JP2017168110A - Information processing apparatus, information processing method, and information processing program - Google Patents

Abstract

An information processing apparatus, an information processing method, and an information processing program capable of reducing the calculation time required for predicting behavior on the Internet performed by a user operating a user terminal are provided.
An information processing apparatus includes: a prediction model for calculating a prediction value related to an execution probability of an action on the Internet performed by a user operating a user terminal; and a group of users classified into groups based on user information. Using the prediction model selected by the model selection unit 24, the model generation unit 23 to generate, the model selection unit 24 that selects a prediction model that matches the user from the prediction model, and the value of the feature information is acquired from the behavior. And a predicted value calculating means 25 for calculating a predicted value. The model selection means 24 selects the prediction model corresponding to the acquired feature information value using the relationship information indicating the relationship between the value of the feature information related to the behavior and the prediction model.
[Selection] Figure 1

Description

  The present invention relates to an information processing apparatus, an information processing method, and an information processing program for predicting behavior on the Internet performed by a user operating a user terminal.

Conventionally, an information providing apparatus that provides an advertisement to a user via the Internet is known (see Patent Document 1).
Among such information providing devices, in order to improve the advertising effect, a prediction value of the probability that the user clicks the advertisement on the web page is calculated, and the advertisement is provided to the user according to the calculated prediction value There is. And the prediction model which calculates a prediction value based on the feature information which shows a user attribute, the relevance of a user attribute, and an advertisement etc. is used for calculation of a prediction value.

JP 2009-193465 A

  However, conventionally, a predicted value is calculated using a common prediction model for all users. For this reason, in order to cope with users having various attributes, it is necessary to include a huge number of feature information in the prediction model. As a result, there is a problem that the calculation time of the predicted value becomes long.

  An object of the present invention is to provide an information processing apparatus, an information processing method, and an information processing program that can reduce a calculation time required for predicting behavior on the Internet performed by a user operating a user terminal.

  The information processing apparatus according to the present invention generates a prediction model for calculating a prediction value related to an execution probability of an action on the Internet performed by a user operating a user terminal for each group in which users are grouped based on user information. A generation unit; a model selection unit that selects a prediction model suitable for a user from the prediction model; and a value of feature information is acquired from the behavior, and the prediction model selected by the model selection unit is used to perform the prediction Prediction value calculation means for calculating a value, and the model selection means corresponds to the value of the acquired feature information by using relationship information indicating a relationship between the value of the feature information related to the behavior and the prediction model. A prediction model to be selected is selected.

In the present invention, the prediction model is generated for each group in which users are grouped based on user information. According to this, since each prediction model calculates a predicted value for a user having a relatively close attribute, the number of feature information used for calculating the predicted value is the same for all users. This can be reduced compared to the case of calculating the predicted value.
And in this invention, since a predicted value is calculated using the prediction model selected from the prediction model produced | generated for every such group, a predicted value is calculated using the prediction model common to all the users. Compared to the case, the calculation time of the predicted value can be shortened.

It is a figure which shows the structure of the information processing apparatus which concerns on 1st Embodiment of this invention. It is a flowchart which shows the model allocation process in 1st Embodiment. It is a flowchart which shows the prediction process in 1st Embodiment. It is a figure which shows the structure of the information processing apparatus which concerns on 2nd Embodiment of this invention.

[First Embodiment]
DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, a first embodiment of the invention will be described with reference to the drawings.
[Information processing device]
FIG. 1 is a diagram illustrating the configuration of the information processing apparatus according to the first embodiment.
The information processing apparatus 1 according to the first embodiment is configured by a computer, and includes a storage device 10 and a control device 20. The information processing apparatus 1 is connected to a network 3 to which a plurality of user terminals 2 can be connected.

[Storage device]
The storage device 10 is configured by an HDD (Hard Disk Drive), a flash memory, or the like, and stores various programs and information necessary for the operation of the control device 20.
The storage device 10 includes a model storage unit 11, a click history storage unit 12, a model assignment information storage unit 13, and a relationship information storage unit 14.

The model storage unit 11 stores a prediction model for calculating a predicted CTR that is a predicted value of a probability (CTR: Click-Through rate) that a specific user clicks a specific advertisement on a web page.
A prediction model is shown by following Formula (1), for example. In the formula (1), A and B are constants. y is shown by following Formula (2). n represents a variable, Xn represents a numerical value of the feature information, and Wn represents a weighting factor that determines a contribution degree by which Xn contributes to the predicted CTR.

