Behavior model creation method

Abstract

To provide a behavior model creation method capable of properly providing a service with implicit meaning in a behavior reflected therein, using action history data of a target person.SOLUTION: A behavior model creation method includes: a first step of constructing a first model which generates feature vectors of input action history data of a plurality of target persons, through learning processing that inputs the action history data indicating time and content of actions chronologically, in place of language data, to a natural language processing model; and a second step of constructing a second model configured to generate predetermined information related to actions, from the first model generated in the first step.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a method for creating a behavior model for predicting the behavior of a target person based on behavior history data.

Background Art

[0002] Conventionally, as a behavior prediction model of this type, as shown in Patent Document 1 below, a behavior prediction device includes a behavior history data storage unit that stores the behavior history data of a user, and a behavior model creation unit that probabilistically creates the behavior model of the user from the behavior history data of the user stored in the behavior history data storage unit. A behavior prediction unit probabilistically predicts the behavior of the user according to a predetermined condition based on the behavior model of the user created by the behavior model creation unit. When the number of data of behavior history data that conforms to a predetermined condition is less than a predetermined threshold, the behavior prediction unit is known to expand the condition and predict the behavior of the user.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, such a conventional behavior prediction model merely probabilistically predicts the behavior of the target person from the behavior history data, and there is a problem that the content meant by the behavior history data, that is, the hidden meaning in the target's behavior is not sufficiently considered.

[0005] This is the same when the behavior history data is used as teaching data for machine learning instead of probability, and there is a problem that the hidden meaning in the target person's behavior is not reflected.

[0006] The present invention has been made in view of such a background, and an object thereof is to provide a method for creating a behavior model that can appropriately provide a service that reflects the meaning hidden in behaviors by using the behavior history data of a target person. **Means for Solving the Problems**

[0007] The behavior model creation method of the first invention includes a first step of constructing a first model that generates a feature vector of the input behavior history data by a learning process of inputting behavior history data of a plurality of target persons, which represents the behavior time and the content of the behavior in chronological order, into a natural language processing model instead of language data. It is characterized by including a second step of constructing a second model that can generate predetermined information related to behaviors from the first model generated in the first step.

[0008] Here, the target person means a person who has behavior history data associated with an ID (in addition to the viewing history data of TV broadcasts, in addition to location data (including GPS data), purchase data, store visit data, event participation (seminar participation) data, and WEB access history data).

[0009] According to the behavior model creation method of the first invention, in the first step, by inputting the behavior history data of the target person into a natural language processing model as an AI model and training it, a first model is constructed that is generated as a feature vector of the behavior history data, so that the hobbies, preferences, habits, etc. of the target person hidden in the behavior history data can be extracted.

[0010] Then, from the actually generated first model, in the second step, predetermined information related to behaviors can be generated.

[0011] Thus, according to the behavior model creation method of the first invention, it becomes possible to appropriately provide a service that reflects the meaning hidden in behaviors such as the hobbies, preferences, and habits of the target person by using the behavior history data of the target person.

[0012] The method for creating an action model according to the second invention is, in the first invention, The second model is a classification model that generates attribute information for each target person from the action history data of each of a plurality of target persons.

[0013] According to the method for creating an action model of the second invention, in the second step, by creating a classification model as the second model, it is possible to generate attribute information for each target person from the action history data of each of a plurality of target persons.

[0014] Thus, according to the method for creating an action model of the second invention, by using the action history data of the target person, it is possible to generate attribute information for each target person as predetermined information regarding actions, and actually and appropriately provide a service that reflects the hidden meaning in actions such as the hobbies, preferences, and habits of the target person.

[0015] The method for creating an action model according to the third invention is, in the first invention, The second model is a model proposal model that generates a schedule of actions of the predetermined content suitable for causing a target person to perform an action of the predetermined content from the action history data of each target person for each of a plurality of items related to the attributes of the target person.

[0016] According to the method for creating an action model of the third invention, in the second step, by creating a proposal model as the second model, for each of a plurality of items related to the attributes of the target person, it is possible to generate a schedule of actions of the predetermined content suitable for causing the target person to perform an action of the predetermined content from the action history data of each target person.

