JP2008244602A - Display of recommended information of broadcast program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a display of recommended information of broadcast program in order to recommend more useful viewing and listening information for the user. <P>SOLUTION: The display of recommended information of broadcast program comprises a means for inputting an evaluation (Good/No-Good) according to the taste of a viewer, a means for inputting an evaluation of the list of genre common to the viewers, a first totalization means for acquiring the evaluation of the list of common genre inputted by each viewer viewing the same program for each broadcast program and totalizing the evaluations, a second totalization means for acquiring the evaluation according to the taste of the viewer for each broadcast program through a work and totalizing the evaluations, and a means for displaying the evaluation of each broadcast program totalized by the first and second totalization means. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a display device that displays recommendation information of a broadcast program such as a television program.

As society has become informatized, the information that we can touch has exploded. The increase in information enables a flexible selection by the user and leads to a richer life. However, too much information far exceeds the information processing ability of the user and conversely hinders the user's life.
For example, in recent years, digitalization of broadcasting has progressed, and TV content has also increased rapidly, enabling users to view programs that meet their needs. However, in the past, it was easy to check what programs were being broadcast by zapping because the number of channels was small, but now that the number of channels has increased, it is only necessary to check what programs are being broadcast. I spend a lot of time.
In addition, a method of searching for a program to be viewed from an information magazine is also a common method, but such a magazine often introduces programs for everyone and does not meet the niche demands of users. In 2011, a transition from analog terrestrial broadcasting to digital terrestrial broadcasting is also planned, and such problems are expected to become more prominent in the future.

  As a conventional technique for solving such a problem, there is an information search technique. This is intended to efficiently find necessary information from a vast amount of information. A typical example is a search engine. Search engines search for and return desired information from a vast amount of information scattered across the Internet. However, if the user does not input an appropriate search expression, the desired information cannot be obtained. In addition, the search result often has a huge amount. In this case, it is necessary to further narrow down desired information from the huge search result. Information retrieval is also highly dependent on user capabilities.

Therefore, there is information recommendation as a method for solving such a problem of information retrieval. Information recommendation is a technique in which a system guesses a user's preference or request, automatically searches for matching information, and presents it to the user. Information recommendation does not depend on the user's ability like information retrieval, so the user can obtain desired information without being aware of anything. Information recommendation technology that places little burden on users is very important in an information-rich society.
However, the conventional information recommendation method mainly places importance on the user's preference and does not consider the quality of information in detail. As a result, there are many cases where good recommendation results cannot be obtained. The quality of information here means how valuable the information is for the user. No matter how valuable that information is, information that is completely useless to the user is worthless. Moreover, since values differ depending on the user, information valuable to a certain user is not necessarily valuable to other users. That is, it can be said that the quality of information is determined by each user.

Based on the above background, the present invention deals with the information recommendation method. In particular, the quality of information that has been neglected in the conventional recommendation method will be taken up, and a method for realizing improvement of the recommendation result will be searched. As a preferred example of the information recommendation target, a TV program is considered in the future digital broadcasting society.
The conventional TV program recommendation system has aimed to guess a program that suits the user's preference. However, from the viewpoint of the quality of information, that is, the value of program content, it can be said that the recommendation system desired by the user recommends a program that suits his / her preference and that is worth seeing for himself / herself.

  The present invention has been made in view of the above-described conventional circumstances, and a problem to be solved by the present invention is to provide a recommended information display device for broadcast programs in order to make it possible to recommend viewing information more useful to the user. is there.

  In order to solve the above problems, the present invention evaluates the broadcast program information with a means for inputting an evaluation (Good / No-Good) according to the viewer's preference and a list of genres common to the viewer. Means for inputting and, for each broadcast program, obtaining evaluations made on the list of common genres inputted by each viewer of viewers viewing the same program through the network, and summing up First tabulation means, and for each broadcast program, an evaluation according to the viewer's preference is obtained through the work, and the second tabulation means for tabulating and the first and second tabulations It has a display means which displays the evaluation of each broadcast program totaled by the means.

  More preferably, the display means displays the evaluation of the broadcast program aggregated by the second aggregation means for the broadcast program being viewed as time series information together with the viewing rate of the program being viewed. .

  The embodiment according to the present invention aims to improve the accuracy of program recommendation in consideration of the quality of the program. Therefore, first, a typical typical recommendation method will be analyzed, and then the effective use of individual user information, which is an important factor in determining the quality of information, will be described. And the problem that appears because the conventional recommendation method does not consider the quality of the program is described.

(Conventional information recommendation method)
There have been many studies on information recommendation [1], and representative methods include content-based recommendation [2] and collaborative recommendation [3].
(References)
[1] Adomavicius, G and Tuzhilin A .: Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions; IEEE Trans. KDE, Vol.17, No.6, pp.734 -749, 2005.
[2] M. Balabanovic and Y. Shoham: Fab: Content-Based, Collaborative Recommendation; Comm. ACM, Vol.40, No.3, pp.66-72, 1997.
[3] P. Resnick, N. Iakovou, M. Sushak, P. Bergstorm, and J. Riedl: GroupLens: An Open Architecture for Collaborative Filtering of Netnews; Proc. Of 1994 ACM Conference on Computer Supported Cooperative Work pp.175- 186., 1994.

  Content-based recommendation is a method of recommending based on user viewing history information. The system learns what the user likes from the viewing history of the user, and presents to the user what matches that preference. Therefore, this method can be expected to recommend information as desired by the user. However, information that deviates from the user's viewing history is not recommended, and for example, it is not possible to experience that the user himself / herself has not yet noticed his / her preference. FIG. 1 schematically shows content-based recommendation in program recommendation. The large circle represents the preference of the viewer A required from the viewing history of the viewer A. Viewer A is recommended a program included in this preference, and a highly accurate recommendation result is obtained.

  On the other hand, the collaborative recommendation is a method of making a recommendation with reference to history information of a user whose preference is similar to that of the user. It recommends information that suits the tastes of users who have similar tastes and that is not yet viewed by them. FIG. 2 schematically shows this. Viewers A and B have similar preferences, and their preferences are similar. Therefore, even if the program is not included in the preference of the viewer A, it is presumed that the viewer A will surely like it if it is included in the preference of the viewer B, and the program is recommended to the viewer A. In this method, since the viewing history of other users is used, information recommendation is effectively performed if there is a wealth of preference information of each user and various users. It also solves the problem of finding new preferences, which is a drawback of content-based recommendation.

However, it has also been pointed out that collaborative recommendation has a “collaborative horizon” of information propagation [4]. In collaborative recommendation, information recommendation is propagated by making recommendations from person to person. However, there is a possibility that a closed group is formed by users with similar preferences, and no recommendation from other users is made to the group (see FIG. 3). To solve this, Forced Social Filtering [4] has been proposed. This is a method in which closed groups are not considered group preferences and are recommended for trials. As a result, the boundary of information propagation is eliminated, and efficient collaborative recommendation is performed as seen by the entire user. However, from the viewpoint of individual users, the recommendation made for the trial is not always successful, and there are still problems.
(References)
[4] Shunichi Tano, Tomoya Kine, Norihiko Ishitani: Activation method of digital information recommendation by global reinforcement learning; IEEJ Transactions C, Vol. 121-C, No.7, pp.1237-1245, 2001.

(Use of individual user information)
In recent years, with the spread of Internet technology, many users have frequently transmitted their information on the Internet. Accordingly, new attempts have been made to obtain useful results using information transmitted by these users. One of these is research that uses messages on Internet bulletin boards posted by various users and uses them to annotate program content [5].
(References)
[5] Naoyuki Okamoto, Takao Takenouchi, Takahiro Kawamura, Akihiko Osuga, Mamoru Maekawa: Development of a viewing support agent that generates public opinion metadata for broadcast programs-Extraction of atmosphere components from the online community and between users Refinement by distribution; IEICE, Vol. J88-DI, No.9, pp.1477-1486, 2005.

This Internet bulletin board is linked to a program being broadcast, and a user who is watching the program posts a message in real time. Therefore, in this research, by calculating the frequency of appearance of a specific posted message, the topic and excitement on the bulletin board at that time are detected in association with the program (see FIG. 4). The data thus obtained can be used as program metadata.
Information that is freely transmitted by individual users is mostly garbage information that is useless, but some information is very valuable. This research extracts valuable information from such information and uses it effectively.

(Problems of conventional information recommendation methods)
Conventional information recommendation systems often use both advantages by combining content-based recommendation and collaborative recommendation. However, the purpose of any recommendation method is to infer user preference information. For TV program recommendations, the goal was to guess as much as possible the user's preferences using EPG (electronic program guide) data of the program, the user's viewing history, and other users' viewing history. .
However, program recommendations using user preferences only look at the superficial overview of the program. Since the user does not actually determine what the user really wants to see by watching the program, the user is often discouraged by seeing the recommended program. Even if a program is watched every week with a favorite program, it is not always an interesting time, and sometimes it feels boring. In such a case, it is only possible to judge by actually watching the program.

  In order to guess the program that the user wants to watch, it is necessary to know the viewing behavior of the user. Users do not always watch programs that suit their preferences. Sometimes I watch a topical program in order to understand the trend of the world, and sometimes I see a program that I don't usually see in talking with friends. As a result, you may discover your new taste. For such users, there is no point in guessing a program that suits their preferences. In the first place, the program that the user wants is different from the program that the recommendation system tries to guess. This is the limit of the conventional recommendation method. The embodiment according to the present invention aims at program recommendation that can cope with such a situation. For this purpose, program recommendation is performed in consideration of each user's request.

