JP6309887B2 - Model generation apparatus, information distribution apparatus, model generation method, information distribution method, model generation program, and information distribution program - Google Patents

Description

  The present invention relates to a model generation device, an information distribution device, a model generation method, an information distribution method, a model generation program, and an information distribution program.

  In recent years, information distribution via the Internet has been actively performed. As an example of such information distribution, the click rate such as CTR (Click Through Rate) is estimated based on the similarity of the advertisement to be distributed, the information of the user who is the advertisement distribution destination, etc. A technique for selecting an advertisement to be distributed according to a click rate is known.

  For example, a technique is known in which CTR is predicted based on a component of an advertisement to be distributed and the similarity between advertisements, and an advertisement to be distributed is selected using the predicted CTR. Further, there is a technique for calculating a virtual CTR based on an attribute of a user who has selected an advertisement, and selecting an advertisement to be distributed according to the attribute of the user as a distribution destination of the advertisement and the calculated virtual CTR. Are known.

JP 2014-006684 A JP 2014-182473 A

  However, in the above-described conventional technology, when the distribution of the advertisement to be distributed and the log indicating the user who has selected the advertisement and the attribute of the user to whom the advertisement is distributed are unknown, the estimation accuracy of the click rate is There is a problem of getting worse.

  This application has been made in view of the above, and model generation that prevents deterioration in the accuracy of estimation of the click rate even when there is no advertisement log or when the attribute of the user to whom the advertisement is distributed is unknown An object is to provide an apparatus, an information distribution apparatus, a model generation method, an information distribution method, a model generation program, and an information distribution program.

  The model generation device according to the present application uses a learning device that calculates a feature amount indicating a feature of the image from an image, a calculation unit that calculates a feature amount of an image having a history selected by a user, and the calculation unit Based on the calculated feature value and the attribute information indicating the attribute of the user, a model for estimating the attribute information of the user who selected the image from the feature value of the image is generated.

  According to one aspect of the embodiment, there is an effect that it is possible to prevent deterioration in the accuracy of estimation of the click rate.

FIG. 1 is a diagram illustrating an example of a distribution server according to the embodiment. FIG. 2 is a diagram illustrating a configuration example of a distribution server according to the embodiment. FIG. 3 is a diagram illustrating an example of a connection coefficient storage unit according to the embodiment. FIG. 4 is a diagram illustrating an example of a model generation image database according to the embodiment. FIG. 5 is a diagram illustrating an example of a user estimation model according to the embodiment. FIG. 6 is a diagram illustrating an example of an advertisement database according to the embodiment. FIG. 7 is a diagram illustrating an example of an effect using CNN. FIG. 8 is a diagram illustrating an example of an effect using attribute information. FIG. 9 is a flowchart illustrating an example of a generation process executed by the distribution server according to the embodiment. FIG. 10 is a flowchart illustrating an example of distribution processing executed by the distribution server according to the embodiment. FIG. 11 is a diagram illustrating another example of attribute information estimated by the distribution server according to the embodiment. FIG. 12 is a diagram illustrating an example of processing in which the distribution server according to the embodiment estimates attribute information based on an image selected from a moving image. FIG. 13 is a diagram illustrating replacement of the estimation method. FIG. 14 is a hardware configuration diagram illustrating an example of a computer that realizes the function of the distribution server.

  Hereinafter, a model generation device, an information distribution device, a model generation method, an information distribution method, a model generation program, and an information distribution program according to the present application (hereinafter referred to as “embodiment”) will be described with reference to the drawings. The details will be described. Note that the model generation device, the information distribution device, the model generation method, the information distribution method, the model generation program, and the information distribution program according to the present application are not limited by this embodiment. In the following embodiments, the same portions are denoted by the same reference numerals, and redundant description is omitted.

[1. Example of distribution server)
First, an example of a distribution server that operates as an example of a model generation device and an information distribution device in the embodiment will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of a distribution server according to the embodiment. In the following description, the distribution server 10 uses the generation process for generating the user estimation model for estimating the attribute information indicating the user's attribute and the like, and the user estimation model generated by the generation process. An example in which attribute information is estimated and a distribution process for distributing an advertisement based on the estimated attribute information will be described.

  Moreover, although the following description demonstrates the example which performs the production | generation process and a delivery process using the advertisement image which is an image which concerns on an advertisement, such an image is not only an advertisement but recruitment of volunteers, public advertisement, public There may be images relating to notifications, product images, product descriptions, and other arbitrary information. In other words, the image described below makes the user interested, information included in the image or information related to the image (for example, information on a web page on which the image is arranged or information on a landing page). Can be applied to any content such as characters, graphics, symbols, sounds, moving images, games, and the like.

  The distribution server 10 is an information processing device such as a web server that performs data communication with the terminal device 100 used by the user. For example, upon receiving an advertisement distribution request for requesting advertisement distribution from the terminal device 100, the distribution server 10 selects and selects an advertisement image to be distributed based on attribute information of a user who uses the terminal device 100. The advertisement image thus transmitted is transmitted to the terminal device 100. Here, the attribute information of the user is various information indicating the profile and identity of the user such as the user's sex, age, address, hobby, object of interest, browsing history of web pages and advertisements.

  The terminal device 100 is a smart device such as a smartphone or a tablet, and is a portable terminal device that can communicate with an arbitrary server device via a wireless communication network such as 3G (Generation) or LTE (Long Term Evolution). . The terminal device 100 performs various operations such as image selection by touching the touch panel with a finger or a dedicated pen.

  The terminal device 100 may be an information processing device such as a desktop PC (Personal Computer), a notebook PC, a netbook, or a server device. In such a case, the terminal device 100 may accept various operations via an information input device such as a mouse or a keyboard.

  Here, when displaying web content such as a web page by a user operation, the terminal device 100 transmits a web content delivery request to a web server that delivers web content. In such a case, when receiving the distribution of the web content, the terminal device 100 transmits a distribution request for the advertisement arranged in the web content to the distribution server 10.

  In such a case, the distribution server 10 distributes the advertisement image to the terminal device 100. And the terminal device 100 arrange | positions and displays the advertisement image received from the delivery server 10 in a web page. In addition, when an operation for selecting an advertisement image is performed, such as when the user taps the advertisement image or clicks the cursor with the advertisement image, the terminal device 100 displays the landing page corresponding to the advertisement image. In addition to displaying, a notification that the advertisement image has been selected is transmitted to the distribution server 10.

  In addition, when the distribution server 10 receives the notification that the advertisement image has been selected, the distribution server 10 performs a billing process for the advertiser. That is, the distribution server 10 performs click-billing advertisement distribution to the terminal device 100.

[2-1. Processing executed by distribution server 10]
Here, in the case where the distribution server 10 distributes the advertisement by the click charge type, the advertisement image to be distributed is selected based on the bid amount set by the advertiser. Specifically, the distribution server 10 calculates the expected return based on the bid amount set for each advertisement image and the click rate that is the probability that each advertisement image is clicked, and the calculated expected return is the highest. Distribute high advertising images with priority.

  Here, the click rate is a click rate estimated on the basis of the attribute information of the user to be distributed as an advertisement and the content of the advertisement image as the distribution target. For example, in the conventional technology, the value of the click rate is estimated according to the correspondence between the characteristics when the advertisement image to be distributed is abstracted and the attribute information of the user to be distributed. To give a more specific example, in the prior art, the advertisement image to be distributed is an image that a man is likely to select, such as a game screen, a vehicle, a product for men, etc., and the gender of the user to be distributed is male. In the case of, the click rate of the advertisement image to be distributed is increased.

  In another conventional technique, when an advertisement image to be distributed is similar to another advertisement image, the attribute information of the user who selected the other advertisement image is acquired, and the acquired attribute information of the user and The value of the click rate is estimated according to the relationship with the attribute information of the user who is the delivery destination. As a more specific example, in the related art, when the user to be distributed is a male, the click rate of the advertising image similar to the advertising image selected by many men is increased.

  However, a newly registered advertisement image or the like may not have an advertisement image log such as attribute information of a user who has selected the advertisement image or a configuration of the advertisement image. In such a case, according to the conventional technique, since the user who can select the advertisement image to be distributed cannot be estimated, the estimation accuracy of the click rate is deteriorated. Also, the estimation accuracy of the click rate is deteriorated even when a sufficient number of logs are not acquired.

