KR20120016723A - The content recommendation method by using menu preferences - Google Patents

Abstract

PURPOSE: A content recommending method using menu preference information is provided to recommend content based on a menu preference and to extract personal preferences by using menu information and using history information. CONSTITUTION: A learning engine extracts menu preference information, popular content, and relation menu by using content watching history information and menu information for using content. The learning engine stores the extracted information in a database(S100). A recommendation engine provides a recommendation list by using content meta data and data stored in the learning engine according to the recommending request of a user(S200).

Description

The content recommendation method by using menu preferences}

The present invention relates to a content recommendation method using menu preference information, and more particularly, in order to recommend content desired by a user in a content providing environment having a menu structure, by reflecting user's menu preference information and recentity. The present invention relates to a content recommendation method using menu preference information for recommending content in a corresponding menu using learned menu preference information.

In an environment in which a large amount of content is provided, recommendation is required as an essential function in terms of user's ease of use of content and promotion of content consumption. There is an increasing need for recommendations that can help to make a choice by providing.

IPTV (Internet Protocol Television) is a content providing environment that enables users to watch various multimedia contents anytime, anytime. As the convergence of broadcasting and communication has arrived, contents of various fields are being provided in IPTV environments.

As the IPTV environment provides more diverse contents than the multimedia contents providing environment of the past, the user can use abundant contents while having difficulty in selecting the contents.

The number of contents provided by service providers continues to increase as the contents produced are accumulated. Providing such a large number of contents is an advantage of IPTV because users have various choices. As the number increases, users have to make an extra effort to find the content they want.

The problem of choice caused by the increase in the number of contents has recently emerged as a problem that must be solved in order for IPTV to be successful due to the demand for personalization and customized content provision. Therefore, users may prefer content in an environment that provides a large number of contents. The development of a recommendation system that can sort out the need for development is required.

The recommendation system currently operated by IPTV operators provides simple popularity-based recommendation that recommends users' most used contents, or offers promotion recommendation that organizes a menu by organizing contents that are currently an issue. .

However, this recommendation method provided by the operators is relatively effective compared to the user's need to find the content that is worth seeing among the numerous contents, but it does not meet the increasing demand for personalization and customized content of the service. It is leading to a decline.

In order to solve the above problems, the present invention extracts individual preferences using menu information and usage history information, which are user's content access information, provides content recommendation based on menu preferences, and learns menu preference information. The purpose of the present invention is to provide a content recommendation method using menu preference information reflecting an associated menu in order to improve the over-specialization problem in which only content corresponding to the preference information can be recommended due to the recommendation.

In order to achieve the object, the method may be configured to extract menu preference information, popular content, and related menus from a learning engine and store them in a database by using content viewing history information used by a user and menu information searched for using the content. And a step 200 of providing a recommendation list in the recommendation engine using data and content metadata stored in the learning engine according to a recommendation request of the user.

As described above, the present invention can extract personal preferences using menu information and usage history information, which are user's content access information, and provide content recommendation based on menu preferences, and use menu preference information reflecting an associated menu. Therefore, there is an effect of improving the over-specialization problem that can only recommend content corresponding to the preference information due to the recommendation through the menu preference information learning.

1 is an embodiment of a content organization in which a hierarchical menu is configured according to content types or characteristics according to the present invention.
2 is a schematic diagram showing a content recommendation structure using menu preference information according to the present invention;
3 is a flowchart of a content recommendation method using menu preference information according to the present invention;
4 is a flowchart illustrating a database storing process in a learning engine using viewing history and menu information according to the present invention.
Figure 5 is an example application of the reduction rate of the content extraction step according to the present invention.
6 is an example of transaction extraction in the viewing block of the association menu extraction step according to the present invention.
7 is a flowchart illustrating a recommendation list providing process using stored data and content metadata according to the present invention.

1 is an embodiment of a content organization in which a hierarchical menu is configured according to content types or characteristics according to the present invention, FIG. 2 is a schematic diagram showing a content recommendation structure using menu preference information according to the present invention, and FIG. 4 is a flowchart illustrating a content recommendation method using menu preference information according to the present invention, FIG. 4 is a flowchart illustrating a database storing process in a learning engine using viewing history and menu information according to the present invention, and FIG. 5 is a similar content extraction step according to the present invention. 6 is an example of transaction extraction in the viewing block of the association menu extraction step according to the present invention, and FIG. 7 is a flowchart illustrating a recommendation list providing process using stored data and content metadata according to the present invention.

Hereinafter, the components will be described with reference to the drawings.