(Approach of this embodiment)
In the present embodiment, a program that a user really wants to watch is estimated by simultaneously performing a conventional program recommendation and a new program recommendation considering the quality of the program.
Since the two types of recommendation methods are different, the information used is also different. Since information is handled differently depending on its nature, information is first classified using these recommendation methods. Based on that, we will examine how to use information suitable for each recommendation method.
In the present embodiment, content-based recommendation is used as a recommendation method. Other methods are also useful, but none of them are used in this embodiment because their essential purpose is the same as content-based recommendation.
As described above, there are two requirements for the program recommendation method considering the quality of the program. The first is that it is necessary to be able to cope with the detailed requirements of the user. This can be dealt with by determining the direction of the program that the user wants to recommend as necessary. Second, it is necessary to actually watch the program in order to know the value of the correct program. This can also be solved by the cooperation of many users. This is because if a large number of users other than yourself watch a program with various values and determine its value, it is very useful information for them. In particular, the evaluation of a user whose values are similar to his / her value is similar to his / her evaluation and can be trusted. Therefore, for the program recommendation considering the quality of the program, a total result obtained by totaling program evaluations by a large number of users is used.

The above is summarized as follows.
・ Clarify what information is handled by the conventional recommendation method and the recommendation method considering the quality of the program ・ Use content-based recommendation as the recommendation method ・ The user decides the direction of program recommendation as needed・ A large number of users evaluate programs and use the evaluation results for program recommendations

[Classification and use of information]
Next, analysis of information used in the conventional recommendation method and the new recommendation method considering the quality of the program will be described.
(Information used for program recommendation)
(Conventional program recommendation method)
The conventional program recommendation method recommends a program using the user's preference. By acquiring and accumulating summary data of programs that the user likes to watch from EPG (electronic program guide) data, the accumulated data becomes data representing user preferences. Therefore, it can be said that the conventional program recommendation method is a recommendation method that relies on EPG data.
This EPG data is provided by a specific organization and is accurate and reliable information. It is not a kind of information that can be freely transmitted by individual users. Such information is called formal information.

(New program recommendation method)
In the new program recommendation method, in order to evaluate the quality of a program, information such as evaluation of a large number of users for the program is used. Unlike EPG, this can be freely transmitted by the user. However, since the user can freely transmit, lie information may be included. Such information is generally called informal information.

With regard to information used for program recommendation, a clear difference was revealed that formal information was used in the conventional recommendation method and informal information was used in the new recommendation method. That is, the conventional recommendation method evaluates program content with formal information, and the new recommendation method evaluates program content with informal information. Therefore, in this embodiment, in order to quantitatively evaluate program content using these two methods, a “genre” that represents user preferences and an outline of the program and a “value” that represents the quality and value of the program are defined, These two are introduced as program content evaluation axes. However, there are a plurality of indicators for the quality and value of the program depending on the user's values. Therefore, there can be a plurality of values.
By the way, informal information is not dealt with by the conventional recommendation method, but is introduced for the first time by a new program recommendation method considering the quality of the program. In the next section, informal information is analyzed.

(Informal information)
Informal information is informal information conveyed through chats, rumors, and word of mouth. Its characteristics are "good luck", "real intention", and "lie".
On the other hand, formal information is contrasted with informal information. Formal information is official information conveyed through prestigious media such as TV, newspapers and magazines. Its characteristics are “accurate”, “building” and “credible”.
Conventionally, formal information has been used in various systems because of its accurate and routine nature. On the other hand, informal information has been shunned because it is indefinite and fine. However, the informal information includes information about the user's real intention, and its utility value is high. Informal information can bring out value if it is not mistakenly handled.

For example, the book “Wisdom of Crowds” introduces many examples that show that the aggregation of subjective opinions of many users under certain conditions is unexpectedly correct [6]. In addition, Wikipedia [7], an encyclopedia on the Internet, says that most of the information posted is correct even though users can freely add and modify it. In this way, it is possible to extract the value of the informal information of a large number of users by aggregating them.
(Reference and reference URL)
[6] James Slowicky, Odaka: "Everyone's opinion is unexpectedly correct", Kadokawa Shoten, 2006.
[7] “Wikipedia”, http://en.wikipedia.org/

  The reason why only formal information has been used in the past is that there was no place where many informal information could be easily shared. Next, we consider sharing informal information with the development of media. And it analyzes whether a lot of informal information can be shared about TV program.

The media has evolved from words, gestures, and letters to printing, radio, telephone, and TV, and further to current PCs, mobile phones, and the Internet. With the development of such media, sharing of formal information and informal information has greatly improved. However, for media such as printing to TV, especially radio and TV, information transmitted by an authorized person occupies most. In other words, the advent of radio and TV has dramatically increased only the sharing of formal information. At this point, the sharing of informal information has not improved significantly. Since then, the Internet has become widespread, and anyone can publish their Web page to an unspecified number of people, and the sharing of informal information has increased rapidly more than formal information. A network that uses P2P (Peer-to-Peer) technology has also appeared to protect the flow of free information so as to support this trend. In addition, the Internet includes functions such as letters, telephone, radio, and TV (such as mail and Internet TV), and even existing media that has been mainly sent by only some people until now is a means of sending personal information. It was changed to. The Internet involved the explosion of informal information, involving existing media [11] [12].
As a result, there are innumerable informal information currently on the Internet. Here, we will examine the sharing of informal information about TV programs. FIG. 5 summarizes the ranking of the number of message postings for a 2005 TV program in a group of anonymous bulletin boards on the Internet for each program genre. In the anonymous bulletin board group, a bulletin board is prepared for each major broadcast key station, and many messages are posted in real time for broadcast programs. FIG. 5 summarizes the number of messages on those bulletin boards.
(References)
[11] Shunichi Tano, Mitsuru Iwata, Tomonori Hashiyama: An example of informal information explosion and user interface utilizing it; Human Interface Symposium, pp.1175-1180, 2006.
[12] Yuzo Koido, Shunichi Tano, Mitsuru Iwata, Tomonori Hashiyama: Proposal of a new viewing support method and implementation of P2P aggregation algorithm of informal viewing information; Human Interface Symposium, 2006.

  From FIG. 5, it can be seen that hundreds of messages are posted every minute for TV programs. In addition, it can be confirmed that the number of posts is large for any program genre, and informal information about TV programs is overflowing. Therefore, it is considered that there is a sufficient possibility that informal information can be used in program recommendation. However, in this embodiment, the informal information on the Internet bulletin board is not used.

(Activation of transmission of informal information)
When using informal information transmitted by the user, how to activate the transmission of the user's informal information is an important issue. This problem is particularly troublesome when a conventional system that has already succeeded in activating information transmission is not used as an information source. In this embodiment, such an existing system is not used as an information source. This is because it cannot be said that such a system still exists. Therefore, in the present embodiment, it is assumed that informal information is acquired from general viewing behavior.

Research related to the activation of informal information generation includes support for informal communication. Informal information is particularly likely to be generated in large quantities from informal communication. This informal communication has long been regarded as important because it plays a role of acquiring and sharing information and promoting intellectual creation activities. For this reason, many studies have been conducted to promote informal communication.In particular, it is thought that familiarity with the other party and the chance of communication are important factors, so that a virtual face-to-face environment can be established and There are many studies that give the opportunity for formal communication. On the other hand, there is also the fact that conversations that do not support the face-to-face environment and that only interact with text data are frequent exchanges that seem to be more authentic than the face-to-face environment.
Originally, these should be analyzed in detail, and a method to activate the generation of informal information in the TV viewing environment should be considered, but this embodiment will not deal with it so far. Let it be an important issue.

(Value in program recommendation)
As described above, it has been found that evaluation of a user's program, that is, informal information can be used for a program recommendation method considering the quality of the program. However, it is not yet known what the user's informal information is. Then, next, specific informal information obtained from the user's viewing behavior will be shown, and values obtained from the information will be described.

The following can be considered as information that can be acquired from the program viewing behavior of the user.
・ Which program is being watched ・ Watching time of the program ・ Whether it was bookmarked ・ Whether it was recorded or reserved recording ・ Number of times the program was viewed ・ Impressions about the program ・ Emotions about the program ・ Annotations about the program

That is, the quality of the program is evaluated from these pieces of information. For example, if the number of times of viewing is large, it can be evaluated that the program is so interesting and high quality. Therefore, in order to realize program recommendation considering the quality of the program, such information is collected from a large number of users. By collecting the collected information, the user can obtain various values. Among them, the following values are easy to use for program recommendation.
・ Program popularity ・ Characteristics of programs ・ Expectations for programs ・ Satisfaction with programs ・ Emotions about programs

The popularity of a program represents how many users like it. This can be calculated from the audience rating of the program and the number of times the program has been viewed. The popularity of programs is effective when you want to watch popular programs in the world.
The nature of the next program represents what characteristics the program has from the user's perspective. For example, conservative programs, radical programs, and programs that are not suitable for children are the characteristics of the programs mentioned here. These require the user to directly input information to that effect.

  The degree of expectation for a program represents how much the user expects a program that has not been broadcast yet. For example, it can be seen from the degree of expectation of the program how much attention the new program that will start broadcasting tonight has attracted attention. In addition to the direct input to the user's program, this can be calculated from information such as whether or not the program has been reserved for recording.

  The degree of satisfaction with the program indicates whether the user is satisfied with the program or thought it was interesting. This can be calculated from "satisfied" or "dissatisfied" information entered directly by the user for the program.

Finally, the feelings about the program indicate what feelings the user had when viewing the program. For example, if you are watching a movie and you are in a moving scene, you will enter “I was moved”. If facial expression recognition technology is used, these can be automatically acquired by the system. The acquired information can be effectively used for a user who wants to watch a moving program or a user who does not want to watch a scary scene because he is not good at it.
Specific usage of these will be described later.

