JP6989434B2 - Programs, devices and methods for estimating potential target users who will reach the predicted target behavior using the access history of the URL. - Google Patents

Description

The present invention relates to a technique for predicting a website browsing behavior by a user.

The user can freely access various websites using a smartphone or a mobile terminal.
On the other hand, as one of the marketing analysis, the URL (Uniform Resource Locator) access history of the website for each user is collected, and the interest and tendency of the user are analyzed. In particular, it is also possible to predict browsing behavior and deliver appropriate targeting advertisements for each user.

For example, there is a technique of guiding a user by learning by a hidden Markov model using a browsing history of a website by a user as input data (see, for example, Patent Document 1). According to this technology, multiple internal states of the learning model are calculated from an ordered set in which the page functions of the browsed page are arranged in chronological order for each user, and the output probability of the page functions that can be browsed in each internal state is determined. The transition probability between internal states is predicted as the browsing behavior of the user.

There is also a technique for estimating a user who will cause a predicted target behavior by inputting a specific behavior and an explanatory behavior by an advertiser (see, for example, Patent Document 2). According to this technology, the action history of a user who is a candidate for content distribution is acquired, and the user who is predicted to take a specific action is extracted based on the action history specified by the content provider.

These conventional techniques build a learning model that calculates the probability of causing a predicted target behavior for the estimated target user by using the behavior logs of a large number of users who have caused the predicted target behavior in the past. A group of users whose probability is equal to or higher than a predetermined threshold is estimated as a target user for advertisement distribution.

Japanese Unexamined Patent Publication No. 2014-160345 Japanese Unexamined Patent Publication No. 2016-0388222

The above-mentioned conventional technique uses a time-series action history (access history) of each user. On the other hand, it does not take into account various related behaviors leading up to the predicted target behavior.
The inventors of the present application wondered if the potential target user should be estimated based on the related behavior that would lead to the predicted target behavior.

Therefore, an object of the present invention is to provide a program, a device, and a method for estimating a potential target user who will reach a predicted target behavior by using the access history of the URL.

According to the present invention, it is a program for operating a computer mounted on a device for estimating a potential target user who will lead to a predicted target behavior.
An access history storage means for accumulating accessed URLs for each user,
With the sensor URL registration means that registered multiple sensor URLs related to the predicted behavior
To make it work
As a learning stage
If the predicted behavior is an access to the specified target URL, the classification condition is whether or not the target URL is accessed (positive example) or not (negative example).
A user group classification means for classifying into a target user group as a positive example and a non-target user group as a negative example with respect to the classification condition by using the access history storage means.
A machine learning engine that learns by inputting the number of accesses of each sensor URL associated with each target user as a positive example and the number of accesses of each sensor URL associated with each non-target user as a negative example.
To make it work
As an estimation stage
The machine learning engine inputs the number of accesses of each sensor URL for the estimated target user, outputs a positive example similarity score, and outputs it.
As a latent target user estimation means that estimates an estimated target user whose similarity score is equal to or higher than a predetermined threshold value as a latent target user.
It is characterized by making the computer function.
According to other embodiments in the program of the invention
For the estimation target user, the learning stage and the estimation stage are executed for each of the plurality of target URLs, and the regular similarity score is estimated for each target URL.
Advertising information distribution means that distributes advertising information based on each target URL to the terminal of the estimation target user for each target URL.
For each target URL, an advertisement reaction detection means that detects whether or not there is a reaction operation of the estimated target user with respect to the advertisement information,
As a normal acceptance score calculation means for calculating the regression coefficient between the presence / absence of an advertising response and the presence / absence of an advertising response as a regular acceptance score based on the regular similarity score and the presence / absence of an advertising response for each target URL.
And make it work further
The latent target user estimation means estimates the potential target user by using the positive similarity score calculated by weighting the positive acceptance score.
It is also preferable to make the computer function as such.
According to the present invention, it is a program for operating a computer mounted on a device for estimating a potential target user who will lead to a predicted target behavior.
An access history storage means for accumulating accessed URLs for each user,
With the sensor URL registration means that registered multiple sensor URLs related to the predicted behavior
To make it work
As a learning stage
If the predicted behavior is access to the URL searched by the specified search key, the classification condition is whether or not the URL searched by the search key is accessed (positive example) or not (negative example). ,
If the predicted behavior is a specified behavior in the behavior history recorded for each user, the classification condition is whether or not the behavior was performed (positive example) or not (negative example).
A user group classification means for classifying into a target user group as a positive example and a non-target user group as a negative example with respect to the classification condition by using the access history storage means.
A machine learning engine that learns by inputting the number of accesses of each sensor URL associated with each target user as a positive example and the number of accesses of each sensor URL associated with each non-target user as a negative example.
To make it work
As an estimation stage
The machine learning engine inputs the number of accesses of each sensor URL for the estimated target user, outputs a positive example similarity score, and outputs it.
As a latent target user estimation means that estimates an estimated target user whose similarity score is equal to or higher than a predetermined threshold value as a latent target user.
It is characterized by making the computer function.

