WO2011036704A1

WO2011036704A1 - Content recommendation device

Info

Publication number: WO2011036704A1
Application number: PCT/JP2009/004812
Authority: WO
Inventors: 中田康太; 村上知子
Original assignee: 株式会社東芝
Priority date: 2009-09-24
Filing date: 2009-09-24
Publication date: 2011-03-31

Abstract

Provided is a content recommendation device capable of presenting content to meet the preference of a user from among large amounts of VoD content. The content recommendation device increases, in a pseudo manner, VoD content which is conceivable that a user wants to view on the basis of viewed VoD content, and selects a case effective for the recommendation of the VoD content from the viewing histories of broadcast content using the increased VoD content, thereby performing the learning of an effective recommendation model.

Description

Content recommendation device

The present invention relates to a content recommendation device for recommending video content to a user in a video on demand service.

With the digitization of content and the spread of the Internet environment, the video that can be used on the Internet has increased dramatically. Since the content and timing of viewing the video on the Internet can be freely selected, the user can enjoy content according to his / her interests and preferences. For example, in a general Video On Demand (hereinafter referred to as VoD) service, high-quality contents such as movies and sports are distributed in large quantities, and the user can view the contents according to his / her situation. On the other hand, as the video content increases, it is necessary for the user to select a video he / she wants to see from a huge amount of videos, and finding a video he / she really wants to see becomes increasingly difficult.

In response to this problem, video distribution sites provide information by expanding the search system based on genre information and presenting popular videos. By using such a method, it has become possible to quickly find a target video when the video that the user wants to see is clear or when the user wants to search for video based on a certain standard such as popularity. Also, development of a broadcast program recommendation system is progressing as a method for presenting video that is more tailored to the user's personality. The broadcast program recommendation system is a system that recommends a program that matches a user's preference from among an increasing number of broadcast programs as the number of broadcast channels increases. In such a recommendation system, it is expected to present a broadcast program that suits the user's preference by recommending the program using the program genre, metadata such as performers, and the user's viewing history.

However, in the conventional search system, the target video can be presented when the content that the user wants to view is clear, but when the content that the user wants to view is not clear, the video that suits the taste can be presented. difficult. In a system that presents a popular video, it is possible to present content that is generally viewed, but since the user's interests are different, it is not always possible to present content that suits the taste. For example, Japanese Patent Application Laid-Open No. 2006-127145 discloses that a viewer's preference at the time of viewing a program is appropriately reflected and presented to the viewer by reflecting a preference that changes with changes in program content ( Patent Document 1).

In such a case, it is considered effective to utilize conventional broadcast program recommendation technology in order to present content suitable for the user. However, when considering a general VoD service, most content is paid for viewing, and the user has fewer opportunities to view VoD content than broadcast programs. For this reason, the viewing history of VoD content is very small compared to a broadcast program. In addition, since there is a large amount of content, it is unlikely that all content will be viewed or not, and there is an imbalance problem that there are a lot of unviewed content compared to the viewed content. End up. Therefore, it is difficult to make highly accurate recommendations, particularly when recommendations are made using the viewing history of VoD content in the initial stage of service introduction.

When a recommendation is made using a large amount of viewing history, a method of using a relatively large amount of viewing history of a broadcast program can be considered as a method for supplementing the viewing history of VoD. However, there are many people who generally have different viewing styles for broadcast programs and VoD content, such as watching mainly for variety in broadcast programs, but mainly for movies in VoD content. Therefore, when the VoD content recommendation is performed using the viewing history of the broadcast program, the recommendation reflecting the user's preference for the broadcast program is performed, and an appropriate VoD content cannot be recommended.

JP 2006-127145 A

The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a content recommendation device capable of presenting content that matches a user's preference from a large amount of VoD content. .

In order to achieve the above object, the content recommendation device of the present invention includes a VoD positive example complementing unit that complements data similar to a positive example from negative examples of VoD learning data, and supplemented VoD learning data. The feature amount importance of each feature amount is calculated using the feature amount importance, and a selection processing unit that selects data suitable for VoD content recommendation from the broadcast learning data, and the complemented VoD learning data are selected. A recommendation model learning unit that learns a recommendation model using broadcast learning data and the calculated feature value importance, a determination unit that determines a recommendation based on a recommendation model learned for a VoD target content, and a recommendation determination result It has an output part which performs control presented to.

According to the present invention, it is possible to present content that suits the user's preference from among a large amount of VoD content.