The present invention predicts the user's preference by using the content viewing history information after using the user's content and menu information searched for by the user. The IPTV environment provides tens of thousands of VOD contents as shown in FIG. In order to enable the user to find the desired content, the hierarchical menu is organized by the type or characteristic of the content to organize the content.

The menu used for learning the user's preference information may be referred to as an item having a hierarchical structure used by the user to find the content. In the present invention, the menu may be classified according to whether the content is directly included or not. Category menu ”, and a menu containing content directly is defined as a“ content organization menu ”.

Menu information used to learn menu preference information is a content organization menu, and a menu containing a series content such as broadcast and infant content is defined as a 'series content organization menu' according to the characteristics of the included content, and a movie and music Menus having fragment content, such as the content, are subdivided by being defined as a 'short content organization menu'.

If the above-described subdivided content menus are simply learned using only simple content usage frequency information, a menu containing a very small number of contents is less likely to be selected as a preferred menu due to the difference in availability. Perform the preprocessing to integrate the fragment content menu to the higher level.

The content recommendation method using the menu preference information finds the user's preference menu from the viewing history, applies various content recommendation methods according to the characteristics of the preference menu, and if the content recommendation cannot be made within the preference menu, the related menu of the preference menu. Provide content recommendation using.

2 is a schematic diagram showing a content recommendation structure using menu preference information of the present invention, comprising a learning engine 10 and a recommendation engine 20, wherein the learning engine 10 uses a menu using a viewing history and menu information. Learning preference information, extracting popular content, and extracting related menus are performed and stored in a database, and the recommendation engine 20 provides a recommendation list according to a recommendation request by using each result and content metadata of the learning engine 10. do.

The content recommendation method using the menu preference information of the present invention configured as described above will be described with reference to FIG. 3. The learning engine 10 uses the content viewing history information used by the user and the menu information searched for using the content. In step 100 of extracting menu preference information, popular content, and related menus from the database and storing them in a database, a recommendation engine using data and content metadata stored in the learning engine 10 according to a user's recommendation request In step 20, the step S200 of providing a recommendation list is included.

The step 100 (S100) is a step 110 (S110) for learning the preference information for the menu based on the content usage frequency information in the menu as shown in Figure 4 (S110), the overall popular content and the menu Step 120 (S120) for ranking in the order of the highest frequency using the popular content and storing in the popular content database, and step 130 (S130) for calculating the similarity between the content from the viewing history data and storing it in the content similarity database; In step 140, the content extraction relation is extracted from the viewing history data as a rule and stored in the content association rule database, and from the time when the user starts watching to the time when the user finishes using the content continuously. Viewing it as a viewing unit and defining it as a viewing block to relate usage of the menu in the viewing block. Extracting and storing the association menu database 150 (S150).

In the database storage of the step 110 (S110) to step 150 (S150), the order of "save menu preference information, popular content storage, similar content storage, related content storage, related menu storage" is carried out in the embodiment of FIG. It is not limited to the example, the database storage order can be changed, it can also be done in parallel flow of the work at the same time.

In the menu preference information learning step of step 110 (S110), the preference information for the menu is learned based on the content usage frequency information in the menu, and users find a desired menu to use the content and select the desired content in the menu. Find and use the search process. The content search process is not only the preference information of the content used by the user, but also the preference information of the menu having the characteristics of the organization information to which the content belongs.

The basic preference for the menu (m) to which the content used by the user (u) belongs is defined as in Equation 1 below based on the frequency of using the content, and an example of the result is shown in Table 1 below.

은 사용자(u)의 메뉴(m)에 대한 선호도,

는 사용자(u)가 메뉴(m)의 콘텐츠를 d시간 이상 이용한 빈도,

는 사용자(u)의 선호로 볼 수 있는 최소 시청시간 임계값을 의미한다.here,

Is the user's (u) preference for menu (m),

Is the frequency with which user u has used the content of menu m for at least d hours,

Denotes a minimum viewing time threshold that can be seen as the preference of the user u.

Since the menu basic preference obtained by Equation 1 is obtained by using accumulated cumulative frequency regardless of time, since the past preference is treated the same as the recent preference, the present menu preference information of the user cannot be accurately represented. In order to learn more accurate menu preference information, we apply the decrease rate of preference over time, which reduces the influence of old preference information such as the characteristics of 'drive' in the past of the series contents over time and makes the recent preference more important. For the sake of simplicity.

