JP5350319B2 - Friend recommendation device, method and program - Google Patents

Description

  The present invention relates to a friend recommendation device, method and program, and in particular, in communication between users provided on a network, such as BBS (Bulletin Boad System), SNS (Social Network Service), QA site, etc. The present invention relates to a friend recommendation device, method and program for introducing friends in a community in order to activate communication within the community.

  In recent years, with the development of the Internet, communication via a network has become widespread. In particular, a mailing list, a bulletin board, a chat, a blog, a QA site, a social networking service, and the like form a community and actively communicate and exchange information regardless of whether or not there is an acquaintance between users.

  In such a networked community, various companies use it to advertise products and services and to collect consumer voices. Users exchange information and collect information about their hobbies and interests, and use it as a place to meet new people.

  A number of methods for activating such communities have been proposed. For example, an apparatus that discovers an important user who is communicating with various users regardless of old or new and presents the community to which the user belongs as an active community is disclosed (for example, see Patent Document 1). This is a method based on the hypothesis that the presence of an important user reduces the anxiety felt in communication with a stranger in the user and creates an environment where communication is easy.

  In addition, a method of finding two users who are likely to be friends and allowing them to communicate with each other has been proposed. Friendship can be modeled as a graph in discrete mathematics. For example, when user A recognizes user B as a friend, A and B are regarded as nodes, and if a directional branch having A as a start point and B as an end point exists, it can be modeled as a directed graph. At this time, there is a growth limitation in a general graph, and there are those that find two nodes that are likely to generate edges, that is, two users who are likely to be friends (for example, Non-patent document 1). A community can be activated by having two users who are likely to become friends using this method communicate.

JP-A-2005-235118

Liben-Nowell, D. et al, `` The link prediction problem for social networks '', CIKM '03 * Proceedings of the twelfth international conference on Information and knowledge management, ACM Press New York, 2003.

  Conventional online community activation techniques focus only on the nature of the community itself or the nature of a single user belonging to the community. However, communication is originally performed by two users.

  For example, in the technique described in Patent Document 1, it is possible to know which community is easy for the user to participate from the nature of the community, but since it does not focus on the nature of the user, it is not known which member of the community can easily communicate with. . Therefore, it is considered that the community cannot be fully activated.

  Further, the method described in Non-Patent Document 1 for finding two users who are friends based on a graph is based on a method applied to a general graph, and does not consider the characteristics of individual graphs. Therefore, there is a problem that it is not possible to find two friends who are friends with sufficient accuracy.

  The present invention has been made in view of such a problem, and the purpose of the present invention is to create a graph of friendships between users in a community, and based on the characteristics of the graph, It is an object to provide a friend recommendation device, method, and program capable of facilitating communication by finding two users who are likely to become and recommending them as friends.

In order to solve the above-mentioned problems, the present invention (Claim 1) is directed to a friend recommendation device for recommending friends in an online community system in which each user has an ID and aims to communicate with each other. Because
Communication information acquisition means for acquiring users in an online community and storing them in a communication storage means;
3-partial graph information acquisition means for classifying the three users in the online community from the communication information of the communication storage means based on the presence / absence pattern of communication, and storing it in the 3-partial graph storage means as 3-node graph information;
From the communication information in the communication storage means, information about the presence or absence of communication for two users in the online community, and from the 3-partial graph information in the three-part graph storage means, for two users in the online community, another one other than two users Edge type information acquisition means for acquiring the frequency of which each user is classified according to the pattern of presence / absence of communication of three users, and storing in the edge type storage means;
For the two users in the online community, using the information on the presence / absence of communication stored in the edge type storage means and the number of other users classified into the pattern of presence / absence of communication of three users, respectively. , Weight parameter acquisition means for calculating weight parameters and storing them in weight parameter storage means;
When a user identifier is given, information on the presence / absence of communication, the number of another user classified as a pattern of presence / absence of communication of three users, and the weight parameter of the weight parameter storage means And a friend recommendation means for recommending another user as a friend to the user.

