Prediction system, prediction method, and computer program

Abstract

To improve prediction accuracy of a viewing amount of a program.SOLUTION: A prediction system inputs, into a deep learning model, input information which includes first information including a broadcast schedule date of a prediction target program, second information including broadcast dates and viewing amounts of a plurality of programs in the same time zone which are a plurality of programs broadcasted in a past time zone corresponding to the broadcast schedule date of the prediction target program, third information including the broadcast schedule date of at least one adjacent program to be broadcasted in a time zone adjacent to the broadcast schedule date of the prediction target program, and fourth information including the broadcast dates and the viewing amounts of a plurality of past adjacent programs being a plurality of programs broadcasted in time zones respectively adjacent to the broadcast dates of the programs in the same time zone, and the prediction system obtains a predictive value of the viewing amount of the prediction target program, as an output of the deep learning model.SELECTED DRAWING: Figure 11

Description

[Technical Field]

[0001] The present disclosure relates to a forecasting system. [Background technology]

[0002] There are known techniques for predicting the audience ratings of programs broadcast on broadcast media such as television and the Internet. For example, Patent Document 1 discloses a system that extracts related programs related to a target program from past programs and predicts the audience ratings of the target program using the related programs. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 7399225 Summary of the Invention [Problem to be solved by the invention]

[0004] However, it has been difficult to predict the audience rating of a target program with high accuracy by simply taking into account the audience rating of related programs, as has been done conventionally. One aspect of the present disclosure is to improve the accuracy of predicting the viewing volume of a program. [Means for solving the problem]

[0005] One aspect of the present disclosure is a prediction system comprising an acquisition unit and a prediction unit. The acquisition unit is configured to acquire first information including the scheduled broadcast date and time of the program to be predicted, second information including the broadcast dates and times and viewing volumes of multiple same-time slot programs, which are multiple programs that were broadcast in a past time slot corresponding to the scheduled broadcast date and time of the program to be predicted, third information including the scheduled broadcast date and time of at least one adjacent program that will be broadcast in a time slot adjacent to the scheduled broadcast date and time of the program to be predicted, and fourth information including the broadcast dates and times and viewing volumes of multiple past adjacent programs, which are multiple programs that were broadcast in time slots adjacent to each of the broadcast dates and times of the multiple same-time slot programs.

[0006] The prediction unit is configured to input input information including the scheduled broadcast date and time of the program to be predicted identified from the first information, the broadcast date and time and viewing volume of multiple programs in the same time slot identified from the second information, the scheduled broadcast date and time of at least one adjacent program identified from the third information, and the broadcast date and time and viewing volume of multiple past adjacent programs identified from the fourth information into the deep learning model, and obtain a predicted value of the viewing volume of the program to be predicted as an output of the deep learning model, thereby predicting the viewing volume of the program to be predicted.

[0007] According to this prediction system, the input information used to predict the viewership volume of the program to be predicted includes the scheduled broadcast dates and times of adjacent programs that may affect the viewership volume of the program to be predicted, as well as the broadcast dates and times and viewership volumes of past adjacent programs, thereby improving the accuracy of the prediction of the viewership volume of the program to be predicted.

[0008] In one aspect of the present disclosure, the first information may include, as the scheduled broadcast date and time of the program to be predicted, information capable of identifying the scheduled broadcast start date and time of the program to be predicted. The second information may include, as the broadcast dates and times of multiple programs in the same time slot, information capable of identifying the broadcast start dates and times of multiple programs in the same time slot. The multiple programs in the same time slot may be programs that were previously broadcast on the same day of the week as the program to be predicted by the same broadcast station, and that had the broadcast start date and time closest to the broadcast start date and time of the program to be predicted in each of the past multiple weeks.

[0009] With this configuration, even if the broadcast start date and time of the program in the same time slot does not exactly match the broadcast start date and time of the program to be predicted, the broadcast start date and time of the program in the same time slot that is closest to the broadcast start date and time of the program to be predicted can be included in the input information, thereby improving the accuracy of predicting the viewing volume of the program to be predicted.

[0010] The aspect of the present disclosure may further include a generating unit configured to generate a first feature amount and a plurality of second feature amounts. The first feature amount is a feature amount of a program to be predicted. The second feature amounts are each a feature amount of a plurality of programs broadcast in the same time slot.

[0011] The first feature amount may be generated based on the scheduled broadcast date and time, identification information of the broadcasting station, and title of each of the program to be predicted and at least one adjacent program. The second feature amounts may be generated based on the broadcast date and time, identification information of the broadcasting station, and title of each of the programs broadcast in the same time slot and the programs adjacent to each other in the past.

[0012] The prediction unit may be configured to input information including the first feature amount and a plurality of second feature amounts as input information to the deep learning model. With this configuration, it is possible to include the broadcast date and time, broadcast station, and title of a program related to the program to be predicted in the input information, and it is possible to relatively increase the influence of programs similar to the program to be predicted on the viewership prediction, thereby improving the accuracy of the viewership prediction for the program to be predicted.

[0013] In one aspect of the present disclosure, the deep learning model may be configured to calculate a predicted value using time series information included in the input information. In one aspect of the present disclosure, the generating unit may be configured to generate the plurality of second feature amounts so as to include, as time series information, information indicating the broadcast order of the plurality of programs broadcast in the same time slot.

[0014] With this configuration, time-series information can be included in the input information, which improves the accuracy of prediction of the viewing volume of the program to be predicted. According to one aspect of the present disclosure, there may be provided a computer-implemented prediction method, which includes: acquiring first information including a scheduled broadcast date and time of a program to be predicted; second information including broadcast dates and times and viewership amounts of multiple concurrent programs that were broadcast in a past time slot corresponding to the scheduled broadcast date and time of the program to be predicted; third information including a scheduled broadcast date and time of at least one adjacent program that will be broadcast in a time slot adjacent to the scheduled broadcast date and time of the program to be predicted; and fourth information including broadcast dates and times and viewership amounts of multiple past adjacent programs that were broadcast in time slots adjacent to each of the broadcast dates and times of the multiple concurrent programs; inputting input information including the scheduled broadcast date and time of the program to be predicted identified from the first information, the broadcast dates and times and viewership amounts of the multiple concurrent programs identified from the second information, the scheduled broadcast date and time of at least one adjacent program identified from the third information, and the broadcast dates and times and viewership amounts of the multiple past adjacent programs identified from the fourth information, into a deep learning model; and obtaining a predicted value of the viewership amount of the program to be predicted as an output of the deep learning model, thereby predicting the viewership amount of the program to be predicted.

[0015] This prediction method provides the same effects as the prediction system described above. In one aspect of the present disclosure, a computer program may be provided for causing a computer to execute, at least in part, the above-described prediction method. [Brief explanation of the drawings]

[0016] [Figure 1] FIG. 1 is a block diagram showing a configuration of a prediction system. [Figure 2] 10 is a flowchart illustrating a learning process. [Figure 3] FIG. 10 is a diagram illustrating past programs aired in the same time slot as a program to be learned that are associated with the program to be learned in the learning process. [Figure 4] FIG. 10 is a diagram illustrating adjacent programs associated with a program to be learned in the learning process. [Figure 5] 10 is a diagram illustrating past adjacent programs associated with a program in the same time slot in the learning process. FIG. [Figure 6]10 is a flowchart showing a process of generating first features for learning. [Figure 7] FIG. 10 is a diagram illustrating a process of generating a first feature for learning. [Figure 8] 10 is a flowchart showing a process of generating second features for learning. [Figure 9] FIG. 10 is a diagram illustrating a process of generating second features for learning. [Figure 10] FIG. 1 is a diagram illustrating a process for generating a deep learning model. [Figure 11] 10 is a flowchart illustrating a prediction process. [Figure 12] FIG. 10 is a diagram illustrating past programs in the same time slot that are associated with a program to be predicted in the prediction process. [Figure 13] FIG. 10 is a diagram illustrating adjacent programs associated with a program to be predicted in the prediction process. [Figure 14] FIG. 10 is a diagram illustrating past adjacent programs associated with a program in the same time slot in the prediction process. [Figure 15] 10 is a flowchart showing a process of generating a first feature for prediction. [Figure 16] FIG. 10 is a diagram illustrating a process of generating a first feature amount for prediction. [Figure 17] 10 is a flowchart showing a process of generating a second feature for prediction. [Figure 18] FIG. 10 is a diagram illustrating a process of generating a second feature for prediction. [Figure 19] FIG. 10 is a diagram illustrating a process for predicting the viewing volume of a program to be predicted. DETAILED DESCRIPTION OF THE INVENTION

[0017] Hereinafter, exemplary embodiments of the present disclosure will be described with reference to the drawings. [1. First embodiment] [1-1.Configuration] [1-1-1. Overall structure] 1 is a system for predicting the viewing volume of a program broadcast on a broadcasting medium such as television or the Internet. The prediction system 100 is installed in an information terminal such as a personal computer.

