CN103942278A - Method for conducting friend recommendation through analysis of user active friends making will - Google Patents

Abstract

The invention discloses a method for conducting friend recommendation through analysis of the user active friends making will and relates to the technical field of computer social networks. The method comprises the steps that a user friends making will characteristic matrix P and a user friends making will characteristic matrix Q are initialized, a directed adjacent matrix H is worked out, inconsistency of the directed adjacent matrix H is calculated through loop iteration, the user friends making will characteristic matrix P and the user friends making will characteristic matrix Q are updated through iteration, and a new directed adjacent matrix H is worked out; according to the directed adjacent matrix H, an initiator and a receiver of each relation pair are distinguished, active friends and passive friends of the user are distinguished, and accordingly, new friends are recommended to each user. According to the method for conducting friend recommendation through analysis of the user active friends making will, the active friends making will of each user in an undirected social relation is reasonably judged, so that personal interest and friends making preference of each user can be more accurately captured, a more accurate friend recommendation service can be provided for the undirected social networks, all kinds of analysis work of the social networks can be facilitated.

Description

Carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances

Technical field

The present invention relates to computing machine social networks technical field, particularly relate to and a kind ofly carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances.

Background technology

Nowadays online social networks has been dissolved in the middle of people's life, by the online social network-i i-platform such as Renren Network, microblogging, micro-letter make friends, sharing information, interaction entertainment, become people's one way of life mode.The network operator of social network-i i-platform often wishes user social contact relation and Social behaviors to carry out deep analysis mining, and then for user provides personalized recommendation service, to strengthen user's liveness and viscosity.Friend recommendation is one of the most typical in social networks, modal recommendation scene.The conventional method of friend recommendation is, by the existing friend of analysis user, to excavate and extract user's individual preference, and then carry out friend recommendation according to user's preference.Therefore, the excavation of the existing friended analysis of user and individual subscriber preference is to provide to friend recommendation accurately and serves important basic step.

Online social networks generally can be divided into directed networks and the large class of undirected network two.Wherein, the friends in directed networks has directive property, not reciprocity as two customer relationships of friend.Such as on microblogging, a user (being called promoter) pays close attention to another user (being called recipient) on one's own initiative, and recipient might not send concern promoter.We claim to be oriented relation with the friends of obvious directive property like this.And in other social networks, friends is without obvious directive property, two customer relationships in the external world as friend are reciprocity.Such as in Renren Network, we can only see that certain two user is friends, and cannot distinguish promoter or recipient wherein.We claim that there is no like this friends of obvious directive property is undirected relation.

For the oriented relation of a pair of friend's composition, we are also called recipient promoter's a positive friend, promoter are called to recipient's a passive friend.Obviously, compare passive friend, actively friend is the friend that user initiatively adds and makes friends with, and has more embodied user's friend-making demand and wish.Therefore, the feature that we should pay attention to all positive friends to user more in oriented social networks is analyzed, and then recommends more actively friend for user.In undirected social networks, we general None-identified user's positive friend and passive friend, therefore traditional recommend method it is generally acknowledged user and its friended relation be all reciprocity.But, in our real life, relation between friend often and non-correspondence, the production process of relation remains directive, different users has different wishes for the generation of friends, just its directivity often cannot be observed, and this has just brought difficulty to the friend recommendation in undirected social networks.If directivity potential in undirected relation and the user wish of making friends is not identified, just possibly cannot really identify user's request, produce low-quality recommendation results.

Therefore, need at present the urgent technical matters solving of those skilled in the art to be exactly: how can innovate and to propose a kind of effectively recommend method, thereby realize the oriented relation being implied between friend in undirected social networks of identifying better, infer user's friend-making wish, carry out friend recommendation to user better.

Summary of the invention

Technical matters to be solved by this invention is to provide a kind ofly carries out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, analyze judgement for the wish that takes the initiative in making acquaintances of the user in undirected social networks and undirected social networks, identify promoter and the recipient of every a pair of relation, distinguish user's positive friend and passive friend, and then friend recommendation service is more accurately provided.