  Also, when an advertisement distribution request is received from a new user, or when a B-cookie or the like is deleted and an ID (Identifier) for identifying the user is unknown, the user attribute information is It may be unknown. In such a case, the conventional technique cannot calculate the predicted click rate of the advertisement image to be distributed.

  In this way, in the conventional technology, when the image log does not exist or when the attribute information of the user is unknown, the click rate cannot be accurately estimated in a so-called cold start state. It cannot be calculated. As a result, in the related art, the advertisement image to be distributed cannot be appropriately selected in the cold start state.

  Therefore, the distribution server 10 calculates the feature amount of the advertisement image having the history selected by the user using a learning device that calculates the feature amount indicating the feature of the image from the image. Then, the distribution server 10 performs a generation process for generating a user estimation model for estimating the user attribute information from the image feature amount based on the calculated feature amount and the attribute information of the user who selected the advertisement image. Run.

  For example, the distribution server 10 uses a learning device to calculate a feature amount of an image having a history selected by the user, and based on the attribute information indicating the attribute of the user in the history and the calculated feature amount. Thus, a model for estimating the attribute information of the user to be estimated for attribute information is generated from the feature amount of the image selected by the user to be estimated for attribute information. Here, the user in the history and the user to be estimated for the attribute target may be the same user or different users.

  In addition, when a user whose attribute information is unknown selects an advertisement image, the distribution server 10 calculates a feature amount of the advertisement image using a learning device. Further, the distribution server 10 estimates user attribute information from the calculated feature amount using a user estimation model. Then, the distribution server 10 calculates a click rate at which a user selects an advertisement image using the estimated attribute information, selects an advertisement image to be distributed based on the calculated click rate, and performs distribution. Execute the distribution process.

[2-2. (Learning device)
Here, a learning device used when the distribution server 10 calculates the feature amount of the advertisement image will be described. The learning device is, for example, a learning device in which a plurality of nodes that output calculation results for input data are connected in multiple layers, and learns image features abstracted by unsupervised learning. More specifically, the learning device is a neural network in which layers having a plurality of nodes are connected in multiple stages, and is a DNN (Deep Neural Network) realized by a so-called deep learning technique. Image features include not only the specific features that appear in the image, such as the presence / absence of characters in the image, color, composition, etc., but also what the object is being imaged and what kind of user the image is It is a concept that also includes the characteristics of an abstracted (meta-) image such as the atmosphere of the image.

  When such a deep learning technique is used, the learning device is generated by the following learning method, for example. For example, in the learning device, the connection coefficient between the nodes is initialized, and images having various characteristics are input. Then, backpropagation is performed to correct the connection coefficient so that an error between the output of the learning device and the input image is reduced. By repeating such processing, the learning device can output an output that can better reproduce the input image, that is, a feature of the input image.

  Note that the learning method of the learning device is not limited to the method described above, and any known technique can be applied. In addition, an advertisement image to be distributed may be used as an image used when learning by the learning device, or a general image data set may be used. In addition, any method can be applied to the image input method to the learning device, the format of data output from the learning device, the content of features that are explicitly learned by the learning device, and the like. That is, the distribution server 10 can use any learning device as long as the distribution server 10 can calculate the feature amount indicating the feature abstracted from the image.

  In the following description, it is assumed that the distribution server 10 calculates the feature amount of the advertisement image using a learning device that repeats convolution and pooling of the local region of the input image, so-called CNN (Convolutional Neural Network). In addition to the function of extracting and outputting features from an image, the CNN has output invariance with respect to positional variations such as characters and imaging objects included in the image. For this reason, the distribution server 10 can calculate the abstract feature of the advertisement image with high accuracy.

[2-3. Generation process)
Next, the generation process executed by the distribution server 10 will be described with reference to FIG. First, the distribution server 10 executes a generation process for generating a user estimation model prior to the distribution process (step S10). Specifically, the distribution server 10 inputs advertisement images C10 to C13 (hereinafter may be referred to as selected images C10 to C13) selected by a user whose attribute information can be specified to the learning device (step S11), feature amounts of the selected images C10 to C13 are calculated (step S12). Then, the distribution server 10 uses the calculated feature amount and the attribute information of the users U10 to U13 who selected the selected images C10 to C13, and the attribute of the user who selected the image from the feature amount of the image. A user estimation model for estimating information is generated (step S13).

  For example, the distribution server 10 collects the distributed advertisement image and the attribute information of the user as the distribution destination from the distribution log of the advertisement image, the attribute information log of the user as the distribution destination, and the like. In the example illustrated in FIG. 1, a user U10 whose attribute information is “male” selects a PC advertisement image C10, and a user U11 whose attribute information is “female” selects a shoe advertisement image C11. In the example shown in FIG. 1, the user U12 whose attribute information is “30's” selects the advertisement image C12 of the accessory, and the user U13 whose attribute information is “10's” selects the advertisement image C13 of the game machine. doing.

  Here, a user having similar attribute information is likely to select an advertisement image having similar characteristics. For example, male users tend to select game machines and advertisement images related to men's fashion, and female users tend to select accessories and advertisement images related to women's fashion. For this reason, even if the user is a user other than the user U10, as in the case of the user U10, when the attribute information is “male”, the other advertisement images having the same characteristics as the advertisement image C10 and the advertisement image C10 Is likely to choose.

  Therefore, the distribution server 10 inputs the advertisement image C10 to the learning device, thereby calculating a feature amount A indicating an abstract feature of the advertisement image C10. More specifically, the distribution server 10 acquires, as the feature amount A of the advertisement image C10, an output when the advertisement image C10 is input to the learning device, that is, a multidimensional amount indicating an abstract feature of the advertisement image C10. . Then, the distribution server 10 uses the same attribute information “male” as the user U10 who has selected the advertisement image C10 from the feature amount whose difference from the feature amount A of the advertisement image C10 is smaller than a predetermined threshold as the user estimation model. Generate a user estimation model to estimate.

  Similarly, the distribution server 10 uses the same attribute information “female” as the user U11 who selected the advertisement image C11 from the feature amount whose difference from the feature amount B of the advertisement image C11 is smaller than a predetermined threshold as the user estimation model. Generate a user estimation model that estimates. Further, the distribution server 10 uses the same attribute information “30's” as the user U12 who selected the advertisement image C12 from the feature amount whose difference from the feature amount C of the advertisement image C12 is smaller than a predetermined threshold as the user estimation model. Generate a user estimation model that estimates. Further, the distribution server 10 uses the same attribute information “10th generation” as the user U13 who selected the advertisement image C13 from the feature amount whose difference from the feature amount D of the advertisement image C13 is smaller than a predetermined threshold as the user estimation model. Generate a user estimation model that estimates.

  For example, in the example illustrated in FIG. 1, the distribution server 10 associates the feature amount A of the advertisement image C10 with the attribute information “male”, associates the feature amount B of the advertisement image C11 with the attribute information “female”, The user estimation model 17 in which the feature amount C of the advertisement image C12 is associated with the attribute information “30's” and the feature amount D of the advertisement image C13 is associated with the attribute information “10's” is generated.

[2-4. (Delivery process)
Next, the distribution process executed by the distribution server 10 will be described with reference to FIG. First, when the advertisement image C14 without a log is selected by the user U20 whose attribute information is unknown (step S20), the distribution server 10 acquires the advertisement image C14 selected by the user U20 (step S21). . In addition, the distribution server 10 inputs the acquired advertisement image C14 to the learning device (step S22), and calculates a feature amount E indicating an abstract feature of the advertisement image C14 (step S23).

  And the delivery server 10 estimates the attribute information of the user U20 using the feature-value E and the user estimation model 17 (step S24). For example, the distribution server 10 determines that the difference between the feature amount B and the feature amount E and the difference between the feature amount C and the feature amount E are larger than a predetermined threshold, and the difference between the feature amount A and the feature amount E and the feature amount. If the difference between D and feature amount E is smaller than a predetermined threshold, “male” and “teenage” are estimated as attribute information of the user U20 (step S25).

  Further, the distribution server 10 calculates a predicted click through rate (pCTR) of the advertisement image C14 from the feature amount E (step S26). For example, the distribution server 10 inputs a feature quantity “x” represented by a multidimensional quantity to a predetermined linear function f (x) defined in advance, and the value of 1 / (1 + exp (−f (x))) Is calculated as pCTR. Then, the distribution server 10 estimates the click rate of the advertisement image to be distributed using the estimated attribute information and the calculated predicted click rate (step S27).

