CN103700018A - Method for dividing users in mobile social network - Google Patents

Abstract

The invention discloses a method for dividing users in a mobile social network. The method comprises the following steps of a, importing user geographical track journal files, and calculating disperse geographical track similarity si,j between any two users; b, establishing a User-User similarity matrix G, a Locate-User matrix L and a User-Locate matrix U based on the disperse geographical track similarity between any two users; c, importing user social activity loggings, establishing a social network model according to mobile phone calling historical records between users, and calculating a social network adjacent matrix F; d, calculating a block matrix and dividing the users. According to the method, the users in a mobile phone network or in a position-based on-line social network are accurately divided, and the method can be applied to a personalized content recommendation system and pushes personalized content to target users.

Description

Crowd's division methods in a kind of mobile community network

Technical field

The present invention relates to data mining and user behavior analysis technology, be specifically related to the crowd's division methods in a kind of mobile community network.By the user in cell phone network or the online social networks of position-based, precisely divide, can be applicable to individualized content commending system.

Background technology

Accurately the corporations in division crowd or cluster are significant for individualized content commending systems such as optimizing location-based Internet service, according to crowd's division result, can to targeted customer, push individualized content exactly.

Existing division crowd's method generally can be divided into based on community network Detecting Community method and the clustering method based on data characteristics.

On the one hand, the method of the Detecting Community based on community network is divided traditional corporations according to the tightness degree of the social networks between individuality, the social networks inside and outside corporations can be reflected very intuitively in the corporations that obtain due to this method, therefore be easy to for analyzing and explain some universal phenomena, can be applied to very easily that pure social networking service is recommended or Study on Problems that other and social networks are closely related in, but crowd might not only have social attribute, other attributes are geographical attribute for example, social propertys etc. have influence on individual behavior and even the division of corporations equally, and then affect its application surface.

On the other hand, the data clustering method based on common trait has had a lot of application at Data Mining.The similarity of the attribute of this method based on Different Individual dimension obtains the cluster of the common trait between individuality, thereby divides dissimilar crowd.This method has been widely applied in personalized recommendation fields such as ecommerce, content of multimedia propelling movement, but this method specific aim is relatively strong, only can analyze and the behavior of detecting user in particular community dimension.

Summary of the invention

For the deficiencies in the prior art, the object of the invention is to propose the crowd's division methods in a kind of mobile community network, the method, based on discrete geographical track similarity, simultaneously in conjunction with individual geographical attribute and social attribute, is divided more accurate.

For realizing above goal of the invention, the present invention by the following technical solutions:

Crowd's division methods in mobile community network, comprises the following steps:

A, the geographical trace logs file of importing user, calculate the discrete geographical track similarity s between any two users _i,j, wherein, given total number of users N, User Activity region is covered completely by M base station, and the geographical trace logs of described user is comprised of the base station sequence number list of user ID and User Activity;

B, the discrete geographical track similarity based on any two users, set up User-User similarity matrix G, Locate-User matrix L and User-Locate matrix U, wherein, and the matrix element PV of N*M dimension User-Locate matrix U _ijrepresent that user i is at the probability of occurrence of position j; The matrix element LV of M*N dimension Locate-User matrix L _ijrepresent the probability that i upper user j in position occurs; N*N dimension User-User similarity matrix G is:

Wherein, the right geographical similarity of user of correspondence position in each matrix element respective user list;

C, importing user social contact activity log, set up social networks model according to user's handset call historical record each other, calculates social networks adjacency matrix F, and wherein, social networks adjacency matrix F is N*N dimension matrix, matrix element V _ijrepresent user i, the weight of the social networks of j;

D, calculate partitioned matrix and divide crowd.

Compared with prior art, the present invention has following technique effect: owing to coming cluster to have the individuality of similar behavior in conjunction with multidimensional data, divide more accurate.

Accompanying drawing explanation

Explanation with reference to below, by reference to the accompanying drawings, can have best understanding to the present invention.In the accompanying drawings, identical part can be represented by identical label.

Fig. 1 is crowd's division methods general frame schematic diagram that the present invention proposes;

Fig. 2 is social networks adjacency matrix calculation flow chart;

Fig. 3 is the process flow diagram of crowd's partiting step.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with accompanying drawing and exemplary embodiment, the present invention is further elaborated.Should be appreciated that exemplary embodiment described herein is only in order to explain the present invention, the scope of application being not intended to limit the present invention.

