JP7330726B2 - MODEL GENERATING DEVICE, MODEL GENERATING METHOD, AND PROGRAM - Google Patents

Description

The present invention relates to a model generation device, a model generation method, and a program.

Conventionally, from the training data, multiple sets are sampled, each containing a predetermined number of elements classified into the minority class and a number determined based on the predetermined number of elements classified into the majority class. A data classification technique comprising a sampling unit and a learning unit that machine-learns each of a plurality of models for classifying the input data into the plurality of classes based on each of the plurality of sampled sets. disclosed (see Patent Document 1).

JP 2009-122851 A

However, with the above-described conventional technology, it may not be possible to switch the prediction item optimized for the index to be improved among the items to be learned or output by the model.

The present invention has been made in consideration of such circumstances. One of the purposes is to provide a generation method and a program.

One aspect of the present invention is an action history of a user, which includes content-related information including at least identification information of the user, location of the user, and information indicating content provided to the user, and details of the content. a learning unit that learns a detailed information DB containing information as learning data; and a learned part that receives a point of interest containing at least the content in the learning data and a correct answer, and outputs a correct answer when information corresponding to the point of interest is input. and a model generation unit that generates a model.

According to one aspect of the present invention, it is possible to generate a trained model that outputs a correct answer when information corresponding to a point of interest is input.

FIG. 2 is a schematic diagram showing a usage environment of the model generating device 100; The figure which shows an example of the content relevant information 210. FIG. The figure which shows an example of the detailed information 220. FIG. FIG. 4 is a diagram for explaining the processing of the model generating device 100; The figure which shows an example of the view point setting 136. FIG. FIG. 4 is a diagram for explaining an output example of a recommendation model M; FIG. 11 is a diagram for explaining another output example of the recommendation model M; The figure which shows another example of the detailed information 220. FIG. 4 is a flowchart showing an example of the flow of processing by the model generating device 100;

Embodiments of a model generation device, a model generation method, and a program according to the present invention will be described below with reference to the drawings.

[overview]
A model generator is implemented by one or more processors. A model generation device is a device for generating a trained model that outputs a correct answer when information corresponding to a point of interest is input.

The model generation device uses content-related information and a detailed information DB including detailed information about content as learning data. A content is an information service provided via a website or an application for information service provided via a network, and includes, for example, a map information service, a town information service, a travel information service, and the like. Town information is, for example, information rooted in a specific city or an area consisting of a plurality of adjacent cities. As tags for classifying the information, names of municipalities, names of stations, etc. can be added. It's about information.

Note that the content may include other services that constitute information about the user, and may provide information about shopping, auctions, finance, news, entertainment, and the like, for example.

The content-related information includes information provided by the content provider to identify the user when the user uses the above services, information set by the user, user's usage history, and the like. The content-related information may also include a purchase history of purchase services such as shopping content.

In the detailed information DB, for example, information such as content identification information, content titles, outlines, and tags that simply indicate the content are collected from a server that provides content (hereinafter referred to as a service server) and stored.

The model generation device performs learning using content-related information and a detailed information DB containing detailed information about content as learning data, and a point of interest including at least the content in the learning data by the administrator of the service server (hereinafter referred to as the service provider). and a correct answer, and generates a trained model (hereinafter referred to as a recommendation model) that outputs the correct answer when information corresponding to the point of interest is input by the service provider.

[overall structure]
FIG. 1 is a schematic diagram showing the usage environment of the model generation device 100. As shown in FIG. The information providing system 1 in the embodiment includes one or more terminal devices 10-1 to 10-n (n is a natural number), an administrator terminal 20, a model generation device 100, and one or more service servers 200-1. ~200-m (m is a natural number) and a detailed information DB300. These devices are connected to each other via a network NW. The network NW includes, for example, wireless base stations, Wi-Fi access points, communication lines, providers, and the Internet. Note that not all combinations of the devices shown in FIG. 1 need to be able to communicate with each other, and the network NW may partially include a local network.

The terminal devices 10-1 to 10-n are terminals used by content users, such as mobile phones such as smart phones, tablet terminals, and personal computers. In the following description, each terminal device 10-1 to 10-n is simply referred to as a terminal device 10 when not distinguished.

The administrator terminal 20 is a terminal used by a service provider, and is, for example, a mobile phone such as a smart phone, a tablet terminal, a personal computer, or the like.

The service servers 200-1 to 200-m are servers that provide content, and are, for example, web servers that provide web pages corresponding to requests from terminal devices 10 operated by users, and terminals on which applications are started. It is an application server or the like that communicates with the device 10 to exchange various information and provide content information. The service servers 200-1 to 200-m, for example, provide news services, shopping services, auction services, matching services, financial settlement services, navigation services, webmail services, etc., to the terminal device 10 via the network NW. It is provided to browsers and dedicated applications. In the following description, the individual service servers 200-1 to 200-m are simply referred to as service servers 200 when they are not distinguished. Also, in the constituent elements of the service server 200, the numbers after the hyphen may be omitted as well.

