KR100631806B1 - Method and system for music recommendation by using graph representing relations between musics - Google Patents

Abstract

A method and a system for recommending music using a graph representing relation among the music are provided to quickly and efficiently recommend the music suitable for user preference based on a graph model while recommending the music suitable for the user preference by using only the suggestive preference such as preferred music lists and music execution information. A preferred music database(100) stores the preferred music list and the preference information of many users. A correlation database(102) stores a correlation graph including the preference of the preferred music calculated from the preferred music list and the preference information, and correlation among the preferred music. A correlation graph updater(103) updates the correlation graph database by changing the data of the correlation graph if the preference information of the user is changed. A preferred music list extractor(104) extracts the preferred music list for the recommended user from the preferred music database if the recommended user is selected. A recommended music provider(112) provides a recommended music list by using an expected preference value calculated in a preference calculator(110).

Description

Method and System for Music Recommendation by Using Graph Representing Relations between Musics}

1 is a block diagram schematically illustrating a music recommendation system using a correlation graph showing a correlation between music according to a preferred embodiment of the present invention;

FIG. 2A is an example data for illustrating an association graph according to a preferred embodiment of the present invention, and FIG. 2B is an association graph created according to the example data of FIG. 2A;

3 is a flowchart illustrating a process of updating an association graph according to a preferred embodiment of the present invention;

4 is a flowchart illustrating a music recommendation process using an association graph showing an association relationship between music according to a preferred embodiment of the present invention.

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

100: preferred music database 102: association database

103: association graph update unit

104: favorite music list extraction unit 106: adjacent music extraction unit

108: calculator 110: preference calculator

112: recommended music provider

The present invention relates to a method and system for recommending music using a graph indicating a correlation between music. More specifically, the present invention relates to a music recommendation method and system for recommending music according to a user's taste by using the user's favorite music list and preference information. The present invention relates to a method and system for recommending music.

As information technology advances to the information society, the Internet is becoming an important and essential medium in all fields such as society, culture, and economy. Accordingly, users using the Internet are receiving multimedia information such as music or movies using the Internet, and a service for providing such multimedia information is increasing.

The multimedia information service also provides services that analyze and recommend personal tastes for items that are not easy to analyze, such as music or movies. For this purpose, conventionally used methods include a contents-based recommendation method, a demographic recommendation method, and a collaborative filtering recommendation method.

Here, the content-based recommendation method is a method of recommending items similar to the preferred item of the user to be recommended using data describing the content of the item, and the demographic recommendation method is demographic data about the user. It is a method of finding similar users and recommending similar items by using the user, and recommending similar items by the cooperative filtering recommendation method. The method refers to a method of finding a user or finding and recommending an item having a similar preference to that of a user to be recommended. Of these three methods, collaborative filtration recommendation is the most commonly used.

The collaborative filtering recommendation method is to recommend an item to a recommendation target user based on the preference item information of a plurality of users. There are a method of finding a similar user and a method of finding a similar item. The method of finding similar users is to compare the preferences of the items of the recommending user u with the preferences of the items of other users, wherein the similarity between users can be obtained using the Pearson Correlation Coefficient. . Similarity calculation using Pearson's correlation coefficient is shown in Equation 1.

는 사용자 u의 선호도 평균을 의미한다.Where W _{u, w} is the similarity between user u and user w , r _{u, i} is user u 's preference (or evaluation value) for item i ,

Denotes a preference average of user u .

On the other hand, the preference prediction value for the item i of the recommendation user u is calculated as Equation 2 as a weighted average of preferences of similar users.

Here, P _{u, i} is a preference prediction value for the item i of the user u using the similarity between the users.

Meanwhile, a method of finding similar items is to compare the preferred items of the recommended target user with other items by user preference, and the similarity between the items is calculated by Equation 3 by Pearson's correlation coefficient similarly to the above.

는 아이템 i에 대한 선호도 평균을 의미한다.Where W _{i, j} is the similarity between item i and j , r _{u, i} is user u 's preference (or rating) for item i ,

Denotes a preference mean for item i .

