JP7249375B2 - Information processing device, information processing method, and information processing program - Google Patents

Description

The present invention relates to an information processing apparatus, an information processing method, and an information processing program for generating a model that can be applied to recommending information to a user.

2. Description of the Related Art Conventionally, a system for recommending information to a user is known (see Patent Document 1, for example). This system acquires the user's action history information from the mobile terminal owned by the user, calculates the usefulness of recommended information in the future using a classification model, and distributes highly useful recommended information to the user. Also, the user's reaction to the recommendation information is acquired as reaction history information.
Then, in this system, a classification model is generated using action history information before a predetermined time and reaction history information after a predetermined time. In model generation, a feature vector is generated that indicates the user's behavior indicated by the behavior history information that constitutes the teacher data. Machine learning is performed as a negative example when it does not correspond to . That is, this classification model receives as input a feature vector corresponding to each action of the user, and outputs the probability that the action corresponding to the recommendation information was performed as the usefulness.

JP 2018-132860 A

A conventional classification model used in a system such as that described in Patent Literature 1 generates a model using individual behavior information of users as training data. In this case, it is possible to calculate the degree of similarity of each action from the feature vector of each action information, but the information is based only on the action information. For example, if an online shopping user browses the content of product A (first action) and then browses the content of product B (second action), vectors for each action are calculated. By doing so, a feature vector for product A and a feature vector for product B can be calculated. However, a series of flow such as seeing the product A and then seeing the product B is not taken into account. In other words, by inputting the user's action history, a product similar to product A for the first action and a product similar to product B for the second action are output from the model. (score). However, it is not possible to recommend the product B or similar products to the first action. In view of such circumstances, there is a demand for a model capable of outputting a wider range of information in response to user behavior.

An object of the present invention is to provide an information processing apparatus, an information processing method, and an information processing program that generate a model capable of outputting a wide range of information regarding user behavior.

An information processing apparatus according to the present invention includes an action history acquisition unit that acquires action history information including an action type that is a type of action of a user, a title of a target of the action, and a time point at which the action is performed; A model generation unit that generates a model that outputs recommended information indicating the target corresponding to the action history information by using the plurality of the action history information along the series as teacher data and inputting the action history information. And prepare.

In the present invention, a model is generated using a plurality of time-series action history information as teacher data. Therefore, it is possible to perform machine learning that considers the transition of successive user action histories. Models generated by such machine learning can broaden the range of recommendations for user behavior. In other words, even if a target is not recommended based on only one action history of the user, the possibility of being recommended is increased by learning the previous and subsequent action history, and the user can be recommended in a wider range. Target can be recommended.

1 is a schematic diagram showing a schematic configuration of an information processing system according to an embodiment of the present invention; FIG. 1 is a block diagram showing a schematic configuration of an information processing apparatus according to an embodiment; FIG. 4 is a flowchart showing target value setting processing according to the information processing method of the present embodiment. 4 is a flowchart showing recommendation information distribution processing according to the information processing method of the present embodiment.

An embodiment according to the present invention will be described below with reference to the drawings.
[1. Overall configuration of information processing system]
FIG. 1 is a schematic diagram showing a schematic configuration of an information processing system according to this embodiment.
As shown in FIG. 1, in an information processing system 1 of the present embodiment, a server device 10, which is an information processing device, and user terminals 20 are communicably connected via a network such as the Internet.
In this information processing system 1, for example, when predetermined content or the like is displayed on the user terminal 20, recommendation information introducing recommended products or the like is distributed from the server device 10 to the user terminal 20 and included in the content. display. The content is information displayed on the display when a predetermined application is executed. Includes images displayed when executed.

