JP2009146314A - Similar user detection system, similar user detection method, and similar user detection program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To correctly detect an unknown similar user in the presence of a plurality of groups handling the same theme. <P>SOLUTION: The system includes a representative group extraction means which extracts a representative group representing a predetermined group based on information showing a link relation between groups; a transform matrix generation means which generates a transform matrix mapping the source of a first vector space having a group as a base to a second vector space having the representative group as a base by approximating the group using the representative group; a transform means which transforms the source of the first vector space showing a user's vector expression to the source of the second vector space by use of the transform matrix generated by the transform matrix generation means; and a similarity calculation means which calculates a similarity between users based on the source of the second vector space transformed by the transform means. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention finds a user similar to a user for a user in a system in which a user and a group (community) represented by a social networking service exist and a link between related items exists between the groups. The present invention relates to a similar user discovery system, a similar user discovery method, and a similar user discovery program.

  There are various methods for recommending content to users in a system in which users and groups (communities) exist, as represented by social networking services. For example, when a friendship is defined between users, there are a method for recommending the content of a user who is in a friendship, a method for recommending the content of a user belonging to the same group, and the like.

  Users are interested in the information of their friends and share the same interests with other users who belong to the same group, so the content of users who are friends or who belong to the same group is recommended Suitable as content. In addition, if you consider another user who shares some of the same interests with you, that user may already know about other things that you may not know yet that you may be interested in. In that sense, the user may be able to obtain unexpectedly good information by searching for other users who have the same interest and interest as him.

  As a method of finding a user similar to the user, there is known a method of finding a similar user by expressing the user by a vector and expressing the similarity between the users by a cosine value of an angle between the vectors. Since the user is considered to belong to a group in which he / she is interested, in order to find a user having an interest and interest similar to the user, each element of the vector may be associated with the group. The user is represented by an N-dimensional vector with N groups, and each element represents the presence / absence of belonging to the group by 1, 0. By the above method, users with similar interests can be easily found.

  In addition, a similarity measuring device that can measure with high accuracy by expressing the similarity in a field of interest between users with a cosine between vectors given to each word has been proposed (for example, Patent Documents). 1).

JP 2004-240505 A (paragraphs 0027-0029)

  However, in finding similar users using the method described above, there is a problem that what is normally determined to be similar cannot be determined to be similar. This is a problem that occurs when multiple groups of the same theme exist. For example, a group whose group name is “Java (registered trademark) program”, “J2SE”, “Java & Eclipse” is considered to be a group with the same Java theme. In such a case, if Mr. A belongs to the “Program in Java” group and Mr. B belongs to the “J2SE” group, the above-described method is applicable to A. And Mr. B are judged to have different attributes. Therefore, even when searching for a user who has an interest similar to Mr. A, Mr. B cannot be found.

  In addition, there is a problem that the result obtained by the above-described method is likely to be a known user or an easily discoverable user. Similar users that can be found by a general method are other users who belong to the same group as themselves. Users who belong to the same group as themselves often have a relatively close link relationship, and are often known or easily found by simply following the group.

  Therefore, an object of the present invention is to provide a similar user discovery system, a similar user discovery method, and a similar user discovery program that can correctly discover similar users even when there are a plurality of groups that handle the same theme.

  The similar user discovery system according to the present invention is based on information indicating a link relationship between groups, representative group extracting means for extracting a representative group representing a predetermined group, and approximating the group using the representative group, A transform matrix generating means for generating a transform matrix for mapping an element of the first vector space having a group as a base to a second vector space having a representative group as a base, and a first vector space indicating a user's vector expression The conversion means for converting the element into the element of the second vector space using the conversion matrix generated by the conversion matrix generation means, and the similarity between the users based on the element of the second vector space converted by the conversion means And a similarity calculation means for calculating.

  The similar user discovery method according to the present invention includes a representative group extraction step of extracting a representative group representing a predetermined group based on information indicating a link relationship between groups, and approximating the group using the representative group. A transformation matrix generating step for generating a transformation matrix for mapping an element of the first vector space based on the group to a second vector space based on the representative group; and a first vector space indicating a user vector representation A transformation step for transforming the element into an element in the second vector space using the transformation matrix generated in the transformation matrix generation step, and similarity between users based on the element in the second vector space transformed in the transformation step And a similarity calculation step for calculating.

  The similar user finding program according to the present invention approximates a group using a representative group, and a representative group extraction process for extracting a representative group representing a predetermined group based on information indicating a link relationship between groups. Thus, a transformation matrix generation process for generating a transformation matrix for mapping an element of the first vector space having a group as a base to a second vector space having a representative group as a base, and a first representing a user vector expression A conversion process for converting an element of the vector space into an element of the second vector space using the conversion matrix generated by the conversion matrix generation process, and a user based on the element of the second vector space converted by the conversion process A similarity calculation process for calculating the similarity is executed.

  According to the present invention, there is an effect that a similar user can be found correctly even when there are a plurality of groups that handle the same theme.

