JP6585998B2 - Content determination device - Google Patents

Description

  The present invention relates to a content determination apparatus that determines content to be presented to a user.

  Conventionally, a system for distributing an online advertisement displayed on a web page, an application, or the like is known. In such a system, a mechanism is known in which content to be distributed to a user is determined based on feature quantities such as purchase history and personality data of the user's past products and services (see Patent Document 1 below). In this mechanism, the response rate (click rate) of each of the plurality of contents is predicted from the feature amount of the user, and the content predicted to have the highest response rate is distributed to the user.

JP 2002-123739 A

  However, since the number of feature quantities (dimensions) of the user is generally enormous, there is a risk that the amount of calculation will be enormous when trying to determine the distribution content based on the feature quantity of the user. On the other hand, contents such as online advertisements are preferably displayed immediately after the user accesses a web page or the like so that the user can touch it as soon as possible. For this reason, the process for determining the distribution content is required to be performed in a short time, that is, the amount of calculation is as small as possible.

  Here, as a technique for reducing the amount of calculation, the principal component analysis is performed based on the feature amount of the user who has accessed the web page in the past specific period offline in advance, and the number of dimensions less than the original number of dimensions. There is a method of reducing the number of dimensions by using the principal component as a new feature amount. However, the new feature amount generated by the offline processing reflects the tendency of the user group that accessed the web page during the specific period, while the user group that accessed the web page during the specific period. It does not reflect the trend of. For this reason, when the tendency of the user group fluctuates dynamically, there is a possibility that appropriate content that matches the user's interest cannot be determined as distribution content.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a content determination device that can determine content to be presented to the user in accordance with a change in the tendency of the user group.

  A content determination device according to an aspect of the present invention is a content determination device that determines content to be presented to a user of a terminal, and acquires a first feature amount including a plurality of items indicating characteristics of the user of the terminal. A feature amount acquisition unit, a first feature amount storage unit that stores the first feature amount acquired by the first feature amount acquisition unit in association with a presentation time corresponding to the presentation of content, and a content to be presented are newly determined. An analysis period selection unit that probabilistically selects one candidate period as an analysis period from among a plurality of candidate periods prepared in advance for the target user, which is a target user, and the first feature amount storage unit Based on the first feature value of the first feature value whose presentation time is included in the analysis period, the first feature value of the target user is converted into a second feature value having a smaller number of items than the first feature value. The second feature quantity of the target user obtained by the quantity conversion means, the feature quantity conversion means, and the second of the past users who are users whose candidate period same as the target user is selected as the analysis period before the target user. The content to be presented to the target user is determined from a plurality of candidate contents based on the feature amount and operation information indicating whether or not the past user has performed an operation on the content presented to the past user. Determining means.

  In the content determination apparatus, one candidate period is stochastically selected as an analysis period from among a plurality of candidate periods prepared in advance for a target user who newly determines a content to be presented. And the 2nd feature-value by which the number of dimensions (number of items) was reduced based on the 1st feature-value of the user who presented the content in the analysis period is obtained. Subsequently, the second feature amount of the target user, the second feature amount of the past user in which the same candidate period as the target user is selected as the analysis period, and operation information indicating whether the past user has performed an operation on the content Based on the above, the content to be presented to the target user is determined. Thus, in the content determination device, the analysis period used to execute the process of reducing the number of dimensions is not uniformly determined for all users, but one of a plurality of candidate periods. Candidate periods are probabilistically assigned as analysis periods. Accordingly, it is possible to prevent content to be presented based on only the analysis result biased toward the tendency of the user group who requested the content during a specific period for all users. As a result, it is possible to determine the content presented to the user according to the change in the tendency of the user group.

  In the content determination device, the analysis period selection unit displays an operation rate indicating a ratio of the user's operation on the content determined by the determination unit for each user group in which the same candidate period is selected as the analysis period. Acquire and select a candidate period that has the maximum operation rate for each user group from among a plurality of candidate periods with a predetermined probability as an analysis period, and randomly select one of the plurality of candidate periods from a plurality of candidate periods. The candidate period may be selected as the analysis period.

  In the content determination device, the operation period (for example, click rate) for each user group indicating the ratio of the user's operation (for example, click operation) to the content determined by the determination means has the maximum candidate period (hereinafter “best”). The candidate period ") is selected with the highest probability. On the other hand, candidate periods other than the best candidate period are also randomly selected with a predetermined probability. According to such a configuration, the process of determining the candidate period other than the best candidate period as the analysis period while performing the process (utilization) of determining the best candidate period with the maximum operation rate as the analysis period at a certain rate. (Search) can also be performed. As a result, for the target user to which the utilization process is applied, it is possible to present content that is likely to be operated by the target user at the present time. In addition, it is possible to grasp the variation of the user tendency based on the operation information obtained for the target user to which the search process is applied.

  In the content determination device, the determination unit estimates, for each of the plurality of candidate contents, a parameter for calculating an expected value for the operation by the target user from the second feature amount of the target user. The content presented to the target user may be determined based on the parameters estimated for each of the candidate contents.

  According to the content determination device, it is possible to appropriately determine the content to be presented to the target user based on the parameters estimated for each of the plurality of candidate contents. For example, by determining the candidate content having the maximum expected value operated by the target user as the content to be presented to the target user, it is possible to increase the possibility that the target user will operate the content.

  In the content determination device, the content presented to the target user may be determined by executing a calculation using a bandit algorithm based on a parameter estimated for each of the plurality of candidate contents.