[0017] Thus, according to the method for creating an action model of the third invention, by using the action history data of a plurality of target persons, it is possible to generate a schedule of actions of the predetermined content suitable for causing an action of the predetermined content as predetermined information regarding actions, generate attribute information for each target person as predetermined information regarding actions, and actually and appropriately provide a service that reflects the hidden meaning in actions such as the hobbies, preferences, and habits of the target person.

[0018] The action model creation method of the fourth invention, in the first invention, The first step has a learning result verification step of verifying the learning processing result of the generated first model, In the learning result verification step, a part of the action history data is masked, and the masked action data is used to predict by the first model, thereby verifying the learning processing result of the first model.

[0019] According to the action model creation method of the fourth invention, for the first model that is the basis for constructing the second model, a part of the action history data is masked, and the masked action data is used to predict by the first model, so that the learning processing result can be verified, and the learning processing of the first model can be ensured.

[0020] In this way, according to the action model creation method of the fourth invention, by using the action history data of a plurality of target persons, a highly reliable learning model related to actions can be constructed, and a service that sufficiently reflects the hidden meanings in actions such as the hobbies, preferences, and habits of the target persons can be actually and appropriately provided.

[0021] The action model creation method of the fifth invention, in the first invention, The first step has an action prediction step of predicting whether the target person will perform a specific action by the generated first model, In the action prediction step, the action history data before the specific action of the target person is used as teacher data, input into the first model and re-learned, so as to predict whether the target person will perform the specific action.

[0022] According to the action model creation method of the fifth invention, the first model is re-learned with the action history data before the specific action to be predicted as teacher data so that the first model can predict the action, so that the specific action can be predicted.

[0023] Thus, according to the action model creation method of the fifth invention, it is possible to make predictions regarding actions using the action history data of a plurality of subjects, and to actually and appropriately provide a service that sufficiently reflects the hidden meaning in actions such as the hobbies, preferences, and habits of the subjects.

Brief Description of the Drawings

[0024]

Figure 1

Embodiments for Carrying Out the Invention

[0025] An embodiment of the present invention will be described with reference to FIG. 1. Referring to FIG. 1, a broadcast processing apparatus in which a broadcast model creation method as an example of the action model creation method of the present embodiment is executed is an apparatus composed of one or more computers or one or more electronic circuit units, or a combination thereof, and as functions realized by its hardware configuration and the installed program (software configuration), it includes a basic model creation processing unit 1, a fine-tuning processing unit 2, a viewing degree information generation unit 3, and an information output unit 4.

[0026] The basic model creation processing unit 1 is a processing unit that creates a basic model (corresponding to the first model of the present invention) that generates a feature vector of the viewing history data from the viewing history data of an arbitrary target viewer, and the learning process (STEP1) in the basic model creation processing unit 1 corresponds to the first step of the present invention.

[0027] In the present embodiment, as the basic model, for example, natural language processing models such as GPT and BERT are used. In this case, the basic model is created by performing a learning process on the natural language processing model using the viewing history data of a plurality of target viewers instead of language data.

[0028] In this learning process, the viewing history data used instead of language data is time - series data of a plurality of unit viewing periods within a predetermined period (for example, a period of 〇 months starting from 〇:〇〇 on 〇 / 〇, or a period from 〇:〇〇 on 〇 / 〇 to 〇:〇〇 on 〇 / 〇). Each unit viewing period is a period of a predetermined unit time width (for example, 5 minutes) during which TV broadcasts are viewed. And for each unit viewing period, its representative time (start time, end time, central time, etc.) is associated with the broadcasting station or channel of the TV broadcast during that unit viewing period, and identification data for each unit viewing period is attached.

[0029] Then, in the learning process of creating the basic model, the entire viewing history data of each target viewer is regarded as text data, and the identification data for each unit viewing period is regarded as a word and input into the natural language processing model. As a result, the basic model is constructed to generate a feature vector representing the characteristics of each terminal viewing period in the viewing history data when the viewing history data of an arbitrary target viewer is input.

[0030] An additional learning result verification step for verifying the learning processing result may be performed on the basic model constructed by the learning process in this way.

[0031] More specifically, in the learning result verification step, a part of the viewing history data is masked, and the process of making the basic model predict a part of the masked viewing data is performed by changing the mask positions in order to verify whether the learning processing result of the basic model has a certain correct answer rate. In this way, the learning processing of the basic model can be ensured.

