JP6549643B2 - Audience rating system, method and computer program - Google Patents

Description

  The present invention relates to a method of estimating the rating of a broadcast program.

  At present, the ratings in television broadcasting are surveyed for each television program of each television station, and are used, for example, to determine the airing time and airing fee of television advertisements. In addition, audience rating data is also used as a measure of television media power and advertising effectiveness when an advertising agency (person), a television station or an advertising company trades in advertising. Particularly in the United States, audience data is used primarily as a currency for trading television advertisements. Currently, ratings surveys are conducted by major companies such as, for example, Nielsen in the United States and Video Research in Japan.

  At present, there are questionnaire methods and mechanical methods as the methods of surveying audience rating. In the questionnaire type audience rating survey, viewing of programs in each time zone is checked based on the filling in of questionnaires and telephone calls. In mechanical ratings surveys, a device connected to a television receiver automatically records the channels viewed in each time zone on the television receiver.

  In these conventional methods, there is a problem that it takes a certain amount of time until all the data are aggregated and the results are announced from the research institute. For example, the data of the current audience rating survey is basically provided on the next day, not on the day of the broadcast, and in particular, on Friday, Saturday, and Sunday, the data is provided on the following Monday. On the other hand, recently, survey and measurement machines have been developed that can obtain preliminary values in "almost real time" in which audience rating data by time zone is obtained 1 to 2 hours after broadcast.

Further, in the conventional method, since it is necessary to count data from each viewer or viewing device, there is a problem that it takes time and cost as a matter of course.
In the following, we would like to give an overview of the ratings estimation method proposed in the prior art document, which is not exhaustive but is listed later. Non-Patent Document 3 describes, as an existing audience rating survey method, a survey based on a people meter system, an online meter system, and a diary questionnaire. Moreover, in the Unexamined-Japanese-Patent No. 2010-130585 (patent document 4), the audience rating prediction apparatus which improved the prediction accuracy of the program audience rating of a non-broadcast television program is proposed. In the audience rating prediction apparatus of Patent Document 4, broadcasted program characteristic information indicating the characteristics of each of a plurality of broadcasted television programs broadcasted in the same time zone in the past, and each of a plurality of viewers have a plurality of broadcasted television programs. A program selection model construction unit for constructing, for each viewer, a program selection model modeling a television program selection behavior based on selected program information indicating a television program selected and viewed from among the programs; Is selected from among the plurality of unbroadcast television programs based on program scheduling information indicating the organization of the plurality of unbroadcast television programs to be selected and unbroadcast program characteristic information indicating the characteristics of each of the plurality of unbroadcast television programs And a selected program determination unit that determines, for each viewer, a predicted viewing program that is an unbroadcast television program predicted to be selected and viewed, using the program selection model. However, acquisition of information necessary for prediction, such as selected program information, is not easy and takes time and cost.

  Further, Non-Patent Document 6 proposes a method of predicting an audience rating of a television drama using only program information before broadcast. There is a description about prediction using a combination of many kinds of feature quantities, such as broadcast stations and time zones, appearing actors, production staff, and topicality of actors. The number of views of Wikipedia (Wikipedia) and the number of posts of Twitter (Twitter) are used as a quantitative indicator of the subjectivity of actors. However, there is a problem that creation of many types of feature quantities is time-consuming, and on the other hand, when only a small number of feature quantities are used, the prediction accuracy is low. In Non-Patent Document 6, an audience rating prediction is made including the number of browsing times of Wikipedia and Twitter related to an actor appearing in a certain drama. However, rather than focusing on the time series data of the number of viewing times, the content described in Wikipedia and Twitter is also taken into consideration, and the calculation amount is large because weighting is performed using many types of prediction factors. It is huge and the load on the system is large.

  Nielsen measures the index (Nielsen social content rating) showing the number of social media (Twitter, Facebook) usage related to TV programs, and shows them to TV companies (networks), sponsors, advertising companies, etc. It is provided as reference information to evaluate the number of viewers' interest and the degree of interest. As the number of uses, a plurality of types of measurement quantities are used, including the number of posts and the number of views (views). The number of posts can be automatically acquired from an interface (API) published as a social media service. On the other hand, the number of views is not public but can be acquired by partnering with a company providing social media services.

  In the inventions described in Patent Document 2 and Non-Patent Document 4, a method of measuring an index indicating the number of messages of social media (Twitter, Facebook) related to a television program, and using it as one of the prediction factors of program ratings Has been devised. In Patent Document 2, a method used for predicting time-shift viewing by recording or the like for one week (Live + 7) immediately after a program broadcast on a television has been devised. Non-Patent Document 4 devised a general learning / prediction method using an index indicating the number of social media (Twitter, Facebook) usage related to television programs as one of the prediction factors for program ratings. ing. However, the information on the number of views (views) of Twitter and Facebook is not generally released to the public, and there is a large workload and cost to acquire. On the other hand, only with the number of posted posts, the correlation with the program viewing rate is weak, and in order to increase the prediction accuracy, it is necessary to combine with a large number of other prediction factors, which takes time and cost.

  Non-patent documents 5 to 12 report researches for estimating television ratings and the number of movie audiences from the number of times of use of social media. Non Patent Literatures 5, 7 and 12 devise a method for predicting the audience rating using the audience rating of the TV drama up to the previous time and the response on the comment on the facebook, etc. Showed that it could improve the prediction of audience rating next time.