Further, the present invention (Claim 2) is a weight parameter calculating means,
Means for calculating a weight parameter by logistic regression analysis based on information on the presence / absence of communication and the number of another user classified into each of the patterns of presence / absence of communication of three users.

Further, the present invention (Claim 3) is a friend recommendation method for recommending a friend in an online community system for which each user has an ID and aims to communicate between users,
Communication storage means, 3-part graph storage means, edge type storage means, weight parameter storage means are stored, communication information acquisition means, 3-part graph information acquisition means, edge type information acquisition means, weight parameter acquisition means, friend recommendation A device having means for:
A communication information acquisition means for acquiring a user in an online community and storing it in a communication storage means;
The 3-partial graph information acquisition unit classifies the three users in the online community from the communication presence / absence pattern based on the communication information stored in the communication storage unit, and stores the information in the 3-partial graph storage unit as 3-node graph information. A subgraph information acquisition step;
The edge type information acquisition means, from the communication information of the communication storage means, about two users in the online community, information about the presence or absence of communication, from the 3-partial graph information of the 3-partial graph storage means, about two users in the online community, An edge type information acquisition step of acquiring the frequency of which one user other than the two users is classified according to the pattern of presence / absence of communication of each of the three users, and storing the frequency in an edge type storage unit;
For the two users in the online community, the weight parameter acquisition means is classified into the information on the presence / absence of communication stored in the edge type storage means and the pattern of the presence / absence of communication of another one user for each of the three users. A weight parameter obtaining step of calculating a weight parameter using the number of and storing the weight parameter in a weight parameter storage means;
When the friend recommendation means is given a user identifier, the information about the presence / absence of communication of the 3-partial graph storage means and another user of the edge type storage means are classified into the patterns of presence / absence of communication of 3 users respectively. The friend recommendation step of recommending another user as a friend to the user is performed using the number of the received parameters and the weight parameter of the weight parameter storage means.

Further, the present invention (Claim 4) is the weight parameter calculating step,
A weight parameter is calculated by logistic regression analysis based on information on the presence / absence of communication and the number of another user classified into the pattern of presence / absence of communication of three users.

  Moreover, this invention (Claim 5) is a friend recommendation program for functioning a computer as each means which comprises the friend recommendation apparatus of any one of Claim 1 thru | or 2.

  The present invention finds a typical change in the structure of friendships in the growth process of a community, particularly paying attention to three users, and using the information 2 is likely to be friends in the community. Discover users. Since typical changes in the structure of friendship are extracted for each community, it becomes possible to recommend friends to users with higher accuracy than the conventional method based on the growth pattern of the network.

It is a block diagram of the friend recommendation apparatus in one embodiment of this invention. It is an example of the communication information table in one embodiment of this invention. It is a graph showing the friend relationship of FIG. 2 in one embodiment of this invention. It is an example of the 3-partial graph information table in one embodiment of this invention. It is the position of each node in the 3-partial graph in one embodiment of this invention. It is all the patterns of the 3 node connection directed graph in one embodiment of this invention. It is an example of the edge type appearance frequency table in one embodiment of this invention. It is an edge type determined by the 3-node connected directed graph in one embodiment of the present invention. It is a flowchart of the process at the time of friend recommendation in one embodiment of this invention. It is a flowchart of the detailed process of step 200 of FIG. 0 in one embodiment of this invention. It is a flowchart of the detailed process of step 220 of FIG. 10 in one embodiment of this invention. It is a communication matrix corresponding to each 3-subgraph in one embodiment of the present invention. It is a flowchart of the detailed process of step 300 of FIG. 9 in one embodiment of this invention. It is a flowchart of the detailed process of step 340 of FIG. 13 in one embodiment of this invention. It is an edge type determination rule table (in the case of f _{u, v} = 0) in an embodiment of the present invention. It is an edge type determination rule table (in the case of f _{u, v} = 1) in an embodiment of the present invention. It is a flowchart of the detailed process of step 400 of FIG. 9 in one embodiment of this invention. It is a flowchart of the process at the time of the weight parameter update in one embodiment of this invention.