In order to solve the problems of the technologies described above, the embodiment of the invention discloses and a kind ofly carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, comprising:

Initialization user make friends wish eigenmatrix P and Q, and calculate initial Digraph adjacent matrix H;

Obtain the comprehensive inconsistency of Digraph adjacent matrix H obtained above;

For the every a pair of friends in social networks or each user, about the respective element in gradient updating matrix P, the Q of P and Q, and obtain the Digraph adjacent matrix H that new comprehensive inconsistency is lower according to Digraph adjacent matrix H;

For the every a pair of friends in social networks, according to the final Digraph adjacent matrix H obtaining, identify promoter and recipient wherein, and then distinguish each user's positive friend and passive friend;

According to the Digraph adjacent matrix H obtaining, make friends with new friend's wish intensity according to each user, produce the new friend's who recommends to each user list.

Preferably, described comprehensive inconsistency comprises degree inconsistency, ternary structural inconsistency, similarity inconsistency, collaborative inconsistency and the complexity of Digraph adjacent matrix H.

Preferably, the comprehensive inconsistency of obtaining Digraph adjacent matrix H obtained above described in adopts loop iteration account form to obtain.

Preferably, described initialization user make friends feature matrix adopt random fashion complete.

Preferably, the respective element in described renewal matrix P, Q adopts iterative manner to carry out.

Compared with prior art, the present invention has the following advantages:

The inventive method is based on sociological principles, the wish that takes the initiative in making acquaintances to user in undirected social networks is carried out legitimate inference, thereby catch more exactly user's personal interest and friend-making preference, avoid not distinguishing the drawback that user makes friends wish, all friend's equities are treated in traditional recommend method, can provide friend recommendation service more accurately for undirected social networks.

Simultaneously, the present invention also provides a kind of general method that general undirected social networks is converted into oriented social networks, all kinds of further analytical work of social networks be can be conducive to, the calculating of social networks user force, community discovery, Praise etc. included, without being limited to.

Brief description of the drawings

Fig. 1 of the present inventionly a kind ofly carries out the schematic flow sheet of the embodiment of the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances;

Fig. 2 is the schematic diagram of introducing scheme of the present invention in embodiment by concrete algorithm.

Embodiment

For above-mentioned purpose of the present invention, feature and advantage can be become apparent more, below in conjunction with the drawings and specific embodiments, the present invention is further detailed explanation.

Referring to Fig. 1, a kind of described in this programme carries out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, and specifically comprises:

Step S101, initialization user make friends wish eigenmatrix P and Q, and calculate initial Digraph adjacent matrix H;

Step S102, obtains the comprehensive inconsistency of Digraph adjacent matrix H obtained above;

Step S103, for the every a pair of friends in social networks or each user, about the respective element in gradient updating matrix P, the Q of P and Q, and obtains the Digraph adjacent matrix H that new comprehensive inconsistency is lower according to Digraph adjacent matrix H;

Step S104, for the every a pair of friends in social networks, according to the final Digraph adjacent matrix H obtaining, identifies promoter and recipient wherein, and then distinguishes each user's positive friend and passive friend;

Step S105, according to the Digraph adjacent matrix H obtaining, makes friends with new friend's wish intensity according to each user, produce the new friend's who recommends to each user list.

For making those skilled in the art understand better the present invention, introduce this programme realization in actual applications below in conjunction with Fig. 2 by concrete algorithm:

(1) use non-directed graph G=<V, E> is illustrated in the undirected social networks of given time t, wherein V represents the set of all users in given undirected social networks, and E represents the set of all friends in given undirected social networks.If the number of users in G is U.For each user i, establish the friend who newly makes friends with after moment t for i gathers, for having common good friend with i but be not the friend's of i user set. in each user d be called the new friend of i, in each user l be called the non-friend of i.