The feature information is various information related to the user and the advertisement. For example, the user attribute, the similarity between the delivery target page and the advertisement, the relevance between the user attribute and the advertisement, the information of the advertisement itself, and the past delivery results. It is information which shows etc.
In addition, as long as the prediction model calculates | requires the predicted value regarding a click probability, it does not need to calculate the click probability itself. For example, y indicated by the expression (2) may be obtained.

Here, the model storage unit 11 includes a prediction model (also referred to as a local model) generated for each group in which users are grouped based on user attributes, and a common prediction model (also referred to as a common model) common to all groups. Is stored).
For example, examples of the local model include a housewife model generated for a housewife group, a rich model generated for a wealthy group, and the like.

The click history accumulation unit 12 (which constitutes the action history accumulation unit of the present invention) is a click history indicating the click result of the distributed advertisement when the user has accessed a predetermined web page in the past (configures the action history of the present invention). Accumulate. The click history includes information such as user ID (IDentifiers: identification information), web page ID, advertisement ID, presence / absence of click, and click date / time.
Here, the click history accumulating unit 12 accumulates the click history in association with any one of the local model and the common model.

  The model assignment information storage unit 13 stores model assignment information in which a user ID is associated with a prediction model assigned to a user by a model assignment process described later.

  The relationship information accumulating unit 14 accumulates relationship information in which a feature information value acquired by a model allocation process described later is associated with a prediction model allocated by the model allocation process.

[Control device]
The control device 20 controls the operation of the information processing device 1 and is configured by an arithmetic device such as a CPU (Central Processing Unit).
In the control device 20, when the arithmetic device executes the information processing program stored in the storage device 10, the model assignment unit 21, the click history accumulation control unit 22, the model generation unit 23, the model selection unit 24, and the predicted value calculation unit 25, functions as distribution means 26.

The model assigning means 21 assigns a prediction model suitable for the user from the common model and the local model generated for each group for each user. In addition, model assignment information that associates the user with the assigned prediction model is stored in the model assignment information storage unit 13.
Further, the model assigning means 21 reads the accumulated click history from the click history accumulation unit 12 for each user, and acquires the value of the feature information based on the read click history. Then, the relationship information storage unit 14 stores the relationship information associating the acquired feature information value with the assigned prediction model.

  When the click history accumulation control means 22 (which constitutes the action history accumulation control means of the present invention) acquires the click history of the distribution advertisement obtained when the user accesses a predetermined web page, the acquired click history will be described later. The click history storage unit 12 stores the prediction model in association with the prediction model selected by the model selection unit 24.

The model generation unit 23 generates each prediction model based on the corresponding click history accumulated in the click history accumulation unit 12 and stores it in the model storage unit 11.
Specifically, the model generation unit 23 acquires a large number of sets of feature information values and click results based on the accumulated click history, and learns the contribution degree (Wn) of the feature information to the predicted CTR. Generate a prediction model.
Here, the model generation unit 23 periodically generates a prediction model, and overwrites and stores the newly generated prediction model in the model storage unit 11.

  The model selection unit 24 selects a prediction model suitable for the user based on the model assignment information stored in the model assignment information storage unit 13 or the relationship information stored in the relationship information storage unit 14.

  When the predicted value calculating unit 25 acquires page access information indicating that the user has accessed a predetermined web page, the predicted value calculating unit 25 calculates a predicted CTR for each advertisement of the distribution candidate using the predicted model selected by the model selecting unit 24. To do.

The distribution unit 26 determines and distributes an advertisement to be distributed to the user based on the predicted CTR calculated by the predicted value calculation unit 25.
For example, an advertisement with the highest predicted CTR among advertisements of distribution candidates is distributed to the user.

[Model assignment process]
Next, the model assignment process will be described.
FIG. 2 is a flowchart showing the model assignment process. The model assignment process is executed at predetermined time intervals such as one week.
As shown in FIG. 2, when the model allocation process is executed, the model allocation unit 21 selects one user from the users whose click history is stored in the click history storage unit 12 (step S11). .

  Next, the model assigning means 21 reads and acquires the click history accumulated for the selected user from the click history accumulation unit 12 (step S12), and further, based on the obtained click history, the value of the feature information is obtained. Obtain (step S13).