On the other hand, the preference prediction value for the item i of the recommendation target user u is calculated as Equation 4 as the weighted average of the preference of similar items.

Where P _{u, i} is a predicted value for the preference the user u of the item i using the degree of similarity between items.

However, this general collaborative filtration recommendation method has a disadvantage in that its application requires an explicit preference for users' items. In addition, when calculating the similarity between users or between items, there is a disadvantage that it takes a lot of time in proportion to the number of users and the number of items. In particular, a new recommendation method is required in order to recommend music in real time using only implicit preferences such as a favorite music list and the number of music executions in a music service site dealing with millions of users and hundreds of thousands of items.

An object of the present invention is to provide a music recommendation method and system for recommending music that matches a user's taste using only implicit preferences such as a user's favorite music list and music execution information. To provide a music recommendation method and a system for performing the same that can recommend the right music quickly and efficiently.

In order to achieve the object as described above, the music recommendation method and system using a graph showing the correlation between music according to the present invention has the following features.

According to an aspect of the present invention, there is provided a music recommendation method using an association graph indicating an association relationship between music, comprising: (a) extracting a favorite music list of a recommendation target user from a favorite music database when a recommendation target user is selected; (b) extracting adjacent music corresponding to adjacent vertices of the preferred music included in the preferred music list extracted in the step (a) with reference to the association graph; (c) extracting a preference for the preferred music with reference to the association graph and calculating an association degree between the favorite music and the adjacent music extracted in step (b); (d) calculating a predicted preference value for the adjacent music using the preference extracted in step (c) and the association calculated in step (c); And (e) providing a recommendation music list to the recommendation target user using the preference prediction value calculated in step (d), wherein the correlation graph includes a plurality of user's favorite music lists and preference information. A music recommendation using a correlation graph indicating an association relationship between music, wherein the association relationship between the preference music of the preferred music and the preferred music calculated by using the mobile device is updated and the preference information of the user is changed. Provide a method.

Also, the present invention provides a music recommendation system using a correlation graph indicating a correlation between music, comprising: a favorite music database for storing a plurality of user's favorite music lists and preference information; An association database that stores the association graph including the association relationship between the preference music list stored in the preference music database, a preference of the preferred music calculated using the preference information, and the preference music; An association graph updating unit for updating the association database by changing data of the association graph when the preference information of the user is changed; A preference music list extracting unit configured to extract a favorite music list of the recommendation target user from the favorite music database when the recommendation target user is selected; An adjacent music extracting unit extracting adjacent music corresponding to adjacent vertices of the preferred music included in the preferred music list extracted by the favorite music list extracting unit with reference to the association graph; A calculator configured to extract a preference for the preferred music by referring to the correlation graph and calculate an association degree between the favorite music and the adjacent music extracted by the adjacent music extractor; A preference calculator configured to calculate a preference prediction value for the adjacent music using the preference extracted by the calculator and the association degree calculated by the calculator; And a recommendation music providing unit providing a recommendation music list to the recommendation target user by using the preference prediction value calculated by the preference calculation unit. to provide.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. First of all, in adding reference numerals to the components of each drawing, it should be noted that the same reference numerals are used as much as possible even if displayed on different drawings. In addition, in describing the present invention, when a detailed description of a related well-known configuration or function is apparent to those skilled in the art, or it is determined that the subject matter of the present invention may be obscured, the detailed description thereof will be omitted.

1 is a block diagram schematically illustrating a music recommendation system using an association graph showing an association relationship between music according to a preferred embodiment of the present invention.

As shown in FIG. 1, a music recommendation system using an association graph showing an association relationship between music according to a preferred embodiment of the present invention includes a preferred music database 100, an association database 102, and an association graph update unit. Reference numeral 103, a preferred music list extractor 104, an adjacent music extractor 106, a calculator 108, a preference calculator 110, a recommended music provider 112, and the like.