  According to the content determination device, the content that is estimated to be the best at the present time (for example, the candidate content having the maximum expected value that is operated by the target user) is presented to the target user by calculation using the bandit algorithm. It is possible to appropriately adjust the ratio of executing the process (utilization) to be determined as and the ratio of executing the process (search) to determine other candidate content as the content presented to the target user. Thereby, it is possible to determine the content presented to the target user so that the above-described utilization and search are executed in a balanced manner.

  According to one aspect of the present invention, content to be presented to a user can be determined according to a change in the tendency of a user group.

It is a figure which shows the function structure of the content determination system containing the content determination apparatus which concerns on this embodiment. It is a figure which shows the hardware constitutions of the content determination apparatus of FIG. It is a figure which shows an example of user information. It is a figure which shows an example of log | history information. It is a figure which shows an example of the total result of log | history information. It is a figure for demonstrating the process of an analysis period selection part and a feature-value conversion part. It is a figure for demonstrating the selection method of the analysis period by (epsilon) -Greyy method. It is a figure for demonstrating the conversion process of the feature-value by principal component analysis. It is a figure which shows an example of the information memorize | stored in a 2nd feature-value memory | storage part. It is a figure which shows an example of content information. It is a flowchart which shows operation | movement of a content determination apparatus.

  Embodiments according to the present invention will be described below with reference to the drawings. Where possible, the same reference numerals are given to parts having similar functions, and duplicate descriptions are omitted.

  FIG. 1 is a block diagram showing a functional configuration of a content determination system including a content determination device according to the present embodiment. A content determination system 1 shown in FIG. 1 includes a terminal 2 owned by a user and a content determination device 3 that is a server device capable of communicating with the terminal 2 via a communication network. The content determination device 3 is connected to each terminal of a plurality of users, and is configured to execute processing described below with each terminal.

  The content determination device 3 is a device that determines content to be presented to the user from among a plurality of candidate contents prepared in advance, and transmits the determined content to the terminal 2. For example, the content determination device 3 determines the content to be presented to the user in response to a request from the user terminal 2. The content request from the terminal 2 to the content determination device 3 is performed in the following scene, for example.

  A web page published on the Internet accessible via a web browser or the like may be provided with a display area for displaying content (for example, advertisement content) including links to various information. . In such a case, for example, when the terminal 2 accesses a web page via a web browser or the like installed in the terminal 2, an acquisition request for content to be displayed in the display area in the web page is It is transmitted from the terminal 2 to the content determination device 3. Thereafter, in the content determination device 3, one content is determined as a content to be presented to the user from among a plurality of candidate contents prepared in advance, and transmitted to the terminal 2. As a result, the content determined by the content determination device 3 is displayed in the display area of the web page displayed on the web browser of the terminal 2.

  In addition, a display area for displaying content including links to various types of information may also be provided on dedicated screens for various applications installed on a smartphone or the like. In such a case, for example, when the user activates an application installed on the terminal 2 and opens a dedicated screen, the terminal 2 sends a content acquisition request to the content determination device 3 by the function of the application. Sent. Thereafter, in the content determination device 3, one content is determined as a content to be presented to the user from among a plurality of candidate contents prepared in advance, and transmitted to the terminal 2. As a result, the content determined by the content determination device 3 is displayed in the display area on the dedicated screen of the application of the terminal 2.

  The “content” in the present embodiment means a display element including links to various information as described above. Such content is generally displayed on the screen of the terminal 2 by a combination of characters, images, and the like. Examples of the content include a web page for purchasing a specific product / service, an advertising content including a link to a web page on which specific information (for example, gourmet information) is posted, and the like. When presenting such content to the user, it is preferable to present the user with content that matches the user's interests and interests from the viewpoint of maximizing the user's convenience and the interests of the advertiser. That is, it is preferable to present to the user content that is likely to be accessed by the link destination information selected by the user and associated with the content.

  Next, components of the content determination system 1 will be described with reference to FIG.

  The terminal 2 is a device that has a one-to-one correspondence with the user, and can communicate with the content determination device 3 via a communication network such as the Internet. Specific examples of the terminal 2 include mobile terminals such as mobile phones, smartphones, tablets, and PDAs, desktop PCs, notebook PCs, and the like. However, the specific form of the terminal 2 is not limited to these. The terminal 2 includes a user information transmission unit 21 and a display unit 22.

  When a content acquisition request is transmitted from the terminal 2 to the content determination device 3, the user information transmission unit 21 includes a user ID for identifying the user of the terminal 2 and a plurality of items indicating the characteristics of the user. This is means for transmitting user information associated with the first feature amount to the content determination device 3. The first feature amount is information including a plurality of items related to, for example, sex, age, living area, presence / absence of purchase of goods (or services), for example, web page visit history for each genre such as gourmet and sports. Such a first feature amount is stored in advance in a storage device or the like provided in the terminal 2, for example.

  The user information may be transmitted directly from the terminal 2 to the content determination device 3 or may be transmitted via another device that mediates between the terminal 2 and the content determination device 3. For example, consider a case where a user accesses a membership website through a web browser installed in the terminal 2. In this case, member user information may be managed in the server device of the provider providing the website. In such a case, when the server device of the business operator accepts the user's access to the website provided by the business operator, the user ID and the first feature amount of the user are transmitted to the content determination device 3. May be.

  The display unit 22 is means for displaying the content transmitted from the content determination device 3 in a display area provided on the above-described web page, a dedicated screen for various applications, or the like. After the content is displayed in the display area by the display unit 22, the user performs an operation on the content displayed in the display area, thereby purchasing information on a link destination set in the content (for example, purchasing a specific product / service). To access web pages, etc.). Here, the “operation for content” is an operation predetermined as an operation for accessing the information of the link destination associated with the content. As a typical example of the operation on the content, the mouse click operation in a state where the position of the mouse cursor is matched with the content displayed on the display area, or the display on the display area when the terminal 2 is a touch panel is displayed. Examples include touch operations on content. In the following description, the operations for the contents in the various modes as described above are collectively referred to as a click operation.