The block diagram which shows the example of whole structure of the content recommendation apparatus based on the Example of this invention. The figure which shows an example of the metadata provided to the VoD content which concerns on an Example. The figure which shows an example of the VoD metadata produced | generated from the metadata of FIG. The figure which shows an example of the VoD viewing-and-listening data which concern on an Example. The figure which shows an example of the VoD learning data based on an Example. The flowchart which shows the specific example of the process in the VoD positive example complementation part which concerns on an Example. The figure which shows an example of the process which reverses the "negative example" of VoD learning data to the "positive example" using four content which concerns on an Example. The flowchart which shows the specific example of the process in the selection process part which concerns on an Example. The figure which shows an example of the importance provided to the feature-value of the performer which concerns on an Example. The figure which shows an example of the importance provided to the feature-value of the genre which concerns on an Example. The figure which shows an example of the importance provided to the feature-value of the keyword which concerns on an Example. The figure which shows an example of the positive example of the broadcast learning data which concerns on an Example. The figure which shows an example of the Bayesian network for recommendation model learning which concerns on an Example. The flowchart which shows the specific example of the process in the recommendation model learning part which concerns on an Example. It is the example which converted the performer feature-value and keyword feature-value of the learning data of FIG. 7 into binary. The flowchart which shows the specific example of the process in the determination part which concerns on embodiment. The figure which shows an example of the VoD object data which concern on embodiment. The figure which shows an example of the output by the output part which concerns on embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an example of the overall configuration of a content recommendation device according to an embodiment of the present invention. The content recommendation device 1 according to the present embodiment includes a VoD learning data storage unit 11, a broadcast learning data storage unit 12, a VoD correct example complementing unit 13, a selection processing unit 14, a recommended model learning unit 15, a VoD target data storage unit 16, A determination unit 17 and an output unit 18 are included.

Hereinafter, a specific configuration and operation of each unit in the content recommendation device 1 will be described.

The VoD learning data storage unit 11 is a storage device for storing VoD learning data. The VoD learning data includes VoD metadata and VoD viewing data. VoD metadata is data obtained from metadata attached to content.

FIG. 2 shows an example of metadata assigned to the VoD content. The metadata includes “content ID, genre, title, performer, time, summary” and the like. The “summary” is expressed as text data. In the present embodiment, the keywords assigned to the VoD metadata are extracted from the “summary”. The keyword is a word having a high frequency and a high TFIDF value from the text data in the “summary” of all contents. The word frequency and the TFIDF value can be easily calculated by using a conventional morphological analysis technique.

FIG. 3 is an example of VoD metadata generated from the metadata of FIG. Through the above-described processing, the keyword extracted from the “summary” is given instead of the “summary” information.

The VoD viewing data used in the present embodiment is information regarding viewing of the user's VoD content. Here, it is assumed that information related to viewing is obtained from the device for all VoD contents.

FIG. 4 shows an example of VoD viewing data. The content ID 41 of the VoD viewing data is an individual ID assigned to each VoD content. The viewing information 42 of VoD viewing data represents viewing-related information. For example, if the user is viewing “back to the xxx” corresponding to the content ID-1 in FIG. 2, the viewing information 42 is “viewing”. On the other hand, when the user has not viewed the contents ID-2 and ID-3, the viewing information 42 is “unviewed”. Depending on the device, it is possible to give “favorite” information to the VoD content that the user wants to see in the future. For example, when information that the user wants to view in the future is given to the content ID-4, the viewing information 42 becomes “favorite”.

FIG. 5 shows an example of VoD learning data. The VoD learning data is obtained by adding VoD viewing data to the VoD metadata of each content. In the recommended model learning in the present embodiment, learning is performed by generating “positive example” and “negative example” based on the VoD viewing data of FIG. 4 as a teacher signal. Here, it is assumed that the “positive example” is what the user desires to watch, and the “negative example” does not want to watch. Therefore, “viewed” and “favorite” of the VoD learning data are set as “positive examples”, and “unviewed” are set as “negative examples”.

Further, in this embodiment, each value of the performer, keyword, and genre of VoD metadata is set as a “feature amount” representing the content. In the example of FIG. 5, the feature amount of the content ID-1 is “Robert 〇 △ 〇 △”, “Michael J. 〇 △ □ ○ □”, “Time”, “Machine”, “Past”, “Movie” -Western film. A set of performers such as [Robert 〇 △ 〇 △, Michael J. ○ △ □ ○ □] in the VoD metadata is called a performer feature amount of the content ID-1. A set of keywords such as [time, machine, past] is called a keyword feature amount of content ID-1. Furthermore, a set of genres such as [movie-foreign film] is called a genre feature amount of content ID-1.

Next, the broadcast learning data storage unit 12 is a storage device for storing broadcast learning data. Broadcast learning data includes broadcast metadata and broadcast viewing data. Broadcast learning data is stored in the same format as VoD learning data.

Since information such as title, genre, performer, and outline can be obtained from the VoD content and the broadcast content, it is possible to generate data similar to the VoD learning data. With regard to broadcast metadata, for example, there may be cases where the genre system and expression differ between VoD content and broadcast content, but this can be solved by associating the genre of broadcast content with the genre of VoD content. Further, it is assumed that the same word as the keyword used in the VoD metadata is also used in the broadcast metadata. Through the above processing, broadcast metadata in the same format as VoD metadata can be obtained for broadcast content.