  Thereafter, the distribution server 10 calculates the expected return using the estimated click rate of each advertisement image and the bid amount of each advertisement image, and the user U20 selects the advertisement image with the highest expected return value. Distribute to the terminal device 100 to be used (step S28).

  As described above, the distribution server 10 calculates the feature amount of the selected images C10 to C13 selected by the users U10 to U13 using the learning device that calculates the feature amount of the image, and the calculated feature amount and the user U10. Based on the attribute information of U13, a user estimation model for estimating the attribute information of the user U20 who selected the image C14 from the feature amount of the advertisement image C14 is generated.

  Further, the distribution server 10 estimates the attribute information of the user U20 who selected the advertisement image C14 using the user estimation model, and the probability that the user U20 selects the advertisement image using the estimated attribute information. Calculate clickthrough rate. Then, the distribution server 10 selects an advertisement image to be distributed based on the calculated click rate.

  For this reason, since the distribution server 10 can estimate the attribute information of the user even in the cold start state, it can appropriately select the advertisement image to be distributed.

  Note that the user estimation model 17 illustrated in FIG. 1 is merely an example, and the distribution server 10 can generate a model of any other format. For example, when the distribution server 10 inputs the value of each dimension of a feature quantity that is a multidimensional quantity as an independent variable, the distribution server 10 generates a function that outputs as a solution how much each attribute information is estimated as a model. May be.

  Hereinafter, an example of a functional configuration of the distribution server 10 that realizes the above selection process will be described.

[3. Distribution server configuration)
Next, the configuration of the distribution server 10 according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram illustrating a configuration example of a distribution server according to the embodiment. As shown in FIG. 2, the distribution server 10 is connected to the terminal device 100 and the web server 200 via a network N so as to be communicable by wire or wirelessly. The distribution server 10 may be connected to a plurality of terminal devices 100 and a plurality of web servers 200.

  The terminal device 100 acquires a web page from the web server 200 according to an operation by the user, and displays the acquired web page. Further, when the advertisement image acquisition command is included in the web page, the terminal device 100 transmits the user attribute information together with the advertisement image distribution request to the distribution server 10 to acquire the advertisement image. Then, the terminal device 100 displays the acquired advertisement image. Further, when the user selects an advertisement image, the terminal device 100 transmits information indicating the selected advertisement image and user attribute information to the distribution server 10.

  On the other hand, the distribution server 10 includes a communication unit 11, a storage unit 12, and a control unit 13, as illustrated in FIG. Note that the internal configuration of the distribution server 10 is not limited to the configuration illustrated in FIG. 2, and may be another configuration as long as the selection process described above is performed.

  The communication unit 11 is realized by, for example, a NIC (Network Interface Card). The communication unit 11 is connected to the network N by wire or wirelessly and transmits / receives information to / from the terminal device 100.

  The storage unit 12 is realized by, for example, a semiconductor memory device such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. The storage unit 12 stores a learning image database 14, a connection coefficient storage unit 15, a model generation image database 16, a user estimation model 17, and an advertisement database 18.

  The control unit 13 is realized by, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or the like executing various programs stored in the storage device inside the distribution server 10 using the RAM as a work area. The The control unit 13 is realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

  Further, the control unit 13 executes the learning unit 19, the generation unit 20, the distribution by executing the learning program, the model generation program, and the distribution program stored in the storage device inside the distribution server 10 using the RAM as a work area. It operates as the unit 30. Here, the generation unit 20 includes a feature amount calculation unit 21, a user information acquisition unit 22, and a model generation unit 23. The distribution unit 30 includes a selection image acquisition unit 31, a user information estimation unit 32, a click rate calculation unit 33, and an advertisement selection unit 34.

[4-1. (Image database for learning)
Next, an example of information stored in the storage unit 12 will be described. In the learning image database 14, a learning image used when the learning device learns the feature amount of the abstracted image is registered. For example, in the learning image database 14, an image data set used when learning DNN or CNN is registered.

[4-2. Connection coefficient storage unit)
In the connection coefficient storage unit 15, a connection coefficient between the nodes included in the learning device is registered. For example, FIG. 3 is a diagram illustrating an example of the connection coefficient storage unit according to the embodiment. As illustrated in FIG. 3, the connection coefficient storage unit 15 includes items such as “neuron ID”, “connection destination neuron ID”, and “connection coefficient”. “Neuron ID” indicates identification information for identifying a neuron that is a node included in the DNN or CNN. “Destination neuron ID” indicates identification information for identifying a neuron coupled to a neuron corresponding to “neuron ID”. “Connection coefficient” indicates a connection coefficient between neurons.

For example, the neuron corresponding to the neuron ID “N ₁₁ ” is connected to the neuron corresponding to the neuron ID “N ₂₁ ” with the connection coefficient “W _A ”. For this reason, to the neuron corresponding to the neuron ID “N ₂₁ ”, a value obtained by integrating the connection coefficient “W _A ” to the output of the neuron ID “N ₁₁ ” is transmitted as an input.

[4-3. (Image database for model generation)
Returning to FIG. 2, the image used for generating the user estimation model 17 is registered in the model generation image database 16. For example, FIG. 4 is a diagram illustrating an example of a model generation image database according to the embodiment. As shown in FIG. 4, in the model generation image database 16, an advertisement image selected by a user who can acquire attribute information and attribute information of the user who selected the advertisement image are registered in association with each other. The

  For example, in the example illustrated in FIG. 4, the model generation image database 16 includes an advertisement image C30 selected by a user whose attribute information is “male” or “teenage”, and attribute information “male” of the user. "10" is registered in association with each other. Similarly, in the model generation image database 16, the advertisement image C31 and the attribute information “male” and “20s” are registered in association with each other, and the advertisement image C32 and the attribute information “female” and “teens” are registered. Are registered in association with each other. In the model generation image database 16, the advertisement image C33 and attribute information “female” and “20's” are registered in association with each other, and the advertisement image C34 and attribute information “10's” are associated with each other. The advertisement image C35 and the attribute information “20's” are registered in association with each other.

[4-4. User estimation model)
Returning to FIG. 2, the user estimation model 17 is a model generated from the feature amount of the image selected by the user who can acquire the attribute information and the attribute information of the user, and is based on the feature amount of the image. This is a model for estimating attribute information of a user who has selected such an image. For example, FIG. 5 is a diagram illustrating an example of a user estimation model according to the embodiment. In the example illustrated in FIG. 5, the user estimation model 17 registers the feature amount of an image and attribute information estimated for a user who has selected an image having a feature amount similar to the feature amount. .

  For example, in the example illustrated in FIG. 4, in the user estimation model 17, the feature amount E and attribute information “male” and “teenage” are registered in association with each other. That is, the user estimation model 17 indicates that the attribute information of the user who has selected the feature amount image that matches or is similar to the feature amount E is “male” or “teenage”. Similarly, the user estimation model 17 includes a feature amount F and attribute information “male” and “20s”, a feature amount G and attribute information “female” and “teen”, a feature amount H and “female”, “20 ”, Feature amount I and attribute information“ 10 ”, and feature amount J and attribute information“ 20 ”are registered in association with each other.

[4-5. Advertising database)
Returning to FIG. 2, advertisement images to be distributed are registered in the advertisement database 18. For example, FIG. 6 is a diagram illustrating an example of an advertisement database according to the embodiment. In the example shown in FIG. 6, items such as “advertiser ID”, “advertisement content”, “bid price”, “CTR”, and “attribute information” are registered in the advertisement database 18 in association with each other.

  “Advertiser ID” indicates identification information for identifying an advertiser. “Advertising content” indicates advertising content submitted by the advertiser, for example, an advertising image. FIG. 4 shows an example in which conceptual information such as “C11” is stored in “advertising content”, but actually, a still image, a moving image, text data, a URL, or a file indicating the storage location thereof. Stores path names and so on.

  “Bid price” indicates an advertising fee designated when an advertiser submits advertising content. For example, the “bid price” corresponds to a unit price paid from the advertiser to the advertisement distributor (for example, the administrator of the distribution server 10) when the advertisement content is clicked once by the user. Further, for example, the “bid price” may be a unit price paid from the advertiser to the advertisement distributor when the advertisement content is displayed.