  In Non-Patent Documents 3 and 10, a method has been devised to count the number of posts (tweets) to a Twitter related to a program in order to estimate the rating, but the content of the tweets is automatically analyzed. It is difficult to determine the relationship between watching a television program, and the work load is large.

  In Non-Patent Document 8, a method for estimating sales on the first day of movie release is devised. The number of users of movie information sites before movie release is used as one index. However, there is a problem that it requires multiple types of information such as the number of screening places. In addition, there is a problem that it is limited to sales on the first day.

U.S. Patent No. 8,887,188 U.S. Patent Application Publication No. 2016/0148228 U.S. Patent No. 8,112,301 Unexamined-Japanese-Patent No. 2010-130585

Audience Ratings Handbook Video Research Inc. 2017/4 <http://www.videor.co.jp/rating/wh/rgb201704.pdf> Social Content Ratings <http://www.nielsensocial.com/socialcontentratings/> R. Subramanyan, “The Relationship Between Social Media Buzz and TV Ratings. Nielsen Media and Entertainment,” (2011). <Http://www.nielsen.com/us/en/insights/news/2011/the-relationship- between-social-media-buzz-and-tv-ratings.html> S. Sereday, J. Cui, “Using machine learning to predict future ratings,” Nielsen Journal of Measurement, Vol. 1 No. 3, pp. 30-40 (2016). M.-H. Cheng, et al., “Television Meets Facebook: The Correlation between TV Ratings and Social Media,” American Journal of Industrial and Business Management, Vol. 6, NO. 3, pp. 282-290 (2016) DOI: 10.4236 / ajibm.2016.63026 Yusuke Fukushima et al., "Rate Forecasting for TV Dramas Using Only Information Before Broadcasting" Journal of the Institute of Image Information and Television Engineers Vol. 70, No. 11, pp. J255-J261 (2016). Yu-Yang Huang et al, “A Weight-Sharing Gaussian Process Model Using Web-Based Information for Audience Rating Prediction,” “Lecture Notes in Computer Science, Vol. 8916, pp 198-208 (2014) DOI: 10.1007 / 978-3 -319-13987-6_19 M. Mestyan, T. Yasseri, and J. Kertesz, “Early Prediction of Movie Box Office Success Based on Wikipedia Activity Big Data,” PLoS ONE 8 (8): e71226 (2013). DOI: 10.1371 / journal.pone.0071226 S. Wakamiya, R. Lee, and K. Sumiya “Towards Better TV Viewing Rate: Exploiting Crowd's Media Life Logs over Twitter for TV Rating,” Proceedings of the 5th International Conference on Ubiquitous Information Management and Communication ICUIMC '11, Article No. 39. (2011) DOI: 10.1145 / 1968613.1968661 W. T. Hsieh, S. T. Chou, Y. H. Cheng, and C. M. Wu, “Predicting TV Audience Rating with Social Media,” IJCNLP 2013 Workshop on Natural Language Processing for Social Media (Social NLP), pp. 1-5 (2013). J. Kleinberg, “Bursty and Hierarchical Structure in Streams,” Data Mining and Knowledge Discovery, Vol. 7 Issue 4, pp. 373-397 (2003). DOI: 10.1023 / A: 1024940293143 AJC Moreira, MY Santos, "Concave hull: A k-nearest approach for the region occupied by a set of points," In: J. Braz, P.-P. Vazquez, JM Pereira (eds.) GRAPP 2007-Proceedings of the Second International Conference on Computer Graphics Theory and Applications Volume GM-R, pp. 61-68 (2007). ISBN 978-972-8865-71-9.

  As described above, in the method of estimating the audience rating according to the prior art, in the case of conducting a questionnaire directed to a potential audience, the actual survey and tallying work require manpower and are the survey target Even when collecting data directly from a home television receiver, a special monitoring device such as a set top box installed on the television receiver is required. In any case, collecting data that is the basis of estimation requires considerable cost. This cost issue can also be inferred from the fact that current ratings estimates are being conducted only by large companies. In addition, since it is not easy to collect basic data, it takes time to obtain final estimation results. This issue of time delay is evident as no ratings are announced today on the day of the show airing.

  In addition, although methods have also been proposed to estimate television program ratings and movie watching numbers based on the number of times social media is used, accuracy is still low. The reality is that television program ratings and movie audience numbers can be used as basic data for calculating advertising rates. In view of such a reality, it is practically difficult to use a low-reliability numerical value due to low accuracy in a situation where such an economic ripple effect is large.

  An object of the present invention is to solve the problems above existing in the prior art and to provide accurate and low rating information quickly and inexpensively.

  According to the present invention, an audience rating estimation system is provided, wherein the audience rating estimation system estimates a query module to which program specific information is input and an input program specific information provided from the query module. Module, and an information site access frequency database required to provide access counts of information related to a program identified by the program identification information from the estimation module in response to provision of the program identification information; A program rating database requested from the estimation module to provide ratings of past programs related to the program specified by the program specification information in response to the provision; The access frequency to the information related to the program specified by the program specific information from the access frequency database And the ratings of past programs related to the program specified by the program identification information from the program rating database, and based on the correlation between the access count history and the ratings of past programs, Estimate the audience rating of the identified program.

In estimation of audience rating, regression analysis including linear regression can be used. Also, a minimum range model may be used in the estimation of the ratings.
Based on the accuracy of the audience rating estimated in the estimation of the audience rating, the determination of the related program in the information site used in the estimation of the rating and the past program related to the program specified by the program specifying information is again performed There is also a possibility.