Next, the model assignment unit 21 tests the plurality of prediction models stored in the model storage unit 11 using the acquired feature information value, and assigns a prediction model suitable for the user (step S14).
Specifically, the model assigning means 21 inputs the acquired feature information value for each prediction model, and calculates a prediction CTR. Then, by comparing the calculated prediction CTR and the actual value obtained from the acquired click history, a prediction model that matches the user is assigned from a plurality of prediction models.
For example, the model assigning means 21 assigns to the user a prediction model in which the calculated prediction CTR is closest to the actual value.
That is, the model allocating unit 21 allocates a suitable local model to the user if any of the local models is matched, and assigns a common model if any of the local models is matched and the common model is matched. .

Next, the model assignment unit 21 stores model assignment information in which the user is associated with the assigned prediction model in the model assignment information storage unit 13 (step S15). When the model assignment information corresponding to the user is already stored in the model assignment information storage unit 13, the old model assignment information is overwritten with the new model assignment information.
Further, the model allocating unit 21 accumulates relationship information in which the acquired feature information value is associated with the allocated prediction model in the relationship information accumulation unit 14 (step S16).

Next, the model assigning means 21 determines whether or not all the users whose click histories are stored in the click history storage unit 12 have been selected (step S17).
If it is determined NO in step S17, the model assigning unit 21 returns the process to step S11 and executes the processes of steps S11 to S17 again. When all the users are selected and it is determined YES in step S17, the model assignment process is terminated.

[Prediction processing]
Next, the prediction process will be described.
FIG. 3 is a flowchart showing the prediction process. The prediction process is executed, for example, when the user accesses a predetermined web page.
As illustrated in FIG. 3, when the prediction process is executed, the predicted value calculation unit 25 acquires page access information indicating that the user has accessed a predetermined web page, and based on the acquired page access information, The value of the feature information is obtained for each distribution candidate advertisement (step S21).

Next, the predicted value calculation means 25 determines whether or not the model assignment information of the target user is stored in the model assignment information storage unit 13 (step S22).
When it determines with YES by step S22, the model selection means 24 selects the prediction model (predictive model of a local model and a common model) related by the said model allocation information (step S23).
On the other hand, when it is determined NO in step S22, the model selection unit 24 selects a prediction model having a high correlation with the value of the acquired feature information based on the relationship information stored in the relationship information storage unit 14. To do.
That is, if there is a local model having a high correlation with the value of the acquired feature information, the model selecting unit 24 selects the local model, and if not, selects the common model.

  After the processing of step S23 or step S24, the predicted value calculation means 25 uses the prediction model selected in step S23 or step S24 for each distribution candidate advertisement to set the acquired feature information value. Based on this, a predicted CTR is calculated (step S25).

  Next, the distribution unit 26 determines an advertisement to be distributed to the user, based on the predicted CTR calculated for each distribution candidate advertisement. For example, the distribution unit 26 determines an advertisement having the highest predicted CTR as a distribution advertisement. And the delivery means 26 delivers the said advertisement to a user (step S26). A plurality of advertisements to be distributed may be selected.

Next, the click history accumulation control unit 22 acquires a click history for the distributed advertisement, and accumulates the acquired click history in the click history accumulation unit 12 (step S27).
Here, when the prediction model selected in Step S23 or Step S24 is a local model, the click history accumulation control unit 22 converts the acquired click history into the common model and the selected local model. Correspondingly accumulates in the click history accumulating unit 12.
In addition, when the selected prediction model is a common model, the acquired click history is stored in the click history storage unit 12 in association with the common model.
And the control apparatus 20 complete | finishes a prediction process.

[Effects of First Embodiment]
Since the prediction CTR is calculated based on the prediction model (local model) generated for each group, the number of feature information used for the calculation can be reduced compared with the case where the prediction CTR is calculated using the common model. CTR calculation time can be shortened.
In addition, since the predicted CTR is calculated using a local model that matches the user, the accuracy of the predicted CTR is improved compared to the case where the predicted CTR is calculated using a common model for users having various attributes. it can.

  The model generation means 23 generates a common model that is common among groups as a prediction model. According to this, the predicted CTR can be calculated by the common model for users who do not belong to any group.

