KR20140140309A - Method for calculating simularity between users and item recommendation method using the same - Google Patents

Abstract

Provided are a method of calculating similarity between users using social relation weightings of users and context information, and an item recommending method using the calculated similarity between users. A method of calculating similarity between users according to an embodiment of the present invention comprises: a user data storing step of storing data related to a user including a social relation weighting of the user; and an inter-user similarity calculating step of calculating the similarity between users by using at least a portion of the stored user data. A similar user group is generated by using the calculated inter-user similarity, and for items without evaluation scores, scores are estimated by using an item-based cooperation filtering based on recommendation scores of similar users in the similar user group where the user belongs, and the estimated values are provided to the items without evaluation scores, such that similar items can be recommended. According to the invention, social relations between users can be reflected on while calculating the similarity between users, thereby calculating similarity between users more accurately.

Description

[0001] The present invention relates to a method for calculating similarity between users and an item recommendation method using the same,

The present invention relates to a method of calculating similarity between users and an item recommendation method using the same. More particularly, the present invention relates to a method of calculating similarity between users using social relation weights and context information of a user, And a method for recommending an item such as a commodity.

With the development of information technology, information retrieval through the Internet is widely used. However, since a large amount of information is provided without being verified, not only the reliability of the information on the online due to unvalidated information is worried but also a considerable time There is no choice but to take this. In order to compensate for this, personalization services that provide valuable data to users have received attention recently.

The recommendation system, which is one of the personalized services, is a system that induces interest of a user (hereinafter, referred to as a 'recommended target user') by recommending information that each user may be interested in from a large amount of information. In order to recommend the most suitable items to the recommendation target user, the recommendation system may include a recommendation target user who has previously selected a recommendation target user's history, a user profile such as interest of the recommendation target user, age, gender, Etc., and provides appropriate recommendation information to the user.

For example, in Korean Patent No. 7,207,612, in addition to the keywords given as user static information, the user's situation information is determined based on the user's dynamic information such as the user's current location, place, search intention, etc., The similarity degree w of the content (or item) is calculated, and a content to be recommended to the user is selected based on the similarity degree of the content.

However, this system considers only the static and dynamic information of the user and does not consider how other users judge the recommended content, in particular, how the user similar to the user is evaluating the content The reliability of the content recommendation is reduced.

Reflecting this point, a user - based collaborative filtering recommendation system that recommends items using similarity between users has been proposed. The user - based collaborative filtering system is a recommendation system that can predict the user 's preference about items by using the preference information obtained from many users.

In other words, it is a method using the fact that the accuracy of item recommendation can be improved by utilizing the information evaluated in advance by a user having a similar tendency to the target user. However, in order for this method to be effective, there is a need for a method that can more accurately evaluate the similarity between users.

In order to more accurately evaluate the similarity between users, Korean Patent Laid-Open Publication No. 2010-0086296 discloses a method of correcting the similarity between a recommendation target user and another user by using the relationship between the recommendation target user and another user's preference information entropy as a weight Korean Patent Laid-Open Publication No. 2004-0042091 discloses a method for improving the accuracy of user's similarity measurement by calculating the weights necessary for calculating the user similarity degree using the commonality and the link strength information of the tag information of the contents posted by users .

However, the conventional method of calculating the similarity between users only uses only the profile of the user, the relationship between the entropy of each user's preference information, the tag information, and the link information of the contents, and does not consider the social relationship between the users at all. There is a problem in that the accuracy and reliability of the recommendation are deteriorated.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a method of deriving similar users by reflecting user context information and social relations, and a method of recommending items using recommendation scores of similar users derived using the method The purpose is to provide.

According to an embodiment of the present invention, there is provided a method of calculating similarity between users, the method comprising: storing user data including a user's social relation weights; calculating a degree of similarity between users using at least some of the stored user data; And calculating the degree of similarity between the users.