[2. Configuration of server device 10]
FIG. 2 is a block diagram showing a schematic configuration of the server device 10. As shown in FIG.
The server device 10 is configured by a computer and functions as an information processing device of the present invention. The server device 10 includes a communication section 11, a storage section 12, a control section 13, and the like. Note that the number of computers constituting the server device 10 is not particularly limited. For example, the server device 10 may be configured by one computer, or a cloud server constructed by connecting a plurality of computers via a network may be used as the server device 10 .
The communication unit 11 is connected to a network (Internet) via a LAN or the like, and communicates with the user terminal 20 or the like.

[2-1. Configuration of storage unit 12]
The storage unit 12 is a data recording device configured by, for example, a memory, a hard disk, or the like. The storage unit 12 includes a user information storage unit 121, a recommendation target storage unit 122, a genre storage unit 123, and the like. An information processing program (software) for controlling the server device 10 is recorded in the storage unit 12 . Here, an example in which a user information storage unit 121, a recommendation target storage unit 122, and a genre storage unit 123 are provided in the storage unit 12 of the server device 10 is shown, but communication with the server device 10 via a network is possible. These various databases may be provided in other connected data servers or cloud storages.

[2-1-1. Information stored in user information storage unit 121]
The user information storage unit 121 is a database that records various kinds of information about users who use this system, and stores user information for each individual user.
This user information includes, for example, the user's own information such as the user ID, user name, and user attributes, as well as the user's action history information.

A user ID is identification information for identifying a user. The user name is the account name of the user identified by the user ID. The user attribute is various information about the user identified by the user ID, and records various information about the user such as the user's sex, age, location, occupation, annual income, and hobbies, for example.

The user's action history information may include the history of each offline action in addition to various online actions on the Internet. Online behaviors include, for example, online shopping behaviors, search processing behaviors, browsing behaviors in predetermined content such as news, follow-up on SNS (Social Networking Service), reading articles in news content, and the like.
Offline actions include, for example, a user's movement history (position history) measured by the user terminal 20, a user terminal operation history, and a payment history by electronic payment performed at a store.
In the present embodiment, as an example, action history information based on actions in online shopping will be given. In addition, as actions in online shopping, in this embodiment, the browsing history when the user browses the product, the registration history when the product is registered as a watch list or a shortcut such as a favorite, the purchase history of purchasing the product is exemplified. Each of these action histories includes the type of action (browsing, registration, purchase, etc.), the title of the target of the action (product name), and the point in time (time stamp) when the action was performed.

[2-1-2. Information stored in recommendation target storage unit 122]
The recommendation target storage unit 122 records target advertisement information to be included in the recommendation information to be transmitted to the user terminal 20 . The advertisement information includes, for example, an advertisement ID, target title, target genre, target detailed information, and the like.
The advertisement ID is identification information that identifies advertisement information.
The target title is the name of the advertisement target (target).
The target genre stores the genre name of each layer when the advertisement target is classified into hierarchically structured genres. Hierarchical structure here refers to multiple hierarchical categories such as multiple first categories, second categories that are subordinate to each first category, and third categories that are subordinate to the second category. It is a structure classified into For example, if the advertisement target is "00 car genuine bumper", the first category is "automobile", the second category is "parts", and the third category is "bumper".
Target detailed information is various kinds of detailed information about the advertisement target, and includes, for example, the price of the target product, the contact information of the seller of the target product, the payment method, the image of the target product, the evaluation of the target product and the seller, etc. may

Advertising information is information registered by the seller who exhibits (sells) products, but it is not necessary for all of the above information to be registered by the seller, as long as at least the target title is included. good. For example, with respect to the target genre, the server device 10 may read the genre corresponding to the target title from the genre storage unit 123 and register it in the advertisement information.

[2-1-3. Information stored in genre storage unit 123]
The genre storage unit 123 stores genres corresponding to the above-described objects.
Specifically, the genres of the hierarchical structure as described above are recorded in the genre storage unit 123, and the corresponding advertisement target is associated with the lowest category. For example, the lowest category may be stored in association with a title to be advertised, or may be stored in association with a product code (manufacturing code) or the like to be advertised.

