JP7499397B1 - BEHAVIOR PREDICTION MODEL CREATING METHOD AND GROUP BEHAVIOR PREDICTION METHOD USING THE BEHAVIOR PREDICTION MODEL CREATING METHOD - Google Patents

Abstract

[Problem] To provide a behavior prediction model creation method capable of appropriately providing a service that reflects hidden meanings in behavior using behavior history data of a target person, and a group behavior prediction method using the behavior prediction model creation method. [Solution] In a viewing behavior prediction model creation method as an example of a behavior prediction model creation method, a viewing behavior generation unit 10 executes a pre-trained model construction step of constructing a pre-trained model that generates a feature vector of the input viewing history data by a learning process in which viewing history data that chronologically represents viewing times and broadcast content viewed as behavior history data of the target viewer is input to a learning process model as behavior history data of the target viewer, and a viewing behavior generation step of generating viewing behavior as predicted behavior of the target viewer from the pre-trained model by inputting metadata of the broadcast time and broadcast content into the pre-trained model constructed in the learned model construction step by a viewing rating processing unit 20. [Selected Figure] Figure 1

Description

The present invention relates to a method for creating a behavior prediction model for predicting the behavior of a subject based on behavior history data, and a group behavior prediction method using the behavior prediction model creation method.

As shown in Patent Document 1 below, a conventional behavior prediction model of this type is known in which a behavior prediction device includes a behavior history data storage unit that stores a user's behavior history data, a behavior model creation unit that probabilistically creates a behavior model of the user from the user's behavior history data stored in the behavior history data storage unit, and a behavior prediction unit that probabilistically predicts the user's behavior according to predetermined conditions based on the user's behavior model created by the behavior model creation unit, and when the number of data items in the behavior history data that meet the predetermined conditions does not meet a predetermined threshold, the behavior prediction unit expands the conditions to predict the user's behavior.

JP 2010-198243 A

However, such conventional behavioral prediction models merely predict the target person's behavior probabilistically from behavioral history data, and have the problem that they do not fully take into account the meaning of the behavioral history data, i.e., the hidden meaning of the target's behavior.

The same problem occurs when machine learning is performed using behavioral history data as training data instead of probabilities, and there is a problem that the hidden meaning of the subject's behavior is not reflected.

The present invention has been made in light of this background, and aims to provide a behavior prediction model creation method that can appropriately provide services that reflect hidden meanings in behavior by using behavior history data of subjects, and a group behavior prediction method that uses the behavior prediction model creation method.

A behavior prediction model creation method of a first invention is a behavior prediction model creation method executed by a computer, comprising the steps of:
a pre-trained model construction step of constructing a pre-trained model that generates a feature vector of the input behavior history data by a learning process in which behavior history data of a subject corresponding to the ID, the behavior history data representing a time and a content of the behavior in a chronological order, is input to a learning process model;
The method is characterized by comprising a predictive behavior generation step of generating predictive behavior from the pre-trained model constructed in the trained model construction step by inputting metadata regarding the event time and content.

The target person refers to a person who has behavioral history data associated with an ID (television broadcast viewing history data, location data (including GPS data), purchase data, store visit data, event participation (seminar participation) data, and web access history data).

According to the behavioral prediction model creation method of the first invention, in the pre-trained model construction step, the behavioral history data of the subject is input to a learning processing model and trained to construct a pre-trained model that generates a feature vector of the behavioral history data, making it possible to extract the subject's hobbies, preferences, habits, etc. hidden in the behavioral history data.

Then, since the event has actually been generated, in the predicted behavior generation step, metadata about the event time and content can be input to generate specific information about the behavior.

In this way, according to the behavioral prediction model creation method of the first invention, it is possible to appropriately provide a service that reflects the hidden meaning of the subject's behavior, such as the subject's hobbies, preferences, and habits, by using the subject's behavioral history data.

A group behavior prediction method according to a second aspect of the present invention is a group behavior prediction method executed by a computer, comprising the steps of:
The method for creating a behavior prediction model according to claim 1 is characterized in that the behavior of a group is predicted from the predicted behaviors of a plurality of subjects.

