CN106372072A - Location-based recognition method for user relations in mobile social network - Google Patents

Abstract

The invention discloses a location-based recognition method for user relations in a mobile social network. The method comprises the steps that the mobile social network is established based on user data, a factor graph model is established through feature extraction of user behaviors, and then model parameter learning and multi-relation concurrent deduction are carried out, so that the user relations can be obtained. Specifically, the user data is pre-processed, so that sampling data is acquired; user behavior features including interactive behavior features and spatial behavior features are extracted according to the sampling data; an interaction factor, a space factor and a group factor of the user relations are obtained; the factor graph model is established; the parameter learning training is conducted aiming at the factor graph model; and the trained factor graph model is used for relation recognition through a multi-relation concurrent deduction method, so that the user multi-relations can be obtained. The method disclosed by the invention considers interactive features of space locations, users' surrounding environments, and interactive effects in family and colleague relations, so that accuracy of the relation recognition can be increased.

Description

A kind of recognition methodss of location-based mobile community network customer relationship

Technical field

The present invention relates to customer relationship recognition methodss, the knowledge of the customer relationship of more particularly, to a kind of location-based mobile community network Other method.

Background technology

Relation recognition is one of key issue of community network research.In community network, people are often because of different classes of relation (household, colleague, friend etc.) links together, and analyzes the type of relation, all has very important meaning in numerous areas Justice.As in marketing domain, by analyzing household and the Peer Relationships of user, recommendation that it accurately can be marketed；In safety Field, by grasping household and the friend relation of offender, then can help relevant departments to find clue, efficiently carry out Suspect investigates.With the extensive popularization of mobile phone, the population coverage of mobile call data already close to 100%, this be based on Mobile data carries out Fiel and the identification of person-to-person social relations can provide natural platform.Meanwhile, mobile subscriber's relation Identification also carry out help be provided for the business of operator itself, as the personalized customization of business such as family's set meal, group's set meal etc..

The essence of relation recognition problem is classification.At present, most of recognition methodss are all for several classes by abstract for relation, such as " strong and weak ", Relation is not given specifically semantic (as household, colleague etc.) by the relation of the class such as " trust and suspect ", " friendly and hostile ". Also certain methods are had to carry out semantic classification to relation, for example " instruct-by directive relationship " or " instruction-guidance-assiatant's relation ", this The relation recognition model that a little methods are set up is the special purpose model of specific area it is impossible to direct set is used in " family-Peer Relationships " classification； Also there is method to be based on the specific set of data such as terrorist's network data and carry out relation recognition it is impossible to direct set is used in mobile call data On collection.

Social networks identification based on mobile community network, needs emphasis to solve several key issues as follows:

One, the extraction of spatial relation characteristics: carry out the recognition methodss of social networking relationships at present, mostly adopt network topology structure Feature is carrying out the judgement of relation, and the impact to relation to be identified lacks consideration to user's space behavior characteristicss；

Two, the identification of the n-tuple relation based on graph model: present relation recognition method typically adopts traditional recognition methodss, such as The methods such as decision tree, svm, do not take into full account the networked Characteristics of relation recognition data；

Three, the judgement of n-tuple relation: present relation recognition method typically judges to single-relation, such as friend and friend non- Relation, does not account for inferring the lifting interacting to Model Identification precision between different relations in identification process.

Content of the invention

In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a kind of location-based mobile community network customer relationship identification Method, is identified according to mobile community network data pluralistic society's relation between user.

Present invention provide the technical scheme that

A kind of location-based mobile community network customer relationship recognition methodss, build mobile community network based on user data, lead to Cross and feature extraction carried out to user behavior, set up factor graph model, then carry out the parallel deduction of model parameter study and n-tuple relation, Identification obtains customer relationship；Specifically include following steps:

1) pretreatment is carried out to user data, obtain the sampled data for relation recognition；

2) sampled data is utilized to extract user behavior feature；Described user behavior feature includes interbehavior feature and spatial behavior is special Levy；By user behavior feature extraction, obtain the interactive factor, steric factor and group's factor of customer relationship；

3) based on the user behavior feature extracted, set up factor graph model, as customer relationship identification model；

4) carry out parameter learning training for factor graph model；

5) using the factor graph model training, relation recognition is carried out by the parallel estimating method of n-tuple relation, obtain the polynary pass of user System.

In above-mentioned location-based mobile community network customer relationship recognition methodss, user data includes master data and assistance data； Described master data is that user's communication is single in detail；Assistance data includes base station information table, family information table and group's information table；Step 1) described pretreatment is specifically included and based on family information table, data is sampled and solves noise jamming by deleting user.

Interbehavior feature includes interaction strength characteristic and interaction stability features；Spatial behavior feature include space co-occurrence feature and Geographical semantics feature.