The social relationship weights can be determined by considering whether the user has commented on the postings of other users. Alternatively, the social relation weights may be SNS Depth, which represents the connection depth of other users' comments on posts posted by the user to the SNS. SNS debs with user A of user B who has commented directly on SNS uploaded by user A is 1dept and SNS debut with user A of user D who has commented to user C with SNS depth N with user A is N + 1 . Alternatively, the social relationship weights can be determined by considering whether the user is registered as a neighbor or friend of another user's blog or SNS.

The data on the user may include user context information and user preference data. The user context information includes the user's sex, residence area, and educational background. In addition, the user context information may include contextual information related to the service provided. The user's preference data is the score the user rated for one or more items.

In the user similarity calculation step, the similarity between users can be calculated using the degree of matching of the user context information and the social relation weights.

The data about the user is preferably stored in a Semantic Ontology Triple form.

When a similarity degree between users is calculated, a similar user group is defined as a group of users whose degree of similarity between users satisfies a predetermined criterion. Then, based on the recommendation scores of similar users in the similar user group to which the user belongs, And a similar item is recommended by giving a predicted score to the item whose score is missing.

The similar user group creation step may include a first step of selecting one user not included in any similar user group and creating a similar user group having the same user, To the group.

The similar item recommendation may comprise constructing a user-item table for items for users in the similar user group, and applying a collaborative filtering algorithm to the configured user-item table to generate a prediction score And recommending to the user an item having a predicted score that is equal to or higher than a predetermined criterion.

A weighted slope-one algorithm may be used as the cooperative filtering algorithm.

In the present invention, the degree of similarity between users can be more accurately calculated by reflecting the social relations between users in calculating the similarities between users. In addition, since the items are recommended by utilizing the item evaluation scores of users having higher similarity, the accuracy and reliability of item recommendation can be improved.

1 is a block diagram illustrating an internal configuration of an item recommendation system according to a preferred embodiment of the present invention.
FIG. 2 is a flowchart illustrating a similarity calculation procedure between users according to a preferred embodiment of the present invention and an item recommendation procedure using the same.
Figure 3 is an example of a user-comment table.
4 is an example of the SNS depth table.
5 is an example of a user-item table.

Hereinafter, a method of calculating similarity between users according to a preferred embodiment of the present invention and an item recommendation method using the same will be described with reference to the drawings.

1 is a block diagram illustrating an internal configuration of an item recommendation system according to a preferred embodiment of the present invention.

The data collecting unit 110 receives context information such as the user's sex, residence area, education, hobby, occupation, marital status, interest field, etc. from the user or collects such information from other sites. In addition, the data collection unit 110 may receive context information related to the provided service or may be provided from other sites. Context information related to the provided service refers to context information related to a movie in the case of a site providing movie information, for example, a favorite movie genre, a favorite actor or director, etc. Context information.

The data collection unit 110 also receives or collects the user's preference data. The user's preference data refers to a score the user has rated for one or more items (or content). In the case of an information providing site such as a movie, music, or book, a rating evaluated by the user for each movie / music / book may be included. The collected information is stored in the user DB 150.

The data collecting unit 110 also collects information (social relationship information) necessary for calculating a user's social relation weights. The user's social relationship information refers to information about other users' comments, 'likes', and affirmative / negative responses to articles posted on a social network system (SNS), a blog or a cafe. In addition, friend or follow relationship information of the SNS, a phone number or an e-mail address stored in the phone book of the user's mobile phone or a web site, and the like can be included in the social relation information of the user. This information may be present within the item recommendation system or may be present in an external separate system (e.g., a facebook site). External system access information (URL, ID, password) of the external system may be stored in the user DB 150 in order to access the external system and retrieve such information.

The ontology transformation unit 120 transforms user context data of the user stored in the user DB 150, social relationship information, and user preference data for each item into a semantic ontology triple form, 160). It is desirable to convert social relationship information into social relation weights and store them. The social relationship weight can be, for example, the SNS deprecated to indicate the connection depth of a user's comment on a post on a social network system (SNS). An embodiment for calculating the SNS depth will be described later.