The application of the preference reduction rate is reflected through a process of dividing the viewing history into a certain period of time to extract the basic preferences, and multiplying the preference reduction rate by the weight by a period. The preference of the entire period reflecting the recentity is calculated as shown in Equation 2 below as the sum of the preferences of all the periods in which the preference reduction rate is multiplied by the weighted product.

는 지난 시간(기간),

는 선호감소율(0＜W≤1),

는 t기간의 선호도,

는 전체 기간의 선호도를 의미한다.here,

Is the last time (period),

Is the decrease rate of preference (0 <W≤1),

Is the preference for period t,

Means the preference of the whole period.

5 is an example of applying the reduction rate, the preference of the menu reflecting the recentity by applying the reduction rate is compared with the frequency-based preference that does not reflect the recentness, and the preference of the recently used menu is reflected as the preference of the higher rank. In addition, the preference of the old 'driven' menu in the past can be seen to have a lowering effect.In the case of applying the reduction rate, the smaller the value of the preference decrease rate, the more important the recent preference is. Has a less important effect.

The extracting of the popular content of step 120 (S120) is to rank all the popular content and the popular content in the menu in order of high frequency of use and store them in a database. The extracted popular content for each menu is the most basic of recommendation of short content. It is used for recommending popular content, and popular content recommendation in exceptional circumstances is impossible in the preference menu and related menu.

Table 2 below is an example of the extraction of the entire popular content, Table 3 shows an example of the extraction of popular content for each menu.

In the extracting of the similar content in step 130 (S130), the similarity between the contents is calculated from the viewing history data and stored in the database. The content vector used for calculating the content similarity has the number of users and the content used by the user. Is 1, unused content is defined as a vector having a binary value of 0, and cosine similarity that is commonly used as in Equation 3 below is used to calculate similarity between vectors.

Equation 3 is a formula of dividing the inner product of two vectors by the product of magnitude, and is calculated by the product of the elements of each vector and the square operation, and this equation is calculated as in Equation 4 below in the operation between binary vectors used in the present invention. It can be simplified and utilized, since a vector element having a value of 0 or 1 can be simplified to an operation for calculating a simple frequency because the value of 0 or 1 has a value of 0 or 1 even if it takes a product or square operation.

는 콘텐츠 i와 j의 유사도(코사인),

는 콘텐츠 i와 j를 모두 이용한 사용자 수,

는 콘텐츠 i를 이용한 사용자 수를 의미한다.here,

Is the similarity (cosine) of content i and j,

Is the number of users using both content i and j,

Denotes the number of users using the content i.

As a result of the similar content extraction step, the similarity value between the two contents is stored in a database, and the form of the result is shown in Table 4 below.

In the extracting of the related content in the step S140, the extracting of content usage relevance from the viewing history data as a rule is stored in a database. In the step of extracting content usage relevance, one transaction is performed by the user during the learning period. It is composed of a usage list, and the related content calculation and filtering process is based on the support, confidence, and lift values calculated from the extracted transactions, filters the thresholds, and ranks and stores the rules based on the confidence values.

In Equation 5, support is a probability of using both contents, and may be referred to as a ratio of users who use both contents to obtain relevance among all users, and as the number of users using both contents increases, the support value between the two contents increases. It becomes larger and has a value of 0 to 1 at the minimum.

Confidence is defined as the ratio of users using one content X to another using content Y. The larger the number of users using content Y, the greater the value, and the minimum value is 0 to 1.

lift is defined as the ratio of the percentage of users who use content Y to the total number of users and the percentage of users who use content Y among the users who use content X. If used a lot, the value becomes large, and it has a minimum value of 0 and a value of 1 when the two ratios are the same. The larger the value of the lift, the greater the relevance. However, if the total number of times or the ratio used is small, the value is rapidly increased because it is a ratio value.

In calculating the association value of two contents, since the number of operations increases to n ² as the number of contents increases, a scalability problem occurs. Therefore, a method of lowering the calculation to a processable level by applying a threshold value when extracting an association rule is required. Contents that are extremely rarely used based on the support value of the content are easy to extract the wrong association rule, so the calculation is not performed to reduce the computational complexity and the primary filtering, and the association values are sorted based on confidence and stored in the database. And, the results are stored only the rules above the threshold of the support, confidence, lift value and the form is shown in Table 5 below.

In the step of extracting the associated menu of step 150 (S150), a view is made from the time point at which the user starts watching to the time point at which the end of watching after using the content is defined as a viewing block and defined as a viewing block. Extract usage associations.

