JP6692739B2 - Extraction device, extraction method, and extraction program - Google Patents

Description

The present invention relates to an extraction device, an extraction method and an extraction program.

  In recent years, a technique for distributing contents such as news and advertisement via the Internet has been known. As an example of a technology for distributing such content, there is known a technology for estimating the importance of the content based on the content content and the newness of the content and preferentially displaying the content with the estimated high importance. There is. For example, in such a technique, the importance based on the content is estimated in advance, and the estimated importance is attenuated over time, so that new content is preferentially displayed.

JP, 2005-99265, A

  However, the above-mentioned conventional technique has a problem that it is not always possible to appropriately estimate the importance of content. For example, contents such as columns are less important than new contents such as news, so that importance does not change much over time. For this reason, if the importance of each content is uniformly attenuated, there may be a difference from the actual importance. Further, a model in which the importance is estimated in advance by considering the decrease of the importance with the passage of time may be considered. However, learning of such a model is difficult because there are contents and the like whose true importance is known after a certain period of time depending on the type of contents.

  The present application has been made in view of the above, and an object thereof is to improve the accuracy of estimating the importance of content.

The extraction device according to the present application inputs the characteristic information indicating the characteristic of the content at a predetermined time based on the predetermined stationary information about the content and the variation information that changes with the passage of time in the information about the content. The present invention is characterized by having an extraction unit that extracts a feature of the information from the extracted information by using a learning device.

  According to one aspect of the embodiment, it is possible to improve the accuracy of estimating the importance of content.

FIG. 1 is a diagram illustrating an example of processing executed by the information distribution processing according to the embodiment. FIG. 2 is a diagram illustrating an example of a functional configuration of the information distribution device according to the embodiment. FIG. 3 is a diagram showing an example of information registered in the content database according to the embodiment. FIG. 4 is a diagram showing an example of information registered in the performance information database according to the embodiment. FIG. 5 is a diagram showing an example of information registered in the learning device database according to the embodiment. FIG. 6 is a diagram showing an example of information registered in the feature information database according to the embodiment. FIG. 7 is a diagram showing an example of information registered in the content database according to the embodiment. FIG. 8 is a diagram illustrating an example of a process in which the information distribution device according to the embodiment extracts feature information. FIG. 9 is a diagram illustrating an example of processing in which the terminal device according to the embodiment displays content according to priority. FIG. 10 is a flowchart showing an example of the flow of processing in which the information distribution apparatus according to the embodiment accepts posting of content. FIG. 11 is a flowchart showing an example of the flow of an estimation process performed by the information distribution device according to the embodiment. FIG. 12 is a flowchart showing an example of information distribution processing by the information distribution device according to the embodiment. FIG. 13 is a diagram illustrating an example of the hardware configuration of a computer that realizes the functions of the information distribution device.

Hereinafter, modes (hereinafter, referred to as “embodiments”) for carrying out the extraction device, the extraction method, and the extraction program according to the present application will be described in detail with reference to the drawings. Note that the extraction device, the extraction method, and the extraction program according to the present application are not limited by this embodiment. Also, in each of the following embodiments, the same parts are designated by the same reference numerals, and duplicate description will be omitted.

[1. Example of information distribution device]
First, an example of the information distribution device 10 which is an example of an estimation device will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of processing executed by the information distribution processing according to the embodiment. In FIG. 1, as an example of a process executed by the information distribution device 10, an estimation process of estimating the importance of content for which a posting is accepted from a content distribution source for each predetermined time period, and a terminal used by the user U01. An example of the information distribution process for causing the device 100 to display the information in the order according to the estimated importance will be described.

  Here, the content distributed by the information distribution apparatus 10 is content registered in the information distribution apparatus 10 by a predetermined client, and is, for example, content including text such as news and columns. In the following description, an example will be described in which the information distribution device 10 distributes content including text, but the embodiment is not limited to this. That is, the information distribution process described below is performed on the terminal device 100 such as contents not including text such as images and moving images, contents relating to commercial or non-profit advertisement, games, mails, tweets, blog articles, etc. It is applicable to any content that can be distributed by

  As shown in FIG. 1, the information distribution device 10 is communicatively connected to a terminal device 100 used by a user U01 and a client terminal 200 used by a client via a network (not shown) such as the Internet. .. The information distribution device 10 is communicatively connected to the terminal device 101 and the log server 102 via a network. The information distribution device 10 can communicate with any number of terminal devices 100 and 101, the log server 102, and the client terminal 200.

  The terminal device 100 is an information processing device used by the user U01, and is used by the user to browse content such as news and columns. As a specific example, the terminal device 100 is a mobile terminal such as a smartphone, a tablet terminal, a PDA (Personal Digital Assistant), a desktop PC (Personal Computer), a notebook PC, or the like.

  The terminal device 101, like the terminal device 100, is an information processing device used by the user U01. The terminal device 101 may be the same device as the terminal device 100. Here, the terminal device 100 transmits various logs related to the content to the log server 102. On the other hand, the log server 102 is an information processing device realized by a server device and the like, and collects and collects various logs related to distributed contents from the terminal device 101, other terminal devices not shown, and the like. The log record information based on the information is transmitted to the information distribution device 10.

  Here, various logs related to content include, for example, whether or not the content has been browsed, whether or not the content has been selected such as being tapped, and a link from a predetermined web page to a content or a web page where the content is arranged. It is information such as whether it has been set. The various logs related to the content indicate whether or not the content was searched by using a web search engine or the like, whether or not information about the content was transmitted (tweet, etc.) by a short message service, a microblog, or the like. It is a log.

  When such a log is acquired, the log server 102 generates performance information for each content. For example, the log server 102 includes the number of times the content is viewed (the number of impressions), the number of times the content is selected, the CTR (Click Through Rate), the number of links set for the content, the number of times the content is searched, and information including the content. The number of transmissions and the like are transmitted to the information distribution device 10 as performance information. That is, the log server 102 transmits various performance information to the information distribution device 10 as variation information, which is information that changes according to the passage of time among the information related to the content.

  The types of logs collected by the log server 102 and the contents of the record information transmitted to the information distribution device 10 are not limited to the above-mentioned examples. That is, the log server 102 may transmit arbitrary information to the information distribution apparatus 10 as the record information, as long as the information changes as time passes.

  The client terminal 200 is a terminal device that registers content to be distributed in the information distribution device 10. For example, the client terminal 200 is an information processing device used by a client that generates, edits, and distributes contents, and is a mobile terminal such as a PDA, a PC, a server, or the like.

  The information distribution device 10 is an information processing device that stores the content registered from the client terminal 200 and, upon receiving a content distribution request from the terminal device 100, distributes the stored content to the terminal device 100. For example, the information distribution device 10 is realized by a physical or virtual information processing device such as a PC, a server, or a cloud system.

  Here, in the content distributed by the information distribution device 10, the importance according to the content, the time elapsed since the content was registered (that is, information freshness), etc. is estimated. Then, the information distribution device 10 preferentially distributes the content of which the estimated importance is high. As a result, the terminal device 100 preferentially displays the content having a higher priority.

[1-1. Information distribution processing]
Here, in the conventional technique, the initial value of the importance of the content is estimated based on the content of the content when the content is registered, and the estimated importance is attenuated over time. However, the contents to be distributed include not only contents such as news where importance is placed on information freshness, but also contents such as columns where importance is not placed on information freshness. In addition, among contents, there is also contents in which, after a certain amount of time has passed after being posted, as a result of being found worth, retweets and the like are frequently performed and the importance is increased. Therefore, when the importance of each content is calculated on a uniform basis, or when the importance is attenuated over time, the value of the calculated importance is regarded as the actual importance of each content. It will be separated.

  In addition, a method of using a model that calculates an initial value of importance in consideration of a change in importance over time when content is posted is possible. However, since the contents to be distributed vary in importance in different modes, it is difficult to learn a model that appropriately calculates the initial value of the importance of each content.

  Therefore, the information distribution device 10 executes the following estimation process. For example, the information distribution device 10 acquires predetermined information about the content as stationary information of the content. For example, the information distribution apparatus 10 may change the contents of information such as text, images, and audio included in the content, the provider such as the content distribution source, the author to which the content belongs, and the like with the passage of time. At least one of a small amount of information is acquired as article information. In addition, the information distribution device 10 is, if the information about the content is less changed with the passage of time as compared with the variable information described below, such as information that can be specified before the content is distributed. , Any information can be adopted as article information.