[0018] The prediction system 100 includes a prediction device 1, a database 2, and an artificial intelligence 3. The prediction device 1 includes a processor 10, a memory 20, a storage 30, a user interface 40, and a communication interface 50.

[0019] The processor 10 is configured to execute processing in accordance with a computer program recorded in the storage 30 . The memory 20 is used as a working area when the processor 10 executes processing. Examples of the memory 20 include a RAM (Random Access Memory), a ROM (Read Only Memory), and a flash memory.

[0020] The storage 30 holds computer programs and data used when executing processes according to the computer programs. Examples of the storage 30 include a hard disk drive (HDD) and a solid state drive (SSD).

[0021] The user interface 40 is a general term for an interface for receiving various input operations from a user and an interface for outputting various information to the user. Examples of the user interface 40 include a keyboard, a mouse, a touch panel, a display, etc.

[0022] The communication interface 50 is an interface capable of communicating various data in accordance with a predetermined standard. The prediction device 1 is configured to be able to communicate with the database 2 and the artificial intelligence 3 through the user interface 40.

[0023] The database 2 includes program information on programs that have been broadcast in the past (hereinafter referred to as past programs) and program information on programs that are scheduled to be broadcast in the future (hereinafter referred to as scheduled programs). The program information includes information indicating the broadcast start date and time of the program, the broadcast end date and time, the title, the broadcast duration, and the broadcast station that will broadcast the program. The program information of past programs further includes information indicating the viewing volume of the program.

[0024] In this embodiment, the viewing amount represents a viewing rate or a parameter related to a viewing rate. The viewing amount may be, for example, a GRP (Gross Rating Point) representing a total viewing rate for each household, or a TRP (Target Rating Point) representing a total viewing rate for each individual.

[0025] The artificial intelligence 3 is configured to generate a deep learning model 31 by learning the input information when information including program information of past programs is input as input information. The deep learning model 31 is configured to calculate and output a predicted value of the viewing volume of the program to be predicted when information including the scheduled broadcast date and time of the program to be predicted and the broadcast dates and viewing volumes of multiple programs that have been broadcast in the past is input as input information.

[0026] The scheduled broadcast date and time is the broadcast date and time of the program scheduled to be broadcast. The broadcast date and time is information indicating the date and time slot on which the program specified by the program information will be broadcast. The broadcast date and time may or may not include information regarding the day of the week on which the program will be broadcast. The broadcast date and time may be explained, for example, by the broadcast start date and time and broadcast end date and time of the program included in the program information, or by the broadcast start date and time and broadcast duration of the program.

[0027] The artificial intelligence 3 may be configured to further use the time series information to generate a deep learning model 31. The deep learning model 31 may be configured to further use the time series information to calculate a predicted value of the viewing amount.

[0028] The time series information is, for example, information based on the broadcast date and time of a program. Examples of the artificial intelligence 3 and deep learning model 31 that perform processing using time series information include Transformer, LSTM (Long Short Term Memory), and RNN (Recurrent Neural Network). Transformer includes a self-attention layer, and evaluates the association between elements including time series information through the self-attention layer. By evaluating the association between elements, the deep learning model 31 using Transformer can predict viewing volume with high accuracy, taking past programs into consideration.

[0029] [1-1-2. Functional configuration] As shown in FIG. 1, a processor 10 functions as an acquisition unit 11, a generation unit 12, and a prediction unit 13 by executing a computer program. The acquisition unit 11 is configured to acquire program information of past programs and program information of programs scheduled to be broadcast. The generating unit 12 is configured to generate feature quantities of a program based on the program information. The prediction unit 13 is configured to input program information into the deep learning model 31 and obtain a predicted value of the viewing volume of the program to be predicted as an output of the deep learning model 31, thereby predicting the viewing volume of the program to be predicted. The specific processes of the acquisition unit 11, generation unit 12, and prediction unit 13 will be described later. [1-2. Processing] In the prediction system 100, the prediction device 1 is configured to execute the learning process shown in FIG. 2 and the prediction process shown in FIG. 11 by mutually communicating with the database 2 and the artificial intelligence 3.

[0030] The learning process is a process in which each of a plurality of past programs from a predetermined period in the past (for example, the past five years) is set as a learning target program, and a deep learning model 31 is generated for calculating a predicted value of the viewing volume of the learning target program based on the program information of each learning target program and the program information of other past programs associated with the learning target program. The learning process generates the deep learning model 31 by having the artificial intelligence 3 perform supervised learning using the viewing volume of the learning target program as training data.

[0031] The prediction process is a process for predicting the viewing volume of a program to be predicted by setting at least one program scheduled to be broadcast in a predetermined future period as a program to be predicted, and calculating a predicted value of the viewing volume of the program to be predicted using a deep learning model 31 generated by the learning process.

[0032] The following describes the processes executed by the processor 10 of the prediction device 1 in each of the learning process and the prediction process. [1-2-1. Learning process] The learning process executed by the processor 10 will be described with reference to the flowchart of Fig. 2 and Fig. 3 to Fig. 10. Below, the process of learning about one program to be learned will be described.

[0033] When a predetermined instruction is input from the user via the user interface 40, the processor 10 starts the learning process shown in FIG. First, in S10, the processor 10 acquires program information of a plurality of past programs from the database 2 via the communication interface 50. For example, the processor 10 acquires program information of past programs for 100 weeks. The processing of S10 corresponds to the processing of the acquisition unit 11.

[0034] Next, in S11, processor 10 selects one program to be studied from the multiple past programs, and associates the program with past programs that aired in the same time slot as the program to be studied (see FIG. 3). In the study process, a program that aired in the same time slot as the program to be studied is a past program that aired on the same day of the week, in the same time slot, and by the same broadcasting station as the program to be studied. In S11, processor 10 associates multiple programs that air in the same time slot with one program to be studied.

[0035] Hereinafter, in the learning process, the program information of the program to be learned is referred to as first learning information 41. The program information of multiple programs broadcast in the same time slot is referred to as second learning information 42. In S11, the first learning information 41 and the second learning information 42 are associated with each other.

[0036] Here, a program that was broadcast in the same time slot as the program to be studied refers to a program that has a broadcast start time that is closest to the broadcast start time of the program to be studied on each day. If there is no past program that was broadcast on the same day of the week and by the same broadcasting station as the program to be studied in a given week and that has a broadcast start time that matches the broadcast start time of the program to be studied, the program with the broadcast start time closest to the broadcast start time of the program to be studied is associated as the program broadcast in the same time slot in that week.If there are multiple programs with the broadcast start time closest to the broadcast start time of the program to be studied, any one of the programs is associated as the program broadcast in the same time slot in that week.

[0037] For example, if the broadcast start date and time of the program to be studied is "2023 / 9 / 3 19:58," and two past programs that were broadcast on the same day of the week and by the same broadcasting station one week before the program to be studied had broadcast start times of "2023 / 8 / 27 18:58" and "2023 / 8 / 27 20:58," respectively, one of the past programs will be associated as the program in the same time slot. In this case, either program may be associated as the program in the same time slot.

[0038] 3 shows an example of the first learning information 41, which includes the broadcast start date and time of the program to be studied (2023 / 9 / 3 19:58), the broadcast end date and time of the program to be studied (2023 / 9 / 3 20:54), the broadcast station of the program to be studied (ZZ TV), the title of the program to be studied (AA News), and the target audience rating of the program to be studied. The target audience rating is the audience rating for each of the C (child) demographic, T (teen) demographic, M (male)1-M3 demographic, and F (female)1-F3 demographic, which are gender and age categories commonly used in the marketing field. As an example, FIG. 3 shows the C demographic audience rating of "3.2" and the T demographic audience rating of "5.5" corresponding to the program to be studied.