The preferred music database 100 according to the preferred embodiment of the present invention stores a plurality of user's favorite music lists and preference information. Here, the preference information refers to information reflecting the number of purchases of music, the number of downloads of music, the number of times of music execution, and the like.

The association database 102 according to the preferred embodiment of the present invention includes association information between the preference music and the preference music of the preferred music calculated using the preference music list and the preference information stored in the favorite music database 100. An association graph is stored according to a preferred embodiment of the present invention. Here, the association between the preferred music means the sum of the frequencies in which two songs are favored simultaneously by the same user.

Meanwhile, in the association graph update unit 103 according to an exemplary embodiment of the present invention, when the user's preference information is changed, such as making a selection list or downloading / executing music, the association relationship stored in the association database 102 is changed. The association database 102 is updated continuously by changing the data in the graph.

FIG. 2A is an example data for illustrating an association graph according to a preferred embodiment of the present invention, and FIG. 2B is an association graph created according to the example data of FIG. 2A.

Referring to FIG. 2A, the user 1's favorite music list is {a, b, c} and the number of executions is {1, 10, 10}. Similarly, user 2's favorite music list is {a, b}, each run number is {1, 1}, user 3's favorite music list is {a, d}, and each run number is {1, 1}. do.

Hereinafter, the preferred information will be described as the execution number. However, as described above, the preferred information according to the preferred embodiment of the present invention is not limited thereto and may correspond to the preference information including the number of music purchases and the number of music downloads. Anything can be anything.

Now each piece of music is called M _i . In the case of FIG. 2A, M _i = {a, b, c, d}. In the graph of FIG. 2B, each vertex M _i represents each music, and the edges M _i and M _j between the vertices represent an association between each music. Each vertex stores the total preferences C ( Mi ) of users for each music, where the preference is a value reflecting the aforementioned preference information. Thus, like the preference information, the preference may be the number of purchases, the number of downloads, the number of executions, or a combination thereof.

As described above, in the examples of FIGS. 2A and 2B, since the preference information is limited to the number of executions, the total preferences are also stored as the sum of the number of executions of FIG. 2A. That is, C (a), which is the total preference of music a, becomes 3, which is the sum of the execution number 1 of user 1, the execution number 1 of user 2, and the execution number 1 of user 3. In addition, C (b), which is the total preference of music b, becomes 11, which is the sum of the number of runs 1 of user 1 and the number of runs 1 of user 2. In this way, the preference total is calculated and stored in the association graph of FIG. 2B.

In addition, each trunk stores a total frequency of simultaneous appearance C ( M _i , M _j ) from the same user of music corresponding to two vertices connected by each trunk. In the present specification, such a co-occurrence total is defined as an association, where the association information is calculated as the sum of pairs of music that are simultaneously favored by the same user.

For example, in the case of FIG. 2A, if user 1 performed music a once, b ten times, and c ten times, C (a) = 1, C (b) = 10, and C (c) = 10 Accumulated at each vertex value, assuming that (a, b) pairs occur once, (a, c) pairs occur once, and (b, c) pairs occur 10 times, C (a, b) = 1, C (a, c) = 1, C (b, c) = 10 is accumulated at each edge value.

As can be seen from Figure 2, the association graph according to a preferred embodiment of the present invention is stored as a list of the number corresponding to the total number of music as an undirected weighted graph (Undirected Weighted Graph).

3 is a flowchart illustrating a process of updating an association graph according to a preferred embodiment of the present invention.

When a user adds or executes a new music M _i to the selection list, the user is selected as the user whose preference is changed (S300), and the user u 's changed music preference value ΔC _u ( M _i ) is selected (S302). .

Next, the association graph update unit 103 changes the total preference C ( M _i ) for the music of the association graph data to C ( M _i ) + ΔC _u ( M _i ) (S304). For all other music M _j in the user's favorite music list, compare C _u ( M _j ) with C _u ( M _i ) (S306), if C _u ( M _j )> C _u ( M _i ) C ( M _i , M _j ) and C ( M _j , M _i ) in the correlation graph data, C ( M _i , M _j ) + ΔC _u ( M _i ) and C ( M _j , M _i ) + ΔC Each change to _u ( M _i ) (S308). If at this time the _{C u (M j)> C} u (M i) The reason for changing the graph data _{C u (M j) <C} u (M i) only if there is increase in C _u (M _i) This is because the frequency of the pair M _i , M _j is not increased.