A means may also be included to reduce the amount of data by quantitating the data used in the estimation of the ratings.
Also, according to the present invention, in a system comprising a query module coupled in a manner that allows mutual data transmission, an estimation module, a program rating database, and an information site access frequency database A method of estimating is also provided, the method comprising the steps of the query module providing the program identification information input to the query module to the estimation module, and the estimation module responding to the provision of the program identification information. On the other hand, the number of accesses to information related to the program specified by the program specifying information is requested from the information site access number database, and on the other hand, the past program related to the program specified by the program specifying information Requesting an audience rating from a program audience database; and an estimation module The access frequency history to the information related to the program specified by the program specific information is received from the access frequency database, and the ratings of the past programs related to the program specified by the program specific information are received from the program rating database And the estimating module estimating the audience rating of the program specified by the program specifying information based on the correlation between the access count history and the rating of the past program.

  As in the case of the system, regression analysis including linear regression can be used in the estimation of the ratings. In addition, a minimum range model may be used in estimation of audience rating.

  Based on the accuracy of the audience rating estimated in the step of estimating the audience rating, determination of the related information in the information site used in the step of estimating the audience rating and the past program related to the program specified by the program specifying information It may be done again.

It may also include the step of reducing the amount of data by quantizing the data used in the step of estimating the ratings.
Furthermore, the present invention can also be realized as a computer program including computer executable instructions that cause a computer to execute the steps included in the audience rating estimation method according to the present invention. The present invention can also be realized as a computer readable storage medium in which such computer program is stored.

  By using the audience rating estimation method according to the present invention, audience rating information having accuracy that can be used for applications with high economic ripple effect is used using a very common personal computer and information published free of charge on the Internet Can be provided quickly.

FIG. 1 is a block diagram showing a configuration of a rating estimation system according to an embodiment of the present invention. When program specific information is designated, it is the figure which shows the example of the procedure which predicts the audience rating of the program. It is a figure which shows the example of the procedure which acquires the number of accesses (number of views) corresponding to a reference program, and the historical data of audience rating, and extracts a audience rating estimation model. It is a figure which shows the example of the procedure which calculates the value of a prediction factor from the access historical data of a program page. It is a figure which shows the procedure in the case of using the "minimum range" model for estimation of audience rating. It is a figure which shows the definition in the case of using a "linear regression" model for estimation of audience rating, and a formula. It is a figure which shows the example of a query and the selected reference program in the case of estimating an audience rating using a linear regression model method. It is a figure which shows the example of the data of a query and the selected reference program in the case of estimating an audience rating using a linear regression model method. The program count number data in the case of estimating the audience rating using the linear regression model method is visualized as a graph in which the UNIX time is taken on the horizontal axis. The table of predictors and ratings for each program and the predictors are plotted on the horizontal axis with ratings on the vertical axis. It is a figure which shows calculation of the coefficient of a linear regression model, visualization of a linear regression model by a graph, and an estimation result. It is a figure which shows the example of the estimation which used the minimum range model. FIG. 7 is a diagram showing queries and model selection in the case of predicting an audience rating using a linear regression model method. It is a figure which shows calculation of the coefficient of a linear regression model, a prediction result, and visualization of the prediction result by a graph. It is a figure which shows the procedure of an example of the process for improving precision. It is a figure showing an example of processing which changes a program classification, in order to improve accuracy. It is a figure which shows an example which deletes an outlier in the process which changes a selection program, in order to improve accuracy. It is a figure which shows another example which deletes an outlier in the process which changes a selection program, in order to improve accuracy. It is a figure which shows an example which changes the calculation method of a prediction factor, in order to improve accuracy. It is a figure which shows another example which changes the calculation method of a predictor, in order to improve accuracy. It is a figure which shows the further another example which changes the calculation method of a predictor, in order to improve accuracy. It is a figure which shows the procedure of the method of compressing time series data using a probability model. It is an example which converts a time-series data format into a BUR data format. It is an example of the time series of count number and the number of levels. It is graphing which made the UNIX time the horizontal axis of an example of FIG. 11 (a). It is an example of estimation using BLS data, and it is a calculation result of a predictor (in the case of predictor = Last, −3, 3). It is an example of estimation using BLS data, and is a calculation result of a linear regression model (predictor = Last, −3, 3). It is an example of the prediction using BLS data, and is a calculation result of a prediction factor (in the case of a prediction factor = Last, -100, -40). It is an example of the prediction using BLS data, and is a calculation result of a linear regression model (predictor = Last, -100, -40). It is an example of the compression effect. It is a module block diagram in the 2nd Example of this invention.

  FIG. 1 is a block diagram showing the configuration of an audience rating estimation system according to an embodiment of the present invention, in which a plurality of modules are shown. The query module 101 receives, from a user desiring a viewer rating estimate of a television program, an input regarding information identifying the television program, and provides the estimation module 102 with the television station specified in the input and the airing time frame. When the estimation of the rating ends in the estimation module 102 according to the algorithm characteristic of the present invention, the query module 101 receives the estimation result from the estimation module 102 and outputs it to the user.

