KR20080043140A - Collaborative filtering system and method - Google Patents

Abstract

A collaborative filtering system and a method thereof are provided to offer a correct recommendation result to users of a minor group by considering propensity of the users belonging to the minor group, and calculate correct relation between user preference and contents by excluding user action inducing an incorrect result. A contents providing server interface(105) receives contents use information of users by interfacing with a contents providing server. A contents use information management database(115) stores and manages the contents use information of the users. A user contents preference calculator(120) calculates contents preference applying weight according to user action and/or use time information of each contents by analyzing the contents use information stored in the database. A contents similarity calculator(130) determines a pair of contents frequently used together by analyzing the contents use information stored in the database and calculates similarity of the contents pair by applying the weight according to a use type of the contents pair and attribute of the paired contents. A user recommendation data generator(140) generates recommendation data for each user by using both the user contents preference calculating results and similarity data.

Description

Collaborative filtering system and method {COLLABORATIVE FILTERING SYSTEM AND METHOD}

1 is a schematic block diagram of a collaborative filtering system according to an embodiment of the present invention;

2 is an example of a table structure for managing content usage information for collaborative filtering according to an embodiment of the present invention;

3 is an example of a time weight value applied when calculating a content preference according to an embodiment of the present invention;

4A to 4C illustrate examples of a table structure for managing user-specific content usage information separated from the content usage information illustrated in FIG. 2;

5 is an example of a user information management table structure that is a property classification criterion of a user for collaborative filtering according to an embodiment of the present invention;

6A to 6B are examples of a table structure for managing content usage information for each user attribute separated from the content usage information illustrated in FIG. 2 based on the user attribute classification criteria illustrated in FIG. 5;

7 is an example of a table structure for managing similarity analysis results between specific contents for collaborative filtering according to an embodiment of the present invention;

8 is a flowchart illustrating a process of providing recommendation data for each user using a collaborative filtering method according to an embodiment of the present invention;

9 is a flowchart illustrating a process of calculating content preference for each user according to an embodiment of the present invention;

10 is a flowchart illustrating a content similarity calculation process according to an embodiment of the present invention.

<Explanation of symbols for the main parts of the drawings>

100: collaborative filtering system 105: content providing server interface unit

110: content usage information storage unit 115: content usage information management DB

120: content preference calculator for each user 125: content preference management DB for each user

130: content similarity calculating unit 135: similarity data management DB

140: recommendation data generation unit for each user 145: recommendation data management DB for each user

150: user recommendation data provider

The present invention relates to a collaborative filtering system and a method thereof, and more particularly, to a collaborative filtering system and a method for enhancing the accuracy of the recommendation by properly reflecting the attributes of the user, the attributes of the content, and the user's time of use of the contents.

With the rapid development of the Internet, a vast amount of information is provided to users, which makes it increasingly difficult for users to find the information they want. In addition, even if the desired information is found, the time required to find the information (ie, information retrieval time) has increased. Therefore, recently, the necessity of a 'personalization service' that reduces information retrieval time and provides only the data desired by the user, that is, valuable data for the user, has emerged.

'Personalization Service' refers to only data that is expected to be valuable to the user by using information implicitly or explicitly provided by the user (for example, user's personal information and preferences), and the selected data to the user. Say the service you provide. Hereinafter, it is defined as 'recommended' to provide the data selected for each user as described above.

"Collaborative filtering" is a kind of "personalization service" and refers to a method of using similarity between users or contents in determining data to be recommended to users. In other words, 'collaboration filtering' refers to 'user to user collaboration filtering', which uses the similarity between users, and 'item to item' collaboration, which uses the similarity between contents. Filtering ”or collaborative filtering that merges the two. The features of 'collaboration filtering between users' and 'collaboration filtering between contents' are as follows.

First, 'collaboration filtering between users' is used to classify users who have a similar pattern by using information provided by the user (for example, user's personal information and preferences), and then purchase information or contents of other users included in the same group or The recommendation data of a specific user is determined using the content usage information. That is, cross recommendation is performed among users included in the same group. For example, in the recommendation system using the 'collaboration filtering between users' method, if 'user 1' and 'user 2' are included in 'group 1' and 'user 1' has purchased 'content 2' Include 'Content 2' in the recommendation data of 'User 2'.

Meanwhile, 'collaboration filtering between contents' determines the recommendation data of a specific user by using a pair of contents simultaneously used by the same user. For example, in a recommendation system using a 'collaboration filtering between contents' method, when 'user 1' purchases 'content 1' and 'content 2' at the same time, if 'user 2' purchases 'content 1', the recommendation system Includes 'content 2' in the recommendation data of 'user 2'.