  Next, the configuration of the content determination device 3 will be described. As shown in FIG. 1, the content determination apparatus 3 includes a user information acquisition unit (first feature amount acquisition unit) 31, a user information storage unit (first feature amount storage unit) 32, a history information storage unit 33, A history information acquisition unit 34, an analysis period selection unit (analysis period selection unit) 35, a feature amount conversion unit (feature amount conversion unit) 36, a second feature amount storage unit 37, and a determination unit (determination unit) 38 A transmission unit 39 and a content information storage unit 39a.

  FIG. 2 is a block diagram illustrating an example of a hardware configuration of the content determination device 3. As shown in FIG. 2, the content determination device 3 physically includes one or a plurality of CPUs (Central Processing Units) 301, a main storage device such as a RAM (Random Access Memory) 302 and a ROM (Read Only Memory) 303. , An input device 304 that is an input device, an output device 305 such as a display, a communication module 306 that is a data transmission / reception device, an auxiliary storage device 307 such as a hard disk drive, and the like. Each function of the content determination device 3 in FIG. 1 is configured such that one or a plurality of predetermined computer software is read on hardware such as the RAM 302 shown in FIG. This is realized by operating the device 305 and the communication module 306 and reading and writing data in the RAM 302 and the auxiliary storage device 307.

  The user information acquisition unit 31 receives the user information transmitted by the user information transmission unit 21 of the terminal 2 (or another device such as the above-described server device of the provider) when the content request is received from the terminal 2. It is a means to acquire. That is, the user information acquisition unit 31 acquires the first feature amount of the user of the terminal 2 every time a content request is received from the terminal 2. The user information acquired by the user information acquisition unit 31 is stored in the user information storage unit 32.

  The user information storage unit 32 is a means for accumulating the user information acquired by the user information acquisition unit 31 in association with the presentation time corresponding to the content presentation. Here, the presentation time point corresponding to the presentation of the content may be a time point arbitrarily determined between the time when the user information is acquired from the terminal 2 and the time when the content is transmitted to the terminal 2, for example. . In the present embodiment, as an example, the presentation time point associated with the user information in the user information storage unit 32 is an acquisition date and time when the user information is acquired by the user information acquisition unit 31.

  FIG. 3 shows an example of user information stored in the user information storage unit 32. For example, the first user information in FIG. 3 is that a user whose user ID is “user_001” is a 44-year-old man who has purchased product A and product B. It is acquired at 10:00 on September 1, 2015. Note that the gender, age, living area, presence / absence of purchase of the product A, and presence / absence of purchase of the product B shown in the example of FIG. 3 are items included in the first feature amount. The sex, age, and living area are features that indicate the attributes of the user, and the presence / absence of purchase of the product A and the presence / absence of purchase of the product B are features that indicate the user's behavior history. In general, the number of items indicating the user's behavior history, such as whether or not a product has been purchased, is as large as the number of products (or services) registered as management targets. Further, the feature indicating the action history includes not only information on whether or not the product is purchased, but also a website visit history for each genre such as a gourmet site or a news site. The feature related to the user's taste / hobby can be grasped by the feature amount related to the visit history.

  The history information storage unit 33 is means for storing history information including click information (operation information) indicating whether or not a click operation has been performed on the content when the content is displayed on the terminal 2 of each user. FIG. 4 is a diagram illustrating an example of history information. As illustrated in FIG. 4, the history information stored in the history information storage unit 33 is, for example, information in which presentation date / time, user ID, analysis period, content ID, and click information are associated with each other.

  “Presentation date and time” is the date and time when the content is presented to the user. As the presentation date and time, the date and time when the content is transmitted to the terminal 2 by the transmission unit 39 to be described later may be stored. “User ID” is a user ID for identifying the user of the terminal 2 to which the content is transmitted. The “analysis period” is an analysis period assigned to the user by the analysis period selection unit 35 described later. “Content ID” is an ID for identifying content. As described above, “click information” is information indicating whether or not a click operation has been performed on the content. For example, in the first history information in FIG. 4, the content “ad_001” is presented to a user “user_001” to which a candidate period C1 described later is assigned as an analysis period at 10:00 on September 1, 2015. It indicates that the content has been clicked by the user.

  FIG. 5 shows an example of a total result of a plurality of history information stored in the history information storage unit 33. Such a tabulation process can be easily performed by, for example, a query process on a database. Therefore, it can be said that the history information storage unit 33 substantially stores not only the history information shown in FIG. 4 but also various pieces of aggregate information obtained by aggregating a plurality of history information as shown in FIG. . The tabulation results shown in FIG. 5 are the results of tabulating the content display count, content click count, and click rate (CTR: Click Through Ratio) for each analysis period.

  Here, the number of times of content display and the number of content clicks for each analysis period are the number of times of content display and the number of times of content display for each user group for which the same candidate period (described later in detail) is selected as the analysis period. This is the total number of content clicks. The click rate for each analysis period is the click rate (operation rate) for each user group in which the same candidate period is selected as the analysis period, and is a value indicating the ratio of the number of content clicks to the number of content displays for each analysis period. . For example, the click rate of the first record in FIG. 5 is the click rate for a user group in which a later-described candidate period C1 is selected as the analysis period.