In this way, according to the group behavior prediction method of the second invention, it is possible to predict group behavior by using behavioral history data from multiple subjects and aggregating behavioral predictions that reflect hidden meanings in the behavior of the subjects, such as their hobbies, preferences, and habits, making it possible to provide appropriate services related to group behavior.

A third aspect of the present invention provides a group behavior prediction method according to the second aspect of the present invention, comprising:
The present invention is characterized in that it predicts the behavior of a target group when the targets are limited to the multiple subjects based on items related to the attributes of the subjects.

According to the group behavior prediction method of the third invention, it is possible to predict the behavior of each subject for each of a number of items related to the subject's attributes. Therefore, it is possible to predict the behavior of a target group by aggregating behavior predictions when the target subjects are limited by attribute items.

In this way, according to the group behavior prediction method of the third invention, it is possible to use behavioral history data from multiple subjects to predict the behavior of target subjects corresponding to attribute-related items as predetermined information regarding their behavior, and, for example, generate an event schedule suited to the reactions and feedback of participation in an event with a predetermined content, and actually provide an appropriate service related to behavior prediction that reflects the hidden meanings of the behavior of the target subjects, such as their hobbies, preferences, and habits.

A behavior prediction model creating method according to a fourth aspect of the present invention is the method according to the first aspect of the present invention,
The pre-trained model construction step includes a learning result verification step of verifying a learning process result of the generated pre-trained model,
The learning result verification step is characterized in that a portion of the behavioral history data is masked and the masked behavioral history data is predicted by the pre-trained model to verify the learning process results of the pre-trained model.

According to the behavioral prediction model creation method of the fourth invention, the learning process results can be verified by masking a portion of the behavioral history data for a pre-trained model and having the pre-trained model predict the masked behavioral history data, thereby ensuring the learning process of the pre-trained model.

In this way, according to the behavioral prediction model creation method of the fourth invention, a highly reliable learning model regarding behavior can be constructed using the subject's behavioral history data, making it possible to more appropriately provide behavioral services that reflect the hidden meanings of the subject's behavior, such as his or her hobbies, preferences, and habits.

FIG. 1 is an explanatory diagram showing an entire system according to an embodiment of the present invention. FIG. 2 is an explanatory diagram showing an example of the configuration of a viewing behavior generation unit in FIG. 1 .

An example of an embodiment of the present invention will be described with reference to FIG.

As shown in FIG. 1, as an example of the behavior prediction model creation method of this embodiment and the group behavior prediction method using the behavior prediction model creation method, a broadcast processing device that executes a viewing behavior prediction model creation method and a viewer rating prediction method using the viewing behavior prediction model creation method will be described.

The broadcast processing device is a system that predicts the viewing behavior of each target viewer using a viewing behavior generation unit 10 that corresponds to each individual target viewer, and calculates the viewing rate by aggregating (accumulating) the predictions.

Specifically, in FIG. 1, the viewing history data, which is the viewing behavior, of Mr. A, one of the target viewers, is input into the learning processing model as behavior history data, and the learning processing model is trained, so that the viewing behavior generation unit 10 constructs a pre-trained model that generates feature vectors of the viewing data (feature vectors of the viewing behavior and feature vectors of meta documents related to programs, which are the viewing contents).

Here, the viewing history of Person A that is input to the learning processing model is the actual viewing behavior and meta information of the program, which is the viewing content. By inputting these, the learning processing model constructs a pre-learning model in the viewing behavior generation unit 10 that generates a feature vector of the viewing behavior and a feature vector of the meta document from the viewing data, based on the viewing behavior pattern based on Person A's viewing behavior and the document pattern based on the meta information of the program.

According to such a pre-trained model for Mr. A, by inputting metadata on the broadcast time and broadcast content, the viewing behavior of the target viewer, Mr. A, is generated from the pre-trained model as a predicted behavior.