The model assigning means 21 assigns a predictive model suitable for the user by testing the predictive model generated for each group using the user's click history. And the model selection means 24 selects the prediction model which the model assignment means 21 assigned to the user.
According to this, since a prediction model can be selected based on a user's click performance, a prediction model suitable for the user can be selected with high accuracy.

When there is no model assignment information of the user in the model assignment information storage unit 13, the model selection unit 24 selects a prediction model suitable for the user based on the relation information in which the value of the feature information is associated with the prediction model. .
According to this, it is possible to select a suitable prediction model even for a user who does not accumulate click history and has no model assignment information.
Further, since it is not necessary to test the prediction model, a prediction model suitable for the user can be immediately selected, and the processing load on the control device 20 for the test can be reduced.

The click history accumulation control unit 22 accumulates the user's click history in the click history accumulation unit 12 in association with the prediction model selected by the model selection unit 24, and the model generation unit 23 stores each prediction model in the click history. Generated based on the corresponding click history stored in the storage unit 12.
According to this, since the learning information for generating each prediction model increases as the click history is accumulated, the prediction accuracy of each prediction model can be improved.

  Each group is grouped based on user attributes. According to this, for example, if the administrator of the information processing apparatus 1 grasps the relationship between the user attribute and the click probability from experience, effective grouping can be performed according to the experience.

[Second Embodiment]
Next, 2nd Embodiment of this invention is described based on drawing.
In the first embodiment, a group of users is determined in advance by, for example, an administrator of the information processing apparatus 1, but in the second embodiment, a user is identified using a Dirichlet process that is a statistical method based on user information. Automatically grouped. Other configurations are the same as those of the first embodiment.

FIG. 4 is a diagram illustrating the configuration of the information processing apparatus according to the second embodiment.
As shown in FIG. 4, the control device 20A of the information processing apparatus 1A according to the second embodiment includes the same model assignment unit 21, click history accumulation control unit 22, model generation unit 23, model selection unit 24, and the like as in the first embodiment. In addition to the predicted value calculation means 25 and the distribution means 26, a grouping means 27 is provided.

The grouping unit 27 groups users based on user information (user attributes, the number of click histories accumulated in the click history accumulation unit 12 (the number of logs), etc.) using the Dirichlet process. Divide.
Here, the grouping means 27 performs a plurality of groupings in which values of parameters (also referred to as hyperparameters) that determine the degree of gathering in the Dirichlet process are set in a plurality of stages.
For example, three types of grouping are performed in which the hyperparameter values are set to “0.01”, “1”, and “100”.

  In the present embodiment, the model generation unit 23 generates a prediction model for each group for each of a plurality of groupings by the grouping unit 27 and stores the generated prediction model in the model storage unit 11. .

[Effects of Second Embodiment]
Since the grouping means 27 groups users based on the user information using the Dirichlet process, which is a statistical technique, the grouping means 27 can automatically perform appropriate grouping according to the user information. Thereby, the prediction accuracy of the prediction model can be improved without depending on the experience of the administrator of the information processing apparatus 1.

[Third Embodiment]
Next, a third embodiment of the present invention will be described.
In the first embodiment, the model assigning means 21 tests a plurality of prediction models using the value of the feature information acquired from the user's click history and assigns a prediction model suitable for the user. In the embodiment, the model allocating unit 21 allocates a prediction model in the same manner as in the first embodiment after the model allocation process is first executed, and after the predetermined period has elapsed, The prediction model is assigned based on the relationship information accumulated in the. Other configurations are the same as those of the first embodiment.

In other words, in the present embodiment, until the predetermined period (for example, two weeks) after the model assignment process is first executed, the model assigning means 21 in step S14, the model storage unit 11 is the same as in the first embodiment. A plurality of prediction models stored in the table are tested using feature information values acquired from the user's click history, so that a prediction model suitable for the user is assigned.
Thereby, in step S16, the relationship information in which the value of the feature information acquired in step S13 and the prediction model assigned in step S14 are related is stored in the relationship information storage unit 14 for the predetermined period. The

  Then, after the predetermined period has elapsed since the model allocation process was first executed, in step S14, the model allocation unit 21 stores the feature information in the relationship information for the predetermined period stored in the relationship information storage unit 14. Based on the regularity of the relationship with the prediction model, a prediction model having a high correlation with the acquired feature information is extracted and assigned to the user.

