiment, information corresponding to the latent factor _zk is defined in advance by the business operator, and is associated with the user and the content. This recommendation method differs from PLSA in that respect.
The method of associating the information corresponding to the latent factor _zk with the user may be, for example, by interviewing the user (specifically, questionnaire). On the other hand, the method of associating the information corresponding to the latent factor _zk with the content may be based on empirical rules, for example.

1-2. Configuration Next, an information processing system according to an embodiment will be described.
FIG. 8 shows an example of the configuration of an information processing system. An information processing system 800 shown in the figure includes a recommendation server 801 and a plurality of user terminals 802 . Each component is connected via a wired or wireless network and can transmit and receive information to and from each other.

A recommendation server 801 configuring this information processing system 800 is a server for recommending WB contents and WB services to users.
On the other hand, the user terminal 802 is a terminal device for receiving the recommendation service provided by the recommendation server 801 .

Each component of the information processing system 800 may be, for example, a portable terminal (mobile terminal) such as a smartphone, tablet, mobile phone, or personal digital assistant (PDA), or a wearable terminal such as a glasses type, a wristwatch type, or a clothing type. good. Also, each component may be a stationary or portable computer, or a server arranged on a cloud or network. Further, each component may be a VR (Virtual Reality) terminal, an AR (Augmented Reality) terminal, or an MR (Mixed Reality) terminal as a function. Alternatively, each component may be a combination of these multiple terminals. For example, a combination of one smart phone and one wearable terminal can logically function as one terminal. Further, each component device may be an information processing terminal other than these.

Each component consists of a processor that executes operating systems, applications, programs, etc., main storage devices such as RAM (Random Access Memory), and auxiliary devices such as IC cards, hard disk drives, SSDs (Solid State Drives), and flash memories. Storage device, communication control unit such as network card, wireless communication module, mobile communication module, input device such as touch panel, keyboard, mouse, voice input, input by motion detection by imaging camera unit, monitor, display, etc. and an output device. The output device may be a device or terminal that transmits information for output to an external monitor, display, projector, printer, device, or the like.

Various programs and applications (modules) are stored in the main storage device, and each functional element of the entire system is realized by the processor executing these programs and applications. Note that each of these modules may be implemented by hardware by integration or the like. Each module may be an independent program or application, or may be implemented in the form of a partial subprogram or function in one integrated program or application.

In this specification, each module is described as an entity (subject) that performs processing, but actually a processor that processes various programs and applications (modules) executes processing.

Various databases (DB) are stored in the auxiliary storage device. A "database" is a functional element (storage) that stores a set of data so as to accommodate any data manipulation (eg, extraction, addition, deletion, overwriting, etc.) from a processor or external computer. The database implementation method is not limited, and may be, for example, a database management system, spreadsheet software, or a text file such as XML or JSON.

The recommendation server 801 in particular among the constituent devices of the information processing system 800 will be described in detail below.

FIG. 9 shows an example of the configuration of the recommendation server 801. As shown in FIG. A recommendation server 801 shown in FIG. 8 is composed of, for example, one or more servers arranged on a cloud.
In the main storage device 901 of this server, modules for overall optimization include a model learning module 911, a scoring module 912, a reranking module 913, an extraction module 914, a screening module 915, a recommendation module 916, an action log management module 917, and the like. program is stored.

In the main memory 901, modules for individual optimization include a model learning module 921, a scoring module 922, a filtering module 923, a reranking module 924, an extraction module 925, a screening module 926, a recommendation module 927, and an action log management module. 928 and other programs are stored.

Further, the main storage device 901 stores programs such as a my page module 931 and a user information management module 932 .
Each functional element of the recommendation server 801 is implemented by the processor 903 executing the program described above. Each module will be described below.

The model learning module 911 learns an estimation model using WB factor information pre-associated with WB content and WB factor score information pre-associated with a user as explanatory variables and reaction probability as an objective variable. The module generates and acquires a learned estimation model through this learning.

The estimation model that this module learns is FM (Factorization Machines). FM has the following features.
• FM is a general discriminator that allows reliable parameter estimation even for very sparse data.
• FM can solve regression, classification and ranking problems.
- FM is a model that can approximately consider interactions between all feature quantities.
・FM has a device to reduce the amount of calculation internally.

This FM mathematical model is generally represented by the following equation.
... (2)
... (3)
y and x in this equation (2) are given parameters and v and w are estimated parameters.

Next, the scoring module 912 inputs WB factor information pre-associated with the WB content and WB factor score information pre-associated with the target user into the learned estimation model for each of the plurality of WB contents. to estimate the reaction probability.

The reranking module 913 excludes the WB contents that the target user has selected in the past from the WB contents whose reaction probabilities have been estimated. This is to avoid recommending previously recommended WB contents again. In addition, if there are similar WB contents with higher response probabilities, the module rearranges the order of one or more WB contents among them. This is to avoid recommending multiple similar WB contents at once.

The extraction module 914 extracts one or more WB contents from the reranked WB contents based on the reaction probability estimated for each WB content. The WB contents to be extracted are WB contents ranked within the top five in response probability.

A screening module 915 excludes inappropriate content from the extracted WB content. Inappropriate content here means WB content that is identified based on a rule base or a model base based on the user attributes of the target user.

A recommendation module 916 recommends the WB content that has passed the screening to the target user.

The action log management module 917 records the target user's action on the recommended WB content.

The model learning module 921 learns an estimation model using factor information pre-associated with a WB service and WB factor score information pre-associated with a user as explanatory variables and reaction probability as an objective variable. At that time, the module learns an estimation model for each individual region. Through this learning, the module generates and acquires a trained estimation model for each individual region. Note that the estimation model that this module learns is FM.