  Further, the information distribution device 10 acquires, as the variation information of the content, information that changes with the passage of time among the information related to the content, that is, the information whose value may change for each predetermined time period. For example, the information distribution device 10 acquires the record information collected by the log server 102 as the variation information. More specifically, the information distribution device 10 determines the number of times the content is viewed (the number of impressions), the number of times the content is selected, the number of links to the content, the number of times the content is searched, the number of times information about the content is transmitted, and the like. At least one of the above is acquired as performance information.

  Then, the information distribution device 10 uses the learning device Le that extracts the characteristic of the information from the input information to newly generate the characteristic information indicating the characteristic of the content at a predetermined time based on the steady information and the variation information. To extract. That is, the information distribution device 10 extracts the characteristic of the content at a predetermined time based on the constant information that is the predetermined information regarding the content and that changes less than the change information with the passage of time, and the change information. To do. Then, the information distribution device 10 estimates the importance of the content at a predetermined time based on the characteristic information.

  For example, when estimating the importance wn of the content at the time tn, the information distribution device 10 uses the learning device Le to create a set of the stationary information of the content and the variation information acquired up to the time tn. Based on this, the characteristic information yn indicating the characteristic of the content at the time tn is extracted, and the importance wn of the content at the time tn is estimated from the extracted characteristic information yn. In addition, when estimating the importance wn + 1 of the content at time tn + 1, the information distribution device 10 uses the learning device Le to create a set of the stationary information of the content and the variation information acquired up to time tn + 1. Based on this, the feature information yn + 1 indicating the feature of the content at the time tn + 1 is extracted, and the importance wn + 1 of the content at the time tn + 1 is estimated from the extracted feature information. More specifically, the information distribution device 10 extracts and extracts the characteristic information yn + 1 indicating the characteristic of the content at the time tn + 1 from the set of the characteristic information yn and the variation information acquired until the time tn + 1. The importance wn + 1 of the content at time tn + 1 is estimated from the characteristic information.

  As described above, the information distribution device 10 estimates the importance from the information that can be specified before the distribution of the content, and does not attenuate the importance with the passage of time, but at each time (or time zone). The feature of the content at that time is re-extracted based on the steady information of the content and the variation information up to that point. More specifically, the information distribution device 10 extracts the feature of the content at a predetermined time based on the steady information and the variation information.

  Therefore, the information distribution device 10 can improve the accuracy of estimating the importance of content. For example, it is considered that the number of impressions and the number of searches for content such as news where importance is attached to information freshness will decrease with time. On the other hand, with regard to contents such as columns where the information freshness is not so important, the number of impressions and the number of searches may be kept constant or may increase regardless of the passage of time.

  Therefore, the information distribution device 10 re-extracts the feature of the content at each time based on the feature of the steady information and the variable information. Then, the information distribution device 10 estimates the importance degree from the feature of the content re-extracted at each time. Therefore, the information distribution apparatus 10 can estimate the importance level that appropriately reflects the importance level, even if the index of the importance level is different for each content or the aspect of the change of the importance level over time is different. it can.

[1-2. Learner]
Here, the information distribution device 10 extracts the features of the stationary information and the variation information from the input information by using the learning device Le that has been learned to extract the features of the information, and extracts the features. The information indicating the selected characteristic is set as the characteristic information of the content. For example, the information distribution device 10 connects a plurality of nodes, such as a neural network, an auto encoder, and deep learning, that output the calculation result for the input data in multiple layers, and compresses the number of dimensions of the input information to input. The feature information is extracted using the learning device Le that extracts the features included in the extracted information.

  A more specific example will be described. For example, the information distribution device 10 performs learning by compressing the number of dimensions of the input information, extracting the characteristic information indicating the characteristic of the information, and restoring the input information from the extracted characteristic information. The automatic encoder is stored as a learning device Le. Such a learning device Le has an input layer that receives input information, an intermediate layer in which a smaller number of nodes are arranged than the input layer, and an output layer that outputs information restored from the feature information. Then, the node of each layer is connected to the node of the adjacent layer via a route in which a predetermined connection coefficient is set.

  Here, since the number of nodes in the intermediate layer is smaller than that in the input layer, the data output by the intermediate layer is data obtained by compressing the dimensionality of the data input to the learning device Le, and the output layer The data output by is the data reproduced from the compressed data of the number of dimensions, and has the same number of dimensions as the input data. Therefore, if the learning device Le can compress the number of dimensions of the input information so as not to lose the characteristics of the input information, the information input to the learning device Le and the information restored by the learning device Le are , The same or similar information is considered.

  Therefore, when the set of the steady information, the record information, and the time information indicating the time is input, the information distribution device 10 extracts the information in which the number of dimensions of the input information set is compressed, and the information is input from the extracted information. The learner Le whose connection coefficient is modified so as to reproduce the set of the obtained information is acquired. Such a learning device Le extracts, in the middle layer, information that is implicitly necessary to reproduce the set of input information, that is, the feature information indicating the feature of the set of the steady information, the record information, and the time information. Will be done.

  For this reason, the information distribution device 10 converts the stationary information, the record information, and the time information into a distributed expression that is multidimensional information, and sets the set of the distributed expressions of the stationary information, the record information, and the time information as the learning device Le. To enter. Then, the information distribution device 10 indicates the information output by the middle layer of the learning device Le, which is the characteristic of the set of the steady information and the record information, that is, the characteristic of the content, at the time or time zone indicated by the time information. Acquire as feature information.

[1-3. About RNN (Recurrent Neural Networks)]
Here, the feature information extracted by the learning device Le from the set of the steady information, the record information, and the time information reflects the feature of the steady information, the feature of the record information, and the feature of the time information. Therefore, the information distribution device 10 selects a predetermined time from the set of the feature information extracted as the information indicating the feature of the content at the time before the predetermined time and the variation information at the time before the predetermined time. The characteristic information of the content in is extracted.

  For example, the information distribution apparatus 10 extracts from the characteristic information extracted as the characteristic information indicating the characteristic of the content at the time “tn-1”, that is, the previously extracted characteristic information “yn-1” and the time “tn-1”. The characteristic information “yn” of the content at the time “tn” is extracted by inputting the record information up to the time “tn” and the time information indicating the time “tn” to the learning device Le. Then, the information distribution device 10 estimates the importance “wn” of the content at the time “tn” based on the extracted characteristic information “yn”.

  When executing the above-described processing, for example, when the information distribution device 10 extracts the characteristic information “yn” indicating the characteristic of the content at the time “tn”, the extracted characteristic information “yn” is stored in a predetermined storage device. By inputting the held characteristic information “yn”, the held performance information from the time “tn” to the time “tn + 1”, and the time information indicating the time “tn + 1” to the learning device Le, the time “ The feature information “yn + 1” of the content at “tn + 1” may be extracted. Further, the information distribution device 10 uses the learning device Le, that is, the RNN, which has a storage layer that temporarily stores the information output by the intermediate layer and transmits the stored information to the input layer, thereby creating a new content. The characteristic information may be extracted.

  For example, when the learning device Le realized by the RNN extracts the characteristic information “yn” of the content at the time “tn” in the intermediate layer, the learning device Le outputs the extracted characteristic information “yn” and holds it in the storage layer. Further, the information distribution device 10 inputs, to the learning device Le, the performance information from the time “tn” to the time “tn + 1” and the time information indicating the time “tn + 1” to the learning device Le. In such a case, the learning device Le, the information accumulated by the accumulation layer, that is, the characteristic information “yn” of the content at the time “tn” and the performance information from the time “tn” to the time “tn + 1”, The characteristic information “yn + 1” of the content at the time “tn + 1” is extracted by using the time information indicating the time “tn + 1” as the input information. Therefore, the information distribution device 10 can extract the characteristic information of the new content by simply inputting the new record information at any time.

[1-4. About estimation method]
Further, the information distribution device 10 can estimate the importance of the content by an arbitrary method when estimating the importance of the content from the extracted feature information of the content. For example, the feature information of the content extracted using the learning device Le is extracted as multidimensional information. Therefore, the information distribution apparatus 10 estimates the feature information indicating the feature of the content from the extracted feature information using a model that estimates the feature indicated by the information of each dimension included in the feature information, and determines the content from the estimated feature information. The importance of content may be estimated using a model for estimating the importance of.

  For example, the information distribution device 10 generates a model for estimating what kind of feature the information of each dimension included in the feature information shows. Then, the information distribution device 10 may use the generated model to estimate the feature information from the feature information and estimate the importance of the content based on the estimated feature information. Further, the information distribution device 10 may use a learning device such as a neural network or a DNN that has been learned so as to estimate the feature information from the feature information and to estimate the importance of the content from the estimated feature information.