According to other embodiments in the program of the invention
For each URL, a computer is used as a sensor URL determining means for determining the URL in which the first probability of access by the target user group and the second probability of access by the non-target user group satisfy a predetermined condition as the sensor URL. It is also preferable to make it work.

According to other embodiments in the program of the invention
It is also preferable that the sensor URL determining means causes the computer to function as a sensor URL to determine a URL in which the odds ratio between the first probability and the second probability is equal to or higher than a predetermined threshold value.

According to the present invention, it is an estimation device that estimates a potential target user who will reach a predicted target behavior.
An access history storage means for accumulating accessed URLs for each user,
It has a sensor URL registration means that registers a plurality of sensor URLs related to the predicted behavior.
As a learning stage
If the predicted behavior is an access to the specified target URL, the classification condition is whether or not the target URL is accessed (positive example) or not (negative example).
When the predicted behavior is access to the URL searched by the specified search key, the classification condition is whether or not the URL searched by the search key is accessed (positive example) or not (negative example). ,
If the predicted behavior is a specified behavior in the behavior history recorded for each user, the classification condition is whether or not the behavior was performed (positive example) or not (negative example).
A user group classification means for classifying into a target user group as a positive example and a non-target user group as a negative example with respect to the classification condition by using the access history storage means.
A machine learning engine that learns by inputting the number of accesses of each sensor URL associated with each target user as a positive example and the number of accesses of each sensor URL associated with each non-target user as a negative example. Have,
As an estimation stage
The machine learning engine inputs the number of accesses of each sensor URL for the estimated target user, outputs a positive example similarity score, and outputs it.
It is characterized by having a latent target user estimation means for estimating an estimation target user whose regular similarity score is equal to or higher than a predetermined threshold value as a latent target user.

According to the present invention, it is an estimation method of a device for estimating a potential target user who will reach a predicted target behavior.
The device is
An access history storage unit that stores accessed URLs for each user,
A sensor URL registration unit that registers multiple sensor URLs related to the predicted behavior, and
Has a machine learning engine and
The device, as a learning stage,
If the predicted behavior is an access to the specified target URL, the classification condition is whether or not the target URL is accessed (positive example) or not (negative example).
When the predicted behavior is access to the URL searched by the specified search key, the classification condition is whether or not the URL searched by the search key is accessed (positive example) or not (negative example). ,
If the predicted behavior is a specified behavior in the behavior history recorded for each user, the classification condition is whether or not the behavior was performed (positive example) or not (negative example).
The eleventh step of classifying into a target user group that is a positive example and a non-target user group that is a negative example for the classification condition using the access history storage unit.
Using the machine learning engine, input the number of access of each sensor URL associated with each target user as a positive example and the number of access of each sensor URL associated with each non-target user as a negative example. Perform the twelfth step of learning and
The device, as an estimation stage,
The 21st step of inputting the number of accesses of each sensor URL for the estimated target user using the machine learning engine and outputting the correct example similarity score, and
It is characterized in that the 22nd step of estimating an estimation target user whose regular similarity score is equal to or higher than a predetermined threshold value as a potential target user is executed.