For above-mentioned location-based mobile community network customer relationship recognition methodss, wherein, the interaction factor is represented by formula 6:

$f(r_{i,j},x_{i,j})=\frac{1}{w_e}\exp \{{\mathrm{\α}}_{e_{i,j}}\·x_{i,j}\}$ (formula 6)

In formula 6, interaction factor f (r_{I, j}, x_{I, j}) description two users between relation r_{I, j}With exchange attribute of a relation x_{I, j}Between relation； w_eIt is for standardized parameter；For every a pair related user i and j, parameterIt is one | x_{I, j}| the ginseng of dimension length Number, and the kth dimension description x of parameter_{I, jk}Contribution for relation between two user i and j；

Steric factor is represented by formula 7:

$g(r_{i,j},s_{i,j})=\frac{1}{w_{es}}\exp \{{\mathrm{\β}}_{e_{i,j}}\·s_{i,j}\}$ (formula 7)

In formula 7, steric factor g (r_{I, j}, s_{I, j}) description two users between relation r_{I, j}And between physical space attribute s_{I, j}Pass System；For each pair related user i and j, parameterKth dimension describe two user's co-occurrences place semanteme for The contribution of relation between user；

Corporations' factor is represented by formula 8:

$h(r_{i,j},c_{i,j,k})=\left\{\begin{array}{c}\frac{1}{w_{c1}}\exp \{{\mathrm{\γ}}_1\·h_1^{\′}(r_{i,k},r_{j,k})\}\\ \frac{1}{w_{c2}}\exp \{{\mathrm{\γ}}_2\·h_2^{\′}(r_{i,k},r_{j,k})\}\\ \frac{1}{w_{c3}}\exp \{{\mathrm{\γ}}_3\·h_3^{\′}(r_{i,k},r_{j,k})\}\end{array}\right.$ (formula 8)

In formula 8, social factor h (r_{I, j}, c_{I, j, k}) description two users between relation r_{I, j}The group being constituted with them and other users Between impact；w_cpFor normalizing parameter；Function h '₁(r_{I, k}, r_{J, k}) it is vector function, with vector function, collection is described in group In group, the relation on other both sides is for the contribution of user i and j Relationship Prediction.

In above-mentioned location-based mobile community network customer relationship recognition methodss, step 3) set up factor graph model include as follows Step:

First, go out mobile social networking topological structure using undirected graph Structure Representation, the nodal community of network is expressed user's row For characteristic vector, in the attribute of network edge expression user between interbehavior characteristic vector；

Then, the conditional probability distribution that an overall situation function describes relation between user is defined based on factor graph model modelling approach Expression, by overall situation function be decomposed into the interactive factor, steric factor and group the factor, be respectively intended to portray extracted mobile subscriber Behavior characteristicss；

Finally, using the method for Maximum-likelihood estimation, try to achieve so that model reaches the parameter of maximum likelihood value, obtain for relation The factor graph model of identification.

Step 3) described factor graph model is an overall probability-distribution function, is described as formula 5:

$p\left(r\right|g,x)={\π}_{e_{i,j}\&element;e}f(r_{i,j},x_{i,j})\×{\π}_{e_{i,j}\&element;e}g(r_{i,j},s_{i,j})\×{\π}_{c_{i,j,k}\&element;g}h(r_{i,j},c_{i,j,k})$ (formula 5)

In formula 5, r represents the relationship type (including family relation, Peer Relationships, friendss) between user；G represents net Network structure chart；X represents eigenmatrix, and every a line of x eigenmatrix represents the feature of a user；r_{I, j}Represent user i, j it Between relation；x_{I, j}Represent user i, the exchange ratio characteristics between j；s_{I, j}Represent user i, the steric factor between j；c_{I, j, k}Generation Table user i, the social factor between j；e_{I, j}Represent in figure i, the connection side of j；c_{I, j, k}Represent that user i, j are constituted with other users Group's factor；

Objective function formula 9 is as the maximum likelihood value of described factor graph model:

$o(\mathrm{\α},\mathrm{\β},\mathrm{\γ})={\mathrm{\σ}}_{e_{i,j}\&element;e}{\mathrm{\α}}_{e_{i,j}}\·x_{i,j}+{\mathrm{\σ}}_{e_{i,j}\&element;e}{\mathrm{\β}}_{e_{i,j}}\·s_{i,j}+{\mathrm{\σ}}_{c_{i,j,k\&element;g}}{\mathrm{\σ}}_{q=1}^3{\mathrm{\γ}}_q{h}_q^{\′}(\·)-\log w$ (formula 9)

In formula 9, o (α, beta, gamma) is the logarithmic function of p (r | g, x)；e_{I, j}Represent user i, the side between j, if user is i, have between j Then it is assumed that there is side between this two users in the interactive actions such as call；E represents the set on all of side in data set；γ_qFor needing the parameter of study, substantially express the weight of different characteristic；x_{I, j}、s_{I, j}、h′_q() represent three kinds because Son, is the exchange factor, steric factor and corporations' factor respectively；W=w_ew_esw_cpIt is global criteria parameter.

Step 4) the described gradient descent method carrying out parameter learning training for factor graph model, specifically adopting classics, described It is formula 10 that each iteration of gradient descent method needs the operation carrying out:

${\mathrm{\θ}}_{new}={\mathrm{\θ}}_{old}+\mathrm{\η}.\frac{\∂o\left(\mathrm{\θ}\right)}{\∂\mathrm{\θ}}$ (formula 10)

In formula 10, θ_newRepresent the new θ value that iteration obtains each time；θ_oldRepresent the θ value each time before iteration, be initially with The θ of machine assignment；η represents the rate value that gradient descent method updates, and η is bigger, updates faster, but fluctuation is also bigger；θ={ α, beta, gamma }, Wherein α, beta, gamma is respectively the interaction factor, steric factor and group's factor；O (θ) is maximum likelihood object function；Represent gradient Gradient selection amount in descent method.

Consider family, colleague and friend three class relation during above-mentioned model parameter study simultaneously.

Step 5) the parallel estimating method of described n-tuple relation specifically includes the method for parameter estimation declining with gradient and based on polynary The method that the probit of relation carries out relation deduction.