For each of a plurality of WB services, the scoring module 922 inputs factor information pre-associated with the WB service and WB factor score information pre-associated with the target user into the learned estimation model to determine the response. Estimate probabilities. Note that the trained estimation model used in this estimation corresponds to the individual area to which the WB service belongs.

A filtering module 923 excludes WB services that do not correspond to the WB content selected by the target user from the WB services whose response probabilities have been estimated. This is to avoid recommending a WB service unrelated to the WB content selected by the target user. A WB service that does not correspond to the WB content selected by the user is a WB service that does not share an individual area with the WB content selected by the user.

A reranking module 924 excludes WB services that the target user has selected in the past from the filtered WB services. This is to avoid re-recommending the WB service recommended in the past. In addition, if there are similar WB services with higher response probabilities, the module rearranges the order of one or more WB services among them. This is to avoid recommending multiple similar WB services at once.

An extraction module 925 extracts one or more WB services from the reranked WB services based on the estimated response probability for each WB service. The WB services to be extracted are the WB services ranked within the top five in response probability.

A screening module 926 excludes inappropriate WB services from among the extracted WB services. Inappropriate services here are WB services that are identified on a rule basis or a model basis based on the user attributes of the target user.

The recommendation module 927 recommends the WB service that has passed the screening to the target user.

The behavior log management module 928 records the target user's behavior with respect to the recommended WB service.

The my page module 931 causes the user terminal 802 to display my page (see FIG. 32) in response to the user's request.

The user information management module 932 manages user attribute information and WB factor scores.

Next, the auxiliary storage device 902 of the recommendation server 801 will be described.
The auxiliary storage device 902 stores various information. Specifically, the apparatus includes user attribute information 1000, factor score information 1100, content action log 1200, content action score information 1300, service action log 1400, service action score information 1500, content information 1600, service information 1700, activity log. It stores information such as 1800.

The device also stores various data such as screen data 941 and image data 942 .
Each piece of information will be described below.

FIG. 10 shows an example of user attribute information 1000. As shown in FIG. User attribute information 1000 shown in the figure is information in a table format indicating attribute information of each user. This user attribute information 1000 has columns of user ID 1001, date of birth 1002, sex 1003, weight 1004, height 1005, subject of interest ID 1006, and the like.

FIG. 11 shows an example of factor score information 1100. FIG. Factor score information 1100 shown in the figure is information in a table format indicating the user's WB factor score. This factor score information 1100 has columns such as user ID 1101, registration date and time 1102, factor score 1103, and the like. Among these columns, the factor score 1103 stores a value indicating the degree to which the user attaches importance to the WB factor. The stored values are, for example, values from "1" to "5".
This factor score information 1100 is generated for each user.

FIG. 12 shows an example of a content action log 1200. As shown in FIG. A content action log 1200 shown in the figure is information in a table format indicating a user's action history with respect to recommended WB content. This content action log 1200 has columns of user ID 1201, content ID 1202, display position 1203, action score 1204, registration date and time 1205, and the like. Among these columns, the display position 1203 stores a value indicating the order in which the WB content is displayed on the recommendation screen (see FIG. 29). Also, the action score 1204 stores a value indicating the probability of the user's reaction to the WB content. Specifically, the action score 1204 stores a value of "1" indicating that the user has selected the WB content, or a value of "0" indicating that the user has not selected the WB content. .

This content action log 1200 is accumulated by randomly distributing WB content to users at the beginning of the service. When the action log is accumulated to a certain extent and model learning becomes possible, the WB contents estimated by the learned model are distributed and accumulated.

FIG. 13 shows an example of content action score information 1300. As shown in FIG. The content action score information 1300 shown in FIG. 13 is information in a table format indicating action scores recorded or estimated for each WB content. In this information, the WB contents are sorted in descending order of action score, and the information indicates the ranking of the WB contents.

This content action score information 1300 has columns of user ID 1301, content ID 1302, action score 1303, and the like. Among these columns, the action score 1303 stores a value indicating the user's reaction probability to the WB content. The stored values are, for example, values from "0" to "1".
This content action score information 1300 is generated for each user.

FIG. 14 shows an example of a service behavior log 1400. As shown in FIG. A service action log 1400 shown in the figure is information in a table format indicating the user's action history with respect to the recommended WB service. This service action log 1400 has columns of user ID 1401, service ID 1402, display position 1403, action score 1404, registration date and time 1405, and the like. Among these columns, the display position 1403 stores a value indicating the order in which the WB service is displayed on the recommendation screen (see FIG. 30). Also, the action score 1404 stores a value indicating the user's reaction probability to the WB service. Specifically, the action score 1404 stores a value of "1" indicating that the user has selected the WB service, or a value of "0" that indicates that the user has not selected the WB service. .

This service action log 1400 is accumulated by randomly distributing the WB service to users at the beginning of the service. Then, when the action log is accumulated to a certain extent and model learning becomes possible, the WB service estimated by the learned model is distributed and accumulated.

FIG. 15 shows an example of service behavior score information 1500 . Service action score information 1500 shown in the figure is information in a table format indicating action scores recorded or estimated for each WB service. In this information, the WB services are sorted in descending order of action score, and the information indicates the ranking of the WB services.

This service action score information 1500 has columns of user ID 1501, service ID 1502, action score 1503, and the like. Among these columns, the action score 1503 stores a value indicating the user's reaction probability to the WB service. The stored values are, for example, values from "0" to "1".
This service behavior score information 1500 is generated for each user.

FIG. 16 shows an example of content information 1600. As shown in FIG. The content information 1600 shown in FIG. 16 is table format information indicating information about each WB content. This content information 1600 has columns of content ID 1601, content 1602, image ID 1603, factor ID 1604, individual area ID 1605, and the like. Among these columns, the image ID 1603 stores the ID of the WB content image (see FIG. 29). The factor ID 1604 stores the ID of the WB factor pre-associated with the WB content. The individual area ID 1605 stores the ID of the individual area associated with the WB content in advance. Here, an individual area is a category to which WB contents or WB services belong. This category includes "sleep", "eat", "play", and the like.

