JP5442806B2 - Content distribution device - Google Patents

Description

  The present invention relates to a technique for distributing content such as advertisements via a network such as the Internet.

  In many services such as information provision on the Internet, an advertisement business is established by displaying advertisement contents together when displaying a Web page according to a user's request.

  In order to increase the advertising effectiveness, it is desirable that the user is interested, such as when the user clicks on the displayed advertisement and jumps to the detail page. There is a mechanism to decide according to the above.

  Conventionally, as a method for determining content to be presented (recommended: recommended) to a user, content having the highest CTR (Click Through Rate) is determined based on the click performance of each past content of the user (for example, Patent Document 1) Reference) and data (for example, refer to Patent Document 2) that determine highly relevant content for users who have a strong relationship in accordance with data (for example, social graph data). .

JP 2012-18571 A Japanese Unexamined Patent Publication No. 2011-233286

  As described above, since the content to be distributed has been determined based on a single index such as the click performance and the relationship between users in the past (other information may be used supplementarily to improve accuracy, Basically, based on a single indicator), there was a limit to improving the accuracy of recommending content.

  The present invention has been proposed in view of the above-described conventional problems, and an object of the present invention is to improve recommendation accuracy of content to be distributed.

  In order to solve the above problems, in the present invention, means for acquiring actual CTR data in which the click rate for each content of the user is associated with the user identifier and the content identifier, Corresponding to means for obtaining content × content matrix data associated with content identifiers, means for obtaining user × user matrix data in which the degree of association between users is associated with two-axis user identifiers, and the actual CTR data Based on the user × content matrix data, the content × content matrix data, and the user × user matrix data, a matrix of predicted values corresponding to these actual value matrices is represented by a product of two decomposed matrices, Convergence calculation is performed so that the loss function value indicating the difference between the actual value and the predicted value is minimized. When there is a request for specifying a user identifier from the service providing device based on the generated predicted CTR data that is the x content matrix data and the generated predicted CTR data, the click rate by the user of the user identifier is the highest. Means for selecting a content identifier of the content expected to be high and responding the corresponding content to the service providing apparatus.

  In the content distribution apparatus of the present invention, the recommendation accuracy of the content to be distributed can be improved by considering a plurality of indices simultaneously.

It is a figure which shows the structural example of the system concerning one Embodiment of this invention. It is a figure which shows the data structure example of a browsing log. It is a figure which shows the example of a data structure of track record CTR data. It is a figure which shows the data structure example of affinity data. It is a figure which shows the example of a data structure of a content list. It is a figure which shows the example of a data structure of similarity data. It is a figure which shows the example of a data structure of interest graph data. It is a figure which shows the data structural example of social graph data. It is a figure which shows the data structure example of prediction CTR data. It is a figure which shows the data structure example of a content database. It is a flowchart (generation of prediction CTR data) which shows an example of processing of an embodiment. It is a figure which shows the example of calculation of the affinity data by logarithmic odds. It is explanatory drawing of the production | generation of prediction CTR data. It is a figure which shows the example of a loss function. It is a flowchart (at the time of content distribution) which shows the process example of embodiment.

  Hereinafter, preferred embodiments of the present invention will be described.

<Configuration>
FIG. 1 is a diagram showing a configuration example of a system according to an embodiment of the present invention.

  In FIG. 1, the system includes a plurality of terminal devices 1 operated by a user, a plurality of service server devices 3 that provide an information providing service to a user who uses the terminal device 1, and a user who uses the terminal device 1. A plurality of SNS (Social Networking Service) server devices 4 that provide social network services to the Internet, and a content distribution server device 5 that distributes content such as advertisements when providing services of the service server device 3 include a network 2 such as the Internet. It is connected to the. The content distribution server device 5 may have a path connected directly or indirectly (including offline data exchange) with the service server device 3 and the SNS server device 4.

  The terminal device 1, the service server device 3, the SNS server device 4, and the content distribution server device 5 have a hardware configuration of a general computer device.