  Further, according to the present invention, it is possible to find a similar user who is unlikely to be known. This is because a plurality of groups that handle the same theme can be identified by the representative group, so that a user who belongs to a group of the same theme as that of the user but who belongs to a group that is not exactly the same can be found.

  First, an outline of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a minimum configuration example of a similar user discovery system according to the present invention. The similar user discovery system illustrated in FIG. 1 includes a representative group extraction unit 101, a conversion matrix generation unit 102, a conversion unit 200, and a similarity calculation unit 201.

  The representative group extracting unit 101 extracts a representative group representing a predetermined group based on information indicating a link relationship between groups.

  The transformation matrix generation means 102 generates a transformation matrix that maps the element of the first vector space based on the group to the second vector space based on the representative group by approximating the group using the representative group. To do.

  The conversion unit 200 converts the element of the first vector space indicating the user's vector expression into the element of the second vector space using the conversion matrix generated by the conversion matrix generation unit 102.

  The similarity calculation unit 201 calculates the similarity between users based on the elements of the second vector space converted by the conversion unit 200.

  If configured as shown in FIG. 1, even when there are a plurality of groups that handle the same theme, it is possible to correctly find an unknown similar user.

  The similar user discovery system according to the present invention is a system for extracting a user similar to a certain user in a system such as a social networking service in which the user belongs to a group. In the present invention, a user is expressed by being characterized by a group to which the user belongs. It is assumed that there are N groups, and each group is referred to as a group 1,.

The user is represented as an element of an N-dimensional vector space U _N spanned by basis vectors e ₁ ,..., E _N representing each group (where e _i is 1 only for the i-th element, and the others are 0) ). That is, when the group j is represented by the basis vector e _j of the N-dimensional vector space U _N , the user sets the coefficient of the basis vector representing the group to which the user belongs to 1 and the coefficient of the basis vector representing the group to which the group does not belong. 0, and it is expressed as the original on the vector space U _N. For example, the user usr ₁ belonging to the group 1 and the group 3 is represented by the following expression.

usr ₁ = 1 * e ₁ + 0 * e ₂ + 1 * e ₃ + 0 * e ₄ + ... + 0 * e _N

Hereinafter, representation as the original on the vector space U _N users, that is, the original vector space U _N indicating a vector representation of the user, may be referred to as "user representation of the vector space U _N".

  When the user is expressed in vector in this way and the similarity of the user is calculated using the vector expression, there arises a problem that the users belonging to the different groups are determined not to be similar. The main reason is that there are multiple groups that handle the same theme.

  Therefore, in the present invention, it is considered that representative groups of each theme are selected from the groups and all groups are represented only by those representative groups. The representative group is, for example, a group that represents a plurality of groups that handle the same theme. By doing so, a plurality of groups that handle the same theme are expressed by the same representative group concerned. Therefore, the first problem does not occur, and the similarity is correctly determined even for users belonging to different groups.

  As for the second problem, some groups who handle the same theme can be identified by the representative group, so some people who belong to a group that handles the same theme as their group but is not exactly the same Because it can be discovered, the possibility that the resulting user is known is low.

Representative Group calculation means 11 finds a representative group from among the plurality of groups (extracted), the transformation matrix generating means 12, the user (the original vector space U _N), with a representative group from the normal vector representation Generate a transformation matrix to convert to a vector representation.

  First, a method in which the representative group calculation unit 11 finds a representative group from among the groups will be described. The groups in the system handled by the present invention have links between related ones. A page rank algorithm is applied to the link structure of related groups.

  The page rank (for example, see http://infolab.stanford.edu/~backrub/google.html) is the number of pages when there is a web surfer that randomly follows the link between pages and the page transition continues indefinitely. Represents the probability of being on (referring to) the page. It can be considered that the link structure between groups has the property that links are gathered to an authoritative group like the web link structure. Therefore, the page rank algorithm is applied to the group link relationship.

  A local maximum point having a high probability is obtained for the periphery, and a group corresponding to the local maximum point can be regarded as a group representing the peripheral group. Therefore, the representative group calculation unit 11 applies a page rank algorithm to the link structure of the group to obtain a maximum point, and extracts a group corresponding to the obtained maximum point as a representative group.

  Next, a method in which the conversion matrix generation unit 12 generates a conversion matrix will be described. A description will be given assuming that the representative group calculation unit 11 has extracted M representative groups.

For the representative group, representative group 1,. . . Suppose that the name is referred to as representative group M. Similar to the vector space U _N , a vector space U _M that represents users by representative groups is defined. The vector space U _M is a vector space spanned by basis vectors f ₁ ,..., F _M representing each representative group (where f _i is 1 only for the i-th element and 0 for others). That is, the representative group i is represented by a base vector f _i of the M-dimensional vector space U _M.

Now consider a linear mapping T from the vector space U _N to U _M. This linear mapping T is defined as follows, and maps (maps) a user (an element of the vector space U _N ) from the vector space U _N to the vector space U _M. The following formula, basis vectors _e 1 on the linear map T vector space is mapped by the _{U N,} ···, _{e N} (group), basis vectors _f 1 on the vector space _{U M,} · · ·, f It is represented by _M (representative group).