  The program broadcast information data module 103 is a database storing broadcast information data of each program, such as a television section of a newspaper or a television guide magazine. This database is also updated automatically when the broadcast information is updated as the broadcast schedule is changed or new programs are announced. The stored broadcast information includes a program title, a broadcast station, broadcast time, performers and the like. Also, the program broadcast information data module 103 stores various attribute information of each program, such as that the broadcast information is part of a series in which the program is present, or includes an episode number in the case of a series. ing.

  The program rating data module 104 is a database in which the ratings of past programs are stored. This database is updated each time a new audience rating for each program is announced. Therefore, in the past audience rating data announced by the audience rating research institute up to that point, there may be a limitation of the period required for estimation in the audience rating estimation system according to the present invention. , All are remembered.

  The information site access number data module 105 is a database that provides the number of accesses to an information site on the Internet that provides information related to each program in chronological order. In the present invention, consider, for example, Wikipedia as this information site. It is also possible to use an information site other than Wikipedia.

  As widely known, Wikipedia is an encyclopedia on the Internet using a system called wiki which can edit web pages on a web browser. It is run by the Wikimedia Foundation, and has the feature that anyone who can access the site can participate in editing freely and freely, under a copyleft license. Although the number of languages supported varies from item to item, it has been expanded in many languages around the world. In Wikipedia, the access (URL, also referred to as "view" or "view") number of times (PVC) to the address (URL) to display the text information (page) for each item of the encyclopedia in hourly or hourly hourly units We have released the recorded data. Public data can be downloaded and saved as a file. The information site access count data module 105 stores, for example, time-series data of the access count (PVC) to this Wikipedia. Note that time-series data means PVC per unit time, and is expressed, for example, as (tj, cj). Assuming that the unit time is one hour, (tj, cj) means that the PVC between one hour before a certain time tj and one hour from the time tj is cj times.

Next, FIG. 2 illustrates a procedure in which the audience rating estimation system according to the present invention estimates the audience rating when the television station and the broadcasting time frame are designated by the user. In step 201, identification of a target program is started. For example, the user inputs, via the query module 101, information specifying a program for which an audience rating estimation is desired. For example, the program name and the airing date, or the airing date and the airing time frame may be input. Then, in step 202, the query module 101 acquires program information corresponding to the input information from the program information data module 103 in response to the input from the user. For example, get the airing time and series name of the program. In the next step 203, the type of program is determined based on the program information obtained from the program information data module 103 in step 202. For example, a plurality of features representing the type of program can be provided, and the type of program can be represented as a list of features. For example, the type of a certain program is [genre = drama, broadcasting station = FNS, airing day = MON, airing start time = JST 21:00, program frame = sunday theater, episode = first time, leading performer = Takuya Kimura, stage =] It can be represented by a list of features, such as a hospital]. With regard to the feature of the episode, the feature of the episode can be relatively expressed as “last time” or “two before final time”.
Next, in step 204, the estimation module 102 acquires, from the program information data module 103, a list of reference programs of the same type as the specified program for use in estimation of the rating. For example, if the program for which the estimation of the audience rating is required is the first episode of a series of dramas set in a hospital, one of the series of dramas set in the hospital that was also aired last year The program that was the second episode may be this reference program. In addition, the first episode of a drama series in which a school broadcasted two years ago is the stage and the same actor as the one for which the estimation of audience rating is sought is also on a different stage However, it can be a reference program.

  In step 205, the number of views of the website including information related to these reference programs, that is, for example, the number of accesses to the page of Wikipedia including the related information and the historical data of the rating, It is acquired from the module 105 and the program rating data module 104, respectively, to extract a rating estimation model. The audience rating extraction model will be described later in relation to FIG. Finally, in step 206, the estimation module 102 calculates a prediction factor from the number of accesses to the website corresponding to the program, and uses the calculated prediction factor and the rating estimation model to estimate the user's desired Calculate the value.

  Next, FIG. 3 shows details regarding the operation performed in step 205 of FIG. 2, that is, details regarding extraction of a rating estimation model. First, in step 301, a calculation method of a prediction factor is set. The predictor means a factor that is highly correlated with the audience rating, and in the case of the present invention, for example, the cumulative total access to items in Wikipedia including information related to the program, and a certain relationship with the airing time of the program. The total number of accesses in a certain period is assumed. Therefore, in step 302, the estimation module 102 acquires the time series of the number of accesses of each reference program from the information site access count data module 105 such as Wikipedia and calculates the value of the predictor. The details of the calculation of the prediction factor will be described later with reference to FIG. In step 303, the rating of each reference program is acquired from the program rating data module 104, and in step 304, the prediction factor and the rating ratio are recorded for all reference programs. Then, in step 305, a rating estimation model is extracted using pairs of prediction factors and ratings of all reference programs. Extraction of the audience rating estimation model will be described later in relation to FIG.

FIG. 4 shows an example of a procedure for calculating a predictor. As mentioned above, the present invention uses, as a predictor, the number of accesses to the information site, for example, an item of Wikipedia including information related to a program for which it is desired to estimate the rating. In order to calculate a predictor, first, in step 401, a measurement period of the predictor is set. In that case, for example, by specifying the period start time T1 and the period end time T2, the measurement period of the prediction factor can be expressed as measurement period = [T1, T2]. Also, based on the program broadcast start time showTime, the period start time T1 = showTime + predBegin, the period end time T2 = showTime + predEnd, and using the relative start time predBegin and the relative end time predEnd, the measurement period is set as (predBegin, predEnd). Can be represented.
Then, in step 402, the total number of accesses within the designated period is calculated from the time-series data stored in the information site access count data module 105, and becomes the value of the prediction factor. In addition, as a calculation method of the value of the predictor, it is possible to specify a weighting function with respect to time to calculate the weighting accumulation with respect to the number of accesses in the measurement period.