These conventional "collaborative filtering" methods have the following problems.

First, the conventional 'collaborative filtering' method does not consider the propensity of users belonging to the minority group and obtains the recommendation results biased to the propensity of the majority users. For example, if a total of 100 users select 'Content 1' and 'Content 2' at the same time and 80 users select 'Content 1' and 'Content 3' at the same time, In the filtering method, the combination of 'content 1' and 'content 2' is four times similar to 'content 1' and 'content 3'. That is, similarity is calculated by the number of users without considering the user's attributes (eg, gender, age, region, etc.). Therefore, 70 females and 10 males among the 80 users who simultaneously selected the 'Content 1' and the 'Content 2' were 5 females among the 20 users who simultaneously selected the 'Content 1' and 'Content 3', If you have 15 men, 15 times more males choose Content 1 and 3 Contents than 1.5 people choose Content 1 and 2 Contents Nevertheless, in the related art, 'content 2' rather than 'content 3' is recommended to a male user who selected 'content 1' according to the result calculated by the total number of users. This shows an example in which the tendency of male users belonging to the minority group does not reflect the tendency of female users who are majority users without considering gender in the analysis of the propensity of 25 male and 100 female users. will be.

In addition, in the conventional 'collaboration filtering' method, users who randomly access the content without using the content selectively according to their preferences when calculating the user's preference and the similarity between the contents (for example, a shopping mall promoter and a spam advertiser) And program developers, etc.), as it is, it is a cause of inaccurate results. For this reason, the conventional "collaborative filtering" method has a problem of low recommendation accuracy.

In addition, the user's preference for the content changes over time, and there is a problem in that it is difficult to apply it properly with the conventional 'collaboration filtering' method. That is, in the related art, the preference tendency of the contents is determined simply by using only the use time of the individual contents, so that when the valid content usage history period is set long, the accuracy decreases. There is a problem.

Lastly, the conventional 'collaboration filtering' method simply refers to the user's content usage history only when calculating the similarity of the content even when the property of the content that can be estimated similar to the content category, color, music genre, etc. is already known. . Therefore, it is difficult to calculate more accurate similarity.

The present invention provides a collaborative filtering system and method for providing accurate recommendation data to solve the above problems.

In addition, the present invention is to provide a collaborative filtering system and method for providing accurate recommendation results to users of the minority group by considering the propensity of users belonging to the minority group.

In addition, the present invention is to provide a collaborative filtering system and method for calculating accurate user preferences and similarity between contents by excluding user's actions that lead to inaccurate results.

In addition, the present invention is to provide a collaborative filtering system and method for accurately calculating the similarity between the contents by using the time interval between the contents by the same user.

In addition, the present invention is to provide a system and method for collaborative filtering that accurately calculates the similarity between contents using the attributes of the content that can guess the similarity of the contents.

In order to achieve the above object, the present invention provides a collaborative filtering method comprising: receiving and storing content usage information of users from a content providing server; Analyzing the content usage information to determine the user's behavior and / or use time information for each content, and calculating the user's preference for each of the content by applying a weight according to the user's behavior and / or use time information process; Analyzing the content usage information to determine a content pair that is often used together, and calculating similarity of the content pairs by applying weights according to the usage form of the content pair and the attributes of each of the contents of the pair; ; And generating user-specific recommendation data using the user-specific preference and the similarity of the content pair.

At this time, the preference calculation process for each user may include: analyzing the content usage information to determine a user's behavior for each content; Calculating a user's behavior preference for the corresponding content by applying a preset weight of preference for each behavior to each of the determined actions of the user; Determining time usage information of each of the contents and calculating time preference for each user by applying time weights according to the usage time information; And calculating a user's preference for the corresponding content by using the user's action preference and the user's time preference.

The user preference calculation process may further include filtering abnormal user information based on a preset abnormal user determination criterion.

In addition, the preference calculation process for each user may further include the step of calculating a preference for each user attribute after separating the user preference calculation result using a user attribute (eg, gender, age, region, etc.).

In addition, in the step of applying the time weight, it is preferable to apply a time weight that is inversely proportional to the interval between the most recent use time of the corresponding content and the current time.

The calculating of the similarity may include analyzing the content usage information to determine a combination of two contents that are selectable from two or more different contents consecutively used by the same user within a preset time interval as a pair of similar contents; Applying a time weight inversely proportional to an interval between usage times used by the same user for the pair of similar contents; And applying attribute weights according to attributes of each of the contents forming the similar content pair.

The similarity calculation step may further include filtering the content pairs when the number of users using the content pairs is less than or equal to a predetermined minority user determination value.