  “CTR” indicates a value obtained by dividing the number of times the advertising content is clicked by the number of times the advertising content is displayed. The CTR of the advertising content that has not been distributed to the terminal device 100 includes a predetermined fixed value, an average value of CTR for all advertising content, and the same advertising category (for example, car, travel). The average value of CTR and the like for all the advertising contents belonging to is stored. “Attribute information” indicates attribute information of a user who is a distribution target of advertisement content.

  For example, FIG. 4 shows an example in which the advertiser identified by the advertiser ID “B10” designates the bid price “100” and submits the advertising content “C11”. FIG. 4 shows an example in which the CTR of the advertising content “C11” is “0.02” and the attribute information of the user to be distributed is “male”.

  Note that “CTR” may store a predicted CTR predicted from a CTR prediction model or the like. Such a predicted CTR is predicted based on, for example, the type of advertising content, the type of web page on which the advertising content is displayed, and the like.

[4-6. Example of processing executed by control unit]
Returning to FIG. 3, the contents of the processing executed by the learning unit 19, the generation unit 20, and the distribution unit 30 of the control unit 13 will be described.

[4-6-1. Example of processing executed by learning unit]
The learning unit 19 performs learning of the learning device. For example, the learning unit 19 is a learning device in which neurons are connected in multiple stages, and learns a learning device that learns the features of an image abstracted by unsupervised learning, that is, DNN. Further, the learning unit 19 performs learning of a learning device that further learns the universality of output with respect to image variation, that is, CNN, by repeating local region convolution and pooling as a learning device.

  For example, the learning unit 19 constructs a learning device in which a predetermined connection relationship is set between neurons. In addition, the learning unit 19 uses the learning image registered in the learning image database 14 to cause the learning device to learn the characteristics of the abstracted image. More specifically, the learning unit 19 uses a known method such as backpropagation to correct the connection coefficient so that the abstracted feature is extracted from the learning image with high accuracy. Then, the learning unit 19 registers the connection relationship between the neurons and the connection coefficient in the connection coefficient storage unit 15.

  As a specific learning method executed by the learning unit 19, a known learning method used when learning DNN or CNN can be employed.

[4-6-2. Example of processing executed by generation unit]
The generation unit 20 executes the generation process by the processes executed by the feature amount calculation unit 21, the user information acquisition unit 22, and the model generation unit 23. Specifically, the feature amount calculation unit 21 calculates a feature amount of an advertisement image having a history selected by a user who can acquire attribute information, using a learning device that calculates a feature amount from an image.

  For example, the feature amount calculation unit 21 acquires a neuron ID and a connection coefficient from the connection coefficient storage unit 15 and configures a learning device. In addition, the feature amount calculation unit 21 acquires the advertisement image C30 from the model generation image database 16. Then, the feature amount calculation unit 21 inputs the advertisement image C30 to the learning device, and acquires the output of the learning device as the feature amount E.

  The user information acquisition unit 22 acquires attribute information of the user who has selected the advertisement image whose feature value has been calculated. For example, when the feature amount calculation unit 21 calculates the feature amount E of the advertisement image C30, the user information acquisition unit 22 generates a model of attribute information “male” and “teenage” associated with the advertisement image C30. Obtained from the image database 16.

  The model generation unit 23 generates a user estimation model based on the feature amount calculated by the feature amount calculation unit 21 and the attribute information acquired by the user information acquisition unit 22. For example, the model generation unit 23 registers the feature amount E of the advertisement image C30 calculated by the feature amount calculation unit 21 and the attribute information acquired by the user information acquisition unit 22 in the user estimation model 17 in association with each other.

  Note that the model generation unit 23 may generate the user estimation model 17 by a more complicated method. For example, in the model generation unit 23, the feature amount calculation unit 21 calculates the feature amount E and the feature amount I, and the user information acquisition unit 22 sets the attribute information “male”, “teenage”, and attribute information “teenage”. When acquired, the difference between the feature quantity E and the feature quantity I may be added to the user estimation model 17 as a feature quantity corresponding to the attribute information “male”. That is, the model generation unit 23 may compare a plurality of feature amounts and generate the user estimation model 17 based on the difference or the common part. Further, the model generation unit 23 may generate, as a model, a function that outputs a value corresponding to the attribute information when the feature amount is input as a variable.

[4-6-3. Example of processing executed by distribution unit]
The distribution unit 30 performs the distribution process by the processes executed by the selected image acquisition unit 31, the user information estimation unit 32, the click rate calculation unit 33, and the advertisement selection unit 34. Specifically, the selected image acquisition unit 31 acquires an advertisement image selected by a user whose attribute information is unknown.

  For example, when the notification of the advertisement image selected by the user is received, the selected image acquisition unit 31 determines whether the attribute information of the user can be specified. Specifically, the selected image acquisition unit 31 is a case where there is no B cookie or there is no user ID, for example, when the user attribute information cannot be specified, or a management system that manages the user attribute information. If the user attribute information is not registered, it is determined that the user attribute information is unknown.

  When the selected image acquisition unit 31 determines that the user attribute information is unknown, the selected image acquisition unit 31 specifies the notified advertisement image from the advertisement database 18 and outputs the specified advertisement image to the user information estimation unit 32. To do. When the user's attribute information can be identified, the selected image acquisition unit 31 associates the notified advertisement image with the identified attribute information and registers them in the model generation image database 16.

  The user information estimation unit 32 uses the user estimation model 17 to estimate the attribute information of the user from the advertisement image selected by the user whose attribute information is unknown. For example, the user information estimation unit 32 uses the information registered in the connection coefficient storage unit 15 to construct a learning device that calculates a feature value indicating the feature of the abstracted image.

  In addition, when the user information estimation unit 32 receives the advertisement image from the selection image acquisition unit 31, the user information estimation unit 32 inputs the received advertisement image to the learning device, and calculates the feature amount of the advertisement image. Then, the user information estimation unit 32 uses the user estimation model 17 to estimate attribute information from the calculated feature amount. For example, when the user estimation model 17 shown in FIG. 5 is registered and the difference between the feature quantity and the feature quantity E is smaller than a predetermined threshold, the user information estimation unit 32 selects the advertisement image used. Attribute information “male” and “teen” is estimated. Then, the user information estimation unit 32 outputs the calculated feature amount and the estimated attribute information “male” and “teenage” to the click rate calculation unit 33.

  The click rate calculation unit 33 uses the attribute information estimated by the user information estimation unit 32 to calculate a click rate at which the user selects an advertisement image to be distributed. For example, the click rate calculation unit 33 calculates a predicted click rate that is a probability that the user selects an advertisement image selected by the user who estimated the attribute information from the feature amount calculated by the user information estimation unit 32. To do. Then, the click rate calculation unit 33 uses the predetermined model with the calculated predicted click rate and the attribute information estimated by the user information estimation unit 32 as parameters, The click rate selected by the user who estimated the attribute information is calculated.

  For example, when the estimated attribute information is “male”, the click rate calculation unit 33 increases the click rate of the advertisement image associated with the attribute information “male” and associates the attribute information “female” with the attribute information “male”. A process such as reducing the click rate of the advertisement image is executed. A known click rate calculation model can be applied to the process in which the click rate calculation unit 33 calculates the click rate of each advertisement image.

  The advertisement selection unit 34 selects an advertisement image to be distributed to the user based on the click rate calculated by the click rate calculation unit 33. For example, the advertisement selection unit 34 accumulates the values of the click rate calculated by the click rate calculation unit 33 and the bid amount of each advertisement image, and selects the advertisement image having the highest accumulated value. And the advertisement selection part 34 distributes the selected advertisement image, when the distribution request is received from the terminal device 100 used by the user whose attribute information is estimated.

  In addition, when the distribution request is received from a user whose attribute information is not unknown, the click rate calculation unit 33 calculates the click rate using the user attribute information. Then, the advertisement selection unit 34 selects an advertisement image using the click rate calculated by the click rate calculation unit 33 and distributes the selected advertisement image.

[5. About calculation of features)
Here, the distribution server 10 described above can appropriately calculate the feature amount of the abstracted image by calculating the feature amount of the advertisement image using the CNN. For example, FIG. 7 is a diagram illustrating an example of an effect using CNN. In FIG. 7, mean absolute error (MAE) obtained by ridge regression to the actual click rate for each advertisement image is plotted for each method, using the feature values extracted by various methods as the features of the model.