Broadcast viewing data is information related to viewing of the broadcast content of the user. Here, it is assumed that information viewed by the user is obtained directly or indirectly from the television receiver for all broadcast contents of all channels. For example, it is possible to obtain data in the same format as the VoD viewing data by setting “viewing” for broadcast content that has viewed 20% or more of a whole broadcast program and “not viewing” for other content. is there.

Broadcast learning data in the same format as VoD learning data can be obtained by using broadcast metadata and broadcast viewing data. The definitions of “positive example” and “negative example” and the definition of the feature amount are the same as those of the VoD learning data.

The VoD correct example complementing unit 13 performs a correct example complement process including a similar title selection process, a similar metadata selection process, and a correct example inversion process. The similar title selection process is a process of selecting a “negative example” similar to the “positive example” title from the “negative examples” of the VoD learning data acquired from the VoD learning data storage unit 11. The similar metadata selection process is a process of selecting a “negative example” in which performers and keywords overlap from the selected “negative example”. The positive example inversion process is a process of converting the selected “negative example” into the “positive example”. The VoD positive example complementing unit 13 is realized by executing a program for the positive example complementing process by the processor.

FIG. 6 is a flowchart showing a specific example of the processing operation of the VoD correct example complementing unit 13. First, the VoD correct example complementing unit 13 determines whether there is a “correct example” that is not selected in the “correct example” of the VoD learning data (step S131). If the “positive example” that has not been selected remains in the VoD learning data, the “positive example” that has not yet been selected is selected (step S132). On the other hand, if the “positive example” of all VoD learning data has been selected, the process ends (No in step S131).

Next, the VoD positive example complementing unit 13 determines whether there is a “negative example” that is not selected in the “negative example” of the VoD learning data (step S133). If the “negative example” that has not been selected remains in the VoD learning data, the “negative example” that has not yet been selected is selected (step S134). On the other hand, if “negative example” of all VoD learning data has been selected, the process returns to step S131.

Then, the VoD positive example complementing unit 13 calculates a title score representing the similarity between the “negative example” of the selected VoD learning data and the “positive example” title of the selected VoD learning data (step S135). . Then, it is determined whether or not the calculated title score is equal to or greater than a threshold set by the user (step S136). If the calculated title score is equal to or greater than the threshold, the process proceeds to step S137, and if less than the threshold, the process returns to step S133.

Similarly, the VoD positive example complementing unit 13 generates a metadata score representing the degree of commonality between the “negative example” of the selected VoD learning data and the performers and keywords of the “positive example” of the selected VoD learning data. Is calculated (step S137). Then, it is determined whether or not the calculated metadata score is equal to or greater than a threshold set by the user (step S138). If it is equal to or greater than the threshold, the process proceeds to step S139, and if it is less than the threshold, the process returns to step S133. The VoD correct example complementing unit 13 inverts the “negative example” of the VoD learning data whose title score and metadata score are both equal to or greater than the threshold value to the label “positive example”, and stores the contents (step S139).

With the above processing, the VoD positive example complementing unit 13 can extract the content that should actually be handled as the “positive example” from the “negative example” of the VoD learning data. Next, a specific operation of the VoD correct example complementing unit 13 will be described using the four contents of FIG.

Formula (1) is a title score calculation example executed in the process of step S135 of FIG. The sub_N (title) in the expression (1) represents an arbitrary N character partial character string in the title. It should be understood that a part of the expression is different between the expression (1) and the expression of the expression (1). The same applies to the explanation of other equations.

In this embodiment, the title score is calculated using the selected title “title_p” of “positive example” and the selected title “title_n” of “negative example”. In the expression (1), a title score that is “1” is used when an arbitrary N-character partial character string of title_p is included in title_n, and “0” is used otherwise. In this embodiment, N = 4 if the title is longer than 4 characters, and N = 2 otherwise.

For example, it is assumed that the content ID-1 in FIG. 7 is selected as the “positive example” and the content ID-2 is selected as the “negative example”. At this time, one of the N = 4 partial character strings “back •” of the title of “correct example” is included at the beginning of the title of content ID-2, and therefore the title score value is “1”. Become. In another example, it is assumed that content ID-1 is selected as “positive example” and content ID-4 is selected as “negative example”. In this example, since the N = 4 partial character string of the title “positive example” is not included in the title “negative example”, the value of the title score is “0”. In the process of step S136 of FIG. 6, if the threshold value is “1”, if the selected “positive example” is content ID-1 and “negative example” is content ID-2, the content ID-2 is It is extracted as a “correct example” candidate and proceeds to step S137. However, if the selected “positive example” is the content ID-1 and the “negative example” is the content ID-4, the content ID-4 is not a “positive example” candidate, and the process goes to step S133. Return.