FIG. 5 shows an example of two methods for estimating the rating based on the values of the predictors of all the reference programs and the rating.
FIG. 5A shows the case of the minimum range model. This is a method using an envelope of a distribution range in a two-dimensional space of a predictor and an audience rating. First, in step 501, the distribution range of polygons is calculated with respect to all reference program pairs in a two-dimensional space whose components are predictors and ratings. In this case, for example, there may be a method using a convex envelope which is the smallest convex set including all pairs, and a method of calculating a concave hull using a k-neighbor. The method of calculating these envelopes is described in Non-Patent Document 12 and is a widely known method. For example, in the case of convex envelope, repeat the operation of selecting the point with the largest right rotation angle with the previous link among all the remaining points, starting from the point where the predictor is the smallest. Allows to construct a polygon representing the distribution range. In the case of a concave envelope, when crossing the previous link starting from the point where the predictor is the smallest, the right rotation angle with the previous link is the k nearest points, excluding that point A polygon representing the distribution range can be constructed by repeating the operation of selecting the largest point.
Then, after a distribution range such as a convex envelope or a concave envelope is calculated in step 501, step 502 has a constant width from the maximum value to the minimum value including the intermediate value for each predictor value. Calculate the audience rating.
FIG. 5 (b) shows an example of using a linear regression model to estimate the audience rating. FIG. 5 (b) also shows an example of a formula for calculating the coefficients of the linear regression model using the ratings of the reference program and the predictors. 6 and 7 show an example of extraction of a rating estimation model using specific program data and count number time series data obtained from Wikipedia.

  FIG. 6A shows an example of a reference program and an extraction result of a rating estimation model corresponding thereto. The query program is a drama series "XYZ" (for example, "beautiful life", "Naoki Hanzawa", etc.), and episode = "last" means the last round of the drama series. The present time is airing start time + 4 days, and (PVC, -3, 3) is set as a calculation method of a prediction factor. That is, this estimation model estimates the audience rating of the final round of "XYZ" using the prediction factor (PVC, -3, 3) at the current time 4 days after the broadcast start time (96 hours later) It means to do. (PVC, -3, 3) means that "the total number of PVCs from 3 days (72 hours) before the broadcast transmission start time to 3 days (72 hours) after the broadcast start time" is used as a predictor . As a reference program selected as a program similar to the query program, the last episode in a plurality of previous series of TBS Sunday theaters in which the ratings and PVC data exist is used, and is listed under FIG. 6 (a) It is done. In the present specification, “estimate” the audience rating means the ex post calculation of the audience rating of the program that has already been broadcasted.

The following FIG. 6 (b) shows an example of data concerning a program series "02-00" which is the second of the TBS Sunday theater selected and listed as a reference program in FIG. 6 (a). . Here, the data is a time series of airing time (start and end time) of each episode, an audience rating, and a count number (PVC) at which a related Wikipedia item is accessed. Here, the time is represented by UNIX time.
FIG. 6C shows, for example, the time series data of the count number of the program series "02-00", which is the second in the list, visualized as a graph with the UNIX time taken on the horizontal axis. At the top of FIG. 6D, a table of predictors and ratings of each program is created. A Pearson correlation of predictors and ratings is calculated, and it can be confirmed that the correlation value of 0.948 is high. On the lower side of FIG. 6 (d), the prediction factor is plotted on the horizontal axis and the ratings on the vertical axis.

  FIG. 6 (e) shows an example in which linear regression is performed based on the above data, and coefficients of the linear regression model are calculated. Shown at the lower left is the estimation result using a linear regression model. On the right side, an example is shown in which the linear regression model is visualized as a straight line. On the other hand, FIG. 6 (f) shows an example using the minimum range model. The upper table shows a list of envelope points selected as convex envelopes. The prediction results of the audience rating corresponding to the prediction factors of the query program are also shown. The minimum value and the maximum value of the viewing rate corresponding to the prediction factor of the query program can be obtained by interpolating the envelope points before and after, and the median value of the minimum value and the maximum value is taken as the estimated value. On the lower side, an example is shown in which the convex envelope and the middle line are visualized.

  An example similar to the example of FIG. 6 but in which the program for which the audience rating is to be known has not yet been broadcast is shown in FIG. Here, the calculation of the ratings of future programs that have not been broadcast yet is referred to as "prediction".

  Here, I would like to briefly explain the relationship between "estimate" and "prediction". As described above, "estimation" is to calculate the rating of a program that has already been broadcast, and "prediction" is to calculate the rating of a program to be broadcast from now. Therefore, in each calculation, although the procedure of calculation is the same, the accumulation range which integrates a predictor in calculation differs. Naturally, in the case of estimation, both pre-broadcast data and post-broadcast data are present, but in the case of prediction, at most, only data up to the current time before broadcast is present It is from. Since the procedure of calculation is the same, in the present specification, both terms may be used without particular distinction unless it causes a problem.

  FIG. 7 (a) shows an example of query and model selection. In the query program, the series name is “PQR” and the final episode is the target. The current time is 40 days (960 hours before) the broadcast start time, and the predictor is (PVC, -100, -40), that is, 100 days before the broadcast start time to the current time from the current time to the related item in Wikipedia It is the total number of accesses. The reference program is the same as in FIG.