T (e ₁ ) = a ₁₁ f ₁ + a ₂₁ f ₂ +... + A _M1 f _M (1-1)
...
T (e _N ) = a _1N f ₁ + a _2N f ₂ +... + A _MN f _M (1-N)

  Formula (1-1) means that the group 1 is approximated from the representative group 1 to the representative group M. The degree of approximation such as primary approximation, secondary approximation,. In the case of the first approximation, non-zero (non-zero value) is assigned only to the coefficient of the representative group having the closest link relationship (that is, the smallest number of steps that can be reached in the shortest), and 0 is assigned to the other representative groups. It is done. Similarly, in the case of quadratic approximation, the coefficients up to the second closest representative group are non-zero, and the coefficients of the other representative groups are zero. The degree of approximation is a parameter of the similar user discovery system according to the present invention. When the order of approximation becomes too large, unrelated groups become related. For this reason, it is assumed that the degree of approximation is low and an optimum one is selected according to the system.

The meaning of a _ij is a numerical value representing the degree of relationship on the link between group j and representative group i. It is assumed that the closer the relationship between the group j and the representative group i is, the larger the value is. If there is no relationship, the value is zero. For example, the value of a _ij can be calculated using an expression such as 1 / (l _ij +1) ^ 2 (where l _ij is the shortest number of steps from group j to representative group i). That is, each group can be approximated by a representative group by calculating 1 / (l _ij +1) ^ 2. In addition, this formula is only an example, and is not limited to this formula.

  Next, the matrix A is defined as follows.

A = (a _ij )

Matrix A, the base _e 1, · · ·, _{e N} and the base _f 1, · · ·, a representation matrix of linear mapping T about _{f M.} By multiplying the user's representation in the vector space U _N by the matrix A, it can be easily converted into the user representation in the vector space U _M. The transformation matrix generation means 12 obtains the matrix A by calculation and stores it in the transformation matrix storage unit 32.

  As described above, after the user's expression is converted into the representation by the representative group using the conversion matrix A, similarities that are not known can be found correctly by calculating the similarity.

  Next, a second similar user discovery system according to the present invention will be described.

In general, when a user similar to a certain user is found, the amount of calculation increases. The reason is that the similarity between the user and all other users must be calculated. In order to find a user similar to a certain user, the similarity between the certain user and all other users must be calculated. In the present invention, the user representation on the vector space U _N must be converted into the vector space U _M and then compared. For this reason, the amount of calculation increases more than usual, and the reduction of the amount of calculation is important. In the present invention, first, users who are likely to be similar are narrowed down, and the calculation amount is reduced by calculating the similarity only with the narrowed-down users. Hereinafter, a method of narrowing down users who may be similar will be described.

Consider the determination of the other users similar to the representation usr _i users with on vector space U _N. First, transformed using the transformation matrix A to usr _i to represent usr _i 'of a vector space _{U M.}

usr _i '= A * us _ri = A * (u ₁ e ₁ + u ₂ e ₂ +... + u _N e _N ) = u ′ ₁ f ₁ + u ′ ₂ f ₂ +... + u ′ _M f _M )

Here, the values of u ′ ₁ ,..., U ′ _M are considered. u ′ ₁ ,..., u ′ _M are coefficients of each basis vector of the vector space U _M. The values of u ′ ₁ ,..., u ′ _M are values obtained by approximating the group to which the user belongs with a representative group. Generally, some of u ′ ₁ ,..., U ′ _M have a value of 0, except when the group to which the user belongs is across all themes. This indicates that the user does not belong to a group related to a representative group having a coefficient value of 0 (a representative group indicated by a base vector having a coefficient of 0). Conversely, if the coefficient is a non-zero value, it indicates that the user belongs to a group that has some relationship with the representative group.

Here, attention is focused on a representative group whose coefficient is a non-zero value. At least one user who is similar to a certain user i should belong to a group related to the theme handled by the group to which the user i belongs. Thus, when finding a user similar to user i, at least one of the non-zero vector components must be similarly non-zero when converted to a user representation of vector space U _M . That is, the component corresponding to at least one of the components other than 0 in the user representation of the vector space U _M for the user i needs to be other than 0 in the user representation of the vector space U _M for a similar user. is there.

Above satisfy the (representative of the user similar to a certain user) and a vector of a vector space U _N is referred to as a similar candidate vector. These similar candidate vectors can be easily obtained based on the transformation matrix A.

Row ₁ row _r nonzero components in the user representation of the vector space _{U M} of a user i, · · ·, if it was _{r n} rows, _{r 1} row of the transformation matrix A, · · ·, the _{r n} rows All row vectors are added to generate a row vector R. Hereinafter, this vector R is referred to as a similar candidate mask.

The row vector nonzero columns c ₁ column (similar candidate mask), ..., when c _m columns, necessary and sufficient condition vector is similar candidate vector vector space U _N is, c ₁ row - .., _cm There are at least one non-zero value in the _m line. Therefore, in order to narrow down users who may be similar from the set of users, a similar candidate mask is obtained from the transformation matrix A, and at least one of the rows (components) corresponding to the non-zero column is non-zero. What is necessary is just to obtain | require the user represented by the vector which is zero. Candidate search means 61 shown in FIG. 6 to be described later narrows down the users who may be similar by the above-described processing. The amount of calculation can be reduced by the method described above.