  The history information acquisition unit 34 is means for acquiring the history information described above and storing it in the history information storage unit 33. For example, when the content is transmitted to the terminal 2 by the transmission unit 39 described later, the history information acquisition unit 34 acquires the transmission date and time of the content as the presentation date and time. In addition, the history information acquisition unit 34 associates the user ID of the user of the terminal 2 with the content ID of the content transmitted to the terminal 2 and refers to the second feature amount storage unit 37 to be described later. Get the analysis period assigned to the user. Further, the history information acquisition unit 34 receives notification from the terminal 2 of click information indicating whether or not a click operation by the user of the terminal 2 has been performed on the content transmitted to the terminal 2 by the transmission unit 39. Etc. In addition, the acquisition method of said various information is an example, and the historical information acquisition part 34 may acquire historical information by methods other than the above.

  The analysis period selection unit 35 probabilistically selects one candidate period as an analysis period from among a plurality of candidate periods prepared in advance for a target user who is a user of the terminal 2 that newly requests content. It is. Here, the analysis period is a period used for dimension reduction of the first feature quantity of the target user by the feature quantity conversion unit 36 described later. Specifically, the analysis period selection unit 35 will be described later each time a target user newly accesses a web page provided with a content display area, a dedicated screen for an application, or the like (hereinafter referred to as “specific page”). An analysis period for performing the conversion process by the feature amount conversion unit 36 is selected from a plurality of candidate periods.

  Processing of the analysis period selection unit 35 will be described with reference to FIG. In the example of FIG. 6, the reference times t1, t2, t3, and t4 are set at a predetermined time interval (one hour interval as an example) from a certain reference time t0. In this example, a case will be described in which an access of the target user to the specific page (that is, a content acquisition request from the terminal 2 of the target user) occurs at a time t between the reference time t3 and the reference time t4. In this embodiment, as an example, the analysis period selection unit 35 manages a period (1 hour) between the reference time tk and the reference time tk + 1 set as described above as one unit period. In addition, the analysis period selection unit 35 sets a plurality of periods that are different from each other by the number of unit periods from the most recent reference time t3 past the access time t to the specific page of the target user as candidate periods.

  In the example of FIG. 6, the analysis period selection unit 35 prepares the following three candidate periods C1 to C3 as candidate periods for the target user who accessed the specific page at time t. However, the number of candidate periods is not limited to three, and may be two or four or more. Candidate period C1 is the longest longest period among the three candidate periods, and is a period that goes back three unit periods starting from reference time t3 (ie, three hours from reference time t0 to reference time t3). Candidate period C2 is the second longest medium period of the three candidate periods, and is a period that goes back by two unit periods with reference time t3 as a base point (that is, two hours from reference time t1 to reference time t3). Candidate period C3 is the shortest short-term period among the three candidate periods, and is a period that goes back one unit period with reference time t3 as a base point (that is, one hour from reference time t2 to reference time t3).

  The analysis period selection unit 35 selects one candidate period among the above-described candidate periods C1 to C3 as an analysis period to be assigned to the target user, for example, by the ε-Greedy method classified as a so-called contextless bandit algorithm. . Specifically, the analysis period selection unit 35 refers to the history information stored in the history information storage unit 33, so that the same candidate period is the most from the viewpoint of the click rate for each user group selected as the analysis period. A good candidate period (hereinafter “best candidate period”) is extracted. More specifically, the analysis period selection unit 35 extracts a candidate period with the maximum click rate for each user group as the best candidate period. In the example of FIG. 6, the user whose one candidate period is selected as the analysis period from the candidate periods C1 to C3 before the target user is a user who accesses a specific page between the reference time t3 and the time t. (Past user). As the time from the reference time t3 elapses, the click information of the past user is acquired as history information, and the total number (content display count, content click count) shown in FIG. 5 increases. Then, the analysis period selection unit 35 selects the best candidate period as the analysis period with a predetermined probability, and randomly selects one candidate period as the analysis period from the plurality of candidate periods C1 to C3 with the remaining probability. .

  Specifically, the analysis period selection unit 35 refers to the history information storage unit 33 to extract a candidate period with the maximum click rate for each user group as the best candidate period. Here, it is assumed that the example of FIG. 5 shows history information (aggregation result) at time t. In this case, the analysis period selection unit 35 specifies the candidate period C3 that maximizes the click rate for each user group as the best candidate period. Then, as shown in FIG. 7, the analysis period selection unit 35 sets the candidate period C3 specified as the best candidate period as an analysis period that is assigned to the target user with a predetermined probability 1−ε (0 <ε <1). select. On the other hand, the analysis period selection unit 35 selects one candidate period randomly extracted from the plurality of candidate periods C1 to C3 with the remaining probability (search probability) ε as an analysis period to be assigned to the target user.

  By such a process, the process (utilization) for selecting the best candidate period (candidate period C3) for which the best result (that is, the maximum click rate) is currently obtained as an analysis period is a certain ratio (1-ε). In addition, a process (search) for selecting an analysis period from all candidate periods C1 to C3 including candidate periods C1 and C2 other than the best candidate period can be performed. For example, when the search probability ε is 0.3, the analysis period selection unit 35 executes the “utilization” process with a probability of 0.7, and executes the “search” process with a probability of 0.3. In the search process, all candidate periods are randomly selected. In this case, the analysis period selection unit 35 selects the candidate period C3 as an analysis period with a probability of 0.8, and each of the candidate periods has a probability of 0.1. C1 and C2 are selected as analysis periods. For a target user to which the “utilization” process is applied, it is possible to present content that is likely to be clicked by the target user at the present time. In addition, it is possible to grasp the variation of the user tendency based on the click information obtained for the target user to which any of the candidate periods C1 and C2 is selected as the analysis period by applying the “search” process. Specifically, it becomes possible to grasp the variation in the click rate of the user group in which the candidate periods C1 and C2 other than the candidate period C3 which is the best candidate period are selected as the analysis period. In this way, instead of applying the “utilization” process to all users, by applying the “search” process at a certain rate, a change in trend (for example, a group of users who can obtain the maximum click rate) It is possible to grasp that the candidate period C3 changes from the user group selected as the analysis period to the user group selected as the analysis period.