  The content distribution server device 5 includes a browsing log acquisition unit 501, a performance CTR data calculation unit 503, an affinity data calculation unit 505, a similarity data calculation unit 508, an interest graph as functional units mainly realized by software (computer program). A data calculation unit 510, a social graph data acquisition unit 512, a predicted CTR calculation unit 514, a content determination unit 516, and a content distribution unit 517 are provided.

  The browsing log acquisition unit 501 has a function of acquiring a browsing log, which is a history of browsing behavior of the user, from the service server device 3 online or offline and holding it as a browsing log 502 in a storage area inside the content distribution server device 5. ing.

  FIG. 2 is a diagram showing an example of the data structure of the browsing log 502, such as “time stamp” “user ID” “content ID of the browsed content” “category corresponding to the browsed content” “content ID of the recommended content”, etc. Have items. In addition to the user ID registered as a member (preliminary login is required to obtain the user ID), the user ID includes a user ID such as an ID (also called bcookie or webcookie) that identifies the terminal device 1 that has accessed. Any one that can be specified can be used.

  Returning to FIG. 1, the actual CTR data calculation unit 503 calculates actual CTR data in which the click rate for each user content is associated with the user ID and the content ID from the browsing log 502, and the content distribution server apparatus as actual CTR data 504. 5 has a function of holding in an internal storage area. When the actual CTR data 504 can be acquired from the service server device 3 online or offline, the actual CTR data calculation unit 503 can be omitted.

  FIG. 3 is a diagram showing an example of the data structure of the actual CTR data 504, and the actual CTR value for each content ID of the content viewed by the user with the user ID is recorded for each user ID. Details of the calculation of the actual CTR data 504 will be described later.

  Returning to FIG. 1, the affinity data calculation unit 505 calculates affinity data indicating the possibility that two contents are browsed at the same time from the browsing log 502, and stores the affinity data 506 in a storage area inside the content distribution server device 5. It has a function to hold. If the affinity data 506 can be acquired from the service server device 3 online or offline, the affinity data calculation unit 505 can be omitted.

  FIG. 4 is a diagram showing an example of the data structure of the affinity data 506. For each content ID, the degree of association between the content of the content ID and other content (including the same content) (the degree of continuous browsing). Degree of occurrence) is recorded. Details of the calculation of the affinity data 506 will be described later.

  Returning to FIG. 1, the similarity data calculation unit 508 obtains similarity data indicating the similarity between two contents from the content list 507 indicating various attributes of the contents held in the storage area inside the content distribution server device 5. It has a function of calculating and storing as similarity data 509 in a storage area inside the content distribution server device 5. If the similarity data 509 can be acquired from the service server device 3 online or offline, the content list 507 and the similarity data calculation unit 508 can be omitted.

  FIG. 5 is a diagram showing an example of the data structure of the content list 507. For each content ID, the content name, category, image, etc. of the content with the content ID are recorded.

  FIG. 6 is a diagram showing an example of the data structure of the similarity data 509. For each content ID, the degree of association (name, category, image, and other attributes) between the content of the content ID and other content (including the same content) Similarity etc.) are recorded. Details of the calculation of the similarity data 509 will be described later.

  Returning to FIG. 1, the interest graph data calculation unit 510 calculates interest graph data indicating the commonality of the content viewed by the two users from the browsing log 502, and stores it as the interest graph data 511 in the content distribution server device 5. It has a function to hold in the area. When the interest graph data 511 can be acquired from the service server device 3 or the SNS server device 4 online or offline, the interest graph data calculation unit 510 can be omitted.

  FIG. 7 is a diagram showing an example of the data structure of the interest graph data 511. For each user ID, the degree of association between the user with the user ID and other users (including the same user) (the degree of commonality of the content category to be browsed, etc.) ) Is recorded. Details of the calculation of the interest graph data 511 will be described later.