[2-2. Functional Configuration of Control Unit 13]
The control unit 13 includes an arithmetic circuit such as a CPU (Central Processing Unit) and a recording circuit such as a RAM (Random Access Memory). The control unit 13 expands the information processing program recorded in the storage unit 12 or the like into the RAM, and executes various processes in cooperation with the program expanded into the RAM. By reading and executing the information processing program recorded in the storage unit 12, the control unit 13, as shown in FIG. It functions as the distribution unit 134 and the like.

The user information registration unit 131 acquires registration information about the user from the user terminal 20 via the Internet, and updates the user information in the user information storage unit 121 or newly registers the user information.
The action history acquisition unit 132 acquires the user's action history information from the user terminal 20 and registers the acquired user information in the user information storage unit 121 as the action information of the corresponding user information.

The model generating unit 133 generates a model for outputting a recommended target (advertising target) to be recommended for the user's behavior based on the acquired user's behavior history information.
The model outputs a recommended target of a vector similar to the input feature vector of the action history information of the user. The recommended targets to be output may be a predetermined number of advertising targets with similar feature vectors, or may be all recommended targets with similarities equal to or greater than a predetermined percentage.

The model generator 133 includes a first generator 133A and a second generator 133B.
133 A of 1st production|generation parts produce|generate the teacher data for producing|generating a model based on the action information of user information, and the genre information of the genre memory|storage part 123. FIG.
In the present embodiment, the first generation unit 133A sorts a plurality of pieces of action history information in chronological order based on the time stamps of the action history information, and generates teacher data to which genres are assigned to the action targets. Generate. A detailed description of the teacher data generated by the first generation unit 133A will be given later.

The second generator 133B generates a model using the teacher data generated by the first generator 133A. In other words, inputting action history information by converting teacher data, which has multiple action histories in chronological order and genres are associated with each action target, into feature vectors and performs machine learning. and generate a model that outputs recommendation information that indicates the recommended target.

The recommendation distribution unit 134 transmits the recommended advertisement information output from the model to the user terminal 20 as recommendation information.

[3. Configuration of user terminal 20]
The user terminal 20 is a terminal device owned by a user, and is configured by a computer such as a smart phone, a tablet terminal, or a personal computer. Although the illustration of the specific configuration of the user terminal 20 is omitted, the user terminal 20 has the basic configuration of a general computer. That is, the user terminal 20 includes an input operation unit that receives user operations, a display that displays image information, a recording device that records various types of information, and an arithmetic circuit (such as a CPU) that performs arithmetic processing on various types of information.

[4. Information processing method]
Next, an information processing method in the server device 10 will be described.
[4-1. Generate model]
FIG. 3 is a flowchart showing model generation processing according to the information processing method of the present embodiment.
In this embodiment, the user accesses the server device 10 from the user terminal 20 and performs user registration processing for using the information processing system 1 or registration update processing for updating user information. , the user registration information is transmitted to the server device 10 . Upon acquiring the user registration information transmitted from the user terminal 20, the user information registration unit 131 of the server device 10 newly registers the user information in the user information storage unit 121 or updates the existing user information (step S1).

Also, when a user who has performed user registration operates the user terminal 20 to perform a predetermined action, action history information is transmitted to the server device 10 together with the user ID. The action history acquisition unit 132 of the server device 10 stores and accumulates the action history information acquired from the user terminal 20 in the corresponding user information (step S2: action history acquisition step).

Next, the model generator 133 generates a model based on the accumulated action history information. The model may be generated, for example, at a timing specified by the administrator of the server device 10, or at predetermined intervals such as every month. Further, deterioration of the model may be determined, and a new model may be generated when it is determined that the model has deteriorated.

First, the model generating unit 133 extracts action history information for generating teacher data for generating a model using the first generating unit 133A (step S3).
Here, the first generation unit 133A may randomly extract a plurality of pieces of user information from the user information stored in the user information storage unit 121 and read the action history information thereof. You can read the information. Alternatively, user information may be clustered into a plurality of clusters based on the user attributes of the user information, a predetermined number of pieces of user information may be randomly extracted from each cluster, and action history information thereof may be read. Also, the action history information acquired from each user information may be, for example, action history for a certain period of time such as the last month, or may be all action history information.