  Hereinafter, with reference to FIG. 1, an example of an estimation process performed by the information delivery apparatus 10 that performs the above-described process and an example of an information delivery process will be described.

[1-5. Example of estimation processing]
First, the estimation process executed by the information distribution device 10 will be described. For example, the information distribution device 10 accepts posting of content from the client terminal 200 (step S1). More specifically, the information distribution device 10 accepts, together with the content of the content, article information “x” including various information indicating content attributes such as a content distribution source, a category to which the content belongs, a content author, and the like.

  In such a case, the information distribution device 10 stores the article information and uses the techniques such as tf-idf (Term Frequency-Inverse Document Frequency), Key-Value method, W2V (Word 2 Vector), and the like, and the article information “x ] To generate multidimensional information. That is, the information distribution device 10 generates the distributed expression "y0" of the article information "x" which is the stationary information (step S2).

  Next, the information distribution device 10 collects various performance information regarding the content such as the number of impressions and the number of selections of each content from the log server 102 (step S3). For example, the information distribution device 10 collects the record information “h1” of the content at the time “t1”.

  Then, the information distribution device 10 inputs the distributed expression “y0”, the time information “t1” indicating the time “t1”, and the actual result information “h1” of the content to the learning device Le, and thus the time “t1”. The feature information “y1” indicating the feature of the content is generated (step S4). Further, the information distribution device 10 estimates the feature information “z1” indicating the feature of the content at the time “t1” from the feature information “y1” by using a predetermined model, a learning device, or the like, and estimates the feature information “z1”. , The estimated importance “w1” which is the importance of the content at the time “t1” is estimated (step S5).

  Further, the information distribution device 10 newly regenerates the characteristic information based on the previously extracted characteristic information and the newly collected performance information for each predetermined time period, and estimates the estimated importance degree each time. (Step S6). More specifically, the information distribution device 10 collects the performance information “h2” of the content from the time “t1” to the time “t2”, and the time information “t2” indicating the time “t2” and the performance information. By inputting “h2” and the previously extracted specific information “y1” to the learning device Le, the characteristic information “y2” indicating the characteristic of the content at the time “t2” is generated. Further, the information distribution device 10 estimates the feature information “z2” indicating the feature of the content at the time “t2” from the feature information “y2” by using a predetermined model, a learning device, or the like, and estimates the feature information “z2”. , The estimated importance level “w2” that is the importance level of the content at the time “t2” is estimated.

  Similarly, the information distribution device 10 collects the performance information “h3” of the content from the time “t2” to the time “t3”, and acquires the time information “t3”, the performance information “h3”, and the previously extracted identification. By inputting the information “y2” to the learning device Le, the characteristic information “y3” indicating the characteristic of the content at the time “t3” is generated. Further, the information distribution device 10 estimates the feature information “z3” indicating the feature of the content at the time “t3” from the feature information “y3”, and determines the importance of the content at the time “t3” from the estimated feature information “z3”. The estimated importance “w3”, which is the degree, is estimated.

  That is, the information distribution device 10 has the performance information “hn” up to the predetermined time “tn” and the previously calculated characteristic information, that is, the characteristic information “yn-1” indicating the characteristic of the content at the time “tn−1”. Based on and, the feature information “yn” indicating the feature of the content at the time “tn” is extracted. Then, the information distribution device 10 estimates the feature information “zn” indicating the feature of the content at the time “tn” from the extracted feature information “yn”, and at the time “tn” from the estimated feature information “zn”. The estimated importance “wn” which is the importance of the content is estimated.

[1-6. Example of information distribution process]
Next, the information distribution process executed by the information distribution device 10 will be described. For example, the information distribution device 10 receives a content distribution request from the terminal device 100 (step S7). In such a case, the information distribution device 10 selects the content to be distributed from the attributes of the user U01, the attribute of the content, and the like, and preferentially distributes the content having a high estimated importance among the selected contents. (Step S8). For example, the information distribution device 10 distributes the selected content in the order of the estimated highest estimated importance at the end. As a result, the terminal device 100 can allow the user U01 to browse the content in an appropriate order by preferentially displaying the content having a high estimated importance (step S9).

[2. Configuration of information distribution device]
Next, the configuration of the information distribution device 10 according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram illustrating an example of a functional configuration of the information distribution device according to the embodiment. As shown in FIG. 2, the information distribution device 10 includes a communication unit 11, a storage unit 12, and a control unit 13. The communication unit 11 is realized by, for example, a NIC (Network Interface Card) or the like. The communication unit 11 is connected to the network N in a wired or wireless manner, and transmits / receives information to / from the terminal device 100, the log server 102, and the client terminal 200.

[2-1. Database stored in storage unit]
The storage unit 12 is realized by, for example, a semiconductor memory device such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk. The storage unit 12 also stores a content database 14, an achievement information database 15, a learning device database 16, a feature information database 17, an estimated model database 18, and an importance database 19.

  In the content database 14, the content accepted from the client terminal 200 is registered. For example, FIG. 3 is a diagram showing an example of information registered in the content database according to the embodiment. As shown in FIG. 3, the content database 14 has items such as “content ID”, “contents content”, “distribution source”, “category”, “author”, and “distributed expression”. In addition to the information shown in FIG. 3, arbitrary information that can be specified at the time of content registration, that is, arbitrary stationary information, may be registered in the content database 14.

  The “content ID” is information that identifies the registered content. Further, the “contents content” is information such as text included in the content. The "contents" may include information other than text, such as moving images, images, and sounds. The “delivery source” is information indicating a content providing source, a content producing source, and a registration source that has registered the content in the information delivery apparatus 10. When the content is an advertisement or the like, the “delivery source” is information indicating an advertisement provider, an advertisement provider such as an advertisement agency, a company or an individual who advertises in the advertisement.

  The “category” is information indicating a category to which the content belongs, such as news and columns. The “author” is information indicating the creator of the content. The “distributed expression” is information extracted from the content and is multidimensional information indicating the content of the content. More specifically, the “distributed expression” is multidimensional information indicating the content of article information, which is the constant information of the content, such as “contents content”, “distribution source”, “category”, “author”, and the like. ..

  For example, in the example shown in FIG. 3, the content ID “ID1”, the content content “content A”, the distribution source “distribution source A”, the category “category A”, the author “author A”, and the distributed expression “distributed expression A” are Corresponding and registered. In this information, the content of the content indicated by the content ID “ID1” is “content A”, the distribution source is “distribution source A”, the author is “author A”, and belongs to the category “category A”. Indicate. Further, this information indicates that the distributed expression of the article information of the content indicated by the content ID "ID1" is "distributed expression A".

  Returning to FIG. 2, the achievement information database 15 stores the achievement information of each content. For example, FIG. 4 is a diagram showing an example of information registered in the performance information database according to the embodiment. As shown in FIG. 4, the record information database 15 has items such as “content ID”, “date and time”, “impression count”, “click count”, and “tweet count”. It should be noted that, in addition to the information shown in FIG. 4, information that changes with time, that is, arbitrary variation information, may be registered in the performance information database 15 in addition to the information shown in FIG.

  The “date and time” indicates the time zone in which the associated variation information is collected. In addition, the “impression number” indicates the number of times the content is browsed in the time period indicated by the associated “date and time”. The “click count” indicates the number of times the content is selected in the time zone indicated by the associated “date and time”. Further, the “tweet number” indicates the number of times the information regarding the content is transmitted in the time period indicated by the associated “date and time”.

  For example, in the example shown in FIG. 4, the content ID “ID1”, the date and time “2015/1/1/00: 00 to 00:59”, the number of impressions “1000”, the number of clicks “300”, and the number of tweets “500” are displayed. Corresponding and registered. Further, in the example shown in FIG. 4, the content ID “ID1”, the date and time “2015/1/1/01: 00 to 01:59”, the number of impressions “300”, the number of clicks “10”, and the number of tweets “5” are displayed. It is registered.

  As for this information, the content indicated by the content ID “ID1” is browsed “1000” times and selected “300” times during the date and time “2015/1/1/00: 00 to 00:59”. Indicates that the message has been tweeted 500 times. Also, regarding this information, the content indicated by the content ID “ID1” is browsed “300” times and selected “10” times during the date and time “2015/1/1/01: 00 to 01:59”. , Indicates that it has been tweeted “5” times. As described above, the content indicated by the content ID “ID1” is considered to be the content whose information freshness is important such as news, because the number of impressions, the number of clicks, and the number of tweets are all significantly decreased with the passage of time. Be done.