According to the program, device and method of the present invention, the access history of the URL can be used to estimate the potential target user who will reach the predicted target behavior.

It is explanatory drawing which shows the relationship between the predicted target behavior and the related behavior. It is a functional block diagram of the estimation apparatus in this invention. It is explanatory drawing which shows the input / output data of the machine learning engine of FIG. It is a functional block diagram of the estimation device which has a receptive reaction learning stage in this invention. It is a functional block diagram of the estimation device which estimates a potential target user based on a positive acceptance score.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

FIG. 1 is an explanatory diagram showing the relationship between the predicted target behavior and the related behavior.

As a marketing analysis, the predicted target behavior is expressed as an access to the "target URL", and the related behavior is expressed as an access to the "sensor URL". Here, the sensor URL is registered in advance so as to access the target URL. For example, when the target URL is "the homepage of the automobile a manufactured by company A", the related URL is like "the homepage of the automobile comparison site".

According to FIG. 1, a sensor URL related to the target URL and a URL other than the target URL are shown centering on the target URL.
Further, each user is classified into a target user group that has accessed the target URL and a non-target user group that has not accessed the target URL as an access history.

The group of users who have accessed the target URL is accessing various related URLs in order to reach the target URL. At this time, it can be assumed that the user group accessing the related URL will access the target URL with a high probability. That is, if the user group accessing the target URL can learn the tendency to access the related URL, the probability that the estimation target user will access the target URL can be estimated from the tendency to access the related URL.
According to the present invention, in order to estimate a potential target user who has a high probability of accessing a target URL, an attempt is made to dare to estimate a user who has a high probability of accessing the related URL.

FIG. 2 is a functional configuration diagram of the estimation device in the present invention.

The estimation device 1 of the present invention uses the access history of the URL to estimate the potential target user based on the related behavior that will lead to the predicted target behavior.
According to FIG. 2, the estimation device 1 has an access history storage unit 101 and a sensor URL registration unit 102.

[Access history storage unit 101]
The access history storage unit 101 stores the accessed URLs in association with each user (ID).
[User ID] <-> [URL]
001 http://www.aaa.com
002 http://www.bbb.com
003 http://www.aaa.com
002 http://www.ccc.com
001 http://www.aaa.com
・ ・ ・ ・ ・ ・ ・ ・ ・

[Sensor URL registration unit 102]
The sensor URL registration unit 102 registers a plurality of sensor URLs related to the predicted behavior.
The sensor URL is a URL that publishes a page related to the predicted behavior (for example, access to the target URL), and may be registered in advance by the operator. Specifically, it may be determined by the sensor URL determination unit 14 described later.

The target URL and the sensor URL are not limited to the entire URL itself, and may be only a domain name or a category name based on clustering.

The estimation device 1 has a user group classification unit 11, a machine learning engine 12, and a sensor URL determination unit 14 as a <learning stage>. Further, as the <estimation stage>, it has a machine learning engine 12 and a latent target user estimation unit 13.
These functional components are realized by executing a program that makes a computer mounted on the device function. Further, the processing flow of these functions can be understood as an estimation method in the apparatus.

<Learning stage>
[User group classification unit 11]
The user group classification unit 11 uses the access history storage unit 101 to classify the target user group as a positive example and the non-target user group as a negative example with respect to the classification condition.