Since the extracted action history information is given a time stamp as described above, it is possible to grasp the actions of each user in chronological order. Then, the first generation unit 133A generates behavior history information in chronological order for each user as teacher data. At this time, the first generation unit 133A adds the category of the target of the action to the action type and the title of the target of the action history information, and arranges a plurality of pieces of action history information for each user in chronological order as teacher data. is generated (step S4).
Specifically, the first generation unit 133A searches the genre storage unit 123 for a genre corresponding to the action target recorded in the action history information. Then, for each piece of action history information, the first generation unit 133A links the type of action in the action history information, the title of the target of the action, and the genre corresponding to the target as text information in a predetermined order.
For example, the action history information of "viewing" product A of "TitleA" and the action history information of "watchlisting" product B of "TitleB" are consecutive. , the first category "categoryα1" and the second category "categoryβ1" are retrieved, and for the product B, the first category "categoryα2" and the second category "categoryβ2" are retrieved. In this case, text information such as "view_TitleA_categoryα1_categoryβ1" is generated for the action history information of viewing product A. Also, text information such as "watchlist_TitleB_categoryα2_categoryβ2" is generated for the action history information in which product B is registered in the watchlist.
Then, the first generating unit 133A generates teacher data in which the text information generated as described above is arranged in chronological order for each user. In this embodiment, as described above, text information in which action types, titles, and genres are linked is arranged in chronological order for each user, and delimiter information is inserted between data of different users. . Therefore, even if the user U1 performs the first action at 10:00, the user U2 performs the second action at 10:01, and the user U1 performs the third action at 10:02, the user U1 The text information based on the action history information of the second action of the user U2 does not enter between the text information based on the action history information of the first action and the third action. Therefore, it is possible to correctly learn the transition of a series of actions of each user.
Note that the above text information shows an example of connecting in the order of action type, action target, first category, and second category, but the order is not particularly limited.

After that, the second generator 133B generates a model using the teacher data generated in step S4 (step S5). Specifically, the second generation unit 133B generates a model for calculating a feature vector for action history information by performing machine learning on the teacher data generated in step S4 using a technique such as fastText, for example. . By inputting action history information, this model calculates a feature vector of the action history information, and indicates a recommendation target (such as a product to be recommended to the user) of the feature vector close to the action history information. Output recommendation information. Steps S4 and S5 correspond to the model generation step of the present invention.

As described above, in the present embodiment, teacher data for generating a model includes chronological action history information and target categories. Therefore, it is possible to bring the feature vectors of the targets selected by the actions before and after closer together, and bring the feature vectors of the targets having the same category closer together.
For example, a case will be described in which action history information in which product B is registered in the watch list is obtained after action history information in which product A is browsed. Conventionally, if only one piece of action history information is used as input data, the feature vectors of the browsed product A and the product B registered in the watchlist cannot be brought close to each other. The feature vectors of B are independent feature vectors. In addition, since the categories of product A and product B are not specified, the range of selection of recommended targets similar to product A and product B is limited. Limited to products with similar titles.
On the other hand, in the present embodiment, as described above, feature vectors are calculated so that the feature vectors of the browsed product A and the product B registered in the watchlist are close to each other. Therefore, it is possible to select not only products similar to the product A but also the product B and products similar to the product B as recommended targets for the user who browses the product A. In addition, by including the categories of product A and product B, other products within the same genre as product A and product B can be selected as recommended targets. That is, even if a product has a completely different title from that of product A or product B, if the product is of the same genre, the product can be recommended.