This operation of FIG. 3 is performed every time a user creates a music selection list or downloads or executes a music, and a graph representing an association between music is created and continuously updated to establish an association according to a preferred embodiment of the present invention. Stored in the database 102. As can be seen in the graph generation process, it can be seen that the larger the value of C ( M _i ) in the correlation graph according to the preferred embodiment of the present invention, the more popular music is favored by many users, and the value of C ( M _i , M _j ) The larger the music is, the more likely it is that both songs will be favored simultaneously.

When the recommendation target user is selected, the preferred music list extractor 104 extracts the preferred music list of the recommendation target user from the preferred music database 100 according to the preferred embodiment of the present invention.

The adjacent music extracting unit 106 according to the preferred embodiment of the present invention refers to the association graph of the association database 102, and determines the preference music included in the favorite music list extracted by the favorite music list extracting unit 104. Extract adjacent music corresponding to adjacent vertices. Herein, the music corresponding to the adjacent vertex of the preferred music vertex in the correlation graph described above may be referred to as the adjacent music.

In the calculation unit 108 according to the preferred embodiment of the present invention, with reference to the association graph of the association database 102, the preference for the preferred music is extracted and the neighboring music extracted from the preferred music and the adjacent music extraction unit 106 is extracted. Calculate the association between music.

In this case, the degree of association can be calculated by using the sum of the frequencies in which the preferred music and the adjacent music are simultaneously favored by the same user, the preferred conditional probability of the adjacent music for the preferred music, or the chi-square of the preferred music and the adjacent music. Can be. Detailed description thereof will be described later with reference to FIG. 4.

The preference calculator 110 according to a preferred embodiment of the present invention calculates a preference prediction value for neighboring music using the preference extracted by the calculator 108 and the correlation calculated by the calculator 108. That is, the preference prediction value is calculated by multiplying the product of the preference extracted by the calculation unit 108 and the association degree calculated by the calculation unit 108 by the number of preferred music and adding the resultant values. A description thereof will be described later with reference to FIG. 4.

The recommendation music provider 112 according to an exemplary embodiment of the present invention provides a recommendation music list to the recommendation target user by using the preference prediction value calculated by the preference calculator 110.

4 is a flowchart illustrating a music recommendation process using an association graph showing an association relationship between music according to a preferred embodiment of the present invention.

First, when the recommendation target user u is selected (S400), the preferred music list extractor 104 according to the preferred embodiment of the present invention extracts the preferred music list of the recommendation target user from the preferred music database 100 (S402). For example, like the audience M ₁ ~ M _n, if n pieces of music stored in the database as preference 100 to preferences of music are extracted the n number of the music list of the M ₁ ~ M _n.

Next, in the neighboring music extraction unit 106 according to the preferred embodiment of the present invention, with reference to the association graph in the association database 102, the adjacent vertices of the preferred music included in the preferred music list extracted in step S402. Adjacent music corresponding to the extracted (S404). Here, the adjacent music corresponds to the music to be recommended, and the music to be recommended has a high correlation with music that is known to be preferred by the user, and they can be obtained from the correlation graph defined above. In other words, the adjacent music is music corresponding to adjacent vertices connected by M ₁ to M _n, and can be directly obtained by reading the adjacent vertex list from each of the vertices M ₁ to M _n . In the following, adjacent music is represented by A ₁ to A _m .

The calculation unit 108 according to the preferred embodiment of the present invention extracts a preference for the preferred music (S406), and calculates an association degree between the preferred music extracted in step S402 and the adjacent music extracted in step S404 (S408). ). In addition, the preference calculation unit 110 according to the preferred embodiment of the present invention calculates a preference prediction value using the preference extracted in step S406 and the correlation calculated in step S408 (S410).