The step 150 (S150) defines the difference between the end of the previous viewing and the start of the subsequent viewing as the viewing blank time from the viewing history data arranged in order of the viewing start time for each user, and extracts the viewing block based on the viewing blank time. Step 151 (S151), a step 152 (S152) of extracting the menu used in the viewing block as a transaction, and using the extracted menu using transaction to calculate the correlation value (support, confidence, lift) between the menu Step 153 (S153), and the step 154 (S154) of filtering the association menu through the threshold value adjustment of the association value and stores the result in the association menu extraction database.

FIG. 6 is an example of extracting a menu usage transaction from the viewing block in step 152 (S152). Rows with different colors indicate a transaction in which the viewing block is extracted, and the menu using transaction in the rows having the same color. Extract and utilize

Table 6 below shows the results of the association menu extraction.

Step 200 (S200) is a step 210 of selecting a preference menu using the menu preference information learned when the recommendation request occurs as shown in Fig. 7 (S210), and finds content that is preferred within the preference menu recommendation method In step 220 of applying (S220), and step 230 of generating and providing a content recommendation list (S230), and if the generated recommendation list does not meet the number of recommendation requests, the relevant menu of the preferred menu is extracted, In step S240, a content recommendation based on characteristics is performed.

The selection of the preference menu is made by using the learned menu preference information. The selected preference menu is divided into a series content organization menu and a short content organization menu and a recommendation method is applied accordingly.

In the present invention, one content is recommended for each preference menu, and the reason is that when the number of recommended contents is too large at a time from the viewpoint of the user receiving the recommendation, the user must select again, and provide a limited number of recommendations as appropriate. In this case, recommending more than one copy in a series narrows down the choice of content.

In the method for recommending content of the preferred menu proposed by the present invention, the preferred menu is selected based on the learned menu preference information. After the user selects the preferred menu, the user selects the preferred content in the preferred menu, and recommends using the preferred menu. As a result of the first filtering, the content in the user's preference menu is a candidate, and a method of recommending content among the latter contents is a recommendation method in the preference menu.

When the preference menu is selected, the recommendation method is determined according to the characteristics of the content belonging to the preference menu. If the preference menu is a series consisting of series contents, a series rule-based recommendation is applied. Apply recommendation method applying content recommendation, item-based collaborative filtering, and item association rules.

The series rule-based recommendation applies a rule that recommends the next side of the previous viewing in consideration of the previous viewing history of the preference menu or re-recommends considering the viewing ratio to the content length of the previous viewing.

The popular content recommendation method in the short content menu is a method of recommending the most popular content used in the menu when a preferred menu is selected as a short content menu and content recommendation within the menu is recommended. In the learning phase, the similarity between similar items of items used by on-line users is calculated using the similarity between items previously calculated through batch operation, and the predicted preference is ranked. Recommendations are made in the manner provided.

Equation 6 below calculates the predicted preferences of the contents which are candidates for recommendation, and calculates the predicted preferences by calculating the sum of similarities with the contents for which the user wants to obtain the predicted preferences. In addition, the method of recommending the prediction preferences of the contents to be recommended and ordering the prediction preferences in the order of preference is applied.

는 사용자 p의 콘텐츠 i에 대한 예측 선호도,

는 사용자 p가 이용한 콘텐츠 j의 선호도(이진 값 0 또는 1),

는 콘텐츠 j와 콘텐츠 i의 유사도를 의미한다.here,

Is the predicted preference for content i of user p,

Is the affinity (content of binary value 0 or 1) of content j used by user p,

Denotes the similarity between the content j and the content i.

The item association rule recommendation is a method of recommending content existing in a higher association rule of content recently used by using an association rule between contents generated as a result of the learning step.

When the recommendation is limited in the preference menu, the recommendation is performed using the related menu. There are two cases in which the recommendation cannot be recommended in the preference menu. Assumes that only content is recommended), and secondly, consumes all content in the preferred menu.

Selection of the association menu for the preference menu is determined through a process of calculating and ranking the preference prediction value of the association menu, as shown in Equation 7 below.

는 사용자 u의 선호메뉴 집합,

는 메뉴 m의 연관메뉴 집합,

는 사용자 u_i의 연관메뉴 am_j의 선호도 예측값,

는 사용자 u_i의 선호메뉴 pm_k의 메뉴 선호도,

는 메뉴 pm_k와 메뉴 am_j의 연관값을 의미한다.here,

Is the user's preference menu set,

Is the set of menus associated with menu m,

Is the predicted preference of the associated menu am _j for user u _i ,

Is the menu preference for user u _i pm _k ,

_Denotes the association of menu pm _k with menu am _j .