(2) establish potential intrinsic dimensionality K(span and be generally 5-50), the user of the respectively random initializtion K × U dimension each element span in wish eigenmatrix P and Q(matrix of making friends is 0-1), use p _irepresent the i row in P, use q _jrepresent the j row in Q.Make Digraph adjacent matrix H=P ^tq.

(3) directed networks metric module is calculated the inconsistency of Digraph adjacent matrix H:

(3-1) calculate according to the following formula, the degree inconsistency C of H _d(H):

$C_d\left(H\right)=\underset{(i,j)\&Element;E}{\mathrm{\&Sigma;}}\left(\mathrm{\&gamma;}(H_{i,j}-H_{j,i},d_j^{\left(\mathrm{in}\right)}-d_i^{\left(\mathrm{in}\right)}+\mathrm{\&gamma;}{H}_{i,j}-H_{j,i},d_i^{\left(\mathrm{out}\right)}-d_j^{\left(\mathrm{our}\right)})\right)$

Wherein Υ ^Υfunction definition is:

$\mathrm{\&gamma;}(x,y)=\left\{\begin{array}{cc}0,& \mathrm{sgn}\left(x\right)=\mathrm{sgn}\left(y\right)\\ -\frac{x}{y},& \mathrm{otherwise}\end{array}\right.,$

with be defined as:

$d_i^{\left(\mathrm{in}\right)}=\underset{u=1}{\overset{U}{\mathrm{\&Sigma;}}}{H}_{u,i},d_i^{\left(\mathrm{out}\right)}=\underset{u=1}{\overset{U}{\mathrm{\&Sigma;}}}{H}_{i,u},$

with definition similar with it.

(3-2) calculate according to the following formula, the ternary structural inconsistency C of H _t(H):

$C_t\left(H\right)=\underset{i\&Element;V}{\mathrm{\&Sigma;}}\underset{m,n\&Element;F\left(i\right)}{\mathrm{\&Sigma;}}\mathrm{\&gamma;}(H_{i,m}-H_{i,n},H_{n,m}-H_{m,n}),$

Wherein F (i) represents all friend's set of i.

(3-3) calculate according to the following formula, the similarity inconsistency C of H _s(H):

$C_s\left(H\right)=\underset{i\&Element;V}{\mathrm{\&Sigma;}}\underset{m,n\&Element;F\left(i\right)}{\mathrm{\&Sigma;}}{(\mathrm{sim}(m,n)\&CenterDot;{(H_{i,m}-H_{i,n})}^2)}^{s(m,n)},$

Wherein sim (m, n) is defined as:

$\mathrm{sim}(m,n)=\frac{|F\left(m\right)\&cup;\mathrm{UF}\left(n\right)|}{|F\left(m\right)\&cup;\mathrm{UF}\left(n\right)|},$

S (m, n) is defined as:

$s(m,n)=\left\{\begin{array}{cc}1,& \mathrm{sim}(m,n)>0.6\\ 0,& 0.4\&le;\mathrm{sim}(m,n)\&le;0.6\\ -1,& \mathrm{sim}(m,n)<0.4\end{array}\right..$

(3-4) calculate according to the following formula, the collaborative inconsistency C of H _c(H):

$C_c\left(H\right)=\underset{i\&Element;V}{\mathrm{\&Sigma;}}\underset{d\&Element;D_i^{+},l{\&Element;}_i^{-}}{\mathrm{\&Sigma;}}{(H_{i,l}-H_{i,d})}^2.$

(3-5) calculate according to the following formula, the complexity Ω (H) of H:

$\mathrm{\&Omega;}\left(H\right)=\underset{(i,j)\&Element;E}{\mathrm{\&Sigma;}}({(H_{i,j}+H_{j,i}-1)}^2+\mathrm{\&Gamma;}\left(H_{i,j}\right)+\mathrm{\&Gamma;}(1-H_{i,j})),$

Wherein Γ ^Γfunction definition is:

$\mathrm{\&Gamma;}\left(x\right)=\left\{\begin{array}{cc}{x}^2,& x<0\\ 0,& \mathrm{otherwise}\end{array}\right.,$

(3-6) calculate according to the following formula, the inconsistency C (H) of H:

C(H)＝λ _d·C _d(H)+λ _t·C _t(H)+λ _s·C _s(H)+λ _c·C _c(H)+λ·Ω(H),

Wherein λ _d, λ _t, λ _s, λ _cbe weight coefficient with λ, general value is: λ _d=λ _t=λ _s=λ _c=1, λ=0.1.