[0039] 3 also shows, as an example of second learning information 42, the broadcast start date and time of the program in the same time slot, the broadcast end date and time of the program in the same time slot, the broadcasting station of the program in the same time slot, the title of the program in the same time slot, and the target audience rating for the program in the same time slot for each of a number of past weeks, similar to first learning information 41. In other words, second learning information 42 includes program information for programs in the same time slot for the past N weeks (N is an integer of 2 or more) from the broadcast date and time of the program to be studied.

[0040] Next, in S12, processor 10 associates at least one adjacent program that was broadcast in a time slot adjacent to the broadcast date and time of the program to be studied with the program to be studied (see FIG. 4). The time slot adjacent to the broadcast date and time of the program to be studied includes at least one of the time slots before and after the broadcast date and time of the program to be studied. The broadcast date and time of the adjacent program need only be included in the time slot adjacent to the broadcast date and time of the program to be studied, so the broadcast start date and end date and time of the adjacent program do not necessarily have to be adjacent to the broadcast date and time of the program to be studied. In other words, the broadcast start date and time of the adjacent program may or may not be adjacent to the broadcast end date and time of the program to be studied. The broadcast end date and time of the adjacent program may or may not be adjacent to the broadcast start date and time of the program to be studied. For example, in FIG. 4, the broadcast end date and time of the program to be studied may be "2023 / 9 / 3 20:54," while the time period adjacent to the broadcast date and time of the program to be studied may be from "2023 / 9 / 3 20:54" to "2023 / 9 / 3 22:00." In this case, the adjacent programs may be programs broadcast between "2023 / 9 / 3 20:54" and "2023 / 9 / 3 22:00." In other words, in FIG. 4, programs with the titles "Music EE," "Drama FF," and "GG Weather" may all be associated with the program to be studied as adjacent programs. For example, processor 10 may associate a program with the title "Drama FF" with the program to be studied as an adjacent program, while not including programs with the titles "Music EE" or "GG Weather."

[0041] Hereinafter, the program information of at least one adjacent program is referred to as third learning information 43. In S12, the first learning information 41 and the third learning information 43 are associated with each other. 4 shows, as an example of third learning information 43, program information for one program that was broadcast before the program to be learned and three programs that were broadcast after the program to be learned. Although a total of four adjacent programs are shown in FIG. 4, the number of adjacent programs is not limited to four. For example, the adjacent programs may be two or more programs that were broadcast before the program to be learned, or two or less or four or more programs that were broadcast after the program to be learned.

[0042] The adjacent programs associated with the program to be studied may be only programs that were broadcast before the program to be studied, or may be only programs that were broadcast after the program to be studied. Next, in S13, processor 10 associates with the same-time-slot program a plurality of past adjacent programs, which are a plurality of programs that were broadcast in time slots adjacent to the broadcast date and time of the same-time-slot program (see FIG. 5). The time slots adjacent to the broadcast date and time of the same-time-slot program include at least one of the time slots before and after the broadcast date and time of the same-time-slot program. The broadcast date and time of the past adjacent program need only be included in a time slot adjacent to the broadcast date and time of the concurrent program, so the broadcast start date and end date and time of the past adjacent program do not necessarily have to be adjacent to the broadcast date and time of the concurrent program. In other words, the broadcast start date and time of the past adjacent program may or may not be adjacent to the broadcast end date and time of the concurrent program. The broadcast end date and time of the past adjacent program may or may not be adjacent to the broadcast start date and time of the concurrent program. For example, in FIG. 5, the broadcast end date and time of the concurrent program may be "2023 / 8 / 27 21:00," while the time slot adjacent to the broadcast date and time of the concurrent program may be from "2023 / 8 / 27 21:00" to "2023 / 8 / 27 22:30." In this case, the past adjacent programs may be programs broadcast from "2023 / 8 / 27 21:00" to "2023 / 8 / 27 22:30." That is, in FIG. 5, programs with the titles "Drama II," "JJ Weather," and "Sports KK" may all be associated with the concurrent program as past adjacent programs. For example, processor 10 may associate a program with the title "Sports KK" with the concurrent program as a past adjacent program, while not including programs with the titles "Drama II" or "JJ Weather" as past adjacent programs.

[0043] Hereinafter, the program information of the past adjacent program is referred to as fourth learning information 44. In S13, the second learning information 42 and the fourth learning information 44 are associated with each other. 5 shows, as an example of the fourth learning information 44, program information for one program that was broadcast before the program in the same time slot and three programs that were broadcast after the program in the same time slot. Although a total of four past adjacent programs are shown in FIG. 5, the number of past adjacent programs is not limited to four. For example, the past adjacent programs may be two or more programs that were broadcast before the program in the same time slot, or two or less or four or more programs that were broadcast after the program in the same time slot.

[0044] The past adjacent programs associated with the program in the same time slot may be only programs that were broadcast before the program in the same time slot, or may be only programs that were broadcast after the program in the same time slot. [1-2-1-1. First feature generation process for learning] Next, in S14, the processor 10 executes a first feature value generation process for learning. The first feature value generation process for learning is a process for generating a first feature value vector for learning 51 representing the feature values ​​of the program to be learned. The first feature value generation process for learning executed by the processor 10 will be described with reference to the flowchart of FIG. 6 and FIG. 7.

[0045] First, in S140, the processor 10 performs element decomposition on the first learning information 41 and the third learning information 43. Element decomposition is a process of decomposing the broadcast start date and time, broadcast end date and time, and broadcast station information in the program information into predetermined elements.

[0046] In this embodiment, as shown in FIG. 7, the element decomposition is performed for each program into the elements of "year," "month," "day," "day of the week," "start time," "start minute," "broadcast minutes," and "broadcast station." "Year," "month," "day," "day of the week," "start time," and "start minute" are values ​​corresponding to the broadcast start date and time of the program. "Broadcast minutes" is the value obtained by subtracting the broadcast start date and time from the broadcast end date and time. "Broadcast station" is the broadcast station represented by the program information.

[0047] Next, in S141, the processor 10 performs normalization and embedding on the first learning information 41. Embedding is a process used in fields such as natural language processing and machine learning to convert text data such as words and sentences into numerical vectors that are easy for the artificial intelligence 3 to handle.

[0048] In this embodiment, the processor 10 first concatenates the broadcast start date and time, the broadcast end time, the broadcast station, and the title in a text format. For example, the processor 10 concatenates the broadcast start date and time, the broadcast end time, the broadcast station, and the title of the first learning information 41 in FIG. 7 to generate text data such as "2023 / 9 / 3 19:58-20:54 ZZ TV AA News."

[0049] At this time, the processor 10 normalizes the text data. Normalization is a process of converting text data of different formats into the same format. Normalization is a process that makes it easier for the artificial intelligence 3 to handle the text data. Examples of normalization include standardizing full-width characters and half-width characters to either one, standardizing the type of parentheses, etc.

[0050] The processor 10 then converts the generated text data into a numerical vector using the artificial intelligence 3. The process of converting the text data into a numerical vector is realized by applying an existing natural language processing model to the artificial intelligence 3. An example of an existing natural language processing model is BERT (Bidirectional Encoder Representations from Transformers).

[0051] If the program information contained in two pieces of text data is similar to each other, the two pieces of text data are converted into two similar numerical vectors. Two similar numerical vectors have a larger inner product than two dissimilar numerical vectors.

[0052] Next, at S142, the processor 10 converts the viewing amount in each of the first learning information 41 and the third learning information 43 into a missing value. A missing value is a value that cannot exist as an actual viewing amount. In FIG. 7, as an example, the C-class viewing rate of "3.2" and the T-class viewing rate of "5.5" corresponding to the program to be studied are converted into the missing value "XXX."

