JP2020140624A - Attribute estimation device, attribute estimation method, and program - Google Patents

Abstract

To estimate attributes of users using a content distribution service using information available to a content distribution company.SOLUTION: A reproduction frequency count unit of an attribute estimation device counts, for each content provider, the reproduction frequency of a content on a terminal of a content user, using provider information indicating a content provider of each content and log data indicating a content reproduced on the terminal of the content user. A model application unit inputs input data indicating the reproduction frequency for each content provider that has been counted, to an estimation model that is a neural network that inputs the reproduction frequency for each content provider and outputs a value indicating the probability that the attribute of the content user is each attribute value. An estimation unit estimates the attribute value of the content user based on a result output by the estimation model in response to the input of the input data.SELECTED DRAWING: Figure 5

Description

The present invention relates to an attribute estimation device, an attribute estimation method and a program.

Content distributors that provide services that distribute content via the Internet distribute notifications, advertisements, etc. (hereinafter referred to as "advertisements, etc.") to the terminals of users who use the service. Advertisements that match the attributes of the user are expected to be more effective than advertisements that do not match.

On the other hand, there is a technique for estimating information on a user's taste and personality from a user's profile image used in an SNS (social networking service) (see, for example, Non-Patent Document 1).

Yudai Yamashita, Junichiro Mori, "Estimating poster attributes from SNS profile images using deep learning", [online], 2016, 2016 National Conference of the Japanese Society for Artificial Intelligence (30th) Proceedings, [2019 2 Search on 21st of March], Internet <URL: https://www.jstage.jst.go.jp/article/pjsai/JSAI2016/0/JSAI2016_4K14/_pdf>

As a method of acquiring the attributes of the user in order to select an advertisement or the like to be delivered to the user, it is conceivable to have the user register as a member including information on the attributes. However, there are many users who want to easily use the service without performing the complicated procedure for membership registration. In addition, as a method of acquiring the user's attribute, it is also conceivable to acquire the access log to the site or the input information to the site so that the attribute can be estimated. However, it is difficult for content distributors to collect a wide range of such information about sites other than their own. In the technology of Non-Patent Document 1, it is necessary to have the user's profile image registered at the time of membership registration, or to make it possible to refer to the profile image of another site.

In view of the above circumstances, the present invention provides an attribute estimation device, an attribute estimation method, and a program capable of estimating the attributes of a user using the content distribution service using information available to the content distribution company. It is an object.

One aspect of the present invention is attribute estimation using provider information indicating a content provider who has provided each of a plurality of contents and log data indicating the content reproduced on the terminal of the content user to be attribute estimation target. A playback count counting unit that counts the number of times the content has been played on the terminal of the target content user for each content provider, and a content user attribute for each attribute by inputting the number of times played for each content provider. A model application unit that inputs input data indicating the number of times of reproduction for each content provider counted by the number of times of reproduction into an estimation model that is a neural network that outputs a value representing a probability of being a value, and the input data. This is an attribute estimation device including an estimation unit that estimates the attribute value of the content user to be attribute estimation based on the result output by the estimation model in response to the input of.

One aspect of the present invention is the attribute estimation device described above, in which the log data includes information on the playback date and time of the content, and the model application unit is the first period indicated by the log data of the content user. The number of playbacks for each content provider in the above is converted into the number of playbacks in the second period shorter than the first period, and the number of playbacks for each content provider in the second period after conversion is equal to the predetermined number of playbacks. The number of playbacks for each content provider when is selected is used as the input data.

One aspect of the present invention is the attribute estimation device described above, in which the number of times the content is played on the terminal of a plurality of the content users is counted for each content provider, and the number of times the content is played is based on the counted number of times played. The model application unit further includes a ranking determination unit that assigns a ranking to the content providers, and the model application unit includes the number of times of playback of each of the content providers whose ranking is higher than a predetermined value, and the said of each of the content providers having a ranking of a predetermined value or less. The total number of playbacks, which is the total number of playbacks, is used as the input data.

One aspect of the present invention is the above-mentioned attribute estimation device, which uses the provider information and the log data of each of the plurality of content users for learning to use the log data of each of the plurality of content users for learning. A learning play count counting unit that counts the number of times the content has been played on the terminal for each content provider, correct attribute values of each of the plurality of learning content users, and a plurality of the learning content users. A learning processing unit that learns the estimation model using input data indicating the number of times of reproduction for each content provider is further provided.

One aspect of the present invention is the attribute estimation device described above, wherein the content is a moving image, and the attribute is age.