  Moreover, the analysis period selection part 35 may change the period which the candidate periods C1-C3 show for every fixed time according to progress of time. Specifically, when the current time has passed the reference time t4, the analysis period selection unit 35 may slide the candidate periods C1 to C3 hour by hour. That is, the analysis period selection unit 35 sets 3 hours from the reference time t1 to the reference time t4 as a new candidate period C1, sets 2 hours from the reference time t2 to the reference time t4 as a new candidate period C2, and sets the reference time t3. One hour from the reference time t4 to the reference time t4 is set as a new candidate period C3. By doing in this way, it is possible to perform feature quantity dimension reduction based on a tendency (feature) of a user who has accessed a specific page in a relatively short time zone from the current time by the feature quantity conversion unit 36 described later. Become.

  The feature amount conversion unit 36 determines the first feature amount of the target user based on the first feature amount of the user (analysis target user) who has accessed the specific page during the analysis period selected by the analysis period selection unit 35. This is means for converting to a second feature quantity having a smaller number of items (dimensions) than the feature quantity. Specifically, the feature amount conversion unit 36 performs principal component analysis based on the first feature amount whose acquisition date and time is included in the analysis period among the first feature amounts of each user accumulated in the user information storage unit 32. Then, a predetermined number of principal components are determined. However, the feature quantity conversion unit 36 may use a technique other than the principal component analysis as a technique for obtaining the second feature quantity by reducing the dimension of the first feature quantity of the target user.

  By such processing, it becomes possible to perform principal component analysis reflecting the tendency of the analysis target user group that has accessed the specific page during the period selected as the analysis period. For example, in the analysis target user group who has accessed a specific page during the period selected as the analysis period, the ratio of users who have a strong interest in gourmet information (users who have visited many websites related to gourmet in the past) is high. In this case, the principal component analysis is executed in such a way that the feature amount related to the interest in the gourmet information is emphasized. Thereby, in the process of the determination part 38 mentioned later, paying attention to the strength of interest with respect to gourmet information, appropriate content (for example, content with high possibility that a user with strong interest in gourmet information will click) is presented to the target user. It can be determined as content. As described above, the analysis target user group used in the principal component analysis changes depending on the selected analysis period, and the converted second feature amount is different. For this reason, depending on which candidate period is selected as the analysis period, which content is likely to be determined by the determination unit 38 to be described later varies.

  The right part of FIG. 6 shows an outline of the process for determining the above-mentioned principal component. As illustrated in FIG. 6, the feature amount conversion unit 36 uses the n-dimensional item axis (F1 to Fn) representing the first feature amount as the first feature amount of the analysis target user who accessed the specific page during the analysis period. By executing the principal component analysis based on this, the number of principal components 1 to principal components 3 is smaller than n (here, three as an example).

  Hereinafter, the feature amount conversion unit 36 will be described in detail. The feature amount conversion unit 36 first quantifies the first feature amount of the analysis target user who accessed the specific page during the analysis period. When the first feature value is digitized in advance, the digitization process can be omitted. The feature amount conversion unit 36 quantifies the sex such as “0” for female and “1” for male. For example, the feature amount conversion unit 36 is grouped by age, for example, 0 to 9 years old is “0”, 10 to 19 years old is “1”, and 20 to 29 years old is “2”. Digitize. The feature amount conversion unit 36 also digitizes other items included in the first feature amount based on a predetermined conversion rule.

Subsequently, as illustrated in FIG. 8, the feature amount conversion unit 36 performs a diagram based on the first feature amount (each value of n items x1 to xn) of the analysis target user who accesses the specific page during the analysis period. Parameters a _ij (i = 1... N, j = 1... K) of the conversion equation shown in FIG. Such a parameter a _ij depends on the analysis period. For example, when there are many users who are strongly interested in gourmet information in the analysis target user group who accessed a specific page during the analysis period as described above, the feature amount related to interest in gourmet information is emphasized more than other feature amounts. Parameter a _ij is determined as follows.

Here, Z _{1 to} Z _k indicate principal components arranged in the order of a small information loss from the original information (information expressed by the items x1 to xn) expressed as the first feature amount. The feature amount conversion unit 36 uses a predetermined number of principal components (three as an example in the present embodiment) in the order from the smallest information loss from the original information among the principal components obtained in this way. In other words, the feature quantity conversion unit 36 uses the parameter a _ij determined by the above processing, so that the first feature quantity (each value of x1 to xn) of the target user is smaller in number than the first feature quantity. Conversion into (three) second feature quantities (principal component 1 to principal component 3). Note that the parameter a _ij can be calculated in advance for each of the candidate periods C1 to C3 before executing the process of converting the first feature amount of the target user. Thereby, when converting the first feature amount of the target user into the second feature amount, the parameter corresponding to the analysis period assigned to the target user among the parameters already obtained for each of the candidate periods C1 to C3 is set. The calculation of the conversion formula shown in FIG. For this reason, the conversion process can be executed in a short time.