Formula (2) is a calculation example of the metadata score executed in the process of step S137 in FIG. [Guest] _ {p, n} in Expression (2) represents a set of performers of “positive example” and “negative example”.

In this embodiment, the metadata score is determined according to the performers of the selected “positive example” and “negative example” and the degree of overlapping of keywords. In Formula (2), the value is “1” when the overlap number of the performers of “positive example” and “negative example” selected is greater than k, and “0” otherwise. Metadata score is used. In this embodiment, k = 1.

Here, it is assumed that the content ID-1 in FIG. 7 is selected as the “positive example” and the content ID-2 is selected as the “negative example”. At this time, since the title score in step S136 is “1” as described above, both metadata scores are calculated in the process in step S137. In this case, since the content ID-1 and ID-2 are common to the two performers, | [guest_p] ∩ [guest_n | = 2], and the metadata score is “1”.

On the other hand, it is assumed that the content ID-3 is selected as the “positive example” and the content ID-4 is selected as the “negative example”. In this case, the title score in step S136 is “1”. However, in step S137, content ID-1 and ID-2 are not the same performers, so | [guest_p] ∩ [guest_n | = 0. The data score is “0”.

In the processing of the next step S138, if the threshold is set to “1”, if the selected “positive example” is the content ID-1 and “negative example” is the content ID-2, the content ID-2 is the content of step S139. A pseudo “positive example” label is given in the process. As a result, the content ID-2 is treated as a pseudo “positive example” in all subsequent processes. However, if the selected “positive example” is the content ID-3 and the “negative example” is the content ID4, the content ID-4 is treated as a “negative example” as it is in all the processes.

As an effect of the VoD positive example complementing process 13, it is possible to perform highly accurate learning by inverting “negative example” that should be handled as a positive example to “positive example”, and at the same time, a small amount of “positive example”. The number can be increased.

As mentioned above, VoD content often includes all content such as movies and drama series and sports tournament videos. At this time, if the content viewed as before is treated as a “positive example” and the unviewed content is treated as a “negative example”, for example, when only the first episode of a series of all 10 episodes is viewed, All are treated as “negative examples”, and it is considered that recommendations with high accuracy cannot be made. For this reason, in the present embodiment, content similar in title representing the same series or the like using the title score is processed as a pseudo “positive example” candidate. On the other hand, for example, content ID-3 and ID-4 are similar in content but contain a lot of completely different content. Therefore, the metadata score is further used to determine whether or not the pseudo “right example” candidate is content that should be treated as a “pseudo“ right example ”.

In the present embodiment, when the performers overlap like content ID-1 and ID-2, the value of the metadata score becomes high and is treated as a pseudo “positive example”. On the other hand, if the performers do not overlap like content ID-3 and ID-4, it is not treated as a pseudo “positive example” because there is no relationship such as series. Therefore, it is expected that the “negative example” that should actually be handled as a positive example can be handled as a “positive example” by the VoD processing complementing unit 13, and a highly recommended recommendation model can be learned.

Also, since VoD content is often distributed by paid services, it is considered that there are a small number of “positive examples”. In the present embodiment, for example, by treating the remaining nine episodes of the series as “positive examples”, it becomes possible to obtain more “positive examples”.

The selection processing unit 14 performs selection processing including feature amount importance calculation processing and case selection processing. The feature amount importance calculation processing is processing for calculating the importance of each feature amount using the VoD learning data acquired from the VoD example complementing unit 13. The case selection process is a process of calculating the reliability of broadcast learning data using the calculated feature value importance and holding the broadcast learning data having a high value as learning data. The selection processing unit 14 is realized by executing a program for selection processing by the processor.

FIG. 8 is a flowchart showing a specific example of the processing operation of the selection processing unit 14. First, the selection processing unit 14 determines whether there is a feature amount that is not selected among the feature amounts of the VoD learning data (step S141). If there is a feature amount that has not yet been selected from the feature amounts of the VoD learning data, the selection processing unit 14 selects it (step S142). If all the feature values have been selected, the process proceeds to step S145.

Next, the selection processing unit 14 calculates the importance of the selected feature amount (step S143). The selection processing unit 14 gives the calculated importance to the feature amount (step S144), and returns to step S141. The processes in steps S142 to S144 are repeatedly executed until there is no unselected feature quantity among the feature quantities of the VoD learning data.

When all the feature values have been selected, the selection processing unit 14 determines whether or not there is a “correct example” not selected in the “correct example” of the broadcast learning data (step S145). If there is a “correct example” that has not yet been selected from the “correct example” of the broadcast learning data, the selection processing unit 14 selects it (step S146). If all broadcast learning data has been selected, the process ends. The selection processing unit 14 calculates the reliability of the selected broadcast learning data (step S147). The selection processing unit 14 determines whether or not the calculated reliability is greater than or equal to a threshold (step S148), and if it is less than the threshold, the process returns to step S145. When the reliability is greater than or equal to the threshold, the selection processing unit 14 assigns the calculated reliability to the broadcast learning data (step S149), and returns to step S145. The processes of steps S146 to S149 are repeatedly executed until there is no “positive example” not selected in the broadcast learning data.