The similarity calculation unit 130 calculates similarities between users based on inferences based on gender, age, region, and the like in the ontology triple DB 160 storing user context data, social relationship information, and item preference data for each item Derives a group of similar users, and stores the derived similar user group information in the similar user group DB 170. A method of calculating the similarity between users and a method of deriving similar user groups will be described later.

The item recommendation unit 140 fetches the user preference information stored in the ontology triple DB 160 for similar users belonging to the similar user group to which the recommendation target user belongs in the similar user group DB 170, The most suitable item is selected and presented to the user. Item recommendation method will also be described later.

Next, referring to FIG. 2, a similarity calculation procedure between users according to a preferred embodiment of the present invention and an item recommendation procedure using the same will be described.

The item recommendation method of FIG. 2 includes a user data storing step (S100) of storing data related to a user including a social relation weight of a user, an inter-user similarity calculating step of calculating a degree of similarity between users using at least some of the stored user data A similar-user-group creation step (S300) of creating a similar-user group, which is a group of users whose similarity between users satisfies a predetermined criterion, a calculation step (S200) (S400) of recommending a similar item by assigning a predicted score to an item in which an evaluation score is missing by predicting by item-based collaborative filtering.

In step S100, the user-related data including the user's social relation weights collected by the data collection unit 110 is converted into a form that is convenient for use at a later stage, such as a semantic ontology triple form, in the ontology conversion unit 120 .

The social relation weight of the user stored in step S100 may be an SNS attribute indicating a connection depth of a comment of another user with respect to an article posted on a social network system by a user. For example, SNS debs with user A of user B who has commented directly on SNS uploaded by user A are used 1 deb, and 2 deb writes with user A of user C who has commented to user B. When this is generalized, the SNS debut with the user A of the user E who has made the comment to the user D having the SNS depth N with the user A becomes N + 1. In this way, the SNS depth between two users can be traced. Since too much deb write is not significant, it is desirable to limit the maximum SNS depth to be tracked. For example, the system can be configured to track up to six depths. Here, SNS may include blogs, Internet cafes, etc. in addition to general SNS such as Facebook.

In addition, various other modifications can be used to obtain the SNS depth. For example, the SNS depth between users who have commented on each other may be set to 1, and the SNS depth between users who have left comments only on one side may be set to 2. It is also possible to reduce the calculated SNS depth by one level for the person registered in the telephone directory of the user, or to set it to at least two depths. It is also possible to assign SNS depth to a person who is registered as a neighbor or friend of a blog or SNS, regardless of whether or not the comment is added, and to assign a SNS depth of 2 to a friend of a friend . It is also possible to allow the user to select the type of SNS used to calculate the social relation weights. Various SNS depth calculation methods are possible as described above, and the present invention is not limited to the specific method of calculating the SNS depth.

One way to obtain the SNS depth is illustrated by way of example.

First, a user-comment table is created to obtain the SNS depth between users who have made comments. The user-comment table can be configured as shown in FIG. 3, where 1 indicates a case where a comment is left, and 0 indicates a case where no comment is left.

Next, referring to the generated table, a SNS depth between users who have not left a comment is generated as follows.

For example, in seeking the SNS depth between user 1 and user 2, the user 1 and user 2 have not commented on each other, so the value of the user-comment table is 0, but when user 1 commented on user 3, 3 has commented on User 2, set the SNS Depth between User 1 and User 2 to 2. In addition, since the user 4 has commented on the user 2 between the user 1 and the user 4 who have not commented on each other, 3, which is obtained by adding 1 to the SNS deprecated 2 between the user 1 and the user 2, SNS depth value.

FIG. 4 shows an example of the SNS depth table generated through the above process. This process is maintained at the latest value whenever the user comments or continuously updates at a certain period.