  The second feature amount storage unit 37 is a means for accumulating the selected analysis period and the second feature amount (principal component 1 to principal component 3) obtained by the feature amount conversion unit 36 for each user. FIG. 9 shows an example of information stored in the second feature amount storage unit 37. Each time the target user accesses the second feature amount storage unit 37, information about the target user (for one record in the table shown in FIG. 9) is accumulated. Information accumulated in the second feature amount storage unit 37 is referred to as appropriate by the history information acquisition unit 34 and the like.

  The determination unit 38 is a unit that determines content to be presented to the target user from among a plurality of candidate contents based on the second feature amount of the target user obtained by the feature amount conversion unit 36. Specifically, the determination unit 38 estimates, for each of the plurality of candidate contents, a parameter for calculating an expected value for a click operation by the target user from the second feature amount of the target user. The content to be presented to the target user is determined based on the parameters estimated for each of the above. In the present embodiment, as an example, the determination unit 38 performs the above-described parameter estimation, and the second feature amount of the target user obtained by the feature amount conversion unit 36 is the same as the target user before the target user. Whether the past user has performed a click operation on the second feature amount of the past user whose analysis period is selected and the content presented to the past user (that is, the content determined by the determination unit 38 in the past) Using the click information indicating these, a calculation process using a linear model is performed. Hereinafter, calculation processing using this linear model will be described. The calculation process described below is based on a known calculation algorithm using a linear model.

The determination unit 38 uses an expected value E [r _i (t) | b _i that is clicked by the target user in the linear model represented by the following formula (1) prepared for each set of the analysis period and the candidate content. (T)] is estimated from the second feature value b _i (t) of the target user. By estimating such a parameter μ with respect to a linear model (Equation (1) below) for each set of the candidate period and the analysis period assigned to the target user, the target user clicks each of the plurality of candidate contents. An expected value E [r _i (t) | b _i (t)] to be operated can be calculated. In addition, by classifying candidate contents in advance for each type (for example, genres such as “gourmet” and “sports”), candidate contents of the same type may be aggregated in one group. In this case, a linear model represented by the following equation (1) is prepared for each set of analysis period and candidate content type.
E [r _i (t) | b _i (t)] = b _i (t) ^T μ (1)

Specifically, the expected value E [r _i (t) | b _i (t)] indicates that the target user is a candidate when the candidate content is presented to the target user having the second feature quantity b _i (t). This is an expected value for performing a click operation on the content. B _i (t) is a vector in which the principal components (three principal components in this embodiment) of the second feature amount of the target user are arranged. In the example of FIG. 8, b _i (t) = (0.11, 0.54, 0.35). The parameter μ is a vector having the same dimension as the second feature amount, and is expressed as μ = (μ ₁ , μ ₂ , μ ₃ ). The determination unit 38 performs a calculation process described below for each set of analysis period and candidate content in order to estimate the parameters μ _{1 to} μ ₃ that satisfy the above formula (1).

  Hereinafter, the process of the determination unit 38 for estimating the expected value for the target user to which the specific analysis period C (any one of C1 to C3) is assigned to perform the click operation on the specific candidate content T will be described. By performing the processing described below for each candidate content in the same manner, it is possible to estimate the expected value for the click operation by the target user for each candidate content.

First, when the dimension of the second feature value b _i (t) is k (k = 3 in the present embodiment), the determination unit 38 determines that the initial value is a k-dimensional vector f having a zero vector. A k × k matrix B whose initial value is a unit matrix and a k-dimensional vector μ ′ whose initial value is a zero vector and is an estimated value of the parameter μ in the above equation (1) are prepared in advance.

When the candidate content T is determined as the content to be presented to one past user U to which the analysis period C is assigned, the determination unit 38 determines the above-described update formulas (2-1) to (2-3) as described above. Update vector f, matrix B, and vector μ ′.
B = B + b _u (t) b _u (t) ^T (2-1)
f = f + b _u (t) r _u (t) (2-2)
μ ′ = B ⁻¹ f (2-3)

In the update equations (2-1) to (2-3), b _u (t) is a vector in which the principal components of the second feature value of the past user U are arranged. Further, r _u (t) is click information indicating whether or not a click operation by the past user U has been performed on the candidate content T. When the past user U has clicked the candidate content T, r _u (t) is “1”, and when the past content U has not been clicked by the past user U, r _u (t) is “0”. Note that the determination unit 38 refers to the history information stored in the history information storage unit 33 (see FIG. 4), thereby indicating whether or not a click operation by the past user U has been performed on the candidate content T. Can be grasped. Each time the candidate content T is determined as the content to be presented to the user to which the analysis period C is assigned, the determination unit 38 obtains the click information of each user as described above, so that each parameter (vector f, matrix) is obtained. B, the vector μ ′) is updated by the update formulas (2-1) to (2-3).

In summary, the determination unit 38 estimates the expected value E [r _i (t) | b _i (t)] at which the candidate content T is clicked by the target user to which the analysis period C is assigned at time t. Calculates the vector μ ′ by the above equation (2-3) using the vector f and the matrix B (may be an initial value or may be updated by the above update equation) at time t. To do. The determination unit 38 substitutes the vector μ ′ calculated in this way for the parameter μ in the above equation (1), so that the expected value E [r _i (t) | b where the candidate content T is clicked by the target user. _i (t)] can be calculated (estimated).

  The determination unit 38 can estimate, for each candidate content, an expected value at which the candidate content is clicked by the target user to which the analysis period C is assigned by executing the calculation described above for each of the plurality of candidate contents. . Hereinafter, the “expected value that the candidate content is clicked by the target user to which the analysis period C is assigned” is referred to as “click expected value”.

(First example)
The determination unit 38 may determine, as content to be presented to the target user, candidate content that maximizes the click expectation value estimated by the above-described calculation processing among the plurality of candidate contents. In this case, it is possible to display appropriate content that matches the interest of the target user (that is, content that is most expected to be clicked by the target user) to the target user.