Through the above processing, the selection processing unit 14 can select data that can be handled in the same way as the VoD learning data from “positive examples” of the broadcast learning data, and can use it as learning data together with the VoD learning data. It is.

In the present embodiment, the feature amount of the VoD learning data uses performers, keywords, and genres, and the selection processing unit 14 can assign importance to all feature amounts. The importance of the feature amount indicates the strength of the correlation between the feature amount and the “positive example”. That is, the feature quantity with high importance strongly reflects the user's preference. In the present embodiment, importance can be given using different criteria for performer feature values, keyword feature values, and genre feature values.

Formula (3) is a calculation example of the importance CR of the guest's feature quantity guest_i. In Expression (3), P (guest_i) is the number of VoD learning data in which guest_i appears, and P (guest_i | viewed) is the number of data in which guest_i appears in the VoD learning data. This score takes a value from 0 to 1, and is a high value when a lot of content in which guest_i appears is viewed.

Expression (4) is an example of the importance GS of the keyword feature quantity keyword_j. Here, P (keyword_j | viewed) is the number of data including keyword_j in the positive example of VoD learning data. P (keyword_j | not viewed) is the number of data including keyword_j in the negative example of VoD learning data. coef is a correction coefficient. This value is a value from 0 to 1 based on the Graham score, and is a high value when a large amount of content including keyword_j is viewed.

Formula (5) is a calculation example of the importance GI of the genre feature quantity genre_l. Here, [genre_viewed] represents the genre set of the content viewed. The importance value is “1” if the genre_l has been viewed, and “0” otherwise.

FIG. 9 shows an example of the importance CR assigned to the feature amount of the performer. In this example, since the content in which “Michael J. ○ △ □ ○ □” appears is often viewed, the importance CR is high. On the other hand, since the content in which “Antonio 〇 △△△ □” appears is rarely viewed, the importance CR is low.

FIG. 10 shows an example of the importance GS assigned to the keyword feature. In this example, since the content including “future” is often viewed, the importance GS is high. On the other hand, since the content including “Legend” is not viewed much, the importance GS is low.

FIG. 11 shows an example of the degree of importance assigned to the genre feature quantity GI. In this example, since the contents of “movie-foreign” and “movie-Japanese” are viewed, the importance GI is “1”. On the other hand, since the content of “animation-domestic animation” is not viewed, the importance GI is “0”.

The selection processing unit 14 calculates the reliability of the broadcast content by the process of step S147 of FIG. 8 using the importance of the feature amount calculated by the above method. In this embodiment, different reliability calculation methods are provided for performers, keywords, and genres.

Formula (6) is a calculation example of the performer reliability calculated by the importance CR of the performer's feature amount. In Expression (6), T_J represents the selected broadcast learning data. The reliability of the performer of the broadcast learning data T_J is obtained by adding the importance of the performer to the broadcast learning data T_J.

Expression (7) is a calculation example of the reliability of the keyword calculated from the importance GS of the keyword feature amount. In Expression (7), the reliability of the keyword of the broadcast learning data T_J becomes higher as T_J includes more important keywords. In this embodiment, the importance level of the keyword is expressed by Expression (4), and the importance level of the keyword included in a large amount of the viewed VoD content increases. Therefore, the more reliable the viewed VoD content and the broadcast learning data T_J keyword, the greater the reliability of the T_J keyword.

Expression (8) is a calculation example of the genre reliability calculated based on the importance of the genre feature quantity GI. In Expression (8), the genre reliability of the broadcast learning data T_J is calculated by taking the sum of the importance levels of the genres included in the broadcast learning data T_J. There may be only one genre of broadcast learning data, but some broadcast learning data has multiple genres such as "Information / Wide Show-Entertainment / Wide Show" and "News / Report-Other". Sometimes given at the same time. Therefore, the sum of importance levels is adopted.

Through the above processing, the performer reliability, keyword reliability, and genre reliability of the selected broadcast learning data are calculated.

The selection processing unit 14 determines whether or not to add the broadcast learning data to the “positive example” depending on whether or not each reliability of the selected broadcast learning data T_J is equal to or greater than a threshold in the process of step S148 of FIG. To do. In the present embodiment, if any one of the reliability levels exceeds the threshold, the broadcast learning data is added to the “positive example”.

FIG. 12 shows an example of a positive example of broadcast learning data. With reference to the example of FIG. 12, the process of holding the “correct example” of the broadcast learning data of the selection processing unit 14 (step S149 in FIG. 8) will be specifically described. In the present embodiment, it is assumed that the performer reliability threshold is 0.7, the keyword reliability threshold is 0.99, and the genre reliability threshold is “1”.