(Information aggregation method)
In the above, the specific informal information obtained from the viewing behavior of the user and the example of the value obtained by totaling the information are shown.
Then, next, the information summarization method which becomes a problem when summarizing informal information such as user's feelings will be described.
Conventionally, a client-server system has been frequently used for collecting information. This is because the information of all users can be totaled easily. However, the information collected this time is informal information. The feature that the informal level may be lowered if the central management mechanism intervenes in the informal information indicates that the compatibility between the client / server system and the informal information is poor. The client / server method is suitable for accurate information aggregation, but it is not necessary to accurately calculate “good or bad” informal information in the first place.
A method that does not require a central management mechanism is the P2P method. The P2P method is a brokerless model in which users connect directly. Since there is no central management mechanism in the middle, it is possible to prevent a decrease in the degree of informal information even when handling informal information. In addition, it has a distributed property, and there is an advantage that the load can be distributed to each node. Furthermore, in P2P, each node aggregates information while cooperating, so there are many cases where the information is aggregated to some extent. This also shows that P2P and informal information are compatible. However, there is a hybrid P2P model that partially incorporates a client / server system in P2P, and it is distinguished from a pure P2P model (pure P2P).

For this reason, in this embodiment, information is aggregated using a P2P method (pure P2P) that is compatible with informal information.
Here, Table 1 summarizes the characteristics of the conventional recommendation method and the new recommendation method aimed at in the present embodiment. The new recommendation method includes the conventional recommendation method.

(Privacy issue)
When dealing with user viewing information, the problem of privacy is a problem. In general, users tend to hate having their viewing history in someone's hand. In other words, some users are worried that their viewing history is automatically counted.
However, in the present embodiment, users cooperate with each other in P2P communication to count viewing information, so that a specific institution does not grasp their viewing history. Of course, information that can identify an individual is not sent in the first place. Also, if the viewing information is encrypted and aggregated, security can be improved.

(Information aggregation algorithm)
Next, we propose an algorithm suitable for collecting informal information and analyze its performance.

(Aggregable information)
To devise an informal information aggregation algorithm, we first clarify what kind of informal information is aggregated.
For example, information acquired from each user is as follows, for example.
・ Which program is being watched ・ Do not want to show to children ・ Is satisfied (Is it interesting)
・ What feelings you have? All of this information can be entered numerically. For example, if “which program you are viewing”, enter 1 for the program you are viewing and 0 for the program you are not viewing. Then, the degree is input numerically for the program. As for “emotion”, it is only necessary to input numerically what kind of feeling you have. Examples of emotions include “joy”, “anger”, “sadness”, and “fun”.

By representing the information numerically in this way, the information can be aggregated simply by adding these values. The information aggregation algorithm proposed here performs the aggregation of the informal information represented by such numerical values.
Note that the tabulation result is calculated as a ratio so that it can be easily used for program recommendation. In other words, obtain information such as what percentage of the program A is watched (viewing rate), how many users are satisfied with program A, and what percentage of users are impressed with the program. For the purpose. Therefore, the number of users (number of samples) can be counted together.

(Requirements)
Next, the requirements for the function and performance of the information aggregation algorithm required in the program recommendation system are summarized.
First, the minimum necessary function is that, as described above, the informal information represented by numerical values can be constantly counted for each program. However, the audience rating, which is the total result, is made to be known at intervals of several seconds. That is, not only the average audience rating of programs but also the transition of audience ratings are tabulated. Also, real-time counting is realized because the rating such as audience rating that changes in real time can be used for program recommendation.

  Next is the assumed network size, and the number of households that currently own TVs in Japan is tens of millions. Therefore, it is necessary to assume a large-scale network also in this embodiment. Therefore, a large-scale network having about hundreds of thousands to millions of nodes is assumed here. Of course, it must work even if the network is small. Also, in order to operate stably at all times, it must be resistant to network failures.

The accuracy such as audience rating must be secured as much as possible. In a current household audience rating survey in Japan, for example, the sample size is set to 200 or 600 depending on the survey area. Samples are extracted by random sampling based on data from the national census, but the legitimacy of the audience rating is also taken into account, excluding facilities such as households, hospitals, offices, etc. with TV station personnel. The error of the audience rating obtained by this is about ± 7% or ± 4% (reliability 95%) when the audience rating is 50%. Therefore, the information aggregation algorithm devised in the present embodiment is also a method that can suppress the error of the aggregation result to at least ± 4% or less.
Further, in the present embodiment, it is considered that each user's TV is connected through a P2P network. In this case, it is difficult to collect information by excluding persons related to the TV station from the user. Therefore, we give up giving accurate sampling, and instead compensate the decline in the validity of the tabulation results by increasing the number of samples over the current survey. Therefore, the information aggregation algorithm devised makes it possible to aggregate information of a large number of users to some extent.

In summary, the following functions and performance are required.
・ Percentage such as audience rating can be calculated for each program. ・ Real-time aggregation is possible. ・ It works even in large networks with hundreds of thousands to millions of users. Less than 4%

(P2P information aggregation algorithm)
Next, we propose an information aggregation algorithm that satisfies the requirements identified above as much as possible.
(Simple information aggregation method)
The simplest and basic method of totaling information on a large number of users by P2P is a method in which one node (user) takes over all the totals (see FIG. 6). Nodes are selected as many as the number of data to be aggregated, and then data is obtained from the selected nodes and aggregated. By broadcasting the aggregation results, all nodes can obtain the aggregation results.
However, such a method cannot actually be used. If the total node load is too large and the node fails, the aggregation will fail at that point.
In view of this, the following provides a method in which all nodes perform aggregation and cooperate with each other in consideration of stability and load distribution.

(Outline of P2P information aggregation algorithm)
The mechanism of the proposed information aggregation algorithm is simply shown in FIG. Arrows between nodes represent the flow of information.

  The basic idea is as follows. Each user regularly collects information from nodes around him and aggregates it. Then, the surrounding nodes further collect and aggregate information from the surrounding nodes. Therefore, the totaling results are also acquired from the nodes around you. For example, in FIG. 7, node 1 collects information from nodes 2-7. Similarly, the node 6 collects information on the nodes 11 to 12, and the node 7 collects information on the nodes 8 to 10. Therefore, in addition to the information of the nodes 2 to 7, the node 1 can simultaneously obtain the information of the nodes 8 to 12 collected by the node 6 and the node 7 (the total result). Furthermore, since the node 8 collects the information of the nodes 13 to 16, the information of the nodes 13 to 16 (the total result) can be collected at the same time. At this time, as viewed from node 1, the information of itself is 0th order information, the information of nodes 2-7 is primary information, the information of nodes 8-12 is secondary information, and the information of nodes 13-16 is tertiary information. I will call it. Of course, there can be more quaternary information and quintic information. Note that the node information referred to here is a numerical value representing the informal information of each user as described above.

  By this method, a huge amount of information on the network can be efficiently aggregated. However, since the aggregation results obtained from other nodes are data that has already been aggregated at that time, there is a time lag in the collected data. Therefore, the secondary information is older than the primary information, and the tertiary information is older than the secondary information. In other words, the information collected from a distant node is older. .

  In addition, duplication may occur in the collected data. For example, assume that node 6 has also collected information on node 8. Then, the node 1 receives the information on the node 8 collected by the node 6 and the information on the node 8 collected by the node 7 in duplicate. This problem will be described later.

(Sending link and receiving link)
Each node can establish a transmission link and a reception link with other nodes (the arrow between the nodes in FIG. 7 corresponds to it). A transmission link is a transmission path when a certain node transmits data to another node. On the contrary, the reception link is a transmission path used when receiving data from another node. Thus, for example, the reception link established between node A and node B is a transmission link to node A for node B.

Each node can establish a predetermined number of transmission links and reception links with other nodes. However, only one link can be extended from one node to another node, and only a receiving link can be extended from one node to another node. With respect to the transmission link, other nodes can have the reception link by setting up a reception link to themselves. Therefore, if another node does not attempt to establish a reception link for itself, it may not have a transmission link at all. It should be noted that it is completely random which node is connected to the receiving link. Further, it is assumed that a reception link cannot be established for a node having transmission links established for a predetermined number.
Each node periodically transmits its own data to all nodes of the transmission link destination. Conversely, data is not requested to the node at the receiving link destination, but only waiting for data to be sent from another node through the receiving link. However, the reception link where data is not sent even after waiting for a certain time is disconnected, and a new reception link is reestablished with another node.

(Data aggregation)
FIG. 8 shows how the data collected from the surrounding nodes is calculated.
In FIG. 8, rectangles divided into four areas represent data held by each node, and 0th order information to 3rd order information are stored in the respective areas. Aggregation calculation is performed by totaling the n-order information of the collected data and storing it in the area of the n + 1-order information of the aggregation result. The latest self information is stored in the 0th order information of the counting result. When data is transmitted to other nodes, the data of the total result is transmitted. Of course, when data is first joined to the network, data has not yet been collected from other nodes, so when data is transmitted, data in which only 0th order information (own information) is stored is transmitted.
In the present embodiment, data having 0th order information to nth order information is referred to as “order n + 1 data”.

(Data delay)
As is clear from the data aggregation method, there is a data delay in the aggregated data. Basically, the n + 1st order information is older than the nth order information. For example, if each node sends data every minute, the 0th order information of the aggregation result is the current information (your information), the primary information is the information one minute ago, and the secondary information is Is the information two minutes ago, and the tertiary information contains the information three minutes ago. That is, if the data transmission interval is t [sec], d-th order information is obtained after t × d [sec].
Since the n + 1 order information is larger than the n order information, older data is used to obtain a more accurate result using more data. Therefore, the proposed information aggregation algorithm cannot perform total real-time aggregation. However, if the data transmission interval of each node is shortened, it is possible to approximate real-time counting. However, it should be noted that as the transmission interval becomes shorter, the network traffic increases accordingly.

