JP6998349B2 - Learning equipment, learning methods, and learning programs - Google Patents

Description

Article 30, Paragraph 2 of the Patent Act Applicable on the website Reiwa June 18, 1st Takahiro Yabe URL: https: arxiv. org / pdf / 1906.07770. pdf

Patent Law Article 30, Paragraph 2 Applicable Reiwa August 6, 1st Takahiro Yabe, Kota Tsubouchi, Toru Shimizu KDD (https://www.kdd.org/kdd2019) Dena'ina Convention Center, Anchorage , Alaska USA

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

Conventionally, a technique for estimating various information about a user has been proposed. As an example of such a technology, there is a technology for estimating the tendency of a search query characteristic of a user from the difference between the tendency of a search query entered by a user and the tendency of a search query entered by another user. Are known.

Japanese Unexamined Patent Publication No. 2018-60469

However, it may not be possible to say that the above-mentioned technology appropriately estimates the user's intention.

For example, in the above-mentioned conventional technique, the tendency of the characteristic search query input by the user is only estimated. Therefore, as a result of not being able to estimate what kind of intention the user entered the search query, it is not possible to provide information according to the user's intention.

The present application has been made in view of the above, and an object thereof is to appropriately estimate the intention of the user.

The learning device according to the present application generates information indicating the user's intention when performing the action indicated by the action information from the acquisition unit that acquires a plurality of action information indicating the user's action and each action information. It is characterized by having a generation unit to be generated and a learning unit to make the first model learn the characteristics of the transition of the intention indicated by the intention information generated by the generation unit.

According to one aspect of the embodiment, the estimation accuracy of the user's intention can be improved.

FIG. 1 is a diagram showing an example of an information providing system according to an embodiment. FIG. 2 is a diagram showing an example of learning of an intention estimation model executed by the information providing device according to the embodiment. FIG. 3 is a diagram showing an example of learning of a transition estimation model executed by the information providing device according to the embodiment. FIG. 4 is a diagram showing an example of a process in which the information providing device according to the embodiment predicts the intention of future actions by using the transition estimation model. FIG. 5 is a diagram showing an example of learning of a transition estimation model executed by an information providing device according to an embodiment using behavior information. FIG. 6 is a diagram showing an example of a process in which the information providing device according to the embodiment predicts an intention from a plurality of types of actions. FIG. 7 is a diagram showing an example of a process in which the information providing device according to the embodiment predicts an intention from a plurality of types of actions. FIG. 8 is a diagram showing a configuration example of the information providing device according to the embodiment. FIG. 9 is a diagram showing an example of information registered in the behavior history database according to the embodiment. FIG. 10 is a diagram showing an example of information registered in the model database according to the embodiment. FIG. 11 is a flowchart showing an example of a flow of processing executed by the information providing apparatus according to the embodiment. FIG. 12 is a diagram showing an example of a hardware configuration.

Hereinafter, the learning device, the learning method, and the embodiment for implementing the learning program (hereinafter referred to as “embodiment”) according to the present application will be described in detail with reference to the drawings. It should be noted that this embodiment does not limit the learning device, learning method, and learning program according to the present application. In addition, each embodiment can be appropriately combined as long as the processing contents do not contradict each other. Further, in each of the following embodiments, the same parts are designated by the same reference numerals, and duplicate explanations are omitted.

[Embodiment]
[1. Information provision system]
First, using FIG. 1, an information providing system 1 having an information providing device 10, a terminal device 100, and a search server 200, which are examples of learning devices, and providing information according to the characteristics of the transition of the user's intention. An example will be described. FIG. 1 is a diagram showing an example of an information providing system according to an embodiment. As shown in FIG. 1, the information providing system 1 includes an information providing device 10, a terminal device 100, and a search server 200. The information providing system 1 may have an arbitrary number of terminal devices 100 and an arbitrary number of information providing devices 10. Further, the information providing system 1 does not have to have the search server 200, for example, and may have various arbitrary external servers.

The information providing device 10 is an information processing device capable of communicating with the terminal device 100 and the search server 200 via a predetermined network N (see, for example, FIG. 7) such as the Internet, and is, for example, a server device or a cloud system. It is realized by such as. The information providing device 10 may be able to communicate with an arbitrary number of terminal devices 100 via the network N, or may be able to communicate with various external servers and external systems (not shown).

The terminal device 100 is a terminal device used by a user, and is realized by, for example, a PC (Personal Computer), a server device, various smart devices, and the like. Further, the terminal device 100 has a camera capable of shooting various subjects, and also has a function of transmitting captured images such as still images and moving images captured by the camera to the information providing device 10. Further, when the terminal device 100 receives the information generated by the information providing device 10 by the providing process described later, the terminal device 100 displays the received information on the screen to provide various information to the user. Shall have.

[1-2. An example of the flow of processing executed by the information providing device 10]
Hereinafter, an example of the flow of the learning process executed by the information providing device 10 and the providing process executed by using the model learned by the learning process will be described with reference to FIG. 1. First, the user U inputs a search query via the terminal device 100 (step S1). In such a case, the search server 200 executes a search process corresponding to the search query and provides the search result to the terminal device 100 (step S2).

Subsequently, the information providing device 10 acquires the history of the search query and the like as behavior information indicating the behavior of the user (step S3). More specifically, the information providing device 10 acquires the history of a plurality of search queries continuously input by the user in a predetermined session. Then, the information providing device 10 learns the intention estimation model M1 that estimates the user's intention when the search query is input from each search query (step S4).

[1-2-1. About the intention estimation model]
Here, the intention estimation model M1 for learning by the information providing device 10 will be described. For example, in order to provide information that is presumed to be of interest to the user, a technique that provides information according to the history of the search query entered by the user can be considered. For example, when the user inputs the search query "Ginza", a method of providing the user with information on the transaction target related to "Ginza" can be considered. However, such technology only provides information on the transaction target corresponding to the search query entered by the user in the past, so it can be said that appropriate information is provided to the user. May not be.

On the other hand, when a certain action performed by the user is detected, the intention of the user who performed the detected action is estimated, and information according to the estimated intention is provided, which is appropriate for the user. It is believed that information can be provided. For example, when the user enters the search query "Ginza" and it is presumed that the user has the intention of "trying to stay in Ginza", the user is asked to use the hotel located in Ginza. By providing the information, it can be considered that the information according to the user's search intention can be provided.

Therefore, the information providing device 10 predicts the user's behavior by learning the model so as to predict the next behavior to be performed by the user after inputting the search query based on the user's search query. Generate a predictive model. That is, the information providing device 10 learns the intention estimation model M1 that estimates the intention when the user inputs the search query (that is, the intention when the user performs a certain action).

For example, FIG. 2 is a diagram showing an example of learning of an intention estimation model executed by the information providing device according to the embodiment. For example, in the example shown in FIG. 2, it is assumed that the user U inputs the search query # 1 with the intention # 1, and then the user U inputs the search query # 2 with the intention # 2. Here, when the input of the search query # 1 and the input of the search query # 2 are continuously performed, for example, within a predetermined same session (for example, within a predetermined time or a session between the terminal device 100 and the search server 200). When the search queries # 1 and # 2 are input in the case where the user's intentions # 1 and # 2 are input, it is considered that the user's intentions # 1 and intentions # 2 are similar.

Therefore, the information providing device 10 is a model for converting the character string of the search query into a vector, and the vector generated from the first search query input by the user and the user input after the first search query. The intention estimation model M1 is trained so as to be similar to the vector generated from the second search query. That is, the information providing device 10 considers that the continuous search queries indicate the same or similar intentions of the users, and the second search that the user inputs after the first search query from the first search query. The intent estimation model M1 that predicts the query and outputs the vector corresponding to the second search query is trained.

For example, the information providing device 10 inputs the search query # 1 to the intention estimation model M1 and causes the intention estimation model M1 to generate the intention vector # 1. Further, the information providing device 10 inputs the search query # 2 into the intention estimation model M1 to generate the intention vector # 2. Then, the information providing device 10 learns the intention estimation model M1 so that the intention vector # 1 and the intention vector # 2 are similar to each other. By executing such processing for each search query continuously input by the user in the same session, the information providing device 10 learns the intention estimation model M1 that estimates the user's intention from the search query. It can be realized.

Here, any kind of model can be adopted as such an intention estimation model. For example, the information providing device 10 may adopt an SVM (Support Vector Machine) or a DNN (Deep Neural Network) as a model. Here, the DNN may be a CNN (Convolutional Neural Network) or an RNN (Recurrent Neural Network). Further, the RNN may be an LSTM (Long short-term memory) or the like. That is, any type of model can be adopted as the intention estimation model. Further, the intention estimation model may be a model realized by combining a plurality of models, for example, a model in which a CNN and an RNN are combined.

On the other hand, considering that the user's intention changes continuously and rarely changes discretely, the user before a certain point in order to better estimate the user's intention. A method of estimating the user's intention from the transition of the search query entered by is also conceivable. For example, the information providing device 10 uses a recurrent neural network, that is, RNN or LSTM, which changes the output according to the contents of information input in the past and the order in which each information is input, as the intention estimation model M1. Therefore, it is considered that the user's intention can be better estimated.

For example, the information providing device 10 prepares an LSTM that is an intention estimation model M1. Then, the information providing device 10 causes the intention estimation model M1 to generate the intention vector # 1 by inputting the search query # 1 to the intention estimation model M1. Subsequently, the information providing device 10 inputs the search query # 2 to the intention estimation model M1 in which the search query # 1 is input, and causes the intention estimation model M1 to generate the intention vector # 2. In such a case, the intention estimation model M1 which is an LSTM inputs the contents of the search query # 1 and the search query # 2 input by the user, and the search query # 2 after inputting the search query # 1. Based on the order, the intention vector # 2 will be generated. Then, the information providing device 10 learns the intention estimation model M1 so that the intention vector # 1 and the intention vector # 2 are similar to each other.