FIG. 17 shows an example of service information 1700 . Service information 1700 shown in the figure is information in a table format indicating information on each WB service. This service information 1700 has columns of service ID 1701, content 1702, image ID 1703, factor ID 1704, individual area ID 1705, and the like. Of these columns, the image ID 1703 stores the ID of the WB service image (see FIG. 30). The factor ID 1704 stores the ID of the WB factor pre-associated with the WB service. The individual area ID 1705 stores the ID of the individual area associated with the WB service in advance.

Here, a method for associating WB factors with WB services will be described.
As a first method, there is a method of associating the WB factor of the WB content corresponding to the WB service with the WB service. Here, the WB content corresponding to the WB service is the WB content that shares an individual area with the WB service.
As a second method, there is a method of picking up the WB factors of the WB contents corresponding to the WB service and manually associating particularly important ones among them. This second method has an advantage over the first method in that WB factors unrelated to the WB service can be reliably eliminated.

Next, FIG. 18 shows an example of an activity log 1800. FIG. An activity log 1800 shown in FIG. 18 is information in a table format showing daily activity records of the user. This activity log 1800 has columns such as user ID 1801, registration date and time 1802, WB degree 1803, memo 1804, and the like. Among these columns, the WB degree 1803 stores a value indicating the user's degree of happiness or satisfaction. A memo 1804 stores information recorded as a diary by the user.

1-3. Operation Next, the operation of the information processing system 800 will be described.
1-3-1. Recommendation Processing FIG. 19 shows an example of the flow of recommendation processing.
First, the user performs pre-registration with the recommendation server 801 before receiving the recommendation service. At that time, the user information management module 932 of the recommendation server 801 causes the user terminal 802 to display a registration screen.

26 to 28 show examples of registration screens. A registration screen 2600 shown in FIG. 26 is a screen for registering a user profile. This screen has a date of birth selection field 2601 , gender selection field 2602 , weight selection field 2603 , height selection field 2604 and a continue button 2605 . When the user selects the date of birth and the like on this screen and selects the continue button 2605, the registration screen 2700 shown in FIG. 27 is displayed next.

The registration screen 2700 is a screen for registering the user's interest. This screen has multiple selection buttons 2701 for selecting subjects of interest and a continue button 2702 . When the user selects the selection button 2701 of his/her interest on this screen and selects the continue button 2702, the registration screen 2800 shown in FIG. 28 is displayed next.

A registration screen 2800 is a questionnaire screen regarding WB factors. This questionnaire screen prompts the user to answer the importance of each WB factor in five levels. The degree of importance answered by the user is registered in the recommendation server 801 as a WB factor score.

This registration screen 2800 has a plurality of answer selection fields 2801 provided for each question and a continue button 2802 . When the user selects an answer to the question on this screen and selects the continue button 2802, the next questionnaire screen (not shown) is displayed.

User registration is completed when the user answers all the questions about the WB factors and selects a send button (not shown).
The questions presented to the user may include questions about attributes of the user other than questions about WB factors. For example, the questions may include questions about the user's diet or illness.

Information selected by the user is registered in the recommendation server 801 . At this time, the user information management module 932 of the recommendation server 801 adds the profile information selected on the registration screen 2600 to the user attribute information 1000 in association with the user ID (step 1901). Here, the profile information is date of birth, sex, weight and height.

The module also associates the ID of the target of interest selected on the registration screen 2700 with the user ID and adds it to the user attribute information 1000 (step 1901).
Also, this module adds the WB factor score selected on the registration screen 2800 to the factor score information 1100 in association with the user ID and the date and time of registration (step 1902).
The above is the description of the user registration.

Next, the action score updating process will be described.
This action score update process is a batch process that is performed periodically. This update process is executed at a frequency such as daily, weekly, or monthly.
Also, this update process is executed for each of the WB content and the WB service. Each update process involves learning an estimation model and estimating an action score based on the learned estimation model. First, the update processing executed for the WB content will be described below.

The model learning module 911 of the recommendation server 801 learns the overall optimum model based on the content action log 1200 and the like (step 1903). At this time, the module learns a global optimum model using each user's attribute information, the initial WB factor score, and the factor information of each WB content as explanatory variables, and the action score of each WB content as an objective variable. The module generates a trained global optimal model as a result of this learning.

FIG. 22 shows an example of input data to be learned by the global optimum model. The input data 2200 shown in the figure employs a data format for FM.
This input data 2200 has columns such as user 2201, item 2202, U_WB 2203, I_WB 2204, and Y 2205. Each column will be described below.

User 2201 stores a flag for identifying the user. For example, a record in which flag "1" is stored in the user1 column is the action log of user1.
Item 2202 stores a flag for identifying the WB content. For example, a record in which flag "1" is stored in the item1 column is an action log for WB content 1. FIG.

U_WB 2203 stores the user's WB factor score. Each WB factor score represents the importance of the WB factor for the user, and is expressed in the following five levels.
1: I don't think so at all 2: I don't think much 3: I can't say either way 4: I agree 5: I strongly agree

I_WB 2204 stores flags for identifying WB factors associated with WB contents. For example, when a flag "1" is stored in the item1 column and a flag "1" is stored in the I_WB1 column for a certain record, it indicates that WB content 1 is associated with WB factor 1. ing.
A plurality of WB factors may be associated with one WB content.

Y2205 stores a flag indicating an action score. The stored flags include a flag "1" indicating that the WB content has been selected and a flag "0" indicating that the WB content has not been selected.