  Embodiments of the present invention will be described below with reference to the drawings.

  In the present invention, similar to Non-Patent Document 1, a friendship between users is modeled in a graph, and two users who are likely to be friends are found based on the growth constraint of the graph. The shape of a subgraph that frequently appears in the local structure of a graph called network motifs (Milo, R. et al., "Network motifs: simple building blocks of complex networks," 2002) It is a method that pays attention to. It is known that network motifs differ depending on the type of graph. Therefore, we considered that it is possible to find two users who are more likely to be friends in the future with higher accuracy than the restriction of general graph growth such as scale-freeness.

  Specifically, in general, the shape of a subgraph that is likely to appear in each graph is determined. Therefore, if a subgraph that is likely to appear can be found, it is possible to find a node between two nodes that cause the subgraph that is likely to appear, that is, two nodes that are likely to generate an edge.

  In the present invention, attention is paid particularly to a subgraph having three nodes. This is because the calculation cost increases as the number of nodes increases. As to whether or not an edge is generated between two nodes, paying attention to the shape of a subgraph consisting of another node, there are only 24 types of subgraphs due to the presence / absence of edges, directionality, and symmetry. That is, it counts to which of 24 types the shape of the subgraph consisting of one node other than the two nodes of interest is classified. Then, the two nodes can be regarded as 24-dimensional, each component being a natural number vector.

  If the presence or absence of an edge between each two nodes of the current graph is used as training data, and the two nodes without an edge are regarded as test data, the conventional vector identification model can be used. That is, the shape of a subgraph that is likely to appear in the graph is learned from the training data, and two nodes that are likely to generate edges are found with high accuracy based on the shape of the subgraph. The community is activated by recommending communication between two users corresponding to the two nodes.

  This will be specifically described below.

  FIG. 1 shows a configuration of a friend recommendation device according to an embodiment of the present invention.

  The friend recommendation device 1 shown in the figure includes a communication information extraction unit 10, a 3-part graph information table construction unit 20, an edge type appearance frequency table construction unit 30, a friend recommendation unit 40, a weight parameter update unit 50, a storage unit 60, The communication unit 70 includes an input / output unit 80. The communication unit 70 is connected to the network 3, and the input / output unit 80 is connected to an external device 4 such as an input device or a display device.

  The 3-partial graph information table construction unit 20 includes a 3-partial graph matching unit 21.

  The storage unit 60 is a storage medium that stores a communication information table 61, a 3-partial graph information table 62, an edge type appearance frequency table 63, and a weight parameter storage unit 64. Each table in the storage unit 60 will be described below.

<Communication information table>
As shown in FIG. 2, the communication information table 61 includes a user ID field and a friend ID field.

  In the user ID field, an identifier for identifying a user is set by the communication information extraction unit 10.

  In the friend ID field, the communication information extraction unit 10 sets a user ID of another user who is a friend of the user in the row.

  The friend relationship between users described in the communication information table 61 is that each user is a node, and a directional branch is extended from a node corresponding to the user described in the user ID field to a node corresponding to the user described in the friend ID field. Can be modeled as a directed graph in discrete mathematics. FIG. 3 shows an example of the communication information table 61 of FIG. 2 modeled as a directed graph.

<3-Partial graph information table>
As shown in FIG. 3, the 3-part graph information table 62 includes a position 1 user ID field, a position 2 user ID field, a position 3 user ID field, and a 3-part graph pattern field.

  In the positions 1 to 3 fields, an identifier for identifying a user corresponding to each node when the friend relationship of the three users in the row is represented by a graph is set by the 3? Partial graph information table construction unit 20. FIG. 4 shows the position of each node in the 3? Subgraph.

  In the 3-partial graph pattern field, a numerical value representing the pattern of the 3-partial graph of the row is set by the 3-partial graph information table construction unit 20.