Subsequently, the information providing device 10 inputs the search query # 3 input by the user after the search query # 2 to the intention estimation model M1 in which the search query # 2 is input, and inputs the intention vector to the intention estimation model M1. Generate # 3. In such a case, the intention estimation model M1 which is an LSTM includes the contents of the search query # 1, the search query # 2, and the search query # 3 input by the user, and the search query # after inputting the search query # 1. The intention vector # 3 is generated based on the order in which 2 is input and then the search query # 3 is input. Then, the information providing device 10 learns the intention estimation model M1 so that the intention vector # 2 and the intention vector # 3 are similar to each other.

When such learning is performed, the information providing device 10 estimates the intention so as to output the intention vector according to the order in which each search query is input, together with the characteristics of each search query continuously input by the user. The model M1 can be trained. As a result, the information providing device 10 can realize learning of the intention estimation model M1 that estimates the user's intention based on, for example, the characteristics and the input order of the search query input by the user in the past.

[1-2-2. About the transition estimation model]
Returning to FIG. 1, the explanation will be continued. For example, the intention estimation model M1 learned in step S4 can estimate the intention when the user inputs the search query from the search query input by the user. On the other hand, if the characteristics of such a transition of the user's intention are further learned, it is considered that a model for estimating the future intention of the user can be realized.

Therefore, when the information providing device 10 learns the intention estimation model, the information providing device 10 causes the transition estimation model M2 to learn the characteristics of the intention transition estimated by the intention estimation model (step S5). More specifically, the information providing device 10 generates intention information indicating the user's intention when performing the action indicated by the action information from the search query input by the user. For example, the information providing device 10 uses the above-mentioned intention estimation model M1 to generate an intention vector indicating the user's intention as intention information. Then, the information providing device 10 causes the transition estimation model M2, which is different from the intention estimation model M1, to learn the characteristics of the intention transition indicated by the generated intention information.

For example, the information providing device 10 causes the transition estimation model M2 to learn the characteristics of the transition of the vector output by the intention estimation model M1 when the behavior information is input to the intention estimation model M1. More specifically, the information providing device 10 inputs a vector generated by the intention estimation model M1 from the first action information to the transition estimation model M2, a vector output by the transition estimation model M2, and a second action. The transition estimation model M2 is trained so that the vector generated by the intention estimation model M1 from the information is similar.

By executing such a learning process for a series of actions of the user (for example, a plurality of search queries continuously input by the user), the information providing device 10 intends to indicate the intention of the user. The characteristics of the vector transition can be trained in the transition estimation model M2. As a result, the information providing device 10 can appropriately estimate the future intention of the user when the user inputs a certain search query by using the intention estimation model M1 and the transition estimation model M2. ..

For example, FIG. 3 is a diagram showing an example of learning of a transition estimation model executed by the information providing device according to the embodiment. In the example shown in FIG. 3, it is assumed that the user inputs the search query in the order of search query # 1, # 2, and # 3. For example, the information providing device 10 generates an intention vector # 1 from the search query # 1 by using the trained intention estimation model M1. Subsequently, the information providing device 10 generates an intention vector # 2 from the search query # 2 for the intention estimation model M1.

Further, the information providing device 10 prepares the LSTM as the transition estimation model M2 in the same manner as the intention estimation model M1. In addition to the LSTM, the information providing device 10 may use a model of any format as the transition estimation model M2. Then, the information providing device 10 has an estimation vector # 1 output by the transition estimation model M2 when the intention vector # 1 is input to the transition estimation model M2, and an intention vector # 1 generated by the intention estimation model M1 from the search query # 2. The transition estimation model M2 is trained so as to be similar to 2.

Subsequently, the information providing device 10 uses the estimation vector # 2 output by the transition estimation model M2 when the intention vector # 2 is input to the transition estimation model M2 in which the intention vector # 1 is input, and the intent estimation model M1 as a search query. The transition estimation model M2 is trained so that the intention vector # 3 generated from # 3 is similar. Further, the information providing device 10 sequentially inputs each search query continuously input by the user into the intention estimation model M1, and changes and estimates the characteristics of the transition of the intention vector # 1 output by the intention estimation model M1 in order. Train model M2. As a result of such processing, the information providing device 10 can make the transition estimation model M2 learn how to gradually change the intention of the user. As a result, the information providing device 10 can realize, for example, a model for estimating the future intention of the user from a plurality of search queries continuously input by the user.

In the example shown in FIG. 3, the information providing device 10 has learned the transition estimation model M2 so as to estimate the intention of inputting the search query next when the user inputs a certain search query. However, the embodiments are not limited to this. As will be described later, the information providing device 10 estimates the intention information indicating the intention when the user performs the predetermined conversion from the transition of the plurality of search queries originally input by the user with the predetermined intention. , The transition estimation model M2 may be trained.

For example, the information providing device 10 sequentially inputs the search queries # 1, # 2, # 3, and # 4 as an intention estimation model when the search queries # 1, # 2, # 3, and # 4 are continuously input. By inputting to M1, the intention vectors # 1, # 2, # 3, and # 4 are generated. Subsequently, the information providing device 10 inputs the intention vectors # 1, # 2, and # 3 in order to the transition estimation model M2, and the estimation vector finally obtained (that is, when the intention vector # 3 is input). The intention estimation model M2 may be trained so that the vector output by the transition estimation model M2) is similar to the intention vector # 4. That is, the information providing device 10 may learn the intention estimation model M2 so as to output an intention vector indicating the intention of the user in the future regardless of the number of steps of the user's action as an input.

That is, in the information providing device 10, when the transition estimation model M2 performs a process for the search query # 3 (a process for outputting the estimated vector # 3 from the intention vector # 3), the search query # 2 and the search query # 1 Etc., the transition estimation model M2 is trained so that the information obtained in the past estimation is inherited as a context, that is, the processing based on the past information is recursively performed.

[1-2-3. About provision processing]
Returning to FIG. 1, the explanation will be continued. Following step S5 described above, the information providing device 10 executes a providing process for providing information according to the future intention of the user. For example, the information providing device 10 predicts the intentions of the future actions of both parties by using the history of the search query and the transition estimation model M2 (step S6).

For example, FIG. 4 is a diagram showing an example of a process in which an information providing device according to an embodiment predicts an intention of a future action by using a transition estimation model. It is assumed that the intention estimation model M1 and the transition estimation model M2 shown in FIG. 4 are models that have been trained by the above-mentioned learning process. For example, the information providing device 10 learns the intention estimation model M1 and the transition estimation model M2 by using the search query input by the user before the latest search query # 10 input by the user. Subsequently, the information providing device 10 inputs the latest search query # 10 input by the user into the intention estimation model M1 and generates the intention vector # 10. Then, the information providing device 10 inputs the intention vector # 10 into the transition estimation model M2 and generates the estimation vector # 10.

Subsequently, the information providing device 10 converts the estimation vector # 10 into action information such as a search query. That is, the information providing device 10 is covered by the user when he / she has the estimated intention from the estimation vector # 10 indicating the future user's intention estimated from the characteristics of the transition of the user's intention. Estimate deaf behavior. For example, the information providing device 10 estimates from the estimation vector # 10 a search query # 11 that the user will input when having the intention (estimation intention # 1) indicated by the estimation vector # 10.

The conversion from such an estimated vector to behavioral information such as a search query is performed by learning the relationship between the estimated vector generated from a certain search query and the search query input next to the search query. It is feasible. For example, the information providing device 10 trains the conversion model so that when the estimation vector # 1 generated from the search query # 1 is input, the search query # 2 input after the search query # 1 is output. By doing so, a transformation model that transforms the estimation model into a search query may be generated.

Subsequently, the information providing device 10 generates the intention vector # 11 by inputting the search query # 11 into the intention estimation model M1. Then, the information providing device 10 inputs the intention vector # 11 into the transition estimation model M2 and generates the estimation vector # 11. Such an estimation vector # 11 is considered to indicate an estimation intention # 2 that is estimated to be possessed by the user after having the estimation intention # 1. Therefore, the information providing device 10 converts the estimation vector # 11 into the search query # 12 by using the conversion model.

Then, the information providing device 10 generates the intention vector # 12 from the search query # 12 using the intention estimation model M1, and generates the estimation vector # 12 from the generated intention vector # 12 using the transition estimation model M2. do. Then, the information providing device 10 estimates the future behavior of the user based on the generated estimation vector # 12. For example, the information providing device 10 converts the estimation vector # 13 into a search query using the future intention # 3 estimated to be held after the future intention # 2 by the user and the estimation vector # 12 using the conversion model. By doing so, the search query that the user will enter when having the intention # 3 in the future is predicted.

That is, the information providing device 10 estimates the transition of the intention of the target user from the behavior information of the target user to be the estimation target of the transition of the intention by using the transition estimation model M2. More specifically, the information providing device 10 estimates the future behavior of the target user based on the transition of the intention of the target user estimated from the behavior information, and the user when the future behavior is performed. The future intention information indicating the intention of the target user is generated, and the future behavior of the target user is estimated from the future intention information by using the transition estimation model M2.

Returning to FIG. 1, the explanation will be continued. When the information providing device 10 predicts the intention of the future action of the user, the information providing device 10 generates information according to the prediction result and provides it to the user U (step S7). For example, the information providing device 10 executes the process shown in FIG. 4 once or repeatedly, so that the user's intention after a predetermined period has elapsed and the action performed by the user (for example, a search input by the user). Predict the query). Then, the information providing device 10 generates the content according to the prediction result, and provides the generated content to the user U.