If an inconsistency threshold value Δ C, if C (H)≤Δ C carries out step (5); Otherwise carry out step (4).

(4) directed networks inference module according to H gradient updating matrix P, the Q about P and Q, calculate new Digraph adjacent matrix H:

(4-1) for every a pair of friends (i, j) in E, if H _i,j-H _j,iwith contrary sign, according to respective column in following various renewal P and Q:

$p_i=p_i-\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;q_j}{d_i^{\left(\mathrm{in}\right)}-d_j^{\left(\mathrm{in}\right)}},p_j=p_j+\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;q_i}{d_i^{\left(\mathrm{in}\right)}-d_j^{\left(\mathrm{in}\right)}},$

$q_i=q_i+\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;p_j}{d_i^{\left(\mathrm{in}\right)}-d_j^{\left(\mathrm{in}\right)}},q_j=q_j-\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;p_i}{d_i^{\left(\mathrm{in}\right)}-d_j^{\left(\mathrm{in}\right)}}.$

(4-2) for every a pair of friends (i, j) in E, if H _i,j-H _j,iwith contrary sign, according to respective column in following various renewal P and Q:

$p_i=p_i-\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;q_j}{d_j^{\left(\mathrm{out}\right)}-d_i^{\left(\mathrm{out}\right)}},p_j=p_j+\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;q_i}{d_j^{\left(\mathrm{out}\right)-d_i^{\left(\mathrm{out}\right)}}},$

$q_i=q_i+\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;p_j}{d_j^{\left(\mathrm{out}\right)}-d_i^{\left(\mathrm{out}\right)}},q_j=q_j-\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_d\&CenterDot;p_i}{d_j^{\left(\mathrm{out}\right)-d_i^{\left(\mathrm{out}\right)}}},$

(4-3) for each user i in V, for each friend m and the n of i, if H _i,n-H _i,mwith H _m,n-H _n,mcontrary sign, according to respective column in following various renewal P and Q:

$p_i=p_i-\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_t\&CenterDot;(q_m-q_n)}{H_{m,n}-H_{n,m}},q_m=q_m-\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_t\&CenterDot;p_i}{H_{m,n}-H_{n,m}},q_n=q_n+\mathrm{\&mu;}\&CenterDot;\frac{{\mathrm{\&lambda;}}_t\&CenterDot;p_i}{H_{m,n}-H_{n,m}}$

(4-4) similarity lower limit threshold values s is set ₁(general value is 0.5-0.7) and distinctiveness ratio upper limit threshold s ₂(general value be 0.3-0.5) for each user i in V, each friend m and n for i:

If sim (m, n) > is s ₁, order

w＝ _μ·λ _s·sim(m,n)·(H _i,m-H _i,n),

If sim (m, n) < is s ₂, order

$w=-2\&CenterDot;\mathrm{\&mu;}\&CenterDot;{\mathrm{\&lambda;}}_s\&CenterDot;\frac{1}{\mathrm{sim}(m,n)\&CenterDot;{(H_{i,m}-H_{i,n})}^3};$

If s ₂≤ sim (m, n)≤s ₁, order

w＝0.

And according to respective column in following various renewal P and Q:

p _i＝p _i-w·(q _m-q _n),q _m＝q _m-w·p _i,q _n＝q _n+w·p _i

(4-5), for each user i in V, for each new friend d and the non-friend l of i, make w=H _i,d+ H _d,i+ H _i,l+ H _l,i, and according to respective column in following various renewal P and Q:

p _i＝p _i-2·μ·λ _c·w·(q _l-q _d),

p _l＝p _l-2·μ·λ _c·w·q _i,

p _d＝p _d+2·μ·λ _c·w·q _i,

q _i＝q _i-2·μ·λ _c·w·(p _l-p _d),

q _l＝q _l-2·μ·λ _c·w·p _i,

q _d＝q _d+2·μ·λ _c·w·p _i.