Note that the input data 2200 shown in FIG. 22 does not include a user attribute column. However, the input data 2200 may include a user attribute column. As a result, the attribute information of each user can be used as explanatory variables to be learned by the global optimum model.

Next, the behavior score estimation using the learned global optimal model will be described.
The scoring module 912 of the recommendation server 801 executes scoring processing using the generated trained global optimum model (step 1904). This scoring process is performed for each user.

FIG. 20 shows an example of this scoring process. In the scoring process 2000 shown in the figure, the module adds the action score of the WB content recommended to the user to the content action score information 1300 (step 2001). Here, the WB content already recommended to the user is the WB content associated with the user in the content action log 1200 . That is, the module identifies the WB content associated with the user in the content action log 1200 and adds the action score of the identified WB content to the content action score information 1300 .

Next, the module estimates the action score of the WB content that has not been recommended to the user by using the learned global optimum model (step 2002). Here, the WB content that has not yet been recommended to the user is the WB content that is included in the content information 1600 and that is not associated with the user in the content action log 1200 . After identifying unrecommended WB content, the module inputs factor information of the identified WB content, attribute information of the target user, and initial WB factor score into the learned global optimum model to estimate the action score. The module adds the estimated behavioral score to the content behavioral score information 1300 .

FIG. 23 shows an example of input data to be input to the trained global optimal model. The input data 2300 shown in the figure employs a data format for FM.
This input data 2300 has columns such as user 2301, item 2302, U_WB 2303, I_WB 2304, and Y 2305. Since the attributes of these columns are common to the columns with the same names in the input data 2200 described above, description thereof will be omitted. Since this input data 2300 is data for estimating the action score, Y2305 does not store any flags.

Incidentally, the input data 2300 shown in FIG. 23 does not include a user attribute column. However, this input data 2300 may include a user attribute column. As a result, the behavior score can be estimated using the target user's attribute information as an explanatory variable.

After estimating the action score, the scoring module 912 sorts the WB contents of the content action score information 1300 in descending order of the action score (step 2003). As a result, the ranking information of the WB contents arranged in descending order of the response probability of the target user is generated.

Next, update processing executed for the WB service will be described.
The model learning module 921 of the recommendation server 801 learns an individual optimum model based on the service behavior log 1400 (step 1905). At that time, the module learns an individual optimal model for each individual region and generates a trained individual optimal model for each individual region. Specifically, for each individual area, the module uses attribute information and initial WB factor scores of each user and factor information of each WB content belonging to the individual area as explanatory variables, is used as the objective variable to learn the individual optimal model. As a result of this learning, the module generates a trained individual optimal model for each individual region.

FIG. 24 shows an example of input data to be learned by the individual optimum model. The input data 2400 shown in the figure employs a data format for FM.
This input data 2400 has columns such as user 2401, item 2402, U_WB 2403, I_WB 2404, and Y 2405. Each column will be described below.

User 2401 stores a flag for identifying the user. For example, a record in which flag "1" is stored in the user1 column is the action log of user1.
Item 2402 stores a flag for identifying the WB service. For example, a record in which a flag "1" is stored in the item1 column is an action log for the WB service1.

U_WB 2403 stores the user's WB factor score. Each WB factor score represents the importance of the WB factor for the user, and is expressed in the following five levels.
1: I don't think so at all 2: I don't think much 3: I can't say either way 4: I agree 5: I strongly agree

I_WB 2404 stores flags for identifying WB factors associated with WB services. For example, if flag "1" is stored in the item1 column and flag "1" is stored in the I_WB1 column for a certain record, it indicates that WB service 1 is associated with WB factor 1. ing.
A plurality of WB factors may be associated with one WB service.

Y2405 stores a flag indicating an action score. The stored flags include a flag "1" indicating that the WB service has been selected and a flag "0" indicating that the WB service has not been selected.

Note that the input data 2400 shown in FIG. 24 does not include a user attribute column. However, the input data 2400 may include a user attribute column. As a result, the attribute information of each user can be used as an explanatory variable to be learned by the individual optimum model.

Next, the behavior score estimation using the trained individual optimum model will be described.
The scoring module 922 of the recommendation server 801 executes scoring processing using the generated trained individual optimum model (step 1906). This scoring process is performed for each user.

FIG. 21 shows an example of this scoring process. In the scoring process 2100 shown in the figure, the module adds the action score of the WB service recommended to the user to the service action score information 1500 (step 2101). Here, the WB service recommended to the user is the WB service associated with the user in the service behavior log 1400 . That is, the module identifies the WB service associated with the user in the service behavior log 1400 and adds the behavior score of the identified WB service to the service behavior score information 1500 .

Next, the module estimates the action score of the WB service that has not yet been recommended to the user using the learned individual optimum model (step 2102). Here, a WB service that has not been recommended to a user is a WB service that is included in the service information 1700 and that is not associated with the user in the service behavior log 1400 . When a WB service that has not yet been recommended is specified, the module inputs the factor information of the specified WB service, the attribute information of the target user, and the initial WB factor score into the trained individual optimum model, and estimates the action score. The trained individual optimum model used in this estimation corresponds to the individual area to which the specified WB service belongs.
The module adds the estimated behavioral score to the service behavioral score information 1500 .

FIG. 25 shows an example of input data to be input to the trained individual optimum model. The input data 2500 shown in the figure employs a data format for FM.
This input data 2500 has columns such as user 2501, item 2502, U_WB 2503, I_WB 2504, and Y 2505. Since the attributes of these columns are common to the columns with the same names in the input data 2400 described above, description thereof will be omitted. Since this input data 2500 is data for estimating the action score, Y2505 does not store any flags.

