KR100839306B1 - System and method for recommendation of digital television program using reflection techniques of program veiwing pattern - Google Patents

Abstract

A system and a method for recommending a digital television program using reflection techniques of a program viewing pattern are provided to calculate the program preference of a user and reflect viewing pattern weight to a program which the user has viewed, thereby improving the program recommendation efficiency of personalized electronic program guide. A viewing history storage unit(102) stores a viewing history. A user model storage unit(111) stores user model information. A user preference updating module(110) calculates and updates a preference value by program property. A viewing pattern weight calculating module(120) applies a viewing pattern weight to each televising program and accumulates the viewing pattern weights of the each televising program. A program preference calculating module(130) extracts a property value by program property from the user model information, applies the accumulated viewing pattern weight of a corresponding program to a value that the extracted property values are integrated, and calculates the program preference of the user. A program recommend module(140) recommends a program to the user in an order that program preference is increased.

Description

System and Method for recommendation of digital television program using reflection techniques of program veiwing pattern}

The following drawings attached to this specification are illustrative of preferred embodiments of the present invention, and together with the detailed description of the invention to serve to further understand the technical spirit of the present invention, the present invention is a matter described in such drawings It should not be construed as limited to

1 is a block diagram of a digital TV program recommendation system using a program viewing pattern reflection technique according to the present invention;

2 is a flowchart illustrating a method of updating a preference value for each program attribute of user model information in the present invention;

3 is a flow chart for a method of accumulating the viewing pattern weight of each broadcasting session in the present invention;

4 is a flowchart sequentially illustrating a process of making a program recommendation to a user by calculating a user's program preference according to the present invention;

5 is a diagram illustrating an example of updating user model information;

6 is a diagram illustrating an example of calculating a viewing pattern weight;

7 is a graph showing the change of Recall according to the change of k value in the program preference equation;

8 is a graph showing the results of evaluating the recommended efficiency of the five experiments A, B, C, D, E in terms of precision and recall;

9 is a graph showing the results of evaluating the recommended efficiency of the five experiments A, B, C, D, E in terms of precision,

10 is a graph showing the results of evaluating the recommended efficiency of the attribute addition A experiment, the attribute addition B experiment, the attribute multiplication C experiment, and the attribute multiplication D experiment in terms of Recall.

11 is a graph showing the results of applying the viewing pattern weight to the result of evaluating the recommended efficiency of the attribute addition B experiment, the attribute multiplication C experiment, and the attribute multiplication D experiment in terms of Recall,

12 is a graph illustrating a result of evaluating a result of applying viewing pattern weights to the results of Experiment E and the results of Experiment E in terms of Recall;

13 is a graph showing a ranking in which a user selects a program among recommendation lists provided to the user;

14 is a graph showing the change of the Recall value according to the reflection ratio α value.

<Explanation of symbols for main parts of the drawings>

100: watch history collection module 101: watch history storage module

102: watch history store 110: user preference update module

111: User Model Repository 120: Viewing Pattern Weight Calculation Module

130: program preference calculation module 131: program information store

140: program recommendation module 150: recommended program output module

The present invention relates to a system for recommending a program of a digital TV, and more particularly, to a system and a method for recommending a program list personalized by a user's program viewing history and a user's viewing pattern weight.

As digital broadcasting has become more popular and the spread of digital TV has become more common, the number of channels and programs that users can select has increased significantly compared to the existing TV environment. Now you can watch the programs you've done.

However, while the channels and programs that have been multiplied by the number of times can watch more diverse and rich programs, there is a problem that it takes more effort and time, such as searching the broadcast schedule to find a program of interest.

Accordingly, as the problem of program selection emerges, DTV (Digital TV) program recommendation techniques that recommend programs suitable for the user's taste and interest through the pEPG (persoalized Electronic Program Guide) have been actively studied.

However, most of the DTV program recommendation techniques currently studied are considering only the preferences of program attributes such as channel, genre, sub-genre, and person, so that only programs that have long-term preferences are recommended. There is a problem that the recommendation for a program that you have is not made.

That is, most of the current DTV program recommendation techniques have a limitation that if the user's viewing pattern is changed, the program recommendation considering the current viewing pattern of the user is not made. In order to solve these problems, it is essential to reflect the program viewing pattern of the user. to be.

The present invention was devised to solve the above problems, and after obtaining a program preference using a user model composed of attributes such as genre, detail genre, channel, and the like, by reflecting a viewing pattern weight for a program having a past viewing history The purpose of the present invention is to provide a digital TV program recommendation system and method using a program viewing pattern reflection technique that can improve the program recommendation efficiency of pEPG.

Objects and advantages according to the invention will be described below and will be appreciated by the embodiments of the invention. In addition, the objects and advantages of the present invention can be realized by the configuration and combination of configurations shown in the claims.