[0053] Next, in S143, the processor 10 generates a training first feature vector 51, which is a feature of the training target program. The training first feature vector 51 is generated based on each element of the program information decomposed into elements in S140, the numerical vector obtained in S141, and the missing value obtained in S142. In other words, the training first feature vector 51 is a vector whose components are each element of the program information decomposed into elements in S140, the numerical vector obtained in S141, and the missing value obtained in S142.

[0054] Thereafter, the processor 10 ends the first feature generation process for learning shown in FIG. In this way, the processor 10 generates the first training feature vector 51 based on the first training information 41 and the third training information 43.

[0055] [1-2-1-2. Second feature generation process for learning] Subsequently, in S15, the processor 10 executes a second feature value generation process for learning. The second feature value generation process for learning is a process for generating a second feature value vector for learning 52 representing the feature values ​​of programs in the same time slot.

[0056] The second feature vectors for training 52 are generated for each of the programs broadcast in the same time slot for the past N weeks (N is an integer of 2 or greater) from the broadcast date and time of the program to be trained. In other words, a total of N second feature vectors for training 52 are generated for each of the programs broadcast in the same time slot from one week before the broadcast date and time of the program to be trained to N weeks before the broadcast date and time of the program to be trained. The second feature vectors for training 52 are generated so as to include, as time series information, information indicating the broadcast order of these programs broadcast in the same time slot for the past N weeks.

[0057] The time series information is, for example, a number assigned to each training second feature vector 52 based on the broadcast start date and time of the program. The assigned number may be understood as an array number when defining N training second feature vectors 52 as an array. As an example, the number "1" is assigned as time series information to the training second feature vector 52 of a program that aired in the same time slot one week before the training program, and the number "2" is assigned as time series information to the training second feature vector 52 of a program that aired in the same time slot two weeks before the training program. In this example, a smaller number indicates a relatively later broadcast start date and time (i.e., closer to the present).

[0058] The second feature generation process for learning executed by the processor 10 will be described with reference to the flowchart of FIG. 8 and FIG. Hereinafter, a program that was broadcast in the same time slot i weeks before the broadcast date and time of the program to be learned is referred to as the program in the same time slot i weeks before, where i is an integer between 1 and N. When processor 10 starts the process of generating second features for learning, it sets the value 1 to variable i as an initial value.

[0059] In the following description of the second learning feature generation process, the second learning information 42 and the fourth learning information 44 refer to the second learning information 42 and the fourth learning information 44 corresponding to the program aired in the same time slot i weeks ago.

[0060] First, in S150, the processor 10 performs element decomposition on the second learning information 42 and the fourth learning information 44. The element decomposition is the same process as in S140. By decomposing the second learning information 42 and the fourth learning information 44 into elements, each element of the program information of the programs in the same time slot and the past adjacent programs is obtained.

[0061] Next, in S151, the processor 10 normalizes and embeds the second training information 42. The normalization and embedding are the same processes as in S141. By normalizing and embedding the second training information 42, a numerical vector corresponding to the second training information 42 is obtained.

[0062] Next, in S152, the processor 10 acquires the viewing amount for each of the second learning information 42 and the fourth learning information 44. Next, in S153, the processor 10 generates a second feature vector for training i weeks ago 52, which is a feature of a program aired in the same time slot i weeks ago. The second feature vector for training 52 is generated based on each element of the program information decomposed into elements in S150, the numerical vector obtained in S151, and the viewing amount obtained in S152. In other words, the second feature vector for training 52 is a vector whose components are each element of the program information decomposed into elements in S150, the numerical vector obtained in S151, and the viewing amount obtained in S152.

[0063] Subsequently, in S154, the processor 10 increments the value of the variable i by one. Subsequently, in S155, the processor 10 determines whether the variable i is equal to or less than the value N. If the processor 10 determines that the variable i is equal to or less than the value N (S155: YES), the processor 10 returns to S150.

[0064] On the other hand, if the processor 10 determines that the variable i is greater than the value N (S155: NO), the processor 10 ends the second feature generation process for learning in FIG. In this way, processor 10 generates a total of N second learning feature vectors 52 for each week from one week ago to N weeks ago based on second learning information 42 and fourth learning information 44 corresponding to programs aired in the same time slot for the past N weeks.

[0065] [1-2-1-3. Generate a deep learning model] Next, at S16, the processor 10 generates a deep learning model 31 by training the artificial intelligence 3 through supervised learning based on input information including a first training feature vector 51 and multiple second training feature vectors 52.

[0066] In S16, in the prediction process described below, the second feature vector for learning 52 used varies depending on how many weeks into the future the scheduled broadcast date and time of the program to be predicted is from the prediction point in time. Specifically, when the program to be predicted is one whose scheduled broadcast date and time is M weeks into the future, the processor 10 uses the second feature vector for learning 52 of a program that airs in the same time slot from M weeks before to N weeks before the program to be learned. M is an integer between 1 and N. In this embodiment, the description will be mainly based on M=1.

[0067] First, the processor 10 arranges in chronological order the first training feature vector 51 and the plurality of second training feature vectors 52 generated in S14 and S15. That is, using the time series information included in the second training feature vector 52, the processor 10 arranges the first training feature vector 51, the second training feature vector 52 of the program airing in the same time slot M weeks ago, the second training feature vector 52 of the program airing in the same time slot (M+1) weeks ago, ..., the second training feature vector 52 of the program airing in the same time slot N weeks ago in that order.

[0068] Next, as shown in FIG. 10, the processor 10 inputs the training first feature quantity vector 51 and a plurality of training second feature quantity vectors 52 to the artificial intelligence 3 in chronological order. The artificial intelligence 3 is configured to predict the viewing volume of a training target program that has been converted into missing values, based on the input training first feature vector 51 and a plurality of training second feature vectors 52. The artificial intelligence 3 is configured to calculate the error between the predicted value, which is the result of the prediction, and the training data, to calculate parameters that minimize the error, and generate a deep learning model 31.

[0069] The teacher data used is the viewing amount of the program to be learned in the first learning information 41 before being converted to missing values ​​in S142. In other words, the viewing amount of the program to be learned in the first learning information 41 before being converted to missing values ​​in S142 is not used as a feature amount of the program to be learned.

[0070] As an example of the error, the mean absolute error (MAE) is used. The mean absolute error is calculated by calculating the absolute value of the difference between each of the multiple predicted values ​​and the true value in the training data, and then dividing the sum of the absolute values ​​corresponding to each of the multiple predicted values ​​by the number of data. The closer the predicted value is to the true value, the smaller the mean absolute error.

[0071] Examples of error calculation methods include batch learning, mini-batch learning, and online learning. After that, processor 10 ends the learning process of FIG. In this way, the processor 10 is configured to perform a learning process to generate a deep learning model 31 based on the first learning information 41, the second learning information 42, the third learning information 43, and the fourth learning information 44.

[0072] [1-2-2. Prediction processing] The prediction process executed by the processor 10 will be described with reference to the flowchart of Fig. 11 and Figs. 12 to 19. In the following, a process will be described in which one broadcast-scheduled program is set as a program to be predicted, and the viewing volume of the program to be predicted is predicted.

[0073] The steps of the prediction process are generally equivalent to the steps of the learning process. Specifically, the processes of S20, S21, S22, S23, S24, S25, and S26 shown in Fig. 11 are generally equivalent to the processes of S10, S11, S12, S13, S14, S15, and S16 shown in Fig. 2.

[0074] When a predetermined instruction is input from the user via the user interface 40, the processor 10 starts the prediction process shown in FIG. First, in S20, processor 10 acquires program information of a plurality of programs scheduled to be broadcast and program information of a plurality of past programs from database 2 via communication interface 50. For example, processor 10 acquires program information of programs scheduled to be broadcast up to one week from the present and program information of past programs for the past 100 weeks from the present. The processing of S20 corresponds to the processing of acquisition unit 11.

[0075] The viewing volume in the program information of a program scheduled to be broadcast is an unknown value and may be the viewing volume to be predicted, so a predetermined missing value is input as the viewing volume. Next, in S21, processor 10 selects one program from the multiple programs scheduled to be broadcast as the program to be predicted, and associates past programs that aired in the same time slot as the program to be predicted with the program to be predicted (see FIG. 12). In the prediction process, a program that airs in the same time slot is a past program that aired on the same day of the week, in the same time slot, and by the same broadcasting station as the program to be predicted. In S21, processor 10 associates multiple programs that air in the same time slot with one program to be predicted.