One aspect of the present invention is an attribute estimation method executed by an attribute estimation device, in which the attribute estimation device provides provider information indicating a content provider who has provided each of a plurality of contents, and a content user to be attribute-estimated. Using the log data indicating the content played on the terminal, the number-of-playback counting step of counting the number of times the content has been played on the terminal of the content user whose attributes are estimated for each content provider, and the above-mentioned The attribute estimation device counts in the playback count counting step in the estimation model, which is a neural network that inputs the number of playbacks for each content provider and outputs a value representing the probability that the attribute of the content user is each attribute value. Based on the model application step of inputting the input data indicating the number of times of playback for each content provider and the result output by the estimation model in response to the input of the input data, the content user whose attributes are estimated. It has an estimation step for estimating an attribute value.

One aspect of the present invention is a program for causing a computer to function as any of the above-mentioned attribute estimation devices.

INDUSTRIAL APPLICABILITY According to the present invention, it is possible to estimate the attributes of a user who is using a content distribution service by using information available to a content distribution company.

It is a block diagram of the attribute estimation system by one Embodiment of this invention. It is a block diagram which shows the structure of the moving image viewing terminal by this embodiment. It is a block diagram which shows the structure of the moving image viewing server by the same embodiment. It is a block diagram which shows the structure of the data collection server by the same embodiment. It is a block diagram which shows the structure of the server for deep learning by the same embodiment. It is a figure which shows the example of the estimation model by the same embodiment. It is a figure which shows the input data to the estimation model by the same embodiment. It is a figure which shows the acquisition method of the moving image viewing count by the same embodiment. It is a flow diagram which shows the estimation model learning process of the attribute estimation system by the same embodiment. It is a flow chart which shows the estimation process of the attribute estimation system by the same embodiment. It is a figure which shows the example of oversampling by the same embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The attribute estimation device of the present embodiment estimates the attributes of the user who is using the content distribution service provided by the content distribution company. In the following, the content distributed by the content distributor is video data (hereinafter, simply referred to as “video”), and the case where the estimated user attribute is age will be described as an example.

FIG. 1 is a configuration diagram of an attribute estimation system 1 according to the present embodiment. The attribute estimation system 1 includes a moving image viewing terminal 3, a moving image viewing server 4, a data collecting server 5, and a deep learning server 6. The video viewing terminal 3, the video viewing server 4, and the data collection server 5 are connected to the network 9. The network 9 is, for example, the Internet. In the figure, one video viewing server 4 and one data collection server 5 are shown, but a plurality of video viewing server 4 and data collection server 5 may be provided. Further, in the figure, the data collection server 5 and the deep learning server 6 are directly connected, but they may be connected via the network 9. Further, a part or all of the video viewing server 4, the data collection server 5, and the deep learning server 6 may be realized by the same computer server.

FIG. 2 is a block diagram showing the configuration of the moving image viewing terminal 3. The video viewing terminal 3 is a computer terminal such as a personal computer, a tablet terminal, or a smartphone used by a video viewer. The video viewer is a user who uses the video distribution service. The moving image viewing terminal 3 includes an input unit 31, a moving image viewing function unit 32, and an output unit 33. The input unit 31 inputs the operation of the moving image viewer. The video viewing function unit 32 transmits the terminal ID and the video distribution request to the video viewing server 4 according to the operation of the video viewer using the input unit 31. The terminal ID is information that uniquely identifies the moving image viewing terminal 3. The address information assigned to the moving image viewing terminal 3 can be used as the terminal ID. Moreover, you may use a user ID as a terminal ID. The user ID is information that uniquely identifies the video viewer. The video distribution request includes information that identifies the target video for which the video viewer requests distribution. The video viewing function unit 32 outputs the video distributed from the video viewing server 4 to the output unit 33 in response to the video distribution request. The output unit 33 is a display and a speaker.

FIG. 3 is a block diagram showing the configuration of the moving image viewing server 4. The video viewing server 4 is used for a video distribution service provided by a video distribution company. The video viewing server 4 includes a storage unit 41, a distribution unit 42, and a log generation unit 43.

The storage unit 41 stores the posted video information. The posted video information includes a video ID, a video, video poster information, and video additional information. The moving image ID is information that uniquely identifies the moving image. The video contributor information includes information such as a video contributor ID and a video contributor name. The video contributor ID is information that uniquely identifies the video contributor. Video contributors are individuals, companies, groups, etc. who provided the videos to be distributed in the video distribution service. The video contributor name and the video contributor ID may be the same. The video additional information indicates information related to the video. The video additional information includes the title of the video, the description of the video used for the search, keywords, and the like.

The distribution unit 42 receives the terminal ID and the video distribution request from the video viewing terminal 3. The distribution unit 42 reads the moving image requested to be distributed by the moving image distribution request from the storage unit 41 and distributes the moving image to the moving image viewing terminal 3 which is the transmission source of the moving image distribution request. The log generation unit 43 generates a moving image viewing log. The video viewing log includes the terminal ID of the video viewing terminal 3, the video ID included in the video poster information of the video distributed to the video viewing terminal 3, the video poster information and the video title, and the viewing start time and viewing end time. It is data including and. The log generation unit 43 transmits the generated video viewing log to the data collection server 5.