(accuracy)
Next, we consider the accuracy of the results obtained when the proposed information aggregation algorithm is used.
(Amount of collected data)
The accuracy of the audience rating, which is the result of the aggregation, is determined by the amount of data collected. If the calculation is performed using more data, the error is reduced accordingly. Therefore, we first estimate the amount of data that can be collected by the proposed algorithm.
The amount of data that can be collected is determined by the number of links and the order. If the number of transmission links and the number of reception links of each node are both L and the order of data is d, the data for L nodes is the primary information for the aggregation results, and the data for L2 nodes is the secondary information for the aggregation results. Therefore, data for Ld nodes can be collected as the d-order information of the aggregation result. In other words, the total amount of data that can be collected is as shown in Equation (5-1).

  For example, if L = 4 and d = 3, the total result of information for 4096 nodes can be obtained as the tertiary information of the total result. However, this is a result when all nodes have four receiving links. Actually, if the maximum value of the number of received links and the number of transmitted links of each node is set to 4, there may be a node that does not have all the received links. This is because, as each node establishes a reception link, the number of nodes having a margin in the transmission link decreases. Therefore, it is necessary to take measures such as setting the maximum value of the number of transmission links of each node to a value larger than the maximum value of the number of reception links.

  Next, it is investigated how much the accuracy of the audience rating changes depending on the amount of collected data. It is known that a sample ratio error (95% reliability) such as audience rating can be obtained from equation (5.2).

  Using equation (5 · 2), the error in sample ratio is as shown in Table 2.

In the current audience rating survey, the number of samples is about 600, but this table shows that the audience rating error is relatively large. On the other hand, if the number of samples is 4096, it can be seen that the audience rating error is less than half of the current error.
In the proposed information aggregation algorithm, if the parameter d is increased, the amount of data that can be collected increases abruptly, so that the audience rating error can be kept sufficiently small.

(Duplicate data)
As described above, there is a possibility that this information aggregation algorithm will aggregate data redundantly. In particular, data duplication is likely to occur when the network is small. However, just because the data is duplicated does not necessarily mean that the tabulation results are random. For example, when totaling the total number of nodes, the result is strange, and this information aggregation algorithm is not useful, but the data you want to find in this embodiment is "how many people are watching the program" or "how many people “Ratio” data such as “How many people were satisfied with the program”. In that case, it is considered that a correct result can be obtained even by the proposed information aggregation algorithm by constructing a random network. This is because even if the data of each node is aggregated to the same extent, the ratio such as the audience rating hardly changes. For example, if 10 out of 100 people are watching a program (viewing rate is 10%), even if data for 1000 people is duplicated, about 100 of them will end up watching it. The audience rating is almost the same 10%.

However, it is inevitable that the error becomes somewhat larger than the original audience rating. Therefore, the data duplication rate is introduced into Equation (5.2) to derive an error when there is duplication of data.
First, suppose that S data including duplication is collected, and that s of them are watching the program. Next, it is assumed that, among the data S, the data collected redundantly is M and the rest is N. Furthermore, assuming that n out of N people and m out of M people are watching a program, the audience rating can be expressed as shown in Equation (5.3).

  Here, R is the audience rating when there is no duplicate data, R ′ is the audience rating of duplicate data, and α is the duplicate rate of collected data. In this paper, the data duplication rate is described as α = M / N. Note that the overlap rate is not M / (N + M).

と導くことができる。この式(5・4)を用いて具体的に誤差を計算すると表３のようになる。なお、図９はN=4096において式(5・4)をグラフ化したものである。重複率が0.1前後のときに最も誤差が大きくなっていることが分かる。 Therefore, the audience rating error E is calculated from Equation (52) and Equation (53).

Can lead to. Table 3 shows a specific error calculation using this equation (5.4). FIG. 9 is a graph of the formula (5 · 4) when N = 4096. It can be seen that the error is the largest when the overlap rate is around 0.1.

  Looking at Table 3, it can be seen that the error hardly increases even if there is duplication in the data. Therefore, it can be seen that the data duplication rate is not a problem in the proposed information aggregation algorithm.

(traffic)
This system is premised on operating on a large-scale P2P network. Therefore, it is important to estimate in advance how many messages will flow through the network when collecting information.
In the proposed information aggregation algorithm, if the number of received links is L, the size of data to be exchanged is m [bit], and d-order information can be obtained after time T [sec], the network traffic of each node is (5 It becomes like 5) type.

Next, the values of L and d that can keep this traffic small are obtained.
First, it can be seen from equation (5.4) that the audience rating error is greatest when the audience rating is 50%. Here, it is considered to suppress the error at that time to about 1% or less. Since the error in the current audience rating survey is about 4%, if the error is 1% or less, it can be said that the accuracy is sufficient. Therefore, when the number of samples N at which the duplication rate is 0.1 and the audience rating (50%) error is 1% is calculated from Equation (5.4), it is about 14000. Therefore, if data for about 15,000 people is collected in consideration of duplicate data, the error can be suppressed to 1% or less. However, in an actual network, the amount of data that can be collected by leaving the node is smaller than the theoretical value, so it is sufficient to set parameters that can collect data of up to about 30,000 people in consideration of this. It is thought that.
Table 4 shows a set of the received link number L and the order d that can collect data of 30000 or more as d-order information.

  From Table 4, the value of the number L of received links with the smallest traffic is 2 or 3. However, if the number of received links L is too small, there is a possibility that the amount of collected data will decrease rapidly just because one receiving link cannot be installed. Therefore, this information aggregation algorithm adopts a set of the number of received links L = 4 and the order d = 8, in which such risks can be distributed and the traffic is small. In that case, if the time T = 60 [sec], the traffic is 0.53 m [bps].

(Fault tolerance)
The information aggregation algorithm according to the present embodiment basically assumes a system that is always connected to a network. Therefore, there are not so many nodes that frequently leave the network, but in order to always operate stably in a large-scale network, it is required to have excellent fault tolerance.
The algorithm according to the present embodiment operates on a pure P2P network. Therefore, it is relatively strong against network failures. For example, it is assumed that the node at the receiving link suddenly leaves the network. In that case, simply search for an alternative node and re-establish the receiving link. If the maximum value of the number of transmission links of each node is set to be large, the number of reception links that can be accepted can be afforded, and an alternative node should be easily found. Even if it takes time to find an alternative node, the aggregated data includes old data (how much old depends on the order of the data). It is also possible to interpolate some old data.

(P2P program recommendation system)
Next, a method for realizing a program recommendation system using genres and values will be described. The recommended programs are broadcast programs, past programs, and unbroadcast programs.

(Program recommendation by genre)
In the present embodiment, content-based recommendation which is a conventional recommendation method is used for program recommendation using a genre. Content-based recommendation is to extract a user's preference from the user's viewing history and recommend a program that meets that preference.

(User preference)
Each user has his / her preference for viewing the program. For example, some users prefer to watch sports programs, while other users prefer to watch documentary programs. In content-based recommendation, such user preferences are used to recommend programs.
In this system, user preferences are expressed in a word list (Table 5). The word list consists of pairs of words and their appearance frequencies, which are created using the EPG of the program. Specifically, a “Good button” and “No-Good button” are prepared on the remote control, and the word sequence of the EPG information of the program for which the Good / No-Good button is pressed is disassembled into words and extracted into the word list. sign up. The word list also has information on the appearance frequency of each word, and the word extracted from the EPG by the Good button increases the appearance frequency by 1, and the word extracted by the No-Good button decreases the appearance frequency by 1. The word list created in this way becomes the user's preference itself. The Good button is also used as an input for value satisfaction.

In order to guess a program of a user's favorite genre, a similarity between the user's word list and a word string of EPG information of the program (which is converted into a word list) is calculated. This similarity calculation is a two-step procedure in which the similarity is first calculated between the words, and then the similarity is calculated between the word lists.
For the calculation of the similarity between words, for example, a concept dictionary and an English word dictionary of an EDR electronic dictionary of the Japan Electronic Dictionary Research Institute are used. The concept dictionary describes the hierarchical relationship of concepts in a tree structure. The English word dictionary describes the correspondence between English words and concepts.
To calculate the similarity between words, first draw the concepts of two words in the English word dictionary, and then calculate the distance between them in the concept dictionary tree. It is known that when this distance is Np and the height of the tree is D, the similarity between words can be calculated by the following equation (6.1).

  Next, in order to calculate the similarity between the word lists, first, from the words in the longer word list, the ones with high similarity to each word in the shorter word list are extracted, and a new word list is created. create. This new word list is the same length as the shorter word list. Then, the inner product is calculated using the new word list (the value obtained by multiplying the value of the number of occurrences of each word by the similarity) and the shorter word list (value of the number of occurrences of each word) as a vector, and the angle formed is calculated . If this angle is small, the program matches the user's preference.

(Program recommendation by value)
Next, the value used for program recommendation will be described in detail.
(Value type)
The following values can be used for program recommendation.
・ Program popularity (viewing rate)
・ Characteristics of programs ・ Expectations for programs ・ Satisfaction with programs ・ Emotions for programs

Of these, popularity, expectation, and satisfaction can be calculated as one value, but nature and emotion can have multiple values. Therefore, in this paper, the types of nature and emotion are defined as follows.
"nature"
・ Emotions you don't want your children to see
・ We are happy, fun, sad, irritated, fear For nature and emotion, we calculate the degree of each one. The degree of popularity is calculated by dividing the number of viewers of the program by the number of samples, and the other expectation, satisfaction, nature, and emotion are calculated by dividing the total value by the number of viewers of the program. That is, a value such as what percentage of users who watched the program was satisfied is obtained.