Embodiment 1. FIG.
A first embodiment of the present invention will be described below with reference to the drawings. FIG. 2 is a block diagram illustrating a configuration example of the similar user discovery system according to the first embodiment. Referring to FIG. 2, the similar user discovery system according to the first embodiment of the present invention includes a transformation matrix generation device 1 that operates under program control, a similarity calculation device 2, a storage device 3 that stores information, and a network. An input device 4 realized by an interface, a keyboard, and the like, and an output device 5 realized by a network interface, a display device, a printing device, and the like are included. The similar user discovery system is specifically realized by an information processing apparatus such as a personal computer that operates according to a program.

  The storage device 3 includes a group information storage unit 31, a transformation matrix storage unit 32, and a user information storage unit 33.

  The group information storage unit 31 stores group information and link information between groups. The link information between groups is information indicating the link relationship between groups, for example, and is information used by the representative group calculation unit 11. The group information is information indicating the rank of each group, for example, and is information stored by the representative group calculation unit 11. The information stored in the group information storage unit 31 is registered and updated in advance by a system administrator, for example.

The transformation matrix storage unit 32 stores the transformation matrix generated by the transformation matrix generation device 1 and information indicating the group and the representative group at the time of the transformation matrix generation. The information is shown group during the transformation matrix generator, for example, a group to be the transformation matrix generator, namely, information indicating a group represented by the basis vectors of the vector space U _N. The information indicating the representative group at the time of generating the transformation matrix is, for example, information indicating the representative group that is the target of the transformation matrix generation, that is, the representative group represented by the basis vector of the vector space U _M.

The user information storage unit 33 stores, for example, user information and group affiliation information. Information of the user is, for example, information including identification information capable user, and a representation of a vector space U _N.

  The transformation matrix generation device 1 includes a representative group calculation unit 11 and a transformation matrix generation unit 12. The representative group calculation unit 11 obtains a representative group using a page rank algorithm based on the link information between groups stored in the group information storage unit 31. Further, the transformation matrix generation means 12 obtains the transformation matrix A by expressing all groups by representative groups, and stores them in the transformation matrix storage unit 32. Further, the transformation matrix generation unit 12 stores the group and the representative group at this time together in the transformation matrix storage unit 32.

The similarity calculation device 2 includes similarity calculation means 21. The similarity calculation means 21 converts the representation of the input user and the representation of the user to be compared from the user representation on the vector space U _N to the user representation on the vector space U _M to calculate the similarity. To do.

  Next, the operation of the first embodiment will be described with reference to the drawings. FIG. 3 is a flowchart showing the operation of the first embodiment.

  The transformation matrix generation device 1 and the similarity calculation device 2 are devices that operate independently. When there is a change in the group information storage unit 31, the conversion matrix generation device 1 repeats the calculation of the conversion matrix and stores it in the conversion matrix storage unit 32. The similarity calculation device 2 calculates the similarity using the transformation matrix stored in the transformation matrix storage unit 32.

  First, the operation of the transformation matrix generation device 1 will be described. FIG. 3A is a flowchart showing the operation of the transformation matrix generation device 1. Note that the conversion matrix generation device 1 executes the following conversion matrix generation processing based on, for example, detecting that the information stored in the group information storage unit 31 has been updated. Moreover, the conversion matrix production | generation apparatus 1 may perform the conversion matrix production | generation process shown below repeatedly, for example for every predetermined period, and may update a conversion matrix at any time.

  The representative group calculation unit 11 obtains the group link relationship (link information between groups) from the group information storage unit 31 and ranks the groups using a page rank algorithm. Next, the representative group calculation means 11 selects a group having a rank higher than that of the link source group or the link destination group and sets it as a representative group (step S11).

  FIG. 4 is an explanatory diagram illustrating an example of information indicating the rank of each group. In FIG. 4, the numbers surrounded by circles indicate each group. Moreover, the arrow between groups shows the link relationship between groups. For example, when there is a link relationship between groups as shown in FIG. 4, the representative group calculation means 11 calculates the value shown below each node (group) as the rank of each group (however, the link transition probability) Is calculated using 0.85 as the parameter of the page rank algorithm as an equal probability). The representative group calculation means 11 selects a predetermined number of groups with a high calculated rank and determines them as representative groups. In the example illustrated in FIG. 4, the representative group calculation unit 11 determines group 3 and group 8 as representative groups.

  Next, the calculated representative group is supplied to the transformation matrix generation means 12. Specifically, the representative group calculation unit 11 stores information indicating the determined representative group in the group information storage unit 31, and the transformation matrix generation unit 12 stores information indicating the representative group stored in the group information storage unit 31. To extract.