  In the example shown in FIG. 7, Gist, SIFT (Scale-Invariant Feature Transform), C-SIFT (Colored-SIFT), Opponent SIFT, RGB-SIFT (Red Green Blue-SIFT), and CNN are used as feature amount extraction methods. Adopted. In addition, as a comparison target with an extraction method that estimates a feature without using an ID feature such as an advertiser ID, an advertisement ID, and an advertisement image ID, a Baseline result that specifies a feature amount using only the ID feature is plotted.

  Here, Gist is a global image feature quantity used when describing scene information of an image, and is a feature quantity having a 960-dimensional number. SIFT, C-SIFT, Opponent SIFT, and RGB-SIFT are methods for calculating local feature amounts. In the example shown in FIG. 7, a 2048-dimensional Fisher vector expressing a distribution of local feature amounts calculated by SIFT, C-SIFT, Opponent SIFT, and RGB-SIFT as a mixed Gaussian distribution is based on a Bag-of-Visual-Words base. It was adopted as the band feature amount.

  In the example shown in FIG. 7, Caffe's Imagenet plain model is adopted as CNN, and 4096-dimensional information output from the final layer of CNN is used as a feature amount. In the example illustrated in FIG. 7, a 20-bit feature amount using Feature hashing is used as Baseline from ID features such as an advertiser ID, an advertisement ID, and an advertisement image ID.

  When comparing the feature values obtained by the above-described methods, it is important whether or not the imaging target can be expressed as a part of the feature value, not information on the local shape in the image. , The average absolute error of the feature quantity extracted by CNN is the lowest value. However, since various ID features cannot be obtained in the cold start state, it is difficult to use Baseline. For this reason, the distribution server 10 can appropriately calculate the feature amount of the image even in the cold start state by using the CNN.

[6. (About estimation of attribute information)
Further, the distribution server 10 described above can estimate the attribute information of the user who has selected the advertisement image, and can calculate the click rate with high accuracy by using the estimated attribute information. For example, FIG. 8 is a diagram illustrating an example of an effect using attribute information. In FIG. 8, logistic regression is performed with “1” for the click and “−1” for only the impression, and the correspondence between the AUC (area under the curve) and the overlap rate is plotted for each method of calculating the click rate. Further, in the example shown in FIG. 8, the click rate is predicted using Baseline, User_add that adds attribute information to Baseline to predict the click rate, and CNN_add that predicts the click rate by adding a feature amount by CNN to User_add. Adopted.

  In the example shown in FIG. 8, it can be understood that the User_add method for predicting the click rate in consideration of the attribute information generally has better prediction performance than the method using Baseline. Therefore, the distribution server 10 can calculate the click rate with high accuracy by estimating the attribute information of the user at the cold start and using the estimated attribute information.

[7. Processing flow of distribution server 10]
Next, an example of the flow of generation processing and distribution processing executed by the distribution server 10 will be described with reference to FIGS. 9 and 10. First, an example of the generation process executed by the distribution server 10 will be described with reference to FIG. FIG. 9 is a flowchart illustrating an example of a generation process executed by the distribution server according to the embodiment.

  As shown in FIG. 9, the distribution server 10 inputs a model generation image, which is an advertisement image selected by a user who can acquire attribute information, to the CNN, and acquires the output of the final layer as a feature amount (step S101). ). Subsequently, the distribution server 10 acquires attribute information of the user who selected the model generation image (step S102). Then, the distribution server 10 generates a user estimation model that estimates the attribute information from the feature amount of the image based on the feature amount and the attribute information (step S103), and ends the process.

  The distribution server 10 can execute the generation process shown in FIG. 9 at an arbitrary timing. For example, the distribution server 10 may be executed every time a model generation image is registered in the model generation image database 16 or may be executed in accordance with a predetermined time interval or a user instruction. That is, the distribution server 10 may execute the generation process at an arbitrary timing as long as the model is generated prior to the process of estimating the user attribute information.

  Next, an example of distribution processing executed by the distribution server 10 will be described with reference to FIG. FIG. 10 is a flowchart illustrating an example of distribution processing executed by the distribution server according to the embodiment. First, the distribution server 10 determines whether or not the advertisement image information selected by the user has been received (step S201). And the delivery server 10 performs step S201 again, when the information of the advertisement image which the user selected is not received (step S201: No).

  On the other hand, when the distribution server 10 receives the information of the advertisement image selected by the user (step S201: Yes), the distribution server 10 inputs the selected advertisement image into the CNN and acquires the feature amount (step S202). And the delivery server 10 estimates attribute information from the acquired feature-value using the user estimation model 17 (step S203). Further, the distribution server 10 calculates a predicted click rate using the feature amount (step S204), and estimates the click rate of the advertisement image using the calculated predicted click rate and the estimated attribute information ( Step S205). Further, the distribution server 10 calculates the expected profit based on the estimated click rate (step S206), and distributes the advertisement image with the maximum expected profit when the distribution request is received (step S207). Exit.

  The distribution server 10 may perform the distribution process shown in FIG. 10 prior to the distribution timing of the advertisement image. For example, the distribution server 10 takes a log of a user who has selected an advertisement image, and when receiving a distribution request from a user whose attribute information is unknown, identifies and identifies the advertisement image selected by the user from the log The attribute information of the user may be estimated from the feature amount of the advertisement image.

  In addition, when the distribution server 10 receives a distribution request from a user whose attribute information is unknown and does not have a log for selecting an advertisement image, the distribution server 10 distributes the advertisement image selected by an arbitrary method. In addition, when such an advertisement image is selected, the distribution server 10 estimates the attribute information of the user from the advertisement image in advance. Then, when the distribution server 10 receives a new distribution request from the user, the distribution server 10 may select an advertisement image using attribute information estimated in advance, and distribute the selected advertisement image.

[8. (Modification)
The distribution server 10 according to the above-described embodiment may be implemented in various different forms other than the above-described embodiment. Therefore, in the following, another embodiment of the distribution server 10 will be described. Of the processes executed by the distribution server 10 described below, the process corresponding to the generation process is executed by the generation unit 20, and the process corresponding to the distribution process is executed by the distribution unit 30.

[8-1. Inferred attribute information variation)
For example, when estimating the attribute information of the user, the distribution server 10 calculates the matching rate between the feature amount registered in the user estimation model 17 and the feature amount of the selected image, and considers the calculated matching rate. The estimated attribute information may be estimated. For example, FIG. 11 is a diagram illustrating another example of attribute information estimated by the distribution server according to the embodiment. In the example illustrated in FIG. 11, a multidimensional vector having a value of each dimension “0” or “1” is described as the feature amount of each of the images C <b> 10 to C <b> 13.

  First, as shown in FIG. 11A, the distribution server 10 calculates the feature amounts of the advertisement images C10 to C13 selected by the user using a learning device. Moreover, the delivery server 10 specifies the attribute information of the user who selected each advertisement image C10-C13, as shown to (B) in FIG. And the delivery server 10 produces | generates the user estimation model 17 which matched the feature-value and attribute information of each advertisement image C10, as shown to (C) in FIG.

  Subsequently, the distribution server 10 acquires the advertisement image C14 selected by the user who cannot acquire attribute information. In such a case, the distribution server 10 calculates the feature amount of the advertisement image C14 using a learning device as shown in (D) of FIG. Further, the distribution server 10 calculates the similarity between the calculated feature quantity and the feature quantity registered in the user estimation model 17, and estimates the click rate using the calculated similarity as a value indicating the certainty of the attribute information. I do. As a specific example, the distribution server 10 compares the calculated feature quantity with the feature quantity registered in the user estimation model 17 for each dimension, and calculates the number of dimensions with the same value as the feature quantity. The value divided by the number of dimensions is taken as the similarity.

  For example, as illustrated in (E) of FIG. 11, when the similarity between the feature quantity of the advertisement image C10 and the feature quantity of the advertisement image C14 is “0.6”, the distribution server 10 The certainty of the corresponding attribute information “male” is estimated to be “60%”. Similarly, when the similarity between the feature amount of the advertisement image C11 and the feature amount of the advertisement image C14 is “0.4”, the distribution server 10 determines the certainty of the attribute information “female” corresponding to the advertisement image C11. Is estimated to be “40 percent”.

  Further, when the similarity between the feature amount of the advertisement image C12 and the feature amount of the advertisement image C14 is “0.1”, the distribution server 10 confirms the attribute information “30's” corresponding to the advertisement image C12. Is estimated to be “10 percent”. Similarly, when the similarity between the feature amount of the advertisement image C13 and the feature amount of the advertisement image C14 is “0.7”, the distribution server 10 confirms the attribute information “10th generation” corresponding to the advertisement image C13. Is estimated to be “70 percent”.