  Here, the pattern of the 3-part graph will be described. First, a 3-partial graph that is not connected is not set in the 3-partial graph information table 62. That is, a 3-partial graph in which there is a node that does not share any branches with the other two nodes is not set. If all of the same types of such directed three-node connected graphs are classified as the same pattern, there are only 13 patterns as shown in FIG. Here, “the graph G is isomorphic to G ′” means that for any branch e = (v, w) of G, all units such that (f (v), f (w)) belong to G ′. This means that there is a projective map f, that is, it becomes exactly the same by exchanging nodes.

<Edge type appearance frequency table>
As shown in FIG. 6, the edge type appearance frequency table 63 includes a user 1 ID field, a user 2 ID field, a friend field, and edge type 1 to 24 appearance frequency fields.

  In the user 1 ID field and the user 2 ID field, all permutations of identifiers that identify users appearing in the user ID field or the friend ID field of the communication information table 61 are set by the edge type appearance frequency table construction unit 30, respectively. . However, the user 1 ID field and the user 2 ID field do not take the same value.

  The friend field is “1” when there is a row in the communication information table 61 where the user ID field has the value of the user 1 ID field of the row and the friend ID field has the value of the user 2 ID field of the row. Otherwise, “0” is set by the edge type appearance frequency table construction unit 30.

  Next, the edge type will be described. Consider all 2-node permutations (u, v) in the graph. Also, all 3-partial graphs including u and v are extracted. At this time, only a 3-partial graph that is connected even if it is assumed that there is no branch starting from u and ending at v is considered. Let another node other than u and v in such a 3-partial graph be now w. In this case, assuming that there is no edge with u as the start point and v as the end point, the entry order (how many branches end) of u, v, and w in the 3-partial graph and the output order (how many) There are only 24 possible values for the input and output orders of u, v, and w, assuming that there is an edge with u as the start point and v as the end point. For simplicity, it is as shown in FIG.

  At this time, the edge type 1-24 appearance frequency field indicates the number of ws classified as edge types 1-24 for the user 1 u and the user 2 v in the row. It is set by the frequency table construction unit 30.

  Next, the flow of processing at the time of friend registration in an embodiment of the present invention will be described. When the user u registers the user v as a friend, the communication information extraction unit 10 adds a line in the communication information table 61 where the value of the user ID field is u and the value of the friend ID field is v.

  Next, the flow at the time of friend recommendation in one embodiment of the present invention will be described with reference to FIG. The timing of recommending a friend may be when the user logs in to the system or when the system administrator explicitly sends a command to recommend a friend to the system.

  Step 100) The communication information extraction unit 10 extracts information described in the communication information table 61 from the storage unit 60, and extracts it from the 3-part graph information table construction unit 20, the edge type appearance frequency table construction unit 30, and friend recommendation. Delivered to part 40.

  Step 200) The 3-partial graph information table construction unit 20 constructs the 3-partial graph information table 62 in the storage unit 60 from the information described in the communication information table 61 delivered from the communication information extraction unit 10, and the edge type. Delivered to the appearance frequency table construction unit 30.

  Step 300) The information described in the communication information table 61 passed by the communication information table construction unit 10 by the edge type appearance frequency table construction unit 30 and the three parts delivered by the three-part graph information table construction unit 20 Using the information described in the graph information table 62, an edge type information table 63 is constructed in the storage unit 60 and transferred to the friend recommendation unit 40.

  Step 400) The friend recommendation unit 40 includes the information described in the communication information table 61 delivered by the communication information table construction unit 10 and the information described in the edge type appearance frequency table 63 delivered by the edge type appearance frequency table construction unit 30. Recommend friends to the user using the information.

  Next, the process of step 200 in FIG. 9 will be described in detail.

  FIG. 10 is a flowchart of detailed processing of step 200 in FIG. 9 according to the embodiment of the present invention.

  Step 210) The 3-part graph information table construction unit 20 has three possible values of the user ID field and friend ID field that appear in the information of the communication information table 61 delivered from the communication information table extraction unit 10. Get all combinations.