To give a more specific example, the information providing device 10 generates an estimation vector # 10 corresponding to the future intention # 1 from the search query # 10 input last by the user, and from the generated estimation vector # 10. , Generates search query # 11 that the user will enter when having intent # 1 in the future. Subsequently, the information providing device 10 generates an estimation vector # 11 corresponding to the estimation intention # 2 estimated to be possessed after the user has the estimation intention # 1 from the search query # 11, and the generated estimation. From the vector # 11, generate a search query # 12 that the user will enter when having intent # 2 in the future. Then, the information providing device 10 generates an estimation vector # 12 corresponding to the estimation intention # 3 that the user will have next to the estimation intention # 2 from the generated search query # 2, and the generated estimation vector #. From 12, generate search query # 13 that the user will enter in the future.

Further, the information providing device 10 collects the content indicating the transaction target corresponding to the search query # 13, the news content corresponding to the search query # 13, and the like. Then, the information providing device 10 provides the collected contents to the user U. As a result of such processing, the information providing device 10 can provide the content according to the intention that the user will have in the future.

[1-3. Variations on the processing executed by the information provider]
In the above example, the information providing device 10 uses the search queries continuously input by the user, considers the intention of the user in the future, and estimates the search query that the user will input in the future. An example has been described. However, the above description is merely an example, and the embodiment is not limited thereto. Hereinafter, variations of the processing executed by the information providing device 10 will be described.

[1-3-1. About conversion]
In the above example, the information providing device 10 generates the intention vector # 1 from the search query # 1 based on the intention # 1, and the estimation vector # 1 generated from the intention vector # 1 is the search query based on the intention # 2. The transition estimation model M2 was trained so as to be similar to the intention vector # 2 generated from # 2. However, the embodiments are not limited to this.

For example, when the transition estimation model M2 is learned so that the estimation vector # 1 and the intention vector # 2 are similar, the information providing device 10 uses the input of the search query # 2 by the user as a conversion and the user searches. The transition estimation model M2 that predicts the user's conversion is learned from the search query # 1 entered before the query # 2 is entered. Here, the information providing device 10 may learn the transition estimation model M2 so as to estimate the user's intention when performing a predetermined conversion that occurs thereafter from the transition of the search query.

For example, the information providing device 10 generates an intention vector from each search query using the intention estimation model M1 until the user inputs a predetermined search query (that is, until the user executes a predetermined conversion). The generated intention vector is input to the sequential transition estimation model M2. Here, the information providing device 10 may train the transition estimation model M2 so that the estimation vectors continuously generated by the transition estimation model M2 are similar. Then, the information providing device 10 learns the transition estimation model M2 so that the intention vector when the user inputs a predetermined search query and the estimation vector generated by the previous transition estimation model M2 are similar. May be good.

When such a process is executed, the information providing device 10 can make the transition estimation model M2 learn the characteristics of the transition of the user's intention leading up to the conversion. As a result, the information providing device 10 learns the transition estimation model M2 that estimates the transition of the user's intention from various behaviors of the user such as a search query input by the user to conversion. be able to.

The information providing device 10 may adopt an arbitrary action as a conversion. For example, the information providing device 10 may adopt the input of a search query including a predetermined keyword, such as the input of a search query including a transaction target, as a conversion. Further, as will be clarified by the explanation described later, the information providing device 10 uses not only a predetermined search query but also, for example, selection and browsing of a predetermined content, movement to a predetermined store, and various payment actions as conversions. You may adopt it.

[1-3-2. About user behavior]
In the above description, the information providing device 10 learns the intention estimation model M1 for estimating the user's intention by using the search query continuously input by the user, and the intention estimation model M1 is used from the search query. The transition estimation model M2 was trained to learn the characteristics of the transition of the generated intention vector. However, the embodiments are not limited to this.

For example, the information providing device 10 estimates the user's intention from various contents selected by the user by clicking or tapping, and the intention estimation model M1 or the transition estimation model M2 has the characteristics of the transition of the estimated intention. You may study. Further, the information providing device 10 estimates the user's intention from the contents such as news and advertisements browsed by the user and posted information of SNS (Social Networking Service), and intends to have the characteristics of the transition of the estimated intention. The estimation model M1 and the transition estimation model M2 may be trained. For example, the information providing device 10 may learn the intention estimation model M1 that converts the content selected or viewed by the user into an intention vector, and may learn the characteristics of the transition of the vector generated by the intention estimation model M1.

For example, the information providing device 10 may learn the intention estimation model M1 that converts the content posted on the SNS or the like by the user into an intention vector, and may learn the characteristics of the transition of the vector generated by the intention estimation model M1. Further, the information providing device 10 provides information on the transaction target purchased by the user and the transaction target in which the user browses related contents such as sales contents (for example, the type, name, category, price, etc. of the transaction target). The intention estimation model M1 to be converted into an intention vector may be learned, and the characteristics of the transition of the vector generated by the intention estimation model M1 may be learned.

Further, the information providing device 10 acquires position information indicating the user's position using a positioning system such as GPS (Global Positioning System), and converts the acquired user's position into an intention vector. It may be learned and the characteristics of the transition of the vector generated from the position information may be learned by using the intention estimation model M1. Further, the information providing device 10 learns the intention estimation model M1 that converts the store visited by the user into an intention vector, and uses the intention estimation model M1 to characterize the transition of the vector generated from the store visited by the user. You may learn.

Further, the information providing device 10 acquires action information indicating an action such as a walking state or movement of the user based on various sensor information possessed by various wearable devices and the terminal device 100, and obtains the acquired action information. The intention estimation model M1 to be converted into an intention vector may be learned, and the characteristics of the transition of the vector generated from the action information may be learned using the intention estimation model M1.

That is, the information providing device 10 learns an intention estimation model M1 that acquires behavior information indicating an arbitrary behavior indicating the user's intention and generates an intention vector indicating the user's intention from the acquired behavior information. Then, the information providing device 10 estimates the future intention of the user from various behaviors of the user by letting the transition estimation model M2 learn the characteristics of the transition of the intention vector generated by using the intention estimation model M1. May be good.

For example, FIG. 5 is a diagram showing an example of learning of a transition estimation model executed by an information providing device according to an embodiment using behavior information. For example, the information providing device 10 continuously performs the intention vector # 1 output by the intention estimation model M1 when the action information indicating the action # 1 is input to the intention estimation model M1 and the action # 1 by the user. The intention estimation model M1 is trained so as to be similar to the intention vector # 2 output by the intention estimation model M2 when the action information indicating the action # 2 is input to the intention estimation model M1.

That is, the information providing device 10 has a similar intention when the user performs the first action and when the user performs the second action within a predetermined period after performing the first action. Intention information (eg, intention vector) is generated. Further, the information providing device 10 assumes that the intention when the user performs the first action is similar to the intention when the user performs the second action continuously performed with the first action. , Intentional information may be generated. More specifically, the information providing device 10 outputs a vector similar to that when the first action information indicating the first action is input and when the second action information indicating the second action is input. Intention information is generated from each behavior information by using the intention estimation model M1 that has been trained in.

Further, the information providing device 10 generates intention vectors # 1 to # 3 from each action # 1 to # 3 by using the trained intention estimation model M1. Then, the information providing device 10 resembles the estimation vector # 1 output by the transition estimation model M2 when the intention vector # 1 is input to the transition estimation model M2, and the transition estimation model M2 so that the intention vector # 2 is similar. The transition estimation model M2 is trained so that the estimation vector # 2 output by the transition estimation model M2 when the intention vector # 2 is input to the transition estimation model M2 and the intention vector # 3 are similar to each other. conduct.

In this way, the information providing device 10 acquires a plurality of behavioral information indicating the behavior of the user, and generates intention information indicating the intention of the user when the behavior indicated by the behavioral information is performed. Then, the information providing device 10 causes the transition estimation model M2 to learn the characteristics of the intention transition indicated by the generated intention information.

As a result of such processing, the information providing device 10 estimates the user's intention from the behavior information indicating various actions performed by the user, and based on the characteristics of the transition of the estimated intention, the future Estimate the user's intention. As a result, the information providing device 10 can provide various kinds of information based on the future intention of the user, so that more appropriate information can be provided to the user.

[1-3-3. About multiple types of actions]
In the above example, the information providing device 10 estimates the user's intention from the search query input by the user as the user's behavior, and causes the transition estimation model M2 to learn the characteristics of the estimated transition of the intention. .. Here, the information providing device 10 may perform learning using a plurality of types of actions as the actions of the user.

For example, when a user searches the Web, he / she may enter a search query and then click one of the links included in the search results to refer to the linked content. In consideration of such a series of user behaviors, the information providing device 10 may make the transition estimation model M2 learn the characteristics of the transition of the user's intention.

For example, FIG. 6 is a diagram showing an example of a process in which an information providing device according to an embodiment predicts an intention from a plurality of types of actions. For example, in the example shown in FIG. 6, after the user inputs the original search query # 1 of the intention # 1, the link is clicked from the original search result of the intention # 4, and then the original search query # of the intention # 3 is performed. It is assumed that 3 is entered.

In such a case, the information providing device 10 outputs the intention vector # 1 when the search query # 1 is input and the intention vector # 1 when the user inputs the action information indicating that the link is clicked. The intention estimation model M1 is trained so as to be similar to 4. The information providing device 10 may use, for example, information indicating a link clicked by the user, information indicating the content of the link destination, or the like as action information.

Then, the information providing device 10 outputs an estimation vector # 1 when the intention vector # 1 is input, which is similar to the intention vector # 4, and an estimation vector # 4 which is output when the intention vector # 4 is input. The transition estimation model M2 is trained so as to be similar to the intention vector # 3 output by the intention estimation model M1 when the search query # 3 is input. When such learning is performed, the intention estimation model M1 and the transition estimation model M2 estimate the user's intention from the behavior on the web such as the user's search query and click behavior, and estimate the future intention. Can be done.