[0032] Furthermore, an additional viewing prediction step for predicting whether a target viewer will perform specific viewing may be performed on the basic model constructed by the learning process.

[0033] More specifically, in the viewing prediction step, by using the viewing history data before the specific viewing of the target viewer as teacher data and inputting it into the basic model for retraining, it is determined whether the target viewer will perform the specific viewing. In this way, the basic model can predict the specific viewing of the viewer, and by accumulating this for a plurality of viewers, predicted viewing data such as a predicted viewing rate can be generated.

[0034] The fine-tuning processing unit 2 creates two types of models (both corresponding to the second model of the present invention) by fine-tuning the basic model constructed as described above. The two types of models are a classification model created by the classification model creation unit 2a and a proposal model created by the proposal model creation unit 2b. The creation process of the classification model (STEP2a) by the classification model creation unit 2a and the creation process of the proposal model (STEP2b) by the proposal model creation unit 2b each correspond to the second step of the present invention.

[0035] The classification model is configured to generate attribute information of an arbitrary target viewer from the viewing history data of each target viewer. In this case, the attribute information of the target viewer is information regarding a plurality of predetermined types of attributes. The plurality of types of attributes are, for example, information representing gender and age, information representing income, information representing occupation, information representing the composition of cohabiting family members, information representing viewing awareness, information representing purchasing awareness, etc. And the classification model is created from the basic model as separate models for each of these plurality of types of attributes.

[0036] The classification model corresponding to each attribute is created from the basic model so that when the viewing history data of an arbitrary target viewer is input, the target viewer can identify the content of the type of attribute corresponding to the classification model. For example, the classification model corresponding to the attribute of gender and age is created so that it can identify whether the gender is male or female and which age group from the input viewing history data. The classification models corresponding to various types of attributes are created by applying machine learning using the viewing history data of the target viewer for whom the corresponding attribute has been identified as teacher data.

[0037] The proposed model is created from the basic model so as to generate a schedule for a broadcast of predetermined content suitable for causing each target audience to view the broadcast of the predetermined content from the viewing history data of each target audience for each of a plurality of target items regarding the attributes of the target audience. In this case, the target item is the content of the attributes of the target audience who wants to view the broadcast of the predetermined content. For example, the content of the attributes regarding gender and age is women aged from 〇 to 〇 years old, and the content of the attributes regarding income is 〇 yen or more per year, and the content of the attributes regarding occupation is a company employee. Such content is adopted as a target item. A plurality of such target items are set in advance.

[0038] The proposed model corresponding to each target item is created from the basic model so as to generate a schedule for a broadcast of predetermined content suitable for causing the target audience corresponding to that target item to view the broadcast of the predetermined content when the viewing history data of the target audience corresponding to that target item is input.

[0039] For example, when the proposed model corresponding to a certain target item tries to cause the target audience corresponding to that target item to view a commercial broadcast of a certain electrical product, it is created so as to be able to specify an appropriate broadcast date and time for the commercial broadcast (a broadcast date and time when the target audience has a high possibility of viewing the commercial broadcast. For example, between 19:00 and 19:20 every Tuesday, etc.). The proposed model corresponding to each target item is created by applying machine learning using the viewing history data of the target audience corresponding to that target item as teacher data.

[0040] The viewing degree information generation unit 3 uses the viewing history data of each of the plurality of target viewers (viewing history data including broadcasts of a predetermined content) and the classification model to generate viewing degree information representing the relationship between the viewing degree of the broadcast of the predetermined content and the attributes of the target viewers. More specifically, the viewing degree information indicates which target viewers with what attributes or target items watched the broadcast of the predetermined content the most (in other words, which attribute or target item has the highest viewing degree (number of viewers) of the broadcast of the predetermined content).

[0041] For example, the viewing degree information generation unit 3 obtains viewing degree information such that the commercial broadcast of electrical products has the highest viewing degree among target viewers with the attributes of male aged XX to XX years old, income of more than XX yen per year, and occupation of company employees.

[0042] The information output unit 4 outputs the information obtained by the proposal model and the viewing degree information generation unit 3 via a display (not shown) or a printer. Specifically, when the user inputs the broadcast content such as a commercial broadcast scheduled for broadcast and the target item, the information output unit 4 outputs to the user an appropriate broadcast schedule for causing the target viewers matching the broadcast content to the target item to watch the broadcast. Also, when the user performs an operation of requesting viewing degree information for the broadcast content scheduled for broadcast, the information output unit 4 outputs the viewing degree information obtained by the viewing degree information generation unit 3 to the user.