Some or all of the components of the service server 200 are realized by executing a program stored in a storage device by a processor such as a CPU (Central Processing Unit). Also, some or all of the components of the model generation device 100 may be realized by hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), or FPGA (Field-Programmable Gate Array). However, it may be realized by cooperation of software and hardware.

The service server 200 includes content-related information 210, detailed information 220, and a recommendation model M, for example. The content-related information 210 is information indicating the user's action history, and includes, for example, information provided by the content provider to identify the user, information set by the user himself, user's usage history, user's IP address. and so on. Also, the content-related information 210 may include a purchase history of purchase services such as shopping content. The detailed information 220 includes, for example, content identification information, content titles, outlines, tags that simply indicate the content, and the like.

The detailed information DB 300 is implemented by, for example, a HDD (Hard Disc Drive), flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), ROM (Read Only Memory), or RAM (Random Access Memory). Detailed information 220 is collected from each of the service servers 200 and stored in the detailed information DB 300 .

The model generation device 100 includes an acquisition unit 110, a region vector derivation unit 120, a learning unit 130, and a model generation unit 140, for example. Some or all of the components of the model generation device 100 are implemented by executing a program stored in a storage device by a processor such as a CPU (Central Processing Unit). Also, some or all of the components of the model generation device 100 may be realized by hardware such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), or FPGA (Field-Programmable Gate Array). However, it may be realized by cooperation of software and hardware.

Acquisition unit 110 acquires content-related information 210 from service server 200 . Further, the acquisition unit 110 acquires detailed information 220 to be learned from the service server 200 . The acquisition unit 110 may acquire detailed information from the detailed information DB 300 . For the following description, the content-related information acquired by the acquisition unit 110 is called "content-related information 112", and the detailed information is called "detailed information 114".

Regional vector derivation unit 120 derives a region vector based on each of content-related information 112 and detailed information 114 acquired by acquisition unit 110 . Regional vectors will be described later. The regional vector derivation unit 120 is an example of a "derivation unit."

Learning unit 130 learns information included in content-related information 112 and detailed information 114 acquired by acquiring unit 110 as learning data.

The model generation unit 140 receives a specification from the service provider regarding a point of interest including at least content and a correct answer in the content-related information 112 and the detailed information 114 learned by the learning unit 130, and generates a recommendation model M. The recommendation model M is a learned model that outputs a correct answer when information corresponding to a point of interest is input.

For example, when the region vector of the content-related information 112 and the detailed information 114 is input as a point of interest, the recommendation model M is information identifying other content having a similar region vector from the content-related information 112 (hereinafter referred to as content ID). The recommendation model M is placed in the service server 200 based on the operation of the service manager.

[Content-related information]
FIG. 2 is an example of the contents of the content-related information 210. As shown in FIG. The content-related information 210 includes, for example, user IDs, content IDs used by users, content IMP (impression) tags, PV (page views), content completion rates, and links to content on microblogs and the like. "Number of times of sharing", "Number of likes" of evaluating content, user's gender, user attributes (Demog), and the like are included. These pieces of information are acquired, for example, by the service provider for the purpose of conducting a marketing survey on the user's usage status. Note that the user attribute (Demog) is an example of information from which a region vector is derived. The user attributes (Demog) include, for example, the user's current location and place of residence.

[Detailed information]
FIG. 3 is a diagram showing an example of the detailed information 220. As shown in FIG. The detailed information 220 includes, for example, the content ID, the title and outline of the content, geographic information included in the details of the content, tag information (set by the service provider) attached to the content, and the like.

[Regional vector]
A region vector is a vector that measures the closeness of pieces of information having regional characteristics contained in the content-related information 112 (content-related information 210) and the detailed information 114 (detailed information 220). The closeness of locality indicated by the area vector expresses the degree of approximation of the characteristics of the areas in addition to the closeness of the physical distance between the areas. For example, the closeness of locality indicated by the regional vector may be defined by the service provider, or may reflect the result of learning by the learning unit 130 .

When the tag "ramen" is set in the information included in the content-related information 112 and the detailed information 114, "Yamagata" and "Kitakata", which are physically close to each other, may be set as having close regional vectors. Based on the result of the learning by the learning unit 130, it is possible to set "Yamagata" and "Kumamoto", which are likely to have similar interests to each other and have similar characteristics, to have close area vectors. Further, for example, when a tag “coffee” is set in the information included in the content-related information 112 and the detailed information 114, “Sumatra” is a famous production area although it is physically distant. "Jamaica" may be set as having a close regional vector.