Incidentally, the input data 2500 shown in FIG. 25 does not include a user attribute column. However, this input data 2500 may include a user attribute column. As a result, the behavior score can be estimated using the target user's attribute information as an explanatory variable.

After estimating the action scores, the scoring module 922 sorts the WB services in the service action score information 1500 in descending order of action scores (step 2103). As a result, the ranking information of the WB services arranged in descending order of response probability of the target user is generated.
The above is the description of the behavior score update processing.

Next, a recommendation flow for WB contents will be described. This recommendation flow starts when the user accesses a site provided by the recommendation server 801 .

First, the reranking module 913 of the recommendation server 801 excludes the WB content selected by the user in the past from the content action score information 1300 of the target user (step 1907). This is to avoid recommending previously recommended WB contents again. The WB content selected by the user in the past is the WB content associated with the user and the action score "1" in the content action log 1200 .

Also, if there are similar WB contents with high action scores, the module rearranges the order of one or more WB contents among them (step 1907). This is to avoid recommending multiple similar WB contents at once. Note that the WB content with the highest action score is, for example, the WB content within the top five ranks. Further, similar WB content is, for example, WB content having a common individual area. Also, rearranging the order means, for example, demoting one of two similar WB contents to the sixth or lower rank.

Next, the extraction module 914 of the recommendation server 801 extracts the top five WB contents from the reranked contents action score information 1300 (step 1908). Note that the number of WB contents to be extracted may be changed as appropriate.

Next, the screening module 915 of the recommendation server 801 excludes inappropriate content from the extracted WB contents (step 1909). Inappropriate content here means WB content specified on a rule basis or a model basis based on the user attribute information 1000 of the target user.

Next, the recommendation module 916 of the recommendation server 801 recommends the WB content that has passed the screening to the target user (step 1910). At that time, the module refers to the content information 1600 to generate a recommendation screen and causes the user terminal 802 to display it. FIG. 29 shows an example of a recommendation screen.

A recommendation screen 2900 shown in the figure presents five WB contents 2901 . Five WB contents 2901 are arranged in the order of action scores, and each is composed of a content 2902 of the WB contents and an image 2903 . By referring to this recommendation screen 2900, the user can obtain advice for improving his/her own well-being.

When the user selects WB content 2901 of particular interest on this recommendation screen 2900, the WB service recommendation flow is started. Further, along with the selection, the action log management module 917 of the recommendation server 801 adds the information of the selected WB content 2901 to the content action log 1200 in association with the user ID and the action score "1" (step 1911).

Also, along with the selection, the user information management module 932 updates the factor score information 1100 of the target user (not shown). Specifically, this module identifies the WB factor of the selected WB content 2901 and increments the score of the identified WB factor by, for example, "0.1". Thus, the user's WB factor score changes according to the WB content 2901 selected.

Returning to the description of the recommendation screen 2900 .
The user can select multiple WB contents 2901 on the recommendation screen 2900 . WB content 2901 that is not selected by the user for a certain period of time on this screen is registered in the content action log 1200 (step 1911). At this time, the action log management module 917 adds the information of the WB content 2901 that has not been selected to the content action log 1200 in association with the user ID and the action score "0".
The above is the description of the WB content recommendation flow.

Next, the recommendation flow of the WB service will be described. This recommendation flow is started by selecting the WB content 2901 on the recommendation screen 2900 illustrated in FIG.

First, the filtering module 923 of the recommendation server 801 excludes WB services that do not correspond to the WB content selected by the user from the target user's service behavior score information 1500 (step 1912). This is to avoid recommending a WB service unrelated to the WB content selected by the target user. A WB service that does not correspond to the WB content selected by the user is a WB service that does not share an individual area with the WB content selected by the user. Such WB service identification can be performed by referring to the content information 1600 and service information 1700 .

Note that the filtering module 923 may prevent the recommendation of the WB service by setting the action score of the WB service to "0" instead of excluding the WB service that does not correspond to the WB content selected by the user. .

Next, the reranking module 924 of the recommendation server 801 excludes the WB services selected by the user in the past from the filtered service action score information 1500 (step 1913). This is to avoid re-recommending the WB service recommended in the past. The WB service selected by the user in the past is the WB service associated with the user and the action score “1” in the service action log 1400 .

Also, if there are similar WB services with high action scores, the module rearranges the order of one or more WB services among them (step 1913). This is to avoid recommending multiple similar WB services at once. Note that the WB services with the highest action scores are, for example, the WB services ranked within the top five. A similar WB service is, for example, a WB service having a common individual area. Also, rearranging the order means, for example, demoting one of two similar WB services to sixth or lower.

Next, the extraction module 925 of the recommendation server 801 extracts the top five WB services from the reranked service behavior score information 1500 (step 1914). Note that the number of WB services to be extracted may be changed as appropriate.

Next, the screening module 926 of the recommendation server 801 excludes inappropriate WB services from the extracted WB services (step 1915). Inappropriate services here refer to WB services specified on a rule basis or a model basis based on the user attribute information 1000 of the target user.

Next, the recommendation module 927 of the recommendation server 801 recommends the WB service that has passed the screening to the target user (step 1916). At that time, the module refers to the service information 1700 to generate a recommendation screen and causes the user terminal 802 to display it. FIG. 30 shows an example of this recommendation screen.

A recommendation screen 3000 shown in the figure presents five WB services 3001 . Five WB services 3001 are arranged in the order of action score, and each is composed of a WB service content 3002 and an image 3003 . By referring to this recommendation screen 3000, the user can know services for improving his/her own well-being.

When the user selects a WB service 3001 of particular interest on this recommendation screen 3000, the recommendation module 927 causes the user terminal 802 to display a detailed screen. FIG. 31 shows an example of this detail screen.