If the selection processing unit 14 selects the content ID-T1, the reliability value of the performer exceeds the performer reliability threshold value 0.7. That is, since the performers of the content ID-T1 include “Michael J. ○ △ □ ○ □”, the importance level 0.8 of the performers in FIG. 9 exceeds the threshold value 0.7. . In the content ID-T1, the reliability of the genre and the reliability of the keyword are less than the threshold, but the reliability of the performer is greater than or equal to the threshold. Therefore, the selection processing unit 14 sets the content ID-T1 to “ It is added to the broadcast learning data as a “positive example”.

If the selection processing unit 14 selects the content ID-T2, the genre reliability value is the same as the genre reliability threshold “1”. That is, since the genre of the content ID-T2 is “movie-foreign film”, the importance “1” and the threshold “1” of the genre in FIG. 11 have the same value. In the content ID-T2, the reliability of the performer and the reliability of the keyword are less than the threshold value. However, since the reliability of the performer is greater than or equal to the threshold value, the selection processing unit 14 sets the content ID-T2 in the process of step S149. It is added to the broadcast learning data as a “positive example”.

However, when the selection processing unit 14 selects the content ID-T3, the reliability of the performer of the content ID-T3, the reliability of the genre, and the reliability of the keyword are all below the threshold. Therefore, the selection processing unit 14 proceeds from step S148 to step S145, and does not add the content ID-T3 as “correct example” to the broadcast learning data. Similarly, the content ID-T4 is not added to the “primary example”.

As an effect of this selection processing unit 14, it is possible to use only “positive examples” that are effective for the recommended model of VoD learning data from the broadcast learning data.

In general, broadcast content and VoD content often have different trends in content viewed by users. Therefore, “positive examples” of the broadcast learning data include “positive examples” having the same tendency as the VoD learning data and “positive examples” having a tendency unique to broadcast contents.

In the broadcast content shown in FIG. 12, “000 market” and “Mr. 00” are viewed as broadcast content, but are not always desired to be viewed when provided as VoD content. Absent. On the other hand, “0000 Roman ...” and “00.000. Of Caribbean” have the same reason as the “positive example” of VoD learning data, and are used when learning a recommended model of VoD learning data. It is considered appropriate as a “positive example”. Accordingly, the selection processing unit 14 can increase the “positive examples” used for learning the recommended model, and at the same time can exclude the “positive examples” peculiar to the broadcast learning data, so that accurate learning can be expected.

Next, the recommendation model learning unit 15 learns the user's VoD content recommendation model using the learning data composed of the “positive examples” of the VoD learning data and the stored broadcast learning data. In this embodiment, the Bayesian network method is used for learning the recommendation model.

FIG. 13 shows an example of a recommended model learning Bayesian network. The structure of the Bayesian network is such that “genre”, “performer”, and “keyword” nodes are connected to the viewing node. Similar to the conventional learning method, the recommendation model can be learned by calculating a conditional probability table of each node from the learning data.

FIG. 14 is a flowchart showing a specific example of processing in the recommended model learning unit 15. First, the recommended model learning unit 15 determines whether there is any learning data that has not yet been selected (step S151). If there is learning data that has not yet been selected from the learning data, the recommended model learning unit 15 selects the learning data (step S152). If all the learning data has been selected, the process proceeds to step S156. The recommended model learning unit 15 determines whether there is a node that has not been selected in the selected learning data (step S153). If there is a node that has not been selected in the selected learning data, the recommended model learning unit 15 selects that node (step S154). If all nodes have been selected, the process ends. Next, the recommended model learning unit 15 converts the feature amount of the selected node into a discrete value (step S155). If there is a node that has not yet been selected in the selected learning data, the processes of steps S153 to S155 are repeatedly executed. And the recommendation model by a Bayesian network is learned using the learning data by which each node was converted into the discrete value (step S156).

In the learning method using the Bayesian network of this embodiment, the nodes are the feature amount of the performer (for example, see FIG. 9) and the feature amount of the keyword (for example, see FIG. 10).

In this embodiment, the recommended model learning unit 15 converts the feature quantity of the performer and the keyword feature quantity of the learning data into binary discrete values of “preference” and “normal” as shown in FIG. In other words, in this embodiment, when the values of the equations (6) and (7) used for the reliability calculation are used as scores, the sum of the performer scores and the keyword scores of the learning data are each equal to or greater than the threshold value. When “preferred” or less than the threshold value, “normal” can be converted into a binary discrete value.

In order to perform this conversion, usually, all the learning data is selected once, and the score of each performer or keyword is calculated. Based on the sum of the performer and keyword scores of each learning data, “preference” “normal” ”Is required. However, in the present embodiment, the recommended model learning unit 15 receives the scores based on the equations (6) and (7) already calculated by the selection processing unit 14, and therefore performs a process of calculating the weight of each performer and keyword. Can be omitted.