  When such an estimation is performed, the distribution server 10 calculates the click rate on the assumption that the user is “male” with a probability of “60%” and “female” with a probability of “40%”. . For example, the distribution server 10 adds “0.6” to the parameter to be considered when the estimated attribute information is “male” among the parameters of the model used when calculating the click rate. “0.4” is added to the parameter considered when the attribute information obtained is “female”.

  As described above, the distribution server 10 can estimate the attribute of the user in more detail by calculating the attribute information in consideration of the matching rate of the feature amount, so that the click rate can be calculated more appropriately. it can. For example, the distribution server 10 can calculate the click rate based on the prediction that the advertisement image for men is selected with a probability of 70% and the advertisement image for women is selected with a probability of 30%.

[8-2. Regarding the learning image used by the distribution server 10]
In the example described above, the distribution server 10 causes the learner to learn the characteristics of the abstracted image using the image registered in the learning image database 14. However, the embodiment is not limited to this. For example, the distribution server 10 may acquire a learning device in which arbitrary learning has been performed by an external server or the like without performing learning processing, and execute generation processing and distribution processing using the acquired learning device. The distribution server 10 may use different learning devices for the generation process and the distribution process. For example, the distribution server 10 may use a learning device optimized for the generation process and a learning device optimized for the distribution process.

  In addition, when learning processing is performed, the number of learning images and the number of learnings can be reduced or the features can be learned with high accuracy by inputting an image in which the features that the learner wants to learn appear. . Therefore, the distribution server 10 may employ an advertisement image registered in the advertisement database 18 as a learning image.

  In addition, the distribution server 10 may perform learning of the learning device by adopting, as a learning image, an image in which a ratio selected by a user having predetermined attribute information is larger than a predetermined threshold. For example, the distribution server 10 may employ, as the learning image, an advertisement image in which the ratio selected by the user whose attribute information is “male” is equal to or greater than the threshold “80%”. By performing such processing, the distribution server 10 can cause the learning device to learn the features with high accuracy, and thus can improve the prediction accuracy of the attribute information.

  Further, the distribution server 10 may employ images having different features as learning images in order to accurately calculate the feature amounts of various advertisement images.

[8-3. Click rate calculation)
In the example described above, the distribution server 10 estimates the predicted click rate and attribute information from the feature amount, and calculates the click rate of each advertisement image using the estimated predicted click rate and attribute information. However, the embodiment is not limited to this. For example, the distribution server 10 may estimate the attribute information from the feature amount without calculating the predicted click rate, and calculate the click rate of each advertisement image based on the estimated attribute information.

  Here, as shown in FIG. 8, in the cold start state, the estimation accuracy of the click rate is greatly increased by estimating the click rate using the attribute information, rather than estimating the click rate using the feature amount itself. Can be improved. For this reason, the distribution server 10 estimates the click rate based on the attribute information without using the predicted click rate, thereby reducing the calculation resources associated with the distribution process while preventing the prediction accuracy of the click rate from deteriorating. Can do.

  Further, the distribution server 10 may calculate the predicted click rate of the advertisement image using the actual click rate of the advertisement image selected by the user and the feature amount of the advertisement image. Further, the distribution server 10 may predict the click rate of each advertisement image based on the actual click rate of each advertisement image, the estimated attribute information, and the predicted click rate.

[8-4. (Advertising image used for attribute information estimation)
For example, it is predicted that the user will more actively select the advertisement image as compared to an image other than an advertisement placed on a web page or the like. Therefore, the distribution server 10 estimates the attribute information of the user with high accuracy by estimating the attribute information of the user using the feature amount of the advertisement image selected by the user. However, the embodiment is not limited to this. For example, the distribution server 10 may estimate the attribute information of the user using the feature amount of the image other than the advertisement image selected by the user. Further, the distribution server 10 may estimate the attribute information using the feature amount of the advertisement image selected by the user and the feature amount of the image other than the advertisement image selected by the user. In such a case, the distribution server 10 may perform weighting that prioritizes the feature amount of the advertisement image.

[8-5. Estimation of attribute information using moving images)
Further, the distribution server 10 may estimate the attribute information of the user using a moving image browsed by the user. For example, when the user pauses the video being played back or when the user moves the cursor to the video being played, the distribution server 10 may display a frame image that the user is expected to be interested in. Acquire as the selected image. Then, the distribution server 10 may estimate the attribute information of the user from the acquired feature amount of the image.

  For example, FIG. 12 is a diagram illustrating an example of a process in which the distribution server according to the embodiment estimates attribute information based on an image selected from a moving image. For example, in the example illustrated in FIG. 12, the terminal device 100 reproduces the moving image C20 and displays a control bar T10 that controls reproduction of the moving image C20. The control bar T10 displays advertisement insertion timings T11 and T12. At the timings T11 and T12, a predetermined advertisement moving image is reproduced.

  Here, when the cursor C21 pauses the reproduction of the moving image C20 by the user's operation, the terminal device 100 transmits the frame displayed at the time of the suspension to the distribution server 10 as the selected image C22. In such a case, the distribution server 10 calculates the feature amount of the image C22 and estimates the attribute information of the user from the calculated feature amount, as shown in FIG. For example, the distribution server 10 uses the feature quantity of the image C22 and the user estimation model 17 generated in advance, as shown in FIG. presume.

  Further, the distribution server 10 switches advertisements inserted at the timings T11 and T12 based on the estimated attribute information. For example, in the example shown in FIG. 12C, the advertisement A for women in their 20s is selected as the advertisement inserted at timing T11, and the advertisement for women in their 30s is selected as the advertisement inserted at timing T12. Ad B was selected. However, the estimated user attribute information is “teens” and “male”. Therefore, the distribution server 10 switches the advertisement inserted at the timings T11 and T12 based on the estimated attribute information, as shown in (D) in FIG.

  For example, as shown in FIG. 12E, the distribution server 10 newly selects an advertisement C directed to a male teenager as an advertisement to be inserted at timing T11, and as an advertisement to be inserted at timing T12. A new advertisement D for teenagers is selected. As a result, the distribution server 10 can distribute an advertisement presumed that the user is interested even when the attribute information of the user is unknown.

  The distribution server 10 may estimate the attribute information of the user in consideration of the position where the image selected by the user is included among the moving images. In addition, when the user selects a plurality of images, the distribution server 10 estimates the attribute information of the user in consideration of the order in which the images are selected, the time series in which the images are displayed in the moving image, and the like. The user estimation model 17 to be generated may be generated.

[8-6. Estimation of attribute information using multiple images)
In the example described above, the distribution server 10 estimates the user attribute information from the feature amount of the image selected by the user. However, the embodiment is not limited to this. For example, the distribution server 10 may estimate user attribute information from feature amounts of a plurality of images selected by the user.

  To explain a specific example, the distribution server 10 calculates the feature amounts of a plurality of advertisement images selected by the user, and uses the user estimation model 17 to estimate user attribute information estimated from the calculated feature amounts. Identify all. And the delivery server 10 is good also considering attribute information which is common among the specified attribute information as a user's attribute information.

  Further, the distribution server 10 may estimate the user attribute information based on a sequence of a plurality of images selected by the user. For example, the distribution server 10 selects the plurality of selected images and the order of selection based on the feature amounts of the plurality of advertisement images selected by the user who can acquire the attribute information and the order of selecting the advertisement images. A user estimation model 17 for estimating user attribute information is generated from the above. Then, the distribution server 10 can estimate the attribute information of the user using the user estimation model 17 from the feature amount of the advertisement image selected by the user and the order in which the user selected the advertisement image. Good. For example, when the user whose attribute information is “male” selects the selection image C10 and then selects the selection image C13, the distribution server 10 selects an advertisement image whose feature amount is similar to the feature amount of the selection image C10. After the selection, the attribute information of the user who selected the advertisement image whose feature amount is similar to the feature amount of the selected image C13 may be estimated as “male”.

[8-7. (Replacement of estimation method)
Further, the distribution server 10 may estimate user attribute information whose acquired attribute information is less than a predetermined threshold. For example, FIG. 13 is a diagram illustrating replacement of the estimation method. In FIG. 13, a session in which an advertisement image is selected is extracted, and a user selected from the advertisement image by logistic regression is selected as a teacher label based on whether the selected session is male or female. The correspondence between the AUC and the duplication rate when estimating the gender was plotted for the estimation method using Baseline and the estimation method using CNN.