In order to achieve the above object, a digital TV program recommendation system using a program viewing pattern reflection technique according to the present invention includes a user ID, a viewing time ID of a viewing program, a viewing start time and end time, and viewing time information. A viewing history store in which history is stored; A user model repository for storing user model information including a user's relative weight for program channels, genres, and sub-genres; A user preference updating module for querying user viewing history from the viewing history store, normalizing the frequency of the remaining attributes by the program attributes with the highest frequency, and calculating and updating a preference value for each program attribute; Calculate the viewing pattern weight that accumulates the viewing pattern weight of each broadcasting session after assigning the viewing pattern weight to the viewing point more recently with respect to each broadcasting session aired during the airing period of the program by referring to the viewing history of the user for each program. module; The attribute value is extracted for each program attribute from the user model information queried in the user model repository by referring to the attribute information for each program, and the value of the corresponding program is accumulated to a value obtained by integrating the extracted attribute values into one value by mathematical operation. A program preference calculation module configured to calculate a program preference of a user by applying a viewing pattern weight; And a program recommendation module for recommending a program to a user in order of high program preference. The user preference updating module may include a property for each program attribute based on an average value of a newly normalized program attribute value and an existing attribute value of user model information. It is characterized by updating the value.

delete

The system according to the present invention may further comprise a user model repository containing user model information including the user's relative weights for program channels, genres and sub-genres. In this case, the user preference updating module may be configured to update the attribute value for each program attribute by the average value of the newly normalized program attribute value and the existing attribute value of the user model information.

Preferably, the user model information further includes user profile information including the user's ID, name, age, gender and education. The initial value of the user's relative weight with respect to the program channel, genre, and sub-genre may be directly input by the user.

Preferably, the viewing pattern weight calculation module may be configured to calculate the viewing pattern weight for each broadcasting session of the program by using the following equation, and accumulate the viewing pattern weight calculated for each broadcasting session.

Equation

(w ' _{u, p} is the weight of the viewing pattern for the program p of the user u,

w _{u, pi} are the viewing pattern weights for the broadcasting session i during the corresponding period of the program p of the user u,

p _i is the program for the show i during that period,

n (p) is the number of episodes of the program p during that period,

α is a constant representing a reflection ratio between past and current viewing history,

history (u) is the watch history for the user u's time period)

The system according to the present invention includes program broadcasting schedule information including program title, PID (Program ID), SID (Sequence ID), class, broadcasting start time, and ending time and broadcasting time information, and channel and genre of the program. And a program information store in which program attribute information including detailed genres is stored. In this case, the program preference calculation module is preferably configured to calculate the user's program preference by the following equation.

Equation

(w ' _{u , p} is the accumulated viewing pattern weight for user u's program p,

k is a positive number greater than zero,

Preference (u, p) is user u's preference for program p,

preference ^u _{gerne (p)} is the preference for the genre of user u's program p,

preference ^u _{channel (p)} is user u's preference for channel of program p,

preference ^u _{subGenre (p)} is the preference for the sub genre of user u's program p)

Preferably, the program preference calculation module may be configured to receive a program recommendation time zone from a user and calculate a user's program preference for a program of each channel broadcasted in the corresponding time zone. The program recommendation module may generate and output a predetermined number of preference high-order program lists.

The system according to the present invention may be further configured to further include a recommendation program output module for receiving the high preference program list to output on the digital TV screen for outputting the above preference higher program list to the digital TV.

Preferably, the user preference updating module queries the viewing history of the user on a weekly basis to update the preference value for each program attribute of the user, and the viewing pattern weight calculation module refers to the viewing history of the user on a weekly basis. It may be configured to accumulate the viewing pattern weight for each program of. The system according to the present invention comprises: a viewing history collection module for monitoring a user's digital TV remote control operation to collect a program viewing history of the user; And a history storage module for receiving the viewing history information from the viewing history collecting module and storing the viewing history information in the viewing history storage.

According to an aspect of the present invention, the program viewing history may be configured to include a user ID, a program ID, reflecting turn information, a viewing start time and an end time, and viewing time information.

According to another aspect of the present invention, the program viewing history may be configured to include a user ID, a program channel, a watching start time and an ending time. In this case, the history storage module may refer to a program information storage including program broadcasting schedule information for each channel, and the program viewing history including the user ID, the program ID, the reflecting turn information, the viewing start time and the end time, and the viewing time information. Can be configured to create and store it in the viewing history repository.

Digital TV program recommendation method using the program viewing pattern reflecting technique according to another aspect of the present invention for achieving the above technical problem, (a) user ID, broadcast time ID of the viewing program, viewing start time and end time and viewing time Providing a viewing history repository in which a viewing history including information is stored, and a user model repository in which user model information including relative weights of users for program channels, genres, and sub-genres is stored; (b) querying a user viewing history from the viewing history store, normalizing the frequency of the remaining attributes by the program attributes with the highest frequency, and calculating and updating a preference value for each program attribute; (c) accumulating the viewing pattern weight of each broadcasting session after assigning the viewing pattern weight to each viewing session that is viewed during the program's airing period to increase as the viewing point becomes more recent with reference to the viewing history of the user for each program. ; (d) Extracting attribute values for each program attribute from the user model information queried in the user model repository by referring to the attribute information for each program, and corresponding program to a value obtained by integrating the extracted attribute values into one value by mathematical operation. Calculating a program preference of the user by applying the accumulated viewing pattern weight of? And (e) recommending the program to the user in order of high program preferences, wherein step (b) is performed for each program attribute by the average value of the newly normalized program attribute value and the existing attribute value of the user model information. And updating the attribute value.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms or words used in the specification and claims should not be construed as having a conventional or dictionary meaning, and the inventors should properly explain the concept of terms in order to best explain their own invention. Based on the principle that can be defined, it should be interpreted as meaning and concept corresponding to the technical idea of the present invention.