  Returning to FIG. 1, the social graph data acquisition unit 512 acquires the interest graph data indicating the interchangeability of the two users from the SNS server device 4 online or offline, and the social graph data 513 is stored in the content distribution server device 5. It has a function of holding in the storage area.

  FIG. 8 is a diagram showing an example of the data structure of the social graph data 513, and the degree of association between the user of the user ID and other users (including the same user) is recorded for each user ID.

  Returning to FIG. 1, the predicted CTR calculation unit 514 calculates predicted CTR data from the actual CTR data 504, the affinity data 506, the similarity data 509, the interest graph data 511, and the social graph data 513, and as predicted CTR data 515. It has a function of holding it in a storage area inside the content distribution server device 5. Details of the calculation will be described later.

  FIG. 9 is a diagram showing an example of the data structure of the predicted CTR data 515, and for each user ID, a predicted CTR value for each content ID of the content viewed by the user with the user ID is recorded.

  Returning to FIG. 1, when there is a content distribution request specifying the user ID from the service server device 3, the content determination unit 516 has the highest click rate by the user of the specified user ID based on the predicted CTR data 515. It has a function of determining content (content ID) expected to be high. Details of the processing will be described later.

  The content distribution unit 517 has a function of acquiring the content acquisition URL determined by the content determination unit 516 or the content data itself from the content database 518 and distributing it to the requesting service server device 3. When the service server device 3 embeds a URL for acquiring content in a Web page, when the Web page is displayed on the terminal device 1, the browser requests the URL for acquiring content to acquire content data.

  FIG. 10 is a diagram showing an example of the data structure of the content database 518, in which the content URL, content data, etc. of the content ID are recorded for each content ID.

<Operation>
FIG. 11 is a flowchart showing an example of processing related to generation of predicted CTR data in the above embodiment.

  In FIG. 11, when processing is started (step S101), the browsing log acquisition unit 501 acquires a browsing log from the service server device 3 (step S102), and holds it as a browsing log 502 (step S103).

  Next, the actual CTR data calculation unit 503 aggregates CTRs from the browsing log 502 to generate actual CTR data (step S104), and holds the actual CTR data 504 (step S105). When the actual CTR data 504 in FIG. 3 is calculated from the browsing log 502 in FIG. 2, the user ID and the content ID of the browsed content are aggregated, and the ratio of the number of times of browsing to the recommended number of times is defined as the actual CTR.

  Next, returning to FIG. 11, the affinity data calculation unit 505 generates affinity data from the browsing log 502 by logarithmic odds or the like (step S106) and holds it as affinity data 506 (step S107).

  FIG. 12 is a diagram illustrating a calculation example of the affinity data 506 using logarithmic odds.

In the Venn diagram of FIG. 12A, X is the number of browsing of all content, A is the number of browsing of content A, B is the number of browsing of content B, and A∩B is browsing content B at the same time following content A. This is the number of browsing. Where:
score = [log ((A∩B) + α) + log (X + (A∩B) -A-B + β) -log (AA∩B + β) -log (BA∩B + α)] / log (Ten)
By calculating the score score according to the above, and using it as the relevance level of the content B to the content A, the affinity data 506 can be generated. Α and β are constants determined by operation.

  In this case, as shown in FIG. 12 (b), even if the number of contents A viewed and the value of A∩B are almost the same, the pattern B is more than the number of contents B viewed as in pattern # 1. The score score is larger when the number of browsing contents B is smaller as in # 2, and the relevance is evaluated to be higher.

  On the other hand, returning to FIG. 11, the similarity data calculation unit 508 generates similarity data based on cosine similarity or the like from the content list 507 held internally or acquired from the outside (step S108), and held as the similarity data 509. (Step S109). When calculating the similarity data 509 of FIG. 6 from the content list 507 of FIG. 5, the similarity between the content names, categories, images, etc. of the content of the two content IDs (if the content name or category, the character string The degree of coincidence, a parent-child relationship / brother relation for categories, a degree of coincidence of image contents for images, etc.) is calculated based on cosine similarity.