[4-2. Delivery processing of recommendation information]
Next, a process for distributing recommendation information will be described.
FIG. 4 is a flowchart showing a process for distributing recommendation information.
The user terminal 20 transmits the action history information to the server device 10 when predetermined action history information is obtained by the user's operation.
As a result, the action history acquisition unit 132 of the server device 10 acquires action history information from the user terminal 20 (step S11), as in step S2.
Then, the recommendation distribution unit 134 inputs the obtained action history information to the model generated in step S5, and acquires recommendation information indicating a recommendation target corresponding to the action history information mainly output from the model (step S12). .
Thereafter, the recommendation distribution unit 134 generates recommendation information including advertisement information corresponding to the acquired recommendation information, and transmits the recommendation information to the user terminal 20 (step S13). As a result, the user terminal 20 that has received the recommendation information can display the advertisement information arranged according to the predetermined sorting conditions as the recommendation information on the display.

[5. Effects of this embodiment]
In the present embodiment, the server device 10 includes a storage unit 12 and a control unit 13, and the control unit 13 reads and executes an information processing program stored in the storage unit 12 to obtain an action history acquisition unit 132, and It functions as the model generator 133 . The action history acquisition unit 132 acquires action history information including an action type, which is the type of action of the user, the title of the target of the action, and the time when the action was performed. The model generating unit 133 uses a plurality of pieces of time-series action history information as teacher data, and generates a model that outputs a recommended target corresponding to the action history information by inputting the action history information.
In this way, by generating a model based on training data obtained by sorting the action history of a plurality of consecutive users rather than one action history of the user, the target selected by a plurality of consecutive action types is learned. be able to. In other words, in a conventional model that learns only one action type and its target, the range of recommended targets to be output is also limited. By learning the data, it is possible to bring the feature vectors of these behavioral targets closer together, and generate a model capable of outputting a wide range of recommended targets.

In this embodiment, text information including and linking the action type of the action history information, the target title, and the target genre (category) retrieved from the genre storage unit 123 is used as teacher data. As a result, the second generation unit 133B can approximate the feature vectors of each linked item (action types, titles, and genres of a plurality of action history information) by the fastText technique.

In this embodiment, the control unit 13 functions as a recommendation distribution unit 134 by reading and executing an information processing program stored in the storage unit 12, and based on the recommendation information output from the model generation unit 133, the user Distribute recommendation information to As a result, as described above, it is possible to distribute recommendation information targeting a wider range of advertising information than in the past with respect to the user's action history information.

[6. Modification]
It should be noted that the present invention is not limited to the above-described embodiments, and includes modifications shown below within the scope of achieving the object of the present invention.

(Modification 1)
In the above embodiment, distribution of recommendations for products in an online shopping service was exemplified, but the present invention can be applied to other services as well. For example, when distributing news, news articles viewed by the user, news articles bookmarked by the user, and the like may be acquired as action history information. In this case, a model can be generated in the same manner as in the above-described embodiment, using each action history information for news content and the genre of the browsed article.

(Modification 2)
In the above-described embodiment, an example of machine learning using chronological action history information as teacher data is shown, but transition information may be generated by associating each piece of action history information with immediately preceding action history information. . For example, the action history information of "viewing" product A of "TitleA" and the action history information of "watchlisting" product B of "TitleB" are consecutive. , the first category "categoryα1" and the second category "categoryβ1" are retrieved, and for the product B, the first category "categoryα2" and the second category "categoryβ2" are retrieved. In this case, the first generation unit 133A connects the immediately preceding action history information to the action history information in which "product B" is "registered in the watchlist", and generates input data of text information "watchlist_TitleB_categoryα2_categoryβ2_view_TitleA_categoryα1_categoryβ1". may be generated.

(Modification 3)
In the above-described embodiment, the first generation unit 133A uses teacher data in which a genre for a target is added to a plurality of time-series action history information. It doesn't have to be. In other words, the type of action and the target of the action in the action history information may be linked as text information, and teacher data obtained by sorting a plurality of the action history information may be used. Even in this case, it is possible to generate a model that considers the transition of the action history, and output a wider range of recommended targets than the conventional model.