  The following Fig. 7 (b) shows data on the final round of each series of TBS Sunday theater selected and listed as a reference program in Fig. 7 (a) and an example of the ratings prediction result using it. It is done. On the right side of FIG. 7 (b), the prediction factor is plotted on the horizontal axis and the ratings on the vertical axis.

  In the estimation of the ratings according to the invention, additional processing can be performed to improve the accuracy. The actual steps of the process to improve the accuracy of the estimation are shown in FIG. Steps 801 to 806 in FIG. 8 correspond to steps 201 to 206 in FIG. However, at step 806, once the ratings have been estimated, additional processing is done to improve accuracy. As shown on the right side of step 806 of FIG. 8, for example, there are three additional processes.

  However, in order to improve the accuracy, it is necessary to evaluate how accurate the estimated value has been obtained by the estimation method according to the present invention. To that end, after the estimate of the audience rating is calculated in step 806 according to the invention, in step 807 the accuracy of the estimation model or the accuracy of the calculated estimation must be evaluated. Depending on the evaluation, a process change or correction method is selected.

  As a method of calculating the accuracy of the model, there is known a method of calculating a Pearson correlation coefficient using a predicted factor calculated for a reference program and an observed rating. If this Pearson correlation coefficient is high, it is evaluated that the accuracy is high. As a method of calculating the accuracy of the model, a model is extracted using some reference programs (for example, one reference program is excluded one by one), and the error between the predicted value and the observed value estimated for the remaining programs The average value is calculated, and the lower the average value of the error, the higher the accuracy. On the other hand, as a method of calculating the accuracy of the prediction value, change the accumulation period of the prediction factor (for example, change the accumulation start time and the accumulation end time by one hour), and calculate the variance of the obtained plurality of prediction values calculate. If the variance is small, the accuracy is evaluated as high. In order to evaluate the accuracy of the calculated estimated model or the accuracy of the estimated value, for example, there is a general method of setting a certain threshold in advance and judging whether or not the predetermined threshold is exceeded. is there. As mentioned above, in the present invention it is possible to use, for example, the three processing directions shown in FIG. 8 (a) in order to increase the accuracy of the estimates. That is, a change in a program classification method, a change in a selected program (a program selected for use in model extraction in the same type of program), a change in a method of calculating a prediction factor.

  As a method of changing the classification of programs in order to improve the accuracy, it is possible to change the determination conditions for classifying the same type using program information. For example, it can be added as a condition for judging that the broadcasting time zone, the broadcasting year and the cast members are the same. If the program classification is to be changed, the process returns to step 803 after step 807. The change of the selected program is to change the program to be selected for use in model extraction within the same type of reference program. In order to change the selected program, it is an example to delete outlier data which is separated from the straight line of linear regression by a certain limit. In this case, after step 807, the process returns to step 804. Then, as a method of changing the calculation method of the prediction factor, for example, the number of accesses is weighted before addition using a weight that attenuates according to a change of the start time of the accumulation period or a time difference from the airing time. It is possible to do. In other words, such weighting is performed assuming that the probability of actually watching the broadcast is higher when the information site is accessed at a point closer to the broadcast time than when it is farther from the broadcast time. . In this case, after step 807, the process returns to step 805.

  Moreover, it is possible to change each parameter used for calculation of a predictor, for example, as follows. That is, by setting the parameter to be automatically changed in accordance with a previously designated procedure, it is possible to maximize the correlation between the prediction factor for the reference program and the audience rating. As an example of automatically changing the value of the parameter according to the specified procedure, the value of the start time T1 is specified from a specified value T1-min (for example, T1-min = 30 days) It is conceivable to increase at each designated time interval (for example, one hour) up to T1-max (for example, T1-min = current time-the broadcast start time day).

  8 (b) to 8 (g) show examples of specific processing for improving the accuracy. FIG. 8 (b) is an example of a process of changing a program classification, and FIG. 8 (c) is an example of deleting an outlier to change a selected program when using a linear regression model. (D) is another example of deleting outliers in order to change the selected program when the minimum range model is used, and FIG. 8 (e) to FIG. 8 (g) are each a calculation method of a predictor There are three examples of processing for changing.

  Above, the processing performed in the present invention to improve the accuracy of the audience rating estimation has been described. Next, a method of reducing the amount of historical data by quantizing time-series data will be described. The quantization of time series data means expressing as a time series of small discrete values. Although it is necessary to use data of a large number of reference programs in order to improve reliability, when the number of reference programs increases, the amount of time-series data also increases as a result, and the storage capacity of the required data Increases, causing access delay to data. Therefore, it is desirable to reduce storage capacity, processing capacity and processing delay by converting time series data into a compressed format.

  However, when the time series data is converted into a compressed format, the correlation between the prediction factor and the audience rating deteriorates. It is necessary to estimate the correlation with the audience rating. Therefore, a compression type time-series data recording method that does not deteriorate the correlation between the prediction factor and the audience rating is required. In order to solve this problem, in the present invention, we would like to propose a time-series data conversion method using a probability model.