The information providing device 10 may reflect the click action in the intention vector or the estimation vector by using a technique such as attention or a transformer that weights the input / output of the model in various ways. For example, the information providing device 10 may modify the value of a predetermined element of each element of the intention vector depending on whether or not the click action following the search query is performed. Further, the information providing device 10 does not need to learn, for example, search queries and click actions alternately. That is, the information providing device 10 learns the characteristics of the user's intention and the transition of the intention by using the action information indicating the action continuously executed by the user in the order in which the user performed the action. It may be realized.

[1-3-4. About search queries]
Further, the information providing device 10 may use information indicating an arbitrary search query as action information. For example, the information providing device 10 may use behavioral information indicating a plurality of search queries separated by one or a space. Further, the information providing device 10 may adopt not only a web search but also a search query input in various electronic shopping streets, a search query input in a map application, or the like. Further, the information providing device 10 may use the search query itself as an input, or may use a technique such as w2v to use information obtained by vectorizing the search query as action information.

For example, when the information providing device 10 adopts the search query as it is, for example, the search query is performed by inputting the character strings constituting the search query one by one from the beginning to the intention estimation model M1 which is RNN or LSTM. The intention vector may be generated from. Further, the information providing device 10 vectorizes one or more words input as a search query by using a conversion model such as w2v that converts words having similar meanings into similar vectors, and combines the vectors. The resulting vector, a composite vector for synthesizing each vector, or the like may be input to the intention estimation model M1 as action information.

[1-3-5. About the structure of the model]
In the above-mentioned example, the information providing device 10 generates intention information indicating an intention such as an intention vector from the behavior information using the intention estimation model M1, and uses the generated intention information to estimate the characteristics of the change of the intention. The model M2 was trained. However, the embodiments are not limited to this. For example, the information providing device 10 generates one model in which the trained intention estimation model M1 is combined with the trained transition estimation model M2, and the generated model is used from the user's behavior information. Information indicating actions based on the future intentions of the user may be output.

Further, the information providing device 10 may adopt various arbitrary neural networks such as a convolutional neural network in addition to the recurrent neural network such as RNN and LSTM. Further, the information providing device 10 receives input of the first intention information indicating the intention when the user performs the first action for the transition estimation model M2 realized by such various neural networks. Learning may be performed so as to output a vector similar to the case where the second intention information indicating the intention when the user performs the second action continuous with the first action is input.

[1-3-6. About projection on hyperbolic space]
Here, the model is trained so that the first vector output by the model when the first information is input and the second vector output by the model when the second information similar to the first information is input are similar. When the above is performed, the information providing device 10 can generate a model that projects the input information onto the space according to the similarity above. For example, when the information providing device 10 trains the model so that the cosine distances of the first vector and the second vector are similar, the information providing device 10 projects the information on the space reflecting the similarity of the information in the cosine distance. The model can be realized. Further, for example, when the information providing device 10 trains the model so that the Euclidean distances of the first vector and the second vector are similar, the information providing device 10 displays the information in the space reflecting the similarity of the information in the Euclidean distance. It is possible to realize a model to project.

However, the embodiments are not limited to this. For example, a technique is known to improve the accuracy of distributed representation of words by projecting distributed representation on hyperbolic space, which is an example of non-Euclidean space (“Poincare Embeddings for Learning Hierarchical Representations”, Maximilian Nickel, Douwe Kiela <Internet> https://arxiv.org/pdf/1705.08039.pdf (searched on September 13, 2019). For example, in Non-Patent Document 1, in the task of predicting words that come before and after a certain word, the distributed expression of the words is learned on the Poincare space, which is a hyperbolic space. In such a Poincare space, the closer to the edge of the space, the more the distance scale increases exponentially, so it is possible to project infinite information (points in space) in a finite space. can.

For example, consider a process of embedding branching information in a Poincare disk, which is an example of such a Poincare space. In such a Poincare disc, the scale of distance increases exponentially from the center toward the circumference. For this reason, when a tree that branches in the circumferential direction from the center of the Poincare disc is embedded, it has an arbitrary number of branches on the Poincare disc while keeping the angle and length of each branch constant. You can embed the tree naturally. Such characteristics of the Poincare disc are retained even in a higher-dimensional Poincare space.

On the other hand, the behavior of the user is considered to change due to the transition of the user's intention, and such a transition of the user's intention is also considered to have a tree structure. Therefore, when the user's behavior or intention is converted into a vector on the Poincare space, the hierarchical structure of the behavior or intention can be naturally embedded (projected) in the space. Further, in the Poincare space, the scale of the distance increases as it approaches the circumferential direction. Therefore, when learning a vector showing a user's behavior or intention in Poincare space, many hierarchical relationships can be reflected in the vector, which improves the accuracy of clustering and reduces the number of dimensions of the vector. It is thought that it can be done.

Therefore, the information providing device 10 learns the transition estimation model M2 so as to output the estimation vector obtained by projecting the intention vector generated by the intention estimation model M1 onto the hyperbolic space. For example, when the intention estimation model M1 generates the first intention vector from the first action and the second intention vector from the second action, the information providing device 10 generates a second estimation vector corresponding to the second intention vector. , The transition estimation model M2 is trained so that the vector branches from the first estimation vector corresponding to the first intention vector on the twin-fold space.

For example, the information providing device 10 projects a distributed expression on a Minkowski space as a hyperbolic space. More specifically, the information providing device 10 generates a distributed representation on a hyperplane whose coordinates on the focal axis are positive among the two hyperplanes of the Hyperboloid of two sheets in Minkowski space. do.

For example, the information providing device 10 sets a three-dimensional space composed of an x-axis, a y-axis, and a z-axis, and sets a two-leaf hyperboloid with the z-axis as a focal axis in the three-dimensional space. Further, the information providing device 10 sets a perpendicular line to the xy plane from the position indicated by the generated two-dimensional intention vector in the set three-dimensional space, and is a curved surface in which the component in the z-axis direction is positive among the two-leaf hyperboloids. The intersection of the vertical line and the vertical line is the projection destination of the distributed expression. That is, the information providing device 10 generates a distributed representation on the hyperbolic space indicating the intention of the user.

Then, the information providing device 10 determines the distance between the distributed expressions on the hyperbolic space. For example, the information providing device 10 generates an intention vector # 1 on the Euclidean space from the intention # 1 and generates an intention vector # 2 on the Euclidean space from the intention # 2. Then, the information providing device 10 calculates the distance in the hyperbolic space between the intention vector # 1 and the intention vector # 2.

Then, when the intention # 1 and the intention # 2 are similar (for example, when the intention corresponds to an action in the same session), the information providing device 10 is between the intention vector # 1 and the intention vector # 2. The transition estimation model M2 is trained so that the distance in the hyperbolic space of is smaller. On the other hand, in the information providing device 10, when the intention # 1 and the intention # 2 are not similar (for example, when the intention does not correspond to the action in the same session), the sentence # 1 and the sentence # 2 are not similar. To train the transition estimation model M2 so that the distance between the intention vector # 1 and the intention vector # 2 on the hyperbolic space becomes larger.

For example, the information providing device 10 fixes the function f that projects the intention vector onto the hyperbolic space, and corrects the connection coefficient that connects the nodes of the transition estimation model M2 by using a learning method such as backpropagation. You may. Further, the information providing device 10 may modify the function f that projects the intention vector onto the hyperbolic space. That is, the information providing device 10 may perform arbitrary learning as long as it learns the transition estimation model M2 that can project the similarity of intentions onto the distance on the hyperbolic space.

Here, the information providing device 10 can adopt any space as a hyperbolic space as long as it generates a distributed representation on the hyperbolic space. For example, the information providing device 10 may generate a distributed representation on a Minkowski space having an arbitrary number of dimensions. Further, the information providing device 10 may generate a distributed representation on an arbitrary surface of a two-leaf hyperboloid.

Further, the information providing device 10 may generate a distributed representation on a disk of any dimension in the Poincare disk model (so-called Poincare disk) or on a spherical surface of any dimension in the Poincare sphere model. That is, the information providing device 10 generates a distributed representation in space to which hyperbolic geometry can be applied, and the similarity of concepts that information has based on the distance using hyperbolic geometry in such space. If it is determined, any hyperbolic space may be adopted.

The information providing device 10 learns the intention estimation model M1 that projects each action information such as a search query onto the hyperbolic space, and further adds the vector generated by the trained intention estimation model M1 to the hyperbolic space. The transition estimation model M2 projected above may be trained.

When such learning is performed, the information providing device 10 learns the characteristics of the transition of the user's intention as the user's intention changes so as to follow the information group having a tree-like structure. It will be. As a result, the information providing device 10 can more accurately estimate the future intention of the user.

For example, FIG. 7 is a diagram illustrating an example of an effect when the information providing device according to the embodiment projects an intention on a hyperbolic space. For example, when a user continuously performs an action such as a web search, it is considered that the continuous action indicates the intention of the user to make a continuous transition. Here, considering that such user's intention changes according to the action and the action result, it can be considered that each user's intention can be expressed by a tree-type structure as shown in FIG. 7. ..

For example, when a user performs a web search under intention # 1, it is considered that the user's intention branches into intention # 1-1 or intention # 1-2 depending on the search result. In addition, when the user performs a web search under the intention # 1-1, the user's intention is a plurality of intentions # 1-1-1 and intention # 1-1-2 depending on the search result. It is thought that it branches to the intention of. In addition, when the user performs a web search under the intention # 1-2, the user's intention is a plurality of intentions # 1-2-1 and intention # 1-2-2 depending on the search result. It is thought that it branches to the intention of. In addition, the intention of the user is that the intention # 1-1-1, the intention # 1-1-2, the intention # 1-2-1, and the intention # 1-2-2 are different depending on the search result. It can be thought that it branches into.