  For example, as shown in FIG. 13, in the estimation method using Baseline, accuracy increases as the number of user logs increases. On the other hand, the estimation method using CNN can perform estimation with higher accuracy than the estimation method using Baseline when the number of user logs is small.

  Therefore, when the attribute information acquired for the user who selected the advertisement image is less than the predetermined threshold, the distribution server 10 obtains the attribute information of the user from the feature amount of the advertisement image selected by the user. If the attribute information already acquired for the user who selected the advertisement image is greater than the predetermined threshold, information other than the feature amount of the image selected by the user, such as ID identity, is used. Attribute information may be estimated. By executing such processing, the distribution server 10 can accurately estimate the attribute information of the user even in a case other than the cold start state.

[8-8. About the model)
The distribution server 10 may generate a model that implements the various processes described above when executing the generation process. For example, the distribution server 10 may generate the user estimation model 17 that estimates the user attribute information according to the feature amounts of a plurality of advertisement images selected by the user. For example, the distribution server 10 receives the feature quantities of a plurality of advertisement images selected by the user, and generates a function-type user estimation model 17 that outputs user attribute information estimated from the feature quantities. May be.

  Further, the distribution server 10 estimates the attribute information and the predicted click rate from the feature amount of the advertisement image, and calculates the click rate at which each advertisement image is selected using the estimated attribute information and the predicted click rate. A user estimation model 17 may be generated. Further, when the attribute information acquired for the user who selected the advertisement image is less than a predetermined threshold, the distribution server 10 obtains the attribute information of the user from the feature amount of the advertisement image selected by the user. If the attribute information already acquired for the user who selected the advertisement image is greater than the predetermined threshold, information other than the feature amount of the image selected by the user, such as ID identity, is used. A user estimation model 17 that estimates attribute information may be generated. That is, the distribution server 10 may generate a model that realizes any part or all of the distribution processing from the feature amount of the advertisement image selected by the user.

[8-9. (Processing after attribute information estimation)
The distribution server 10 described above distributed the advertisement image using the estimated attribute information. However, the embodiment is not limited to this. For example, the distribution server 10 may recommend various types of information, switch content to be distributed, and the like using the estimated attribute information. Further, the distribution server 10 may estimate the user's identity by comparing the attribute information estimated using the user estimation model 17 and the attribute information of the user registered in advance. Further, the distribution server 10 may restore the connection such as the deleted B cookie based on the estimated user identity. That is, the distribution server 10 can realize arbitrary processing based on the user attribute information estimated using the user estimation model 17.

[8-10. Timing of attribute information estimation)
Further, the distribution server 10 may estimate the attribute information from the feature amount of the image selected by the user at any time until the log of the user attribute information or the like is sufficiently accumulated. Further, the distribution server 10 may update the estimated attribute information as needed based on the estimated history of attribute information. Further, the distribution server 10 may generate a user estimation model 17 that realizes such processing.

[8-11. About the operation subject)
In the above-described embodiment, an example in which the distribution server 10 executes the learning process, the generation process, and the distribution process has been described. However, the embodiment is not limited to this. For example, the process executed by the distribution server 10 estimates the attribute information of the user using the learning device that executes the learning process, the model generation device that performs the generation process, and the user estimation model 17 generated by the model generation device. It may be realized by a distribution device that distributes an advertisement image using the estimated attribute information.

[8-12. Others]
In addition, among the processes described in the above embodiment, all or part of the processes described as being automatically performed can be performed manually, or the processes described as being performed manually can be performed. All or a part can be automatically performed by a known method. In addition, the processing procedures, specific names, and information including various data and parameters shown in the document and drawings can be arbitrarily changed unless otherwise specified. For example, the various types of information illustrated in each drawing is not limited to the illustrated information.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured.

  In addition, the above-described embodiments can be appropriately combined within a range in which processing contents are not contradictory.

[8-13. program〕
Further, the distribution server 10 according to the embodiment described above is realized by a computer 1000 having a configuration as shown in FIG. 14, for example. Hereinafter, the distribution server 10 will be described as an example. FIG. 14 is a hardware configuration diagram illustrating an example of a computer that realizes the function of the distribution server. The computer 1000 includes a CPU 1100, RAM 1200, ROM 1300, HDD 1400, communication interface (I / F) 1500, input / output interface (I / F) 1600, and media interface (I / F) 1700.

  The CPU 1100 operates based on a program stored in the ROM 1300 or the HDD 1400 and controls each unit. The ROM 1300 stores a boot program executed by the CPU 1100 when the computer 1000 is started up, a program depending on the hardware of the computer 1000, and the like.

  The HDD 1400 stores programs executed by the CPU 1100 and data used by the programs. The communication interface 1500 receives data from other devices via the network N, sends the data to the CPU 1100, and transmits the data generated by the CPU 1100 to the other devices.

  The CPU 1100 controls an output device such as a display and a printer and an input device such as a keyboard and a mouse via the input / output interface 1600. The CPU 1100 acquires data from the input device via the input / output interface 1600. In addition, the CPU 1100 outputs the generated data to the output device via the input / output interface 1600.

  The media interface 1700 reads a program or data stored in the recording medium 1800 and provides it to the CPU 1100 via the RAM 1200. The CPU 1100 loads the program from the recording medium 1800 onto the RAM 1200 via the media interface 1700, and executes the loaded program. The recording medium 1800 is, for example, an optical recording medium such as a DVD (Digital Versatile Disc) or PD (Phase change rewritable disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), a tape medium, a magnetic recording medium, or a semiconductor memory. Etc.

  For example, when the computer 1000 functions as the distribution server 10 according to the embodiment, the CPU 1100 of the computer 1000 implements the function of the control unit 13 by executing a program loaded on the RAM 1200. The HDD 1400 stores data in the storage unit 12, that is, the learning image database 14, the connection coefficient storage unit 15, the model generation image database 16, the user estimation model 17, and the advertisement database 18. The CPU 1100 of the computer 1000 reads these programs from the recording medium 1800 and executes them. However, as another example, these programs may be acquired from other devices.

[9. effect〕
As described above, the distribution server 10 calculates the feature amount of the image having the history selected by the user using the learning device that calculates the feature amount from the image. Then, the distribution server 10 estimates the attribute information of the user who selected the image from the image feature amount based on the calculated feature amount and the attribute information indicating the attribute of the user who selected the image. A model 17 is generated.

  For this reason, the delivery server 10 can estimate user attribute information. In particular, the distribution server 10 can accurately estimate the attribute information of the user even when the attribute information of the user to whom the advertisement is distributed is unknown, such as in a cold start state, so that the accuracy of estimating the click rate is improved. The advertisement to be distributed can be appropriately selected.

  Here, a method for estimating the attribute information of the user by calculating the feature amount of the image selected by each user and comparing the calculated feature amount is also conceivable. Since it is necessary to refer to the session log, the calculation cost increases. However, since the distribution server 10 only has to calculate the feature amount of the image selected by the user who estimates the attribute information by generating the user estimation model 17, the calculation cost required for the estimation of the attribute information is reduced. be able to.

  In addition, the distribution server 10 is a learning device in which a plurality of nodes that output calculation results for input data are connected in multiple layers as a learning device, and learning devices that have learned image features abstracted by unsupervised learning. Use. For this reason, the distribution server 10 can estimate the user's attribute information with high accuracy.

  Further, the distribution server 10 uses a learning device, that is, a so-called CNN that further learns the universality of output with respect to image variations by repeating convolution and pooling of local regions. For this reason, since the distribution server 10 can calculate the feature amount according to the distinction of the imaging target regardless of the position of the imaging target in the image, it is possible to estimate the attribute information of the user with higher accuracy.

  Further, the distribution server 10 uses, as a learning device, a learning device that learns the characteristics of an abstracted image using an image in which the ratio selected by a user having a predetermined attribute is greater than a predetermined threshold value. . As a result, since the distribution server 10 calculates the feature amount of the selected image using a learning device that can accurately extract the feature of the abstracted image, it can estimate the attribute information of the user with higher accuracy. Can do.