[0043] Since the broadcast processing device of the present embodiment is configured as described above, by using the viewing history data of television broadcasts by a plurality of viewers, it is possible to appropriately provide services related to television broadcasts without requiring a great deal of man-hours for the user. In addition, the classification model and the proposal model required for the service can be easily constructed using known natural language processing models.

[0044] In addition, in this embodiment, the viewing history data may be transmitted not only through broadcasts (including television, radio, BS, and CS), but also through communication (including IP broadcasts and connected TVs such as Netflix (registered trademark)).

[0045] In addition, in this embodiment, commercial broadcasts (video commercials) are used as an example for explanation, but the present invention is not limited thereto. In addition to video works such as programs, movies, anime, CG, plays, games, and Internet distribution, radio commercials or radio programs without video, photos, novels, comics, books, music (vocal music, performances), NFTs, art works (fine arts, architecture, design, etc.), art expressions (performances), etc. may also be used.

[0046] In the above embodiment, viewing history data, which is an example of behavior history data, is used to construct the classification model and the proposal model. However, instead of the viewing history data, predetermined types of history data related to the behavior of the target person may be used.

[0047] As the predetermined types of history data (behavior history data), various types of history data can be adopted as long as they are data linked to the ID of the target person. For example, as the history data, in addition to location data, purchase data (such as camera purchase history data), store visit data, event participation (seminar participation) data, and web access history data may be used.

[0048] For each of these various action history data, the classification model is created from the basic model as a separate model for each of multiple types of attributes. Also, the proposed model is created from the basic model so as to generate a schedule of actions of the predetermined content suitable for causing the actions of the predetermined content from the various action history data of each target person for each of the plurality of target items regarding the attributes of the target person. For example, in the case of position data, an appropriate movement (visit) timing, in the case of purchase data, an appropriate purchase timing, in the case of store visit data, an appropriate store visit timing, in the case of event participation (seminar participation) data, an appropriate event participation timing, and in the case of web access history data, a schedule of appropriate web access timing is generated.

[0049] In addition, in the embodiment described above, the basic model is created using natural language processing, and the classification model and the proposed model are created from the basic model. However, the classification model and the proposed model may be created without using the natural language processing model.

Description of Signs

[0050] 1... Basic model creation processing unit, 2... Fine-tuning processing unit, 2a... Classification model creation unit, 2b... Proposed model creation unit, 3... Audience degree information generation unit, 4... Information output unit.

Claims

1. A first step of constructing a first model that generates a feature vector of the input behavior history data by a learning process of inputting, instead of language data, behavior history data of a plurality of subjects that represents behavior time and behavior content in a time series into a natural language processing model; A second step of constructing a second model that can generate predetermined information regarding behavior from the first model generated in the first step, wherein the method for creating a behavior model is characterized by comprising these steps.

2. In the method for creating a behavior model according to Claim 1, the second model is a classification model that generates attribute information for each subject from the behavior history data of each of the plurality of subjects, wherein the method for creating a behavior model is characterized by this.

3. In the method for creating a behavior model according to Claim 1, the second model is a model proposal model that generates a schedule of the behavior of the predetermined content suitable for causing the behavior of the predetermined content for each of a plurality of items regarding the attributes of the subject from the behavior history data of each subject, wherein the method for creating a behavior model is characterized by this.

4. In the method for creating a behavior model according to Claim 1, the first step has a learning result verification step of verifying the learning process result of the generated first model, and in the learning result verification step, the learning process result of the first model is verified by masking a part of the behavior history data and causing the masked behavior data to be predicted by the first model, wherein the method for creating a behavior model is characterized by this.

5. In the method for creating a behavior model according to Claim 1, the first step has a behavior prediction step of predicting whether the subject will perform a specific behavior by the generated first model, and in the behavior prediction step, by using the behavior history data before the specific behavior of the subject as teacher data and inputting it into the first model for re-learning, it is predicted whether the subject will perform the specific behavior, wherein the method for creating a behavior model is characterized by this.

Images