FIG. 15 shows an example of converting the feature quantity of the performer in FIG. 9 and the feature quantity of the keyword in FIG. 10 into binary for the VoD content in FIG. For example, according to Expression (6), the score of the performer “Michael J. ○ Δ □ ○ □” of the VoD learning data of the content ID-1 is 0.8. If the threshold value is 0.7 as in the case of the broadcast content, the score is equal to or higher than the threshold value, so the feature amount of the performer is converted to “preference”. On the other hand, the performer score of the VoD learning data of content ID-4 is 0.1, which is less than the threshold value, so the performer feature value is converted to “normal”. As described above, the learning data is converted using the importance of the feature amount calculated in advance, and the CPT is created using the learning data in the same manner as in the conventional Bayesian network learning method. Can learn the recommended model by the process of step S156 in FIG. 14 without recalculating the weight of each feature quantity.

Next, the VoD target data storage unit 16 is a storage device that stores VoD learning data to be recommended. The VoD target data is stored in the same format as the VoD metadata.

Next, the determination unit 17 selects unselected VoD target data from the VoD target data stored in the VoD target data storage unit 16, calculates a filter value for the selected VoD target data, and the filter value is The VoD target data that is equal to or greater than the threshold value is determined by the recommended model learned by the recommendation model learning unit 15, and the determination result is given to the VoD target data. The determination unit 17 is realized by executing a program for the series of processes by the processor.

FIG. 16 is a flowchart illustrating a specific example of processing in the determination unit 17. The determination unit 17 determines whether there is data that is not selected in the VoD target data (step S171). If there is VoD target data that has not been selected in the VoD target data, the determination unit 17 selects the VoD target data (step S172). If all the VoD target data has been selected, the process ends. Next, the determination unit 17 calculates a filter value of the selected VoD target data (step S173). The determination unit 17 determines whether or not the calculated filter value is greater than or equal to a threshold value (step S174). If the calculated filter value is less than the threshold value, the process proceeds to step S176. When the calculated filter value is equal to or greater than the threshold value, the determination unit 17 determines the VoD target data using the recommended model learned by the recommendation model learning unit 15 (step S175). The determination unit 17 assigns the determination result to the VoD target data (step S176), and returns to step S171.

The determination unit 17 calculates the filter value in the process of step S173, and determines only the VoD target data whose filter value is equal to or greater than the threshold using the recommended model in the process of step S175. Generally, since it takes time to make a determination in a model learned by a Bayesian network, it is possible to eliminate the “negative example” VoD target data that is obvious by setting a filter value with a short calculation time in advance, and to reduce the overall calculation cost. Can be reduced. In the present embodiment, the importance of the feature amount calculated by the selection processing unit 14 can be used for calculating the filter value in the process of step S173.

Equation (9) shows an example of the filter value. Where id_obj, guest_obj, keyword_obj, and genre_obj are any elements such as the content ID, performer, keyword, and genre of the selected VoD target data, and [guest] ^ 5_1, [keyword] ^ 5_1, and [genre] ^ 5_1 , Performers, keywords, and genres, the set of the top five importance levels calculated by the selection processing unit 14.

Such filter values include any of the top five performers with the highest importance of performers, the top five keywords with the highest importance of keywords, and the top five genres with the highest importance of genres. The value is “1” when the value is “0”, and “0” otherwise.

If the threshold value is set to “1” in the process of step S174 in FIG. 16, the VoD target data including any one of the performers, keywords, and genres with high importance is recommended model in the process of step S175. More detailed discrimination is performed using. However, it is assumed that the VoD learning data that does not include any of them is not determined by the recommended model because it is clearly not recommended.

In the next step S176, the determination unit 17 gives a determination result to the VoD target data. Here, when the result is obtained from the recommended model, the numerical value of the result is given, and when the result is not obtained from the recommended model, 0.0 is given as an obvious “negative example”.

FIG. 17 is an example of VoD target data. When the determination unit 17 selects the content ID-R1, since the content ID-R1 includes “movie-foreign film” having a higher genre importance, the filter value is “ 1 ". Therefore, the determination unit 17 performs detailed determination using the recommended model in the process of step S175. When the viewing probability calculated by the determination is 0.45, for example, the value is assigned to the VoD target data.

Further, when the determination unit 17 selects the content ID-R2, the content ID-R2 does not include higher ranks in the importance of the performer, the importance of the keyword, and the importance of the genre. The process proceeds from step S174 to step S176. In step S176, the determination unit 17 assigns a value of 0.0 to the VoD target data.

Through the above processing, the determination unit 17 removes data that is clearly determined as “negative example” from the VoD target data in advance by the filter value, thereby reducing the number of determinations by the recommended model and reducing the calculation cost. By using the importance of the feature amount calculated by the selection processing unit 14 as the filter value, it is not necessary to perform a calculation for the filter value, so that a new calculation cost does not occur. For this reason, it is possible to perform recommendation with low calculation cost even for VoD target data that generally exists in large quantities.