[0076] Hereinafter, in the prediction process, the program information of the program to be predicted is referred to as first prediction information 61. Program information of multiple programs broadcast in the same time slot is referred to as second prediction information 62. In S21, the first prediction information 61 and the second prediction information 62 are associated with each other.

[0077] Here, a program that is broadcast in the same time slot as the program to be predicted means a program that has a broadcast start time that is closest to the broadcast start time of the program to be predicted on each day. If there is no program scheduled to be broadcast on the same day of the week and by the same broadcasting station as the program to be predicted in a certain week whose broadcast start time matches that of the program to be predicted, the program with the broadcast start time closest to the broadcast start time of the program to be predicted is associated as the program to be broadcast in the same time slot in that week.If there are multiple programs with the broadcast start time closest to the broadcast start time of the program to be predicted, any one of the programs is associated as the program to be broadcast in the same time slot in that week.

[0078] For example, if the broadcast start date and time of the program to be predicted is "2023 / 10 / 22 19:58," and two past programs that were broadcast on the same day of the week and by the same broadcast station one week prior to the program to be predicted have broadcast start times of "2023 / 10 / 15 18:58" and "2023 / 10 / 15 20:58," respectively, one of the past programs will be associated as the program in the same time slot. In this case, either program may be associated as the program in the same time slot.

[0079] 12 shows, as an example of the first prediction information 61, the broadcast start date and time of the program to be predicted, "2023 / 10 / 22 19:58," the broadcast end date and time of the program to be predicted, "2023 / 10 / 22 20:54," the broadcast station of the program to be predicted, "ZZ TV," the title of the program to be predicted, "Drama LL," and the target audience rating of the program to be predicted. The target audience ratings are the audience ratings for the above-mentioned C demographic, T demographic, M1-M3 demographic, and F1-F3 demographic. The missing value "XXX" is entered for the target audience rating of the program to be predicted.

[0080] 12 shows, as an example of the second prediction information 62, the broadcast start date and time of the program in the same time slot, the broadcast end date and time of the program in the same time slot, the broadcast station of the program in the same time slot, the title of the program in the same time slot, and the target audience rating for the program in the same time slot for each of a plurality of past weeks, similar to the first prediction information 61. In other words, the second prediction information 62 includes program information for the programs in the same time slot for the past N weeks (N is an integer of 2 or more) from the broadcast date and time of the program to be predicted.

[0081] Next, in S22, the processor 10 associates at least one adjacent program that was broadcast in a time slot adjacent to the broadcast date and time of the program to be predicted with the program to be predicted (see FIG. 13). The time slot adjacent to the broadcast date and time of the program to be predicted includes at least one of the time slots before and after the broadcast date and time of the program to be predicted. The broadcast date and time of the adjacent program need only be included in a time slot adjacent to the broadcast date and time of the program to be predicted, so the broadcast start date and end date and time of the adjacent program do not necessarily have to be adjacent to the broadcast date and time of the program to be predicted. In other words, the broadcast start date and time of the adjacent program may or may not be adjacent to the broadcast end date and time of the program to be predicted. The broadcast end date and time of the adjacent program may or may not be adjacent to the broadcast start date and time of the program to be predicted. For example, in FIG. 13 , for a broadcast end date and time of a program to be predicted, "2023 / 10 / 22 20:54," the time period adjacent to the broadcast date and time of the program to be predicted may be from "2023 / 10 / 22 20:54" to "2023 / 10 / 22 22:00." In this case, the adjacent program may be a program that is broadcast from "2023 / 10 / 22 20:54" to "2023 / 10 / 22 22:00." That is, in FIG. 13 , programs with the titles "Music QQ," "Drama RR," or "SS Weather" may all be associated with the program to be predicted as adjacent programs. For example, processor 10 may associate a program with the title "Drama RR" with the program to be predicted as an adjacent program, while not treating programs with the titles "Music QQ" or "SS Weather" as adjacent programs.

[0082] Hereinafter, the program information of at least one adjacent program is referred to as third information for prediction 63. In S22, the first information for prediction 61 and the third information for prediction 63 are associated with each other. 13 shows, as an example of the third information for prediction 63, program information for one program that will be broadcast before the program to be predicted and three programs that will be broadcast after the program to be predicted. Although a total of four adjacent programs are shown in FIG. 13, the number of adjacent programs is not limited to four. For example, the adjacent programs may be two or more programs that will be broadcast before the program to be predicted, or may be two or less or four or more programs that will be broadcast after the program to be predicted.

[0083] The adjacent programs associated with the program to be predicted may be only programs that are broadcast before the program to be predicted, or may be only programs that are broadcast after the program to be predicted. Next, in S23, processor 10 associates with the same-time-slot program a plurality of past adjacent programs, which are a plurality of programs that were broadcast in time slots adjacent to the broadcast date and time of the same-time-slot program (see FIG. 14). The time slots adjacent to the broadcast date and time of the same-time-slot program include at least one of the time slots before and after the broadcast date and time of the same-time-slot program. The broadcast date and time of the past adjacent program need only be included in a time slot adjacent to the broadcast date and time of the concurrent program, so the broadcast start date and end date and time of the past adjacent program do not necessarily have to be adjacent to the broadcast date and time of the concurrent program. In other words, the broadcast start date and time of the past adjacent program may or may not be adjacent to the broadcast end date and time of the concurrent program. The broadcast end date and time of the past adjacent program may or may not be adjacent to the broadcast start date and time of the concurrent program. For example, in FIG. 14 , for a concurrent program whose broadcast end date and time is "2023 / 10 / 15 21:54," the time slot adjacent to the broadcast date and time of the concurrent program may be from "2023 / 10 / 15 21:54" to "2023 / 10 / 15 23:25." In this case, the past adjacent programs may be programs broadcast between "2023 / 10 / 15 21:54" and "2023 / 10 / 15 23:25." That is, in FIG. 14 , programs with the titles "Music UU," "Information VV," and "Variety WW" may all be associated with the concurrent program as past adjacent programs. For example, processor 10 may associate programs with the titles "Music UU" or "Variety WW" as past adjacent programs with the concurrent program, while not including a program with the title "Information VV" as a past adjacent program.

[0084] Hereinafter, the program information of the past adjacent program is referred to as the fourth prediction information 64. In S23, the second prediction information 62 and the fourth prediction information 64 are associated with each other. 14 shows, as an example of the fourth prediction information 64, program information for one program that was broadcast before the program in the same time slot and three programs that were broadcast after the program in the same time slot. Although a total of four past adjacent programs are shown in FIG. 14, the number of past adjacent programs is not limited to four. For example, the past adjacent programs may be two or more programs that were broadcast before the program in the same time slot, or two or less or four or more programs that were broadcast after the program in the same time slot.

[0085] The past adjacent programs associated with the program in the same time slot may be only programs that were broadcast before the program in the same time slot, or may be only programs that were broadcast after the program in the same time slot. [1-2-2-1. First feature generation process for prediction] Subsequently, in S24, the processor 10 executes a first feature quantity for prediction generation process. The first feature quantity for prediction generation process is a process for generating a first feature quantity for prediction vector 71 representing the feature quantities of the program to be predicted. The process of S24 corresponds to the process performed by the generation unit 12.

[0086] The process of generating the first feature for prediction executed by the processor 10 will be described with reference to the flowchart of FIG. 15 and FIG. First, in S240, the processor 10 performs element decomposition on the first prediction information 61 and the third prediction information 63. The element decomposition is the same process as in S140 and S150. In this embodiment, as shown in FIG. 16, the element decomposition is performed for each program into the elements of "year," "month," "day," "day of the week," "start time," "start minute," "broadcast minutes," and "broadcast station." "Year," "month," "day," "day of the week," "start time," and "start minute" are values ​​corresponding to the broadcast start date and time of the program. "Broadcast minutes" is the value obtained by subtracting the broadcast start date and time from the broadcast end date and time. "Broadcast station" is the broadcast station represented by the program information.