(Second example)
In addition, the determination unit 38 performs a calculation using the ε-Greydy method (bandit algorithm) based on the parameter (vector μ ′) estimated for each of the plurality of candidate contents, similarly to the analysis period selection unit 35 described above. By executing, content to be presented to the target user may be determined. Specifically, the determination unit 38 calculates the expected click value for each of the plurality of candidate contents using the above-described equation (1) using the vector μ ′ estimated by the calculation process described above. Then, the determination unit 38 determines the candidate content having the maximum click expectation value among the plurality of candidate contents as the content to be presented to the target user with a predetermined probability, and from among the plurality of candidate contents with the remaining probability One candidate content may be determined as content to be presented to the target user at random.

(Third example)
Further, the determination unit 38 performs a calculation using a so-called bandit algorithm with a context, which is higher than the above-described ε-Greedy method, based on the parameter (vector μ ′) calculated for each of the plurality of candidate contents. Thus, the content to be presented to the target user may be determined. Since the calculation process itself by the above-described bandit algorithm with context is a conventionally known calculation algorithm, only the outline will be described.

The determining unit 38 uses the vector μ ′ calculated by the above equation (2-3) and the matrix B ⁻¹ to use the normal distribution N (average μ ′, variance v ² B ⁻¹ shown in the following equation (3). Extract random number μ '' according to multi-dimensional normal distribution). In the following equation (3), v is a dispersion parameter that determines the degree of search in the bandit algorithm. v is an arbitrary real value, and is arbitrarily set according to a desired degree of search, for example, by an operator of the content determination device 3 or the like.
μ ″ to N (μ ′, v ² B ⁻¹ ) (3)

Subsequently, the determination unit 38 calculates a score value b _i (t) ^T μ ″ for the candidate content based on the extracted random number μ ″. The determination unit 38 similarly applies the above-described score value calculation process (a series of calculation processes using the above equations (2-1) to (2-3) and (3) for one candidate content) for each candidate content. By doing so, a score value for each candidate content can be obtained. And the determination part 38 determines the candidate content with the largest score value among several candidate content as a content shown to an object user.

  When the content to be presented to the target user is determined by executing the calculation using the bandit algorithm as in the second example and the third example, the content that is estimated to be the best at the present time (that is, the above formula The ratio of executing the process (utilization) for determining the candidate content having the maximum expected value calculated in (1) as the content to be presented to the target user and the other candidate content are determined as the content to be presented to the target user. The ratio of executing the process (search) can be adjusted appropriately. Thereby, it is possible to determine the content to be presented to the target user so that the above-described utilization and search are executed in a balanced manner.

  The transmission unit 39 is means for transmitting (distributing) the content determined by the determination unit 38 to the terminal 2. In this embodiment, as an example, information (content information) of each candidate content including a content ID for identifying the content is stored in advance in the content information storage unit 39a. The transmission unit 39 is notified of the content ID for identifying the candidate content determined as the content to be presented to the target user from the determination unit 38, acquires the corresponding content information from the content information storage unit 39a using the content ID as a key, The content information is transmitted to the terminal 2. As shown in FIG. 10, the content information includes various information (stores) associated with the content in addition to the information indicating the content ID and the content (characters, images, link destinations, etc. displayed on the information presentation screen). Type, genre, and advertisement creation company).

  Next, the operation of the content determination device 3 will be described with reference to FIG. Specifically, an operation from when a target user accesses a specific page until content to be presented to the target user is determined and transmitted will be described. First, the user information acquisition unit 31 acquires user information (including a first feature amount) of a target user who accesses a specific page from the terminal 2 of the target user (step S1).

  Subsequently, the analysis period selection unit 35 selects an analysis period to be assigned to the target user from among a plurality of candidate periods C1 to C3 (see FIG. 6) prepared in advance (step S2). For example, the analysis period selection unit 35 selects an analysis period using the ε-Greyy method. Specifically, the analysis period selection unit 35 refers to the history information storage unit 33 to extract a candidate period (candidate period C3 in the example of FIG. 5) having the maximum click rate for each user group. Then, the analysis period selection unit 35 selects a candidate period having the maximum click rate with a predetermined probability 1−ε as an analysis period, and randomly selects one candidate period from the plurality of candidate periods with the remaining probability ε. Is selected as the analysis period (see FIG. 7).

  Subsequently, the feature quantity conversion unit 36 performs the principal component analysis based on the analysis period selected by the analysis period selection unit 35, whereby the first feature quantity of the target user is more than the number of items ( Conversion into the second feature quantity (number of dimensions) having a small number of dimensions (in this embodiment, principal component 1 to principal component 3 as an example) (step S3). Specifically, the feature amount conversion unit 36 performs a principal component analysis based on the first feature amount of the analysis target user who has accessed the specific page during the analysis period, so that a predetermined number (three as an example) is obtained. Determine the principal component.

  Subsequently, the determination unit 38 determines content to be presented to the target user based on the second feature amount of the target user obtained by the feature amount conversion unit 36 (step S4). For example, the determination unit 38 can determine the content to be presented to the target user using the linear model and the bandit algorithm as described above.

  Subsequently, the transmission unit 39 transmits the content information determined by the determination unit 38 to the terminal 2 of the target user (step S5). Thereby, the content is displayed in the display area of the specific page by the display unit 22 of the terminal 2 of the target user. After the content is transmitted to the target user in step S5, the history information acquisition unit 34 acquires click information indicating whether the click operation by the target user has been performed on the displayed content. Accumulated in the storage unit 33. Then, the determination unit 38 updates the vector f, the matrix B, and the vector μ ′ by referring to the history information storage unit 33 and executing the calculations according to the update formulas (2-1) to (2-3). To do.