(Enter information)
In order to obtain the value, in addition to the information obtained from the normal viewing history, it is necessary to aggregate information such as how satisfied each user is with the program and what feelings the program has.
As a method of acquiring such information, a method of acquiring text from voice input such as text input using a keyboard or voices such as laughter and facial expressions can be considered. However, facial expression recognition is technically difficult at this stage, and voice input is covered with TV voice. Also, not everyone can use the keyboard, and in the first place it is not suitable for a keyboard that requires aggressiveness for a passive media TV. The input means that are suitable for TV is something that anyone can easily input like a remote control. For example, if there is a “Good button”, the degree of satisfaction with the program can be easily input, and if a “happy button” or a “sad button” is prepared on the remote control, the degree of satisfaction or feelings about the program can be easily input. The user's informal information can be easily digitized from the information of which button has been pressed and how many times. Therefore, in the present embodiment, a remote controller is assumed as information input means. The target of information input may be the entire program or a part of the program (pinpoint), but in this embodiment, information is input to the pinpoint as “this moment of this program”.
By the way, in recent years, advanced functions of TVs have advanced, and buttons on the remote control have become very complex. In addition to the usual remote control, there are cases where a remote control with only the minimum necessary buttons is prepared. Therefore, it is not desirable to simply increase the buttons.

(User profile)
The value obtained by evaluating a program by a large number of users is a good index for knowing the consensus of all users, but more effective recommendation can be made by considering the user's profile at the same time. For example, for male users, it may be more effective to exclude female user data with different preferences. Alternatively, when a user wants to know a program popular with users of the same generation or wants to know a program popular with local users, the user's profile can be used.
The user profiles that can be effectively used for program recommendation are as follows.
・ Age, gender, and region Information can be summed up for each profile, such as popularity / satisfaction / emotion for teens, popularity / satisfaction / emotion for teens, and so on. However, the total data size is n times if the profile length (number of dimensions) is n.
Many profiles other than those listed here can be considered, but this is a problem deeply related to the privacy of the user, and information that can identify an individual must be avoided.

(Data level clustering)
In the above, it was stated that there is value in use by age, gender, and region. Here, in addition to these, a method for calculating the value according to the preference of the user will be described.
In program recommendation by value, it is very effective to use information of a user whose preference is similar to that of himself / herself. This is because program evaluation by users who have similar preferences to the user is more similar to the user and more reliable than program evaluations by users with various preferences. In the following, we will consider how to aggregate such information.

Some conventional P2P file sharing software performs preference clustering at the network level to increase file sharing efficiency. That is, persons with similar preferences are arranged at positions close to each other in terms of network. By performing such clustering, it is possible to collect information from users with similar preferences.
However, in the proposed information aggregation algorithm, it does not make sense to perform such clustering. This is because this information aggregation algorithm assumes a random network. Even if clustering is performed, if the network is biased, a correct tabulation result cannot be obtained. Therefore, in this embodiment, a “common genre list” is introduced. The common genre list is a column of preference words and values common to all users determined in advance (see Table 6). For example, words such as “SF” and “soccer” are set as a common genre.

Each user creates a common genre list as follows.
First, each user converts the degree of preference for the common genre from his / her preference word list to a value of 0 to 1 and sets it as the value of the common genre list. This is called a temporary common genre list. Next, a common genre list is created with 1 being the largest value and 0 being otherwise (see FIG. 10). That is, each user sets their favorite common genre value to 1. The reason for converting the value to 0 or 1 is to make the user's preference for the common genre more distinct.

  The information is totaled by adding the values of the common genre list (see FIG. 11). The counting method is the same as when using a user profile. The aggregation of each value is further divided into tabulations for each common genre. For example, tabulation is performed such as audience rating for common genre 1, audience rating for common genre 2, satisfaction for common genre 1, and so on.

From the results obtained in this way, it is possible to obtain information such as how many users who like the common genre 1 among users who watch the program A, how many users like the common genre 2, and so on. . This also means that professional value can be obtained. For example, it can be seen how a user who likes SF movies for a movie evaluates this movie.
Furthermore, it is also possible to obtain a value in accordance with each user's preference by using the total result. To do this, the user's temporary common genre list is used. Since the temporary common genre list represents the user's preference for the common genre, by taking the inner product of the total result of the common genre list and the temporary common genre list, a value considering the user's preference can be obtained. That is, the value calculated from the information of the user whose preference is similar to the user can be approximated by such a method.
Increasing the length (number of dimensions) of the common genre list will provide more accurate value, but it should be noted that the size of the aggregate data also increases in that case. If the number of dimensions of the common genre list is m, the data size is m times the original data (excluding the user profile), and if the number n of dimensions of the user profile is also considered, the data size is n + m times.

(Aggregated data)
The following data is a part of the total data related to the broadcast program exchanged at the time of information totalization. User profiles and common genres are not considered. Each user records his / her viewing history every 15 seconds. Therefore, 4 data are created per minute. Also, if the data transmission interval is 60 seconds and the order is 3, the data up to 3 minutes before is totaled, so there are 12 data in total.

<Program ID> ABC
<Number of samples> 1,2,3,4,5,6,7,8,9,10,11,12
<Number of viewers> 1,2,3,4,5,6,7,8,9,10,11,12
<Satisfaction> 1,2,3,4,5,6,7,8,9,10,11,12
<Expectation> 1,2,3,4,5,6,7,8,9,10,11,12
<Happy> 1,2,3,4,5,6,7,8,9,10,11,12
<Fun> 1,2,3,4,5,6,7,8,9,10,11,12
<Sad> 1,2,3,4,5,6,7,8,9,10,11,12
<Irritation> 1,2,3,4,5,6,7,8,9,10,11,12
<Fear> 1,2,3,4,5,6,7,8,9,10,11,12
<I don't want to show my kids> 1,2,3,4,5,6,7,8,9,10,11,12

Twelve numeric data are being compiled (the numbers shown here have no meaning). The leftmost data is 3 minutes ago, and the rightmost data is the current data. Since the order is 3, the data more than 3 minutes ago is truncated. The total data includes such data as many as the number of programs being broadcast.
If the number of received links is set to 4, you will receive 4 such aggregated data in 60 seconds. The received aggregated data is aggregated by adding data between the same programs. The popularity, satisfaction, emotion, and property values of the program are calculated from the totaled results.
In the case of data related to a recorded program rather than a program being broadcast, it is necessary to count data from the start to the end of the program, not data for three minutes. This is because the recorded program is viewed separately. Some users may have watched recorded program A 30 minutes ago, while others may have just started watching recorded program A. In that case, there is no point in counting only 3 minutes of data.
As for the recorded program, the viewing data should be already totaled when the program is being broadcast. Therefore, the data obtained as a total of recorded programs is added to the total results obtained during the broadcast already held. As a result, for example, the satisfaction level may be low during broadcasting, but the satisfaction level may be very high after several months.
On the other hand, since an unbroadcast program can only calculate the expected degree in the first place, only the data of <number of samples> and <expected degree> are collected.
If a user profile and a common genre are introduced into the data shown above, the aggregated data is as follows.

<Program ID> ABC
<Number of samples> 1,2,3,4,5,6,7,8,9,10,11,12
<Number of viewers> 1,2,3,4,5,6,7,8,9,10,11,12
<Male> 1,2,3,4,5,6,7,8,9,10,11,12
<Women> 1,2,3,4,5,6,7,8,9,10,11,12
<Teens> 1,2,3,4,5,6,7,8,9,10,11,12
<20's> 1,2,3,4,5,6,7,8,9,10,11,12
<Common Genre 1: SF> 1,2,3,4,5,6,7,8,9,10,11,12
<Common Genre 2: Soccer> 1,2,3,4,5,6,7,8,9,10,11,12
:
<Satisfaction> 1,2,3,4,5,6,7,8,9,10,11,12
<Male> 1,2,3,4,5,6,7,8,9,10,11,12
<Women> 1,2,3,4,5,6,7,8,9,10,11,12
:

The number of profiles and common genres to be introduced must be determined in consideration of how much the data size (traffic) will increase. When counting the data of recorded programs, the number of programs is innumerable unlike a broadcast program in which the number of programs is determined, so it is necessary to pay particular attention to the size of the data.
In the first place, it is impossible to count all the innumerable recorded programs. For example, if data for 10,000 people is collected, the worst 10,000 recorded program data will be aggregated. In such a case, a process of truncating the program data when the number of programs reaches a certain number during data aggregation is performed. The program data necessary for the user is kept, and the program data unnecessary is discarded. The program necessary for the user is a program for which sufficient viewing data cannot be collected when the program is being broadcast for some reason. Conversely, programs that are not necessary are programs for which sufficient viewing data has already been collected.
The total data of the recorded program is considerably larger than the total data of the program being broadcast even if the number of programs is reduced. However, unlike broadcast programs, recorded programs do not need real time. Therefore, it is not necessary to set the same data transmission interval between the program being broadcast and the recorded program, and the data transmission interval of the program being broadcast is set short and the data transmission interval of the recorded program is set long, thereby reducing traffic. Information can be aggregated.

(Screen interface)
Next, a program recommendation interface in an actual TV viewing scene will be described.
(browsing)
Browsing is to take out information that the user wants while viewing a list of information. Visualizing information is effective for efficient browsing.
In this system, for each program, the genre and value of the program are displayed in a graph, so that the user can find out which program suits his / her preference and which program has received high evaluation from which user. Try to understand at a glance. As a result, the user can easily find a program that the user wants to watch. Such browsing-based display is mainly divided into three scenes in this system. The first is the display while viewing the program, the second is the display in the recorded program list, and the third is the display when viewing the program guide.