Compared with prior art, the invention has the beneficial effects as follows:

The present invention provides a kind of location-based mobile community network customer relationship recognition methodss, according to mobile community network data pair Between user, pluralistic society's relation is identified；Technical solution of the present invention considers the impact to relation for the spatial information, utilizes factor graph Model simultaneously carries out n-tuple relation parallelism recognition, and its advantage is embodied in following several respects:

First, the user of different relations possesses different spatial interaction features；Existing for customer relationship know the most base of method for distinguishing Carry out in the website that facebook etc. possesses positional information；And the present invention is specifically based on " mobile call data " and considers space The impact to relation for the position interaction feature；

Second, existing relation recognition method carries out relation recognition using traditional classifier mostly, does not account for the friendship between user Mutually network structure；And the impact of user's surrounding enviroment is taken into account by the present invention using factor graph model；

3rd, existing relation recognition method is mostly based on unit relation serial identification, is carried out by identifying twice.Such as Carry out the identification of " family relation " and " friendss " first, carry out again " Peer Relationships " and " friendss " afterwards Identification；And the present invention carries out polynary concurrency relation identification, an identification process obtains final result；Compared to existing method, The present invention considers the interaction between family relation and Peer Relationships, can improve the accuracy rate of relation recognition.

Brief description

Fig. 1 is the FB(flow block) of the location-based mobile community network customer relationship recognition methodss that the present invention provides；

Wherein, solid border represents each step；Dotted border represents user behavior characteristics/properties；Feature extraction is substantially Extract three factors；Three factor solid arrows are connected to next step " foundation of factor graph model "；Step " user behavior feature Extract " connected to next step " foundation of factor graph model " with dotted arrow.

Fig. 2 is the FB(flow block) of user behavior characteristic extraction step in the inventive method.

Fig. 3 is the schematic diagram that the factor graph model for customer relationship identification that the present invention sets up includes three factors.

Specific embodiment

Below in conjunction with the accompanying drawings, the present invention is further described by embodiment, but limit the scope of the present invention never in any form.

The present invention provides a kind of location-based mobile community network customer relationship recognition methodss, is built based on call detailed list data and moves Dynamic community network, extracts the interbehavior feature of mobile subscriber and spatial behavior feature, set up comprise the three class factors (the interaction factor, Steric factor and group's factor) graph model, and then realize the parallel deduction of family, colleague and friendss；Fig. 1 is the present invention Overall flow, method and step includes mobile data pretreatment, user behavior feature extraction, opening relationships identification model and relation Study, then infer and be identified result.

The present embodiment selectes the user of training set and test set, is respectively as follows: training set and includes 10 users, be respectively user a, User b ... user j；Test set includes 8 users, is user a, user b... user h respectively.The present embodiment is carrying out feature Carry out Relationship Prediction identification, the data of use includes master data and assistance data on the basis of extraction and model training, wherein, Master data is that user's communication is single in detail, such as user a and user b made a call several times.For conversing each time, call is in detail Dan Zhonghui record corresponding " calling telephone number ", " called phone number ", " call time of origin ", " call duration ", The information such as " captaincy place base station ", " called place base station ".Remove beyond essential information, also include some other auxiliary Information such as base station information table, family information table and group's information table.Wherein, have recorded " number of base stations ", " base in base station information table The longitude and latitude stood ", " place of inside of base station is semantic " (i.e. how many school, hospital, commercial street etc. in this base station)；Family Have recorded the information such as " family id ", " user id " in the information table of front yard, can confirm whether two users belong to one by this table Family；Have recorded the information such as " company id ", " user id " in group's information table, whether two users can be confirmed by this table Belong to a company.

The present embodiment mainly carries out data processing and feature extraction using sql language, obtains the value of multiple factors.Based on above-mentioned Data, the present embodiment realizes location-based mobile community network customer relationship recognition methodss, and concrete pressing operates enforcement as follows:

1) mobile data pretreatment；

Because population sample data is excessive, have such problems as to be inconvenient to be tested, so needing to carry out sampling of data.Data is pre- Process the integrity issue mainly solving sampling of data, and solve noise jamming.

11) based on family relation table, data is sampled, solves the integrity issue of sampling of data；

Discounting for data integrity, if in overall sample, completely random extracts, we can obtain not complete net User's exchange figure of network structure is it is impossible to be analyzed processing.So needing by data preprocessing method, obtaining and knowing for relation Other sampled data.

The present embodiment is sampled to data based on family relation table, to solve the integrity issue of sampling of data；Concrete grammar bag Include following steps:

A) from family relation table, completely random extracts a number of family id；

B) according to family id, the whole users belonging to this family are extracted；These users constitute sampling set, the most to be identified User data.

12) solve noise jamming by deleting user；

In the present embodiment, the user types of deletion include: in one month, total call quantity is less than 10 user；In one month, The user more than 500 for the number of contacts exposing.

2) user behavior feature extraction；

Mobile subscriber's behavior characteristicss according to the present invention extract and are related to two classes: the first is interbehavior feature, and it lays particular emphasis on description The exchange feature of relation between two users, the feature such as such as air time, call intensity；Second is spatial behavior feature, Lay particular emphasis on and describe co-occurrence behavior characteristicss spatially between two users (the local distribution of such as their Chang Tongxian), with And geographical semantics feature (the place co-occurrence such as such as household often stays at home, market, and work together then in unit office space position co-occurrence).