[0087] Next, in S241, the processor 10 performs normalization and embedding on the first prediction information 61. In this embodiment, the processor 10 first concatenates the broadcast start date and time, the broadcast end time, the broadcast station, and the title in a text format. For example, the processor 10 concatenates the broadcast start date and time, the broadcast end time, the broadcast station, and the title of the first prediction information 61 in FIG. 16 to generate text data such as "2023 / 10 / 22 19:58-20:54 ZZ TV Drama LL."

[0088] At this time, the processor 10 normalizes the text data. The normalization is the same process as in S141 and S151. Thereafter, the processor 10 performs embedding on the generated text data to convert it into a numerical vector using the artificial intelligence 3. The embedding is the same process as in S141 and S151.

[0089] Subsequently, in S242, the processor 10 acquires the viewing amount (that is, the missing value "XXX") in each of the first prediction information 61 and the third prediction information 63. Next, in S243, the processor 10 generates a first feature vector for prediction 71, which is a feature of the program to be predicted. The first feature vector for prediction 71 is generated based on each element of the program information decomposed into elements in S240, the numerical vector obtained in S241, and the missing value obtained in S242. In other words, the first feature vector for prediction 71 is a vector whose components are each element of the program information decomposed into elements in S240, the numerical vector obtained in S241, and the missing value obtained in S242.

[0090] Thereafter, the processor 10 ends the process of generating the first feature for prediction in FIG. In this way, the processor 10 generates the first feature quantity vector for prediction 71 based on the first information for prediction 61 and the third information for prediction 63.

[0091] [1-2-2-2. Generation of second feature values ​​for prediction] Subsequently, in S25, the processor 10 executes a process of generating a second feature for prediction. The process of generating a second feature for prediction vector 72 representing the feature of the program in the same time slot. The process of S25 corresponds to the process performed by the generation unit 12.

[0092] The second feature vector for prediction 72 is generated for each of the programs that air in the same time slot for the past N weeks (N is an integer of 2 or greater) from the broadcast date and time of the program to be predicted. In other words, a total of N second feature vectors for prediction 72 are generated for each of the programs that air in the same time slot from one week before the broadcast date and time of the program to be predicted to N weeks before. The second feature vector for prediction 72 is generated so as to include, as time series information, information indicating the broadcast order of these programs that air in the same time slot for the past N weeks.

[0093] The time series information is, for example, a number assigned to each prediction second feature vector 72 based on the broadcast start date and time of the program. As an example, the number "1" is assigned as time series information to the prediction second feature vector 72 of a program that airs in the same time slot one week before the program to be predicted, and the number "2" is assigned as time series information to the prediction second feature vector 72 of a program that airs in the same time slot two weeks before the program to be predicted. In this example, a smaller number indicates a relatively later broadcast start date and time (i.e., closer to the present).

[0094] The process of generating second features for prediction executed by the processor 10 will be described with reference to the flowchart of FIG. 17 and FIG. Hereinafter, a program that was broadcast in the same time slot i weeks before the broadcast date and time of the program to be predicted is referred to as the program in the same time slot i weeks before, where i is an integer between 1 and N. When starting the process of generating second features for prediction, processor 10 sets the value 1 to variable i as an initial value.

[0095] In the following description of the process of generating the second feature for prediction, the second information for prediction 62 and the fourth information for prediction 64 refer to the second information for prediction 62 and the fourth information for prediction 64 corresponding to the program aired in the same time slot i weeks ago.

[0096] First, in S250, the processor 10 performs element decomposition on the second prediction information 62 and the fourth prediction information 64. The element decomposition is the same process as in S140, S150, and S240. By decomposing the second prediction information 62 and the fourth prediction information 64 into elements, each element of the program information of the concurrent time slot program and the past adjacent program is obtained.

[0097] Subsequently, in S251, the processor 10 normalizes and embeds the second information for prediction 62. The normalization and embedding are the same processes as those in S141, S151, and S241. By normalizing and embedding the second information for prediction 62, a numeric vector corresponding to the second information for prediction 62 is obtained.

[0098] Subsequently, in S252, the processor 10 acquires the viewing amount in each of the second prediction information 62 and the fourth prediction information 64. Next, in S253, the processor 10 generates a second feature vector for prediction 72, which is a feature of a program aired in the same time slot i weeks ago. The second feature vector for prediction 72 is generated based on each element of the program information decomposed into elements in S250, the numerical vector obtained in S251, and the viewing amount obtained in S252. In other words, the second feature vector for prediction 72 is a vector whose components are each element of the program information decomposed into elements in S250, the numerical vector obtained in S251, and the viewing amount obtained in S252.

[0099] Subsequently, in S254, the processor 10 increments the value of the variable i by one. Subsequently, in S255, the processor 10 determines whether the variable i is equal to or less than the value N. If the processor 10 determines that the variable i is equal to or less than the value N (S255: YES), the processor 10 returns to S250.

[0100] On the other hand, if the processor 10 determines that the variable i is greater than the value N (S255: NO), the processor 10 ends the process of generating second feature amounts for prediction in FIG. In this way, the processor 10 generates a total of N second feature vectors 72 for prediction, one for each week from one week ago to N weeks ago, based on the second information 62 for prediction and the fourth information 64 for prediction corresponding to programs airing in the same time slot for the past N weeks.

[0101] [1-2-2-3. Predicting the viewing volume of the target program] Next, in S26, the processor 10 inputs input information including the first feature vector for prediction 71 and the second feature vector for prediction 72 to the deep learning model 31, and obtains a predicted value of the viewing amount of the program to be predicted as an output of the deep learning model 31, thereby predicting the viewing amount of the program to be predicted. The processing of S26 corresponds to the processing of the prediction unit 13.

[0102] In S26, the second feature vector for prediction 72 used varies depending on how many weeks from now the scheduled broadcast date and time of the program to be predicted is. Specifically, if the scheduled broadcast date and time of the program to be predicted is M weeks from now, the processor 10 uses the second feature vector for prediction 72 of programs in the same time slot from M weeks ago to N weeks ago based on the program to be predicted (M is an integer between 1 and N). In this embodiment, the description will mainly assume M=1.

[0103] First, the processor 10 arranges in chronological order the first prediction feature vector 71 and the multiple second prediction feature vectors 72 generated in S24 and S25. That is, using the time series information included in the second prediction feature vector 72, the processor 10 arranges the first prediction feature vector 71, the second prediction feature vector 72 of the program airing in the same time slot M weeks ago, the second prediction feature vector 72 of the program airing in the same time slot (M+1) weeks ago, ..., the second prediction feature vector 72 of the program airing in the same time slot N weeks ago in that order.

[0104] Next, as shown in FIG. 19, the processor 10 inputs the first feature amount vector for prediction 71 and the plurality of second feature amount vectors for prediction 72 to the deep learning model 31 in chronological order. The deep learning model 31 is configured to predict the viewing amount of the program to be predicted, which is a missing value, based on the input first feature amount vector for prediction 71 and a plurality of second feature amount vectors for prediction 72.

[0105] Thereafter, the processor 10 ends the prediction process of FIG. In this way, the processor 10 is configured to predict the viewing volume of the program to be predicted by inputting the first prediction information 61, the second prediction information 62, the third prediction information 63, and the fourth prediction information 64 as input information to the deep learning model 31 and obtaining a predicted value of the viewing volume of the program to be predicted as the output of the deep learning model 31.

[0106] [1-3.Effects] According to the embodiment described above, the following actions and effects can be obtained. (1a) Generally, the viewing volume of a program may be affected by adjacent programs broadcast in adjacent time slots.

[0107] For example, for popular programs with relatively high viewing volumes, viewers tend to start watching shortly before the popular program begins and continue watching the same broadcasting station even after the popular program has finished. For this reason, the viewing volumes of programs broadcast before and after a popular program are easily influenced by the popular program. In particular, mini-programs with relatively short broadcast minutes (for example, about 5 or 10 minutes) are more likely to be influenced by the popular program.

[0108] In the above embodiment, the first feature vector for learning 51 is generated based on the program information of the program to be learned and the program information of the adjacent program corresponding to the program to be learned. The second feature vector for learning 52 is generated based on the program information of the program in the same time slot corresponding to the program to be learned and the program information of the past adjacent program corresponding to the program to be learned.