As shown in Figure 1, the crowd's division methods in the mobile community network of the embodiment of the present invention comprises the following steps:

A, the geographical trace logs file of importing user, calculate the discrete geographical track similarity between any two users.

Wherein, the geographical trace logs of user is comprised of the base station sequence number list of user ID and User Activity.Given total number of users N, User Activity region is covered completely by M base station, any user i, the discrete geographical track similarity between j is expressed as the cosine similarity indices sum of both all common location:

$S_{\mathrm{ij}}=\underset{l\∈L,i,j\∈N}{\mathrm{\Σ}}\frac{p_{i,l}{p}_{j,l}}{\left|\right|p_{i,l}\left|\right|\×\left|\right|p_{j,l}\left|\right|}$

P wherein _i,lthe probability that user occurs at position l, || p _i,l|| be p _i,lnorm.In the present embodiment, this norm is 2 norms.

Step a is specially: first import any two users' historical base station sequence number list, contrast both common base station numbers and account for ratio separately, calculate the probability of occurrence p of each comfortable all total base station location l _i,l, then calculate corresponding cosine similarity indices, finally summation obtains any two users' discrete geographical track similarity.

B, the discrete geographical track similarity based on any two users, set up User-User similarity matrix G, Locate-User matrix L and User-Locate matrix U.

Step b specifically comprises: according to the discrete geographical track similarity s between individuality _i,jset up User-User similarity matrix G, set up User-Locate matrix U and Locate-User matrix L simultaneously, utilize symmetrical Algorithms of Non-Negative Matrix Factorization (SNMF) to carry out matrix decomposition to User-User similarity matrix G, obtain the Preliminary division of individual geographic track, i.e. the degree of adhering to separately of affiliated each partitioned matrix.

Wherein, the matrix element PV of N*M dimension User-Locate matrix U _ijrepresent that user i is at the probability of occurrence of position j; The matrix element LV of M*N dimension Locate-User matrix L _ijrepresent the probability that i upper user j in position occurs; N*N dimension User-User similarity matrix G is:

Wherein, the right geographical similarity of user of correspondence position in each matrix element respective user list.

C, importing user social contact activity log, set up social networks model according to user's handset call historical record each other, calculates social networks adjacency matrix F.

Wherein, social networks adjacency matrix F is N*N dimension matrix, matrix element V _ijrepresent user i, the weight of the social networks of j.In the present embodiment, described weight can be user i, the social interaction frequency of j, i.e. user i, the number of communications between j and the duration of call.

Step c is specially: import N user's doings daily record, initialization N*N social networks adjacency matrix F=0, targeted customer corresponding in each user's doings daily record is read in circulation, according to the weight of social networks between individuality, upgrade corresponding matrix element in social networks adjacency matrix, thereby set up the social networks adjacency matrix of colony, as shown in Figure 2.

D, calculate partitioned matrix and divide crowd.

Steps d is specially: import the social networks adjacency matrix F in User-User similarity matrix G, User-Locate matrix U, Locate-User matrix L and the step c in step b, the initial value X of target setting partitioned matrix carries out to User-User similarity matrix G degree of adhering to separately the matrix that matrix decomposition obtains in step b, adopt associating Algorithms of Non-Negative Matrix Factorization (JNMF) to do matrix decomposition, iterate until target partitioned matrix before and after meeting, the mould of the difference of twice iteration is less than predetermined value, exports each node affiliated each partitioned matrix and degree of adhering to separately thereof, obtains the dividing condition of colony, as shown in Figure 3.

According to such scheme, input above-mentioned four matrix G, U, L, F and do associating Non-negative Matrix Factorization, being calculated as follows of its target partitioned matrix (N*K matrix):

${\hat{X}}_{\mathrm{ik}}\←{\hat{X}}_{\mathrm{ik}}{\left(\frac{[U+2\mathrm{\αFX}+X\hat{L}}{2(\mathrm{\α}+\mathrm{\β}){\left[{\mathrm{XX}}^{T}X\right]}_{\mathrm{ij}}}\right)}^{1/4}$ Wherein $\hat{L}=2\mathrm{\βL}-1$

In the present embodiment, preset value α=β=0.5, k=1,2 ..., K, K is default geographical corporations number, scalar 1 is the unit matrix of corresponding dimension (the diagonal element of 2 β L all deducts 1).After final the stablizing that iterate, obtain required target partitioned matrix wherein the capable representative of consumer i of i is for the degree of membership of each geographical corporations.