FIG. 4 is a block diagram showing the configuration of the data collection server 5. The data collection server 5 collects the video viewing log and provides it to the deep learning server 6. The data collection server 5 includes a log collection function unit 51 and a log extraction function unit 52. The log collection function unit 51 collects and stores the video viewing log from the video viewing server 4. The log extraction function unit 52 reads the moving image viewing log stored in the log collection function unit 51 and outputs it to the deep learning server 6.

FIG. 5 is a block diagram showing the configuration of the deep learning server 6. The deep learning server 6 estimates the age of the video viewer using the estimation model learned by deep learning. The estimation model is a neural network with multiple hidden layers. Each node of the input layer corresponds to the order when the video contributors are arranged in the order of the number of times the video is viewed on all the video viewing terminals 3. Each node in the output layer corresponds to an age.

The deep learning server 6 includes a ranking creation processing unit 61, a learning data creation processing unit 62, a deep learning learning processing unit 63, a learning model function unit 64, and an age estimation function unit 65. The ranking creation processing unit 61 counts the number of playbacks, which is the number of times the video of each video contributor has been played on all the video viewing terminals 3, based on the video viewing logs of all the video viewing terminals 3. The ranking creation processing unit 61 may sample a part of all the video viewing terminals 3 and count the number of playbacks of each video contributor based on the sampled video viewing log of the video viewing terminal 3. The ranking creation processing unit 61 assigns rankings to video contributors in descending order of the number of playbacks. This ranking is referred to as "ranking". The training data creation processing unit 62 creates training data for use in deep learning of the estimation model. The deep learning learning processing unit 63 learns an estimation model by deep learning using the learning data created by the learning data creation processing unit 62. The learning model function unit 64 generates input data of the video viewer to be age-estimated and inputs it to the trained estimation model. The age estimation function unit 65 estimates the age of the moving image viewer based on the output result of the estimation model, and outputs the estimation result.

The deep learning server 6 may be realized by a plurality of computer devices connected to the network. In this case, which of these plurality of computer devices is used to realize each functional unit of the deep learning server 6 can be arbitrary. For example, the ranking creation processing unit 61, the learning data creation processing unit 62 and the deep learning learning processing unit 63, and the learning model function unit 64 and the age estimation function unit 65 may be realized by different computer devices.

FIG. 6 is a diagram showing an example of an estimation model. The neural network used as an estimation model consists of an input layer, a hidden layer and an output layer. The number of hidden layers shown in FIG. 6 is three, but the number of hidden layers may be two or four or more.

The input layer consists of M nodes. Each node in the input layer is labeled from 1 to M. The label numbers correspond to the rankings. The value input to the node of the label m (m is an integer of 1 or more and M-1 or less) is the number of times the video viewer has viewed the video of the video contributor of the ranking m. The value input to the node of the label M is the number of times the video viewer has viewed the video of the video contributor having the ranking M or lower.

The output layer consists of K nodes. Each node in the output layer corresponds to the age. For example, each node in the output layer corresponds to under 10 years old, teens, 20s, .... The range of ages corresponding to one node of the output layer can be arbitrary, such as 1 year old, 2 years old, and 5 years old. In addition, the range of ages corresponding to each node may be the same, and some or all of them may be different. The value of each node in the output layer represents the probability of the age corresponding to the node.

In the estimation model shown in FIG. 6, each layer of the hidden layer and the output layer are fully connected layers. When the input layer is the first layer, the hidden layer is the second to fourth layers, and the output layer is five layers, each node of the l layer (l is an integer of 2 or more and 5 or less) is all of the (l-1) layer. Combined with all nodes. In the learning of the estimation model by deep learning, the bond strength (weight) between each node of the l layer and each node of the (l-1) layer is determined.

FIG. 7 is a diagram showing input data to the estimation model. The ranking creation processing unit 61 refers to the video viewing logs of all the video viewing terminals 3 and totals the number of video playbacks for each video contributor. The ranking creation processing unit 61 assigns rankings to video contributors in descending order of the number of times the video is played. It is preferable to use the number of times the video is viewed by all the video contributors as the input data, but when the number of video contributors is large, a large amount of resources of the deep learning server 6 are consumed. In addition, the information of video contributors who have a large number of video views is mainly useful for mass identification, but the information of video contributors who have a small number of video views is closely related to niche tastes and is useful for personal identification. It is believed that there is. Therefore, when the deep learning learning processing unit 63 generates the input data used for learning the estimation model, the deep learning learning processing unit 63 collectively assigns the ranking M to the video contributors whose ranking is the threshold value M or less. In the figure, an example of M = 1001 is shown. The threshold value M can be arbitrarily determined. For example, the estimation model may be trained for each different threshold value M, and the smallest M having a correct answer rate equal to or higher than a predetermined value may be used. The threshold value M may be a value determined by the ratio to the total number of contributors, or may be a fixed value.