User behavior feature extraction is by using the initial data extraction feature such as call detailed list.Input call detailed list, family relation table, Group's relation table and base station information table；By user behavior feature extracting method, obtain customer relationship three factors (interaction because Son, steric factor and group's factor).

21) interbehavior feature extraction: two category features of relationship strength and stability are portrayed in main inclusion as follows:

(1) interaction strength characteristic: in sociology relationship description is bonding strength, bonding strength tie strength is defined by formula 1, It is used for representing call exchange feature, there is in identification process the effect that auxiliary improves accuracy rate；We are special for different calls Levy (i.e. parameter k1～k4) and give different weights:

(formula 1)

In formula 1, parameter k1～k4 represents " two people's exchange number of times ", " two people's busy exchange number of times ", " two people's idles respectively Exchange number of times " and " two people exchange number of times at weekend "；The value of parameter k1～k4 represents the weights of different conversational nature, can be utilized Regression fit is calculated.

Between the user of different relations, have different conversational nature；Such as domestic consumer is more in idle and Sunday call； And the user of Peer Relationships is more in busy exchange number of times.We train k={ k1, k2, k3, k4 } value with svm training aidss, Purpose is the discrimination expanding further between different relations.

Concrete training process comprises the steps:

A) first according to the detailed forms data of user's communication, obtain ten users of a～j data between any two in training set, comprising: " two People's exchange number of times ", " two people's busy exchange number of times ", " two people's idles exchange number of times " and " two people exchange number of times at weekend ".

Illustrate, according to call detailed list, we can obtain the call time of origin of user, according to the different air times, Sum up computing, you can draw above-mentioned data, as conversational nature.

B) according to family information table and group's information table, obtain the relation between user；

C) will " two people's exchange number of times, two people's busy exchange number of times, two people's idle exchange number of times, two people's weekends exchange number of times " four , as x value, the relation between them, as y value, carries out svm training for attribute, finally train corresponding k=k1, k2, K3, k4 } value, this value can be concentrated use in test.

Specifically, in the present embodiment, because " raw information table " is divided into call list and note table, each step feature extraction It is required for being extracted twice operation, obtain " note total quantity " and " total call amount " respectively.Four feature integrations are struck a bargain The mutually factor.Start us and do not know the weight of different characteristic, so classification power can be tried to achieve using svm grader in training set Weight.

Input: x (four basic features x1, x2, x3, x4), y (attribute of a relation between user)

Process: using svm grader, be iterated classifying

Output: finally in the high result of classification accuracy, obtain object vector k (k1, k2, k3, k4)

Categorizing process particularly as follows:

K (k1, k2, k3, k4) value and the corresponding feature obtaining in feature obtains that input obtains in learning process；

Processing procedure: using formula tie strength=k₁*x₁+k₂*x₂+k₃*x₃+k₄*x₄, calculate the interactive factor Numerical value.Wherein x₁The total note bar number of the total talk times of=two people's talk times+two people's note bar number；x₂=busy talk times two people is led to Words total degree+busy note bar number two people's note bar number；x₃=idle talk times two people call total degree+idle note bar number two People's note total number；x₄=Sunday call number of times two people call total degree+weekend note bar number two people's note total number；

It is output as interaction intensity level.

(2) interaction stability features: the call behavior of different relational users has diversity in time, and work pass on the whole The user of system between call concentrate on working hour, and family's air time randomness is stronger.In order to quantify this feature, carry Go out the concept of entropy of conversing, for weighing the stability of different relational users calls.The computing formula of call entropy entropy is as follows:

$entropy=-{\mathrm{\σ}}_{i=1}^{t}p\left(x_i\right)\·{\log }_{2}p\left(x_i\right)$ (formula 2)

Wherein, p (x_i) it is user to the talk probability within i-th hour period, i=1,2 ..., t；Typically the value of t is 24, corresponding to 24 hour periods in one day.

22) spatial behavior feature extraction: main include portraying two category features of space Concurrent Pattern and geographical semantics:

(1) geographical semantics feature:

In known base station information, each base station possesses respective semanteme i.e. in the coverage of this base station, has Several hospitals, school, recreational facilities etc..Test is carried out i.e. if the user while this base station goes out based on simplest summation Existing co-occurrence, then all semantemes in this base station range are added up respectively.But, so finally give a milli The regional distribution of facilities result of no discrimination.

For avoiding the appearance of this phenomenon, we have introduced tf-idf method.The method as a kind of statistical method, for assessing Words is for the significance level of a copy of it file in a file set or a corpus.The importance of words is with it in literary composition The number of times occurring in part is directly proportional increase, the decline but frequency that can occur in corpus with it is inversely proportional to simultaneously.In this patent In mainly for assessment of the semantic importance in certain place.Such as " school " this key word occurs in that 100 in whole city Secondary, the idf that we just define " school " is 1/100；User x, y meet in base station l, and in l, " school " this key word goes out 5 times are showed, then we define during this meets, and the tf of " school " is 5；May finally try to achieve during this meets, " school's language Weights shared by justice ": tf*idf=5*1/100=0.05.Finally we represent this base station with the semanteme of this base station maximum weight.

In the present embodiment, according to call detailed list, the location information of user can be obtained, for example, user a in the xxx time in yyy Base station occurs once to converse.We can simplify and are interpreted as: user a occurs in yyy place in the xxx time.