In addition, the region vector derivation unit 120 uses the user's current location estimated from the IP address in addition to the information such as the address and work location preset by the user among the information included in the content-related information 112 to provide the user with a An associated region vector may be derived.

[Processing of Model Generating Device]
Details of the process of generating the recommendation model M by the model generation device 100 will be described below. FIG. 4 is a diagram for explaining the processing of the model generation device 100. As shown in FIG.

Acquisition unit 110 includes, for example, data integration module 116 and storage unit 118 . The storage unit 118 stores content-related information 112, detailed information 114, and the like. Storage unit 118 is implemented by, for example, an HDD, flash memory, EEPROM, ROM, or RAM.

Acquisition unit 110 outputs content-related information 210 , detailed information 220 , and information acquired from detailed information DB 300 to area vector derivation unit 120 . The regional vector deriving unit 120 derives regional vectors of various information output from the obtaining unit 110 .

The data integration module 116 of the acquisition unit 110 outputs the acquired various information and the area vectors derived by the area vector derivation unit 120 to the learning unit 130 . The data integration module 116 uses, for example, an ETL (Extract/Transform/Load) tool such as TALEND (registered trademark). The data integration module 116 outputs part or all of the information stored in the storage unit 118 (for example, update differences and information that is not preferable as learning data is excluded) to the learning unit 130 at a predetermined cycle. Information that is not preferable as learning data is, for example, information from which the regional vector derivation unit 120 could not derive a regional vector.

The learning unit 130 includes, for example, learning data 132 , a learning framework 134 , and a point-of-interest setting 136 . The learning data 132 is implemented by, for example, an HDD, flash memory, EEPROM, ROM, or RAM.

The learning data 132 stores content-related information 132A and detailed information 132B transmitted by the data integration module 116 . The content-related information 132A stores the content-related information 210 and the area vector derived by the area vector derivation unit 120. FIG. The detailed information 132B stores the detailed information 220 (or the detailed information DB 300) and the area vectors derived by the area vector derivation unit 120. FIG.

FIG. 5 is a diagram showing an example of the point-of-interest setting 136. As shown in FIG. The service provider sets one or more points of interest via the administrator terminal 20 and instructs the learning policy of the learning unit 130 . Points of interest that can be set in the point-of-interest setting 136 may be the items themselves of information stored in the content-related information 132A and the detailed information 132B as shown in the points of interest 1 and 2 in FIG. 5, a condition to be satisfied by the information stored in the content-related information 132A and the detailed information 132B may be set.

The learning framework 134 refers to the point-of-interest setting 136 and performs learning using the learning data 132 , especially the area vectors included in the learning data 132 . The point-of-interest setting 136 is, for example, a point of interest set in advance by the service provider via the administrator terminal 20 or each time learning is performed. The learning framework 134 is a framework (machine learning library) for performing numerical calculations in parallel using a GPU (Graphics Processing Unit). DNN (Deep Neural Network) calculations such as CNN (Convolutional Neural Network) are performed.

The model generation unit 140 generates a recommendation model M based on the learning result of the learning unit 130 .

[Recommendation example 1]
FIG. 6 is a diagram for explaining an output example of the recommendation model M. FIG. For example, when the user U1's content-related information 132A is input to the recommendation model M1 in which the places of interest are set to be "residence" and "current location", the recommendation model M1 is "weekend event in Kanagawa". , “Weekend events in Tokyo” and other recommended information is output.

When the user U1's content-related information 132A is input to the recommendation model M2 in which the points of interest are set to be "current location", "residence", and "interest", the recommendation model M2 is "current location Recommended ramen”, “Weekend events in Kanagawa”, etc. are output. The recommendation information output by the recommendation model M2 may include the same recommendation information as the recommendation model M1.

When the user U1's content-related information 132A is input to the recommendation model M3 in which the points of interest are set to be "current location", "place of residence", "interest", and "purchase history", the recommendation model M3 is , “Osaka Ramen Events”, “Recommended Ramen Around Current Location”, etc.

The output from the recommendation model M3 reflects the influence of the "purchase history" to a greater extent than the outputs from the recommendation models M1 and M2. The recommendation models M1 and M2 produce output influenced by the "current location" and the "residential location". This reduces the possibility that the information of the area is output.

On the other hand, the output of the recommendation model M3 is, for example, the output from the service provider, ``Users who satisfy the condition ``purchase history includes travel goods'' may go on a gourmet trip, a trip for photography, etc.'' , reflects the hypothesis that ``users who satisfy the condition that ``purchasing history includes travel goods'' will collect information about things they are interested in, even if it is information about an area far from their living area.