In the determining of the similar content pair, if the number of times that two or more different contents are consecutively used by the same user within a preset time interval is greater than or equal to a preset number, the two content combinations selectable among the contents are determined as the similar content pairs. It is desirable to.

The applying of the time weight may include determining an interval between usage time points used by the same user for the pair of similar contents; Selecting time weights corresponding to the intervals between the use time points among predetermined time weights in inverse proportion to the time intervals at predetermined time intervals; And applying the selected time weight.

In addition, the attribute weight applying step preferably applies an attribute weight proportional to the similarity of the category to which the contents belong.

Meanwhile, in order to achieve the above object, the collaborative filtering system provided by the present invention provides a content providing server interface unit for receiving content use information of users from the content providing server through an interface with a content providing server providing a content service to users. ; A content usage information management database unit which stores / manages content usage information of the users; A content preference calculator for each user, which analyzes content usage information of users stored in the content usage information management database unit and calculates content preference for each user to which a weight is applied according to user's behavior and / or usage time information for each content; Analyzes content usage information of users stored in the content usage information management database unit to determine a content pair that is frequently used together, and applies a weight according to the usage form of the content pair and the attributes of each of the paired contents. A content similarity calculation unit configured to calculate similarity of the content pairs; And a user-specific recommendation data generation unit configured to generate recommendation data for each user by using the content preference calculation result of the user of the preference calculator and the similarity calculation result of the content pairs of the content similarity calculator.

In this case, the content preference calculator for each user analyzes the content usage information to determine a user's behavior for each content, and calculates a user's behavior preference by applying a preset preference weight for each behavior to each of the determined user's behaviors. After that, it is preferable to calculate the content preference including the user preference for each user.

In addition, the content preference calculator for each user analyzes the content usage information to determine the use time information of each of the contents, and calculates the time preference for each user by applying a time weight according to the use time information. It is preferable to calculate the content preference including the time preference for each user.

The content preference calculator for each user may be configured to calculate abnormal user information after filtering abnormal user information based on a predetermined abnormal user determination criterion.

In addition, the content preference calculator for each user may be configured to calculate a preference for each user property after separating the user preference calculation result using the user property.

In addition, the content preference calculator for each user calculates the content preference for each user by applying a time weight that is inversely proportional to the interval between the most recent use time of the corresponding content and the current time.

In addition, the content similarity calculating unit analyzes the content usage information to determine two content combinations that can be selected among two or more different contents consecutively used by the same user within a preset time interval as a similar content pair and then the similar content pair It is preferable to calculate the similarity obtained by applying a time weight that is inversely proportional to the interval between the use time points used by the same user.

In addition, the content similarity calculator may determine that two or more different contents are consecutively used by the same user within a preset time interval and determine a combination of two selectable contents among the contents as a pair of similar contents. desirable.

In addition, the content similarity calculation unit preferably calculates a similarity by applying attribute weights according to attributes of each of the contents constituting the pair of similar contents.

The content similarity calculator may filter the pair of similar contents when the number of users using the pair of similar contents is less than or equal to a predetermined minority user determination value.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. At this time, it should be noted that the same elements in the accompanying drawings are represented by the same reference numerals wherever possible. In addition, detailed descriptions of well-known functions and configurations that may unnecessarily obscure the subject matter of the present invention will be omitted.

1 is a schematic block diagram of a collaborative filtering system according to an embodiment of the present invention. Referring to FIG. 1, the collaborative filtering system 100 according to an exemplary embodiment may include a content providing server interface 105, a content usage information storage 110, and a content usage information storage management database (DB). 115, content preference calculator 120 for each user, content preference manager DB 125 for each user, content similarity calculator 130, similarity data management DB 135, recommended data generator 140 for each user, Similarity data management DB 145 and the user-specific recommendation data providing unit 150 is included.

The content providing server interface 105 performs an interface with a content providing server (not shown) that provides a content service to users. For example, the content providing server interface 105 receives the content usage information of the users from the content providing server and delivers the content usage information to the content usage information storing unit 110, and the user recommendation data providing unit 150 for each user. Receives the recommendation data and delivers it to the content providing server. At this time, the content providing server interface 105 preferably receives the content usage information of the users or provides the recommendation data for each user on a predetermined time basis.

The content usage information storage unit 110 stores / manages content usage information of users. For example, the content usage information received from the content providing server interface 105 is stored in the content usage information management DB 115.

The content usage information management DB 115 stores / manages content usage information of users delivered through the content usage information storage 110. At this time, the content usage information management DB 115 preferably stores the behavior of the user's content in a three-dimensional array divided by actions. An example of a table structure for managing the content usage information in the content usage information management DB 115 is illustrated in FIG. 2.