When the transition of the intention of the user having such a tree-shaped structure is projected on the hyperbolic space, the distance between the intentions belonging to different branches on the hyperbolic space is kept at a predetermined value or more, and the same branch is formed. It is possible to reduce the distance between the intentions to which they belong. Therefore, the information providing device 10 can accurately train the transition estimation model M2 of the transition of the intention of the user having the tree-shaped structure.

Further, when the user's intention is projected on the hyperbolic space in this way, the information providing device 10 can estimate the user's future intention more accurately. For example, if the intent # 1-1-1 and the intent # 1-1-2 are presumed to indicate a web search-related intent that is a web search-related intent, the information providing device 10 is intended by the user. From the actions taken under 1-1, it is possible to presume that the future intention of the user is a web search-related intention, and automatically provide the result of the web search in advance. Further, for example, when the intention # 1-2-1 and the intention # 1-2-2 are presumed to indicate a purchase-related intention which is an intention related to the purchase of the transaction target, the information providing device 10 is used by the user. It is possible to presume that the future intention of the user is a purchase-related intention from the action performed under the intention # 1-2, and provide the result of the automatic product search in advance.

[1-3-7. About the learning unit of the model]
The information providing device 10 may learn a model common to all users, or may learn a dedicated model for each user. Further, for example, the information providing device 10 learns the intention estimation model M1 and the transition estimation model M2 that are common to all users based on the history of search queries of all users. Then, the information providing device 10 may tune the intention estimation model M1 and the transition estimation model M2 for each user. Further, the information providing device 10 distributes the figure estimation model M1 and the transition estimation model M2 common to all users to the terminal device 100 of each user, and the terminal device 100 side distributes the intention estimation model M1 and the transition estimation model M1 and the transition estimation. The model M2 may be tuned.

Here, the relationship between the user's behavior and intention is the position of the user, the date and time when the user took the action, whether the user was with the family or with a friend when the action was taken, etc. , It is considered that it changes according to the context of the user's state. For example, even if the user inputs a predetermined search query, the intention of the user and the transition of the intention are considered to be different depending on whether the user is at work or private. Be done.

Therefore, the information providing device 10 may learn the intention estimation model M1 and the transition estimation model M2, which are different for each state when the user performs an action. For example, the search server 200 may use not only a search query but also position information indicating a position acquired from a terminal device 100 using a positioning system such as GPS (Global Positioning System), an acceleration sensor or a gyro possessed by the terminal device 100. Information detected by various sensors such as sensors, temperature sensors, and illuminance sensors is acquired as context information. The search server 200 may acquire the context information acquired by the wearable device worn by the user via the terminal device 100.

Subsequently, the information providing device 10 acquires a search query and context information indicating the user's context when the search query is input from the search server 200. Then, the information providing device 10 classifies the search queries into groups according to the context of the user indicated by the context information, and uses the search queries of each group individually to learn the intention estimation model M1 and the transition estimation model M2. You may go. As a result of such processing, the information providing device 10 can learn the intention estimation model M1 and the transition estimation model M2 for each context.

The information providing device 10 may estimate the user's context from the context information by using any known technique. Further, even if the information providing device 10 acquires context information indicating the user's context estimated from various sensor information by various external servers (not shown), the terminal device 100, and the wearable device worn by the user. good.

Further, the information providing device 10 may provide information to the user by using the intention estimation model M1 that has been learned for each context information. That is, the information providing device 10 may estimate the user's intention for each context of the user and provide information according to the estimated intention without learning the transition estimation model M2. That is, the information providing device 10 acquires a plurality of action information indicating the user's action and state information indicating the state when the user performs each action, and the first action showing the user's first action. Indicates the second action performed by the user when the action information is input and when the user performs the second action after the first action in a state similar to that when the first action is performed. The second model may be trained so as to output a vector similar to that when the second action information is input.

[1-3-8. Applicable target]
Here, the information providing device 10 may estimate the user's intention from any type of behavior information and provide any type of information according to the estimated intention.

For example, the information providing device 10 may acquire the information selected by the user as action information. When learning the intention estimation model M1 and the transition estimation model M2 that estimate the user's intention from such behavior information, the information providing device 10 has the intention of the user selecting various information and the intention of the user. It is possible to learn the transition and estimate the future intention of the user from the information selected by the user. For example, the information providing device 10 indicates the intention of what kind of hotel or store the user is looking for from the click history on the travel site, gourmet site, or hotel reservation site, and what kind of hotel or store in the future. It is possible to estimate future intentions such as whether to make a reservation.

Further, for example, the information providing device 10 may acquire information related to the information browsed by the user as action information. When learning the intention estimation model M1 and the transition estimation model M2 that estimate the user's intention from such behavior information, the information providing device 10 determines the intention that the user selected the information and the transition of the intention. In addition to learning, the future intention of the user can be estimated from the information browsed by the user. For example, the information providing device 10 determines from the history of news the intention of the user when browsing the news, what kind of intention the user who browses the news has in the future, and what kind of news is browsed. The estimation can be realized.

Further, for example, the information providing device 10 may acquire information indicating a transaction target selected by the user as action information. When the intention estimation model M1 and the transition estimation model M2 that estimate the user's intention from such behavior information are learned, the information providing device 10 has the intention that the user has selected the transaction target and the transition of the intention. It is possible to estimate the future intention of the user from the transaction target that the user browses or purchases. For example, the information providing device 10 is the user's intention when the user purchases the transaction target or browses the transaction target sales page from the sales page of the transaction target viewed by the user, a search query in an electronic shopping district, or the like. It is possible to estimate (for example, whether you are worried about purchasing a transaction target, whether you are trying to purchase it, whether you are trying to collect information, etc.) and realize an estimation of what kind of transaction target you will purchase in the future. can.

In addition to the above-mentioned processing, the information providing device 10 may adopt arbitrary behavior information and estimate the user's intention and future intention based on the arbitrary behavior information. For example, the information providing device 10 uses various metadata indicating web contents (electronic books, music, movies, etc.) that the user has browsed or purchased as behavioral information, and is a web content that the user is interested in in the future. Predict metadata. Then, the information providing device 10 may propose to the user the web content of the metadata similar to the predicted metadata. For example, the information providing device 10 may adopt arbitrary information such as authors, producers, composers, directors, rankings, review contents, distributors, etc. of web contents as metadata. Further, the information providing device 10 may adopt the information of the influencer who proposes the content selected, purchased, and viewed by the user as metadata, and estimate the content that the user may be interested in in the future.

Further, the information providing device 10 learns the transition estimation model M2 by using the behavior information indicating the behavior until the user performs a predetermined conversion. Subsequently, as shown in FIG. 4, the information providing device 10 sequentially estimates future actions from the action information indicating the latest action executed by the user until the user is estimated to perform conversion. .. The number of times such an estimation is made can be considered to indicate the stage of intention until the user makes a conversion. Therefore, the information providing device 10 sequentially estimates future behavior until it is estimated that the user will perform conversion, and estimates the state of the user's intention until conversion is performed based on the number of times of estimation. However, information according to the estimated state may be provided.

For example, it is considered that there is a stage in which the user recognizes the transaction target, considers the purchase, and then purchases the transaction target before the user purchases the transaction target. For this reason, if the purchase of the transaction target by the user is converted, the user's intention is estimated from each user's behavior leading up to the conversion, and the transition of the estimated intention is learned from the model, the model is used. It is thought that they are learning the characteristics of the transition of intentions leading to conversion, such as recognizing the transaction target, considering the purchase, and then purchasing.

Using such a model, the information providing device 10 estimates the user's future intention step by step from the user's latest behavior, thereby determining the stage of the user's intention up to conversion. Can be estimated. For example, when the information providing device 10 estimates the future intention of the user one step or two steps ahead and the user is estimated to lead to conversion, the user considers the purchase of the transaction target. You may provide information that suggests the purchase of the transaction target. Further, for example, when the information providing device 10 estimates the future intention of the user five steps ahead and the user is estimated to lead to conversion, the user recognizes the transaction target. It may be estimated and provide a detailed explanation of the transaction target, reviews of other users, and the like.

In this way, the information providing device 10 may estimate the stage of the user's intention leading up to the conversion, and may provide information according to the estimated stage. When such a process is executed, the information providing device 10 can provide more appropriate information to the user.

The information providing device 10 may estimate the stage from the action information indicating the action of any type to the conversion of any type. For example, the information providing device 10 may estimate the stage from the history of search queries and the history of clicks to the time when the user makes a purchase for a transaction target or makes a reservation for a store. Further, for example, the information providing device 10 specifies the visit to the actual store and the use of the actual store based on the position history of the terminal device 100 and the payment history of the user, and the specified visit and use are converted. Then, the information providing device 10 learns the model by using the content, the search query, and the like browsed by the user before the conversion as action information. In such a case, the information providing device 10 can learn a model that estimates the stage from the content browsed by the user or the search query to the user visiting or using the store. The information providing device 10 may use the user's position history, store usage history, and the like as action information.

[2. Configuration of information providing device]
Hereinafter, an example of the functional configuration of the above-mentioned information providing device 10 will be described. FIG. 8 is a diagram showing a configuration example of the information providing device according to the embodiment. As shown in FIG. 8, the information providing device 10 includes a communication unit 20, a storage unit 30, and a control unit 40.

The communication unit 20 is realized by, for example, a NIC (Network Interface Card) or the like. Then, the communication unit 20 is connected to the network N by wire or wirelessly, and transmits / receives information to / from the terminal device 100.

The storage unit 30 stores reference information based on a reference image, which is an image of an object taken from a predetermined position. For example, the storage unit 30 is realized by a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk. Further, the storage unit 30 stores the action history database 31 and the model database 32.