The same process is performed for target viewers B, C, etc. other than A, and the viewing behavior of these multiple target viewers (for example, whether or not to watch a program with a specific content in a specific time period) is accumulated in the viewership rating processing unit 20 to calculate various indicators such as viewership ratings.

The above is an overview of the broadcast processing device, and an example of the configuration of the viewing behavior generation unit 10 is shown in Figure 2.

As shown in FIG. 2, the viewing behavior generation unit 10 includes a pre-learning model creation processing unit 11, a fine-tuning processing unit 12, a viewing degree information generation unit 13, and an information output unit 14.

The pre-learned model creation processing unit 11 is a processing unit that creates a pre-learned model that generates a feature vector of the viewing history data of any target viewer from the viewing history data, and the learning process (STEP 1) in the pre-learned model creation processing unit 11 corresponds to the pre-learned model construction step of the present invention.

In this embodiment, a case where a natural language processing model such as GPT or BERT is used as the learning processing model is shown as an example. In this case, a pre-learning model is created by performing a learning process on the natural language processing model using viewing history data of multiple target viewers instead of language data. Note that, in addition to natural language processing models, various learning processing models (AI models) such as image generation AI and emotion visualization models can be used as the learning processing model.

In this learning process, the viewing history data used instead of language data is time series data of multiple unit viewing periods during a specified period (e.g., the period from xx:xx on xx month to xx month, or the period from xx:xx on xx month to xx:xx on xx month). Each unit viewing period is a specified unit time span (e.g., 5 minutes) during which television broadcasts were viewed. Each unit viewing period is associated with its representative time (start time, end time, center time, etc.) and the broadcasting station or channel of the television broadcast that was aired during that unit viewing period, and is assigned identification data for each unit viewing period.

In the learning process for creating the pre-learning model, the entire viewing history data of each target viewer is treated as text data, and the identification data for each unit viewing period is treated as words, and these are input into the natural language processing model. In this way, the pre-learning model is constructed so that when the viewing history data of any target viewer is input, it generates a feature vector that represents the characteristics of each terminal viewing period within the viewing history data.

In addition, a learning result verification step may be additionally performed on the pre-learning model constructed in this manner by the learning process to verify the results of the learning process.

More specifically, in the learning result verification step, a part of the viewing history data is masked, and the pre-learning model is made to predict the masked part of the viewing data, with the masking position being changed in sequence, and it is verified whether the learning process result of the pre-learning model has a certain accuracy rate. This makes it possible to guarantee the learning process of the pre-learning model.

Furthermore, a viewing behavior generation step is executed in which metadata of the broadcast time and broadcast content is input to the pre-trained model constructed by the learning process, and viewing behavior of the target viewer (the target viewer corresponding to the constructed pre-trained model) is generated from the pre-trained model.

Here, in the viewing behavior generation step, in addition to simply inputting metadata about the broadcast time and broadcast content into the pre-learning model to generate viewing behavior (whether or not to watch a particular program), viewing history data of the target viewer prior to the particular viewing may be input as teacher data into the pre-learning model and re-learned to predict whether the target viewer will engage in the particular viewing. This allows the pre-learning model to more accurately predict viewing behavior for a particular viewing.

In this way, the viewership data processing unit 20 can accumulate the predicted viewing behavior (whether or not to watch a particular program) of each target viewer for multiple target viewers to generate predicted viewing data such as predicted viewership ratings.

The fine-tuning processing unit 12 creates two types of models by fine-tuning the pre-trained model constructed as described above. The two types of models are a classification model created by the classification model creation unit 12a and a proposed model created by the proposed model creation unit 12b.

The classification model is configured to generate attribute information for any target viewer from the viewing history data of each of the target viewers. In this case, the attribute information for the target viewer is information on a predetermined number of types of attributes. The multiple 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 attitudes, information representing purchasing attitudes, etc. Then, the classification model is created from the pre-learning model as a separate model for each of these multiple types of attributes.