As described above, the input data to the estimation model is the number of times that the video viewer has viewed the video of each of the video contributors in the rankings 1 to M. The deep learning learning processing unit 63 and the learning model function unit 64 acquire the number of times the video poster of each ranking is viewed based on the video viewing log of the video viewer, and use it as input data. That is, the input data is represented as (P1, P2, ..., PM), where Pm is the number of times the video contributor in the ranking m (m is an integer of 1 or more and M or less) has viewed the video. Pm is an input value to the node of the label m in the input layer. For example, as shown in FIG. 7, it is assumed that the video contributors of rankings 1, 2, 3, ..., 998, 999, 1000 are A, B, C, ..., X, Y, Z. Then, from the video viewing log of the terminal a, which is a certain video viewing terminal 3, the videos of the video contributors A, A, A, A, A, B, B, ..., X, X, X, Y, Z, Z are displayed. Suppose you get what you watched. In this case, the input data for estimating the age of the moving image viewer who viewed the moving image using the terminal a is (5,2,0, ..., 0,3,1,2,0).

FIG. 8 is a diagram showing a method of acquiring the number of times of viewing a moving image used for input data. FIG. 8A is a diagram showing video contributors of videos viewed by video viewers in the past. In the figure, even if the video is not actually watched, the video of the video contributor is shown for the period between the time when the video viewer watched the video of the same video contributor. It is included in the viewing period. In general, video viewers tend to watch videos of the same video contributor in succession. For example, when a video viewer watches a drama composed of 10 episodes of video, the video viewer may continuously watch the video over one to several days. Therefore, when the number of times of video viewing for the past one to several days is acquired from the video viewing log and the input data is generated, the number of times of video viewing by a specific video contributor increases, and the accuracy of dating estimation deteriorates. there is a possibility.

Therefore, the deep learning learning processing unit 63 and the learning model function unit 64 shuffle the video viewing log for a long period (term T1) in units of a target period (period T2 <period T1) for obtaining the number of video views. To do. FIG. 8B is a diagram showing a moving image viewing log after shuffling. Shuffle is to calculate the ratio of the number of video views of each video contributor based on the video viewing log of period T1, and distribute the number of video views of each video contributor for each period T2 while maintaining the calculated ratio. Is. In other words, shuffle refers to the number of video views of each video contributor over a long period (for example, period T1 such as one year, half a year, several months), and the target period (for example, one day, one week, etc.). It is to convert to the number of times of watching a moving image per T2).

For example, it is assumed that a person who watches a lot of videos of video contributors A and C is in his thirties, and a person who watches a lot of videos of video contributors B is in his twenties. At this time, information on the video viewed by the video viewer during the past one day (period T2) is randomly acquired for a predetermined number of times from the video viewing log before shuffling shown in FIG. 8A, and the acquired video is obtained. Suppose that the input data is created based on the information in. In this case, in reality, although the video viewer is watching a lot of the video of the video poster B, the number of times the video of the video poster A is viewed is large in the input data. As a result, the correct answer may be estimated to be in the twenties, even though it is in the thirties.

On the other hand, after shuffling shown in FIG. 8B, the video viewing log for the past one day reflects the ratio of the number of times the video is viewed over the past long period (period T1). Therefore, if the information of the video viewed by the video viewer in the past one day (period T2) is randomly acquired for a predetermined number of times from the video viewing log after shuffling, the number of times the video of the video poster A has been viewed is larger than the number of times the video is viewed. The number of times the video of the video poster B is viewed increases. When input data is created based on the acquired video information, it is estimated that the input data is in the twenties because it represents a long-term viewing tendency. In this way, by temporally shuffling the moving image viewing log when generating the input data, it is possible to eliminate the bias of the input data and improve the estimation accuracy.

FIG. 9 is a flow chart showing the estimation model learning process in the attribute estimation system 1. In the learning data creation process, the log extraction function unit 52 acquires the video viewing log of all (or sampled) video viewing terminals 3 collected and stored by the log collecting function unit 51. At this time, the log extraction function unit 52 may acquire the moving image viewing log for the past predetermined period. The log extraction function unit 52 outputs the acquired video viewing log to the deep learning server 6 (step S105).

The ranking creation processing unit 61 of the deep learning server 6 refers to all the video viewing logs received from the log extraction function unit 52, and counts the number of times the video of each video contributor has been viewed on the video viewing terminal 3. .. The ranking creation processing unit 61 assigns rankings to video contributors in descending order of the number of video views counted, and creates a ranking (step S110). The ranking creation processing unit 61 associates the video poster ID of the video poster with the ranking ranking and outputs the data to the deep learning learning processing unit 63.