  Further, the distribution server 10 estimates the attribute information of the user who has selected the image as the user estimation model 17 from the feature amount of the image, and the user selects another image using the estimated attribute information. Generate a model to calculate the rate. For this reason, the distribution server 10 can estimate the click rate with high accuracy even in the cold start state.

  Further, the distribution server 10 estimates the attribute information of the user who has selected the image and the predicted click rate at which the user selects the image as the user estimation model 17 from the feature amount of the image. Using the click rate, a model for calculating a click rate for the user to select another image is generated. For this reason, the distribution server 10 can estimate the click rate with higher accuracy even in the cold start state.

  In addition, the distribution server 10 uses, as the user estimation model 17, the attribute information of the user who has selected the moving image from the feature amount of the image displayed when the user selects the image included in the moving image. Generate a model to estimate. For this reason, the distribution server 10 can estimate the attribute information of the user who browses the moving image.

  Further, the distribution server 10 generates, as the user estimation model 17, a model that estimates the user attribute information from the feature amounts of a plurality of images selected by the user. For this reason, the distribution server 10 can estimate the user's attribute information with higher accuracy.

  Further, the distribution server 10 generates, as the user estimation model 17, a model for estimating the user attribute information from the feature amounts of a plurality of images selected by the user and the order in which the user selects each image. To do. For this reason, the distribution server 10 can estimate the user's attribute information with higher accuracy.

  Further, the distribution server 10 generates a model that estimates the attribute information of the user who selected the advertisement image from the feature amount of the advertisement image as the user estimation model 17. That is, since the distribution server 10 estimates the attribute information of the user from the feature amount of the image actively selected by the user, the attribute information of the user can be estimated with higher accuracy.

  In addition, the distribution server 10 generates, as the user estimation model 17, a model that estimates the attribute information of the user whose acquired attribute information is less than a predetermined threshold from the feature amount of the image selected by the user. For this reason, the distribution server 10 can estimate the attribute information of the user with high accuracy even in a case other than the cold start state.

  Further, the distribution server 10 estimates the user attribute information from the feature amount of the image selected by the user using the generated user estimation model 17, and becomes a distribution target using the estimated attribute information. An image click rate is calculated, and an image to be distributed is selected based on the calculated click rate. For this reason, since the delivery server 10 can select an image that is estimated to be of interest to the user even in the cold start state, the reward for image delivery can be increased.

  In addition, the distribution server 10 estimates a predicted click rate at which the user who has selected the image selects the image using the feature amount of the image, and uses the estimated attribute information and the predicted click rate. The click rate at which the person selects another image is calculated. For this reason, since the delivery server 10 can select an image that is estimated to be of interest to the user even in the cold start state, the reward for image delivery can be increased.

  As described above, some of the embodiments of the present application have been described in detail with reference to the drawings. However, these are merely examples, and various modifications, including the aspects described in the disclosure section of the invention, based on the knowledge of those skilled in the art, It is possible to implement the present invention in other forms with improvements.

  Moreover, the above-mentioned “section (module, unit)” can be read as “means”, “circuit”, and the like. For example, the generation unit 20 can be read as generation means or a generation circuit.

DESCRIPTION OF SYMBOLS 10 Distribution server 11 Communication part 12 Storage part 13 Control part 14 Learning image database 15 Connection coefficient memory | storage part 16 Model production | generation image database 17 User estimation model 18 Advertising database 19 Learning part 20 Generation part 21 Feature-value calculation part 22 User information acquisition Unit 23 model generation unit 30 distribution unit 31 selected image acquisition unit 32 user information estimation unit 33 click rate calculation unit 34 advertisement selection unit 100 terminal device 200 web server

Claims (17)

A calculation unit that calculates a feature amount of an image having a history selected by the user using a learning device that calculates a feature amount indicating the feature of the image from the image;
A generating unit that generates a model for estimating the attribute information of the user who selected the image from the feature amount of the image based on the feature amount calculated by the calculating unit and the attribute information indicating the attribute of the user; A model generation apparatus characterized by comprising:   The calculation unit is a learning device in which a plurality of nodes that output calculation results for input data are connected in multiple layers as the learning device, the learning device learning the features of the image abstracted by unsupervised learning. The model generation apparatus according to claim 1, wherein the model generation apparatus is used.   3. The model generation according to claim 2, wherein the calculation unit uses, as the learning device, a learning device that further learns the universality of output with respect to image variation by repeating convolution and pooling of a local region. apparatus.   The calculation unit uses, as the learning device, a learning device that learns the features of the abstracted image using an image in which a ratio selected by a user having a predetermined attribute is greater than a predetermined threshold value. The model generation device according to claim 2 or 3, wherein   The generation unit estimates the attribute information of the user who has selected the image as the model from the feature amount of the image, and the probability that the user selects another image using the estimated attribute information The model generation apparatus according to claim 1, wherein a model for calculating the model is generated.   The generation unit estimates, as the model, attribute information of a user who has selected the image and a probability that the user will select the image from a feature amount of the image, and the estimated attribute information and the probability. The model generation apparatus according to claim 1, wherein the model generation apparatus generates a model for calculating a probability that the user selects another image.   The generation unit estimates the attribute information of the user who has selected the moving image as the model from the feature amount of the image displayed when the user selects the image included in the moving image. The model generation apparatus according to claim 1, wherein the model generation apparatus generates a model.   The said generation part produces | generates the model which estimates the attribute information of the user who selected the said several image from the feature-value of the several image which the user selected as the said model. The model production | generation apparatus as described in any one.   The generation unit estimates the attribute information of the user who has selected the plurality of images from the feature amounts of the plurality of images selected by the user and the order in which the user has selected each image as the model. The model generation device according to any one of claims 1 to 8, wherein   The said generation part produces | generates the model which estimates the attribute information of the user who selected the image which concerns on the said advertisement from the feature-value of the image which concerns on the said model as any one of Claims 1-9 characterized by the above-mentioned. The model generation apparatus as described in any one.   The generating unit generates, as the model, a model that estimates user attribute information whose acquired attribute information is less than a predetermined threshold from a feature amount of an image selected by the user. The model production | generation apparatus as described in any one of Claims 1-10. Created based on feature quantities calculated from an image having a history selected by the user using a learning device that calculates feature quantities indicating the features of the image from the image, and attribute information indicating the attributes of the user The estimation unit that estimates the attribute information of the user who selected the image from the feature amount of the image using the model and the attribute information estimated by the estimation unit, and the user selects the image to be distributed. A calculation unit for calculating a probability of selection;
An information distribution apparatus comprising: a selection unit that selects an image to be distributed to the user based on the probability calculated by the calculation unit.   The calculation unit estimates the probability that the user who has selected the image selects the image using the feature amount of the image, and uses the estimated attribute information and the probability, the user can select another image. The information distribution apparatus according to claim 12, wherein a probability of selecting is calculated. A model generation method executed by a model generation device,
A calculation step of calculating a feature amount of an image having a history selected by a user using a learning device that calculates a feature amount indicating the feature of the image from the image;
A generating step of generating a model for estimating the attribute information of the user who selected the image from the feature amount of the image based on the feature amount calculated in the calculating step and the attribute information indicating the attribute of the user; A model generation method characterized by including. An information distribution method executed by the information distribution apparatus,
Created based on the feature amount calculated from the image having a history selected by the user using a learning device that calculates the feature amount indicating the feature of the image from the image, and attribute information indicating the attribute of the user An estimation step of estimating attribute information of a user who has selected the image from the feature amount of the image using the model that has been performed;
A calculation step of calculating a probability that the user selects an image to be distributed using the attribute information estimated in the estimation step;
A selection step of selecting an image to be distributed to the user based on the probability calculated in the calculation step. A calculation procedure for calculating a feature amount of an image having a history selected by a user using a learning device that calculates a feature amount indicating the feature of the image from an image on a computer;
A generation procedure for generating a model for estimating the attribute information of the user who selected the image from the feature amount of the image based on the feature amount calculated in the calculation procedure and the attribute information indicating the attribute of the user. A model generation program characterized by being executed. On the computer,
Created based on feature quantities calculated from an image having a history selected by the user using a learning device that calculates feature quantities indicating the features of the image from the image, and attribute information indicating the attributes of the user An estimation procedure for estimating the attribute information of the user who selected the image from the feature amount of the image using
A calculation procedure for calculating a probability that the user selects an image to be distributed using the attribute information estimated in the estimation procedure;
A selection procedure for selecting an image to be distributed to the user based on the probability calculated in the calculation procedure.

Images