  A display example during viewing of the first program is shown in FIG. FIG. 13 shows a screen of a prototype system currently being prototyped. On the program broadcast screen, as shown in FIG. 12, the genre and the instantaneous value (popularity here) of the program being viewed and the program of other channels are displayed in the right corner of the screen. While viewing the current program, the user can check how the value of the program of the other channel changes every moment. Furthermore, the transition of the popularity (viewing rate) of the program being viewed and other values (satisfaction and emotions) are displayed at the bottom of the screen. If another user presses the satisfaction button or the emotion button, the result is reflected in the graph at the bottom of the screen almost in real time, so that it can be confirmed how the other user thinks about the current program. In particular, sports programs and the like can feel a sense of unity with other users by inputting buttons in a lively scene. If you are not satisfied with the current program, you can check the genre and value of other channels and find a program that you like and that is highly appreciated by other users.

  The display in the second recorded program list and the display at the time of viewing the third program guide only display the information displayed during program viewing for each program. That is, a bar graph of the genre and value of the program is displayed next to each program in the list so that the genre and value of each program can be understood at a glance. Furthermore, when any program is selected, the transition of the value of the program is displayed at the bottom of the screen, so that the user can easily find out which scene of which program is interesting. This allows the user to “pinch” only really interesting scenes.

(Program recommendation interface)
Browsing was to search for a program to watch from a list of programs. However, if the user has an ambiguous request that “I want to watch such a program”, the program should be recommended accordingly. Therefore, in this embodiment, a program recommendation interface as shown in FIG. 14 is proposed.

  The user designates what program he / she wants to recommend by checking each check box arranged around. In the example of FIG. 14, an attempt is made to recommend a program having a high degree of popularity, satisfaction, preference (genre), and emotion (fun) among men in their twenties. The recommended programs are displayed in the center in descending order of evaluation. In this way, the user can find the program he / she wants to watch more efficiently than simple browsing. It is also effective to apply the conditions set here to the browsing screen display. For example, when viewing “total value of popularity, satisfaction, and emotion (fun) among men in their twenties” as a value during viewing a program (FIG. 12), a list of recorded programs, and viewing a program guide, browsing In particular, the user can efficiently find the program he / she wants to watch.

(Simulation experiment)
Next, we conduct a simulation experiment of the information aggregation algorithm proposed above and evaluate its performance.

(Purpose of the experiment)
It is impossible to experiment with hundreds of thousands of users to evaluate the proposed information aggregation algorithm. Therefore, in this experiment, we evaluate the information aggregation algorithm by examining the traffic, how the aggregation results are obtained at each node, and the accuracy of the aggregation results. An overview of the simulator created for this purpose is shown below.
・ Processes equivalent to the proposed information aggregation algorithm among a large number of virtual nodes ・ At that time, each node randomly selects the connection partner (receiving link destination node) from all nodes ・ Maximum number of links, degree, Data transmission interval is common to all nodes ・ Data transmission is performed simultaneously by all nodes ・ Data transmission interval is one step of simulation ・ Other nodes as many as the number of connection attempts determined for each step・ You can set the node withdrawal rate and participation rate for each step. ・ You can control the user's viewing behavior for each step.
(For example, 100 people watch program A and 200 people watch program B)
・ Each node can calculate various values of the program from the data aggregated for each step. This simulator can analyze the network at the time of information aggregation and check the possibility of calculating the value at each node. Designed.

(Number of connected links)
As described above, it was found that the maximum number of received links should be set to 4, but the maximum number of transmitted links has not been determined. If both the maximum number of received links and the maximum number of transmitted links are set to 4, theoretically, all nodes can establish four receive links and four transmit links. However, in reality, there are some nodes that cannot easily find a node with a sufficient transmission link, and it is difficult for such a node to establish all four reception links. If so, the amount of collected data is reduced at those nodes, and the accuracy of the audience rating is reduced.
This problem can be solved by setting the maximum transmission link number to a value larger than the maximum reception link number. On the other hand, if the maximum number of transmission links is too large, a node that holds a large number of transmission links and a node that hardly holds a transmission link appear, and the network is biased. If there is such a bias, there will be a situation where only a large amount of data of a specific node is duplicated. This is also a factor that reduces the accuracy of the audience rating. Therefore, it is considered that the maximum number of transmission links is larger than the maximum number of reception links and a value as small as possible is good.
Based on the above, we confirmed through experiments how the number of received links at each node changes depending on the maximum number of transmitted links. 15 to 19, the number of nodes is 100000, the maximum number of received links of each node is 4, and the number of connection attempts to other nodes is 12 (that is, attempts to establish received links to 12 nodes every step) ), And the number of received links at each node. The value is calculated as an average value when 10 steps of simulation are performed. In addition, about 1% of nodes left the network and joined the network every step.

  15 to 19, when the maximum number of transmission links is increased by 1 from the maximum number of received links, the number of nodes holding four reception links increases rapidly, and conversely three reception links. It can be seen that the following nodes are decreasing rapidly. In addition, even if the maximum number of transmission links was set to 6 or more, there was no significant change. For this reason, it is considered that the maximum number of transmission links should be set to 5, which is one more than the maximum number of reception links. Or, normally, the maximum number of transmission links may be set to the maximum number of reception links, and the maximum number of transmission links may be dynamically changed according to the number of connection failures of each node. For example, for a node in which only one incoming link was established even when trying to connect to another node 10 times out of 12 connection attempts, the maximum number of outgoing links 4 is usually increased to 5. A special connection is allowed.

(Data collection amount)
Next, the amount of data collected at each node is examined when considering the departure of the node from the network and the participation in the network.
FIG. 20 shows simulation results when the number of nodes is 100,000, the maximum number of received links is 4, the maximum number of transmission links is 5, the degree is 8, and the number of connection attempts to other nodes is 12. The vertical axis represents the collected data amount, and the horizontal axis represents the node withdrawal rate and node participation rate. The node leave rate and participation rate represent what percentage of all nodes leave or join the network. Here, it is assumed that the node leaves and joins each time the simulation is advanced by one step. The node leaving rate and the node participation rate were set to the same value.
FIG. 20 shows an average of the eighth data amount of an arbitrary node when the simulation is executed for 10 steps. The ideal amount of collected data at this time is 48 = 65536, but the collected data amount is 40,000 or more even when the node leaving rate is 1%, and the collected data amount is 40000 or more. I understand.

(Network traffic)
The graph shown in FIG. 21 shows the average data transmission count of each node per step when the simulation is performed under the same conditions as described above.
Since the maximum number of received links is 4, each node has four transmission links on average, and theoretically, each node transmits data four times per step. FIG. 21 shows the result when the withdrawal rate is changed. However, even if the withdrawal rate increases, the number of data transmissions does not decrease so much, so that data aggregation is performed stably.
If one step of the simulation is s [sec] and the data size is m [bit], the communication bandwidth required for each node can be estimated to be about 4 m / s [bps]. For example, if s = 10 [sec] and m = 100K [byte], the communication amount is 320K [bps]. Therefore, if the data size is large, the network line may be compressed. Therefore, it is necessary to devise a method for compressing data or collecting only necessary information.

(Data duplication rate)
Subsequently, it is examined how the data duplication rate changes depending on the number of nodes.
The simulation is performed under the condition that the number of received links is 4, the number of transmitted links is 5, the degree is 8, the number of connection attempts is 12, and the node leaving rate (= node participation rate) is 1%. As for the duplication rate, the duplication rate concerning the 8th data shall be examined. The eighth data of the tabulation results is a collection of information on nodes that are 8 hops away. Therefore, by tracing the reception link from a specific node and examining how much the 8 hop ahead node overlaps, it is possible to estimate the approximate overlap rate of the 8th data of that specific node. If the network link structure does not change with no node leaving or joining, the obtained duplication rate is an accurate data duplication rate. FIG. 22 shows the overlap rate when simulation is performed with the number of nodes changed.

  From this result, it can be seen that the data duplication rate is about 1 / n times when the number of nodes is n times. Therefore, we found that the overlap rate is almost negligible in large networks. In this embodiment, the ratio of data that is almost unaffected by data duplication has been obtained as a result of aggregation, but it has been confirmed that there is almost no data duplication in a large-scale network. There is a possibility that data can be obtained.

(accuracy)
Here, the accuracy of the audience rating obtained by the aggregation is examined.
First, it is confirmed how the accuracy of the audience rating obtained at each node changes with the passage of time. The simulation conditions are the same as in Section 7.5. However, the number of nodes was 100,000, and the user's viewing situation was set as follows. The number in parentheses is the true audience rating of the program.
・ Program A: 5000 people (viewing rate 5%)
・ Program B: 10000 people (viewing rate 10%)
・ Program C: 20000 people (20% audience rating)
・ Program D: 50000 people (viewing rate 50%)
・ Unviewed: 15,000 people Note that the user does not change the viewing program during the simulation.
The simulation results are shown in Tables 7 and 8 and FIG. Table 7 summarizes the average audience rating based on each degree data of 100 nodes extracted at random in any step of the simulation. Table 8 shows the basic statistics for the 8th data. FIG. 23 is a graph showing the standard deviation of the audience rating based on the data of each order of 100 nodes.

  From Tables 7 and 8, it can be confirmed that a sufficient audience rating can be obtained by calculating using the eighth data even if the node is detached. In FIG. 23, it can be seen that the accuracy of the audience rating is improved by increasing the order. Also, since the audience rating with good accuracy is obtained for the 6th data and 7th data, the audience rating can be obtained in 6 steps or 7 steps without obtaining the 8th data over 8 steps. it can.

(State of collecting results)
The information aggregation algorithm proposed in the present embodiment has a feature that the accuracy of the aggregation result increases with the passage of time. Here, the viewing behavior of the user is controlled along the time axis, and such a change in the counting result is confirmed.
The simulation conditions are as follows: the number of nodes is 10,000, the maximum number of received links is 4, the maximum number of transmitted links is 5, the degree is 8, the number of connection attempts is 12, the node leaving rate and the node participation rate are 1%, and the data transmission interval is 10 [sec]. The viewing behavior of the user is set as shown below.