Therefore, the embodiments described in the specification and the drawings shown in the drawings are only the most preferred embodiment of the present invention and do not represent all of the technical idea of the present invention, various modifications that can be replaced at the time of the present application It should be understood that there may be equivalents and variations.

1 is a block diagram showing an embodiment of a digital TV program recommendation system using a program viewing pattern reflection technique according to the present invention.

Referring to the drawings, the digital TV program recommendation system using the program viewing pattern reflecting technique of the present invention is the viewing history collection module 100, viewing history storage module 101, viewing history storage 102, user preference update module 110 ), Including the user model store 111, the viewing pattern weight calculation module 120, the program preference calculation module 130, the program information store 131, the program recommendation module 140, and the recommendation program output module 150. It is composed.

The viewing history collecting module 100 monitors the user's digital TV remote control operation to collect the viewing history of the program of the user, and the history storing module 101 receives the viewing history information from the viewing history collecting module through a communication network. Store in city hall history storage 102. The communication network is a wired / wireless internet or TV cable network.

The program information store 131 includes program broadcasting schedule information including program title, PID (Program ID), SID (Sequence ID), class, broadcasting start time and end time, and broadcasting time information, and channel, genre and Program attribute information including detailed genres is stored and managed.

The program viewing history is information for knowing the viewing history of the user. According to an aspect of the present invention, the program viewing history may be configured to include a user ID, a program channel, a viewing start time and an end time, and viewing time information. Can be.

In this case, the collection of the program viewing history refers to the user ID, the program channel, the viewing start time and the ending time, and the viewing time by referring to the broadcasting schedule information and the program attribute information of the program information storage in the viewing history collecting module 100. The program viewing history, which is created to include information, may be configured to be stored in the viewing history storage 102 through the viewing history collecting module 101.

Further, according to another aspect of the present invention, the program viewing history may be configured to include a user ID, a program channel, a viewing start time and an end time.

In this case, the history storage module 101 creates a program viewing history by referring to the program information storage 131 including program broadcasting schedule information for each channel.

That is, the history storage module 101 receives only information on a user ID, a program channel, a viewing start time and an ending time from the viewing history collecting module 100, and receives the received information and the program broadcasting schedule for each channel. The program is configured to create a program viewing history including the user ID, the program ID, the reflecting turn information, the viewing start time and the ending time, and the viewing time information in association with the information.

The user ID may be configured by combining numbers, letters, symbols, and the like as information for identifying a user. The program ID is generated based on the PID (Program ID), and the broadcast episode ID is generated based on a sequence ID (SID) stored in the program information store 131 as information about a program's turn.

The broadcast episode ID may be configured by combining numbers, letters, symbols, and the like as the user ID in the order of airing. As a specific example, when the SID of the 9 o'clock news aired at 9 o'clock is 'C501', the drama Jumong aired at 10 o'clock is shown as 'C502', and the Sisa talk show aired at 11 o'clock is shown as 'C503'. Can be configured to be granted.

Table 1 is a table showing an example of the viewing history stored in the viewing history storage 102. Referring to Table 1, Hong Gil-dong watched the P000 program aired at 21 o'clock on October 18, 2006 for 60 minutes, and watched the P001 program aired at the same 22 o'clock in the same month for 55 minutes.

The programs P000 and P001 can be confirmed that the broadcast PIDs are 'C501' and 'C502', respectively, so that the program P000 is broadcast before the program P001. In addition, the broadcast episode ID is a broadcast date, such as '061018-21' (shows October 18, 2006-21 o'clock in the case of program P000) in order to store the viewing history more accurately. It may be configured to include information about time.

The user preference update module 110 updates the preference value for each program attribute of the user model after normalizing the user's viewing frequency for each program attribute by referring to the viewing history of the user.

That is, the user preference updating module 110 reads out a user ID, a title, a program ID, an antireflection ID, a channel of a program, a genre, a detailed genre watching start time and an end time, and viewing time information stored in the viewing history store. To calculate and update the user model information.

For example, the user preference updating module 110 inquires the viewing history store by the user ID, normalizes the attributes of other programs using the frequency of the program attribute with the highest frequency, and then program attribute of the user model for each program attribute. Update the star preference value.

The updating of the preference value for each attribute refers to the user model information including the user's relative weight for the program channel, genre, and detail genre stored in the user model repository 111. Preferably, the attribute value is updated for each program attribute by the average value of the attribute values. A detailed calculation process of the preference value for each attribute will be described later.

Preferably, the user model information may be configured to further include user profile information including the user's ID, name, age, gender and education. The initial value of the user's relative weight for the program channel, genre, and sub-genre may be configured to be input by the user.

The program channel, genre, and sub-genre included in the user model information are numerical values indicating the degree of correspondence between the broadcast program and the user's taste. Examples of channel attributes may include 'MBC', 'KBS', and the like. Examples of genre attributes may be described as 'news' or 'drama'.