Here, the "classification condition" is a condition for classifying into positive / negative cases, and can be variously defined as follows.
(Condition 1) Whether or not the specified target URL has been accessed For example, whether or not the specified target URL has been accessed (normal example) or not (negative example).
(Condition 2) Whether or not the URL searched by the specified search key was accessed as the target URL For example, the URL searched by the specified search key "Company A""Homepage of car a manufactured by Company A" is targeted. Whether or not it was accessed as a URL (positive example) (negative example)
(Condition 3) The action history is recorded for each user, and whether or not the specified action is performed For example, the action history of "participating in the car presentation" is recorded in the action history for each user (condition 3). Positive example) or not (negative example)

[Sensor URL determination unit 14]
The sensor URL determination unit 14 determines, for each URL, the URL in which the first probability of access by the target user group and the second probability of access by the non-target user group satisfy a predetermined condition as the sensor URL. ..

Specifically, the sensor URL determination unit 14 determines as the sensor URL a URL in which the odds ratio between the first probability and the second probability is equal to or higher than a predetermined threshold value.
The "Odds ratio" is a statistical measure that compares the likelihood of an event between two groups. It is expressed as follows for the probability p that a certain event occurs.
Odds ＝ p / (1-p), Odds ≧ 0
For a certain event, the odds ratio between the event occurrence probability p in the first group and the event occurrence probability q in the second group is expressed as follows.
{p / (1-p)} / {q / (1-q)}
Odds ratio = 1 means that the probability of event occurrence is the same in both groups, and greater than (or smaller) than 1 means that the event is more likely to occur in the first group (or second group). Means that.

For example, for the access history of 100 people, it is calculated as follows, for example.
Number of people who accessed the sensor URL by the target user group = 80 people
Odds = 80 / (100-80) = 4
Number of non-target users accessing the sensor URL = 20
Odds = 20 / (100-20) = 0.25
Odds ratio = 4 / 0.25 = 16

The odds ratio is calculated for each URL, and the URL whose odds ratio is equal to or higher than a predetermined threshold value or a certain number of higher ranks is determined as the sensor URL.

FIG. 3 is an explanatory diagram showing input / output data of the machine learning engine of FIG.

[Machine learning engine 12 (learning stage)]
The machine learning engine 12 inputs the following data output from the user group classification unit 11 as a learning stage.
Access count of each sensor URL for each target user as a positive example Access count for each sensor URL for each non-target user as a negative example Specifically, each sensor URL is elemented for each user, and the access count is an element. Configuration data such as a vector as a value is input to the machine learning engine 12.
Sensor URL: [URL1] [URL2] [URL3] ・ ・ ・ ・ ・ [URLn]
Target user 1: 1 0 2 ・ ・ ・ ・ ・ ・ 0
Target user 2: 0 1 0 ・ ・ ・ ・ ・ ・ 2
・ ・ ・ ・: ・ ・ ・ ・ ・
Non-target user m: 0 1 0 ・ ・ ・ ・ ・ ・ 1
・ ・ ・ ・: ・ ・ ・ ・ ・

The machine learning engine 12 is based on supervised learning. The teacher data x and y (binary classification problem) are input in association with each other, and the mapping function between the x and y data is approximately learned. Thereby, the prediction data y can be obtained by inputting the unknown data x. Specifically, a support vector machine, a neural network, or the like can be used.

<Estimation stage>
[Machine learning engine 12 (estimation stage)]
The machine learning engine 12 inputs the number of accesses of each sensor URL for the estimation target user.
Sensor URL: [URL1] [URL2] [URL3] ・ ・ ・ ・ ・ [URLn]
Estimated target user: 2 1 1 ・ ・ ・ ・ ・ ・ 0
Then, the machine learning engine 12 outputs a regular analogy score.
The "normal example score" is a score that will reach the predicted target behavior (normal example) for the estimated target user. The score is a probability value (score) output from a general machine learning engine.

[Potential target user estimation unit 13]
The latent target user estimation unit 13 estimates an estimation target user whose regular similarity score is equal to or higher than a predetermined threshold value as a latent target user.
Specifically, an estimation target user having a regular similarity score of 70% or more is estimated as a potential target user.

FIG. 4 is a functional configuration diagram of the estimation device having the acceptance reaction learning stage in the present invention.