As shown in the figure, when the recommendation models M1 to M3 with different set points of interest are used for the content-related information 132A of the same user U1, different outputs occur.

[Recommendation example 2]
FIG. 7 is a diagram for explaining another output example of the recommendation model M. FIG. In the recommendation model M (hereinafter referred to as recommendation model M4) used in the example of FIG. The degree of influence of

FIG. 8 is another example of the detailed information 220. As shown in FIG. As shown in the figure, the Yamagata ramen special feature has tag information such as "outing", "ramen", and "Yamagata" set in advance by the service provider. Note that these pieces of information are the items set as the point of interest in the recommendation model M4, or are related to the items set as the point of interest.

Returning to FIG. 7, for example, when the points of interest of the recommendation model M4 are "place of residence", "interest", and "purchase history", and the user U2's content-related information 132A is input, the recommendation model M4 is Yamagata Ramen. Give priority to special features. The recommendation model M4 outputs as shown in the drawing based on the area vector of the current location or residence of the user U3 and the area vector of "Yamagata ramen special" of the detailed information 220. FIG.

In addition, the recommendation model M4 preferentially displays "Yamagata ramen special" even when the user U3's content-related information 132A is input. This output reflects the user U3's "interest" and "purchasing history includes travel goods". The recommendation model M4 outputs as shown in the drawing based on the area vector of the current location or residence of the user U3 and the area vector of "Yamagata ramen special" of the detailed information 220. FIG.

[Processing flow]
FIG. 9 is a flowchart showing an example of the flow of processing by the model generating device 100. As shown in FIG.

First, the learning unit 130 receives the setting of the target part setting 136 input by the service provider via the administrator terminal 20 (step S100). Next, the acquisition unit 110 acquires the detailed information 220 (or the detailed information DB 300) submitted by the service provider via the administrator terminal 20 to the service server 200 (step S102).

Next, the acquisition unit 110 acquires content-related information including the user's action history (step S104). Next, the regional vector deriving unit 120 derives a regional vector (step S106). Next, the learning unit 130 learns the learning data 132 to which the area vectors derived by the area vector deriving unit 120 are added (step S108). Next, the model generation unit 140 outputs the recommendation model M based on the learning result of the learning unit 130 (step S110). This completes the description of the processing of this flowchart.

As described above, in the model generation device 100 of the embodiment, the learning unit 130 is the action history of the user, content-related information including the user ID, the user's position, and information indicating the content provided to the user; The detailed information DB 300 including detailed information of the content is learned as learning data, and the model generating unit 140 generates the recommendation model M based on the learning result of the learning unit 130, thereby identifying the point of interest from the service provider and the correct answer. It is possible to provide a service provider with a recommendation model M that accepts designation and outputs a correct answer when information corresponding to a point of interest is input. As a result, the service provider can use the recommendation model M to select content to be provided and to make efforts to further increase user satisfaction in service provision.

As described above, the mode for carrying out the present invention has been described using the embodiments, but the present invention is not limited to such embodiments at all, and various modifications and replacements can be made without departing from the scope of the present invention. can be added.

10 terminal device 20 administrator terminal 100 model generation device 110 acquisition unit 120 area vector derivation unit 130 learning unit 140 model generation unit 200 service server 300 detailed information DB

Claims (4)

A user's action history, which includes at least identification information of the user, location of the user, information indicating content used by the user , content-related information including tags of the content , and at least tag information. a learning unit that generates a learned model that outputs a correct answer when information corresponding to a point of interest is input using content detailed information and detailed information including a tag of the content or a region vector related to the tag information as learning data;
a model generation unit that accepts designation of the correct answer that includes at least the focused point in the learning data and the correct answer, conveys the focused point and the correct answer to the learning unit, and causes the learning unit to perform learning;
A model generation device comprising : The content-related information further includes the user's purchase history,
The model generation device according to claim 1. the computer
A user's action history, which includes at least identification information of the user, location of the user, information indicating content used by the user , content-related information including tags of the content , and at least tag information. generating a learned model that outputs a correct answer when information corresponding to a point of interest is input using detailed information about the content and detailed information including a tag of the content or a region vector related to the tag information as learning data;
Receiving designation of the correct answer including the point of interest in the learning data and at least the content, and transmitting the point of interest and the correct answer to a function that generates the trained model to perform learning;
Model generation method. to the computer,
A user's action history, which includes at least identification information of the user, location of the user, information indicating content used by the user , content-related information including tags of the content , and at least tag information. generating a trained model that outputs a correct answer when information corresponding to a point of interest is input using detailed information about content and detailed information including a tag of the content or a region vector related to the tag information as learning data;
Receiving designation of the point of interest in the learning data and the correct answer including at least the content, and transmitting the point of interest and the correct answer to the function that generates the trained model to perform learning;
program .

Images