The learning data creation processing unit 62 processes data for deep learning on the moving image viewing log (step S115). For example, the learning data creation processing unit 62 deletes data (for example, a video title) unnecessary for deep learning from the video viewing log. The learning data creation processing unit 62 associates the terminal ID of the video viewing terminal 3 which is the teacher data with the information indicating the age which is the correct answer, and outputs the data to the deep learning learning processing unit 63 (step S120). Information on the correct age of each moving image viewing terminal 3 may be input by an input unit (not shown) of the deep learning server 6. Alternatively, the learning data creation processing unit 62 may read information on the correct answer age of each moving image viewing terminal 3 from another device such as a database server that stores member information.

The learning data creation processing unit 62 totals the number of times of video playback for each video contributor during the video log acquisition period (for example, one year, half a year, several months, etc.) for each video viewing terminal 3. The learning data creation processing unit 62 generates learning data by adding information associated with the video contributor ID of each video contributor and the number of times the video has been played to the video viewing log processed for deep learning (step). S125). The video viewing terminal 3 for which the learning data is generated includes, in addition to the video viewing terminal 3 which is the teacher data, the video viewing terminal 3 to which the age information is not given. The learning data creation processing unit 62 outputs the generated learning data to the deep learning learning processing unit 63.

In the model creation process, the deep learning learning process unit 63 initializes the information required for the process (step S205). First, the deep learning learning processing unit 63 initializes the variable i and the variable j with the value 0. The variable i indicates the number of times the estimation model has been trained. The variable j is a value representing the accuracy of the model. Further, the deep learning learning processing unit 63 acquires the upper limit learning number N times and the threshold value p% of the accuracy of the estimation model. The deep learning learning processing unit 63 may acquire the upper limit of the number of learnings N times input by the input unit (not shown) of the deep learning server 6 and the threshold value p% of the accuracy of the estimation model, and is a storage unit (not shown). You may read from.

The deep learning learning processing unit 63 cuts off the video contributors with lower rankings (step S205). That is, the deep learning learning processing unit 63 refers to the ranking information associated with each video contributor ID, and selects a predetermined ratio or a predetermined number of video contributor IDs having a higher ranking. The deep learning learning processing unit 63 assigns a ranking rank as a label to the video contributor IDs for the selected (M-1) people, and assigns a label M to all the video contributor IDs not selected. (Step S210).

The deep learning learning processing unit 63 shuffles the long-term video viewing log included in the learning data for each video viewer, and eliminates the bias in the target period for acquiring the number of video viewings (step S215). The deep learning learning processing unit 63 generates input data for each video viewer based on the shuffled video viewing log (step S220). That is, the deep learning learning processing unit 63 randomly acquires the video ID of the video viewed from the target period of the shuffled video viewing log for each video viewer for a predetermined number of viewing times, and corresponds to the acquired video ID. Identify the attached video contributor ID label. The deep learning learning processing unit 63 counts the number of times of viewing a moving image for each label, and arranges the number of times of viewing a moving image counted in the order of labels as input data.

The deep learning learning processing unit 63 determines whether or not the amount of teacher data is uniform in each generation (step S225). For example, the deep learning learning processing unit 63 reads out the information of the age corresponding to each terminal ID from the teacher data, counts the number of terminal IDs for each age, and uses it as the number of teacher data. The deep learning learning processing unit 63 compares the number of teacher data for each age group, and determines that the number of teacher data is not uniform if there is an age group smaller than a predetermined age group.

When the deep learning learning processing unit 63 determines that the data is not uniform (step S225: NO), oversampling is performed so as to increase the teacher data of the age group in which the number of teacher data is smaller than a predetermined value as compared with other age groups (step S225: NO). S230). In oversampling, the shortage of teacher data is supplemented based on a small number of teacher data.

FIG. 11 is a diagram showing an example of oversampling. In this embodiment, statistical methods are used to evenly supplement the shortage of teacher data. FIG. 11A shows a plot of teacher data (input data of known age) before oversampling. Although it is shown in two dimensions in FIG. 11 for simplicity, it is actually plotted in M dimensions. As shown in FIG. 11A, the number of teacher data of ages B, C, and D is smaller than a predetermined number of teacher data of age A. Therefore, the deep learning learning processing unit 63 complements the teacher data for each of the ages B, C, and D.

For example, when complementing the teacher data of the age B, the deep learning learning processing unit 63 selects one of the teacher data of the age B, and the other teacher data of the age B in the vicinity of the selected teacher data. Randomly generate input data with. This generated input data is used as the complementary teacher data of the age B. The deep learning learning processing unit 63 of the age B before completion until the difference between the total number of teacher data and the number of supplemented teacher data of the age B before completion and the number of teacher data of age A is within a predetermined range. Select teacher data to complement the teacher data. The deep learning learning processing unit 63 performs this processing for each of the age C and the age D. FIG. 11B shows a plot of teacher data after oversampling.