(2) space Concurrent Pattern feature:

If user a, b are in hours, all occur in same base station c, then record co-occurrence once.According to the time, Co-occurrence feature is divided three classes, is night co-occurrence respectively, the co-occurrence on daytime on working day, weekend co-occurrence etc..Then fixed Adopted co-occurrence formula:Wherein a ∩ b represents the space co-occurrence number of times of user a and user b；A ∪ b represents user a space Position exposes number of times and b locus expose the summation of number of times.Illustrate: extract night co-occurrence, now formulaTime be defined to night, have a ∩ b to represent at night, the space co-occurrence number of times of user a and user b；A+b represents At night, user a locus expose number of times and b locus expose the summation of number of times.

According to the location information of user, we can obtain the co-occurrence information of user.Such as user a is in 10 points of appearance today In y base station, user b also appears in y base station at 10 points in today, then it is considered that two user's once co-occurrences.

Three co-occurrence information are mainly used in model:

First be user's co-occurrence frequency, such as " total degree of user a and user's b co-occurrence " divided by " they respectively with Existing number of times sum "；

Second is that user's co-occurrence place is semantic, and each base station has corresponding place semantic, such as base station y coverage Under, there are 10 hospitals, 2 schools etc.；Place with tf-idf method calculation base station is semantic；

3rd be user's co-occurrence the regularity of distribution, Distribution Entropy is calculated according to the space time information of user's co-occurrence.

3) set up factor graph model, as relation recognition model；

With g=(v, c, r, s), mobile social networking to be described, wherein v is the set of | v | in network=n user,Group between user in statement network, r is used for representing the relation between two users, point friend, Tong Shihe Family three class.S describes relation spatially between two users.It is the base attribute relation between two users of description with x Matrix；Each of x x_{I, j}Represent is one for describing attribute of a relation between user i and user j | x_{I, j}| dimensional feature vector.

A given social networkies g=(v, c, r, s) and relationship characteristic attribute matrix x, our target is study such as minor function:

F:g=(v, c, r, s), x → (r) (formula 3)

It is used for carrying out differentiating the social networks between user.

For making full use of the information included in mobile community network data set, designed model not only customer relationship to be considered The attribute itself having, it is also contemplated that user's behavior characteristicss spatially, and Social behaviors feature.Incorporate for comprehensive State information, the present invention builds factor graph model to carry out n-tuple relation identification.Factor graph model modeling comprises the steps:

First, go out mobile social networking topological structure using undirected graph Structure Representation, the nodal community of network is expressed user certainly The characteristic vector of body behavior, expresses the characteristic vector of interbehavior between user in the attribute of network edge；

Then, theoretical based on factor graph model modeling, define an overall situation function and divide come the conditional probability to describe relation between user The expression of cloth, and then overall situation function is decomposed into three factors, it is respectively intended to portray extracted mobile subscriber's behavior characteristicss；

Finally, using the method for Maximum-likelihood estimation, try to achieve so that model reaches the parameter of maximum likelihood value, that is, complete relation The structure of identification model.

In factor graph model, define an overall situation function and make the conditional probability describing customer relationship reach maximum, by overall letter Number is decomposed into local functions product, the present invention by overall situation function be decomposed into three factors (the interaction factor, steric factor, social because Son):

$\begin{array}{c}p\left(r\right|g,x)=\frac{p(x,g|r)p\left(r\right)}{p(x,g)}\&proportional;p\left(r\right|g)p\left(x\right|r)\\ \&proportional;{\π}_{e_{i,j\&element;e}}p\left(x_{i,j}\right|r_{i,j}){\π}_{c_{i,j,k\&element;e}}p\left(s_{i,j}\right|r_{i,j}){\π}_{c_{i,j,k\&element;e}}p\left(c_{i,j,k}\right|r_{i,j})\end{array}$ (formula 4)

Specifically, three factors are respectively as follows: 1) interact the exchange feature that the factor describes relation between two users, when such as conversing Between, the call feature such as intensity；2) steric factor describes the contact between two users in physical space, and such as they are same Existing local distribution；3) relation that corporations' factor describes between user is affected by affiliated corporations.

Therefore, entirely total overall probability distribution can be described by formula below:

$p\left(r\right|g,x)={\π}_{e_{i,j}\&element;e}f(r_{i,j},x_{i,j})\×{\π}_{e_{i,j}\&element;e}g(r_{i,j},s_{i,j})\×{\π}_{c_{i,j,k}\&element;g}h(r_{i,j},c_{i,j,k})$ (formula 5)

Wherein, r represents the relationship type (including family relation, Peer Relationships, friendss) between user, and g represents network Structure chart；X represents eigenmatrix, such as 10 users, then x eigenmatrix just has 10 row, and every a line represents a user's Feature；r_{I, j}Represent user i, the relation between j；x_{I, j}Represent user i, the exchange ratio characteristics between j；s_{I, j}Represent user i, j Between steric factor feature；c_{I, j, k}Represent user i, the social factor between j；e_{I, j}Represent in figure i, the connection side of j；c_{I, j, k}Table Show user i, the group that j is constituted with other users.