According to FIG. 4, when a predicted target behavior is recommended (recommended) to a user by, for example, advertisement distribution, the potential target user is determined based on the behavior occurrence probability of whether or not each user causes the predicted target behavior. It is something to estimate.

For example, a user who has a good experience with advertisement distribution (recommendation) in the past is interested in the subsequent advertisement distribution and has high receptivity, whereas a user who has an unpleasant experience in the past is an advertisement. He has no interest in delivery and has low receptivity.

According to the above-described embodiments of FIGS. 2 and 3, it is attempted to calculate the occurrence probability of whether or not to cause the predicted target action based on the access history of the user's own action, but according to the embodiment of FIG. We are trying to calculate the probability of occurrence of whether or not to take the predicted behavior based on the user's acceptability to the recommendation.
That is, even if the potential target user is suitable for the predicted target behavior, if the user has low receptivity to the recommendation, the potential target user cannot be used.

According to FIG. 4, as compared with the estimation device 1 of FIG. 2, the advertisement information distribution unit 15, the advertisement reaction detection unit 16, the regular acceptance score calculation unit 17, and the regular acceptance score are set as the acceptance response learning stage. It further has a storage unit 18. These functional components are also realized by executing a program that makes the computer mounted on the device function.

According to FIG. 4, a learning step is executed for each of a plurality of target URLs for the estimation target user, and the machine learning engine 12 constructs a different learning model for each target URL.
Then, the access history of the sensor URL in one estimation target user is input for each learning model of different target URLs, and the estimation step is executed. As a result, the machine learning engine 12 estimates the regular similarity score for each target URL. These target URLs and correct example similarity scores are output to the latent target user estimation unit 13.
[Target URL] <-> [Similar score]
http://www.trg1.com <-> 60%
http://www.trg2.com <-> 50%
http://www.trg3.com <-> 30%

[Advertising information distribution unit 15]
The advertisement information distribution unit 15 distributes advertisement information based on each target URL for each target URL to the terminal of the estimation target user.
For example, the advertisement information in which the target URL is described may be transmitted by e-mail or a message to the terminal of the estimation target user.

[Advertising reaction detection unit 16]
The advertisement reaction detection unit 16 detects whether or not there is a reaction operation of the estimation target user with respect to the advertisement information for each target URL.
The reaction operation includes, for example, a tap operation for a URL and a reply operation for an e-mail (or a message). The reaction operation is not limited to the one directly notified from the user's terminal to the estimation device 1, and may be collected as an access history.

Here, the advertisement information distribution unit 15 may describe the URL based on the redirect request to the estimation device 1 instead of describing the target URL itself in the advertisement information. As a result, the user taps the URL, and the redirect request is transmitted to the advertisement reaction detection unit 16 of the estimation device 1. The advertisement reaction detection unit 16 returns a redirect response including the target URL to the user's terminal. This makes it possible to detect the user's advertising reaction.

The advertisement reaction detection unit 16 stores the acceptance reaction for each user.
[Target URL] <-> [Similar score] <-> [Advertising response]
http://www.trg1.com <-> 60% <-> 1 (Yes)
http://www.trg2.com <-> 50% <-> 1 (Yes)
http://www.trg3.com <-> 30% <-> 0 (none)

[Regular acceptance score calculation unit 17]
The positive example acceptance score calculation unit 17 has a regression coefficient I of whether or not there is an advertising response for the estimated target user, based on the regular example similarity score a and the ad reaction presence / absence r (= 0,1) for each target URL. Is calculated as the positive acceptance score. Specifically, it may be a simple multiplication as follows.
I = Σ _i (a * r) / j
i: Target URL
a: Normal similarity score
r: With / without advertising reaction r (= 0,1)
j: Number of advertising reactions

According to the above-mentioned positive acceptance score storage unit 18, the positive acceptance score is calculated as follows, for example.
I = (60% * 1 + 50% * 1 + 30% * 0) / 2 = 55%
That is, it is calculated that the positive acceptance score = 55% for this estimated target user.