In the example of FIG. 2, a table structure for storing the content usage information and a result of determining a preference applied to each of the content usage information items together is illustrated. In particular, the example of FIG. 2 illustrates an example of arranging behavior information of a user's content in chronological order. Referring to FIG. 2, the content usage information includes content usage time for any content, user information using the content, and content information targeted for the content use, and the user's behavior information (eg, behavior) for the content. Number, type of action, and weight for each action). In this case, time preference, weight for each action, behavior preference, and preference score, which are not mentioned, are calculated by the content preference calculation unit 120 for each user with reference to the content usage information, and operate on the content preference calculation unit 120. Items not mentioned above in the description will be described in more detail.

Referring back to FIG. 1, the content preference calculator 120 for each user uses content usage information (eg, content usage time, information on the user, and contents targeted for use) as illustrated in FIG. 2. The user preference for each content is calculated. For example, the content preference calculator 120 for each user analyzes content usage information of users stored in the content usage information management DB 115, and then calculates content preference for each user to which the weighted value is applied. In particular, the content preference calculator 120 for each user calculates a user's preference for each content and a time preference in consideration of content usage time information, and then calculates a user's preference considering both values. .

To this end, the content preference calculator 120 for each user analyzes the content usage information to determine a user's behavior for each content, and applies a preset preference weight for each behavior to each of the determined user's behaviors. Calculate your preferences. FIG. 2 calculates behavioral preferences for each user's behavior for a specific content when the weight for each activity is set to '1', '5', and '3' for the behavior types 'read', 'purchase', and 'write review'. An example of one case is shown. That is, in the example of FIG. 2, the content usage information registered in the first item is the case where 'user 1' performs all the actions of 'read', 'purchase', and 'write review' for 'content 1'. When the weight '9', which is the sum of all set weights, is calculated as the behavior preference, and the content usage information registered in the second item is 'user 3' performing 'read' and 'purchase' actions on 'content 1'. Therefore, the weight '6', which is the sum of the weights set for each of them, was calculated as the behavior preference.

In addition, the content preference calculator 120 for each user analyzes the content usage information to determine the use time information of each of the contents, and calculates the time preference for each user for the corresponding content by applying a time weight according to the use time information. do. In this case, the use time information may be 'an interval between a most recent use time of a specific content and a current time'. That is, the content preference calculator 120 for each user may calculate the time preference for each user by applying a time weight that is inversely proportional to the 'interval between the most recent use time of a specific content and the current time' as illustrated in FIG. 3. . This is to apply a high weight to recently used content.

In addition, the behavior preference and the time preference are applied together to calculate the user's preference for the corresponding content. 2 illustrates an example in which the sum of the behavior preference and the time preference is calculated as a preference score.

Meanwhile, the content preference calculator 120 for each user presets an abnormal user determination criterion and filters the content usage information of the user who satisfies the criterion. That is, the content usage information of the user determined as the abnormal user is not considered when calculating the content preference for each user. If some users of the service access the content randomly without using the content selectively according to their own preferences, such as displaying advertisements to an unspecified number of users such as spam advertisements, the association between the contents This is to avoid inducing incorrect results in the calculation.

To this end, the content preference calculator 120 for each user may preset a criterion for determining an abnormal user based on the user's attribute or behavior type and filter the user's content usage information that satisfies the criterion. In this case, the content preference calculator 120 for each user may be an owner of an online / offline shopping mall that sells the contents, a program developer who develops the contents, or a user whose nationality is a country other than the target country for selling the contents. If that user is determined to be an abnormal user. In addition, by analyzing the behavior of the user, a user who exhibits the behavior of accessing the same content more than the maximum number of accesses within the minimum access time interval or determines the user who performs the same behavior more than the maximum number of executions within the minimum execution time interval as an abnormal user. . In particular, when the target of the same action is an unspecified number of users, it is preferable to determine the user as an abnormal user.

In the example of FIG. 2, if the time interval of 'yy / mm / dd hh: mm: ss 6_1 to yy / mm / dd hh: mm: ss 6_3' is within a preset minimum access time interval and the maximum number of accesses is three times. In addition, the content usage information of the sixth to eighth items of the table of FIG. 2 corresponds to the 'action type of accessing the same content more than the maximum number of times within a minimum access time interval.' Therefore, the content preference calculator 120 for each user determines the item as a filtering target.

In addition, the content preference calculator 120 for each user calculates a preference for each user property after separating the user preference calculation result using a user property in order to solve a problem in which the characteristics of a user belonging to a minority group are not reflected. desirable. For example, it is preferable to separate the preference calculation result for each user based on a combination of at least one user attribute among gender, age, and region.