A detailed screen 3100 shown in the figure has a detailed explanation column 3101 and a try button 3102 for the WB service 3001 . When the user selects a try button 3102 on this screen, a transition is made to the site of the company that provides the WB service 3001 .
Note that the recommendation module 927 may directly transition from the recommendation screen 3000 to the provider's site without displaying the detailed screen 3100 .

As the WB service 3001 is selected on the recommendation screen 3000, the action log management module 928 of the recommendation server 801 associates the information of the selected WB service 3001 with the user ID and the action score "1", and displays the service action. Add to log 1400 (step 1917).

The user can select multiple WB services 3001 on the recommendation screen 3000 . A WB service 3001 that is not selected by the user for a certain period of time on this screen is registered in the service action log 1400 (step 1917). At this time, the action log management module 928 adds the information of the unselected WB service 3001 to the service action log 1400 in association with the user ID and the action score "0".
The above is the description of the recommendation flow of the WB service.

According to the recommendation processing described above, first, overall domain optimization using WB contents is attempted, and then individual domain optimization using WB services is attempted. This enables cross-domain recommendations.

1-3-2. My Page Display Processing Next, the my page display processing will be described.
Here, my page is a screen for displaying and recording the user's daily activities. The my page module 931 of the recommendation server 801 displays my page on the user terminal 802 in response to the user's request. At that time, the module generates my page based on the target user's factor score information 1100 and activity log 1800 .

FIG. 32 shows an example of My Page. My page 3200 shown in the figure has a content confirmation button 3201 , a factor score display area 3202 , a calendar 3203 , a my page button 3204 and a recommendation button 3205 . Each display element will be described below.

The content confirmation button 3201 is a button for displaying a WB content recommendation screen (see FIG. 29).

A factor score display area 3202 is an area where the user's WB factor score is displayed. Category selection tabs 3206-3208, value selection buttons 3209-3211 and radar chart 3212 are displayed in this area. Among these display elements, category selection tabs 3206 to 3208 are tabs for selecting the category of WB factors to be displayed on the radar chart 3212 from "wellness", "newness", and "community". Value selection buttons 3209 to 3211 select the WB factor values displayed on the radar chart 3212 from the current value (in other words, the latest value), the initial questionnaire value (in other words, the first registered value), and the average value. button for selecting from
By referring to this factor score display area 3202, the user can confirm the balance and change of his own WB factors.

A calendar 3203 indicates the user's WB degree on a daily basis. A user's WB degree is represented by a face mark 3213 .
My page button 3204 is a button for displaying my page 3200 .
The recommendation button 3205 is a button for displaying a WB content recommendation screen (see FIG. 29).

When the user selects any date in the calendar 3203, the my page module 931 causes the user terminal 802 to display a detailed screen. At that time, the module generates a detailed screen based on the target user's activity log 1800 , service behavior log 1400 and service information 1700 . FIG. 33 shows an example of a generated detail screen.

A detailed screen 3300 shown in the same figure has a slider 3301 indicating the degree of WB, a memo display field 3302 and a selected service display field 3303 . By referring to this screen, the user can confirm the WB degree, diary, and selected WB service on the selected date.

The contents of this detail screen 3300 can be edited on the edit screen. FIG. 34 shows an example of an edit screen.

An edit screen 3400 shown in the figure has a slider 3401 for designating the WB degree, a memo entry field 3402 and a record button 3403 . When the record button 3403 is selected, the my page module 931 adds the WB degree specified on this screen and the input memo to the activity log 1800 in association with the user ID.

2. Modifications The above embodiment may be modified as follows.
(1) In the above examples, the WB factor score is evaluated in five stages. However, this five-level evaluation is not essential. For example, the WB factor score may be evaluated in 2 stages or 10 stages.

(2) In the above embodiment, the recommended WB content or WB service is evaluated in two stages of "0" or "1" (see step 1911 in FIG. 19). However, this two-stage evaluation is not essential. For example, the user may evaluate the WB content or WB service on a scale of 5 and record the evaluation value as an action log. The evaluation value recorded here corresponds to the reaction probability indicating the user's reaction to the WB content or WB service.

In addition, evaluation of five steps is expressed as follows as an example.
1: Not interested at all 2: Not very interested 3: Neither 4: Somewhat like 5: Very much like

(3) In the model learning of the above embodiment, user attribute information is learned by the estimation model (see steps 1903 and 1905 in FIG. 19). However, learning of user attribute information is not essential and may be omitted.

(4) In the above example, FM is used as the estimation model (see steps 1903 and 1905 in FIG. 19). However, the use of FM is not mandatory. Regression models other than FM may be employed in the above example.

(5) The model learning module 921 learns an individual optimal model for each individual region. However, this learning method is not required. The model learning module 921 may be made to learn individual optimal models across individual regions. In this case, a single trained individual optimal model will be generated.

(6) In the above recommendation process, the recommendation server 801 outputs the WB content that passed the screening to the user terminal 802 (see step 1910 in FIG. 19). However, the server may output the information at a pre-screening stage in order to cause another device to perform the output. For example, the server may output any one of the content action score information 1300, the content action score information 1300 reranked in step 1907, and the WB content extracted in step 1908 to another device.

Similarly, in the pre-screening stage, the server stores the service behavior score information 1500, the service behavior score information 1500 filtered in step 1912, the service behavior score information 1500 reranked in step 1913, and the WB extracted in step 1914. Any of the services may be output to other devices.

(7) In the above recommendation processing, scoring processing for overall optimization is performed periodically (see step 1904 in FIG. 19). However, this scoring process may be executed at the timing when the user accesses the site.

Similarly, the scoring process for individual optimization may be executed at the timing when the user selects the WB content. In this case, the recommendation server 801 may generate the service action score information 1500 only for the WB service corresponding to the WB content selected by the user. Here, a WB service corresponding to WB content is a WB service that shares an individual area with WB content. By doing so, the filtering process in step 1912 can be omitted.