Hereinafter, an example of information registered in the storage unit 30 will be described with reference to FIGS. 9 and 10. Behavior information indicating the behavior of the user is registered in the behavior history database 31. For example, FIG. 9 is a diagram showing an example of information registered in the behavior history database according to the embodiment. As shown in FIG. 9, the action history database 31 is associated with information such as "user ID (Identifier)", "user information", "context information", "behavior information", and "date and time information". It is registered.

Here, the "user ID" is an identifier for identifying a user. Further, the "user information" is information indicating various attributes such as a user's name, gender, age, hobbies and preferences, and a user's demographic attribute and psychographic attribute. Further, the "context information" is information indicating in what context (state) the action indicated by the associated action information was performed. Further, "behavior information" is information indicating the behavior of the user. Further, the "date and time information" is information indicating the date and time when the action indicated by the associated action information was performed or the date and time when the action information was acquired.

For example, in the example shown in FIG. 9, the action history database 31 contains the user ID "user # 1", the user information "user information # 1", the context information "context # 1", and the action information "behavior information". Information such as "# 1-1" and date and time information "date and time information # 1-1" is registered. Such information indicates that the various attributes of the user indicated by the user ID "user # 1" are the user information "user information # 1". Further, such information can be used by the user indicated by the user ID "user # 1" to perform the action indicated by the action information "behavior information # 1-1" in the state indicated by the context information "context # 1". Date and time information Indicates that the service was performed at the date and time indicated by "date and time information # 1-1".

In addition to the information shown in FIG. 9, arbitrary information may be registered in the action history database 31. Further, in the example shown in FIG. 9, conceptual such as "user # 1", "user information # 1", "context # 1", "behavior information # 1-1", and "date and time information # 1-1". However, in reality, a character string or numerical value for identifying the user, a character string indicating the user's attribute, a character string or numerical value for identifying the context, or a numerical value indicating the user's behavior. , A character string, a character string entered by the user, a numerical value indicating the date and time, etc. will be registered.

The intention estimation model M1 and the transition estimation model M2 are registered in the model database 32. For example, FIG. 10 is a diagram showing an example of information registered in the model database according to the embodiment. As shown in FIG. 10, information such as "model ID", "model type", "corresponding user", and "corresponding context" is registered in the model database 32 in association with each other.

Here, the "model ID" is an identifier for identifying the model. Further, the "model type" is information indicating whether the model is an intention estimation model or a transition estimation model. Further, the "corresponding user" is information indicating a user who estimates the intention and the transition of the intention using the associated model. Further, the "correspondence context" is information indicating the user's context when estimating the intention or the transition using the associated model.

For example, in the example shown in FIG. 10, information such as the model ID "model # 1", the model type "transition estimation model", the corresponding user "user # 1", and the corresponding context "context # 1" are registered in association with each other. Has been done. For such information, the model type indicated by the model ID "model # 1" is the model type "transition estimation model", and the user indicated by the corresponding user "user # 1" is the corresponding context "context # 1". It is shown that the model is a model in which learning is performed based on the actions performed in the state shown in 1 ”.

In addition to the information shown in FIG. 10, arbitrary information may be registered in the model database 32. Further, in the example shown in FIG. 10, conceptual values such as "model # 1", "user # 1", and "context # 1" are described, but in reality, the model and the user are identified. A character string, a numerical value, a character string, a numerical value, etc. for identifying a context will be registered.

Returning to FIG. 8, the explanation will be continued. The control unit 40 is a controller, and for example, various programs stored in a storage device inside the information providing device 10 by a processor such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit) are RAM or the like. Is realized by executing as a work area. Further, the control unit 40 is a controller, and may be realized by, for example, an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

As shown in FIG. 8, the control unit 40 includes an acquisition unit 41, a generation unit 42, a learning unit 43, a reception unit 44, an estimation unit 45, and a provision unit 46.

The acquisition unit 41 acquires a plurality of behavior information indicating the behavior of the user. For example, the acquisition unit 41 acquires behavior information indicating the behavior performed by the user on the Web. To give a more specific example, the acquisition unit 41 acquires information input by the user, such as a search query and posted information, as action information. Further, the acquisition unit 41 may acquire various information such as the content clicked by the user as action information, and acquires information related to the content such as the content viewed by the user and the metadata of the content as action information. You may. Further, the acquisition unit 41 may acquire information indicating a transaction target selected by the user, such as a transaction target viewed or purchased by the user, as action information.

The acquisition unit 41 acquires the action information, as well as the date and time information indicating the date and time when the action indicated by the action information is performed, and the context information indicating the user's context when the action is performed. Then, the acquisition unit 41 registers the user ID indicating the user, the action information, and the context information in the action history database 31 in association with each other. Here, when the context information is various sensor information, the acquisition unit 41 estimates which of the predetermined states the user is in, and associates it with the context information indicating the estimated state. Behavior information may be registered.

The generation unit 42 generates intention information indicating the intention of the user when the action indicated by the action information is performed from each action information. For example, the generation unit 42 acquires the behavior information of the user to be learned of the transition estimation model M2 from the behavior history database 31, and estimates the intention of the user to be learned of the transition estimation model M2 from the model database 32. Read model M1. When the generation unit 42 learns the transition estimation model M2 for each combination of the user and the context, the generation unit 42 uses the behavior information associated with the combination of the user and the context to be learned and the intention estimation model M1. It may be read out.

Then, the generation unit 42 generates the intention vector by inputting the action information into the intention estimation model M1. More specifically, the generation unit 42 inputs the action information to the intention estimation model M1 which is the LSTM in the order in which the user performed (for example, the date and time indicated by the date and time information is in the oldest order), thereby inputting the intention vector. Is generated.

Here, in the learning unit 43, the intention when the user performs the first action is similar to the intention when the user performs the second action within a predetermined period after performing the first action. As a result, the intention estimation model M1 is trained. That is, the learning unit 43 assumes that the intention when the user performs the first action is similar to the intention when the user performs a predetermined second action after the first action. The intention estimation model M1 such as LSTM is learned.

Therefore, the generation unit 42 outputs a vector when the first action information is input and a vector output when the second action information is continuously input to the first action information as the intention estimation model M1. Intention information is generated using a recursive neural network trained to be similar to. For example, the generation unit 42 inputs the action information into the intention estimation model M1 and generates the vector output by the intention estimation model M1 as the intention information. That is, the generation unit 42 makes the first intention vector generated from the first action information indicating the first action similar to the second intention vector generated from the second action information indicating the second action. The intention vector will be generated.

The learning unit 43 causes the transition estimation model M2 to learn the characteristics of the intention transition indicated by the intention information. For example, the learning unit 43 prepares a recurrent neural network such as LSTM as a transition estimation model M2 to be learned. Then, the learning unit 43 causes the transition estimation model M2 to learn the characteristics of the transition of the vector generated by the intention estimation model M1. For example, the learning unit 43 trains the transition estimation model M2 by using various known techniques such as backpropagation.

For example, the learning unit 43 generated a vector generated by the transition estimation model M2 when the first intention vector generated by the intention estimation model M1 was input from the first behavior information, and an intention estimation model M1 generated from the second behavior information. The transition estimation model M2 may be trained so as to be similar to the second intention vector. More specifically, the learning unit 43 uses the transition estimation model M2 as a first vector to be output when the first intention vector generated by the intention estimation model M1 is input when the first action information is input. , Recursive so that it is similar to the second vector output when the second intention vector generated by the intention estimation model M1 is input when the second action information is input in succession to the first intention vector. You may also learn the type neural network.

Further, in the learning unit 43, the first action generates a plurality of first intention vectors generated by the intention estimation model M1 from the first action information indicating the plurality of first actions performed before the user performs the second action. The transition estimation model M2 so that the estimation vector generated by the transition estimation model M2 when input to the transition estimation model M2 in the order in which it was performed is similar to the second intention vector generated by the intention estimation model M1 from the second behavior information. You may study. That is, the learning unit 43 may make the transition estimation model M2 learn the characteristics of the transition of intention until the user performs a predetermined conversion. In other words, the learning unit 43 has a second intention indicating the intention when the user performs the second action when the first intention information indicating the intention when the user performs the first action is input. The transition estimation model M2 may be trained to output a vector similar to the case where the information is input.

The learning unit 43 may learn the transition estimation model M2 so as to output a vector obtained by projecting the vector generated by the intention estimation model M1 onto the hyperbolic space. For example, the learning unit 43 may learn the transition estimation model M2 so as to project the second intention vector from the first vector to the branching vector on the hyperbolic space.

Further, the learning unit 43 may learn the transition estimation model M2, which is different for each user or for each state in which the user has performed an action. Further, the learning unit 43 may perform learning of the intention estimation model M1 for each user and for each state in which the user has performed an action.

The reception unit 44 receives the behavior information of the target user who is the target of estimating the transition of the intention. For example, when the reception unit 44 receives the designation of the target user from an external server or the like (not shown), the reception unit 44 receives the action information within a predetermined period (for example, the latest action information, etc.) among the action information executed by the target user. ) Is accepted from the search server 200 or the like.

The estimation unit 45 estimates the transition of the intention of the target user from the behavior information received by the reception unit 44 by using the transition estimation model M2 learned by the learning unit 43. For example, the estimation unit 45 reads out the intention estimation model M1 and the transition estimation model M2 of the target user from the model database 32. The estimation unit 45 may read the intention estimation model M1 and the transition estimation model M2 according to the user's state from the model database 32.