  On the other hand, returning to FIG. 11, the interest graph data calculation unit 510 generates interest graph data from the browsing log 502 based on the number of browsing categories common among users (step S <b> 110), and holds the interest graph data 511 (FIG. 11). Step S111). When the interest graph data 511 in FIG. 7 is calculated from the browsing log 502 in FIG. 2, for example, the relevance degree is set to “1” for two users who have a predetermined number (for example, 8) or more of common category browsing histories. (Default is “0”). Moreover, it is good also as a numerical value according to the number of browsing histories of a common category, without setting it as a value only "1" "0".

  On the other hand, returning to FIG. 11, the social graph data acquisition unit 512 acquires social graph data from the SNS server device 4 (step S112), and holds it as social graph data 513 (step S113).

  Next, the predicted CTR calculation unit 514 inputs the actual CTR data 504, the affinity data 506, the similarity data 509, the interest graph data 511, and the social graph data 513, and the actual CTR data 504 is directly used as user × content matrix data. deal with.

  As for the affinity data 506 and the similarity data 509, content × content matrix data (matrix data in which the degree of association between contents is associated with a biaxial content ID) is selected by selecting one of them or averaging both. Generate (step S114). For interest graph data 511 and social graph data 513, user × user matrix data (matrix data in which the degree of association between users is associated with a biaxial user ID) by selecting one of them or averaging both, etc. Is generated (step S115).

  Next, the predicted CTR calculation unit 514 is a product of matrixes obtained by decomposing the respective matrixes of predicted values for the user × content matrix data of the actual CTR data, the content × content matrix data, and the user × user matrix data. And the convergence calculation is performed so that the value of the loss function indicating the difference between the actual value and the predicted value (also including the regularization term) is minimized, and predicted CTR data (user × content matrix data) is generated from the converged result ( Step S116), the data is stored as predicted CTR data 515 (Step S117), and the process is terminated (Step S118).

  FIG. 13 is an explanatory diagram of generation of predicted CTR data 515.

In FIG. 13, the element of the user × content matrix data corresponding to the actual CTR data 504 is y _ij , the element of the content × content matrix data corresponding to the affinity data 506 and the similarity data 509 is s _ii , and the interest graph data 511 Elements of user × user matrix data corresponding to the social graph data 513 are indicated by a _{jj ′} .

Also, the user × content matrix data element corresponding to the predicted CTR data 515 is f _ij , the predicted value content × content matrix data element is h _ii , and the predicted value user × user matrix data element is g _{jj ′} . Each is represented as the product of two matrices.

  From these, the value of the formula shown in the lower part of FIG. When the change disappears, it is determined that the data has converged), and the predicted CTR data 515 corresponding to the user × content matrix data is acquired.

  Note that λ is a constant determined in operation. Moreover, l (el) is a loss function, and a function as shown to Fig.14 (a)-(e) is employable. The regularization term is a value that depends on the norm (length) of the decomposed matrix.

  Next, FIG. 15 is a flowchart showing an example of processing at the time of content distribution in the above embodiment.

  In FIG. 15, the content distribution server device 5 receives the content distribution request accompanied by the user ID from the service server device 3, and starts processing (step S201).

  The content determination unit 516 of the content distribution server device 5 refers to the record of the corresponding user ID from the predicted CTR data 515 and acquires the content ID having the highest CTR performance value (step S202).

  Next, the content distribution unit 517 acquires the URL or content data of the corresponding content ID from the content database 518 based on the content ID determined by the content determination unit 516, and responds to the requesting service server device 3 (step S203). ), The process is terminated (step S204).

  Upon receiving the URL or content data, the service server device 3 embeds the URL or content data at a predetermined position on the Web screen provided by the service, and provides page data to the terminal device 1.

  Content such as advertisements embedded in the Web screen is expected to be easily clickable by the user of the terminal device 1 and is expected to be viewed (clicked) with high probability.