  On the other hand, in the example shown in FIG. 4, the content ID “ID2”, the date and time “2015/1/1/00: 00 to 00:59”, the number of impressions “300”, the number of clicks “50”, and the number of tweets “5” are displayed. Corresponding and registered. Further, in the example shown in FIG. 4, the content ID “ID2”, the date and time “2015/1/1/01: 00 to 01:59”, the number of impressions “350”, the number of clicks “60”, and the number of tweets “50” are It is registered.

  As for this information, the content indicated by the content ID “ID2” is browsed “300” times and selected “50” times during the date and time “2015/1/1/00: 00 to 00:59”. Indicates that it was tweeted 5 times. In addition, regarding this information, the content indicated by the content ID “ID2” is browsed “350” times and selected “60” times during the date and time “2015/1/1/01: 00 to 01:59”. , Indicates that it has been tweeted “50” times. As described above, the content indicated by the content ID “ID2” has become more popular with the lapse of time, and the number of impressions, the number of clicks, and the number of tweets have increased, so that the information freshness is not so important. Conceivable.

  Returning to FIG. 2, in the learning device database 16, the learning device Le used when the information distribution device 10 extracts the characteristic information of the content based on the constant information and the variation information of the content is registered. More specifically, the learning device database 16 is a learning device in which a plurality of nodes that output calculation results for input data are connected in multiple layers, and information is stored by compressing the number of dimensions of input information. The connection coefficient set in the path connecting the nodes of the learning device Le that extracts the characteristic information indicating the characteristic is registered.

  For example, FIG. 5 is a diagram illustrating an example of information registered in the learning device database according to the embodiment. As shown in FIG. 5, the learning device database 16 has items such as “neuron ID”, “connection destination neuron ID”, and “connection coefficient”. The “neuron ID” indicates identification information for identifying a neuron that is a node included in the learning device Le. The "connection destination neuron ID" indicates identification information for identifying another neuron that is connected to the neuron corresponding to the "neuron ID". The "connection coefficient" indicates the connection coefficient set in the path connecting the neuron indicated by the "neuron ID" and the neuron indicated by the "connection destination neuron ID".

  For example, in the example shown in FIG. 5, in the learning device database 16, the neuron ID “N11”, the neuron ID “N21”, and the connection coefficient “WA” are registered in association with each other. This information indicates that the neuron corresponding to the neuron ID “N11” is connected to the neuron corresponding to the neuron ID “N21” through the route in which the connection coefficient “WA” is set. Therefore, the value corresponding to the output of the neuron ID “N11” multiplied by the connection coefficient “WA” is transmitted as an input to the neuron corresponding to the neuron ID “N21”.

  Returning to FIG. 2, the characteristic information database 17 stores the characteristic information of each content in each time period. For example, FIG. 6 is a diagram showing an example of information registered in the characteristic information database according to the embodiment. As shown in FIG. 6, in the characteristic information database 17, “content ID”, “use date / time”, and “characteristic information” are registered in association with each other.

  Here, the "use date and time" is information indicating in which time zone the feature indicated by the associated feature information is the feature of the content, and is the record used when extracting the associated feature information. It is the time zone when the information was collected. Further, the “feature information” is feature information of the content in the time zone next to the time zone indicated by the associated use date and time.

  For example, in the example shown in FIG. 6, the feature information database 17 is associated with the content ID “ID1”, the usage date and time “2015/1/1/00: 00 to 00:59”, and the feature information “feature information y1”. Have been registered. This information is the characteristic information of the content indicated by the content ID "ID1", and is the characteristic extracted based on the performance information collected during "2015/1/1/00: 00-00: 59". It indicates that the information is “feature information y1”.

  Returning to FIG. 2, the estimation model database 18 estimates the feature information of the content from the feature information of the content and estimates the estimated importance degree from the estimated feature information. For example, in the estimation model database 18, when the information distribution device 10 uses a neural network similar to the learning device Le as a model, the same information as the learning device database 16 illustrated in FIG. 5 is registered. Note that the estimation model database 18 is registered with information about a model in any format as long as it is a model for estimating the feature information of the content from the feature information of the content and estimating the estimated importance degree from the estimated feature information. You can stay.

  The importance database 19 stores the importance of each content in each time zone. For example, FIG. 7 is a diagram showing an example of information registered in the content database according to the embodiment. As shown in FIG. 7, “content ID”, “use date / time”, and “estimated importance” are registered in the importance database 19 in association with each other. Here, the “estimated importance” is a value indicating the importance of the feature information of the content.

  For example, in the example shown in FIG. 7, the importance database 19 stores the content ID “ID1”, the use date and time “2015/1/1/00: 00 to 00:59”, and the estimated importance “estimated importance w1”. Corresponding and registered. This information is the characteristic information of the content indicated by the content ID "ID1", and is the characteristic extracted based on the performance information collected during "2015/1/1/00: 00-00: 59". It indicates that the estimated importance estimated from the information is “estimated importance w1”.

[2-2. Example of configuration of control unit]
Returning to FIG. 2, the description will be continued. The control unit 13 is stored in a storage device inside the information distribution device 10 by, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or the like. It is realized by executing various programs using the RAM as a work area.

  As illustrated in FIG. 2, the control unit 13 includes a reception unit 20, a collection unit 21, a generation unit 22, an extraction unit 23, an estimation unit 24, and a distribution unit 25, and has functions and actions of information processing described below. Realize or execute. Note that the internal configuration of the control unit 13 is not limited to the configuration shown in FIG. 2, and may be another configuration as long as it is a configuration for performing information processing described later. The connection relationship between the processing units included in the control unit 13 is not limited to the connection relationship illustrated in FIG. 2 and may be another connection relationship.

  The reception unit 20 receives registration of content from a client. For example, the accepting unit 20 accepts regular information that is predetermined information about the content such as the content of the content, information indicating a distribution source, a category, an author, and the like. In such a case, the reception unit 20 registers the regular information of the received content in the content database 14.

  The collection unit 21 collects variation information that changes with time, out of the information about the content. For example, the collection unit 21 accesses the log server 102 and collects performance information regarding the content such as the number of impressions, the number of clicks, and the number of tweets. Then, the collection unit 21 registers the generated achievement information as variation information in the achievement information database 15 in association with the time period when the logs are collected.

  The generation unit 22 generates a distributed expression indicating stationary information of content. For example, when new content is registered, the generation unit 22 uses multi-dimensional information (for example, a vector) from the stationary information of the newly registered content by using a technique such as tf-idf or W2V. Generate a distributed representation. After that, the generation unit 22 registers the generated distributed expression in the content database 14 in association with the content ID of the corresponding content.

  The extraction unit 23 extracts the characteristic information indicating the characteristic of the content at a predetermined time using the learning device Le that extracts the characteristic from the input information, based on the stationary information of the content and the variation information of the content. More specifically, the extraction unit 23 selects, from a set of feature information extracted as information indicating a feature of content at a time earlier than a predetermined time and variation information at a time earlier than the predetermined time, Feature information indicating the feature of the content at a predetermined time is extracted.

  For example, the extraction unit 23 refers to the characteristic information database 17 and identifies the content for which the characteristic information is not registered. Subsequently, the extraction unit 23 acquires the distributed expression of the specified content from the content database 14 and inputs the acquired distributed expression to the learning device Le, thereby indicating the characteristic of the distributed expression, that is, the characteristic of the content. Extract feature information. Then, the extraction unit 23 registers the extracted characteristic information in the characteristic information database 17.

  Further, when extracting the characteristic information indicating the characteristic of the content at the predetermined time, the extraction unit 23 reads the variation information at the time before the predetermined time from the record information database 15. More specifically, the extraction unit 23 reads the variation information in the time zone after the characteristic information indicating the characteristic of the content is extracted from the record information database 15. The extraction unit 23 also reads the previously extracted characteristic information from the characteristic information database 17. Then, the extraction unit 23 extracts the characteristic information indicating the characteristic of the content at a predetermined time based on the read variation information and the characteristic information, and registers the extracted characteristic information in the characteristic information database 17.

Hereinafter, an example of the process executed by the extraction unit 23 will be described. In the following explanation,
An example will be described in which the date and time included in the date and time “2015/1/1/02: 00-02: 59” is “t2”. For example, the extraction unit 23 determines that the performance information collected by the date and time “t2”, that is, the feature information “y2” extracted by using the use date and time “2015/1/1/01: 00: 00 to 01:59”. The unregistered content is specified from the characteristic information database 17.