The building process of three saturations:

The interaction factor: usage factor f (r_{I, j}, x_{I, j}) describing relation r between two users_{I, j}With exchange attribute of a relation x_{I, j}Between Relation；The interactive factor (function) to be described with an index linear function (formula 6):

$f(r_{i,j},x_{i,j})=\frac{1}{w_e}\exp \{{\mathrm{\α}}_{e_{i,j}}\·x_{i,j}\}$ (formula 6)

It is the parameter needing in model to be learnt；w_eFor standardized parameter, for every a pair related user i And j, parameterIt is one | x_{I, j}| the parameter of dimension length, and the kth dimension of parameter describes x_{I, jk}For relation between two users Contribution.Such as x_{I, j}Kth dimension represents is call intensity between the two, then parameterKth dimension then describe Call intensity between the two is for the contribution of the relation between them.The Main Function of the interaction factor is used to describe two users Between call attribute for whole customer relationship impact.

Steric factor: with factor g (r_{I, j}, s_{I, j}) describing relation r between two users_{I, j}And between physical space attribute s_{I, j}'s Relation, the frequency using co-occurrence place (both simultaneously appear in certain place) is portrayed, and describes this with a linearized index function The factor (function):

$g(r_{i,j},s_{i,j})=\frac{1}{w_{es}}\exp \{{\mathrm{\β}}_{e_{i,j}}\·s_{i,j}\}$ (formula 7)

Wherein,It is the parameter needing in model to be learnt, w_esIt is used to the parameter being standardized.Relevant for each pair User i and j of system, parameterKth dimension describe two user's co-occurrences place semanteme for their relations contribution.Such as, s_{I, j}Kth dimension the frequency that two users occur in company describe, then if this frequency reach high if, then they two It is likely to be Peer Relationships between person.Also such as two users are simultaneously higher in the frequency of cell appearance, then between them very It is likely to be family relation.

Corporations' factor: social factor h (r_{I, j}, c_{I, j, k}) description two users between relation r_{I, j}Constituted with them and other users Impact between group.More specific this factor that described with a function:

$h(r_{i,j},c_{i,j,k})=\left\{\begin{array}{c}\frac{1}{w_{c1}}\exp \{{\mathrm{\γ}}_1\·h_1^{\′}(r_{i,k},r_{j,k})\}\\ \frac{1}{w_{c2}}\exp \{{\mathrm{\γ}}_2\·h_2^{\′}(r_{i,k},r_{j,k})\}\\ \frac{1}{w_{c3}}\exp \{{\mathrm{\γ}}_3\·h_3^{\′}(r_{i,k},r_{j,k})\}\end{array}\right.$ (formula 8)

Wherein, function h '₁(r_{I, k}, r_{J, k}) it is vector function, w_cpFor normalizing parameter.With vector function, group is described in group In in addition both sides relation for user i and j Relationship Prediction contribution.Such as, if other twice is family relation, then very Being likely to require the side being predicted is also family relation.Three kinds of different letters can be constructed according to the type difference on two other side Number.

In this enforcement, the acquisition of corporations' factor specifically: input original talk is single in detail, and processing procedure is with sql language, takes out Take all of ternary group；Such as user a and user b, c have call, also there is call between user b, c simultaneously, So a, b, c constitute ternary group；It is output as all ternary groups information.

The comprehensive three above factor, objective function as the maximum likelihood value (log-likelihood) of proposed model,

$o(\mathrm{\α},\mathrm{\β},\mathrm{\γ})={\mathrm{\σ}}_{e_{i,j}\&element;e}{\mathrm{\α}}_{e_{i,j}}\·x_{i,j}+{\mathrm{\σ}}_{e_{i,j}\&element;e}{\mathrm{\β}}_{e_{i,j}}\·s_{i,j}+{\mathrm{\σ}}_{c_{i,j,k\&element;g}}{\mathrm{\σ}}_{q=1}^3{\mathrm{\γ}}_q{h}_q^{\′}(\·)-\log w$ (formula 9)

Wherein, o (α, beta, gamma) is the logarithmic function of p (r | g, x)；e_{I, j}Represent user i, the side between j, if user is i, have logical between j Then it is assumed that there is side between this two users in the interactive actions such as words；E represents the set on all of side in data set；γ_qFor needing the parameter of study, substantially express the weight of different characteristic；x_{I, j}、s_{I, j}、h′_q() represent three kinds because Son, is the exchange factor, steric factor and corporations' factor respectively；W=w_ew_esw_cpIt is global criteria parameter.

4) pass through relational learning and deduction, be identified result.

At present, often adopted Relationship Prediction method is predicted respectively to family, colleague and friendss respectively.But, During deduction, the mutually collaborative lifting contributing to accuracy of identification of different relations, the therefore present invention adopts polynary social networks Parallel estimating method, identifies that the result obtaining will be more reasonable.

The present invention to realize the parallel deduction of n-tuple relation mobile community network from two angles: one, the ginseng declining with gradient Three kinds of relation classifications in factor graph model are carried out parametric inference by number estimation method simultaneously；Two, the probit based on n-tuple relation Carry out relation deduction: for any a line in network, inferred with factor graph model on the basis of parameter learning when Wait, three kinds of relations can obtain a probability, then choose the relation classification corresponding to maximum probability value as the knot of relation recognition Really.

N-tuple relation study and parallel deduction, comprising:

41) model parameter study

By contextual definition attribute of a relation r ∈ { 0,1,2 } of any two user in graph model, wherein 0 represents friendss to this patent, 1 represents Peer Relationships, and 2 represent family relation.Because family relation is a kind of relationship type more higher than Peer Relationships relation intensity, So in the present invention, learn the stage in model parameter, will be that the relationship type again worked together of family is identified only as family pass System.