  FIG. 9 shows a procedure for converting time-series data using a probability model from input of time-series data of step 901 to output of period representation data of step 906. This procedure is characterized in setting the cost function in step 903 after a plurality of discrete levels have been set in step 902. As for the level cost, a cost function based on the probability function for each discrete level is set, but a cost function that reduces the cost according to the magnitude of the probability is used. Then, with regard to the level transition cost, a cost function that raises the cost according to the amount of increase in level is used for time transition from a low level to a high level. As a set of discrete level time series, a time series that minimizes the total cost of level cost and level transition cost for the entire time series is determined. That is, although there is a trade-off between reduction of time-series data volume and correlation between predictors and ratings, in order to give an optimal solution regarding the correlation between these two, it is based on a probability model A method called time series cost function is introduced.

As an example of the cost function, it is possible to use, for example, the function described in Non-Patent Document 6. That is, as the level cost function, a binomial distribution B (P, N) using the probability P and the total number N as parameters is used. The level index J is 0, 1, 2, 3,. . Let P 0 * 2 ^J be the probability P of the binomial distribution. Here, the defined probability P0 and the total number N are common to all level cost probability functions. Cost Level J for counting the number of X at a certain time, the log (X) -log (B ( P0 * 2 J, N)).

  Next, how to determine the defined probability P0 and the total number N used to calculate the level cost will be described. Although the reference time series is selected, the time series of the program with the highest audience rating can be used, and the time series of the number of accesses of the entire information site can also be used. N is the maximum value of time series. P0 is the ratio of the average value to N. N can be set as a function of time with a period of 24 hours or a week. N can also be determined to correspond to the type of program.

  Next, after the cost function is determined in step 904, the values at each time of the input time series are represented as (time, discrete level) in step 904, and the time series with the smallest cost is determined. In step 905, for the time series of discrete levels which is the output of step 904, the same level value represents the continuation period by (start time, end time, level value), and is converted into a period representation. In the final step 906, compressed data of period representation is output.

  10 and 11 show examples of output data. That is, in FIG. 10, input at the time of changing to PVC (PVC data format) data for each time of input time series data and discrete level time series output data (BLS data format) obtained in step 904, and An example with the output is shown. An example in which the BLS data format is converted to period representation data output (BUR format) in step 905 is also shown. FIGS. 11 (a) and 11 (b) show an example of the actual calculation result and a graph plotting UNIX time on the horizontal axis.

FIG. 12 (a) is an example of an estimation model using BLS data, and is a calculation result in the case of a prediction factor = (Last, −3, 3). FIG. 12 (b) is an example of estimation using BLS data, which is the calculation result of a linear regression model using a predictor = (Last, −3, 3). Further, FIG. 13 (a) is an example of prediction using BLS data, and is a calculation result in the case of a prediction factor = (Last, -100, -40). FIG. 13 (b) is an example of prediction using BLS data, which is the calculation result of a linear regression model using a predictor = (Last, −100, −40).
FIG. 14 illustrates the effect of time series data compression. As described above, by compressing data using, for example, a probability model, necessary data compression becomes possible without losing the correlation between the predictor and the rating, and the rating estimation according to the present invention It will be possible to run on a very common personal computer without using a large and expensive system.

The above has described an embodiment focusing on an algorithm for estimating (predicting) the audience rating according to the present invention. Next, refer to FIG. 15 for a second embodiment relating to, for example, the implementation of a system capable of estimating (predicting) the rating according to the algorithm according to the present invention when the necessary conditions are input. I would like to explain by
FIG. 15 shows the configuration of each element constituting the second embodiment. First, the user interface 1501 receives a program arbitrarily designated by the user as an input, and provides the input information to the query module 1502. This user interface 1501 may be implemented as a client application, a website or an application for a smartphone, for example, when a user desiring to estimate the rating accesses the system according to the present invention via the network. The user interface 1501 also receives the estimated ratings from the estimation module 1503 when the estimation according to the algorithm of the present invention is made. The estimation module 1503 records a plurality of estimation models and returns estimation results for the type of program specified by the user interface 1501. For example, an estimation model records, as a model, a coefficient of a linear regression model or a list of convex envelope points of a minimum range model as a model for each of a plurality of reference program lists, and a predictor of a designated program is accessed as an information site The estimated audience rating for the predictor is obtained from the estimated model, and the result is returned to the user interface 1501. The estimation module obtains time-series data from the information site access number data 1504 to calculate a prediction factor. Also, the estimation model obtains a new model from the learning module 1505 when there is no model corresponding to the specified program. The learning module 1505 is a module that learns an estimated model using data of a reference program. The learning result is reported to the estimation module 1503. The time-series analysis module 1508 is a module that calculates a prediction factor from time-series data, and converts the format of compressed data of time-series data as needed. The data acquisition module 1509 is a module for acquiring information site access count (PVC) data.

  Also, as another usage method, the query module 1502 returns, from the information site access number data 1504, a search result having a time series similar to the program specified by the user interface 1501.

Further, the user interface 1501 can receive a program specified by the user as an input, and can present the result of each process described above to the user. For example, for a specified program, another program having high time-series correlation can be searched and the result can be returned to the user.