  In addition, the extraction unit 23, among the performance information of the specified contents, the performance information collected by the date and time “t2”, that is, between the date and time “2015/1/1/01: 00: 00-01: 59”. The collected record information “h2” is acquired from the record information database 15. In addition, the extraction unit 23 extracts the previously extracted characteristic information among the characteristic information of the specified contents, that is, the record information “collected between the dates and times“ 2015/1/1/00: 00 to 00:59 ”. The feature image method “y1” extracted based on “h1” is extracted from the feature information database. Then, the extraction unit 23 inputs the set of the characteristic information “y1”, the record information “h2”, and the time information “t2” to the learning device Le, so that the characteristic information ““ indicating the characteristic of the content at the time “t2” is displayed. y2 ”is extracted.

  When extracting new characteristic information based on the previously extracted characteristic information, the extraction unit 23 may extract the characteristic information using the RNN. For example, FIG. 8 is a diagram illustrating an example of a process in which the information distribution device according to the embodiment extracts feature information. For example, when the content is registered, or until some time elapses after the registration, it is considered that the record information of the content does not exist. Therefore, the extraction unit 23 inputs the distributed expression “y0” of the content to the learning device Le as illustrated in (A) of FIG.

  In such a case, the learning device Le compresses the number of dimensions of the distributed representation input to the intermediate layer and restores the original distributed representation from the information output by the intermediate layer, that is, the compressed representation of the number of dimensions. .. In such a learning device Le, by modifying the value of the connection coefficient so that the input distributed expression and the restored distributed expression are the same, the characteristic of the distributed expression, that is, the content characteristic information Will be extracted accurately.

  Therefore, as shown in (B) in FIG. 8, the extraction unit 23 outputs the information output by the middle layer of the learning device Le when the distributed expression “y0” of the content is input to the learning device Le, at the date and time “t1”. It is assumed to be the characteristic information “y1” of the content in the time zone indicated by. Then, the extraction unit 23 registers the characteristic information “y1” in the characteristic information database 17.

  Here, when the learning device Le is realized by the RNN, the feature information “y1” is held in the storage layer connected to the node of the input layer. Then, as shown in (C) of FIG. 8, when the extraction unit 23 inputs the time information “t2” and the performance information “h2” of the content into the learning device Le, the characteristic information “held in the storage layer” “Y1” is input to the input layer of the learning device Le together with the time information “t2” and the performance information “h2”. As a result, the learning device Le extracts, as the characteristic information “y2”, the characteristic of the set of the characteristic information “y1”, the time information “t2”, and the performance information “h2”, as shown in (D) of FIG. To do.

  Returning to FIG. 2, the description will be continued. The estimation unit 24 estimates the importance of the content at a predetermined time based on the extracted characteristic information at the predetermined time. For example, the estimation unit 24 reads the model from the estimation model database 18. Further, the estimation unit 24 identifies the latest feature information from the feature information database 17 for each content, estimates the feature information indicating the feature of the content from the identified feature information using the read model, and the feature information. Therefore, the estimated importance level of the content at the predetermined time is estimated by using the model for estimating the importance level of the content at the predetermined time. Then, the estimation unit 24 registers the estimated estimated importance and the use date and time associated with the specified characteristic information in the importance database 19 in association with each other. Note that, for example, the estimation unit 24 may estimate the estimated importance degree by compressing the number of dimensions of the input information and inputting the feature information or the feature information to the learning device that extracts the feature.

  The distribution unit 25 preferentially distributes the content to the user U01 from the content having the higher estimated importance. For example, when the distribution unit 25 receives the distribution request, the distribution unit 25 performs targeting based on the regular information of the content database 14 and the attribute of the user U01, and selects the content to be distributed.

  Further, the distribution unit 25 specifies the estimated importance degree when the distribution request of the selected content is received from the importance degree database 19, and distributes the selected content to the terminal device 100 in the order of the specified estimated importance degree. For example, in the time zone of date / time “2015/1/1/02: 00-02: 59”, the importance associated with the use date / time “2015/1/1/01: 00: 00-01: 59” is the latest. Is the estimated importance of. Therefore, when the distribution unit 25 receives the distribution request in the time zone of “2015/1/1/02: 00: 00 to 02:59” on the date and time “2015/1/1/02: 00 to 02:59”, the use date / time “2015/1/1/01: 00: 00”. The latest estimated importance associated with “01:59” is specified, and the content is distributed based on the specified estimated importance.

  Note that the distribution unit 25 distributes the content and the identified estimated importance to the terminal device 10 together, and controls the terminal device 100 so that the content having the higher estimated importance is displayed preferentially. The content with high frequency may be preferentially provided to the user U01.

[2-3. Example of display on terminal device 100]
Next, with reference to FIG. 9, an example of a process in which the terminal device 100 displays the content with the priority according to the estimated importance will be described. FIG. 9 is a diagram illustrating an example of processing in which the terminal device according to the embodiment displays content according to priority. In the example shown in FIG. 9, an example of a screen displayed by the terminal device 100 when news D and news E are distributed after news A, news B, and column C are distributed as contents is described. .. Further, in the following description, the terminal device 100 is supposed to display up to four contents.

  For example, in the terminal device 100, the news A whose posting date is “2015/1/1/10: 00”, the news B whose posting date is “2015/1/1/09: 00”, and the posting date are “ Column C “2015/1/1/11: 00” will be delivered. Here, when the estimated importance degree is calculated such that the importance degree of news is higher than that of the column and the importance degree of the content whose post date is newer is higher, for example, the information distribution device 10 estimates the news A. It is assumed that the importance is “100”, the estimated importance of news B is “90”, and the estimated importance of column C is “30”. As a result, as shown in the first state of FIG. 9, the terminal device 100 displays each content in the order of estimated importance of each content, that is, news A, news B, and column C. ..

  Here, information freshness is an important feature in contents such as news. Therefore, when the information freshness of the news A and the news B decreases, the information distribution device 10 reduces the estimated importance of the news A and the news B. On the other hand, in the content of the column or the like, the information freshness is not a very important feature, and therefore the information distribution apparatus 10 does not reduce the importance of the column C so much even if time passes. In addition, the information distribution device 10 newly recalculates the degree of importance of the column C based on the performance information when, for example, the number of impressions in the column C increases after a lapse of a predetermined time after being posted. ..

  Therefore, for example, the information distribution device 10 sets the estimated importance of news A to “80”, the estimated importance of news B to “60”, and the estimated importance of column C to “50”. In addition, the information distribution device 10 sets the estimated importance degree of the new news D to “100”. As a result, as shown in the second state of FIG. 9, the terminal device 100 displays news D, news A, news B, and column C in this order.

  Here, if the column C becomes a popular content as a result of the increase in the number of impressions and the number of tweets in the column C, the information distribution device 10 recalculates the estimated importance degree of the column C to a higher value. For example, the information distribution device 10 recalculates the estimated importance degree of column C from “50” to “70”. On the other hand, the content of news A, B, D, etc., has the number of impressions and the number of tweets decreased over time. Therefore, the information distribution device 10 sets the estimated importance of news A to “50”, the estimated demand degree of news B to “40”, and the estimated importance degree of news D to “80”. Further, the information distribution device 10 sets the estimated importance degree of the new news E to “100”.

  As a result, the terminal device 100 displays the contents in the order of news E, news D, column C, and news A, as shown in the third state of FIG. That is, the terminal device 100 cancels the display of the news B, and gives priority to the column C, which is the content that has become popular after the passage of time, over the news A whose estimated importance is higher than that of the column C at the time of posting. indicate.

  In this way, the information distribution device 10 recalculates the estimated importance level of the content for each time period based on the stationary information and the time information. Therefore, for example, the information distribution apparatus 10 determines that the content to be distributed is a content whose information freshness is important, a content whose information freshness is not important, a content which becomes popular after a lapse of time after posting, a certain content. Even when the importance of the content changes in various ways, such as the content that keeps popular, the content that gradually decreases in popularity, etc., the estimated importance can be made to follow the true importance.

  That is, the importance of content varies in different ways over time. Therefore, when the estimated importance of the content is calculated using only the stationary information as in the related art, the estimated importance cannot follow the true importance. Further, when the importance is estimated using only the variation information such as the browsing record as in the conventional technique, the importance of the content immediately after the submission for which the sufficient browsing record is not obtained cannot be accurately estimated.