<1 to 3 steps program viewing status>
・ Program A: 500 (viewing rate 5%)
・ Program B: 1000 people (viewing rate 10%)
・ Program C: 2000 people (20% audience rating)
・ Program D: 5,000 people (viewing rate 50%)
・ Not watched: 1500 people

<4-5 steps program viewing status>
・ Program A: 1500 people (15% audience rating)
・ Program B: 800 people (viewing rate 8%)
・ Program C: 2500 people (viewing rate 25%)
・ Program D: 4000 people (40% audience rating)
・ Not watched: 1200 people

<6-step program viewing status>
・ Program A: 2000 people (20% audience rating)
・ Program B: 900 people (9% audience rating)
・ Program C: 2200 people (viewing rate 22%)
・ Program D: 3800 people (viewing rate 38%)
・ Not watched: 1100 people

  In the above, what is shown in parentheses is the true audience rating at that time. The results of this simulation are shown in FIGS. The figure is a graphical representation of the total results of arbitrary nodes obtained at the 8th, 10th, 12th and 14th steps of the simulation. The horizontal axis represents time, and the right end is the current time. The black vertical line in the graph indicates the time six steps before the current time, and the red vertical line indicates the time eight steps before the current time. The “E” graph displayed in purple represents the percentage of users who are not watching TV.

Pay attention to the graph of program D. The true audience rating of program D is 50% when 0 ≦ t <30, 40% when 30 ≦ t <50, and 38% when 50 ≦ t, where t is time.
24 to 27, the data on the right side of the black vertical line 6 steps before the present is not reliable data with a small number of samples.
In FIG. 24, the audience rating at t = 0, which is 8 steps old, is obtained with high accuracy, but the error rate of t = 20, which is 6 steps older, has a large error of 1.5% or more. However, in FIG. 25 after the two steps, it can be confirmed that the error is as small as about 0.2%. Similarly, the audience rating at t = 40 in FIG. 25 has a large error of about 2.7%, but in FIG. 26 after two steps, the error is almost zero. In FIG. 26, the audience rating of 30 ≦ t <70 is random in FIG. 24, but the error is small and the graph is clear. From these results, it is possible to confirm the characteristic of the present information aggregation algorithm that the accuracy of the audience rating increases as time passes.

Furthermore, in this simulation, we set some users to press the Good button at the 4th step. The number of users who pressed the Good button is as follows.
<User who pressed the Good button at the 4th step>
・ Program A: 200 people (13.3%)
・ Program B: 300 people (37.5%)
・ Program C: 400 people (16%)
・ Program D: 500 people (12.5%)

In the above, the numerical value in parentheses is the ratio of users who pressed the Good button to the number of viewers of each program. These numbers are the true satisfaction for each program.
The tabulation results obtained by the simulation are shown in FIGS. The figure is a graph of the total results of arbitrary nodes obtained at the 8th, 10th, 12th and 14th steps. The horizontal axis and the vertical lines of red and black are the same as the above audience rating graph.

  Looking at program B, it can be seen that as the steps are repeated, the true satisfaction level approaches 37.5%. However, the final result was 36.430% with an error of about 1%. During the simulation, 1% of all nodes are repeatedly leaving the network and joining the network at each step, so it is probable that a nearby node has accidentally left. However, an accuracy of 1% is still sufficient.

(Summary of experiment)
In the simulation experiment, first, the value of the maximum number of transmission links that can hold as many reception links as possible was examined. As a result, it was confirmed that a value that is 1 larger than the maximum number of received links is suitable for the maximum number of transmitted links (see FIGS. 15 to 19).
Next, the amount of data collected when the node departure rate was changed was examined. As a result, it was confirmed that a sufficient amount of data could be aggregated even if the node leaving rate was set to a high value. This means that data can always be counted stably.
Next, the number of data transmissions per step when the node leaving rate was changed was examined. As a result, there was no significant change in the number of data transmissions even when the node leaving rate was set to a high value (see FIG. 21). Therefore, like the above experimental result (see FIG. 20), it can be said that the present information aggregation algorithm is excellent in fault tolerance and can be stably aggregated.
Next, the data duplication rate when the number of nodes was changed was estimated. As a result, it was confirmed that there was almost no duplication of aggregated data in a large-scale network (see FIG. 22).
Next, we confirmed how accurately the audience rating can be obtained when incorporating node detachment. As a result, it was confirmed that the audience rating can be obtained with considerably high accuracy (see Tables 7 to 8 and FIG. 23).
Next, the progress of counting when the simulation was executed was confirmed, and it was confirmed that the accuracy of the counting results changed with the passage of time. Although it is a simulation, it has succeeded in calculating an audience rating with sufficient accuracy with a short number of steps, and has shown the possibility of calculating an audience rating with high accuracy in a short time even in a large-scale distributed environment (FIG. 24). ~ 31).

Next, a summary of the present embodiment and future problems will be described.
(Summary)
In the present embodiment, a “program quality” that has not been handled by the conventional program recommendation method is taken in, and an object of the present invention is to realize a new program recommendation method that can present a useful program to the user.
First, the information used for program recommendation was analyzed. As a result, it was found that the EPG that has been used for program recommendation in the past is formal information, whereas the quality of the program is informal information that gathers honest information from the user. Therefore, “genre” corresponding to formal information and “value” corresponding to informal information are defined as content evaluation axes. The genre represents the superficial outline of the content, and the value represents the quality of the inner surface.
Informal information from a large number of users must be aggregated in order to calculate value. Therefore, in the present embodiment, an information aggregation algorithm that can efficiently aggregate informal information of a large number of users using P2P suitable for the aggregation of informal information has been proposed. This algorithm can aggregate information efficiently in a large-scale distributed environment. In small and medium-sized network environments, duplication occurred in the aggregated data, but it was found that there was almost no error when the aggregated results were obtained as a ratio.
Next, the realization method of the P2P program recommendation system using value was shown, and the screen interface that could become a new viewing style of TV was proposed.
In the final simulation experiment, the performance of the information aggregation algorithm was evaluated, and the calculation process of the aggregation results such as audience rating was observed to show the feasibility of an actual program recommendation system.

(Future tasks)
The proposed P2P program recommendation system still has some issues regarding information input. The first is that all input of information relies on the buttons on the remote control. As a means for inputting information instantly, easily and easily, button input by a remote controller is very excellent. However, if buttons are prepared for only the type of information to be input, the number of buttons will increase and the user will be confused. One way to solve this problem is for the system to observe the user and automatically enter information. Many studies have been conducted to obtain user emotions from user voices and facial expressions. If such technology develops, the system will automatically input the user's feelings.
The second problem is whether the user frequently inputs the Good button or the emotion button when watching TV. In order for a user to input information to a passive TV, a device that increases the motivation for the user to input information is required. For example, if the user can interact with other users by inputting information, the user's motivation may increase. Further analysis is needed on this point.
The third problem is when one TV is viewed by multiple people. The information recommendation system as proposed in this embodiment is basically intended for individual users. This is because if a plurality of users are used at the same time, the user's exact preference cannot be acquired. In addition, there is a privacy problem that other users are aware of their preferences. A reasonable way to solve this would be to provide a user function on the information input device (remote control). This allows the system to know who entered the information and to switch processing for individual users. Currently, it is becoming possible to operate home appliances with a mobile terminal, but since the mobile terminal is mainly carried by individuals, it may be used as an input device.

Next, another example of the information aggregation algorithm according to the present invention will be described.
Here, the specific algorithm of information aggregation by P2P is explained. The purpose is to gather information from thousands of users in a casual and efficient manner. Here, the information to be aggregated is “how many channels each user was watching for 1 minute”. In other words, the information obtained as a result of aggregation is the number of viewers per minute of each channel.

First, information held by each node is as shown in FIG. (The numbers with circles in FIG. 32 are represented as (1), (2)...)
In (1), (2),... Of the data to be exchanged, the number of viewers is counted for each channel.
Here, data with d-order information is referred to as “data of order d + 1”.
I will call it.
The data aggregation algorithm is as follows.
“On startup”
・ Connect to the server and register your notebook address with the server.
・ Ask the server to randomly tell X already registered node addresses.
-Register K node addresses of them in the receiving link. On the other hand, your node address is registered in the transmission link of the partner node. (In other words, the transmission link is not established until there is a request from another node.)
“Send”
-Every minute, the data created by “Aggregation” is sent to the destination node.
• At this time, set the TTL of the data to 1.
"transfer"
About data received from the node at the receiving link destination ・ Data coming from 1 hop next (TTL = 1)
Increase the TTL by 1 and forward to the destination node.
・ Data coming from 2 hops next (TTL = 2 data) go to “Total”.
"Aggregate"
・ Set your information to (1) of the data to be sent.
-Add all the received data (1) and set it to (2) of the data to be transmitted.
・ Add all the received data (2) and set it to (3) of the data to be transmitted.
:
-The sum of the d-order information of the received data is the total result obtained (see FIG. 33).
However, since the d-order information is calculated using the d-1 order information one minute before, the aggregated result includes information up to d minutes ago.

This algorithm can efficiently collect information for K ^{(d + 1) · TTL} people. The size of data to be exchanged basically depends only on the order of the data.
The reason for passing the data that comes from the node next to one hop through “forwarding” is to prevent duplicating the data. However, it is easy to imagine that it still cannot completely prevent duplication of data. This overlap rate will be described later.