The sub-genre attribute means that the genre attribute is further subdivided, and is described as 'drama' among the genre attributes, for example, 'comic drama', 'historical drama', 'melo drama', and the like. Can be.

The example of quantifying the user model information for each attribute is a channel attribute. When a user enjoys MBC, the channel attribute of the user for the channel MBC is set to '1', and the channel attribute for the less-favored channel 'KBS' is relative. It can be explained by setting as low as '0.7'.

The channel attribute set in the above manner is used so that MBC can be recommended with a higher probability than KBS when recommending a program.

The viewing pattern weight calculation module 120 refers to the viewing history of the user for each program, and assigns the viewing pattern weight to each viewing session weight so that the value increases as the viewing time becomes more recent. Accumulate the viewing pattern weight.

The viewing pattern weight is information for program recommendation to be made by reflecting the user's recent tendency toward a program that the user is interested in recently or a program that the user has watched most recently, and depends only on the value of the user's relative weight for each property of the program. This is a concept newly proposed by the present inventors to improve the conventional problem of making a recommendation.

For example, if a user who likes sports and rarely watches a drama (for genre attribute: sports = 1, drama = 0.1) is watching a drama that is gaining popularity, according to the surrounding invitation If a program is recommended according to the value of the user's relative weight for each attribute of the program, the recent trend is not reflected, and thus, only the sport is recommended as in the previous program without drama.

However, if the program recommendation is made by reflecting the viewing pattern weight in the user's relative weight for the program channel, genre, and sub genre, the recent trend is reflected so that the drama that has been recently enjoyed with the sport is added to the recommendation list. can do. For the above reason, it is preferable that the viewing pattern weight is configured such that the value of the viewing pattern weight is increased as the viewing time is more recent for each broadcasting session that is aired during the program airing period.

The program preference calculation module 130 extracts the attribute values for each program attribute from the user model information by referring to the attribute information for each program, and integrates the extracted attribute values into one value by a mathematical operation. The program preference of the user is calculated by applying the accumulated viewing pattern weights.

That is, the program preference calculation module 130 calculates the relative weight for each program attribute of the user for the program channel, genre, and sub genre calculated and updated by the user preference update module 110 and the viewing pattern weight calculation module 120. Program preferences, which are information obtained by integrating the viewing pattern weights of each program accumulated by a) into one value by a mathematical operation, are generated.

At this time, the target of calculating the program preference is a program of each channel that is broadcasted at the program recommendation time zone specified by the user. That is, the program preference calculation module 130 is configured to calculate a program preference of each channel that is broadcasted in a desired time zone by the user whenever there is a program recommendation request of the user so as to recommend a program suitable for the personal preference.

This is because the program is recommended by grasping the recent trends of the user. The information about the program of each channel that is broadcasted in the corresponding time zone is configured to read and grasp the information of the program information store 131.

The program recommendation module 140 recommends a program to a user in order of high program preference. Preferably, the program recommendation module 140 creates a predetermined number of preference higher program lists and outputs them to the recommendation program output module 150 through a communication network.

Then, the recommended program output module 150 receives the higher preference program list and outputs the same to the digital TV screen.

According to an aspect of the present invention, the user preference updating module 110 queries the viewing history of the user on a weekly basis to update the preference value for each program attribute of the user, and the viewing pattern weight calculation module 120 is 1 It is configured to accumulate the user's viewing pattern weight for each program by referring to the user's viewing history on a weekly basis. However, the present invention is not limited by the calculation periods of the above modules 110 and 120.

On the other hand, the viewing pattern weight calculation process is more preferably configured to reflect the recent trend of the user, it may be configured to accumulate the viewing pattern weight for each program of the user whenever the user requests a program recommendation.

After the description of the configuration of the digital TV program recommendation system using the program viewing pattern reflecting technique according to the present invention has been described, the following is a mathematical operation process of calculating the preference value, viewing pattern weight, program preference for each program attribute during the processing It will be described in more detail.

In the user preference update module 110, a user's preference, which is a target of calculation, is calculated based on a program channel, a genre, and a detailed genre based on the user viewing history. The user preference calculation and user model update process is as follows.

First, the viewing frequency of the corresponding period is calculated for each attribute value (eg, 'news', 'drama', etc.) in the user's history. The affinity is defined as the Max value of the property value with the highest frequency among the properties, and then normalized based on the Max value such as (attribute value frequency) / (Max value), which is used as the user model update value. do.

Table 2 is a table showing the viewing frequency of each user genre. Taking Table 2 as an example, looking at the frequency of viewing by user A genre, the frequency of current events / documentary-6 times, news-4 times, culture / information-3 times, education-8 times, movies-4 times. In this case, the Max value is 8 (education) with the highest frequency, and the normalized preference values for each genre are current / doc-0.75, news-0.5, culture / information-0.37, education-1.0, and movie-0.5. do.

5 is a diagram illustrating an example of updating a user model. Referring to the drawing, the affinity value for the genre 'news' is updated to 0.2 (news news value of the existing user model) +0.5 (normalized news preference value) /2.0 = 0.35 (news news value of the updated user model) , The preference value for the genre 'drama' is updated to 0.7 (news news value of the existing user model) + 0 (normalized drama preference value) /2.0 = 0.35 (news user news value of the updated user model).