(8) The recommendation server 801 performs both modeling and scoring on the server (steps 1903-1906 in FIG. 19). However, modeling is not required. The server may cause another device to perform the modeling and obtain the generated trained model from the device.

(9) In the above examples, the WB factor is defined in advance as a latent factor on the business side from the viewpoint of well-being. At that time, the business operator may define the WB factor by the following method.
First, the operator defines temporary WB factors. Next, the provider conducts a questionnaire to the user regarding the defined temporary WB factors. Next, the operator performs factor analysis on the questionnaire results and extracts latent factors. Finally, the operator defines the extracted latent factors as formal WB factors.
An operator may define the WB factor in this manner.

(10) In the recommendation process described above, when a recommended WB content is selected by a user, the user's factor score information 1100 is updated. The updated WB factor scores may then be reflected in subsequent modeling and scoring. Specifically, the model learning module 911 may update the learned global optimum model by having the global optimum model learn the updated WB factor score as an explanatory variable. The scoring module 912 may also input the updated WB factor scores into the trained global optimal model to estimate response probabilities.

Similarly, the model learning module 921 may update the learned individual optimum model by having the individual optimum model learn the updated WB factor score as an explanatory variable. The scoring module 922 may also input the updated WB factor scores into the trained individual optimal model to estimate response probabilities.

As a result of these processes, there is a possibility that the WB content and WB service that will be recommended to the target user next time will change.

In addition, the present invention is not limited to the above-described embodiments, and includes various modifications. For example, the above-described embodiments have been described in detail in order to explain the present invention in an easy-to-understand manner, and are not necessarily limited to those having all the described configurations. In addition, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Moreover, it is possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.

Further, each of the above configurations, functions, processing units, processing means, and the like may be realized by hardware, for example, by designing them in an integrated circuit. Moreover, each of the above configurations, functions, etc. may be realized by software by a processor interpreting and executing a program for realizing each function. Information such as programs, tables, and files that implement each function can be stored in recording devices such as memories, hard disks, SSDs (Solid State Drives), or recording media such as IC cards, SD cards, and DVDs.

Further, the control lines and information lines indicate those considered necessary for explanation, and not all control lines and information lines are necessarily indicated on the product. In practice, it may be considered that almost all configurations are interconnected.
It should be noted that the above embodiments disclose at least the configurations described below .
(1)
A first learned estimation generated by learning a first estimation model using factor information associated with content and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. a first model acquisition unit that acquires a model;
For each of a plurality of target contents, factor information associated with the target content and factor score information associated with the target user are input to the first trained estimation model to estimate the reaction probability. 1 estimator;
a first extraction unit that extracts one or more target contents from among the plurality of target contents based on the reaction probability estimated for each of the plurality of target contents;
A recommendation server with
(2)
A second learned estimation generated by learning a second estimation model using factor information associated with a service and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. a second model acquisition unit that acquires a model;
For each of a plurality of target services, factor information associated with the target service and factor score information associated with the target user are input to the second trained estimation model to estimate the reaction probability. a second estimation unit;
a second extraction unit for extracting one or more target contents associated with the one or more target contents from among the plurality of target services based on the reaction probability estimated for each of the plurality of target services; and
The recommendation server according to (1), further comprising:
(3)
The recommendation server according to (1) or (2), wherein the factor is a factor defined as expressing a potential psychological state of the user.
(4)
The recommendation server according to (1) or (2), wherein the content is advice information defined as improving the user's happiness.
(5)
The recommendation server according to (1) or (2), wherein the factor score information associated with the target user is information indicating the degree to which the target user attaches importance to the factor.
(6)
The second model acquisition unit acquires a second trained estimation model for each individual region to which the service belongs,
For each of the plurality of target services, the second estimation unit calculates factor information associated with the target service and factor score information associated with the target user for an individual region to which the target service belongs. Input to the obtained second trained estimation model to estimate the reaction probability
The recommendation server according to (2), characterized by:
(7)
When one of the one or more target contents is selected by the target user, factor score information associated with the target user, the score of the factor associated with the selected target content. The recommendation server according to (1) or (2), further comprising an update unit that updates information.
(8)
The first model acquisition unit updates the first trained estimation model by causing the first estimation model to learn the updated factor score information as an explanatory variable, The recommendation server described in 7).
(9)
The first estimation unit further inputs factor information associated with the target service and the updated factor score information into the first trained estimation model for each of a plurality of target services, The recommendation server according to (7), characterized by estimating a reaction probability.
(10)
A computer-implemented recommendation method comprising:
A first learned estimation generated by learning a first estimation model using factor information associated with content and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. obtaining a model;
A step of inputting factor information associated with the target content and factor score information associated with the target user into the first pre-trained estimation model for each of a plurality of target contents, and estimating the reaction probability. and,
extracting one or more target contents from among the plurality of target contents based on the reaction probability estimated for each of the plurality of target contents;
A recommendation method with
(11)
A second learned estimation generated by learning a second estimation model using factor information associated with a service and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. obtaining a model;
For each of a plurality of target services, factor information associated with the target service and factor score information associated with the target user are input to the second trained estimation model to estimate the reaction probability. a step;
extracting one or more target contents associated with the one or more target contents from among the plurality of target services based on the reaction probability estimated for each of the plurality of target services;
Further having, the recommendation method according to (10).
(12)
The recommendation method according to (10) or (11), wherein the factor is a factor defined as improving the user's happiness level.
(13)
The recommendation method according to (10) or (11), wherein the content is advice information defined as improving the user's happiness.
(14)
The recommendation method according to (10) or (11), wherein the factor score information associated with the target user is information indicating the degree to which the target user attaches importance to the factor.
(15)
The step of acquiring the second trained estimation model is a step of acquiring the second trained estimation model for each individual region to which the service belongs,
In the step of estimating the reaction probability, for each of the plurality of target services, factor information associated with the target service and factor score information associated with the target user are obtained in an individual region to which the target service belongs. estimating the reaction probability by inputting to the second trained estimation model generated for
The recommendation method according to (11), characterized by:
(16)
When one of the one or more target contents is selected by the target user, factor score information associated with the target user, the score of the factor associated with the selected target content. The recommendation method according to (10) or (11), further comprising a step of updating information.
(17)
The method according to (16), further comprising updating the first trained estimation model by having the first estimation model learn the updated factor score information as an explanatory variable. Recommendation method.
(18)
inputting the factor information associated with the target service and the updated factor score information into the first trained estimation model for each of a plurality of target services, and estimating the reaction probability; The recommendation method according to (16), characterized by:
(19)
A program for causing a computer to function as each part described in (1) or (2).