To this end, the content preference calculator 120 for each user first divides the content usage information management table displaying the content preference for each user as illustrated in FIG. 2 for each user as illustrated in FIGS. 4A to 4C. That is, FIG. 4A illustrates an example of separating content usage information and preference information of 'user 1' from the table illustrated in FIG. 2, and FIG. 4B illustrates content usage information and preference information of 'user 3' from the table illustrated in FIG. 4C illustrates an example of separating content usage information and preference information of 'user 4' from the table illustrated in FIG. 2.

When the content usage information and the preference information for each user are separated as described above, the content preference calculator 120 determines the attributes of the users from the user information management table as illustrated in FIG. 5, and then the contents for the users having the same specific user attributes. The content usage information and preference information management table for each user attribute is generated by merging the usage information and the preference information table.

6A and 6B illustrate examples of the content usage information and preference information management table for each user attribute generated by the above process. That is, FIG. 6A illustrates a table managing content preferences of women in Seoul / Gyeonggi area, and FIG. 6B illustrates a table managing content preferences of men in Chungcheong area.

As such, the user-specific content preference calculation result calculated by the user-specific content preference calculation unit 120 is stored / managed in the user-specific content preference management DB 125.

The content similarity calculating unit 130 analyzes the content usage information of the users stored in the content usage information management DB 115 to determine a content pair that is often used together, and the usage form of the content pair and the content forming the pair. Similarity of the content pairs is calculated by applying weights according to the respective attributes.

To this end, the content similarity calculating unit 130 analyzes the content usage information and determines two content combinations that can be selected from two or more different contents consecutively used by the same user within a preset time interval as similar content pairs. . Similarity is calculated by applying a time weight that is inversely proportional to the interval between use times used by the same user for the similar content pair. For example, if 10 users used 'Content A' and 'Content B' on the same day 30 days ago, and 10 users used 'Content C' 10 days ago and 'Content D' 1 day ago The second case is more recent, but 'Content A' and 'Content B' happen on the same day, so 'Content A' and 'Content B' are more similar than 'Content C' and 'Content D'. It is weighted to be calculated.

In this case, the content similarity calculating unit 130 compares two content combinations that can be selected from among two or more different contents when the number of consecutive use by the same user within a preset time interval is equal to or more than a preset number. It is preferable to determine in pairs. Meanwhile, the content similarity calculating unit 130 determines the similar content pair by clustering, and the clustering method uses all algorithms used in general collaborative filtering such as Pearson correlation, popularity difference method, and cosine similarity method as is well known. It is possible to. Therefore, the detailed description of the clustering process for generating a pair of similar content will be omitted in the present specification.

In addition, the content similarity calculator 130 may calculate similarity by applying property weights according to attributes of each of the contents constituting the similar content pair. That is, the content similarity calculator 130 may apply an attribute weight proportional to the similarity of the categories to which the contents of the similar content pair belong. For example, "Content A", "Content B", "Content C", and "Content D" are works of art, and the similarity between pure "Content A" and "Content B" calculated by the record of conduct is "0.3", When the similarity between 'Content C' and 'Content D' is the same as '0.3', 'Content A' is Western painting, 'Content B' is Oriental painting, 'Content C' is sculpture, and 'Content D' is video art. If so, since there is already known similar property that 'content A' and 'content B' are pictures, the content similarity calculating unit 130 weights the similarity scores of 'content A' and 'content B' to 'content C'. And higher similarity than 'content D'.

Meanwhile, the content similarity calculator 130 filters the similar content pairs when the number of users using the similar content pairs is equal to or smaller than a predetermined minority user determination value. This is to prevent the similarity between contents intentionally calculated by a small number of users, which is one of the methods for calculating the more accurate content similarity. At this time, the content similarity calculating unit 130 compares the attributes of the minority users, and if the attributes are equal to or more than a preset number, the content similarity calculating unit 130 transmits the attributes of the minority users and the filtered content pair information to the content preference calculator for each user, corresponding user. It is desirable to manage the preference information of the group. This is because even a small number of user groups need to consider their propensity if the users have attributes in common with each other.

The similarity calculation results of the content pairs calculated by the content similarity calculator 130 are stored / managed in the similarity data management DB 135.

FIG. 7 is an example of a table for storing / managing such similarity data. The similarity data according to the present invention is a value obtained by adding a time weight and an attribute weight to a general clustering result as described above. It is possible to.

The user recommendation data generation unit 140 uses the user's content preference calculation result stored in the user's content preference management DB 125 and the user's recommendation data using the similarity data of the pairs of contents stored in the similarity data management DB 135. Create At this time, the method for generating the recommendation data for each user may use various known techniques. The user recommendation data generation unit 140 stores the user recommendation data generation result in the user recommendation data management DB 145.