Alternatively, the first generating unit 133A may generate, as teacher data, text information to which a target genre is added for one piece of action history information. In other words, even if a plurality of pieces of action history information are not arranged in chronological order, the time points included in each piece of action history information may be added and used as teacher data.

The first generating unit 133A generated teacher data by sorting the action history information for each user in chronological order. Alternatively, a model may be generated for each user. In this case, action history information of the target user in chronological order may be used as teacher data.

DESCRIPTION OF SYMBOLS 1... Information processing system 10... Server apparatus (information processing apparatus) 12... Storage part 13... Control part 20... User terminal 121... User information storage part 122... Recommendation object storage part 123... Genre storage part , 131... User information registration unit, 132... Action history acquisition unit, 133... Model generation unit, 133A... First generation unit, 133B... Second generation unit, 134... Recommendation distribution unit.

Claims (7)

an action history acquisition unit that acquires action history information including an action type that is a type of action of the user, a title of the target of the action, and a time point at which the action was performed;
A plurality of pieces of the action history information along the time series are used as training data for one input action history information , machine learning is performed on the relevance of the series of the action history information along the time series, and the action history information is obtained. is input, a model generation unit that generates a model that outputs recommended information indicating the target corresponding to each of the input action history information and a plurality of the action history information related to the action history information ; ,
Information processing device. A genre storage unit that stores genres of each hierarchy when the object is classified into genres of a hierarchical structure,
The information processing apparatus according to claim 1, wherein the model generation unit generates, as the teacher data, text information in which the action type, the title, and the genre corresponding to the object of each of the action history information are connected. an action history acquisition unit that acquires action history information including an action type, which is the type of action of the user, and a title of the target of the action;
a genre storage unit that stores the genres of each hierarchy when the object is classified into hierarchically structured genres;
Linking the action type, the title, and the genre corresponding to the object, which are generated based on the plurality of the action history information along the time series and the genre corresponding to the object of each of the action history information. Using the text information obtained as training data for one action history information input , machine learning is performed on the relevance of the series of action history information along the time series, and the action history information is input, a model generation unit that generates a model that outputs recommended information indicating the target corresponding to each of the input action history information and a plurality of pieces of action history information related to the action history information ;
Information processing device. a distribution unit that distributes recommended information to the user based on the recommended information output from the model;
The information processing apparatus according to any one of claims 1 to 3. An information processing method for generating a model by a computer,
The computer includes an action history acquisition unit and a model generation unit,
an action history acquisition step in which the action history acquisition unit acquires action history information including an action type that is the type of action of the user, a title of the target of the action, and a time point at which the action was performed;
The model generation unit uses the plurality of time-series action history information as teacher data for one piece of input action history information , and automatically calculates the relationship between the series of time-series action history information. learning and inputting the action history information, outputting recommendation information indicating the target corresponding to each of the input action history information and a plurality of the action history information related to the action history information; a model generation step for generating a model;
Information processing method for implementing An information processing method for generating a model by a computer,
The computer includes an action history acquisition unit, a model generation unit, and a genre storage unit,
The genre storage unit stores the genre of each hierarchy when the target of user's action is classified into genres of hierarchical structure,
an action history acquisition step in which the action history acquisition unit acquires action history information including an action type, which is the type of action of the user, and the title of the target of the action;
The model generation unit generates the action type, the title, and the object generated based on the plurality of pieces of the action history information in chronological order and the genre corresponding to the object of each of the action history information. Using the text information that connects the corresponding genres as teacher data for one input action history information , machine learning is performed on the relevance of the series of action history information along the time series, and the action history information is a model generation step of generating the model for outputting recommended information indicating the target corresponding to each of the input action history information and a plurality of pieces of action history information related to the input action history information; ,
Information processing method for implementing A computer-readable and executable information processing program,
An information processing program that causes the computer to function as the information processing apparatus according to any one of claims 1 to 4.

Images