Step 220) The 3-partial graph matching unit 21 uses the user ID field combination (u ₁ , u ₂ , u ₃ ) passed from the 3-partial graph information table construction unit 20 to determine the user ID. A pattern _{u1, u2, u3} obtained by discriminating which of the 13 types of patterns the 3-partial graph corresponding to the combination (u ₁ , u ₂ , u ₃ ) matches (u ₁ , u ₂ , u _{In 3} ), user IDs p ₁ , p ₂ , and p _{3 corresponding} to position 1, position 2, and position 3 of the 3-part graph are acquired.

Step 230) The 3-part graph information table construction unit 22 determines that the value of the position 1 user ID field is p ₁ , the value of the position 2 user ID field is p ₂ , and the value of the position 3 user ID field is p ₃ . A row in which the value of the graph pattern field is pattern _{u1, u2, u3} is inserted into the 3-partial graph information table 62.

Next, the process of step 220 in FIG. 10 will be described in detail.
FIG. 11 is a flowchart of detailed processing of step 220 in FIG. 10 according to the embodiment of the present invention.

Step 221) The 3-partial graph matching unit 21 makes all six orders (u ₁ , u ₂ , u ₃ ), (u ₁ , u ₃ ) corresponding to the combinations (u ₁ , u ₂ , u ₃ ) of the user ID fields. u ₃ , u ₂ ), (u ₂ , u ₁ , u ₃ ), (u ₂ , u ₃ , u ₁ ), (u ₃ , u ₁ , u ₂ ), (u ₃ , u ₂ , u ₁ ) To get.

Step 222) Based on the information in the communication information table 61 passed from the 3-partial graph information table construction unit 20, the 3-partial graph matching unit 21 orders the user ID fields (p ₁ , p ₂ , p ₃ ) A 3 × 3 communication matrix M _{p1, p2, p3} corresponding to ₃ ) is created. Specifically, each component of the communication matrix M _{p1, p2, p3} is the value of the diagonal elements *, the value of the user ID field the value of x row y column is the p _x, and the value of friends ID field If is the line such that p _y are present in the communication information table 61 provides "1", if not "0". Note that the value of the diagonal component of the communication matrix M _{p1, p2, p3} is given by *, but this may be any other value.

FIG. 12 shows communication matrices M _k (k = 1,..., 13) in 13 types of patterns for the 3-partial graph matching process. For each component of each communication matrix, the value of the diagonal component is given as *, and the value of x rows and y columns is given as “1” when there is an edge from the xth node to the yth node, and “0” when there is no edge. . Note that the value of the diagonal component of M _k is set to match M _{p1, p2, p3} .

  Step 223) The 3-partial graph matching unit 21 prepares a temporary variable k and initializes its value to 1.

Step 224) The 3-partial graph matching unit 21 represents the k-th pattern and _{the 3} × ₃ communication matrix M _{p1, p2, p3} corresponding to the permutation (p ₁ , p ₂ , p ₃ ) of the user ID field. It is determined whether or not the communication matrix M _k matches.

Step 225) The 3-partial graph matching unit 21 _sets the value pattern _{u1, u2, u3} of the 3-partial graph pattern field corresponding to the permutation (u ₁ , u ₂ , u ₃ ) of the user ID field to k, The data is transferred to the subgraph information database construction unit 22.

  Step 226) The 3-partial graph matching unit 21 increments the temporary variable k.

  Next, the processing of step 300 in FIG. 9 will be described in detail.

  FIG. 13 is a flowchart of detailed processing of step 300 in FIG. 9 according to the embodiment of the present invention.

  Step 310) The edge type appearance frequency table construction unit 30 appears in the information described in the communication information table 61 delivered from the communication information table extraction unit 10, and includes two possible values of the user ID field and the friend ID field. Get all permutations.

Step 320) For the permutation (u, v) of the two user IDs passed from the communication information table extraction unit 10 by the edge type appearance frequency table construction unit 30, the value of the user ID field in the communication information table 61 is u, If there is a row where the value of the friend ID field is v, then f _{u, v} = 1, _and if there is no row _, f _{u, v} = 0.