The user preference updating module 110 updates the preference value for the program channel and sub genre according to the viewing pattern of the user through the above-described calculation process.

Next, the viewing pattern weight calculation module 120 calculates the viewing pattern weight for each episode of the program by the following equation, and accumulates the viewing pattern weight calculated for each episode.

(w ' _{u , p} is the viewing pattern weight for user p's program p,

w _{u , pi} Is the viewing pattern weight for the show i during the corresponding period of program p of user u,

p _i is the program for the show i during that period,

n (p) is the number of episodes of the program p during that period,

α is a constant representing a reflection ratio between past and current viewing history,

history (u) is the watch history for the user u's time period)

Reflection of the viewing pattern weight using Equation 1 is performed only when the user has a history of viewing the program. The viewing pattern weights w ' _{u, p} for the user _p 's program p have the sum of the viewing pattern weights w _{u, pi} for the broadcasting society i during the corresponding period of the program p of the user u. w _{u, pi} has a value of α ^{n-i + 1 if} the user u watched the program p of the broadcasting society i during the period, and 0 otherwise.

6 is a diagram illustrating an example of calculating a viewing pattern weight. Referring to the drawings, programs (gagya, Jumong, StarCraft) that are aired at the same time are shown. These programs have the same airing date and three viewing histories of three weeks of viewing history.

Gagya has the viewing history of 'Month', 'Tue' and 'Tue' of the 1st week, in which case the viewing pattern weight is the viewing pattern weight for each airing day for the whole airing day (n = 6). , Which adds up to 0.96 ⁶ + 0.96 ⁵ + 0 + 0.96 ³ + 0 + 0 = 2.4827. Jumong and StarCraft are calculated the same way.

Looking at gagya and starcraft, which have similar viewing histories for 1 week and 2 weeks, although the viewing history of StarCraft is continuous, it can be seen that the weight of the viewing pattern for gagya, which was recently viewed in week 2, is higher. If you compare gagya, StarCraft, and Jumong, you can see that Jumong's viewing pattern weight is higher when viewing history is more recent.

Table 3 shows a result of applying the viewing pattern weight to a program having the same program preference. The final program preference is calculated by multiplying the program preference by the viewing pattern weight. Even with the same program preferences, as a result of applying the viewing pattern weights, it can be seen that 'Jumong' is recommended because many recent viewing patterns are reflected.

The program preference calculation module 130 calculates a user's program preference by the following equation.

(w ' _{u , p} is the accumulated viewing pattern weight for user u's program p,

k is a positive number greater than zero,

Preference (u, p) is user u's preference for program p,

preference ^u _{gerne (p)} is the preference for the genre of user u's program p,

preference ^u _{channel (p)} is user u's preference for channel of program p,

preference ^u _{subGenre (p)} is the preference for the sub genre of user u's program p)

The preference of the program is calculated by multiplying each user's preference by adding k to the user's preference for the program genre, the preference for the detailed genre, and the channel preference. The reason why k is added to the preference value is to prevent this because if the value of a specific item is 0, the preference for the whole program becomes 0. In some cases, k may be 0.

Table 4 shows viewing pattern weights and user model information of the 'gagya', 'jumong', and 'starcraft' programs of the user 'Hong Gil-dong'. With reference to Table 4 will be described the actual operation of the program preference using the equation. (Constants are calculated as 1)

Calculation of program preferences for gags

2.48 × (0.7 + 1) × (0.8 + 1) × (0.7 + 1) ≠ 12.9

Calculation of the program preference of Jumong

2.77 × (0.7 + 1) × (0.6 + 1) × (0.4 + 1) ≠ 10.55

Calculation of StarCraft's Program Preference

2.45 × (0.5 + 1) × (0.8 + 1) × (0.5 + 1) ≠ 9.92

If only the user model information (program channel + genre + sub-genre) of the user 'Hong Gil-dong' is compared, 'gagya' 2.2 (0.7 + 0.8 + 0.7), 'jumong' 1.7 (0.7 + 0.6 + 0.4), 'starcraft' At 1.8 (0.5 + 0.8 + 0.5), gagya has the highest user preference and lowest jumong.

However, looking at the above calculation results, 'Jumong', which was recently watched a lot by applying a viewing pattern weight, has a higher program preference than 'Starcraft'.

The above results show that even if a digital TV program recommendation system using a program viewing pattern reflecting technique is used to watch a program that does not match the user's usual taste as an unusual case, effective program recommendation can be made by reflecting the recent trend.

In the digital TV program recommendation system using the program viewing pattern reflecting method according to the present invention, after finishing the description of the process of calculating preference value, viewing pattern weight, and program preference for each user model, viewing the program according to the present invention. A method of recommending a digital TV program using a pattern reflection technique will be described.

In the digital TV program recommendation method using the program viewing pattern reflecting method according to the present invention, the viewing history collection module 100 monitors the user's digital TV remote control operation to collect the program viewing history of the user.

Then, the viewing history storage module 101 receives the collected viewing history from the viewing history collecting module 100 and stores the collected viewing history in the viewing history storage 102. The specific details of the information stored in the viewing history storage 102 have already been described above.