A classification model corresponding to each attribute is created from a pre-learning model so that when viewing history data of any target viewer is input, the target viewer can identify the content of the type of attribute corresponding to the classification model. For example, a classification model corresponding to attributes such as gender and age is created so that it is possible to identify whether the gender is male or female and what age group the target viewer is from the input viewing history data. A classification model corresponding to each type of attribute is created by applying machine learning using viewing history data of target viewers whose corresponding attributes are known as training data.

The proposed model is created from the pre-learning model so as to generate, for each of a plurality of target items related to the attributes of the target viewers, a schedule of a broadcast of a predetermined content suitable for viewing the broadcast of the predetermined content from the viewing history data of each target viewer. In this case, the target items are the contents of the attributes of the target viewers to whom it is desired that the broadcast of the predetermined content be viewed. For example, the contents adopted as target items are the contents of the attributes related to gender and age being female between xx years old and xx years old, the contents of the attributes related to income being xx yen or more per year, and the contents of the attributes related to occupation being an office worker. A plurality of such target items are set in advance.

The proposed model corresponding to each target item is created from the pre-learning model so that, when the viewing history data of the target viewer corresponding to that target item is input, a schedule of broadcasts of the specified content suitable for viewing the broadcasts of the specified content by the target viewer is generated.

For example, a proposed model corresponding to a certain target item is created so that when a target viewer corresponding to the target item is to watch a commercial broadcast for a certain electrical product, an appropriate broadcast date and time for the commercial broadcast (a broadcast date and time when the target viewer is likely to watch the commercial broadcast, for example, between 19:00 and 19:20 every Tuesday) can be identified. The proposed model corresponding to each target item is created by applying machine learning using the viewing history data of the target viewer corresponding to the target item as training data.

The viewing degree information generating unit 13 uses the viewing history data (viewing history data including a broadcast of a specified content) of each of the multiple target viewers and the classification model to generate viewing degree information that indicates the relationship between the viewing degree of the broadcast of a specified content and the attributes of the target viewers. More specifically, the viewing degree information is information that indicates which attribute or target item of target viewers most frequently viewed the broadcast of the specified content (in other words, which attribute or target item has the highest viewing degree (number of viewers) of the broadcast of the specified content).

This viewing degree information generating unit 13 can obtain viewing degree information such as, for example, that commercial broadcasts for electrical products are most highly viewed by target viewers with the following attributes: males of gender and age between XX and XX years old, annual income of XX yen or more, and occupation of company employee.

The information output unit 14 outputs the information obtained by the proposed model and the viewing degree information generation unit 13 via a display or printer (not shown). Specifically, when a user inputs the broadcast content, such as a scheduled commercial broadcast, and target items, the information output unit 14 outputs to the user a broadcast schedule appropriate for allowing the target viewers who match the target items to view the broadcast content, based on the proposed model. In addition, when a user performs an operation to request viewing degree information for the scheduled broadcast content, the information output unit 14 outputs the viewing degree information obtained by the viewing degree information generation unit 13 to the user.

In addition, in conjunction with the processing by the proposed model creation unit 12b of the viewing behavior generation unit 10, it is preferable that the viewing rate processing unit 20 calculates a target viewing rate from the viewing behavior predicted for target viewers who match specific target items.

More specifically, the viewership rating processing unit 20 selects target viewers who match the target items from among multiple target viewers, and predicts the target viewership rating by accumulating the predicted viewing behavior (whether or not to watch a particular program) of only the selected target viewers.

This makes it possible to predict the viewing behavior of each target viewer for each of multiple items related to the attributes of the target viewer. Therefore, it is possible to predict the target audience rating by aggregating the viewing behavior predictions when the target audience is limited by the attribute items.

In other words, as the specified information regarding broadcast viewing, it is possible to predict the target viewership rating, which is a prediction of the viewership rating of the target audience corresponding to the attribute-related items, and in combination with the proposed model of the proposed model creation unit 12b, it is possible to generate a broadcast schedule that takes into account the viewership rating and is suitable for viewing broadcasts of specified content, and it is possible to actually provide services related to broadcast viewing appropriately.