According to a preferred embodiment of the present invention, the preference predicted value P _u ( A _j ) of the recommended target user for the adjacent music A _j extracted in step S404 is determined in step S406 for the preferred music M _i already known to be preferred by the user. The sum of the product of the extracted preference and the correlation between M _i and A _j calculated in step S408 may be expressed by Equation 5.

Where P _u ( A _j ) is the predicted preference value, n is the number of extracted preferred music, C _u ( M _i ) is user u 's preference for M _i , and W ( M _i , A _j ) is M _i and A _It shows the degree of association of _j . The calculation of Equation 5 is performed by the number m of extracted adjacent music.

Here, W ( M _i , A _j ) is C ( M _i , A _j ), which is the sum of the preferred frequencies of the preferred music and the adjacent music simultaneously from the same user, and Prob ( M _i , the preferred conditional probability of the adjacent music with respect to the preferred music. , A _j ), χ ² ( M _i , A _j ), which is the chi-square between the preferred music and the adjacent music, but is not limited thereto. These calculation methods can be calculated directly from the values stored in the correlation graph according to the preferred embodiment of the present invention.

On the other hand, in a preferred embodiment of the present invention proposes a method for calculating W ( M _i , A _j ) assuming MI ( M _i , A _j ). MI ( x, y ) is a value representing how statistically the variables x and y are statistically defined as in Equation 6.

From the MI ( M _i , A _j ) defined in Equation 6 and the values stored in the correlation graph according to the preferred embodiment of the present invention, the value of W ( M _i , A _j ) can be obtained as shown in Equation 7.

Where TEC ( M ) is the sum of all edge values connected to vertex M , and N is the sum of edge values of the entire graph, which is an accumulated value as the correlation graph according to the preferred embodiment of the present invention is accumulated. This is not necessary.

When the preference prediction values P _u ( A ₁ ) to P _u ( A _m ) are obtained through the above-described process, the recommendation music provider 112 arranges adjacent music using the calculated preference prediction value (S412). By selecting the top K pieces of music to be provided, the final recommended music list is provided and recommended (S414).

On the other hand, the use of the present invention greatly reduces the calculation time required to select the music to be recommended as compared to the cooperative filtration recommendation method.

Specifically, the time complexity in the case of directly calculating using the similarity between users or the similarity between items as the preference data according to the cooperative filtration recommendation method, the total number of items I , the total number of users U , the similar selected When the number of users is S , O ( U · I ) is used to calculate the similarity with all users, O ( I · S ) is used to calculate the preference prediction for all items, and O ( I · log I ) is used for sorting. The total execution time is O ( U · I + I · log I ). On the other hand, when using the method proposed in the present invention, the average number of preferred items of each user is P , the average number of edges of each vertex in the graph is E , and the average number of adjacent vertices of each user is A (< P · E ). , to modify the data in the chart O (P), in adjacent music extraction O (P · E), the affinity prediction value calculation O (P · E), in O (a · log a) aligned when the preference information is changed when the The total execution time is O ( P · E + A · log A ) because it takes time.

Comparing the time O (log U · I · I + I) and the other time Ta O (P · A · log E + A) in the method of the invention according to the collaborative filtering recommendation method, generally U · I >> Since I > P · E and I > A , it can be said that the method of the present invention has a relatively very fast execution time.

The above description is merely illustrative of the present invention, and those skilled in the art to which the present invention pertains may various modifications without departing from the essential characteristics of the present invention. Accordingly, the embodiments disclosed herein are not intended to limit the present invention but to describe the present invention, and the spirit and scope of the present invention are not limited by these embodiments. The protection scope of the present invention should be interpreted by the following claims, and all the technologies within the equivalent scope should be interpreted as being included in the scope of the present invention.