と見積もることが出来る。と見積もることが出来る。上のアルゴリズムはp を小さく取っても多くの情報を集められるので、パラメータの適切な設定によりトラフィックはある程度小さく抑えられる。 (traffic)
This system is premised on operating on a large-scale P2P network. Therefore, it is important to estimate in advance how many messages will flow through the network when collecting information.
In the case of the above algorithm, if the total number of nodes is N, the number of received links is K, and data is transmitted up to p hops away (TTL = p), the number of messages that flow in one time step (1 minute) is

Can be estimated. Can be estimated. Since the above algorithm can collect a lot of information even if p is small, traffic can be kept small to some extent by setting parameters appropriately.

(simulation)
Next, in the above algorithm, the number of duplicated data is examined by simulation.
FIG. 34 shows simulation results when the number of received links and the number of transmitted links are 4, the TTL is 2, and the data order is 3. The data duplication rate when the number of nodes is 5000 to 12000 is shown. The number of data collected at this time was about 4000 for any number of nodes.
This system assumes a large-scale network with hundreds of thousands or millions of nodes. From the simulation results, it is expected that there will be little overlap in such a large network. In the first place, it is aimed at summarizing informal informal information, so it does not make much sense to eliminate duplication completely.
Of course, in a network with a small initial number of nodes, the duplication rate cannot be ignored.

(Efficiency)
In the above algorithm, all nodes work equally. However, it is clearly more efficient to divide roles according to node performance. Therefore, consider a node that reduces the calculation of n-th order information. In this case, since the n + 1 order information should exist in the received data, it can be copied as it is and used as the total result of the n order information. However, since old information is diverted, naturally, the total information that is finally obtained includes old information accordingly. As the number of nodes that reduce computation increases, the information that flows through the network becomes only old information. Therefore, it is necessary to take measures such as controlling the number of nodes to be stochastically controlled, or reducing the number of transmission links of nodes to be reduced to suppress the distribution of old information. This will be analyzed in the future.

(P2P information recommendation system)
In this embodiment, the system is being mounted on a set-top box (STB) that supports digital broadcasting.
Below, the concrete realization method of this system is demonstrated.

(Program recommendation by genre)
In order to guess a user's favorite program, it is necessary to know the user's preference for the program. Therefore, in this system, the “Good button” and “No-Good button” are assigned to the STB remote control, and the user who is watching the program inputs whether they like the program or not.
When these buttons are pressed, the system breaks up the word string of EPG information of the program into words and registers them in the word list. For words that are already registered, the number of appearances is updated. (In the case of the Good button, the number of appearances is increased by 1, and in the case of the No-Good button, it is decreased by 1.)
The word list is a list of words and their number of appearances, and the maximum size is 6000 words. By this method, the word list becomes the user's preference itself. That is, words with a high frequency of appearance represent the features of the user's favorite program.

Matching user preferences and programs In order to infer programs of the user's favorite genre, the similarity between the user's word list and the word string of the EPG information of the program (converted into a word list) is calculated. This similarity calculation is a two-step procedure in which the similarity is first calculated between the words, and then the similarity is calculated between the word lists.
To calculate the similarity between words, the EDR electronic dictionary concept dictionary and English word dictionary are used. The concept dictionary describes the hierarchical relationship of concepts in a tree structure. The English word dictionary describes the correspondence between English words and concepts.
However, the English words actually used are limited to the high frequency of about 6000 words as shown on the STB.
To calculate the similarity between words, first draw the concepts of two words in the English word dictionary, and then calculate the distance between them in the concept dictionary tree. When this distance is Np, and the maximum height of the tree is D, the similarity between words is calculated by the above-described equation (6.1).
Next, in order to calculate the similarity between the word lists, first, from the words in the longer word list, those having high similarity to each word in the shorter word list are extracted, and a new word list is extracted. Create This new word list is the same length as the shorter word list. Then, the inner product is calculated using the new word list (value obtained by multiplying the number of appearances of each word and the similarity) and the shorter word list (value of the number of appearances of each word) as vectors, and the angle formed by the inner product is obtained. If this angle is small, the program matches the user's preference.

(Program recommendation by value)
Various values can be considered depending on the information collected from the user. For example, in this system, the following factors can be considered: • audience rating, viewing time, viewing count, number of times that the Good button or No-Good button was pressed, whether recording was performed, whether bookmarked, etc. These can be easily obtained using the information aggregation algorithm in Chapter 4.
By presenting these values and their transitions for each program, the user can know the program from various angles, and can select the program from a different viewpoint.

(Clustering)
Clustering is a technique for bringing together people with similar preferences together. By introducing this, the user can obtain more useful information. In other words, in addition to the value of the general public, it is also possible to obtain the value of people who have similar tastes. This is the same as a friend who has a hobby like you provides useful information.
It is also very useful to obtain information from a cluster other than the cluster to which it belongs. This is because a cluster is a group of people who are familiar with the field. It is wise to refer to the opinions of experts in a field that you are not familiar with.
It is not so difficult to introduce clustering in this system. This is because each node holds its own preference as a word list, so that it is only necessary to calculate the similarity between each node and the preference and selectively connect to a partner with a high similarity.
But there are problems. When clustering is done, there are only people who are similar to you, so you can only collect information from people who are similar to you. In other words, the value of the general public will not be obtained. Therefore, for example, the following two methods can be considered.
Method 1: The connection destination node is managed by two node lists, a preference node list and a random node list.
Method 2: Extract information on nodes that are similar to you from the surrounding nodes.

(Method 1)
Method 1 is a method in which a clustered network and a non-clustered network exist at the same time. A node address having a high preference similarity with itself is registered in the preference node list, and a random node address is registered in the random node list. Then, the information aggregation algorithm is executed independently and in parallel with these two node lists.
This method enables flexible and accurate clustering. And you can collect information on the general public as well as information on people with similar preferences. However, two information aggregation algorithms are executed in parallel, and traffic and calculation costs are doubled. Also, information about other clusters cannot be obtained.

(Method 2)
In Method 2, it is necessary to introduce a common genre for all users. Each user calculates the degree of preference for a common genre from his / her preference word list and converts it to a value between 0 and 1. (= Temporary common genre list) Then, a common genre list is created in which 1 is high and 0 is otherwise (see FIG. 35). That is, the favorite common genre is 1. Note that there are about several tens of common genres.

The common genre list is totaled while adding together users who were watching the same program. Then, it is possible to know the number of users who like the common genre1, the number of users who like the common genre2, among the users who have watched the program. In other words, it is possible to obtain information on common genreX experts. This is the same as clustering users by common genre.
In addition, information about a person similar to you can be obtained in a pseudo manner by adding together the values of the tabulated common genre list and your temporary common genre list.
In Method 2, since a common genre is determined in advance, flexible and highly accurate clustering becomes impossible, for example, when a new genre comes out, it is impossible to cope.

It is a schematic diagram which shows the concept of content base recommendation. It is a schematic diagram which shows the concept of collaborative recommendation. It is a schematic diagram which shows the concept of collaborative horizon. It is a graph which shows the relationship between time progress and excitement about the topic of the bulletin board synchronized with the broadcast program. It is the graph which totaled the order of the number of message postings to the bulletin board for TV programs for each program genre. It is a schematic diagram which shows an example of the information totalization method. It is a schematic diagram which shows the structure of an information totalization algorithm. It is a schematic diagram which shows an example of the totaling method of data. It is the graph which visualized the error of the audience rating in consideration of duplication data. It is a schematic diagram which shows the example of creation of a common genre list. It is a schematic diagram which shows the concept of totalization of a common genre list. It is a figure which shows an example of the screen during program display typically. It is a figure which shows an example of the screen during an actual program display. It is a figure which shows an example of a program recommendation interface. 6 is a graph showing the relationship between the maximum number of transmission links and the number of nodes when the number of reception links is zero. 6 is a graph showing the relationship between the maximum number of transmission links and the number of nodes when the number of reception links is 1. 6 is a graph showing the relationship between the maximum number of transmission links and the number of nodes when the number of reception links is two. 6 is a graph showing the relationship between the maximum number of transmission links and the number of nodes when the number of reception links is 3. It is a graph which shows the relationship between the maximum transmission link number and the number of nodes when the number of reception links is four. It is a graph which shows an example of the relationship between a node leaving rate and the amount of collected data. It is a graph which shows an example of the relationship between a node leaving rate and the number of times of data transmission. It is a graph which shows an example of the relationship between the number of nodes and a data duplication rate. It is a graph which shows the standard deviation of the audience rating by the data of each order of 100 nodes. It is a graph which shows an example of the total result after 8 steps about audience rating. It is a graph which shows an example of the total result after 10 steps about audience rating. It is a graph which shows an example of the total result after 12 steps about audience rating. It is a graph which shows an example of the total result after 14 steps about audience rating. It is a graph which shows an example of the total result after 8 steps about satisfaction. It is a graph which shows an example of the total result after 10 steps about satisfaction. It is a graph which shows an example of the total result after 12 steps about satisfaction. It is a graph which shows an example of the total result after 14 steps about satisfaction. It is a schematic diagram which shows an example of the data exchanged. It is a schematic diagram which shows an example of the totaling method of data. It is a graph which shows an example of the relationship between the number of nodes and a data duplication rate. It is a schematic diagram which shows the example of creation of a common genre list.

Claims (2)

A means to input the evaluation (Good / No-Good) according to the viewer's preference in the broadcast program information,
A means for entering ratings against a list of genres common to viewers;
For each broadcast program, a first tabulation is performed to obtain and tabulate evaluations made on the common genre list input by the viewers of the viewers who are viewing the same program via the network. Means,
For each broadcast program, while obtaining an evaluation according to the viewer's preference through the work, a second counting means for counting,
A broadcast program recommendation information display device comprising: display means for displaying an evaluation of each broadcast program aggregated by the first and second aggregation means.   2. The display unit displays the evaluation of the broadcast program aggregated by the second aggregation unit for the broadcast program being viewed together with the audience rating of the program being viewed as time-series information. Of recommended information for broadcast programs in Japan.

Images