Step 330) The edge type appearance frequency table construction unit 30 sets u and v in the position 1 to 3 user ID fields in the 3-partial graph information table 62 delivered from the 3-partial graph information table construction unit 20. Get the row set _{Tu, v} that contains it.

Step 340) The edge type appearance frequency table construction unit 30 updates the edge type appearance frequency table 63 based on u, v, f _{u, v} , and _{Tu, v} .

  Next, the process of step 340 in FIG. 13 will be described in detail.

  FIG. 14 is a flowchart of detailed processing of step 340 in FIG. 13 according to the embodiment of the present invention.

Step 341) Edge Type frequency table constructing unit 30, unprocessed 3- subgraph information t _{u, v} ∈T _u, for each _v, the value of the row t _{u, v} in 3 subgraph pattern field In pattern _{u, v} , the position p _u (∈ {1,2,3}) the value of the user ID field is u, the position p _v (∈ {1,2,3}), the value of the user ID field is v Suppose that

Step 342) The edge type appearance frequency table construction unit 30 uses the edge type determination rules shown in FIGS. 15 and 16 based on f _{u, v} , p _u , p _v , pattern _{u, v} to determine the edge type e of u, v. _{u and v} are determined. In this embodiment, it is assumed that the edge type determination rule is stored in advance in a memory (not shown).

15 and 16 show the edge type determination rule in step 340 of FIG. FIG. 15 shows an edge type determination rule when f _{u, v} = 0. For example, the second row and the seventh column indicate that the edge type e _{u, v} is determined to be “2” when f _{u, v} = 0, p _u = 3, p _v = 2 and pattern _{u, v} = 1. . Similarly, FIG. 16 shows an edge type determination rule when f _{u, v} = 1.

Step 343) When the edge type appearance frequency table construction unit 30 does not include a row in the edge type appearance frequency table 63 where the user 1 ID field is u and the user 2 ID field is v, the value of the user 1 ID field is u, the user 2 ID field Is inserted in the edge type appearance frequency table 63. The row of the friend field value is f _{u, v} and the edge type 1 to 24 field value is 0.

Step 344) The edge type appearance frequency table construction unit 30 obtains a row in which the value of the user 1 ID field is u and the value of the user 2 ID field is v in the information described in the edge type appearance frequency table 63, and the edge type e _{u, v} Increment the value of the field.

  Next, the process of step 400 in FIG. 9 will be described in detail.

  FIG. 17 is a flowchart of detailed processing of step 400 in FIG. 9 according to the embodiment of the present invention.

  Step 410) The friend recommendation unit 40 acquires all two permutations of the values that can be taken by the user ID field and the friend ID field that appear in the information described in the communication information table 61 delivered from the communication information extraction unit 10. .

  Step 420) The friend recommendation unit 40 determines whether or not there exists a row in which the value of the user ID field is u and the value of the friend ID field is v in the communication information table 61 delivered from the communication information extraction unit 10. To do.

Step 430) The friend recommendation unit 40 generates an edge type i (i = 1,..., 24) in the row where the value of the user 1 ID field is u and the value of the user 2 ID field is v in the edge type appearance frequency table 62. Get the value x _{u, v, i} of the frequency field.

Step 440) The friend recommendation unit 40 acquires the weight parameter λ _i (i = 1,..., 24) from the weight parameter storage unit 64.

Step 450) The friend recommendation unit 40 calculates a recommendation score score _{u, v} of v for the user u using the following equation (1).