The input of model learning (training) is the physical relationship attribute r between the data and user that feature extraction obtains；This enforcement Example carries out model training in training set.The input training set data of model learning, the tool between all factors of extraction and user Body relation.For training set a～j, and the attribute of a relation r between them known to us (such as ab is one family, then r_ab=2； Work together as 1；Regular friend is 0), (need input AC feature x after the data is entered_{I, j}, space characteristics s_{I, j}, ternary interaction Feature c_{I, j, k}) it would be desirable to obtain θ={ α, beta, gamma } value (for determining the weight of the different factors) in p (r | g, x) to determine Model.

The target of model learning is to find suitable value θ={ α, beta, gamma } to make maximum likelihood target in given training set Function o (θ) reaches maximum.I.e. θ^*=argmaxo (θ).By the θ of model learning output category model={ α, beta, gamma } value.

To solve the problems, such as to train using the method that classical gradient declines.The operation that as each iteration needs are carried out below:

${\mathrm{\θ}}_{new}={\mathrm{\θ}}_{old}+\mathrm{\η}\·\frac{\∂o\left(\mathrm{\θ}\right)}{\∂\mathrm{\θ}}$ (formula 10)

In formula 10, θ_newRepresent the new θ value that iteration obtains each time；θ_oldRepresent the θ value each time before iteration, be initially with The θ of machine assignment；η represents the rate value that gradient descent method updates, and η is bigger, updates faster, but fluctuation is also bigger；Table Show the gradient selection amount in gradient descent method.

θ={ α, beta, gamma }, during beginning, random assignment θ, afterwards, it is iterated computing with gradient descent method.

42) identification of n-tuple relation:

After completing parameter value θ estimation, obtain corresponding parameter value it is possible to the relation classification for side unknown in network is carried out Identification.

During n-tuple relation study based on graph model and deduction: 1) consider family, Tong Shihe during model parameter study simultaneously Friend three class relation；2) during relation deduction, infer three class relations (family, colleague and friend) simultaneously, employing Method is the probit of the three class relations calculating according to each edge, takes its relation corresponding to maximum probability value.For in network Any a line when inferred with factor graph model, three kinds of relations can obtain a probability, then choose probability The corresponding relation classification of big value is as the result of relation recognition.

For test set user a～h, our final goals are to obtain their attribute of a relation r.Similar to model training, by surveying The call detailed list of examination collection data obtains the ternary social networks of user, inputs three factors, is simultaneously entered and tries to achieve in training process θ={ α, beta, gamma } value, finds the r making p (r | g, x) value maximum in the function represented by formula 5, r is exactly the relation between user.

It should be noted that the purpose publicizing and implementing example is that help further understands the present invention, but those skilled in the art It is understood that various substitutions and modifications are all possible without departing from the present invention and spirit and scope of the appended claims. Therefore, the present invention should not be limited to embodiment disclosure of that, and the scope of protection of present invention is defined with claims Scope is defined.

Claims (9)

1. a kind of location-based mobile community network customer relationship recognition methodss, build mobile community network based on user data, By feature extraction is carried out to user behavior, set up factor graph model, then carry out pushing away parallel of model parameter study and n-tuple relation Disconnected, identification obtains customer relationship；Specifically include following steps:

1) pretreatment is carried out to user data, obtain the sampled data for relation recognition；

2) sampled data is utilized to extract user behavior feature；Described user behavior feature includes interbehavior feature and spatial behavior is special Levy；By user behavior feature extraction, obtain the interactive factor, steric factor and group's factor of customer relationship；

3) based on the user behavior feature extracted, set up factor graph model, as customer relationship identification model；

4) carry out parameter learning training for factor graph model；

5) using the factor graph model training, relation recognition is carried out by the parallel estimating method of n-tuple relation, obtain the polynary pass of user System.

2. location-based mobile community network customer relationship recognition methodss as claimed in claim 1, is characterized in that, described user Data includes master data and assistance data；Described master data is that user's communication is single in detail；Described assistance data includes base station information Table, family information table and group's information table；Step 1) described pretreatment specifically included and based on family information table, data taken out Sample and by delete user solve noise jamming.

3. location-based mobile community network customer relationship recognition methodss as claimed in claim 1, is characterized in that, described interaction Behavior characteristicss include interaction strength characteristic and interaction stability features；Described spatial behavior feature includes space co-occurrence feature and geography Semantic feature.

4. location-based mobile community network customer relationship recognition methodss as claimed in claim 1, is characterized in that,

The described interactive factor is represented by formula 6:

$f(r_{i,j},x_{i,j})=\frac{1}{w_e}\exp \{{\mathrm{\α}}_{e_{i,j}}\·x_{i,j}\}$ (formula 6)

In formula 6, interaction factor f (r_{I, j}, x_{I, j}) description two users between relation r_{I, j}With exchange attribute of a relation x_{I, j}Between relation； w_eIt is for standardized parameter；For every a pair related user i and j, parameterIt is one | x_{I, j}| the ginseng of dimension length Number, and the kth dimension description x of parameter_{I, j k}Contribution for relation between two user i and j；

Described steric factor is represented by formula 7:

$g(r_{i,j},s_{i,j})=\frac{1}{w_{es}}\exp \{{\mathrm{\β}}_{e_{i,j}}\·s_{i,j}\}$ (formula 7)

In formula 7, steric factor g (r_{I, j}, s_{I, j}) description two users between relation r_{I, j}And between physical space attribute s_{I, j}Pass System；For each pair related user i and j, parameterKth dimension describe two user's co-occurrences place semanteme for The contribution of relation between user；