<Summary>
As described above, according to the present embodiment, the recommendation of the content to be distributed by simultaneously considering a plurality of indicators such as the degree of association between the contents and the degree of association between the users, as well as the click results for each content of the user. Accuracy can be improved.

  The present invention has been described above by the preferred embodiments of the present invention. While the invention has been described with reference to specific embodiments, various modifications and changes may be made to the embodiments without departing from the broad spirit and scope of the invention as defined in the claims. Obviously you can. In other words, the present invention should not be construed as being limited by the details of the specific examples and the accompanying drawings.

DESCRIPTION OF SYMBOLS 1 Terminal apparatus 2 Network 3 Service server apparatus 4 SNS server apparatus 5 Content delivery server apparatus 501 Browsing log acquisition part 502 Browsing log 503 Actual CTR data calculation part 504 Actual CTR data 505 Affinity data calculation part 506 Affinity data 507 Content list 508 Similarity Data calculation unit 509 Similarity data 510 Interest graph data calculation unit 511 Interest graph data 512 Social graph data acquisition unit 513 Social graph data 514 Prediction CTR calculation unit 515 Prediction CTR data 516 Content determination unit 517 Content distribution unit 518 Content database

Claims (5)

Means for acquiring actual CTR data in which a click rate for each content of a user is associated with a user identifier and a content identifier;
Means for acquiring content × content matrix data in which the degree of association between contents is associated with a biaxial content identifier;
Means for acquiring user × user matrix data in which the degree of association between users is associated with a two-axis user identifier;
Based on the user × content matrix data corresponding to the actual CTR data, the content × content matrix data, and the user × user matrix data, a matrix of predicted values corresponding to these actual value matrices is decomposed 2 Convergence calculation is performed so that the value of the loss function indicating the difference between the actual value and the predicted value is minimized, and predicted CTR data that is user × content matrix data of the predicted value is generated from the converged result. Means,
When there is a request specifying the user identifier from the service providing apparatus based on the generated predicted CTR data, the content identifier of the content that is expected to have the highest click rate by the user of the user identifier is selected and the corresponding content And a means for responding to the service providing apparatus. The content delivery apparatus according to claim 1,
The content × content matrix data is either affinity data indicating the possibility that two contents are browsed at the same time, similarity data indicating the similarity of two contents, or data obtained by combining them. A feature content distribution apparatus. The content distribution apparatus according to claim 2, wherein
The affinity data includes A as the number of times the first content has been viewed, B as the number of times the second content has been viewed, and A as the number of times the second content has been viewed following the first content. ∩B, constants α, β, log odds score score
score = [log ((A∩B) + α) + log (X + (A∩B) -A-B + β) -log (AA∩B + β) -log (BA∩B + α)] / log (Ten)
Is a degree of relevance of the second content to the first content. In the content delivery apparatus as described in any one of Claims 1 thru | or 3,
The user × user matrix data is either interest graph data indicating the commonality of content viewed by two users, social graph data indicating the interchangeability of two users, or data obtained by combining these. A content distribution apparatus characterized by the above. A step in which the content distribution apparatus acquires the actual CTR data in which the click rate for each content of the user is associated with the user identifier and the content identifier;
The content distribution device acquires content × content matrix data in which the degree of association between content is associated with a biaxial content identifier;
The content distribution apparatus acquires user × user matrix data in which the degree of association between users is associated with a biaxial user identifier;
Based on the user × content matrix data corresponding to the actual CTR data, the content × content matrix data, and the user × user matrix data, the content distribution apparatus has predicted values corresponding to these actual value matrixes. The matrix is expressed by the product of the two decomposed matrices, and the convergence calculation is performed so that the value of the loss function indicating the difference between the actual value and the predicted value is minimized. Generating predicted CTR data;
When there is a request for specifying a user identifier from the service providing apparatus based on the generated predicted CTR data, the content distribution apparatus is configured to select a content identifier of the content that is predicted to have the highest click rate by the user of the user identifier. Selecting and responding the corresponding content to the service providing apparatus.

Images