As shown in FIG. 9, when the deep learning learning processing unit 63 determines in step S225 that the amount of teacher data is uniform in each generation (step S225: YES) or after oversampling (step S230), The process of step S235 is performed. That is, when one teacher data is selected, the deep learning learning processing unit 63 learns the estimation model by the deep learning method using the selected teacher data (step S235). Specifically, the deep learning learning processing unit 63 inputs the input data indicated by the selected teacher data to the current estimation model, and obtains the value of each node of the output layer as the estimation result. The deep learning learning processing unit 63 calculates the difference between the estimation result and the value of the output layer when the age indicated by the teacher data is correct by the loss function. When the age indicated by the teacher data is correct, the value of the node corresponding to the correct age is 100%, and the value of the node corresponding to the non-correct age is 0%. The deep learning learning processing unit 63 updates the coupling strength between the nodes of the estimation model by using a weighted loss function that weights the calculation result by the loss function according to the amount of data.

The deep learning learning processing unit 63 adds 1 to the number of learning times i and sets the value j of the loss function to the value p'of the loss function calculated in the immediately preceding step S230 (step S240). The deep learning learning processing unit 63 determines whether the learning number i has reached the upper limit learning number N, or the value j of the loss function has reached the accuracy threshold value p% or more (step S245). Here, the larger the value of the loss function, the better the accuracy. When the smaller the value of the loss function is, the better the accuracy is, instead of "whether the value j of the loss function is equal to or more than the threshold value p% of the accuracy", "the value j of the loss function is the threshold value of the accuracy". "Is it p% or less?" Is the judgment condition.

When the deep learning learning processing unit 63 determines that the number of learnings i has not reached the upper limit of the number of learnings N and the value j of the loss function is less than the threshold value p% of accuracy (step S245: YES). ), Semi-supervised learning in step S250 is performed (step S250). That is, the deep learning learning processing unit 63 inputs the input data that is not the teacher data (the correct age is not given) among the input data generated in step S220 to the current estimation model being trained, and the output layer. The value of is obtained as the estimation result. When the deep learning learning processing unit 63 obtains an estimation result indicating that only one of the nodes in the output layer has a predetermined probability or more, the input data at the time when the estimation result is obtained. Is added as teacher data. The correct age of the added teacher data is the age corresponding to the node for which a value equal to or higher than a predetermined probability is obtained. For example, the output layer of the estimation model consists of 6 nodes corresponding to each of teens or younger, 20s, 30s, 40s, 50s, 60s or older, and the values of each node are 99%, 0.1%, and so on. It is assumed that 0.5%, 0.2%, 0.1%, and 0.1% are obtained. In this case, the deep learning learning processing unit 63 adds the input data to the teacher data, and sets the age of the correct answer to the teenager or younger corresponding to the node in which 99% or more of the predetermined probability is obtained. The deep learning learning processing unit 63 repeats the processing from step S225.

When the deep learning learning processing unit 63 determines that the number of learnings i has reached the upper limit of the number of learnings N or the value j of the loss function is equal to or more than the threshold value% of accuracy (step S245: YES), the deep learning learning processing unit 63 has learned. The estimated model is output to the learning model function unit 64 (step S255). The learning model function unit 64 saves the trained estimation model (step S260).

FIG. 10 is a flow chart showing the estimation process in the attribute estimation system 1. The video viewing terminal 3 accesses the video viewing server 4 via the network 9 (step S305). The video viewing function unit 32 of the video viewing terminal 3 transmits a video distribution request to the video viewing server 4 according to the operation performed by the video viewer by the input unit 31. The distribution unit 42 of the video viewing server 4 reads the video requested to be distributed by the video distribution request from the storage unit 41 and distributes it to the video viewing terminal 3. The video viewing function unit 32 of the video viewing terminal 3 reproduces the video distributed from the video viewing server 4 and outputs it to the output unit 33. The log generation unit 43 of the video viewing server 4 includes the terminal ID of the video viewing terminal 3, the video ID of the video distributed to the video viewing terminal 3, the video poster information and the video title of the distributed video, and the viewing start time. A moving image viewing log including the viewing end time and the viewing end time is generated and transferred to the data collection server 5 (step S310). The log collection function unit 51 of the data collection server 5 stores the transferred moving image viewing log.

The learning model function unit 64 of the deep learning server 6 refers to the video viewing log of the video viewer to be dating, which is collected by the log collection function unit 51 of the data collection server 5, and is capable of dating. It is detected that the viewing log has been collected (step S315). For example, when the number of times the video is viewed reaches a predetermined value or more, the age can be estimated.

The learning model function unit 64 generates input data by the same processing as the deep learning learning processing unit 63, using the video viewing log of the video viewer whose age is to be estimated. That is, the learning model function unit 64 shuffles the video viewing log when the number of video playbacks for each video contributor during the video log acquisition period is totaled. The learning model function unit 64 randomly acquires the video ID of the viewed video for a predetermined number of viewing times from the target period of the shuffled video viewing log, and the video poster ID associated with the acquired video ID. Identify the label. The learning model function unit 64 counts the number of times the video is viewed for each label, and arranges the number of times the video is viewed in the order of the labels as input data (step S320). The learning model function unit 64 inputs the input data to the input layer of the trained estimation model and calculates the estimation result (step S325).