  However, since the information distribution device 10 calculates the estimated importance level of the content based on the stationary information and the variation information, it is possible to estimate the importance level of the content immediately after the submission. Further, when the variation information is acquired, the information distribution device 10 recalculates the estimated importance of the content based on the acquired variation information and the steady information. As a result, the information distribution device 10 can compare the importance of the content immediately after submission and the content for which the variation information can be acquired. It is possible to realize the distribution of contents according to the actual importance at time.

[2-4. Example of learner and model]
Next, an example of the learning device Le and the model used by the information distribution device 20 will be described using mathematical expressions. In the following description, the performance information at the time indicated by the time information “k” will be described as “hk”, the distributed expression or feature information as “yk”, the feature information as “zk”, and the estimated importance as “wk”. .. Moreover, in the following description, the article information is described as “x”. Note that “x” is a sparse vector indicating article information.

  For example, in the case of “k = 0”, the distributed expression “y0” of the content is calculated from the article information “x” by the following formula (1). Here, “A” and “a” shown in Expression (1) are predetermined matrices. That is, the information distribution device 10 inputs the value obtained by adding the matrix “a” to the product of the matrix “A” and the article information “x” to the function represented by the equation (2), and thus the distributed expression “y0”. To calculate.

  Then, in the case of “k = 1”, that is, when the first record information “h1” is collected, the information distribution device 20 displays the distributed expression “y0”, the time information “t1”, and the record information. The feature information “y1” will be extracted from the set of “h1”. In the case of “k ≧ 2”, the information distribution device 20 selects the characteristic information “yk” from the previously extracted set of the characteristic information “yk−1”, the time information “k”, and the performance information “hk”. Extract.

  Specifically, the information distribution device 20 will extract the characteristic information “yk” using the following equation (3). Here, “B” and “b” shown in Expression (3) are predetermined matrices. That is, the information distribution device 10 sets the transpose matrix of the matrix in which the transposed matrix of the feature information “yk−1”, the transposed matrix of the time information “k”, and the transposed matrix of the performance information “hk” are arranged. The feature information “yk” is calculated by inputting the sum of the value accumulated in “B” and the matrix “b” into the function shown in Expression (2).

  Subsequently, the information distribution device 20 calculates the feature information “zk” at the time “k” by using the equation (4) shown below. Here, “C” and “c” shown in Expression (4) are predetermined matrices. That is, the information distribution device 10 inputs the value obtained by adding the matrix “c” to the product of the matrix “C” and the feature information “yk” to the function shown in Expression (5), and thus at the time “k”. The feature information “zk” is calculated.

  Subsequently, the information distribution device 20 calculates the estimated importance level “wk” at the time “k” from the feature information “zk” at the time “k” using the formula (6) shown below. Here, “D” and “d” shown in Expression (6) are predetermined matrices. That is, the information distribution device 10 calculates a value obtained by adding the matrix "d" to the product of the matrix "D" and the feature information "zk" as the estimated importance "wk".

  A process of correcting the connection coefficient of the learning device Le that extracts the feature information based on the stationary information and the variation information, and a model that estimates the estimated information from the feature information and estimates the estimated importance degree from the estimated information The learning process is a process of adjusting the matrices “A” to “D” and the matrices “a” to “d” included in the above-described formulas (1) to (6) to appropriate values. Therefore, the information distribution device 20 learns the values of the matrices “A” to “D” and the matrices “a” to “d” using a method such as stochastic gradient descent (SGD). As a result, the estimation accuracy of the estimation importance can be improved.

[3. Procedure of processing executed by information distribution device]
Next, an example of the flow of processing executed by the information distribution device 10 according to the embodiment will be described with reference to FIGS. 10 to 12. First, with reference to FIG. 10, a flow of processing executed when the posting of content is accepted will be described. FIG. 10 is a flowchart showing an example of the flow of processing in which the information distribution apparatus according to the embodiment accepts posting of content. For example, the information distribution device 10 accepts posting of content from the client terminal 20 (step S101). In such a case, the information distribution device 10 generates a distributed expression indicating the content of the stationary information of the content (step S102), and ends the process.

  Next, the flow of processing for estimating the estimated importance level of content will be described with reference to FIG. FIG. 11 is a flowchart showing an example of the flow of an estimation process performed by the information distribution device according to the embodiment. For example, the information distribution device 10 collects the record information of each registered content (step S201).

  Subsequently, the information distribution device 10 determines whether or not the previously extracted characteristic information exists for each content (step S202). Then, when the previously extracted characteristic information does not exist (step S202: No), the information distribution device 10 extracts the characteristic information of a set of the distributed representation of the stationary information of the content and the record information (step S203). On the other hand, when the previously extracted characteristic information is present (step S202: Yes), the information distribution device 10 extracts the characteristic information of the set of the previously extracted characteristic information and the performance information (step S204).

  Further, the information distribution device 10 estimates the feature information indicating the feature of the content from the feature information extracted in step S203 or step S204 (step S205), and estimates the estimated importance degree from the feature information (step S206). Then, the information distribution device 10 registers the estimated importance degree (step S207) and ends the process.

  Next, an example of the flow of information distribution processing executed by the information distribution device 10 according to the embodiment will be described with reference to FIG. FIG. 12 is a flowchart showing an example of information distribution processing by the information distribution device according to the embodiment. For example, the information distribution device 10 determines whether or not a distribution request has been received from the terminal device 100 (step S301). Then, when the distribution request has not been received (step S301: No), the information distribution device 10 waits until the reception.

  On the other hand, when the information distribution device 10 receives the distribution request (step S301: Yes), the information distribution device 10 selects the content that is a candidate for distribution (step S302). Further, the information distribution device 10 identifies the estimated importance degree of the selected content (step S303). Then, the information distribution device 10 distributes the content and the estimated importance to the terminal device 100, displays each content in the order of priority according to the estimated importance (step S304), and ends the process.

[5. Modification example)
The information distribution device 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 information distribution device 10 will be described.

[5-1. Learner]
In the above description, the information distribution device 10 extracts the characteristic information using the learning device Le, which is an auto encoder having three layers of an input layer, an intermediate layer, and an output layer. However, the embodiment is not limited to this. For example, the information distribution device 10 may extract the characteristic information by using the learning device Le having a plurality of intermediate layers, that is, the learning device Le having an arbitrary number of layers.

  Here, when the learning device Le has a plurality of intermediate layers, the intermediate layer having the smallest number of nodes, that is, the intermediate layer having the smallest number of dimensions of information to be output, directly shows the characteristics of the input information. It is believed that Therefore, when using the learning device Le having a plurality of intermediate layers, the information distribution apparatus 10 may use the information output by the intermediate layer having the smallest dimension number of the output information as the characteristic information.

  Further, the information distribution device 10 may execute the above-described information distribution process using any learning device Le other than the automatic encoder. In addition to the neural network and the deep learning, the information distribution device 10 can extract feature information that appropriately indicates the feature of the content in a predetermined time zone based on the steady information and the variation information. , An arbitrary model may be learned, and the estimation process described above may be executed using the model.

  Further, the information distribution device 10 estimates the feature information from the feature information and the learning device Le that extracts the feature information indicating the feature of the content at a predetermined time based on the steady information and the variation information, and estimates from the feature information. The model for estimating the importance may be realized by one learning device Le or a model. The learning device Le and the model as described above realize the processes represented by the above-described formulas (1) to (6).

  In the above description, the information delivery device 10 inputs the set of the previously extracted characteristic information, time information, and performance information to the learning device Le. When such an input is performed, it is considered that the learning device Le outputs the feature information reflecting the feature based on the difference between the last time information included in the previously extracted feature information and the newly input time information. Be done. However, the embodiment is not limited to this, and for example, the information distribution device 10 may input the set of the previously extracted characteristic information and the record information to the learning device Le. That is, the information distribution device 10 may use the learning device Le in any manner as long as the feature of the content is extracted each time based on the set of the stationary information and the variation information.

[5-2. Regarding fluctuation information]
In the above description, the information distribution device 10 has extracted the characteristic information based on the steady information and the variation information. Here, when the time interval at which the information distribution device 10 estimates the estimated importance is shorter than the time interval at which the record information is collected, the variation information may not exist in the initial estimation process. Even in such a case, the information distribution device 10 may estimate the feature information from the steady information and estimate the estimated importance degree from the estimated feature information by executing the above-described estimation process.