The user-specific recommendation data providing unit 150 provides the user-specific recommendation data generated by the generation unit stored in the user-specific recommendation data management DB 145 to provide the user-specific recommendation data to the user. To pass.

8 is a flowchart illustrating a process of providing recommendation data for each user using a collaborative filtering method according to an embodiment of the present invention. That is, FIG. 8 illustrates an example of a collaborative filtering process by the collaborative filtering system illustrated in FIG. 1.

1 and 8, a collaborative filtering method according to an embodiment of the present invention will be described.

The present invention calculates the user preferences and the similarity of the content according to the user's behavior information and the usage form of the content for each of the content, and generates the recommendation data for each user based on the result. Therefore, the collaborative filtering method of the present invention should be preceded by a process (not shown) for receiving and storing content usage information of users from a content providing server.

When the content usage information is stored, the collaborative filtering system 100 analyzes the content usage information to determine user's behavior and / or usage time information for each content, and according to the user's behavior and / or usage time information. The user preference for each content is calculated by applying the weight (S100). The content usage information is analyzed to determine a content pair that is often used together, and the similarity of the content pairs is calculated by applying a weight according to the usage form of the content pair and the attributes of each of the contents of the pair. (S200).

When the user preferences and the similarity of the content pairs are calculated as described above, the collaborative filtering system 100 generates user-specific recommendation data using the user preferences and the similarity of the content pairs (S300). When the content recommendation data providing condition for each user is satisfied, the content recommendation data is provided to the corresponding user (S400). At this time, the content recommendation data providing condition is preferably when the user accesses the content providing server.

9 is a flowchart illustrating a process of calculating content preference for each user according to an embodiment of the present invention. Referring to FIG. 9, the content preference calculation process S100 illustrated in FIG. 8 is as follows.

First, the collaborative filtering system 100 illustrated in FIG. 1 analyzes pre-stored content usage information and stores user's behavior information for each content (S110). That is, the user's behavior information for each content is determined and the result is stored. In addition, an action preference to which an action weight is applied according to the action type is calculated (S120). To this end, the collaborative filtering system 100 preferably applies a preset preference weight for each action to each of the determined user's actions to perform the user's action preference for the corresponding content.

In addition, the user preference calculation process for each content of the present invention calculates the time preference to which the time weight is applied according to the content use time (S130). To this end, the collaborative filtering system 100 of the present invention may determine the use time information of each content, and calculate a time preference for each user for the corresponding content by applying time weights according to the use time information.

The user preference for the corresponding content is calculated using the behavior preference for each user and the time preference for each user.

Meanwhile, the collaborative filtering system 100 of the present invention, which calculates user preferences as described above, filters the abnormal user information (S140) and analyzes the content preferences of the user attributes (S150). In this case, the process of filtering abnormal user information and analyzing content preference for each user attribute is similar to that mentioned in the operation description of the content preference calculator 120 for each user with reference to FIG. 1, and thus a detailed description thereof will be omitted.

The content preference calculation process for each user of the present invention has an advantage of calculating accurate recommendation data by considering content usage tendency of a small group while ignoring content usage history information of an abnormal user.

10 is a flowchart illustrating a content similarity calculation process according to an embodiment of the present invention. Referring to FIG. 10, the process of calculating the similarity degree (S200) of the content pairs illustrated in FIG. 8 is as follows.

First, as illustrated in FIG. 1, the collaborative filtering system 100 analyzes pre-stored content usage information, and selects two content combinations among two or more different contents consecutively used by the same user within a preset time interval. Is determined as a similar content pair (S210).

In addition, a time weight is inversely proportional to an interval between times of use for the clustered content pairs (S220), and an attribute weight value corresponding to the attribute is assigned to the clustered content pairs (S230). Calculate the similarity. In this case, a detailed description of the clustering process (S210) and the time and attribute weight application process (S220, S230) are similar to those mentioned in the description of the operation of the content similarity calculator 130 with reference to FIG. Omit.

In this case, the collaborative filtering system 100 of the present invention, which calculates the similarity as described above, filters the content pairs when the number of users using the content pairs classified as the similar content pairs is less than or equal to a predetermined minority user determination value (S240). . This is to prevent the similarity between contents intentionally calculated by a few users differently from the actual one.

Meanwhile, the collaborative filtering system 100 of the present invention compares the attributes of the minority users and transmits the attributes of the minority users and the filtered content pair information to the content preference calculator for each user when the attributes are equal to or more than a preset number. It is preferable to manage the preference information of the corresponding user group. This is because even a small number of user groups need to consider their propensity if the users have attributes in common with each other.