(4-6) make H '=H, and according to new P and Q, calculate new directed networks adjacency matrix H:

H＝P ^TQ.

Calculate the variable quantity of H: δ H=||H-H ' ||, and establish a variable quantity threshold value Δ H(span and be generally 0-1), if δ H≤Δ H carries out step (5); Otherwise return to step (3).

(5) promoter and the recipient of the every a pair of relation of directed networks generation module identification, differentiation user's positive friend and passive friend.

(5-1) initialization directed networks add therein whole user V of undirected network G.

(5-2) a general value of wish difference limen value δ h(being set is 0-0.3), the every a pair of friends (i, j) in E: if H _i,j-H _j,i> δ h, thinks that i is promoter, and j is recipient, directed edge i->j of middle interpolation, represents that j is a positive friend of i, and i is a passive friend of j; If H _j,i-H _i,j> δ h, thinks that j is promoter, and i is recipient, directed edge j->i of middle interpolation, represents that i is a positive friend of j, and j is a passive friend of i; If | H _j,i-H _i,j|≤δ h, think that i and j are common promoter, two directed edge i->j of middle interpolation and j->i, represent that j is a positive friend of i, and i is also a positive friend of j.

(6) friend recommendation module is that each user recommends new friend according to the directed networks adjacency matrix H obtaining.For user i, each capable element value H of i in H _i,jembody the wish intensity that i initiatively makes friends with corresponding user j.By the friend's of each non-i user j according to its corresponding H _i,jvalue is sequence from high to low, as the list of friends of recommending to i.

Each embodiment in this instructions all adopts the mode of going forward one by one to describe, and what each embodiment stressed is and the difference of other embodiment, between each embodiment identical similar part mutually referring to.

A kind ofly be described in detail by the analysis user method that wish carries out friend recommendation that takes the initiative in making acquaintances provided by the present invention above, applied specific case herein principle of the present invention and embodiment are set forth, the explanation of above embodiment is just for helping to understand method of the present invention and core concept thereof; , for one of ordinary skill in the art, according to thought of the present invention, all will change in specific embodiments and applications, in sum, this description should not be construed as limitation of the present invention meanwhile.

Claims (5)

1. carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, it is characterized in that, comprising:

Initialization user make friends wish eigenmatrix P and Q, and calculate initial Digraph adjacent matrix H;

Obtain the comprehensive inconsistency of Digraph adjacent matrix H obtained above;

For the every a pair of friends in social networks or each user, about the respective element in gradient updating matrix P, the Q of P and Q, and obtain the Digraph adjacent matrix H that new comprehensive inconsistency is lower according to Digraph adjacent matrix H;

For the every a pair of friends in social networks, according to the final Digraph adjacent matrix H obtaining, identify promoter and recipient wherein, and then distinguish each user's positive friend and passive friend;

According to the Digraph adjacent matrix H obtaining, make friends with new friend's wish intensity according to each user, produce the new friend's who recommends to each user list.

2. as claimed in claim 1ly carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, it is characterized in that, described comprehensive inconsistency comprises degree inconsistency, ternary structural inconsistency, similarity inconsistency, collaborative inconsistency and the complexity of Digraph adjacent matrix H.

3. as claimed in claim 1ly carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, it is characterized in that, described in obtain Digraph adjacent matrix H obtained above comprehensive inconsistency adopt loop iteration account form to obtain.

4. as claimed in claim 1ly carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, it is characterized in that, the described initialization user feature matrix of making friends adopts random fashion to complete.

5. as claimed in claim 1ly carry out the method for friend recommendation by the analysis user wish that takes the initiative in making acquaintances, it is characterized in that, the respective element in described renewal matrix P, Q adopts iterative manner to carry out.