ステップ４６０）友人推薦部４０が、前記ユーザuに対するvの推薦スコア score_u,v を用いてユーザuにvを友人として推薦すべきかどうか決定する。その際、推薦スコアの値が大きいものから順に予め設定した任意の人数N(N=1,2,3,…) 人を推薦すべき友人としてもよいし、推薦スコアに対し、ある閾値を設けて、推薦スコアがその値を超えるものを推薦するべき友人としてもよい。

Step 460) The friend recommendation unit 40 determines whether or not to recommend v to the user u as a friend using the recommendation score score _{u, v} of _v for the user u. At that time, an arbitrary number N (N = 1,2,3, ...) set in advance in descending order of the recommendation score value may be recommended friends, and a certain threshold is set for the recommendation score. A friend whose recommendation score exceeds that value may be recommended.

  Next, details of the processing at the time of updating the weight parameter will be described.

  FIG. 18 is a flowchart of processing when updating the weight parameter according to the embodiment of the present invention. The weight parameter may be automatically calculated using a logistic regression model described in Reference 2 (Hosmer, DW et al., “Applied logistic regression”, 2000) as in the present embodiment, or Reference 3 ( Cristianini, N. et al., "An introduction to Support Vector Machines: and other kernel-based learning methods", 2000) A weight parameter may be set. The update of the weight parameter may be started by an explicit input from the system administrator, or may be periodically batch processed.

  Step 510) The weight parameter updating unit 50 acquires the permutation of all two users appearing in the user ID field and friend ID field in the communication information table 61 delivered from the communication information extracting unit 10.

Step 520) The weight parameter updating unit 50 sets the row of the user type 1 ID field to u and the value of the user 2 ID field to v in the edge type appearance frequency table 63 for the permutation (u, v) of each of the two users. The value f _{u, v} of the friend field, the edge type i, and the value x _{u, v, i} of the appearance frequency field are acquired.

Step 530) The probability P (f _{u, v} = 1 | u) that the weight parameter updating unit 50 uses the equation (2) shown below for the permutation (u, v) of the two users. , v).

上記の式（２）は式（１）の値を確率化したもので、式（１）の値が高ければ高いほど値が１に近づき、式（１）の値が低ければ低いほど値が０に近づく。

The above formula (2) is a probability value of the formula (1). The higher the value of the formula (1), the closer the value is to 1, and the lower the value of the formula (1), the lower the value. Approaching zero.

Step 540) The weight parameter update unit 50 determines the likelihood of the weight parameter λ _i (i = 0,..., 24) for the data from the probability that u makes v a friend for the permutation (u, v) of all two users. The degree L (λ) is calculated using the following equation (3).

上記の式（３）は、重みパラメータが与えられたデータ（本実施の形態の場合、u、v、f_u,v）にどれだけ適合しているかを表す尺度であり、式（３）で算出される値を最大にするλ_i が最も適切な重みパラメータとなる。

The above equation (3) is a measure representing how much the weight parameter is matched with the given data (in this embodiment, u, v, f _{u, v} ). Λ _i that maximizes the calculated value is the most appropriate weight parameter.

Step 550) The weight parameter updating unit 50 uses the quasi-Newton method or the like described in Reference 4 (Liu, DC et al., “On the limited memory BFGS method for large scale optimization”, 1989). Λ _i that maximizes L (λ) represented by

Step 560) The weight parameter updating unit 50 records the calculated weight parameter λ _i in the weight parameter storage unit 64.

  As described above, according to the present embodiment, the calculation cost is limited to three users (nodes), and the three nodes can be classified into 24 types by paying attention to the edge, directionality, and objectivity of the three nodes. By determining the frequency of actual friendships that occur in each of the 24 types of patterns, it is possible to determine which pattern is likely to cause friendships. It is possible to detect a pattern in which a friend relationship is likely to occur, calculate a recommendation score, and output a user having a high recommendation score as a friend to be recommended.

  In addition, the operation | movement of each component of the friend recommendation apparatus shown in FIG. 1 of said embodiment is constructed | assembled as a program, installed in the computer used as a friend recommendation apparatus, is performed, or distribute | circulates via a network. It is possible.