  Further, when the information distribution device 10 receives the distribution request for the content, the information distribution device 10 collects the record information of the content to be distributed, and a set of the collected record information and the steady information of the content, or the extracted feature information. It is also possible to extract new characteristic information from the set of the collected performance information and estimate the estimated importance degree from the extracted characteristic information. That is, the information distribution device 10 may execute the above-described estimation process at any timing.

  Further, the information distribution device 10 may use the variation information of the content in any mode. For example, in the above-described example, when the characteristic information “yk” is extracted, the characteristic information is extracted based on the record information between “tk−1” and “tk”. However, the information distribution device 10 extracts the characteristic information based on, for example, the performance information between “t0” and “tk”, that is, the performance information during a predetermined time after the content is registered. good. Further, the information distribution device 10 may use the record information (provided that i <k) between “ti” and “tk”. That is, the information distribution device 10 may extract the characteristic information based on the record information in any time interval. When such a process is executed, for example, the learning device Le extracts the feature of the content according to the aspect of the transition of the variation information up to that time, such as the transition of the number of impressions, the transition of the number of clicks, the transition of the CTR, and the like. be able to. As a result, the information distribution device 10 can improve the estimation accuracy of the estimation importance degree.

  Further, the information distribution device 10 can adopt any information as the variation information as long as it is the information that changes with the passage of time among the information regarding the content. For example, the information distribution device 10 may include the number of impressions, the number of clicks, the number of links to the content, the number of times the content is searched, the number of tweets, the user's evaluation of the content, and the like among the information that changes with time. At least one of them may be adopted as the variation information, and the characteristic information may be extracted based on the variation information.

[5-3. About steady information]
In addition, the information distribution device 10 can employ arbitrary information as stationary information, as long as it is information that does not change with the passage of time among information related to content. For example, the information distribution device 10 includes at least one of the content details, the provider such as the distribution source, the category to which the content belongs, the author, the text included in the content, the image, the moving image, and the unchanging information such as sound. One may be adopted as the variation information, and the characteristic information may be extracted based on the variation information.

  Further, in the above-described example, the information distribution device 10 sets the stationary information to the stationary information, out of the information about the content, the predetermined information that is hard to change with the passage of time. Here, the stationary information may be information that does not change with the passage of time, such as text included in the content. Further, the information distribution device 10 may use, as the steady information, any information that can be specified in a state where the record information is not collected, such as when the content is distributed.

[5-4. About extraction of characteristic information]
The information distribution device 10 described above uses the learning device Le capable of extracting the feature of the content from the set of the stationary information and the variation information. However, the embodiment is not limited to this, and the information distribution device 10 may perform learning of the learning device Le.

  For example, the information distribution device 10 compares the estimated importance degree estimated last time for each content with the performance information newly collected for each content, and calculates the order of the estimated importance degree of the content and the order of the performance information (for example, impressions). The learning device le or the model for estimating the estimated importance degree from the feature information may be modified so that the order of the numbers is the same. Further, the information distribution device 10 corrects the connection coefficient of the learning device Le so that the set of the input feature information, time information, and performance information matches the information restored from the extracted feature information. You can go. Such modification may be realized by a method such as a back propagation method.

[5-5. About targeting]
Here, the information distribution device 10 may use the learning device Le to select the content to be distributed. For example, the information distribution device 10 obtains a distributed expression indicating the attribute information of the user U01, such as demographic attributes such as income, occupation, and educational background, and psychographic attributes such as lifestyle and preferences of the user U01, and the learning device. Using Le, the characteristic information indicating the characteristic of the set of the attribute information of the user U01, the stationary information of the content, and the performance information is extracted. Then, the information distribution device 10 may estimate the estimated priority according to the attribute of the user U01 who is the distribution destination of the content by estimating the estimated priority from the extracted feature information.

  By executing such processing, the information distribution device 10 can improve the accuracy of targeting. In addition, when performing such a process, the information delivery apparatus 10 inputs the attribute information, the steady-state information, and the performance information that have been input from the feature information indicating the characteristics of the set of the input attribute information, the steady-state information, and the performance information. The learning device Le may be learned so as to restore the pair of and.

[5-6. About content]
In the above description, the information distribution device 10 has extracted the characteristic information from the distributed representation of the content including text such as news and columns. However, the embodiment is not limited to this. That is, the information distribution device 10 may generate a distributed expression from content such as an image, a moving image, and audio, and may extract feature information of the content based on the generated distributed expression. That is, the information distribution device 10 may use the distributed representation of images, moving images, audio, etc. as the stationary information of the content, and extract the characteristic information of the content based on the set of the stationary information and the record information.

  That is, the information delivery apparatus 10 estimates not only the text information but also the above-mentioned estimation of any content as long as the content can extract feature information by compressing the dimensionality of the information such as an image or a moving image. Processing or information distribution processing can be adopted.

[5-7. Advertising submission]
When the content is an advertisement, the advertiser may request the agency to submit the advertisement content. In this case, it is the agency that submits the advertising content to the information distribution device 10. In this specification, when the content is an advertisement, the notation such as “client” is a concept that includes not only the advertiser, but also the agency, such as a client where advertisement is performed by the advertisement content, and the notation such as “client terminal” is The concept includes not only the device used by the advertiser but also the device used by the agency.

[5-8. Other]
Further, of the processes described in the above embodiment, all or part of the processes described as being automatically performed may be manually performed, or the processes described as manually performed may be performed. All or part of the process can be automatically performed by a known method. In addition, the processing procedures, specific names, information including various data and parameters shown in the above-mentioned documents and drawings can be arbitrarily changed unless otherwise specified.

  Further, each constituent element of each illustrated device is functionally conceptual, and does not necessarily have to be physically configured as illustrated. That is, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or part of the device may be functionally or physically distributed / arranged in arbitrary units according to various loads and usage conditions. It can be integrated and configured.

  For example, the databases 14 to 19 shown in FIG. 2 may be held in the storage server or the like without being held by the information distribution device 10. In this case, the information distribution device 10 acquires the content and the like by accessing the storage server.

  Further, for example, the above-described information distribution device 10 receives the distribution request from the terminal device 100, distributes the content and the estimated importance level, and the front-end server side that mainly executes the distribution process, and It may be distributed to the back-end servers that mainly perform the estimation process of extracting the feature information based on the stationary information and the variation information and estimating the estimated importance degree from the extracted feature information. That is, the information distribution device 10 may be realized by an estimation device that estimates the estimated importance and an information distribution device that distributes content according to the estimated importance.

[5-9. Hardware configuration]
The information distribution device 10 according to the above-described embodiment is realized by, for example, a computer 1000 having a configuration shown in FIG. FIG. 13 is a diagram illustrating an example of the hardware configuration of a computer that realizes the functions of the information distribution device. The computer 1000 has a CPU 1100, a RAM 1200, a ROM 1300, an HDD 1400, a communication interface (I / F) 1500, an input / output interface (I / F) 1600, and a 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 starts up, a program dependent on the hardware of the computer 1000, and the like.

  The HDD 1400 stores programs executed by the CPU 1100, data used by the programs, and the like. The communication interface 1500 receives data from another device via the communication network 500 (corresponding to the network N) and sends the data to the CPU 1100, and also transmits the data generated by the CPU 1100 to the other device via the communication network 500. To do. The CPU 1100 controls output devices such as a display and a printer and input devices 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. The CPU 1100 also outputs the data generated via the input / output interface 1600 to the output device.

  The media interface 1700 reads a program or data stored in the recording medium 1800 and provides the program or data 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 information distribution device 10, the CPU 1100 of the computer 1000 realizes the function of the control unit 13 by executing the program loaded on the RAM 1200. Further, the HDD 1400 stores each data in the storage unit 12. The CPU 1100 of the computer 1000 reads these programs from the recording medium 1800 and executes them, but as another example, these programs may be acquired from other devices via a communication network.

[6. effect〕
As described above, the information distribution device 10 provides the characteristic information indicating the characteristic of the content at the predetermined time based on the stationary information that is the predetermined information related to the content and the variation information that changes with the passage of time. Extraction is performed using the learning device Le that extracts the characteristics of the input information from the input information. Then, the information distribution device 10 estimates the importance of the content at a predetermined time based on the extracted characteristic information. Therefore, the information distribution device 10 can improve the accuracy of estimating the importance of content.