On the other hand, when calculating the similarity of another user to the user A in calculating the SNS depth, it is also possible to ignore the comment left by the other user to the user A and to calculate the SNS depth by considering only the comment left by the user A It is possible.

The social relation weights can be used as they are, or they can be transformed into social relation weights using conversion tables. For example, in one embodiment of the present invention, the inverse of the SNS depth value is used as the social relation weight. That is, if the SNS depth is 1, the social relation weight is 1, and if the SNS depth is 2, the social relation weight is 0.5.

In addition to the social relation weights, the user-related information stored in step S100 may further include user context information and user preference data.

 The user context information may include various information regarding the user such as the user's sex, residence area, education, hobby, occupation, marital status, area of interest, and the like. Depending on the characteristics of the site, the user context information to be stored may be different.

 In addition, the user context information may include contextual information related to the service provided. Context information related to the provided service refers to context information related to a movie in the case of a site providing movie information, for example, a favorite movie genre, a favorite actor or director, etc. Context information. Examples of such context information may include a favorite genre and a favorite musician (artist) in the case of a music or book-related site. In the case of a photo site, a camera model to be used and a type of photograph (portrait, landscape, In the case of sites related to the science fair such as the goldberg competition, the materials (dice, dolls, beads, etc.) and scientific principles (kinetic energy, location energy, Etc.) may be included.

The user's preference data refers to a score the user has rated for one or more items (or content). In the case of an information providing site such as a movie, music, or book, a rating evaluated by the user for each movie / music / book may be included.

Thus, it is preferable that the user-related data stored in step 100 is stored in a semantic ontology triple form. The ontology transformation unit 120 performs ontology modeling using, for example, modeling based on RDBMS (Relational Data Base Management System) modeling, and provides the object data to the object data through the ontology modeling rule as <Subject, Predicate, Object >, And save it.

In the user similarity calculation step (S200), it is preferable that the similarity between users is calculated by using the degree of matching of the user context information and the social relation weights. For this purpose, for example, the following formula can be used to calculate the similarity Sim (U1, U2) between user 1 (U1) and user 2 (U2)

Sim (U1, U2) = (Sex + Education + Area + Generation + SRC + SRW) / 6

In this case, Sex is the gender match (1 = matched), Education is the difference of education (same = 1, 1 = 0.8, 2 = 0.6, ...) SRC is a service-related context in which the age-related differences (age <5 years = 1, 5 to 10 years = 0.7, 11-20 years = 0.4, and more = 0) And SRW are social relations weights.

The above formula is illustrative and can be appropriately modified in accordance with the characteristics of a service, a site, a subscription member, etc. provided. It is also possible to configure each user context information to have a weight. For example, in sites where user's academic background is important, we can assign weight to academic achievement, and in a site where user's age is important, we can assign weight to age group.

In this way, the degree of similarity between each user is calculated for all users. Alternatively, in order to reduce the amount of calculation, it is also possible to calculate the similarity only among users whose social relation weights are more than a certain level, or to calculate the similarity only between users whose specific context information (for example, high-weighted context information) .

On the other hand, it is possible to calculate asymmetrically in calculating the similarity between users. That is, the degree of similarity of user A to user B and the degree of similarity of user B to user A are separately calculated. For example, when calculating the social relation weights of user A with respect to user B, only the comments that user B left in the posting of user A are considered. In the case where the user A does not consider comments left in posting by user B, This asymmetry can occur if one is a famous person and the other one is unilaterally reading the person's post.

When the degree of similarity between users is calculated, a similar user group is created using the calculated similarity degree (step S300). The similar user group can be grouped into the users whose similarities between users are equal to or higher than a certain score. An example of such a grouping method is to select a single user (user A) not included in any similar user group to create a similar user group having the user A, and to assign a user satisfying a predetermined condition to the user A There is a way to add to a user group. The predetermined condition may be a predetermined number of users (for example, 10 users) having the highest degree of similarity to the user A, a user having a similarity degree of the user A having a predetermined value (for example, 0,95) . Alternatively, when the number of users whose similarity degree is equal to or higher than a predetermined value is equal to a predetermined maximum number, for example, 100 or more, to user A, only up to 100 users can be added to one similar user group.