In this case, the detailed processing of the filtering is similar to that mentioned in the description of the operation of the content similarity calculator 130 with reference to FIG. 1, and thus the detailed description thereof is omitted.

In the above, specific preferred embodiments of the present invention have been illustrated and described. However, the present invention is not limited to the above-described embodiments, and various modifications can be made by any person having ordinary skill in the art without departing from the gist of the present invention attached to the claims. .

As described above, the present invention generates user recommendation data using content preferences and similarities between users, and calculates user's attributes, contents' attributes, and user's content use time when calculating user's content preferences and similarities. Reflect. That is, the present invention applies the weight using the content usage information and the user attribute information of the user and filters the user's behavior determined to be not a universally valid user. Because of this, the present invention has the advantage of increasing the accuracy of the recommendation. In particular, it has the advantage of providing accurate recommendation results to a small group of users. It also has the effect of extending the recipient of the recommendation.

Claims (36)

In the collaborative filtering method, Receiving and storing content usage information of users from a content providing server; Analyzing the content usage information to determine the user's behavior and / or use time information for each content, and calculating the user's preference for each of the content by applying a weight according to the user's behavior and / or use time information process; Analyzing the content usage information to determine a content pair that is often used together, and calculating similarity of the content pairs by applying weights according to the usage form of the content pair and the attributes of each of the contents of the pair; ; And And generating recommendation data for each user by using the user preference and the similarity of the content pair. The method of claim 1, wherein the user's preference calculation process is performed. Analyzing the content usage information to determine a user's behavior for each content; Calculating a user's behavior preference for the corresponding content by applying a preset weight of preference for each behavior to each of the determined actions of the user; Determining time usage information of each of the contents and calculating time preference for each user by applying time weights according to the usage time information; And And calculating a user preference for the corresponding content by using the action preference for each user and the time preference for each user. The method of claim 2, wherein the preference calculation process for each user is performed. And filtering the abnormal user information based on a predetermined abnormal user determination criterion. The method of claim 3, wherein the filtering of abnormal user information is performed. And determining a user who satisfies a predetermined abnormal user attribute as an abnormal user to determine an abnormal user. The method of claim 4, wherein the abnormal user attribute is Collaborative filtering method, characterized in that the user's occupation is an occupation that performs at least one of the action of creating, developing, processing, selling, advertising the content. The method of claim 4, wherein the abnormal user attribute is Collaborative filtering method, characterized in that the nationality of the user is a country other than the country of sale of the content. The method of claim 4, wherein the abnormal user attribute is A method for collaborative filtering, characterized in that it is determined that a user exhibiting a behavior of approaching the same content more than the maximum number of times within a minimum access time interval or a user performing the same behavior more than the maximum number of times within a minimum execution time interval is an abnormal user. The method of claim 7, wherein the abnormal user attribute is And determining the user as an abnormal user when the target of the same action is an unspecified number of users. The method of claim 2, wherein the preference calculation process for each user is performed. And separating the preference calculation result for each user using a user attribute and calculating a preference for each user attribute. The method of claim 9, wherein the calculating of the content preference for each user attribute comprises And separating the preference calculation result for each user based on a combination of at least one user attribute among gender, age, and region. 10. The method of claim 2 or 9, wherein the time weighting step is And a time weight that is inversely proportional to the interval between the most recent use time of the corresponding content and the current time. The method of claim 1, wherein the similarity calculation process Analyzing the content usage information to determine a combination of two contents that are selectable among two or more different contents consecutively used by the same user within a preset time interval as a similar content pair; Applying a time weight inversely proportional to an interval between usage times used by the same user for the pair of similar contents; And And applying an attribute weight according to an attribute of each of the contents constituting the pair of similar contents. The method of claim 12, wherein the similarity calculation process And filtering the content pairs when the number of users using the content pairs is less than or equal to a predetermined minority user determination value. The method of claim 13, wherein the similarity calculation process is performed. And comparing the attributes of the minority users and managing the filtered content pairs as preference information of the corresponding user group when the attributes are equal to or more than a preset number. The method of claim 12 or 13, wherein determining the similar content pair If two or more different contents are consecutively used by the same user within a preset time interval, the collaborative filtering method comprising determining a combination of two contents that are selectable among the contents as a pair of similar contents. . The method of claim 12 or 13, wherein the time weighting step Determining an interval between usage time points used by the same user for the pair of similar contents; Selecting time weights corresponding to the intervals between the use time points among predetermined time weights in inverse proportion to the time intervals at predetermined time intervals; And And applying the selected time weight. The method of claim 12 or 13, wherein the attribute weighting step And applying an attribute weight proportional to the similarity of the category to which the contents belong. In a collaborative filtering system, A