Steps d obtains after crowd's division result, and individualized content commending system can push individualized content to targeted customer according to the crowd who divides.

The foregoing is only preferred embodiment of the present invention, not in order to limit the present invention, all any modifications of doing within the spirit and principles in the present invention, be equal to and replace and improvement etc., within all should being included in protection scope of the present invention.

Claims (7)

1. the crowd's division methods in mobile community network, comprises the following steps:

A, the geographical trace logs file of importing user, calculate the discrete geographical track similarity s between any two users _i,j, wherein, given total number of users N, User Activity region is covered completely by M base station, and the geographical trace logs of described user is comprised of the base station sequence number list of user ID and User Activity;

B, the discrete geographical track similarity based on any two users, set up User-User similarity matrix G, Locate-User matrix L and User-Locate matrix U, wherein, and the matrix element PV of N*M dimension User-Locate matrix U _ijrepresent that user i is at the probability of occurrence of position j; The matrix element LV of M*N dimension Locate-User matrix L _ijrepresent the probability that i upper user j in position occurs; N*N dimension User-User similarity matrix G is:

Wherein, the right geographical similarity of user of correspondence position in each matrix element respective user list;

C, importing user social contact activity log, set up social networks model according to user's handset call historical record each other, calculates social networks adjacency matrix F, and wherein, social networks adjacency matrix F is N*N dimension matrix, matrix element V _ijrepresent user i, the weight of the social networks of j;

D, calculate partitioned matrix and divide crowd.

2. division methods according to claim 1, wherein, user i arbitrarily, the discrete geographical track similarity between j is expressed as the cosine similarity indices sum of both all common location:

$S_{\mathrm{ij}}=\underset{l\∈L,i,j\∈N}{\mathrm{\Σ}}\frac{p_{i,l}{p}_{j,l}}{\left|\right|p_{i,l}\left|\right|\×\left|\right|p_{j,l}\left|\right|}$

P wherein _i,lthe probability that user occurs at position l, || p _i,l|| be p _i,lnorm, be preferably 2 norms.

3. division methods according to claim 2, wherein, step a is specially:

First import any two users' historical base station sequence number list, contrast both common base station numbers and account for ratio separately, calculate the probability of occurrence p of each comfortable all total base station location l _i,l, then calculate corresponding cosine similarity indices, finally summation obtains any two users' discrete geographical track similarity.

4. division methods according to claim 1, wherein, step b also comprises:

Utilize symmetrical Algorithms of Non-Negative Matrix Factorization (SNMF) to carry out matrix decomposition to User-User similarity matrix G, obtain the Preliminary division of individual geographic track, be i.e. the degree of adhering to separately of affiliated each partitioned matrix.

5. division methods according to claim 1, wherein, described user i, the weight of the social networks of j is specially user i, the social interaction frequency of j, i.e. user i, the number of communications between j and the duration of call.

6. division methods according to claim 1, wherein, step c is specially: import N user's doings daily record, initialization N*N social networks adjacency matrix F=0, targeted customer corresponding in each user's doings daily record is read in circulation, according to the weight of social networks between individuality, upgrade corresponding matrix element in social networks adjacency matrix, thereby set up the social networks adjacency matrix of colony.

7. division methods according to claim 1, wherein, steps d is specially: import the social networks adjacency matrix F in User-User similarity matrix G, User-Locate matrix U, Locate-User matrix L and the step c in step b, the initial value X of target setting partitioned matrix carries out to User-User similarity matrix G degree of adhering to separately the matrix that matrix decomposition obtains in step b, adopt associating Algorithms of Non-Negative Matrix Factorization (JNMF) to do matrix decomposition, iterate until target partitioned matrix

before and after meeting, the mould of the difference of twice iteration is less than predetermined value, exports each node affiliated each partitioned matrix and degree of adhering to separately thereof, obtains the dividing condition of colony.

Images