The described corporations factor is represented by formula 8:

$h(r_{i,j},c_{i,j,k})=\left\{\begin{array}{c}\frac{1}{w_{c1}}\exp \{{\mathrm{\γ}}_1\·h_1^{\′}(r_{i,k},r_{j,k})\}\\ \frac{1}{w_{c2}}\exp \{{\mathrm{\γ}}_2\·h_2^{\′}(r_{i,k},r_{j,k})\}\\ \frac{1}{w_{c3}}\exp \{{\mathrm{\γ}}_3\·h_3^{\′}(r_{i,k},r_{j,k})\}\end{array}\right.$ (formula 8)

In formula 8, social factor h (r_{I, j}, c_{I, j, k}) description two users between relation r_{I, j}The group being constituted with them and other users Between impact；w_cpFor normalizing parameter；Function h '₁(r_{I, k}, r_{J, k}) it is vector function, described with vector function in group In group, the relation on other both sides is for the contribution of user i and j Relationship Prediction.

5. location-based mobile community network customer relationship recognition methodss as claimed in claim 1, is characterized in that, step 3) Described factor graph model of setting up comprises the steps:

First, go out mobile social networking topological structure using undirected graph Structure Representation, the nodal community of network is expressed user's row For characteristic vector, in the attribute of network edge expression user between interbehavior characteristic vector；

Then, the conditional probability distribution that an overall situation function describes relation between user is defined based on factor graph model modelling approach Expression, by overall situation function be decomposed into the interactive factor, steric factor and group the factor, be respectively intended to portray extracted mobile subscriber Behavior characteristicss；

Finally, using the method for Maximum-likelihood estimation, try to achieve so that model reaches the parameter of maximum likelihood value, obtain for relation The factor graph model of identification.

6. location-based mobile community network customer relationship recognition methodss as claimed in claim 1, is characterized in that, step 3) Described factor graph model is an overall probability-distribution function, is described as formula 5:

$p\left(r\right|g,x)={\π}_{e_{i,j}\&element;e}f(r_{i,j},x_{i,j})\×{\π}_{e_{i,j}\&element;e}g(r_{i,j},s_{i,j})\×{\π}_{c_{i,j,k}\&element;g}h(r_{i,j},c_{i,j,k})$ (formula 5)

In formula 5, r represents the relationship type (including family relation, Peer Relationships, friendss) between user；G table Show network structure；X represents eigenmatrix, and every a line of x eigenmatrix represents the feature of a user；r_{I, j}Represent and use Family i, the relation between j；x_{I, j}Represent user i, the exchange ratio characteristics between j；s_{I, j}Represent user i, the sky between j Between the factor；c_{I, j, k}Represent user i, the social factor between j；e_{I, j}Represent in figure i, the connection side of j；c_{I, j, k}Represent user Group's factor that i, j are constituted with other users；

Objective function formula 9 is as the maximum likelihood value of described factor graph model:

$o(\mathrm{\α},\mathrm{\β},\mathrm{\γ})={\mathrm{\σ}}_{e_{i,j}\&element;e}{\mathrm{\α}}_{e_{i,j}}\·x_{i,j}+{\mathrm{\σ}}_{e_{i,j}\&element;e}{\mathrm{\β}}_{e_{i,j}}\·s_{i,j}+{\mathrm{\σ}}_{c_{i,j,k\&element;g}}{\mathrm{\σ}}_{q=1}^3{\mathrm{\γ}}_q{h}_q^{\′}(\·)-\log w$ (formula 9)

In formula 9, o (α, beta, gamma) is the logarithmic function of p (r | g, x)；e_{I, j}Represent user i, the side between j, if user i, j it Between have the interactive actions such as call then it is assumed that there is side between this two users；E represents the set on all of side in data set；γ_qFor needing the parameter of study, substantially express the weight of different characteristic；x_{I, j}、s_{I, j}、h′_q() represents three kinds The factor, is the exchange factor, steric factor and corporations' factor respectively；W=w_ew_esw_cpIt is global criteria parameter.

7. location-based mobile community network customer relationship recognition methodss as claimed in claim 1, is characterized in that, step 4) the described gradient descent method carrying out parameter learning training for factor graph model, specifically adopting classics, under described gradient It is formula 10 that each iteration of fall method needs the operation carrying out:

${\mathrm{\θ}}_{new}={\mathrm{\θ}}_{old}+\mathrm{\η}\·\frac{\∂o\left(\mathrm{\θ}\right)}{\∂\mathrm{\θ}}$ (formula 10)

In formula 10, θ_newRepresent the new θ value that iteration obtains each time；θ_oldRepresent the θ value each time before iteration, initially It is the θ of random assignment；η represents the rate value that gradient descent method updates, and η is bigger, updates faster, but fluctuation is also bigger； θ={ α, beta, gamma }, wherein α, beta, gamma is respectively the interaction factor, steric factor and group's factor；O (θ) is maximum likelihood target letter Number；Represent the gradient selection amount in gradient descent method.

8. location-based mobile community network customer relationship recognition methodss as claimed in claim 7, is characterized in that, described Consider family, colleague and friend three class relation during model parameter study simultaneously.

9. location-based mobile community network customer relationship recognition methodss as claimed in claim 1, is characterized in that, step 5) the parallel estimating method of described n-tuple relation specifically includes the method for parameter estimation declining with gradient and based on n-tuple relation The method that probit carries out relation deduction.