[Effects of Third Embodiment]
The model allocating means 21 allocates a prediction model suitable for the user based on the relationship information after a predetermined period of time has elapsed since the model allocation process was first executed, so there is no need to test the prediction model. Calculation time required for allocation can be shortened.

  In addition, this invention is not limited to embodiment mentioned above, In the range which can achieve the objective of this invention, the deformation | transformation shown below is also included.

[Modification 1]
In each said embodiment, although the prediction model calculated the predicted value regarding the probability that a user clicks an advertisement, this invention is not limited to this. That is, what is necessary is just to calculate the predicted value regarding the execution probability of the action on the Internet performed by the user operating the user terminal 2. Examples of such behavior include purchase of goods.
Then, the information processing apparatus 1 may perform various recommendations and suggestions to the user based on the calculated predicted value, for example.

[Modification 2]
In each of the above embodiments, the model assignment process is performed for all users whose click history is stored in the click history storage unit 12, but the present invention is not limited to this. That is, it may be performed only for some users.

[Modification 3]
In the second embodiment, the grouping means 27 performs grouping using the Dirichlet process, but the present invention is not limited to this. For example, grouping may be performed using other statistical methods.

[Modification 4]
In the third embodiment, the model allocating unit 21 tests a plurality of prediction models using the feature information values acquired from the user's click history until a predetermined period from when the model allocating process is first executed. Thus, although a prediction model suitable for the user is assigned, the present invention is not limited to this. For example, if the relationship information in which the feature information and the prediction model are related in advance is created based on the experience of the administrator of the information processing apparatus 1, the relationship information from the time when the model allocation process is first executed. A prediction model may be extracted based on the above and assigned to a user.

  DESCRIPTION OF SYMBOLS 1 ... Information processing apparatus, 2 ... User terminal, 10 ... Storage device, 11 ... Model storage part, 12 ... Click history storage part, 13 ... Model allocation information storage part, 14 ... Relation information storage part, 1A ... Information processing apparatus, DESCRIPTION OF SYMBOLS 20 ... Control apparatus, 20A ... Control apparatus, 21 ... Model allocation means, 22 ... Click history accumulation control means, 23 ... Model generation means, 24 ... Model selection means, 25 ... Predicted value calculation means, 26 ... Distribution means, 27 ... Grouping means.

Claims (7)

A model generation means for generating a prediction model for calculating a prediction value related to an execution probability of an action on the Internet performed by a user operating a user terminal for each group into which users are grouped based on user information;
Model selection means for selecting a prediction model suitable for the user from the prediction model;
Obtaining a value of feature information from the behavior, and using a prediction model selected by the model selection unit, and a prediction value calculation unit that calculates the prediction value,
The model selection means selects a prediction model corresponding to the acquired feature information value using relationship information indicating a relationship between the value of the feature information related to the behavior and the prediction model. apparatus. The information processing apparatus according to claim 1,
The information processing apparatus, wherein the model generation unit generates a common model common to all the groups as the prediction model. The information processing apparatus according to claim 1 or 2,
Comprising an action history accumulation control means for accumulating an action history as a history of the action in a storage device;
The behavior history accumulation control means accumulates the behavior history in the storage device in association with the prediction model selected by the model selection means,
The information processing apparatus, wherein the model generation unit generates the prediction model based on a corresponding action history stored in the storage device. The information processing apparatus according to any one of claims 1 to 3,
The model generation means generates the prediction model for each group in which users are grouped based on user attributes. The information processing apparatus according to any one of claims 1 to 3,
A grouping means for grouping users based on user information using the Dirichlet process, which is a statistical method,
The grouping means sets a parameter value that determines the degree of gathering in the Dirichlet process in a plurality of stages, performs a plurality of groupings,
The information processing apparatus, wherein the model generation unit generates the prediction model for each group for each of the plurality of groupings by the grouping unit. An information processing method executed by a computer,
A prediction model for calculating a prediction value related to an execution probability of an action on the Internet performed by a user operating a user terminal is generated for each group in which users are grouped based on user information,
Obtain the value of the feature information from the action,
From the prediction model, using the relationship information indicating the relationship between the value of the feature information related to the behavior and the prediction model, a prediction model corresponding to the value of the acquired feature information is selected,
An information processing method, wherein the predicted value is calculated using a selected prediction model. An information processing program for causing a computer to function as the information processing apparatus according to any one of claims 1 to 5.

Images