[Regular acceptance score memory 18]
The positive acceptance score storage unit 18 stores the positive acceptance score for each estimation target user.

FIG. 5 is a functional configuration diagram of an estimation device that estimates a potential target user based on a positive acceptance score.

According to FIG. 5, the positive acceptance score calculation unit 17 is further provided as the acceptance response estimation step. These functional components are also realized by executing a program that makes the computer mounted on the device function.

The latent target user estimation unit 13 weights the normal case acceptance score of the estimation target user to the normal case similarity score calculated for the target URL by using the access history to the sensor URL in the estimation target user.
A = a * R
a: Normal similarity score
R: Positive case acceptance score Then, the latent target user estimation unit 13 estimates the estimated target user whose regular case similarity score is equal to or higher than a predetermined threshold value as the potential target user.

As described above in detail, according to the program, device and method of the present invention, it is possible to estimate the potential target user who will reach the predicted target behavior by using the access history of the URL. Further, according to the present invention, for the positive / negative conditions for the user group, the positive example acceptance score in the estimated target user is weighted to the positive example similarity score in the machine learning engine, so that the advertising target user group Can be narrowed down with high accuracy, and it can be expected to improve the advertising effect.

Various modifications, modifications and omissions of the above-mentioned various embodiments of the present invention within the scope of the technical idea and viewpoint of the present invention can be easily made by those skilled in the art. The above explanation is just an example and does not attempt to limit anything. The present invention is limited only to the scope of claims and their equivalents.

1 Estimator 101 Access history storage unit 102 Sensor URL registration unit 11 User group classification unit 12 Machine learning engine 13 Latent target user estimation unit 14 Sensor URL determination unit 15 Advertisement information distribution unit 16 Advertisement reaction detection unit 17 Correct case acceptance score calculation unit 18 Positive acceptance score memory

Claims (7)