Next, the output unit 18 selects the VoD target data having a higher viewing probability from the VoD target data to which the determination result is given by the determination unit 17, and the feature having high importance included in the selected VoD target data. The amount is selected as a reason, and control is performed to present the selected VoD target data and the reason to the user. The output unit 18 is realized by executing a program for the series of processes by the processor.

FIG. 18 is an example of output by the output unit 18. Here, the title, genre, performer, and outline are displayed with the data having the highest viewing probability among the VoD target data as “new content recommendation”. If the VoD target data includes a higher importance level of the feature amount calculated by the selection processing unit 14, it is considered that this is an important reason for recommending the VoD learning data. It is possible to emphasize and display the most important feature amount (for example, coloring, underlining, bold text, shading, etc.). With such a display, the user can find content that suits each taste without work such as search.

In addition, the VoD learning data that has become “pseudo positive example” in the VoD correct example complementing unit 13 has not been viewed by the user. Therefore, by displaying the “pseudo positive example” at the same time, the user can find not only new contents but also contents of series that are missed at the same time.

Such a display method is possible because the importance of the feature amount and the pseudo positive example are calculated in advance by the VoD correct example complementing unit 13 and the selection processing unit 14.

The content recommendation apparatus according to the present embodiment is effective for recommending VoD content from a viewing history of broadcast content using a pseudo increase of the VoD content that the user would like to view based on the viewed VoD content. By selecting a case, it is possible to learn an effective recommendation model.

The content recommendation device according to the present embodiment may be realized by installing a program in a computer device in advance. Further, the program may be realized by storing the program in a storage medium such as a CD-ROM or distributing the program through a network and installing the program in a computer apparatus as appropriate. The VoD learning data storage unit 11, the broadcast learning data storage unit 12, and the VoD target data storage unit 16 are a memory, a hard disk or a CD-R, a CD-RW, a DVD-RAM, It can be realized by appropriately using a storage medium such as a DVD-R.

In addition, the said Example is not limited as it is, In an implementation stage, a component can be deform | transformed and embodied in the range which does not deviate from the summary. Various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the above embodiments.

DESCRIPTION OF SYMBOLS 11 ... VoD learning data storage part 12 ... Broadcast learning data storage part 13 ... VoD example supplement part 14 ... Selection processing part 15 ... Recommended model learning part 16 ... VoD object data storage part 17 ... Determination part,
18 Output section

Claims

A VoD learning data storage unit for storing a plurality of VoD learning data having metadata of a plurality of VoD contents and data relating to viewing of the plurality of VoD contents;
A broadcast learning data storage unit for storing a plurality of broadcast learning data having metadata of a plurality of broadcast contents and data relating to viewing of the plurality of broadcast contents;
A VoD positive example complementing unit that converts data similar to a positive example among positive examples of the plurality of VoD learning data stored in the VoD learning data storage unit;
For each metadata of the VoD learning data supplemented by the VoD positive example complementing unit, a feature amount importance indicating the strength of correlation with the positive example is calculated, and the broadcast learning is performed using the calculated feature amount importance. A selection processing unit for selecting broadcast learning data suitable for recommendation of VoD content from the data;
A recommended model learning unit that learns a recommended model of VoD content using the selected broadcast learning data and the calculated feature value importance;
A VoD target data storage unit storing metadata of a plurality of VoD contents to be recommended;
A determination unit configured to determine recommendation for a plurality of VoD contents to be recommended based on a recommendation model of the VoD content;
An output unit for outputting a determination result of recommendation by the determination unit;
A content recommendation device comprising:
The VoD positive example complement part is:
A similar title selection process for selecting a negative example similar to a positive example title from negative examples of the VoD learning data;
Similar metadata selection processing for selecting a negative example in which performers and keywords overlap among negative examples selected in the similar title selection processing,
A positive example inversion process for converting a negative example selected in the similar metadata selection process into a positive example;
The content recommendation device according to claim 1, wherein:
The selection processing unit
A feature amount importance calculation process for calculating the importance of each feature amount using the VoD learning data;
Example selection processing for calculating the reliability of the broadcast learning data using the feature amount importance calculated in the feature amount importance calculation processing, and selecting broadcast learning data appropriate for recommendation of the VoD content based on the value When,
The content recommendation device according to claim 1, wherein:
The determination unit
A filtering process for selecting only VoD target data including a feature quantity having a high importance using the calculated feature quantity importance;
A determination process for performing detailed determination on the selected VoD target data using the recommended model and giving a result thereof;
The content recommendation device according to claim 1, wherein:
The content recommendation device according to claim 1, wherein the output unit presents a result of a recommendation higher than the recommendation reason using the determination result and the feature amount importance.
The content recommendation device according to claim 1, wherein the output unit presents VoD learning data having similar titles together with a recommendation reason using the complemented VoD learning data.