The age estimation function unit 65 estimates the age of the video viewer based on the calculation result obtained in step S325 (step S330). That is, the age estimation function unit 65 sets the age corresponding to the node having the largest value among the values of each node of the output layer as the estimation result. The age estimation function unit 65 outputs the age of the estimation result. The output may be displayed on a display, transmitted to a device connected by a network, or recorded on a recording medium.

In the above, the video viewing server 4 generates the video viewing log, but when the video viewing terminal 3 can generate the video viewing log, the data collection server 5 directly from each video viewing terminal 3 or the video. The moving image viewing log may be collected via the viewing server 4.

As described above, the attribute estimation system 1 has a base for accumulating the video viewing log of the video viewer. The attribute estimation system 1 analyzes the video viewing log using deep learning and estimates the age of the video viewer. When estimating the age by deep learning, the video viewing log must be in a suitable form as input data. In deep learning, how to learn teacher data is also important. Therefore, the attribute estimation system 1 calculates the ranking of video contributors based on all video viewing logs. The attribute estimation system 1 replaces the name of the video poster or the video poster ID in the past video viewing log with the ranking. The attribute estimation system 1 inputs the number of times of video viewing by the video contributor of the video viewer whose age is to be estimated into the estimation model by arranging them in the order of ranking based on the viewing logs of all the video viewers. At this time, the attribute estimation system 1 inputs the number of times of video viewing obtained by shuffling the video viewing log into the estimation model. The attribute estimation system 1 displays the age estimated using the estimation model.

The attribute estimation system 1 can estimate the age of the moving image viewer by using the estimation model learned by deep learning. In addition, the attribute estimation system 1 aggregates the video viewing logs in a ranking format, uses the individual video viewing counts as input data for the top video contributors that affect accuracy, and video viewing for other video contributors. Add up the number of times. As a result, the resources required for learning can be reduced. Further, since the attribute estimation system 1 shuffles the moving image viewing log at the time of input, it is possible to eliminate the bias of the input data and improve the estimation accuracy. Therefore, when the moving image viewer views a plurality of moving images, the attribute estimation system 1 can estimate the age of the moving image viewer. In this way, the business operator providing the video distribution service can estimate the age of the video viewer by using the log of the video distribution to the video viewing terminal.

According to the embodiment described above, the attribute estimation device includes a reproduction count counting unit, a model application unit, and an estimation unit. For example, the attribute estimation device is a deep learning server 6, the reproduction count counting unit and the model application unit are the learning model function unit 64, and the estimation unit is the age estimation function unit 65. The play count counting unit uses the provider information indicating the content provider who provided each of the plurality of contents and the log data indicating the content played on the terminal of the content user whose attribute is estimated, to estimate the attribute. The number of times the content is played on the content user's terminal is counted for each content provider. The model application unit inputs the number of times played by each content provider and outputs a value representing the probability that the attribute of the content user is each attribute value. The content counted by the number of times played unit counts in the estimation model, which is a neural network. Input input data indicating the number of playbacks for each provider. The estimation unit estimates the attribute value of the content user whose attribute is to be estimated based on the result output by the estimation model in response to the input of the input data.

For example, the content is a video, the content provider is a video poster, the content user is a video viewer, the provider information is the video poster information, and the content user's terminal is the video viewing terminal 3. Yes, the log data is a video viewing log. Also, for example, the attribute is age and the attribute value is age.

The log data may include information on the playback date and time of the content. The model application unit converts the number of times played by each content provider in the first period (period T1) indicated by the log data of the content user into the number of times played in the second period (period T2), which is shorter than the first period. The model application unit uses the number of playbacks for each content provider when a predetermined number of playbacks is selected from the number of playbacks for each content provider in the second period after conversion as input data.

The attribute estimation device may further include a ranking determination unit. The ranking determination unit is, for example, a ranking creation processing unit 61. The ranking determination unit counts the number of times the content has been played on the terminals of a plurality of content users for each content provider using the provider information and the log data of a plurality of content users, and is based on the counted number of times played. Gives ranking to content providers. The model application unit uses input data as the number of times played by each content provider whose rank is higher than the predetermined value and the total number of times played by each content provider whose rank is lower than the predetermined value.

The attribute estimation device may further include a learning reproduction count counting unit and a learning processing unit. For example, the learning reproduction count counting unit and the learning processing unit are the deep learning learning processing unit 63. The learning playback count counting unit uses the provider information and the log data of each of the plurality of content users for learning to determine the number of times the content has been played on each terminal of the plurality of content users for learning for each content provider. Count. The learning processing unit uses the attribute values of the correct answers of each of the plurality of content users for learning and the input data indicating the number of times of playback for each content provider of each of the plurality of content users for learning to generate an estimation model. learn.