Subsequently, the estimation unit 45 inputs the behavior information received by the reception unit 44 into the intention estimation model M1, and inputs the intention vector generated by the intention estimation model M1 into the transition estimation model M2, whereby the future of the user is reached. Estimate the intent of the action. Subsequently, the estimation unit 45 estimates the future behavior of the target user from the estimated intention. For example, the estimation unit 45 estimates future behavior information from the estimation vector by using a conversion model that converts the estimation vector into behavior information. In such a case, the generation unit 42 further generates future intention information indicating the intention of the future action by inputting the estimated behavior information into the intention estimation model M1. Then, the estimation unit 45 estimates the future behavior of the target user from the generated future intention information by using the transition estimation model M2. The estimation unit 45 may further estimate future intentions and actions by repeatedly executing such processing.

The providing unit 46 provides information according to the estimated future intention or behavior of the target user. For example, the providing unit 46 generates information according to the future behavior or intention estimated by the estimation unit 45, and transmits the generated information to the terminal device 100. For example, when the estimation unit 45 estimates a search query to be input by the user in the future, the providing unit 46 may provide the content corresponding to the estimated search query in advance.

Further, for example, the estimation unit 45 repeatedly executes the above-mentioned process until the user is estimated to execute a predetermined conversion, and the number of times the process is executed until the user is estimated to execute the predetermined conversion. Count. Then, the estimation unit 45 estimates the stage until the user reaches a predetermined conversion according to the number of times counted. In such a case, the providing unit 46 may provide information according to the estimated stage.

[3. An example of the flow of processing executed by the information providing device]
Subsequently, the flow of processing executed by the information providing apparatus 10 will be described with reference to FIG. FIG. 11 is a flowchart showing an example of a flow of processing executed by the information providing apparatus according to the embodiment.

First, the information providing device 10 determines whether or not it is the execution timing of the learning process (step S101), and if it is determined that it is the execution timing (step S101: Yes), acquires the action information (step S102). ). Subsequently, the information providing device 10 learns the intention estimation model M2 so that a similar vector is output when the behavior information indicating continuous behavior is input (step S103). Then, the information providing device 10 causes the transition estimation model M2 to learn the characteristics of the vector transition output by the intention estimation model M2 (step S104).

Subsequently, the information providing device 10 determines whether or not it is the execution timing of the providing process (step S105). Then, when the information providing device 10 determines that it is the execution timing of the providing process (step S105: Yes), the information providing device 10 uses each model to generate an estimation vector corresponding to the future intention from the latest action information. (Step S106). Subsequently, the information providing device 10 estimates the behavior corresponding to the estimation vector (step S107), and further estimates the future user's intention from the behavior estimated for each model (step S108). Then, the information providing device 10 selects the information to be provided to the user (step S109) based on the estimated intention and the corresponding action, provides the selected information (step S110), and ends the process.

If the information providing device 10 determines that it is not the execution timing of the learning process (step S101: No), the information providing device 10 executes step S105. If the information providing device 10 determines that it is not the execution timing of the providing process (step S105: No), the information providing device 10 ends the process.

[4. Modification example]
In the above, an example of the processing by the information providing apparatus 10 has been described. However, the embodiments are not limited to this. Hereinafter, variations of the processing executed by the information providing device 10 will be described.

[4-1. Execution subject]
In the above example, the information providing device 10 has executed the learning of each model. However, the embodiments are not limited to this. For example, the above-mentioned learning process may be executed standalone by the terminal device 100, or may be executed by the terminal device 100 and the information providing device 10 in cooperation with each other, for example. For example, the terminal device 100 may tune the model generated by the information providing device 10 for each of a plurality of users on the terminal device 100 side for each user, or may tune according to the user's context. May be done. That is, the terminal device 100 may execute a learning process of modifying the shared model generated by the information providing device 10 according to the user or the user's own context by using the edge computing technique. In such a case, the information providing device 10 does not need to acquire a large amount of behavior information for all users, so that the privacy of each user can be protected.

[4-2. others〕
Further, among the processes described in the above-described embodiment, all or a part of the processes described as being automatically performed can be manually performed, and conversely, the processes described as being manually performed. It is also possible to automatically perform all or part of the above by a known method. In addition, information including processing procedures, specific names, various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified. For example, the various information shown in each figure is not limited to the information shown in the figure.

Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of them may be functionally or physically distributed / physically in any unit according to various loads and usage conditions. Can be integrated and configured.

In addition, the above-described embodiments can be appropriately combined as long as the processing contents do not contradict each other.

[4-3. program〕
Further, the information providing device 10 according to the above-described embodiment is realized by, for example, a computer 1000 having a configuration as shown in FIG. FIG. 12 is a diagram showing an example of a hardware configuration. The computer 1000 is connected to the output device 1010 and the input device 1020, and the arithmetic unit 1030, the primary storage device 1040, the secondary storage device 1050, the output IF (Interface) 1060, the input IF 1070, and the network IF 1080 are connected by the bus 1090. Have.

The arithmetic unit 1030 operates based on a program stored in the primary storage device 1040 or the secondary storage device 1050, a program read from the input device 1020, or the like, and executes various processes. The primary storage device 1040 is a memory device that temporarily stores data used by the arithmetic unit 1030 for various operations such as RAM. Further, the secondary storage device 1050 is a storage device in which data used by the calculation device 1030 for various calculations and various databases are registered, such as a ROM (Read Only Memory), an HDD (Hard Disk Drive), and a flash memory. Is realized by.

The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various information such as a monitor and a printer. For example, USB (Universal Serial Bus), DVI (Digital Visual Interface), and the like. It is realized by a connector of a standard such as HDMI (registered trademark) (High Definition Multimedia Interface). Further, the input IF 1070 is an interface for receiving information from various input devices 1020 such as a mouse, a keyboard, a scanner, and the like, and is realized by, for example, USB.

The input device 1020 is, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), a PD (Phase change rewritable Disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), or a tape. It may be a device that reads information from a medium, a magnetic recording medium, a semiconductor memory, or the like. Further, the input device 1020 may be an external storage medium such as a USB memory.

The network IF 1080 receives data from another device via the network N and sends it to the arithmetic unit 1030, and also transmits the data generated by the arithmetic unit 1030 to the other device via the network N.

The arithmetic unit 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic unit 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040, and executes the loaded program.

For example, when the computer 1000 functions as the information providing device 10, the arithmetic unit 1030 of the computer 1000 realizes the function of the control unit 40 by executing the program loaded on the primary storage device 1040.

[5. effect〕
As described above, the information providing device 10 acquires a plurality of behavioral information indicating the behavior of the user, and generates intention information indicating the intention of the user when the behavior indicated by the behavioral information is performed. Then, the information providing device 10 causes the transition estimation model M2 to learn the characteristics of the intention transition indicated by the generated intention information. As a result of such processing, the information providing device 10 can realize the transition estimation model M2 for estimating the future intention of the user. Therefore, for example, as a result of realizing the information provision according to the future intention of the user, more. It is possible to provide appropriate information.

For example, in the information providing device 10, the intention when the user performs the first action is similar to the intention when the user performs the second action within a predetermined period after performing the first action. Generate intent information as it does. Further, for example, in the information providing device 10, the intention when the user performs the first action is similar to the intention when the user performs a predetermined second action after the first action. As a thing, the intention information is generated. As a result of such processing, the information providing device 10 can generate intention information that appropriately indicates the intention of the user, and as a result, the transition of the intention of the user can be learned accurately.

Further, the information providing device 10 learns to output a vector similar to the input of the first action information indicating the first action and the input of the second action information indicating the second action. Using the obtained intention estimation model M1, intention information is generated from each action information acquired by the acquisition unit. For example, the information providing device 10 outputs a vector when the first action information is input and a vector output when the second action information is continuously input to the first action information as the intention estimation model M1. Intention information is generated using a recursive neural network trained to be similar to. More specifically, the information providing device 10 inputs the behavior information into the intention estimation model M1 to generate the vector output by the intention estimation model M1 as the intention information, and the intention estimation model M1 is added to the transition estimation model M2. Learn the characteristics of the transition of the generated vector. Further, in the information providing device 10, the vector generated by the transition estimation model M2 from the vector generated by the intention estimation model M1 from the first action information and the vector generated by the intention estimation model M1 from the second action information are similar to each other. The transition estimation model M2 is trained.

As a result of each of the above-mentioned processes, the information providing device 10 can make the transition estimation model M2 learn the transition of the user's intention.

Further, the information providing device 10 performs the first action on a plurality of vectors generated by the intention estimation model M1 from the first action information indicating the plurality of first actions performed before the user performs the second action. The transition estimation model M2 is trained so that the vector generated by the transition estimation model M2 when input to the transition estimation model M2 in the same order and the vector generated by the intention estimation model M1 from the second behavior information are similar. Further, the information providing device 10 has, as the transition estimation model M2, a first vector to be output when the first intention vector generated by the intention estimation model M1 is input when the first action information is input, and a first vector. A recursive neural network so that it is similar to the second vector output when the second intention vector generated by the intention estimation model M1 is input when the second action information is input in succession to the intention vector. To learn.

As a result of each of the above-mentioned processes, the information providing device 10 can, for example, make the transition estimation model M2 learn the characteristics of the transition of the intention until the user performs the conversion.

Further, the information providing device 10 learns the transition estimation model M2 so as to output a vector obtained by projecting the vector generated by the intention estimation model M1 onto the hyperbolic space. Further, for example, the information providing device 10 learns the transition estimation model M2 so as to project the second intention vector from the first vector to the branching vector on the hyperbolic space. As a result of such processing, the information providing device 10 can make the transition estimation model M2 learn the characteristics of the transition of the user's intention in consideration of the branching of the user's intention. Therefore, the transition estimation model M2 It is possible to further improve the estimation accuracy by.

Further, the information providing device 10 has a second intention indicating the intention when the user performs the second action when the first intention information indicating the intention when the user performs the first action is input. The transition estimation model M2 is trained so as to output a vector similar to the case where the information is input. Therefore, the information providing device 10 can realize the transition estimation model M2 that estimates the intention of the future user.