Claims (10)

A query module into which program specific information is input;
The estimation module provided from the query module, the inputted program specific information;
An information site access number database which is requested by the estimation module to provide the number of accesses to information associated with the program specified by the program specific information in response to the provision of the program specific information;
A program rating database requested to provide ratings of past programs related to the program specified by the program specifying information from the estimation module in response to the provision of the program specifying information;
Equipped with
The estimation module receives, from the information site access frequency database, a history of access frequency to information related to a program specified by the program specification information, and a program specified by the program specification information from the program rating database In a system for receiving an audience rating of a past program related to the above and estimating the audience rating of a program specified by the program specifying information based on the correlation between the access count history and the rating of the past program.
The number of times of access to information associated with the program specified by the program specifying information stored in the information site access frequency database is quantized to reduce the amount of data thereof,
The quantization is
Multiple discrete levels set
Set a cost function having a cost based on the probability function for each discrete level and a level transition cost that raises the cost according to the amount of level increase for time transition from low level to high level,
The level cost and the level transition cost of the entire access count history to the information related to the program specified by the program specifying information stored in the information site access count database as the set discrete level time series Find a time series that minimizes the overall cost,
Made by,
A system characterized by A query module into which program specific information is input;
The estimation module provided from the query module, the inputted program specific information;
An information site access number database which is requested by the estimation module to provide the number of accesses to information associated with the program specified by the program specific information in response to the provision of the program specific information;
A program rating database requested to provide ratings of past programs related to the program specified by the program specifying information from the estimation module in response to the provision of the program specifying information;
Equipped with
The estimation module receives, from the information site access frequency database, a history of access frequency to information related to a program specified by the program specification information, and a program specified by the program specification information from the program rating database In a system for receiving an audience rating of a past program related to the above and estimating the audience rating of a program specified by the program specifying information based on the correlation between the access count history and the rating of the past program.
A minimum range model is used in the estimation of the audience rating of the program specified by the program specifying information performed by the estimation module;
The minimum range model is
The envelope of the distribution range in a two-dimensional space of the correlation between the access count history and the rating of the past program for a plurality of past programs related to the program specified by the program specification information is convexly enveloped or concaved It is calculated by the envelope,
The estimation module estimates an audience rating of a program specified by the program specifying information performed by the estimation module, the value on the envelope of the distribution range of the correlation between the access count history and the rating of the past program. Output as maximum and minimum values of,
A system characterized by The system according to claim 1, wherein regression analysis including linear regression is used in the estimation of the audience rating.   Based on the accuracy of the rating estimated in the estimation of the rating, the determination of the related information in the information site used in the estimation of the rating and the past program related to the program specified by the program specifying information is again performed The system according to any one of the preceding claims, which is performed. What is claimed is: 1. A method of estimating program ratings in a system comprising: a query module coupled in a mutually data-transferable manner, an estimation module, a program rating database, and an information site access frequency database,
The query module provides the program identification information input to the query module to the estimation module;
The estimation module, in response to the provision of the program identification information, requests the access frequency to the information associated with the program identified by the program identification information from the information site access frequency database, and Requesting an audience rating of a past program associated with the program identified by the program identification information from a program audience database;
The estimation module receives, from the information site access frequency database, a history of access frequency to information related to a program specified by the program specification information, and a program specified by the program specification information from the program rating database Receiving ratings of past programs associated with the
The estimation module estimating the rating of the program specified by the program specifying information based on the correlation between the access count history and the rating of the past program;
In a method that includes
The number of times of access to information associated with the program specified by the program specifying information stored in the information site access frequency database is quantized to reduce the amount of data thereof,
The quantization is
Multiple discrete levels set
Set a cost function having a cost based on the probability function for each discrete level and a level transition cost that raises the cost according to the amount of level increase for time transition from low level to high level,
The level cost and the level transition cost of the entire access count history to the information related to the program specified by the program specifying information stored in the information site access count database as the set discrete level time series Find a time series that minimizes the overall cost,
Made by,
A method characterized by What is claimed is: 1. A method of estimating program ratings in a system comprising: a query module coupled in a mutually data-transferable manner, an estimation module, a program rating database, and an information site access frequency database,
The query module provides the program identification information input to the query module to the estimation module;
The estimation module, in response to the provision of the program identification information, requests the access frequency to the information associated with the program identified by the program identification information from the information site access frequency database, and Requesting an audience rating of a past program associated with the program identified by the program identification information from a program audience database;
The estimation module receives, from the information site access frequency database, a history of access frequency to information related to a program specified by the program specification information, and a program specified by the program specification information from the program rating database Receiving ratings of past programs associated with the
The estimation module estimating the rating of the program specified by the program specifying information based on the correlation between the access count history and the rating of the past program;
In a method that includes
A minimum range model is used in the estimation of the audience rating of the program specified by the program specifying information performed by the estimation module;
The minimum range model is
The envelope of the distribution range in a two-dimensional space of the correlation between the access count history and the rating of the past program for a plurality of past programs related to the program specified by the program specification information is convexly enveloped or concaved It is calculated by the envelope,
The estimation module estimates an audience rating of a program specified by the program specifying information performed by the estimation module, the value on the envelope of the distribution range of the correlation between the access count history and the rating of the past program. Output as maximum and minimum values of,
A method characterized by The method according to claim 5, wherein in the estimation of the audience rating, regression analysis including linear regression is used.   Related information in an information site used in the step of estimating the rating based on the accuracy of the rating estimated in the step of estimating the rating, and a past program related to a program specified by the program specifying information The method according to any one of claims 5 to 7, wherein the determination of is made again.   A computer program comprising computer executable instructions that cause a computer to perform the steps included in the method according to any one of claims 5 to 8.     A computer readable storage medium having stored thereon a computer program comprising computer executable instructions for causing a computer to perform each step included in the method according to any one of claims 5 to 8.