  That is, the information distribution device 10 re-extracts the feature of the content at each time based on the feature of the stationary information and the variable information, using the learning device Le that extracts the feature of the input information. Then, the information distribution device 10 estimates the importance degree from the feature of the content re-extracted at each time. For this reason, the information distribution device 10 determines the content based on the characteristics of the stationary information and the variation information even when the index of the importance is different for each content or the aspect of the change of the importance with the passage of time is different. Extract the feature again. Further, the information distribution device 10 newly estimates the estimated importance degree from the performance information collected later even when the estimated importance degree estimated first is different from the actual importance degree. The importance of can be followed. Therefore, the information distribution device 10 can estimate the estimated importance level that appropriately reflects the actual importance level.

  In addition, the information distribution device 10 sets the information, which has less change with the passage of time than the change information, as the steady information. Therefore, the information distribution device 10 can extract the feature information indicating the feature of the content, based on the information about the content, which has little change over time.

  In addition, the information distribution device 10 selects a predetermined time from a set of the feature information extracted as the information indicating the feature of the content at the time before the predetermined time and the variation information at the time before the predetermined time. The feature information indicating the feature of the content is extracted. For example, the information distribution device 10 extracts new characteristic information from the set of the previously extracted characteristic information and the newly collected variation information.

  Here, it is considered that the previously extracted characteristic information includes cumulatively the characteristics of the variation information collected so far. Therefore, the information distribution device 10 can extract the characteristic information in consideration of all the variation information collected so far, from the set of the previously extracted characteristic information and the newly collected variation information. It is possible to extract feature information that accurately indicates the feature of the content in a certain time zone. As a result, the information distribution device 10 can improve the estimation accuracy of the estimation importance degree.

  In addition, the information distribution device 10 extracts new characteristic information from the set of the variation information in the time zone after the characteristic information is extracted last time and the characteristic information extracted last time. In this way, the information distribution device 10 re-extracts the characteristic information based on the new variation information in each time zone, so that the estimation accuracy of the estimated importance degree in each time zone can be improved.

  The information distribution device 10 also extracts the characteristic information of the new content based on the variation information collected so far. For this reason, the information distribution device 10 can estimate the estimated importance degree in the future based on the collected variation information, and thus can accurately perform CTR prediction and the like, for example.

  In addition, the information distribution device 10 extracts the characteristic information indicating the characteristic of the content, based on the variation information after the content is registered. Therefore, the information distribution device 10 extracts the characteristic information of the content based on the transition mode of the variation information, such as the variation of the number of impressions, for example, so that the estimation accuracy of the estimation importance degree can be improved.

  In addition, the information distribution device 10 is a learning device Le in which a plurality of nodes that output a calculation result for input data are connected in multiple layers, and the characteristic that the information has by compressing the dimension number of the input information is shown. The feature information of the content is extracted using the learning device Le that extracts the information. Therefore, the information distribution device 10 can appropriately extract the characteristic information of the content in each time period.

  Further, the information distribution device 10 includes a learning device Le having an intermediate layer that outputs the characteristic information extracted from the input information, an input layer that receives the characteristic information output by the intermediate layer, and the variation information, that is, a learning device Le. The feature information of the content is extracted using the RNN. Here, as described above, it is considered that the characteristic information includes cumulatively the characteristics of the variation information collected so far. Therefore, the information distribution device 10 can extract the feature information that accurately indicates the feature of the content in the next time zone, and thus the estimation accuracy of the estimated importance can be improved.

  In addition, the information distribution device 10 includes a content description, a content provider, a category to which the content belongs, an author of the content, a text included in the content, and a text, an image, a moving image, or a sound included in the content. Among them, the feature information of the content is extracted based on at least one of them. Therefore, the information distribution device 10 can extract the characteristic information based on the stationary information included in the content.

  Further, the information distribution device 10 sets at least one of the number of impressions, the number of selections, the number of links to the content, the number of searches of the content, the number of times information about the content such as the number of tweets has been transmitted, and the content evaluation. Based on this, the characteristic information of the content is extracted. Therefore, the information distribution device 10 can extract the characteristic information included in the content based on the variation information up to that point.

  Further, the information distribution device 10 estimates the feature importance indicating the feature of the content from the feature information, and estimates the estimated importance of the content using the model that estimates the estimated importance of the content from the feature information. Therefore, the information distribution device 10 can improve the estimation accuracy of the estimated importance of the content in each time zone.

  In addition, the information distribution device 10 determines a time later than the predetermined time from the set of the feature information extracted as the information indicating the feature of the content at the predetermined time and the variation information at the time later than the predetermined time. The feature information indicating the feature of the content is extracted. For example, the information distribution device 10 extracts new characteristic information from the set of the previously extracted characteristic information and the newly collected variation information. In this way, the information distribution device 10 re-extracts the characteristic information based on the new variation information in each time zone, so that the estimation accuracy of the estimated importance degree in each time zone can be improved.

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

  Also, the above-mentioned "section (module, unit)" can be read as "means" or "circuit". For example, the delivery unit can be read as delivery means or delivery circuit.

10 information distribution device 11 communication unit 12 storage unit 13 control unit 14 content database 15 result information database 16 learning device database 17 feature information database 18 estimation model database 19 importance database 20 reception unit 21 collection unit 22 generation unit 23 extraction unit 24 estimation Part 25 Distribution part 100, 101 Terminal device 102 Log server 200 Client terminal

Claims (13)

Based on the predetermined stationary information regarding the content and the variation information that changes with the passage of time among the information regarding the content, the characteristic information indicating the characteristic of the content at a predetermined time is changed from the input information. extraction apparatus characterized by comprising an extraction unit for extracting by using a learning device for extracting features with. The extraction unit extracts the characteristic information as the predetermined stationary information, based on the information about the content, which has a smaller change with time than the variation information. The extraction device described in 1. The extraction unit selects the predetermined value from a set of feature information extracted as information indicating a feature of the content at a time earlier than the predetermined time and variation information at a time earlier than the predetermined time. The extraction device according to claim 1 or 2, wherein characteristic information indicating a characteristic of the content at a time is extracted . The extraction unit displays variation information in a time period after extracting characteristic information indicating characteristics of the content at a time earlier than the predetermined time and characteristics of the content at a time earlier than the predetermined time. The extraction device according to claim 3, wherein feature information indicating a feature of the content at the predetermined time is extracted from a set with the feature information indicated. The extraction unit extracts, as the variation information, characteristic information indicating a characteristic of the content at the predetermined time, based on variation information at a time before the predetermined time. The extraction device according to any one of 4 to 4. The extraction unit extracts, as the variation information, characteristic information indicating a characteristic of the content at the predetermined time, based on the variation information from the registration of the content to the predetermined time. The extraction device according to any one of claims 1 to 5. The extraction unit is a learning device in which a plurality of nodes that output a calculation result for input data are connected in multiple layers, and the feature information indicating the feature of the information is extracted by compressing the dimension number of the input information. The extraction device according to any one of claims 1 to 6, wherein characteristic information of the content is extracted using a learning device that performs the extraction . The extraction unit uses, as the learning device, a learning device having a layer for outputting the characteristic information extracted from the input information, a characteristic information output by the layer, and a layer for receiving the input of the variation information. The extraction device according to claim 7, wherein the characteristic information of the content is extracted . The extraction unit uses, as the stationary information, details of the content, a provider of the content, a field to which the content belongs, an author of the content, a text included in the content, and a text, an image, and a video included in the content. The feature information of the content is extracted based on at least one of an image and a sound. The extraction device according to any one of claims 1 to 8, wherein the feature information of the content is extracted . The extraction unit uses, as the variation information, the number of times the content has been browsed, the number of times the content has been selected, the number of links to the content, the number of times the content has been searched, and the number of times information about the content has been transmitted. , And the evaluation of the content, based on at least one of them, the characteristic information of the content is extracted . The extraction apparatus according to claim 1. The feature of the content at a time later than the predetermined time is shown based on the feature information indicating the feature of the content at the predetermined time and the variation information that changes with the passage of time in the information about the content. An extraction device having an extraction unit for extracting characteristic information. Based on the predetermined stationary information regarding the content and the variation information that changes with the passage of time among the information regarding the content, the characteristic information indicating the characteristic of the content at a predetermined time is changed from the input information. An extraction method characterized by including an extraction step of extracting using a learning device for extracting the feature. On the computer,
Based on the predetermined stationary information regarding the content and the variation information that changes with the passage of time among the information regarding the content, the characteristic information indicating the characteristic of the content at a predetermined time is changed from the input information. An extraction program characterized by executing an extraction procedure for extraction using a learning device for extracting a feature having the feature.

Images