The functions of the moving image viewing terminal 3, the moving image viewing server 4, the data collecting server 5, and the deep learning server 6 in the above-described embodiment are realized by a computer. In that case, a program for realizing these functions may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read by a computer system and executed. The term "computer system" as used herein includes hardware such as an OS and peripheral devices. Further, the "computer-readable recording medium" refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, or a CD-ROM, or a storage device such as a hard disk built in a computer system. Further, a "computer-readable recording medium" is a communication line for transmitting a program via a network such as the Internet or a communication line such as a telephone line, and dynamically holds the program for a short period of time. It may also include a program that holds a program for a certain period of time, such as a volatile memory inside a computer system that serves as a server or a client in that case. Further, the above-mentioned program may be a program for realizing a part of the above-mentioned functions, and may be a program for realizing the above-mentioned functions in combination with a program already recorded in the computer system.

Although the embodiments of the present invention have been described in detail with reference to the drawings, the specific configuration is not limited to this embodiment, and the design and the like within a range not deviating from the gist of the present invention are also included.

1 ... Attribute estimation system, 3 ... Video viewing terminal, 4 ... Video viewing server, 5 ... Data collection server, 6 ... Deep learning server, 9 ... Network, 31 ... Input unit, 32 ... Video viewing function unit, 33 ... Output Unit, 41 ... Storage unit, 42 ... Distribution unit, 43 ... Log generation unit, 51 ... Log collection function unit, 52 ... Log extraction function unit, 61 ... Ranking creation processing unit, 62 ... Learning data creation processing unit, 63 ... Deep Learning learning processing unit, 64 ... learning model function unit, 65 ... age estimation function unit

One aspect of the present invention is the attribute estimation device described above, wherein the log data includes information on the playback date and time of the content, and the playback count counting unit is the first indicated by the log data of the content user. The playback period for each content provider in the period is converted into a playback ratio in the second period shorter than the first period, and a predetermined value is randomly determined from the playback ratio for each content provider in the second period after conversion. The number of playbacks for each content provider when the number of playbacks is selected is counted .

Claims (7)

Using the provider information indicating the content provider that provided each of the plurality of contents and the log data indicating the content played on the terminal of the content user of the attribute estimation target, the content user of the attribute estimation target A playback count counting unit that counts the number of playbacks of the content on the terminal for each content provider,
The content provider counted by the play count unit in an estimation model that is a neural network that inputs a play count for each content provider and outputs a value indicating the probability that the attribute of the content user is each attribute value. A model application unit that inputs input data indicating the number of times of reproduction for each
An estimation unit that estimates the attribute value of the content user to be attribute-estimated based on the result output by the estimation model in response to the input of the input data, and an estimation unit.
Attribute estimation device including. The log data includes information on the playback date and time of the content.
The model application unit converts the number of times played by each content provider in the first period indicated by the log data of the content user into the number of times played in a second period shorter than the first period, and after conversion. The number of playbacks for each content provider when a predetermined number of playbacks is selected from the number of playbacks for each content provider in the second period is used as the input data.
The attribute estimation device according to claim 1. Using the provider information and the log data of the plurality of content users, the number of times the content is played on the terminal of the plurality of content users is counted for each content provider, and the counted reproduction is performed. Further provided with a ranking determination unit that assigns a ranking to the content provider based on the number of times.
The model application unit uses the number of times played by each of the content providers having a higher rank than a predetermined number and the total number of times played by each of the content providers having a rank lower than a predetermined value as the input data. ,
The attribute estimation device according to claim 1 or 2. Using the provider information and the log data of each of the plurality of content users for learning, the number of times the content is played on the terminal of each of the plurality of content users for learning is counted for each content provider. Playback count counting unit for learning and
The estimation model is learned using the attribute values of the correct answers of each of the plurality of content users for learning and the input data indicating the number of times of playback of each of the plurality of content users for learning for each content provider. Further equipped with a learning processing unit,
The attribute estimation device according to any one of claims 1 to 3. The content is a moving image
The attribute is age,
The attribute estimation device according to any one of claims 1 to 4. It is an attribute estimation method executed by the attribute estimation device.
The attribute estimation device uses the provider information indicating the content provider that provided each of the plurality of contents and the log data indicating the content reproduced on the terminal of the content user to be attribute estimation, and the attribute estimation target. The playback count counting step of counting the playback count of the content on the terminal of the content user for each content provider, and
In the playback count counting step, the attribute estimation device is a neural network in which the playback count for each content provider is input and a value representing the probability that the content user's attribute is each attribute value is output. A model application step for inputting input data indicating the number of times of playback for each of the counted content providers, and
An estimation step of estimating the attribute value of the content user to be attribute-estimated based on the result output by the estimation model in response to the input of the input data, and an estimation step.
Attribute estimation method with. Computer,
The program for functioning as the attribute estimation device according to any one of claims 1 to 5.

Images