The information providing device 10 may learn a transition estimation model M2 that is different for each user, or may train a transition estimation model M2 that is different for each state when the user takes an action. When such a process is executed, the information providing device 10 can make the transition estimation model M2 learn the characteristics of the transition of the user's intention for each user or each user's state. The estimation accuracy by the transition estimation model M2 can be further improved.

Further, the information providing device 10 may acquire behavior information indicating the behavior performed by the user on the Web, or may acquire the information input by the user as behavior information. Further, the information providing device 10 may acquire the information selected by the user as action information, or may acquire the information related to the information browsed by the user as action information. Further, the information providing device 10 may acquire information indicating a transaction target selected by the user as action information. As a result, the information providing device 10 can estimate the user's intention from the behaviors of various users, and let the transition estimation model M2 learn the characteristics of the transition of the intention.

Further, when the information providing device 10 receives the behavior information of the target user to be the estimation target of the change of intention, the information providing device 10 estimates the transition of the intention of the target user from the behavior information by using the transition estimation model M2. For example, the information providing device 10 estimates the future behavior of the target user based on the transition of the intention of the target user estimated from the received behavior information, and is used when the estimated future behavior is performed. The future intention information indicating the intention of the person is generated, and the future behavior of the target user is estimated from the future intention information by using the transition estimation model M2. As a result of such processing, the information providing device 10 can proactively provide information according to the future intentions and actions of the user.

Further, the information providing device 10 acquires a plurality of action information indicating the user's action and state information indicating the state when the user performs each action. Then, the information providing device 10 is a second action performed by the user after inputting the first action information indicating the user's first action and after the first action, and the first action is performed. The intention estimation model M1 is learned so as to output a vector similar to that when the second action information indicating the second action performed by the user is input in a state similar to that when the user is asked. As a result of such processing, the information providing device 10 can estimate the user's intention more accurately.

Although some of the embodiments of the present application have been described in detail with reference to the drawings, these are examples, and various modifications are made based on the knowledge of those skilled in the art, including the embodiments described in the disclosure column of the invention. It is possible to carry out the present invention in other modified forms.

Further, the above-mentioned "section, module, unit" can be read as "means" or "circuit". For example, the imparting unit can be read as a specific means or a specific circuit.

10 Information provision device 20 Communication section 30 Storage section 31 Action history database 32 Model database 40 Control section 41 Acquisition section 42 Generation section 43 Learning section 44 Reception section 45 Estimating section 46 Providing section 100 Terminal device 200 Search server M1 Intention estimation model M2 Transition Estimated model

Claims (24)

An acquisition unit that acquires multiple behavioral information indicating the user's behavior,
The acquisition unit uses a second model that generates intention information indicating the intention of performing the action indicated by the action information from the action information assuming that the action performed within a predetermined period is performed with a similar intention. A generator that generates intention information from the acquired behavior information,
A learning device characterized by having a learning unit that causes a first model to learn the characteristics of the intention transition indicated by the intention information generated by the generation unit. The generation unit has similar intentions when the user performs the first action and when the user performs the second action within a predetermined period after performing the first action. The learning device according to claim 1, wherein the learning device is characterized by generating the intention information. The generation unit is learned to output similar vectors when the first action information indicating the first action is input and when the second action information indicating the second action is input. The learning device according to claim 2 , wherein the intention information is generated from each action information acquired by the acquisition unit using the second model. As the second model, the generation unit includes a vector output when the first action information is input and a vector output when the second action information is input in succession to the first action information. The learning device according to claim 3 , wherein the intention information is generated by using a recursive neural network trained so as to be similar to the above. By inputting the action information into the second model, the generation unit generates the vector output by the second model as the intention information.
The learning device according to claim 3 or 4 , wherein the learning unit causes the first model to learn the characteristics of the transition of the vector generated by the second model. In the learning unit, the vector generated by the first model from the vector generated by the second model from the first behavior information and the vector generated by the second model from the second behavior information are similar to each other. The learning device according to claim 5 , wherein the learning of the first model is performed. The learning unit performs the first action on a plurality of vectors generated by the second model from the first action information indicating the plurality of first actions performed by the user before performing the second action. The learning of the first model is performed so that the vector generated by the first model when input to the first model in the same order and the vector generated by the second model from the second behavior information are similar. The learning device according to claim 5 or 6 . The learning unit has a first vector to be output when the first intention vector generated by the second model is input when the first action information is input as the first model, and the first intention vector. In succession with, when the second action information is input, the recursive neural network is similar to the second vector output when the second intention vector generated by the second model is input. The learning device according to claim 6 or 7 , wherein learning is performed. The learning device according to claim 8 , wherein the learning unit trains the first model so as to output a vector obtained by projecting a vector generated by the second model onto a hyperbolic space. The ninth aspect of the present invention is characterized in that the learning unit trains the first model so as to project the second intention vector from the first vector to a branching vector on the hyperbolic space. Learning device. The learning unit receives second intention information indicating the intention when the user performs the second action when the first intention information indicating the intention when the user performs the first action is input. The learning apparatus according to any one of claims 1 to 10, wherein the first model is trained so as to output a vector similar to the case where is input. The learning device according to any one of claims 1 to 11, wherein the learning unit performs learning of the first model, which is different for each user. The learning device according to any one of claims 1 to 12, wherein the learning unit performs learning of the first model that differs depending on the state when the user performs the action. .. The learning device according to any one of claims 1 to 13, wherein the acquisition unit acquires behavior information indicating an action performed by the user on the Web. The learning device according to any one of claims 1 to 14, wherein the acquisition unit acquires information input by the user as the action information. The learning device according to any one of claims 1 to 15 , wherein the acquisition unit acquires information selected by the user as the action information. The learning device according to any one of claims 1 to 16 , wherein the acquisition unit acquires information related to the information browsed by the user as the behavior information. The learning device according to any one of claims 1 to 17 , wherein the acquisition unit acquires information indicating a transaction target selected by the user as the action information. The reception department that receives the behavior information of the target user who is the target of estimation of the change of intention,
Claims 1 to 18 characterized by being an estimation unit that estimates the transition of the intention of the target user from the behavior information received by the reception unit using the first model learned by the learning unit. The learning device according to any one of them. An acquisition unit that acquires multiple behavioral information indicating the user's behavior,
From each action information, a generator that generates intention information indicating the user's intention when performing the action indicated by the action information, and a generation unit.
A learning unit that causes the first model to learn the characteristics of the intention transition indicated by the intention information generated by the generation unit.
The reception department that accepts the behavior information of the target user who is the target of estimation of the change of intention,
Using the first model learned by the learning unit, an estimation unit that estimates the transition of the intention of the target user from the behavior information received by the reception unit.
Have,
The estimation unit estimates the future behavior of the target user based on the transition of the intention of the target user estimated from the behavior information received by the reception unit.
The generation unit generates future intention information indicating the user's intention when the future action estimated by the estimation unit is performed.
The estimation unit is a learning device characterized in that the future behavior of the target user is estimated from the future intention information generated by the generation unit by using the first model. It is a learning method executed by the learning device.
The acquisition process to acquire multiple behavior information indicating the user's behavior,
By the acquisition step, the second model is used to generate intention information indicating the intention of performing the action indicated by the action information from the action information assuming that the action performed within the predetermined period is performed with a similar intention. A generation process that generates intention information from the acquired behavior information,
A learning method comprising a learning step of causing a first model to learn the characteristics of the intention transition indicated by the intention information generated in the generation step. Acquisition procedure to acquire multiple behavior information showing user behavior,
According to the acquisition procedure, the second model is used to generate intention information indicating the intention of performing the action indicated by the action information from the action information assuming that the action performed within the predetermined period is performed with a similar intention. The generation procedure to generate the intention information from the acquired behavior information,
A learning program for causing a computer to execute a learning procedure for causing the first model to learn the characteristics of the intention transition indicated by the intention information generated in the generation procedure. It is a learning method executed by the learning device.
The acquisition process to acquire multiple behavior information indicating the user's behavior,
From each action information, a generation process for generating intention information indicating the user's intention when performing the action indicated by the action information, and a generation process.
A learning process in which the first model learns the characteristics of the intention transition indicated by the intention information generated by the generation process.
The reception process that accepts the behavior information of the target user who is the target of estimation of the change of intention,
Using the first model learned by the learning process, an estimation process of estimating the transition of the intention of the target user from the behavior information received by the reception process.
Including
The estimation process estimates the future behavior of the target user based on the transition of the intention of the target user estimated from the behavior information received by the reception process.
The generation step generates future intention information indicating the user's intention when the future action estimated by the estimation step is performed.
In the estimation step, the future behavior of the target user is estimated from the future intention information generated by the generation step by using the first model.
A learning method characterized by including. Acquisition procedure to acquire multiple behavior information showing user behavior,
From each action information, a generation procedure for generating intention information indicating the user's intention when performing the action indicated by the action information, and a generation procedure.
A learning procedure that causes the first model to learn the characteristics of the intention transition indicated by the intention information generated by the generation procedure.
The reception procedure for accepting the behavior information of the target user who is the target of estimation of the change of intention,
Using the first model learned by the learning procedure, an estimation procedure for estimating the transition of the intention of the target user from the behavior information received by the reception procedure.
Is a learning program to make a computer execute
The estimation procedure estimates the future behavior of the target user based on the transition of the intention of the target user estimated from the behavior information received by the reception procedure.
The generation procedure generates future intention information indicating the user's intention when the future action estimated by the estimation procedure is performed.
In the estimation procedure, the future behavior of the target user is estimated from the future intention information generated by the generation procedure using the first model.
A learning program to make a computer run.

Images