The broadcast processing device of this embodiment is configured as described above, and can appropriately provide services related to the viewing of television broadcasts using television broadcast viewing history data from multiple viewers without requiring a large amount of work from the user. In addition, the classification model and proposal model required for the service can be easily constructed using a publicly known natural language processing model.

In this embodiment, the viewing history data may be obtained via broadcasting (including television, radio, and BS and CS) as well as communication (including IP broadcasting and connected TV such as Netflix (registered trademark)).

In addition, in this embodiment, commercial broadcasts (video commercials) on television broadcasts have been described as an example, but the present invention is not limited to this and may include video works such as programs, movies, animation, CG, plays, games, and internet distribution, as well as radio commercials and radio programs that do not involve video, photographs, novels, manga, books, music (vocals, performances), NFTs, works of art (fine arts, architecture, design, etc.), and artistic expressions (performances).

In addition, in the above embodiment, viewing history data was used as an example of behavioral history data to construct the classification model and the proposed model, but a predetermined type of history data regarding the subject's behavior may be used instead of viewing history data.

As the predetermined type of history data (behavioral history data), various types of history data may be used as long as they are linked to the subject's ID. For example, the history data may be location data, purchase data (such as camera purchase history data), store visit data, event participation (seminar participation) data, and web access history data.

Even in the case of such behavioral history data, the pre-trained model generates predicted behavior (whether or not to participate in the event) for each of the subjects A, B, and C by inputting metadata about the event's time and content. Then, by aggregating (accumulating) the predicted behavior of each of these subjects, it is possible to predict the behavior of the group as a whole.

In the case of these various types of behavioral history data, classification models are also created from the pre-learning model as separate models for each of the multiple types of attributes. Furthermore, the proposed model is created from the pre-learning model so as to generate a schedule of predetermined behavior appropriate for causing a predetermined behavior from the various behavioral history data of each subject for each of multiple target items related to the subject's attributes. For example, a schedule is generated of appropriate movement (visit) timing in the case of location data, appropriate purchase timing in the case of purchase data, appropriate store visit timing in the case of store visit data, appropriate participation timing in the case of event participation (seminar participation) data, and appropriate web access timing in the case of web access history data.

In this way, it will be possible to provide more appropriate behavioral services that reflect the hidden meanings of a person's behavior, such as their hobbies, preferences, and habits.

In the embodiment described above, the pre-learning model was created using natural language processing, and the classification model 12a and the proposed model 12b were created from the pre-learning model, but the classification model 12a and the proposed model 12b may be created without using a natural language processing model.

10...viewing behavior generation unit, 11...pre-learning model creation processing unit, 12...fine-tuning processing unit, 12a...classification model creation unit, 12b...proposed model creation unit, 13...viewing degree information generation unit, 14...information output unit.

Claims (4)

A computer-implemented method for creating a behavioral prediction model, comprising:
a pre-trained model construction step of constructing a pre-trained model that generates a feature vector of the input behavior history data by a learning process in which behavior history data of a subject corresponding to the ID, the behavior history data representing a time and a content of the behavior in a chronological order, is input to a learning process model;
A method for creating a behavior prediction model, comprising: a predictive behavior generation step of generating predictive behavior from the pre-trained model constructed in the trained model construction step by inputting metadata regarding the time and content of an event. 1. A computer-implemented method for predicting collective behavior, comprising:
A group behavior prediction method, comprising: predicting group behavior from predicted behaviors of a plurality of subjects by using the behavior prediction model creation method according to claim 1. The method for predicting group behavior according to claim 2, characterized in that the behavior of a target group is predicted when the targets of the multiple subjects are limited based on items related to the attributes of the subjects. 2. The behavior prediction model creation method according to claim 1,
The pre-trained model construction step includes a learning result verification step of verifying a learning process result of the generated pre-trained model,
A method for creating a behavioral prediction model, characterized in that in the learning result verification step, a portion of the behavioral history data is masked and the masked behavioral history data is predicted by the pre-trained model to verify the learning process results of the pre-trained model.

Images