A program that activates a computer installed in a device that estimates potential target users that will lead to predicted behavior.
An access history storage means for accumulating accessed URLs for each user,
It functions as a sensor URL registration means for registering a plurality of sensor URLs related to the predicted behavior.
As a learning stage
If the predicted behavior is an access to the specified target URL, the classification condition is whether or not the target URL is accessed (positive example) or not (negative example).
Using the access history storage means, with respect to the classification condition, and the target user group to be positive cases, and the user group classification means for classifying into a non-target user group to be negative cases,
As a machine learning engine that learns by inputting the access count of each sensor URL associated with each target user as a positive example and the access count of each sensor URL associated with each non-target user as a negative example. Make it work,
As an estimation stage
The machine learning engine inputs the number of accesses of each sensor URL for the estimation target user, outputs a correct example similarity score, and outputs the same number.
A program characterized in that a computer functions as a latent target user estimation means for estimating an estimated target user whose regular similarity score is equal to or higher than a predetermined threshold value as a latent target user. For the estimation target user, the learning stage and the estimation stage are executed for each of the plurality of target URLs, and the regular similarity score is estimated for each target URL.
An advertisement information distribution means that distributes advertisement information based on each target URL to the terminal of the estimation target user for each target URL.
For each target URL, an advertisement reaction detection means that detects whether or not there is a reaction operation of the estimated target user with respect to the advertisement information,
Further, as a normal acceptance score calculation means for calculating the regression coefficient between the presence / absence of an advertising response and the presence / absence of an advertising response as the regular acceptance score for the estimated target user based on the regular similarity score and the presence / absence of an advertising response for each target URL. Make it work,
The potential target user estimation means, wherein said calculated positive cases receiving score Te weighted with the positive cases similar score, according to claim 1, characterized by causing a computer to function so as to estimate the potential target user Program. A program that activates a computer installed in a device that estimates potential target users that will lead to predicted behavior.
An access history storage means for accumulating accessed URLs for each user,
It functions as a sensor URL registration means for registering a plurality of sensor URLs related to the predicted behavior.
As a learning stage
When the predicted behavior is access to the URL searched by the specified search key, the classification condition is whether or not the URL searched by the search key is accessed (positive example) or not (negative example). ,
If the predicted behavior is a specified behavior in the behavior history recorded for each user, the classification condition is whether or not the behavior was performed (positive example) or not (negative example).
Using the access history storage means, and the target user group to be positive cases with respect to the classification condition, the user group classification means for classifying into a non-target user group to be negative cases,
As a machine learning engine that learns by inputting the access count of each sensor URL associated with each target user as a positive example and the access count of each sensor URL associated with each non-target user as a negative example. Make it work,
As an estimation stage
The machine learning engine inputs the number of accesses of each sensor URL for the estimation target user, outputs a correct example similarity score, and outputs the same number.
A program characterized in that a computer functions as a latent target user estimation means for estimating an estimated target user whose regular similarity score is equal to or higher than a predetermined threshold value as a latent target user. A computer as a sensor URL determining means for determining a URL in which a first probability of access by a target user group and a second probability of access by a non-target user group satisfy a predetermined condition for each URL as the sensor URL. The program according to any one of claims 1 to 3, wherein the program is made to function. The fourth aspect of the present invention is characterized in that the sensor URL determining means causes a computer to function as the sensor URL to determine a URL in which the odds ratio between the first probability and the second probability is equal to or higher than a predetermined threshold value. The listed program. An estimation device that estimates potential target users that will lead to predicted behavior.
An access history storage means for accumulating accessed URLs for each user,
It has a sensor URL registration means for registering a plurality of sensor URLs related to the predicted behavior.
As a learning stage
If the predicted behavior is an access to the specified target URL, the classification condition is whether or not the target URL is accessed (positive example) or not (negative example).
When the predicted behavior is access to the URL searched by the specified search key, the classification condition is whether or not the URL searched by the search key is accessed (positive example) or not (negative example). ,
If the predicted behavior is a specified behavior in the behavior history recorded for each user, the classification condition is whether or not the behavior was performed (positive example) or not (negative example).
Using the access history storage means, and the target user group to be positive cases with respect to the classification condition, the user group classification means for classifying into a non-target user group to be negative cases,
A machine learning engine that learns by inputting the number of accesses of each sensor URL associated with each target user as a positive example and the number of accesses of each sensor URL associated with each non-target user as a negative example. Have,
As an estimation stage
The machine learning engine inputs the number of accesses of each sensor URL for the estimation target user, outputs a correct example similarity score, and outputs the same number.
An estimation device comprising a latent target user estimation means for estimating an estimation target user whose regular similarity score is equal to or higher than a predetermined threshold value as a latent target user. It is an estimation method of a device that estimates a potential target user that will lead to a predicted target behavior.
The device is
An access history storage unit that stores accessed URLs for each user,
A sensor URL registration unit that registers a plurality of sensor URLs related to the predicted behavior, and a sensor URL registration unit.
Has a machine learning engine and
The device is used as a learning stage.
If the predicted behavior is an access to the specified target URL, the classification condition is whether or not the target URL is accessed (positive example) or not (negative example).
When the predicted behavior is access to the URL searched by the specified search key, the classification condition is whether or not the URL searched by the search key is accessed (positive example) or not (negative example). ,
If the predicted behavior is a specified behavior in the behavior history recorded for each user, the classification condition is whether or not the behavior was performed (positive example) or not (negative example).
Using the access history storage unit, and the target user group to be positive cases with respect to the classification condition, and the eleventh step of classifying into a non-target user group to be negative cases,
Using the machine learning engine, input the access count of each sensor URL associated with each target user as a positive example and the access count of each sensor URL associated with each non-target user as a negative example. And perform the twelfth step to learn
The device is used as an estimation stage.
Using the machine learning engine, the 21st step of inputting the number of accesses of each sensor URL for the estimation target user and outputting the correct example similarity score, and
A method for estimating an apparatus, which comprises performing the 22nd step of estimating a user whose estimation target user has a regular similarity score equal to or higher than a predetermined threshold value as a potential target user.

Images