As another method, there is a method of selecting and storing a predetermined number of users most similar to the user for each user. That is, for example, 10 persons who are most similar to user A are found and stored in association with user A. This is useful in the case of embodiments that calculate the degree of similarity asymmetrically.

There are various methods for creating the similar user group, and the present invention is not limited to the specific similar user group creation method.

In the similar item recommendation step (S400), an item-based collaborative filtering is predicted based on a recommendation score of similar users in a similar user group to which the user belongs, and a similar item is recommended by giving a predicted score to the item will be. In an exemplary embodiment of the present invention, the similar item recommendation step (S400) includes the steps of: constructing a user-item table for items for users in the similar user group; applying a collaborative filtering algorithm to the configured user- A step of assigning a predictive score to an item which is not evaluated by the user, and a step of recommending to the user an item having a given predictive score equal to or higher than a predetermined criterion.

The user-item table is a table in which users who are similar to a specific user, that is, users belonging to a similar user group, as shown in FIG. The question mark in FIG. 5 indicates that the user has not evaluated the item.

The collaborative filtering algorithm uses the score of user A on two items i, j and the score of user B on item i, As an algorithm for predicting the score, a typical cooperative filtering algorithm is a Slope-One algorithm. Hereinafter, a method of assigning item prediction scores using a weighted slope-one scheme, which is one of the slope original algorithms, will be described.

The weighted slope original algorithm calculates a deviation of all the pairs of items and a step of assigning a prediction score to an item that the user has not evaluated using the calculated deviation to give a prediction score do.

Sura card user given a rating for both _{(S j, i (X)} ) the item i, j, and, u _i and Let u _j be the evaluation score of the user's items i and _j , respectively, and u ∈ S _{j , i} ( X ) be all users belonging to the similar user group. The deviation of the item pairs can be calculated by Equation (1).

Thus, if the deviation of the item pairs is calculated for all pairs of items, Equation 2 is used to give a predictive score for the item that the user has not evaluated.

Here, c _{j , i} = card ( S _{j , i} ( X )) and i ∈ S (u) - {j} denote all items to which the user has given an evaluation score.

In the above, a method of assigning a predictive score to an item not evaluated by a user by applying a collaborative filtering algorithm is described as an example of a weighted slope original algorithm. However, the present invention is not limited to a specific collaborative filtering algorithm, It is possible to provide a prediction score for an item which is not evaluated by the user by various algorithms.

If a prediction score for an item that is not evaluated by the user is given through the above process, the user is recommended to the user with an item with a predicted score that is equal to or higher than a predetermined criterion, thereby recommending the item that the user is most interested in.

While the present invention has been described with reference to exemplary embodiments, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. That is, within the scope of the present invention, all of the components may be selectively coupled to one or more of them. In addition, although all of the components may be implemented as one independent hardware, some or all of the components may be selectively combined to perform a part or all of the functions in one or a plurality of hardware. As shown in FIG. The codes and code segments constituting the computer program may be easily deduced by those skilled in the art. Such a computer program can be stored in a computer-readable storage medium, readable and executed by a computer, thereby realizing an embodiment of the present invention. The storage medium of the computer program may include a semiconductor recording medium, a magnetic recording medium, an optical recording medium, a carrier wave medium, and the like.

Furthermore, the terms "comprises", "comprising", or "having" described above mean that a component can be implanted unless otherwise specifically stated, But should be construed as including other elements.

The foregoing description is merely illustrative of the technical idea of the present invention and various changes and modifications may be made by those skilled in the art without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are intended to illustrate rather than limit the scope of the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. The scope of protection of the present invention should be construed according to the following claims, and all technical ideas within the scope of equivalents should be construed as falling within the scope of the present invention.