The content recommendation in the related menu is selected by combining and ranking the preferences of the preference menu and the related values of the related menu of the preferred menu, and apply a method to supplement the insufficient recommendation result through the selected related menu.

Accordingly, the present invention extracts individual preferences using menu information and usage history information, which are user's content access information, provides content recommendation based on menu preferences, and learns menu preference information using menu preference information reflecting an associated menu. Due to the recommendation through, it improves the over-specialization problem that can only recommend content corresponding to the preference information.

Although the present invention has been shown and described with respect to specific embodiments thereof, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined by the appended claims. Anyone who can afford it will know.

10: learning engine 20: recommended engine

Claims (8)

In the content recommendation method using the menu preference information,
Step 100 (S100) of extracting menu preference information, popular content, and a related menu from the learning engine 10 using the content viewing history information used by the user and the menu information searched for using the content, and storing them in a database; ;
In accordance with a recommendation request of the user using the data stored in the learning engine 10 and the content metadata using the menu preference information, characterized in that it comprises a step (200) (S200) for providing a recommendation list in the recommendation engine 20 How to recommend content. The method of claim 1, wherein the first step S100 is
Step S110 of learning preference information for the menu based on the content usage frequency information in the menu and storing the preference information for the menu in a menu preference information database;
Step 120 (S120) of ranking the most frequently used contents and the most popular contents in the menu in order of popularity and storing the contents in a popular contents database;
Step S130 of calculating similarity between contents from the viewing history data and storing the similarity between contents in a contents similarity database;
Step 140 (S140) of extracting a content use relation from a viewing history data as a rule and storing the content usage association in a content association rule database;
A view from a point in time at which a user starts watching to a point in which a user finishes using a content and then ends viewing, defined as a viewing block, extracts the use association of the menu in the viewing block, and stores the same in a related menu database; Content recommendation method using the menu preference information, characterized in that it comprises a step (S150). The method of claim 2,
In order to learn the accurate menu preference information of step 110 (S110), a preference reduction rate according to time is applied, but the application of the preference reduction rate is performed by dividing the viewing history into a predetermined period to extract basic preferences, and weight the preference reduction ratio to the basic preferences by period. The preference of the whole period reflected by multiplying by, and reflecting the recency is the sum of the preferences of all periods by applying the percentage reduction ratio to the basic preference of each period.

, Where

Is the user,

Silver menu,

Is the last time (period),

Is the decrease rate of preference (0 <W≤1),

Is the preference for period t,

The content recommendation method using the menu preference information, characterized in that the preference of the entire period. The method of claim 2,
The extracting of the popular content of the 120th step (S120) is used for recommending the most basic popular content of the short content recommendation among the extracted popular content for each menu, and the overall popular content in the exceptional situation in which the recommendation is not possible in the preferred menu and the related menu. Content recommendation method using the menu preference information, characterized in that used for recommending popular content. The method of claim 2,
The content vector used for calculating the content similarity in operation 130 (S130) has a number of users, the content used by the user is defined as a vector having a binary value of 1, and the content not used has a cosine similarity. Leverage

, Where

Is the similarity (cosine) of content i and j,

Is the number of users using both content i and j,

The content recommendation method using the menu preference information, characterized in that the number of users using the content i. The method of claim 2,
Extracting the content usage association in step 140 (S140) consists of a usage list of contents used by the user during a transaction learning period, and the related content calculation and filtering process is based on the support, confidence, and lift values calculated from the extracted transaction. The method of recommending content using menu preference information, wherein the threshold value is applied and filtered, and then the rules are ranked and stored based on the confidence value. The method of claim 2, wherein the 150 step (S150)
Extracting a viewing block from the viewing history database (S151);
Extracting a menu used together in a viewing block as a transaction (S152);
Calculating an association value between menus by using the extracted menu using transaction (S153);
And filtering the associated menu by adjusting threshold values of the associated values and storing the result in the associated menu extraction database (S154). The method of claim 1, wherein step 200 (S200)
Step 210 of selecting the preferred menu using the learned menu preference information when the recommendation request occurs;
Step S220 of applying a recommendation method by finding content that is preferred in the preference menu;
Step 230 of generating and providing a content recommendation list;
If the generated recommendation list does not meet the number of recommendation requests, the relevant menu of the preferred menu is extracted, and content using the menu preference information comprising the step S240 of making a content recommendation according to the characteristics of the relevant menu. Recommended way.

Images