[0109] The first feature vector for prediction 71 is generated based on the program information of the program to be predicted and the program information of the adjacent program corresponding to the program to be predicted. The second feature vector for prediction 72 is generated based on the program information of the program in the same time slot corresponding to the program to be predicted and the program information of the past adjacent program corresponding to the program to be predicted.

[0110] Therefore, by using the program information of adjacent programs and the program information of past adjacent programs to predict the viewing volume in the learning process and prediction process, it is possible to improve the accuracy of predicting the viewing volume of a program.

[0111] (1b) In a certain week, if there is no program whose broadcast start time matches that of the learning program or the prediction program, the program whose broadcast start time is closest to the learning program or the prediction program is associated as the program broadcast in the same time slot as the learning program or the prediction program in that week. If there are multiple programs whose broadcast start time is closest to the learning program or the prediction program, one of the programs is associated as the program broadcast in the same time slot in that week.

[0112] According to this process, even if the broadcast start time of a program in the same time slot does not exactly match the broadcast start time of the program to be learned or predicted in the learning process and prediction process, it is possible to use the program information of the program in the same time slot whose broadcast start time is closest to the broadcast start time of the program to be learned or predicted in the prediction process, thereby improving the accuracy of the prediction of program viewing volume.

[0113] (1c) The first training feature vector 51, the second training feature vector 52, the first prediction feature vector 71, and the second prediction feature vector 72 are generated based on the broadcast date and time and broadcast station of the program information. The deep learning model 31 is generated based on the first training feature vector 51 and the second training feature vector 52. The deep learning model 31 predicts the viewing volume of the program to be predicted based on the first prediction feature vector 71 and the second prediction feature vector 72.

[0114] According to this type of processing, in the learning process and prediction process, it is possible to generate a deep learning model 31 or predict the viewing volume of the program to be predicted based on the broadcast date and time and broadcast station of the program information.

[0115] For example, it is possible to relatively increase the influence on the predicted viewing volume of past programs that were broadcast on the same day of the week, in the same time slot, and by the same broadcasting station as the program to be learned or predicted, thereby improving the accuracy of the prediction of the viewing volume of the program.

[0116] (1d) The first feature quantity vector for training 51, the second feature quantity vector for training 52, the first feature quantity vector for prediction 71, and the second feature quantity vector for prediction 72 are generated based on the title of the program information.

[0117] According to this process, in the learning process and prediction process, it is possible to relatively reduce the influence on the prediction of the viewing amount of a program having a title that is less relevant to the learning target program or the prediction target program.

[0118] In addition, it is possible to determine, based on the title of the program information, whether a past program is a special program (a so-called special number), which is a one-off program that is broadcast irregularly, unlike regular programs that are broadcast periodically. In general, the viewing volume of special programs tends to be higher than the viewing volume of regular programs. Therefore, by considering whether a past program is a special program, it is possible to improve the accuracy of predicting the viewing volume of a program.

[0119] (1e) In the learning process, when the program to be predicted is one whose scheduled broadcast date and time is M weeks from now, the processor 10 uses the second learning feature vectors 52 of programs in the same time slot from M weeks ago to N weeks ago (M is an integer between 1 and N).

[0120] In the prediction process, if the target of prediction is a program whose scheduled broadcast date and time is M weeks in the future, feature amounts from 1 week before to (M-1) weeks before the program scheduled to be broadcast cannot be used. This type of learning process makes it possible to train the artificial intelligence 3 under the same conditions as when making predictions. This improves the accuracy of predicting the program's viewing volume.

[0121] (1f) The deep learning model 31 is configured to be able to calculate a predicted value of the viewership volume using time-series information included in the input information. For example, the deep learning model 31 is configured to perform calculations such that the closer the broadcast date and time of a past program is to the present, the greater the influence of the program information of that program on the predicted viewership volume. The second learning feature vector 52 included in the input information is generated so as to include, as time-series information, information indicating the broadcast order of programs broadcast in the same time slot for the past N weeks.

[0122] By using such a deep learning model 31, the accuracy of predicting the viewing volume of a program can be improved by including a second learning feature vector 52 containing time series information in the input of the deep learning model 31.

[0123] [1-4. Correspondence between terms] In the above embodiment, the first prediction information 61 corresponds to an example of the first information, the second prediction information 62 corresponds to an example of the second information, the third prediction information 63 corresponds to an example of the third information, and the fourth prediction information 64 corresponds to an example of the fourth information. The first prediction feature vector 71 corresponds to an example of the first feature, and the second prediction feature vector 72 corresponds to an example of the second feature. The target audience rating corresponds to an example of the viewership.

[0124] 2. Other Embodiments Although the embodiments of the present disclosure have been described above, it goes without saying that the present disclosure is not limited to the above-described embodiments and can take on various forms.

[0125] (2a) In the above embodiment, the first training feature vector 51 is generated based on the broadcast start date / time, broadcast end time, broadcast station, title, and missing values ​​contained in the first training information 41 and the third training information 43. The second training feature vector 52 is generated based on the broadcast start date / time, broadcast end time, broadcast station, title, and viewing volume contained in the second training information 42 and the fourth training information 44.

[0126] However, the first training feature vector 51 does not need to be generated based on all of this information, but only needs to be generated based on at least the broadcast date and time.The second training feature vector 52 does not need to be generated based on all of this information, but only needs to be generated based on at least the broadcast date and time and the viewing volume.

[0127] (2b) In the above embodiment, the program information includes information indicating the broadcast start date and time of the program, the broadcast end date and time, the title, the broadcast minutes, and the broadcast station that will broadcast the program. However, the program information does not need to include all of this information, but only needs to include at least the broadcast date and time of the program. The program information of past programs only needs to further include the viewing volume of the program.

[0128] The program information may include, as the broadcast date and time of the program, information that can identify the broadcast start date and time of the program. (2c) In the above embodiment, the programs in the same time slot used in the learning process are past programs that were broadcast on the same day of the week, in the same time slot, and by the same broadcasting station as the program to be learned. The programs in the same time slot used in the prediction process are past programs that were broadcast on the same day of the week, in the same time slot, and by the same broadcasting station as the program to be predicted. However, these same-time programs may be any past programs that were broadcast in a time slot corresponding to the scheduled broadcast date and time of the program to be studied or predicted. In other words, the same-time programs may be past programs that were broadcast in at least the same time slot as the program to be studied or predicted, and may be past programs that were broadcast on a different day of the week or by a different broadcasting station than the program to be studied or predicted. (2d) In the above embodiment, the viewing volume represents a viewing rate or a parameter related to the viewing rate. However, the viewing volume is not limited to a viewing rate or a parameter related to the viewing rate. For example, the viewing volume may be the number of views reached to a specific viewer or other number of views. The viewing volume may also be the percentage of a specific viewer attribute relative to the number of viewers.

[0129] (2e) Multiple functions possessed by one component in the above embodiments may be realized by multiple components, or one function possessed by one component may be realized by multiple components. Also, multiple functions possessed by multiple components may be realized by one component, or one function realized by multiple components may be realized by one component. Also, part of the configuration of the above embodiments may be omitted. Also, at least part of the configuration of the above embodiments may be added to or substituted for the configuration of another of the above embodiments.

[0130] (2f) The present disclosure may be realized in various forms other than the prediction system described above, such as a system including the prediction system as a component, a computer program for causing a computer to function as the prediction system, a non-transitory tangible recording medium such as a semiconductor memory on which the computer program is recorded, a prediction method, etc.

[0131] [Technical idea disclosed in this specification] [Item 1] 1. A prediction system comprising: an acquisition unit; a prediction unit; Equipped with The acquisition unit First information including a scheduled broadcast date and time of a program to be predicted; Second information including broadcast dates and times and viewing amounts of a plurality of programs that were broadcast in the same time slot, which are a plurality of programs that were broadcast in the past in a time slot corresponding to the scheduled broadcast date and time of the program to be predicted; Third information including a scheduled broadcast date and time of at least one adjacent program that will be broadcast in a time slot adjacent to the scheduled broadcast date and time of the program to be predicted; Fourth information including broadcast dates and times and viewing amounts of a plurality of past adjacent programs, which are a plurality of programs that were broadcast in time slots adjacent to each of the broadcast dates and times of the plurality of same-time slot programs; configured to obtain The prediction unit a scheduled broadcast date and time of the program to be predicted, which is identified from the first information; the broadcast dates and times and viewing amounts of the plurality of programs in the same time slot identified from the second information; a scheduled broadcast date and time of the at least one adjacent program identified from the third information; and the broadcast dates and times and viewing amounts of the plurality of past adjacent programs identified from the fourth information; to a deep learning model, and obtain a predicted value of the viewing volume of the program to be predicted as an output of the deep learning model, thereby predicting the viewing volume of the program to be predicted. Prediction system.