The transformation matrix generation means 12 generates the transformation matrix A by representing all groups as representative groups. Specifically, first, the transformation matrix generation means 12 determines the coefficients a _ij of the expressions (1-1),..., Expression (1-N). In the first embodiment, first-order approximation is used, and 1 / (l _ij +1) ^ 2 is used in the expression of the coefficient a _ij . That is, the transformation matrix generation unit 12 calculates the coefficient a _ij by calculating 1 / (l _ij +1) ^ 2. Here, l _ij is, for example, the shortest number of steps from the group j to the representative group i. However, the formula for calculating the coefficient a _ij is merely an example, and is not limited thereto.

When the coefficient a _ij calculated in the example shown in FIG. 4 is applied, Expressions (1-1),..., Expression (1-N) are as follows.

T (e ₁ ) = 1/9 * f ₁ + 0 * f ₂
T (e ₂ ) = 1/4 * f ₁ + 0 * f ₂
T (e ₃ ) = 1 * f ₁ + 0 * f ₂
T (e ₄ ) = 1/4 * f ₁ + 0 * f ₂
T (e ₅ ) = 1/4 * f ₁ + 0 * f ₂
T (e ₆ ) = 1/9 * f ₁ + 1/9 * f ₂
T (e ₇ ) = 0 * f ₁ + 1/4 * f ₂
T (e ₈ ) = 0 * f ₁ + 1 * f ₂
T (e ₉ ) = 0 * f ₁ + 1/4 * f ₂
T (e ₁₀ ) = 0 * f ₁ + 1/4 * f ₂

The transformation matrix A is a matrix representation of the basis vector coefficients in the equations (1-1),..., Equation (1-N). That is, the conversion matrix generation unit 12 generates the conversion matrix A based on the calculated coefficient a _ij . FIG. 5 is an explanatory diagram illustrating an example of the transformation matrix A. The transformation matrix generation means 12 stores in the transformation matrix storage unit 32 information indicating the generated transformation matrix, the group used to obtain the transformation matrix, and the representative group (step S12).

  Next, the operation of the similarity calculation device 2 will be described. FIG. 3B is a flowchart showing the operation of the similarity calculation device 2. Note that the similarity calculation device 2 executes the similarity calculation processing shown below, triggered by, for example, a user performing a similarity calculation execution instruction operation.

Input similarity calculation device 2 is a user representation in a vector space U _N of the user supplied from the input device 4. In other words, the similarity calculation unit 21 inputs the user representation of a vector space U _N from the input device 4. The similarity calculation means 21 converts the input user expression on the vector space U _N to the user expression on the vector space U _M by applying the conversion matrix A, and similarly converts other users. Specifically, the similarity calculation unit 21 converts an element of the vector space U _N into an element of the vector space U _M using a conversion matrix.

Next, the similarity calculation means 21 calculates the similarity between the respective users based on the user expression on the vector space U _M (step S21). In the first embodiment, the cosine value of the angle between vectors is used for calculating the similarity. Note that the method of using the cosine value for calculating the similarity is merely an example, and is not limited to this method. In addition, various methods such as a method using a distance between vectors can be applied.

  Finally, the similarity calculation means 21 outputs the similarity to each group (user set) obtained in step S21 to the output device 5 in descending order of similarity (step S22).

  Here, as an example, the operation of the similarity calculation means 21 when there is a link relationship between groups shown in FIG. 4 will be described. Hereinafter, a case where the similarity between the user 1 and the user 2 and the similarity between the user 1 and the user 3 are specifically calculated using cosine values and a user similar to the user 1 is extracted will be described.

Assume that user 1 belongs to groups 3 and 6, user 2 belongs to groups 2 and 4, and user 3 belongs to groups 9 and 10. In this case, each expression a vector space U _N users 1,2,3, [0; 0; 1; 0; 0; 1; 0; 0; 0; 0], [0; 1; 0; 1 ; 0; 0; 0; 0; 0; 0], [0; 0; 0; 0; 0; 0; 0; 0; 1; 1] (Hereafter, the column vector [a; b; c] And the row vector as [a, b, c].)

Similarity calculating means 21, the representation of a vector space U _N users 1, 2, 3, by applying a transformation matrix illustrated in FIG. 5, represented in a vector space U _M [10/9; 1 / 9], [1/2; 0], [0; 1/2] are obtained. The similarity calculation means 21 obtains a cosine value by using the user expression on the calculated vector space U _M. For example, the value of the cosine of the angle between the vectors of the user 1 and the user 2 is 0.995, and the value of the cosine of the angle between the vectors of the user 1 and the user 3 is 0.0995. Finally, the similarity calculation means 21 outputs the value for identifying the group (between users) for which the similarity is calculated and the value of the similarity to the output device 5 in descending order of similarity. In this example, similarity values are output to the output device 5 in the order of user 2 and user 3.

  Next, the effect of the first embodiment will be described. In the first embodiment, a representative group of each theme is selected from the groups, and all the groups are expressed by the representative group, so that a plurality of groups of the same theme are expressed by the same related representative group. It becomes like this. Therefore, even when there are a plurality of groups that handle the same theme, the similarity can be calculated correctly.