content providing server interface unit configured to receive content usage information of users from the content providing server through an interface with a content providing server providing content services to users; A content usage information management database unit which stores / manages content usage information of the users; A content preference calculator for each user, which analyzes content usage information of users stored in the content usage information management database unit and calculates content preference for each user to which a weight is applied according to user's behavior and / or usage time information for each content; Analyzes content usage information of users stored in the content usage information management database unit to determine a content pair that is frequently used together, and applies a weight according to the usage form of the content pair and the attributes of each of the paired contents. A content similarity calculation unit configured to calculate similarity of the content pairs; And And a user-specific recommendation data generator configured to generate recommendation data for each user by using the user's content preference calculation result of the preference calculator and the similarity calculation result of the content pairs of the content similarity calculator. The method of claim 18, A content preference management database unit for storing the content preference calculation result for each user; A similarity management database unit for storing similarity calculation results of the content pairs; And And a user-specific recommendation data management database unit for storing the user-specific recommendation data. The method of claim 18 or 19, Further comprising a user-specific recommendation data providing unit for transmitting the user-specific recommendation data to the content providing server through the interface to provide the user-specific recommendation data generated by the user-specific recommendation data generation unit to the corresponding user. Collaborative Filtering System. 19. The method of claim 18, wherein the content preference calculator for each user The content usage information is analyzed to determine the user's behavior for each content, and the user's behavior preference is calculated by applying a preset preference weight for each behavior to each of the determined user's behaviors, and including the user's behavior preference. Collaborative filtering system, characterized in that for calculating the content preference. 19. The method of claim 18, wherein the content preference calculator for each user Analyzing the content usage information to determine the use time information of each of the contents, calculates a time preference for each user by applying a time weight according to the use time information content including the time preference for each user Collaborative filtering system, characterized in that to calculate the preference. 19. The method of claim 18, wherein the content preference calculator for each user And filtering the abnormal user information based on a preset abnormal user determination criterion to calculate the preference for each user. 24. The method of claim 23, wherein the content preference calculator for each user Collaborative filtering system, characterized in that the user's job is a job that performs at least one of the creation, development, processing, sales, advertising of the content. 24. The method of claim 23, wherein the content preference calculator for each user If the nationality of the user is a country other than the country of sale of the content, the collaborative filtering system, characterized in that the user is determined to be an abnormal user. 24. The method of claim 23, wherein the content preference calculator for each user Collaborative filtering system, characterized in that the user showing the behavior of accessing the same content more than the maximum number of times within the minimum access time interval or the user performing the same behavior more than the maximum number of times within the minimum execution time interval as an abnormal user. 27. The method of claim 26, wherein the content preference calculator for each user And the user is determined to be an abnormal user when the target of the same action is an unspecified number of users. 19. The method of claim 18, wherein the content preference calculator for each user And separating the preference calculation result for each user using a user attribute and calculating a preference for each user attribute. 29. The method of claim 28, wherein the content preference calculator for each user And a result of calculating preference preference for each user based on a combination of at least one user attribute among gender, age, and region. 19. The method of claim 18, wherein the content preference calculator for each user Collaborative filtering system, characterized in that to calculate the content preference for each user by applying a time weight inversely proportional to the interval between the most recent use time of the content and the current time. 19. The method of claim 18, wherein the content similarity calculating unit The content usage information is analyzed to determine two content combinations that are selectable among two or more different contents consecutively used by the same user within a preset time interval as the similar content pairs, and then to the similar content pairs by the same user. And a similarity applied to a time weight that is inversely proportional to the interval between the used points of view. 32. The method of claim 31, wherein the content similarity calculator The collaborative filtering system of claim 2, wherein two or more different contents are consecutively used by the same user within a preset time interval, and a combination of two contents that are selectable among the contents is determined as a pair of similar contents. . 19. The method of claim 18, wherein the content similarity calculating unit And calculating similarity by applying attribute weights according to attributes of each of the contents constituting the pair of similar contents. The method of claim 33, wherein the content similarity calculating unit And a property weight that is proportional to the similarity of the categories to which the contents of the similar content pair belong. 19. The method of claim 18, wherein the content similarity calculating unit And if the number of users using the similar content pair is less than or equal to a predetermined minority user determination value, filtering the similar content pair. 36. The method of claim 35, wherein the content similarity calculator When the attributes of the minority users are compared and the attributes are equal to or more than a preset number, the attributes of the minority users and the filtered content pair information are transmitted to the content preference calculator for each user to be managed as the preference information of the corresponding user group. Collaborative filtering system.

Images