110 data collection unit, 120 ontology conversion unit,
130 similarity calculation unit, 140 item recommendation unit,
150 user DB, 160 ontology triple DB,
170 similar user group DB.

Claims (23)

A user data storing step of storing data on a user including a user's social relation weight,
An inter-user similarity calculation step of calculating a degree of similarity between users using at least a part of the stored user data;
A similar user group creation step of creating a similar user group, which is a group of users whose similarity between users satisfies a predetermined criterion;
Based on the recommendation score of similar users in the similar user group to which the user belongs, it is predicted by item-based collaborative filtering, and a similar item recommending step of recommending a similar item by giving a predicted score to the item in which the evaluation score is missing
&Lt; / RTI &gt;
The method according to claim 1,
Wherein the social relation weight is an SNS depiction indicating a connection depth of a comment of another user with respect to an article posted by the user on the SNS.
3. The method of claim 2,
SNS debs with the user A of the user B who directly commented on the SNS uploaded by the user A are used for 1 deb, and the SNS debut with the user A of the user D who has made the comment to the user C with the SNS depth N with the user A is N + The item recommendation method comprising:
3. The method of claim 2,
Wherein the data on the user further includes user context information and user preference data.
5. The method of claim 4,
Wherein the user context information includes a sex, a residence area, and an education level of a user.
5. The method of claim 4,
Wherein the user context information includes contextual information related to a provided service.
5. The method of claim 4,
Wherein the user's preference data is a score the user has rated for one or more items.
5. The method of claim 4,
Wherein the degree of similarity between users in the step of calculating the degree of similarity between users is calculated using the degree of matching of the user context information and the weight of the social relation.
9. The method according to any one of claims 1 to 8,
Wherein the data about the user is stored in a semantic ontology triple form.
9. The method according to any one of claims 1 to 8,
A first step of selecting one user not included in any similar user group and creating a similar user group having the same user,
A second step of adding the users satisfying a predetermined condition to the similar user group to the user
&Lt; / RTI &gt;
9. The method according to any one of claims 1 to 8,
Constructing a user-item table for items for users in the similar user group,
Applying a collaborative filtering algorithm to a configured user-item table to give a predictive score for an item that the user has not evaluated;
A step of recommending to the user an item whose predicted score is equal to or greater than a predetermined criterion
&Lt; / RTI &gt;
12. The method of claim 11, wherein the collaborative filtering algorithm is a Weighted Slope-One algorithm.
A user data storing step of storing data on a user including a user's social relation weight,
Calculating user similarity using at least a part of the stored user data,
And calculating a similarity between users.
14. The method of claim 13,
Wherein the social relation weight is an SNS depiction indicating a connection depth of a comment of another user to a user's post on the SNS.
15. The method of claim 14,
SNS debs with the user A of the user B who directly commented on the SNS uploaded by the user A are used for 1 deb, and the SNS debut with the user A of the user D who has made the comment to the user C with the SNS depth N with the user A is N + And calculating the similarity between users.
14. The method of claim 13,
Wherein the social relation weights are determined by considering whether the user has commented on postings of other users.
14. The method of claim 13,
Wherein the social relation weights are determined in consideration of whether a user is registered as a blog or a neighbor of a user or a friend of the SNS.
18. The method according to any one of claims 13 to 17,
Wherein the data on the user further includes user context information and user preference data.
19. The method of claim 18,
Wherein the user context information includes a sex, a residence area, and a scholastic ability of a user.
19. The method of claim 18,
Wherein the user context information includes context information related to a provided service.
19. The method of claim 18,
Wherein the user's preference data is a score the user has rated for one or more items.
19. The method of claim 18,
Wherein the degree of similarity between users in the step of calculating the degree of similarity between users is calculated using the degree of matching of the user context information and the weight of the social relation.
18. The method according to any one of claims 13 to 17,
Wherein the data about the user is stored in a semantic ontology triple form.

Images