  Further, there is a problem that a result obtained by a general method for finding a similar user is likely to be a known user or a user who can be easily found. Similar users that can be found by a general method are other users who belong to the same group as themselves. Users who belong to the same group as themselves often have a relatively close link relationship, and are often known or easily found by simply following the group. On the other hand, in the first embodiment, a group that handles a plurality of the same themes can be identified by the representative group, so that it belongs to a group of the same theme as itself but belongs to a group that is not exactly the same. Users can also discover. Therefore, it is possible to find a similar user who is unlikely to be known.

  In the first embodiment, the case where the similarity calculation unit 21 calculates the cosine value using all the dimensions (components) of the vector is exemplified. However, as a modification, the following method can be considered. For example, the similarity calculation unit 21 calculates the similarity by using only a non-zero dimension (component) of a vector obtained by converting the user representation of a predetermined user input to the similarity calculation unit 21 using a conversion matrix. Also good. By doing so, it can be expected that the accuracy can be improved because it can be compared with other users only in the portion of interest of the input user (predetermined user).

  In the first embodiment described above, the storage device 3 is described as including the group information storage unit 31, the transformation matrix storage unit 32, and the user information storage unit 33. However, another device connected to the network may be used. The group information storage unit 31 and the user information storage unit 33 may be provided. In addition, as for other functional means, it is only necessary that each functional means described above is realized by any one or a plurality of devices connected on a network, and which device has which functional means. The configuration is not limited to the above.

Embodiment 2. FIG.
The configuration of the second embodiment of the present invention will be described. FIG. 6 is a block diagram illustrating a configuration example of the similar user discovery system according to the second embodiment. As shown in FIG. 6, the similar user discovery system according to the second embodiment is different in that the similarity calculation device 6 includes candidate search means 61 in addition to the configuration of the similarity calculation device 2 shown in FIG. The same components as those in the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

  The similarity calculation device 6 in the second embodiment includes candidate search means 61 and similarity calculation means 21. Candidate search means 61 narrows down users who may be similar to the input user. Thereafter, the similarity calculation means 21 calculates the similarity only with the users narrowed down by the candidate search means 61 and outputs the result to the output device 5.

  Next, the operation of the second embodiment will be described. FIG. 7 is a flowchart showing the operation of the second embodiment. FIG. 7A is a flowchart showing the operation of the transformation matrix generation device 1. FIG. 7B is a flowchart showing the operation of the similarity calculation device 6. About the operation | movement similar to 1st Embodiment, the code | symbol same as FIG. 3 is attached | subjected and detailed description is abbreviate | omitted.

  The similarity calculation device 6 in the second embodiment is different from the first embodiment in that the candidate search unit 61 narrows down the targets for calculating the similarity and calculates the similarity only with the narrowed-down target.

Candidate search means 61 receives the user representation usr of a vector space U _N from the input device 4 as an input. Candidate search means 61 multiplies conversion matrix A acquired from conversion matrix storage unit 32 and vector usr to convert user expression usr on vector space U _N into expression usr ′ on vector space U _M.

  Next, the candidate search unit 61 adds all the rows (row vectors) of the transformation matrix A corresponding to the non-zero rows (components) of usr ′ to generate a similar candidate mask.

From the set of users is represented as the original on the vector space U _N, in order to narrow down the users who may be similar, candidate retrieval section 61, a non-zero dimension (components) corresponding to those of the similar candidate mask ( Filtering is done by checking if there is at least one non-zero value in the row (component)). Then, the candidate search unit 61 supplies (outputs) the narrowed-down result to the similarity calculation unit 21 (step S31).

Similarity calculation means 21 receives a representation of a vector space U _N of users candidate searching unit 61 is extracted, converted into the user representation of a vector space U _M. The similarity calculation means 21 calculates the similarity between the user 4 and the user extracted by the candidate search means 61 based on the user expression on the vector space U _M (step S21), and the similarity is in descending order. It outputs to the output device 5 (step S22).

  Here, the operation of the candidate search means 61 will be described using a specific example. Assume that a user similar to the user 4 belonging to the groups 2, 4 and 5 is found when there is a link relationship between the groups shown in FIG.

Expression of a vector space U _N users 4 (usr ₄₎ is [0; 0 1; 0; 1; 1; 0; 0; 0; 0]. When this is multiplied by a transformation matrix A (see FIG. 5) and transformed into an expression (usr ₄ ′) on the vector space U _M , [3/4; 0] is obtained. Thus, the non-zero row of usr ₄ ′ is the first row. Since there is only the first row, it is not necessary to add the row vectors of the transformation matrix A, and the first row of the transformation matrix A becomes a similar candidate mask. That is, the similar candidate mask is [1/9; 1/4; 1; 1/4; 1/4; 1/9; 0; 0; 0;

The non-zero columns of similar candidate masks are the first through sixth columns. Thus, the user that may be similar to the user 4, the sixth row from the first row in the representation of a vector space U _N, at least one is a user who has a non-zero value. The candidate search means 61 extracts users satisfying this condition from the user information storage unit 33 and supplies (outputs) them to the similarity calculation means 21.