Next, the user preference update module 110 normalizes the user's viewing frequency for each program attribute by referring to the stored viewing history information of the user and calculates and updates a preference value for each program attribute of the user model information.

2 is a flowchart illustrating a method of updating a preference value for each program attribute of user model information in the present invention.

First, the user preference updating module 110 reads out a viewing history of a user accumulated and stored in the viewing history storage 102 (S200). Next, preferably, the user's preference value is normalized for each program property (S210).

Thereafter, the user preference update module 110 reads existing user model information stored in the user model repository 111 (S220). The preference value is updated in association with the normalized user preference value for each program attribute based on the existing user model information (S230). The process of normalizing the preference value and updating the preference value for each program attribute has already been described above.

On the other hand, the viewing pattern weight calculation module 120 refers to each viewing session broadcasted during the airing period of the program by referring to the viewing history of the user for each program from the viewing history store when there is a program recommendation request from the user or at predetermined intervals. The viewing pattern weight is added to increase the value as the viewpoint is more recent, and then the viewing pattern weight of each broadcasting session is accumulated.

3 is a flowchart illustrating a method of accumulating the viewing pattern weights of the broadcast episodes in the present invention.

Referring to the drawings, first, the viewing pattern weight calculation module 120 reads out the viewing history of the user (S300). Then, the viewing pattern weights are assigned to each of the broadcast episodes of the program (S310), and thereafter, the viewing pattern weights of each broadcast episode are accumulated (S320). The calculation process of the viewing pattern weights and the accumulation process of the calculated viewing pattern weights of each broadcasting session program have been described above.

It is preferable that the user preference calculation and update process and the viewing pattern weight accumulation process are interpreted as being performed in parallel rather than interpreted as being performed with temporal success.

4 is a flowchart sequentially illustrating a process of making a program recommendation to a user by calculating a program preference of a user according to the present invention.

Referring to the drawing, first, the program preference calculation module 130 checks the presence or absence of a program recommendation request (S400), and if there is a program recommendation request, performs subsequent processing. Although it is possible to select a program to recommend in advance at regular intervals, it may be more preferable to select a program to be recommended when a program recommendation request is made in order to apply the user's most recent viewing trend.

Next, the program preference calculation module 130 reads a program list from the program information store 131 (S410), calculates a user preference value (S420), and accumulates viewing pattern weights (S430). do.

The step S420 may be configured to be performed at a predetermined cycle as described above, but may be configured to be performed whenever a program recommendation request is made as described above.

Thereafter, the preference of each program is calculated based on the data generated in the above process (S440). The preference calculation of each program has already been described above.

Next, the program recommendation module 140 creates and outputs a predetermined number of preference higher program lists (S450), and the recommended program output module 150 receives the higher preference program list and outputs the same to the digital TV screen. (S460).

Hereinafter, a result of performing comparative experiments with 100 viewers for evaluating the efficiency of the digital TV program recommendation system and method using the program viewing pattern reflection technique according to the present invention will be described.

The experiment consisted of 39 weeks of training data, 13 weeks of test data, and 13 weeks of test data. The user model information update and program recommendation cycles were made on a weekly basis, and the number of recommendations was set at five. The experimental data are actual viewing records (corresponding to information stored in the viewing history repository) provided by Nielsen Media Research and program information (corresponding to information contained in the program information repository) provided by EPG Co., Ltd. ].

As a measure of system efficiency, precision and recall were used in the field of IR. The calculation method for Precision and Recall is as shown in Equations 3 and 4, respectively. In the following Equations 3 and 4, the recommendation program is a recommendation list provided by the recommendation system, and the viewing program refers to a TV program watched by a user.

7 is a graph showing a change in Recall according to a change in k value in a program preference calculation equation (Equation 2). Referring to the graph, referring to the graph, it can be seen that the value representing the best recommended efficiency among the k values between 0 and 1.0 used in the experiment is 1.0.

Hereinafter, Experiment A, Experiment B, Experiment C, Experiment D, and Experiment E will be described by way of example. In Experiment A, as a content-based recommendation without applying the viewing pattern weight, the program preference was calculated using only the genre, the detailed genre, and the channel preference of the user. The equation used for the program preference calculation is shown in Equation 5 below.

(preference (u, p) is the u's preference for program p,

preference ^u _{gerne (p) is the} user's preference for program p's genre,

preference ^u _{channel (p) is} the preference for the channel of program p of user u,

preference ^u _{subGenre (p) is} the preference for subGenre of user u's program p)

Experiment B applied only the user model information, and recommended only the programs that were viewed in the past in order of preference. If there were less than 5 programs to be recommended, programs with high preference were recommended, even if the program was never watched. The equation used for the program preference calculation is shown in Equation 6 below.

Step 1: calculate preferences,

Step 2: If viewingRate _{u , p} > 0 only add to recommendation list,

Step 3: If you have less than 5 recommendation lists, even the ones you haven't watched are recommended.

(viewingRate _{u , p} : viewer rating of program p for user u)

Experiment C recommended the programs in the highest viewer ratings. The audience rating is calculated as (total program viewing time) / (total program airing time). If there are less than 5 programs to be recommended, a program with high preference is further recommended among programs for calculating preferences by applying only user model information. The equation used for the program preference calculation is shown in Equation 7 below.