  In the content determination device 3 described above, one candidate period among a plurality of candidate periods C1 to C3 prepared in advance for a target user whose content to be presented is newly determined by the analysis period selection unit 35. Is stochastically selected as the analysis period. Then, the feature quantity conversion unit 36 obtains a second feature quantity in which the number of dimensions (number of items) is reduced based on the first feature quantity of the user whose content is presented during the analysis period. Specifically, for example, a principal component obtained by principal component analysis is obtained as the second feature amount. Then, based on the second feature amount of the target user, the second feature amount of the past user in which the same candidate period as the target user is selected as the analysis period, and the click information of the past user by the determination unit 38, The content to be presented to the target user is determined. As described above, in the content determination device 3, the analysis period used for executing the process of reducing the number of dimensions is not uniformly determined for all users. One of the candidate periods is probabilistically assigned as the analysis period. As a result, it is possible to prevent content to be presented for all users based on only the analysis result (principal component analysis result) biased toward the tendency of the user group that requested the content during a specific period. As a result, it is possible to determine the content to be presented to the user according to the change in the tendency of the user group.

  For example, consider a case where the click rate of the user group in which the candidate period C1 which is the long period shown in FIG. That is, when the feature amount is reduced by principal component analysis with the candidate period C1 as an analysis period, and the content to be presented is determined based on the feature amount after the dimension reduction (second feature amount), another candidate period Consider a case where a better click rate is obtained than when dimensions are reduced using C2 and C3. In this case, if only the best candidate period C1 is assigned as the analysis period to the target user who accesses the specific page after that time, the candidate periods C2 and C3 other than the candidate period C1 thereafter. The click rate cannot be measured when is used as the analysis period. As a result, even if the tendency of the target user accessing the specific page changes and the click rate of the user group selected with the candidate period C3 as the analysis period is actually maximized, Unable to grasp changes. On the other hand, the content determination device 3 executes “utilization” and “search” at a predetermined ratio by a bandit algorithm (ε-Greedy method) regarding selection of an analysis period used for dimension reduction of a feature amount. A candidate period other than the best candidate period estimated to be the best at a certain point in time is also selected with a predetermined probability. Therefore, it is possible to measure the click rate when the candidate period other than the best candidate period is the analysis period with a certain degree of accuracy (accuracy according to the search rate). Therefore, it is possible to appropriately grasp a change in the situation such as a change in the best candidate period, and it is possible to select an analysis period to be assigned to the target user in accordance with such a change.

  Further, in the content determination device 3, even when the determination unit 38 determines the content to be presented to the user, “use” and “search” are executed at a predetermined ratio for the content by executing the bandit algorithm. Thereby, also from the viewpoint of content, it is possible to determine the content to be presented to the target user in accordance with a change in the tendency of the user group (for example, a change in the interest of the user due to a change in time).

  The present invention has been described in detail based on the embodiments. However, the present invention is not limited to the above embodiment. The present invention can be variously modified without departing from the gist thereof.

  DESCRIPTION OF SYMBOLS 1 ... Content determination system, 2 ... Terminal, 3 ... Content determination apparatus, 31 ... User information acquisition part, 32 ... User information storage part, 33 ... History information storage part, 34 ... History information acquisition part, 35 ... Analysis period selection part , 36 ... feature amount conversion unit, 37 ... second feature amount storage unit, 38 ... determination unit, 39 ... transmission unit, 39a ... content information storage unit.

Claims (4)

A content determination device for determining content to be presented to a user of a terminal,
First feature quantity acquisition means for acquiring a first feature quantity comprising a plurality of items indicating characteristics of a user of the terminal;
First feature quantity storage means for storing the first feature quantity acquired by the first feature quantity acquisition means in association with a presentation time point corresponding to the presentation of the content;
An analysis period selecting means for probabilistically selecting one candidate period as an analysis period from a plurality of candidate periods prepared in advance for a target user who is a target user who newly determines content to be presented;
Based on the first feature amount of the first feature amount accumulated in the first feature amount storage means, the presentation feature is included in the analysis period, the first feature amount of the target user is determined as the first feature amount. Feature quantity conversion means for converting to a second feature quantity having a smaller number of items than
The second feature amount of the target user obtained by the feature amount conversion means and the second feature of the past user who is the user who has selected the same candidate period as the target user as the analysis period before the target user. Based on the feature amount and the operation information indicating whether or not an operation by the past user has been performed on the content presented to the past user, the content to be presented to the target user is selected from a plurality of candidate contents. A decision means to decide;
A content determination apparatus comprising: The analysis period selecting means is
An operation rate indicating a rate at which a user performs an operation on the content determined by the determination unit is acquired for each user group in which the same candidate period is selected as the analysis period,
With a predetermined probability, select a candidate period that maximizes the operation rate for each of the user groups among the plurality of candidate periods as the analysis period,
With the remaining probability, randomly select one candidate period from the plurality of candidate periods as the analysis period.
The content determination apparatus according to claim 1. The determination unit estimates, for each of the plurality of candidate contents, a parameter for calculating an expected value that is operated by the target user on the content from a second feature amount of the target user. Determining content to be presented to the target user based on the parameters estimated for each of the candidate content
The content determination apparatus according to claim 1 or 2. The determination means determines content to be presented to the target user by executing a calculation using a bandit algorithm based on the parameters estimated for each of the plurality of candidate contents.
The content determination apparatus according to claim 3.

Images