  Next, effects of the second embodiment will be described. In general, when a user similar to a certain user is found, the amount of calculation increases. The reason is that the similarity between the user and all other users must be calculated. On the other hand, in the second embodiment, the similar candidate mask obtained from the transformation matrix A is used to narrow down users who may be similar, and the similarity is calculated only for the narrowed-down users. In calculation, the amount of calculation can be reduced. That is, unlike the first embodiment, the number of similarities calculated by the similarity calculation unit 21 can be reduced because the candidate search unit 61 narrows down the users who may be similar.

Next, a modification of the second embodiment will be described. In the modified example, the fact that the user similar to the user A is not related to the representative group not related to the user A is used for narrowing down. Based on this idea, attention is paid to the zero row in the expression (usr _A ′) on the vector space U _M of the user A. Candidate search means 61 adds the rows of the transformation matrix corresponding to the zero rows in usr _A ′ to generate a similar candidate mask.

  When selecting a similar candidate using the similar candidate mask in the modification, the candidate search means 61 selects a user whose portion (row) corresponding to a non-zero element (column) of the similar candidate mask is zero. By doing so, the objects whose similarity is calculated can be narrowed down to those who are also not interested in what the user A is not interested in.

  In the embodiment described above, a similar user discovery system having characteristic configurations as shown in the following (1) to (5) is shown.

  (1) Representative group extracting means for extracting a representative group representing a predetermined group based on information indicating a link relationship between groups (for example, realized by the representative group calculating means 11), and representing the group as a representative group Transformation matrix generating means for generating a transformation matrix that maps the element of the first vector space with the group as a base to the second vector space with the representative group as a base by using the approximation (for example, the transformation matrix generating means) And a conversion unit that converts an element of the first vector space indicating the user's vector representation into an element of the second vector space using the conversion matrix generated by the conversion matrix generation unit (for example, Similarity calculation means 21) and similarity calculation means for calculating similarity between users based on the elements of the second vector space converted by the conversion means For example, similar user finding system comprising: a to) and implemented by the similarity calculating unit 21.

  (2) The similar user finding system, wherein the similarity calculating means calculates the similarity using only components other than zero among the elements of the second vector space indicating the vector representation of the predetermined user. The similar user discovery system configured as described above can compare the degree of similarity with other users only for a portion of interest of a predetermined user.

  (3) generating a similar candidate mask based on a row of a conversion matrix corresponding to the original non-zero component of the second vector space indicating the vector representation of the predetermined user converted by the conversion means, and generating the generated similar candidate mask Candidate extraction means (for example, realized by the candidate search means 61) that extracts the element of the first vector space in which at least one of the components corresponding to the non-zero component is non-zero as the vector representation of the similar candidate user The conversion means converts the element of the first vector space indicating the vector representation of the similar candidate user extracted by the candidate extraction means into the element of the second vector space, and the similarity calculation means The similar user discovery system which calculates the similarity between a predetermined user and a similar candidate user based on the element of the converted 2nd vector space. Since the similar user discovery system configured as described above narrows down the users who may be similar in advance, the processing of the similarity calculation means can be reduced.

  (4) The candidate extraction unit generates a similar candidate mask based on the row of the transformation matrix corresponding to the original zero component of the second vector space indicating the vector representation of the predetermined user transformed by the transformation unit, The similar user discovery system which extracts the element of the 1st vector space whose component corresponding to the non-zero component of the similar candidate mask is zero as a vector representation of a similar candidate user. The similar user discovery system configured as described above can extract, as similar candidates, users whose representative groups that are not related are the same.

  (5) A similar user discovery system including a transformation matrix storage unit that stores a transformation matrix generated by the transformation matrix generation unit.

  The present invention can be effectively applied to find a user similar to a certain user in a system having a link relationship between groups.

It is a block diagram which shows the minimum structural example of the similar user discovery system by this invention. It is a block diagram which shows the structural example of the similar user discovery system of 1st Embodiment. It is a flowchart which shows operation | movement of 1st Embodiment. It is explanatory drawing which shows the example of the information which shows the rank of each group. It is explanatory drawing which shows the example of the conversion matrix A. It is a block diagram which shows the structural example of the similar user discovery system of 2nd Embodiment. It is a flowchart which shows operation | movement of 2nd Embodiment.

Explanation of symbols

101 representative group extraction means 102 conversion matrix generation means 200 conversion means 201 similarity calculation means

Claims (15)