  In addition, the constructed program can be stored in a portable storage medium such as a hard disk, a flexible disk, or a CD-ROM, and can be installed or distributed in a computer.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

DESCRIPTION OF SYMBOLS 1 Friend recommendation apparatus 2 External terminal 3 Network 4 External apparatus 10 Communication information extraction part 20 3-Partial graph information table construction part 21 3-Partial graph matching part 30 Edge type appearance frequency table construction part 40 Friend recommendation part 50 Weight parameter update part 60 Storage Unit 61 Communication Information Table 62 3-Partial Graph Information Table 63 Edge Type Appearance Frequency Table 64 Weight Parameter Storage Unit 70 Communication Unit 80 Input / Output Unit

Claims (5)

A friend recommendation device for recommending friends in an online community system where each user has an ID and aims to communicate between users,
Communication information acquisition means for acquiring a user in the online community and storing it in a communication storage means;
3-partial graph information acquisition means for classifying the three users in the online community from the communication information in the communication storage means, based on the presence / absence pattern of communication, and storing it in the 3-partial graph storage means as 3-node graph information;
From the communication information in the communication storage means, information regarding the presence or absence of communication for two users in the online community, and from the 3-partial graph information in the 3-partial graph storage means, for two users in the online community, other than two users Edge type information acquisition means for acquiring the frequency of which one of the other users is classified according to the pattern of presence / absence of communication of three users, and storing it in the edge type storage means;
Information regarding the presence / absence of communication stored in the edge type storage means for two users in the online community, and the number of the other one user classified as a pattern of presence / absence of communication of three users, respectively A weight parameter obtaining means for calculating a weight parameter and storing the weight parameter in a weight parameter storage means,
When a user identifier is given, information on the presence / absence of communication, the number of which the other one user is classified by the pattern of presence / absence of communication of three users, and the weight of the weight parameter storage means Using a parameter, friend recommendation means for recommending another user as a friend to the user,
A friend recommendation device characterized by comprising: The weight parameter calculating means includes
A means for calculating a weight parameter by logistic regression analysis based on the information on the presence / absence of communication and the number of each of the other users classified according to the presence / absence pattern of the communication of three users. Item 3. The friend recommendation device according to Item 1. In an online community system where each user has an ID and aims to communicate between users, a friend recommendation method for recommending friends,
Communication storage means, 3-part graph storage means, edge type storage means, weight parameter storage means are stored, communication information acquisition means, 3-part graph information acquisition means, edge type information acquisition means, weight parameter acquisition means, friend recommendation A device having means for:
The communication information acquisition unit acquires a user in the online community, and stores the communication information in the communication storage unit;
The 3-partial graph information acquisition unit classifies the three users in the online community from the communication information in the communication storage unit, and stores them in the 3-partial graph storage unit as 3-node graph information. A 3-partial graph information acquisition step;
The edge type information acquisition means obtains information about the presence / absence of communication for two users in the online community from the communication information in the communication storage means, and the online community from the 3-partial graph information in the 3-partial graph storage means. An edge type information acquisition step of acquiring a frequency of whether another user other than the two users is classified according to a communication presence / absence pattern of the three users, and storing the frequency in the edge type storage unit ,
The weight parameter acquisition means is the information regarding the presence / absence of communication stored in the edge type storage means for the two users in the online community and the pattern of the presence / absence of communication of the other one user for each of the three users. A weight parameter obtaining step of calculating a weight parameter using the number of classification and storing it in the weight parameter storage means;
When the friend recommendation means is given a user identifier, information on the presence / absence of the communication in the 3-partial graph storage means, and whether the other one user in the edge type storage means is the presence / absence of communication of three users, respectively. A friend recommendation step of recommending another user as a friend to the user by using the number of the patterns classified into the pattern and the weight parameter of the weight parameter storage means;
Friend recommendation method characterized by performing. In the weight parameter calculating step,
4. The weight parameter is calculated by logistic regression analysis based on the information on the presence / absence of communication and the number of each of the other users classified according to the pattern of presence / absence of communication of three users. Friend recommendation method.   The friend recommendation program for functioning a computer as each means which comprises the friend recommendation apparatus of any one of Claim 1 thru | or 2.

Images