As described above, according to the present invention, unlike the conventional methods, implicit preference information can be used by expressing and storing a correlation between music as a graph model, and calculating a calculation time required for selecting a recommendation target music. It has the advantage of greatly shortening. In particular, when targeting millions of users and hundreds of thousands of items, there is an effect that plays a decisive role in increasing the utilization.

Claims (11)

In the music recommendation method using a correlation graph showing the correlation between music, (a) if the recommendation target user is selected, extracting a favorite music list of the recommendation target user from a preferred music database; (b) extracting adjacent music corresponding to adjacent vertices of the preferred music included in the preferred music list extracted in the step (a) with reference to the association graph; (c) extracting a preference for the preferred music with reference to the association graph and calculating an association degree between the favorite music and the adjacent music extracted in step (b); (d) calculating a predicted preference value for the adjacent music using the preference extracted in step (c) and the association calculated in step (c); And (e) providing a recommendation music list to the recommendation target user by using the preference prediction value calculated in step (d) The association graph may include a plurality of user's favorite music lists, the association relationship between the preference music of the preferred music calculated using the preference information and the preferred music, and the preference information of the user is changed. The music recommendation method using the correlation graph showing the correlation between music, characterized in that when the update. The method of claim 1, And the preference information includes at least one of a purchase count, a download count, and an execution count. The method of claim 1, And said association between said preferred music is a sum of frequencies in which any two songs are preferred by the same user at the same time. The method of claim 1, The association graph is an undirected weighted graph, which is stored as a list of the number corresponding to the total number of the preferred music. The method of claim 1, The preference prediction value is calculated by performing a product of the preference extracted in the step (c) and the correlation calculated in the step (c) by the number of the preferred music and adding the resultant values. A music recommendation method using an association graph showing an association relationship between music. The method according to claim 1 or 5, The association degree is a sum of a frequency at which the preferred music and the adjacent music are simultaneously favored by the same user, a preferred conditional probability of the adjacent music with respect to the preferred music, and chi-square of the preferred music and the adjacent music. Or a correlation graph indicating a correlation between music, wherein the preferred music and the adjacent music are calculated in at least one of mutual information. The method of claim 5, The mutual information

Calculated as above

Denotes a sum of frequencies in which the preferred music and the adjacent music are favored simultaneously from the same user.

Represents a sum of edge values of the correlation graph, and

Represents the sum of all edge values connected to the vertices representing the preferred music,

Is a sum of all interference values connected to the vertices representing the adjacent musics. In the music recommendation system using the correlation graph showing the correlation between music, A preference music database for storing a plurality of user's favorite music lists and preference information; An association database that stores the association graph including the association relationship between the preference music list stored in the preference music database, a preference of the preferred music calculated using the preference information, and the preference music; An association graph updating unit for updating the association database by changing data of the association graph when the preference information of the user is changed; A preference music list extracting unit configured to extract a favorite music list of the recommendation target user from the favorite music database when the recommendation target user is selected; An adjacent music extracting unit extracting adjacent music corresponding to adjacent vertices of the preferred music included in the preferred music list extracted by the favorite music list extracting unit with reference to the association graph; A calculator configured to extract a preference for the favorite music by referring to the correlation graph and calculate an association degree between the favorite music and the neighboring music extracted by the neighboring music extracting unit; A preference calculator configured to calculate a preference prediction value for the adjacent music using the preference extracted by the calculator and the association degree calculated by the calculator; And A recommendation music provider for providing a recommendation music list to the recommendation target user by using the preference prediction value calculated by the preference calculator. Music recommendation system using a correlation graph showing the correlation between music, comprising a. The method of claim 8, And the preference information stored in the favorite music database includes at least one of a number of purchases, a number of downloads, and a number of executions. The method of claim 8, And said association between said preferred music stored in said association database is a sum of frequencies in which any two songs are preferred simultaneously by the same user. The method of claim 8, The preference calculator calculates the preference by performing a product of the preference extracted by the calculator and the correlation calculated by the calculator by the number of the preferred music and adding the resultant values. A music recommendation system using an association graph showing an association relationship between music.

Images