Representative group extraction means for extracting a representative group representing a predetermined group based on information indicating a link relationship between groups;
Transformation matrix generation means for generating a transformation matrix for mapping a group of elements in a first vector space based on the group to a second vector space based on the representative group by approximating the group using the representative group;
Conversion means for converting an element of a first vector space indicating a vector representation of a user into an element of a second vector space using the conversion matrix generated by the conversion matrix generation means;
A similarity user discovery system comprising: similarity calculation means for calculating a similarity between users based on an element of the second vector space converted by the conversion means.   The similar user discovery system according to claim 1, wherein the similarity calculation means calculates the similarity using only a non-zero component of the elements of the second vector space indicating the vector representation of the predetermined user. A similar candidate mask is generated based on the row of the conversion matrix corresponding to the original non-zero component of the second vector space indicating the vector representation of the predetermined user converted by the conversion means, and the generated similar candidate mask is not zero Candidate extraction means for extracting an element of the first vector space in which at least one of the components corresponding to the components of is non-zero as a vector representation of a similar candidate user,
The converting means converts an element of the first vector space indicating the vector representation of the similar candidate user extracted by the candidate extracting means into an element of the second vector space,
The similar user according to claim 1, wherein the similarity calculating unit calculates a similarity between the predetermined user and the similar candidate user based on an element of the second vector space converted by the converting unit. Discovery system.   The candidate extraction unit generates a similar candidate mask based on the row of the conversion matrix corresponding to the original zero component of the second vector space indicating the vector representation of the predetermined user converted by the conversion unit, and the generated similar candidate The similar user discovery system according to claim 3, wherein an element of the first vector space in which a component corresponding to a non-zero component of the mask is zero is extracted as a vector representation of similar candidate users.   The similar user discovery system according to any one of claims 1 to 4, further comprising a transformation matrix storage unit that stores a transformation matrix generated by the transformation matrix generation unit. A representative group extracting step of extracting a representative group representing a predetermined group based on information indicating a link relationship between the groups;
A transformation matrix generating step for generating a transformation matrix for mapping the element of the first vector space based on the group to the second vector space based on the representative group by approximating the group using the representative group;
A conversion step of converting an element of a first vector space indicating a vector representation of a user into an element of a second vector space using the conversion matrix generated in the conversion matrix generation step;
A similarity calculation step of calculating a similarity between users based on an element of the second vector space converted in the conversion step.   7. The similar user finding method according to claim 6, wherein in the similarity calculation step, the similarity is calculated using only components other than zero among the elements of the second vector space indicating the vector expression of the predetermined user. A similar candidate mask is generated based on the row of the conversion matrix corresponding to the original non-zero component of the second vector space indicating the vector representation of the predetermined user converted in the conversion step, and the generated similar candidate mask is non-zero A candidate extracting step of extracting an element of the first vector space in which at least one of the components corresponding to the components of is non-zero as a vector representation of the similar candidate user,
In the conversion step, the element of the first vector space indicating the vector representation of the similar candidate user extracted in the candidate extraction step is converted into the element of the second vector space,
The similarity user according to claim 6 or 7, wherein, in the similarity calculation step, the similarity between the predetermined user and the similar candidate user is calculated based on the element of the second vector space converted in the conversion step. Discovery method.   In the candidate extraction step, a similar candidate mask is generated based on the row of the conversion matrix corresponding to the original zero component of the second vector space indicating the vector representation of the predetermined user converted in the conversion step, and the generated similar candidate The similar user finding method according to claim 8, wherein an element of the first vector space in which a component corresponding to a non-zero component of the mask is zero is extracted as a vector representation of the similar candidate user.   The similar user discovery method according to any one of claims 6 to 9, further comprising a transformation matrix registration step of storing the transformation matrix generated in the transformation matrix generation step in a transformation matrix storage unit. On the computer,
Representative group extraction processing for extracting a representative group representing a predetermined group based on information indicating a link relationship between groups;
A transformation matrix generating process for generating a transformation matrix for mapping a group of elements in a first vector space based on the group to a second vector space based on the representative group by approximating the group using the representative group;
A conversion process for converting an element of a first vector space indicating a vector representation of a user into an element of a second vector space using the conversion matrix generated by the conversion matrix generation process;
A similar user finding program for executing a similarity calculation process for calculating a similarity between users based on an element of the second vector space converted by the conversion process. On the computer,
The similar user discovery according to claim 11, wherein the similarity calculation process executes a process of calculating a similarity using only a non-zero component of elements of a second vector space indicating a vector representation of a predetermined user. program. On the computer,
A similar candidate mask is generated based on the row of the conversion matrix corresponding to the original non-zero component of the second vector space indicating the vector representation of the predetermined user converted by the conversion process, and the generated similar candidate mask is not zero A candidate extraction process for extracting an element of the first vector space in which at least one of the components corresponding to the components of is non-zero as a vector representation of a similar candidate user,
In the conversion process, the process of converting the element of the first vector space indicating the vector representation of the similar candidate user extracted in the candidate extraction process into the element of the second vector space is executed,
12. The similarity calculation process for executing a process for calculating a similarity between a predetermined user and a similar candidate user based on an element of the second vector space converted by the conversion process. Item 13. A similar user discovery program according to item 12. On the computer,
In the candidate extraction process, a similar candidate mask is generated based on the row of the conversion matrix corresponding to the original zero component of the second vector space indicating the vector representation of the predetermined user converted by the conversion process, and the generated similar candidate The similar user discovery program according to claim 13, for executing processing for extracting an element of a first vector space in which a component corresponding to a non-zero component of a mask is zero as a vector representation of a similar candidate user. On the computer,
The similar user discovery program according to any one of claims 11 to 14, for executing a transformation matrix registration process for storing a transformation matrix generated by the transformation matrix generation process in a transformation matrix storage unit.

Images