[0132] [Item 2] The prediction system according to item 1, the first information includes information that can identify a scheduled broadcast start date and time of the program to be predicted as a scheduled broadcast date and time of the program to be predicted, the second information includes, as broadcast dates and times of the plurality of simultaneous time slot programs, information capable of identifying broadcast start dates and times of the plurality of simultaneous time slot programs; The plurality of programs in the same time slot are programs that have been broadcast in the past on the same day of the week as the program to be predicted and by the same broadcasting station, and have broadcast start dates and times that are closest to the broadcast start date and time of the program to be predicted in each of the past plurality of weeks. Prediction system.

[0133] [Item 3] The prediction system according to item 1 or 2, a generating unit configured to generate a first feature amount that is a feature amount of the prediction target program and a plurality of second feature amounts that are feature amounts of each of the plurality of concurrent programs, the first feature amount is generated based on a scheduled broadcast date and time, identification information of a broadcasting station, and a title of each of the prediction target program and the at least one adjacent program; the plurality of second feature amounts are generated based on broadcast dates and times, identification information of broadcasting stations, and titles of the plurality of concurrent programs and the plurality of past adjacent programs, the prediction unit is configured to input information including the first feature amount and the plurality of second feature amounts to the deep learning model as the input information. Prediction system.

[0134] [Item 4] The prediction system according to any one of items 1 to 3, The deep learning model is configured to calculate the predicted value using time series information included in the input information. Prediction system.

[0135] [Item 5] Item 3. The prediction system according to item 3, The deep learning model is configured to calculate the predicted value using time-series information included in the input information; the generation unit is configured to generate the plurality of second features so as to include information representing a broadcast order of the plurality of programs broadcast in the same time slot as the time series information. Prediction system.

[0136] [Item 6] 1. A computer-implemented prediction method comprising: To obtain and To predict and Including, The obtaining includes: First information including a scheduled broadcast date and time of a program to be predicted; Second information including broadcast dates and times and viewing amounts of a plurality of programs that were broadcast in the same time slot, which are a plurality of programs that were broadcast in the past in a time slot corresponding to the scheduled broadcast date and time of the program to be predicted; Third information including a scheduled broadcast date and time of at least one adjacent program that will be broadcast in a time slot adjacent to the scheduled broadcast date and time of the program to be predicted; Fourth information including broadcast dates and times and viewing amounts of a plurality of past adjacent programs, which are a plurality of programs that were broadcast in time slots adjacent to each of the broadcast dates and times of the plurality of same-time slot programs; obtaining the The predicting step comprises: a scheduled broadcast date and time of the program to be predicted, which is identified from the first information; the broadcast dates and times and viewing amounts of the plurality of programs in the same time slot identified from the second information; a scheduled broadcast date and time of the at least one adjacent program identified from the third information; and the broadcast dates and times and viewing amounts of the plurality of past adjacent programs identified from the fourth information; inputting input information including the above into a deep learning model, and obtaining a predicted value of the viewing volume of the program to be predicted as an output of the deep learning model, thereby predicting the viewing volume of the program to be predicted. Forecasting methods.

[0137] [Item 7] Item 7. A computer program for causing a computer to execute the prediction method according to Item 6. [Explanation of symbols]

[0138] 1...Prediction device, 3...Artificial intelligence, 31...Deep learning model, 10...Processor, 61...First information for prediction, 62...Second information for prediction, 63...Third information for prediction, 64...Fourth information for prediction, 71...First feature vector for prediction, 72...Second feature vector for prediction, 100...Prediction system.

Claims

1. 1. A prediction system, comprising: an acquisition unit; a prediction unit; Equipped with The acquisition unit First information including a scheduled broadcast date and time of a program to be predicted; second information including broadcast dates and times and viewing amounts of a plurality of programs that were broadcast in a time slot corresponding to the scheduled broadcast date and time of the program to be predicted; third information including a scheduled broadcast date and time of at least one adjacent program that will be broadcast in a time slot adjacent to the scheduled broadcast date and time of the program to be predicted; fourth information including broadcast dates and times and viewing amounts of a plurality of past adjacent programs, which are a plurality of programs that were broadcast in time slots adjacent to each of the broadcast dates and times of the plurality of same-time slot programs; configured to obtain The prediction unit a scheduled broadcast date and time of the program to be predicted, which is identified from the first information; the broadcast dates and times and viewing amounts of the plurality of programs in the same time slot identified from the second information; a scheduled broadcast date and time of the at least one adjacent program identified from the third information; and the broadcast dates and times and viewing amounts of the plurality of past adjacent programs identified from the fourth information; to a deep learning model, and obtain a predicted value of the viewing volume of the program to be predicted as an output of the deep learning model, thereby predicting the viewing volume of the program to be predicted. Prediction system.

2. The prediction system of claim 1, the first information includes, as a scheduled broadcast date and time of the program to be predicted, information capable of identifying a scheduled broadcast start date and time of the program to be predicted; the second information includes, as broadcast dates and times of the plurality of simultaneous time slot programs, information capable of identifying broadcast start dates and times of the plurality of simultaneous time slot programs; The plurality of programs in the same time slot are programs that have been broadcast in the past on the same day of the week as the program to be predicted and by the same broadcasting station, and have broadcast start dates and times that are closest to the broadcast start date and time of the program to be predicted in each of the past plurality of weeks. Prediction system.

3. The prediction system according to claim 1 or 2, a generating unit configured to generate a first feature value that is a feature value of the prediction target program and a plurality of second feature values ​​that are feature values ​​of each of the plurality of same-time slot programs, the first feature amount is generated based on a scheduled broadcast date and time, identification information of a broadcasting station, and a title of each of the prediction target program and the at least one adjacent program; the plurality of second feature amounts are generated based on broadcast dates and times, identification information of broadcasting stations, and titles of the plurality of concurrent programs and the plurality of past adjacent programs, the prediction unit is configured to input information including the first feature amount and the plurality of second feature amounts to the deep learning model as the input information; Prediction system.

4. The prediction system according to claim 3, The deep learning model is configured to calculate the predicted value using time series information included in the input information. Prediction system.

5. 5. The prediction system of claim 4, the generation unit is configured to generate the plurality of second feature amounts so that the time series information includes information indicating a broadcast order of the plurality of programs broadcast in the same time slot. Prediction system.

6. 1. A computer-implemented prediction method comprising: To obtain and To predict and Including, The obtaining includes: First information including a scheduled broadcast date and time of a program to be predicted; second information including broadcast dates and times and viewing amounts of a plurality of programs that were broadcast in a time slot corresponding to the scheduled broadcast date and time of the program to be predicted; third information including a scheduled broadcast date and time of at least one adjacent program that will be broadcast in a time slot adjacent to the scheduled broadcast date and time of the program to be predicted; fourth information including broadcast dates and times and viewing amounts of a plurality of past adjacent programs, which are a plurality of programs that were broadcast in time slots adjacent to each of the broadcast dates and times of the plurality of same-time slot programs; obtaining the The predicting step comprises: a scheduled broadcast date and time of the program to be predicted, which is identified from the first information; the broadcast dates and times and viewing amounts of the plurality of programs in the same time slot identified from the second information; a scheduled broadcast date and time of the at least one adjacent program identified from the third information; and the broadcast dates and times and viewing amounts of the plurality of past adjacent programs identified from the fourth information; inputting input information including the above into a deep learning model, and obtaining a predicted value of the viewing volume of the program to be predicted as an output of the deep learning model, thereby predicting the viewing volume of the program to be predicted. Forecasting methods.

7. A computer program for causing a computer to execute the prediction method according to claim 6.

Images