Step 1: Recommend 5 shows that show viewers the most viewers,

Step 2: If you have less than 5 recommendations, calculate your preferences and make the remaining recommendations.

Experiment D is the method proposed in this paper. After calculating the user's preference for the previously watched program, we recommend it by applying the viewing pattern weight. The equation used for the program preference calculation is shown in Equation 8.

(w ' _{u , p} : cumulative viewing pattern weight for user p's program p)

Experiment E was recommended using the KWL user profile [Xizowei Shi, Jin Hua, An Adaptive Preference Learning Method for Future Personalized TV, KIMAS, pp. 260-264, 2005.]. The configuration of user model information and a user preference calculation equation are as follows. The user profile has a keyword for the attribute, a weight for each attribute, and a preference value (a preference formula is shown in Equation 9). The update of the user model information is performed by adding a value obtained by multiplying a learning rate and a broadcasting viewing rate by an existing preference value (see Equation 10 for the equation). The TV program preference is obtained by calculating the similarity between the user profile and the program (see Equation 11 for the equation).

(m is the number of attributes in the profile,

t is the property i,

w _{i is} the weight _of attribute i,

ld _{i is} the affinity value _of attribute i)

(ld _i 'is the updated preference,

ld _{i is} affinity,

α is the learning rate,

β is the broadcast viewing rate,

T is the time of program viewing,

Tr is the show time,

δ is the threshold for viewing a program)

(P is a vector of weights w and affinity ld of the top m attributes in the user profile,

C is a vector of attributes u and weights v for n programs)

The results of the experiment is a graph showing the results of evaluating the recommended efficiency of the five experiments A, B, C, D, E methods in terms of precision and recall, and the eight experiments A, B, C, D, 9 is a graph showing the results of evaluating the recommendation efficiency of the methods of E in terms of precision. As a result, it was confirmed that the experiment D using the viewing pattern weights showed higher performance than the experiments A, B, C, and E.

When the user preference is calculated using Equation 12, the experimental results obtained by adding each property are compared with the experimental result obtained by multiplying each property using Equation 13. In the experiment of multiplying each property, when k is 0 and 1.0, the experiment of adding each property is performed by multiplying each property by the property weight, and the sum of the weights is 1.

(w ₁ , w ₂ , w ₃ are attribute weights)

'Properties plus A experiment' reflects genre (w ₁ = 0.33), sub genre (w ₂ = 0.33), and channel (w ₃ = 0.33) in a ratio of 1: 1: 1, and 'property plus B experiment' Genre (w ₁ = 0.4), subgenre (w ₂ = 0.2), and channel (w ₃ = 0.4) were reflected in a ratio of 2: 1: 2.

The attribute multiply C experiment was performed by setting k to 0 and the attribute multiply D experiment by setting k to 1.0. The results show that the "property times C experiment" shows high performance and the other three experiments show low performance. Experimental results are shown in FIG. 10, which is a graph showing the results of evaluating the recommended efficiency of the attribute addition A experiment, the attribute addition B experiment, the attribute multiplication C experiment, and the attribute multiplication D experiment in terms of Recall.

In order to increase the recommendation performance of the above experiments, we examined the Recall values for each by reflecting the viewing pattern weight. The viewing pattern weight was applied to the attribute addition B experiment, the attribute multiplication C experiment, and the attribute multiplication D experiment, which showed the higher results among the attribute plus A experiment and the attribute plus B experiment.

Referring to FIG. 11, which is a graph showing the result of applying the viewing pattern weight to the result of evaluating the recommended efficiency of the attribute addition B experiment, the attribute multiplication C experiment, and the attribute multiplication D experiment from the perspective of Recall, the value of Recall reflecting the viewing pattern is 0.7 ~. It has a value between 0.9 and it can be seen that much increased compared to the Recall value of FIG. This result indicates that even if the recommendation engine is low in efficiency, reflecting the viewing pattern can improve the efficiency.

One interesting experimental result is the rank reversal of the attribute times C experiment and the attribute times D experiment. If we look at the results of applying the viewing pattern weights to the attribute multiply C experiment and the attribute multiply D experiment, and not the results, the results before applying the viewing pattern weight show that the attribute multiply C experiment shows better efficiency. Later results show that the attribute multiply D experiment yields better efficiency.

In experiments that apply only user model information, it is better to recommend programs with high overall preferences, and in experiments with viewing pattern weights, recommendations that target only programs with high overall preferences are less efficient. have. In other words, when the user's viewing pattern changes, it is different from the learned user model information. Therefore, in order to make an appropriate recommendation at the time of the change, it is necessary to consider the preference of a certain field together rather than the overall preference. will be.

The experiment E was reflected in the viewing pattern weight. In experiments reflecting the KWL user model information, genre, subgenre, and channel attributes were used for the preference calculation. 12 is a graph illustrating a result of evaluating a result of applying viewing pattern weights to the results of Experiment E and the results of Experiment E in terms of Recall.

Referring to the drawing, this experimental result shows the efficiency difference of 1/2 times that of the experiment result reflecting the viewing pattern weight. Although the KWL user model experiment is different from the experiment using the recommendation applying only the user model information, it can be considered that the preference calculation attribute is limited to the genre, the genre, and the channel, and the recent user's viewing pattern is not considered.