800... Information processing system, 801... Recommendation server, 802... User terminal, 911... Model learning module, 912... Scoring module, 913... Reranking module, 914... Extraction module, 915... Screening module, 916... Recommendation module, 917 Behavior log management module 921 Model learning module 922 Scoring module 923 Filtering module 924 Reranking module 925 Extraction module 926 Screening module 927 Recommendation module 928 Behavior log management module , 931... My page module, 932... User information management module, 941... Screen data, 942... Image image data, 1000... User attribute information, 1100... Factor score information, 1200... Content action log, 1300... Content action score information, 1400... Service action log 1500... Service action score information 1600... Content information 1700... Service information 1800... Activity log

Claims (15)

A first learned estimation generated by learning a first estimation model using factor information associated with content and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. a first model acquisition unit that acquires a model;
For each of a plurality of target contents, factor information associated with the target content and factor score information associated with the target user are input to the first trained estimation model to estimate the reaction probability. 1 estimator;
a first extraction unit that extracts one or more target contents from among the plurality of target contents based on the reaction probability estimated for each of the plurality of target contents ;
The factor is a factor defined as expressing a potential psychological state of the user,
The factor score information associated with the target user is information indicating the degree to which the target user places importance on the factor.
A recommendation server characterized by : A second learned estimation generated by learning a second estimation model using factor information associated with a service and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. a second model acquisition unit that acquires a model;
For each of a plurality of target services, factor information associated with the target service and factor score information associated with the target user are input to the second trained estimation model to estimate the reaction probability. a second estimation unit;
a second extraction unit for extracting one or more target contents associated with the one or more target contents from among the plurality of target services based on the reaction probability estimated for each of the plurality of target services; The recommendation server of claim 1, further comprising: and. 3. The recommendation server according to claim 1, wherein said content is advice information defined as improving the user's happiness. The second model acquisition unit acquires a second trained estimation model for each individual region to which the service belongs,
For each of the plurality of target services, the second estimation unit calculates factor information associated with the target service and factor score information associated with the target user for an individual region to which the target service belongs. 3. The recommendation server according to claim 2, wherein the response probability is estimated by inputting to the obtained second trained estimation model. When one of the one or more target contents is selected by the target user, factor score information associated with the target user, the score of the factor associated with the selected target content. 3. The recommendation server according to claim 1, further comprising an update unit that updates information. The first model acquisition unit updates the first trained estimation model by causing the first estimation model to learn the updated factor score information as an explanatory variable. Item 5. The recommendation server according to Item 5 . The first estimation unit further inputs factor information associated with the target service and the updated factor score information into the first trained estimation model for each of a plurality of target services, 6. The recommendation server according to claim 5 , which estimates reaction probability. A computer-implemented recommendation method comprising:
A first learned estimation generated by learning a first estimation model using factor information associated with content and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. obtaining a model;
A step of inputting factor information associated with the target content and factor score information associated with the target user into the first pre-trained estimation model for each of a plurality of target contents, and estimating the reaction probability. and,
extracting one or more target contents from among the plurality of target contents based on the reaction probability estimated for each of the plurality of target contents ;
The factor is a factor defined as improving the user's happiness,
The factor score information associated with the target user is information indicating the degree to which the target user places importance on the factor.
A recommendation method characterized by : A second learned estimation generated by learning a second estimation model using factor information associated with a service and factor score information associated with a user as explanatory variables and reaction probability as an objective variable. obtaining a model;
For each of a plurality of target services, factor information associated with the target service and factor score information associated with the target user are input to the second trained estimation model to estimate the reaction probability. a step;
extracting one or more target contents associated with the one or more target contents from among the plurality of target services based on the reaction probability estimated for each of the plurality of target services. , the recommendation method according to claim 8 . 10. The recommendation method according to claim 8 , wherein said content is advice information defined as improving the user's happiness. The step of acquiring the second trained estimation model is a step of acquiring the second trained estimation model for each individual region to which the service belongs,
In the step of estimating the reaction probability, for each of the plurality of target services, factor information associated with the target service and factor score information associated with the target user are obtained in an individual region to which the target service belongs. 10. The recommendation method according to claim 9 , wherein the step of estimating the reaction probability by inputting to the second trained estimation model generated for. When one of the one or more target contents is selected by the target user, factor score information associated with the target user, the score of the factor associated with the selected target content. 10. The recommendation method according to claim 8 , further comprising a step of updating information. 13. The method according to claim 12 , further comprising updating the first trained estimation model by having the first estimation model learn the updated factor score information as an explanatory variable. Recommendation method. inputting the factor information associated with the target service and the updated factor score information into the first trained estimation model for each of a plurality of target services, and estimating the reaction probability; 13. The recommendation method according to claim 12 , characterized by: A program for causing a computer to function as each unit according to claim 1 or 2.

Images