13 is a graph showing a ranking in which a user selects a program among recommendation lists provided to the user. Referring to the graph, when the user selects a program in the recommendation list provided to the user, the first and second recommendation programs can be seen a lot. From the 40th to the 52nd week, the probability of selection is 38.32% for the 1st referral program, 36.04% for the 2nd referral program, 18.34% for the 3rd referral program, 7.84% for the 4th referral program, and 3.32 for the 5th referral program. Appears in% order. In other words, the user may select a program having a high recommendation program preference ranking.

One of the considerations when reflecting the viewing pattern is how much the past viewing history and the current viewing history are reflected. This reflection ratio was designated as α, and in the present invention, an optimized value was obtained through experiments. 14 is a graph showing the change of the Recall value according to the reflection ratio α value. Referring to the graph, it can be seen that the highest recommended efficiency is achieved when α = 0.96.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited thereto, and the technical spirit of the present invention and the following will be understood by those skilled in the art to which the present invention pertains. Various modifications and variations are possible, of course, within the scope of equivalents of the claims to be described.

As described above, the digital TV program recommendation system and method using the program viewing pattern reflecting technique according to the present invention, after obtaining the program preferences using a user model composed of attributes such as genre, detail genre, channel, etc. The program recommendation efficiency of pEPG can be improved by reflecting the viewing pattern weight for the existing program.

Claims (5)

delete A viewing history storage configured to store a viewing history including a user ID, an airtime ID of a viewing program, a viewing start time point and an ending time point, and viewing time information; A user model repository for storing user model information including a user's relative weight for program channels, genres, and sub-genres; A user preference updating module for querying user viewing history from the viewing history store, normalizing the frequency of the remaining attributes by the program attributes with the highest frequency, and calculating and updating a preference value for each program attribute; Calculate the viewing pattern weight that accumulates the viewing pattern weight of each broadcasting session after assigning the viewing pattern weight to the viewing point more recently with respect to each broadcasting session aired during the airing period of the program by referring to the viewing history of the user for each program. module; The attribute value is extracted for each program attribute from the user model information queried in the user model repository by referring to the attribute information for each program, and the value of the corresponding program is accumulated to a value obtained by integrating the extracted attribute values into one value by mathematical operation. A program preference calculation module configured to calculate a program preference of a user by applying a viewing pattern weight; And Includes a program recommendation module for recommending a program to the user in order of high program preference; The user preference update module is a digital TV program recommendation system using a program viewing pattern reflection method, characterized in that for updating the attribute value for each program attribute by the average value of the newly normalized program attribute value and the existing attribute value of the user model information. The method of claim 2, The viewing pattern weight calculation module calculates the viewing pattern weight for each broadcasting session of the program according to the following equation, and accumulates the viewing pattern weight calculated for each broadcasting session. Program referral system. Equation

(w ' _{u, p} is the weight of the viewing pattern for the program p of the user u, w _{u, pi} are the viewing pattern weights for the broadcasting session i during the corresponding period of the program p of the user u, p _i is the program for the show i during that period, n (p) is the number of episodes of the program p during that period, α is a constant representing a reflection ratio between past and current viewing history, history (u): watch history for the user u's time period) The method of claim 2, Program airing schedule information including the program title, PID (Program ID), SID (Sequence ID), class, airing start and end time, and airing time information, and program attribute information including the channel, genre, and detailed genre of the program. Further comprises a stored program information store, The program preference calculation module is a digital TV program recommendation system using a program viewing pattern reflecting technique, characterized in that for calculating the user preferences for each program by the following equation. Equation

(w ' _{u , p} is the accumulated viewing pattern weight for user u's program p, k is zero or a positive number greater than zero, Preference (u, p) is user u's preference for program p, preference ^u _{gerne (p)} is the preference for the genre of user u's program p, preference ^u _{channel (p)} is user u's preference for channel of program p, preference ^u _{subGenre (p)} is the preference for the sub genre of user u's program p) (a) the viewing history repository in which the viewing history is stored, including the user ID, the broadcasting episode ID of the viewing program, the starting and ending viewing times, and the viewing time information; and the user's relative weight for the program channel, genre, and genre. Providing a user model repository in which user model information is stored; (b) querying a user viewing history from the viewing history store, normalizing the frequency of the remaining attributes by the program attributes with the highest frequency, and calculating and updating a preference value for each program attribute; (c) accumulating the viewing pattern weight of each broadcasting session after assigning the viewing pattern weight to each viewing session that is viewed during the program's airing period to increase as the viewing point becomes more recent with reference to the viewing history of the user for each program. ; (d) Extracting attribute values for each program attribute from the user model information queried in the user model repository by referring to the attribute information for each program, and corresponding program to a value obtained by integrating the extracted attribute values into one value by mathematical operation. Calculating a program preference of the user by applying the accumulated viewing pattern weight of? And (e) recommending the program to the user in order of high program preference; In the step (b), the method of recommending a digital TV program using a program viewing pattern reflection method, characterized in that the attribute value is updated for each program attribute by the average value of the newly normalized program attribute value and the existing attribute value of the user model information.

Images