CN105791085A - Friend recommending method in position social network based on positions and time - Google Patents

Abstract

The invention discloses a friend recommending method in a position social network based on positions and time. The method mainly is used for solving the problem of low recommending precision resulting from the face that the existing friend recommending scheme ignores time information and sign-in position semantic information in sign-in information. The method comprises following steps of 1, building a communication system framework composed of users, positioning facilities and a position server; 2, sending the longitudes and latitudes of the positions and current time information to a social network server by the users, thus generating the sign-in information; 3, preprocessing massive stored sign-in information by the social network server; 4, calculating the word frequency-inverse document frequency value of each user to each place class by the social network server; 5, calculating similarities among the users by the social network server; and recommending friends to the users with relatively high similarities. According to the method, the available data sparsity of a recommending system is mitigated; the recommending precision is improved; and the method can be applied in the wireless social network service based on the positions.

Description

Friend recommendation method based on position Yu time in the social networks of position

Technical field:

The invention belongs to radio network technique field, relate to the friend recommendation of wireless social networks, can be applicable to based on position Wireless social networking service.

Background technology:

Location-based social networks helps user to share their real time position online so that user is it appeared that feel emerging Also make friends in interesting position.Such as, user can find oneself place interested by the position that good friend shares, or Person makes friends with new good friend by searching the user sharing similar place with oneself.Additionally, popularizing along with smart mobile phone, Its built-in GPS module can detect the position of user more accurately so that user can share respectively more easily From position.Therefore, this service of registering has attracted increasing user.How to use substantial amounts of information of registering for using Family carries out friend recommendation to be needed to be paid attention to.

Commending system plays important role in social networks and e-commerce website, in e-commerce website, existing Some commending systems generally use the purchaser record of user to analyze the preference of user, as e-commerce website purchase remember Record, user's history of registering in location-based social networks also contains the substantial amounts of information about user preference.

When the information of registering using user carries out friend recommendation, owing to user the most only can sub-fraction in database Place is carried out registering behavior, and the data available therefore carrying out friend recommendation is the most sparse.In order to alleviate the problem of Sparse, Existing scheme is divided into the most several: classify according to the density in place of registering, the label being closer in some geographical position Same place of registering is regarded as to place；The data of registering using user good friend are filled into the registering in data of user.But, Close place, geographical position can represent the hobby that user is different, and equally, the hobby of good friend can not generation completely For the hobby of user self, both behaviors reduce the precision of friend recommendation the most to a certain extent.

In terms of considering the commending system of temporal information, existing system is used in e-commerce website and film recommended website more, Recommended by the preference watching record analysis user analyzing the long-term purchaser record of user and film.Compared to length The trend of phase, the behavior of registering of user can provide more information reflected user preferences the concrete time of one day.And In existing social networks commending system, only by the register time attribute in place of analysis, place of registering is classified, and The Annual distribution of behavior of not registering user is analyzed to carry out friend recommendation, and the accuracy rate therefore recommended is not ideal enough.

Summary of the invention

Present invention aims to the deficiency of above-mentioned prior art, propose a kind of base in the social networks of position In the friend recommendation method of position Yu time, to improve the accuracy rate recommended.

The present invention, by the place having similar semantic information is classified as a class, solves data available in commending system sparse Problem, is used the interest in term frequency-inverse document frequency equilibrium hot topic place and user self, is registered by analysis user simultaneously The behavior regularity of distribution in time, is adjusted the Similarity Measure of user, it is achieved friend recommendation more accurately.Its Implementation includes the following:

(1) a communication system framework being made up of user, location facilities, social network server is set up, wherein:

User, for communicating with location facilities and social network server by mobile cellular network or WiFi；

Location facilities, realizes location for collaborative user's GPS module；

Social network server, for storing the positional information of user, and carries out friend recommendation to user on this basis；

(2) longitude and latitude of position and current temporal information are sent to social network server by user, generate letter of registering Breath；

(3) the storage information of registering is pre-processed by social network server:

(3a) for the longitude and latitude in the information of registering, the position semantic database searching server obtains and registers place pair The semantic information answered, then adds up, by the semantic information classification in its place of registering, number of times of registering by the information of registering of each user；

(3b) for the temporal information in the information of registering, on the basis of step (3a), each user is added up each Behavior of registering under the classification of place distribution in time；

(4) according to user at the number of times of registering of each place classification, social network server calculates each user u for often Term frequency-inverse document frequency score TF-IDF of individual place classification c_u,c；

(4a) according to user u number of times of registering under the classification of each place, social network server calculates user u each The word frequency score value TF of information of registering under classification c of place_u,c；

(4b) according to all users number of times of registering under the classification of each place, social network server calculates each location category The inverse document frequency score value IDF of other c_c；

(4c) social network server is by user u word frequency score value TF under classification c of place_u,cInverse with place classification c Document frequency score value IDF_cIt is multiplied, obtains its term frequency-inverse document frequency score under this place classification TF-IDF_u,c=TF_u,c×IDF_c；

(5) similarity between social network server calculating user:

(5a) use user to register under the classification of each place behavior distribution in time, calculate two users in the same manner Register under some classification the relative entropy D (P Q) being distributed, and wherein P and Q is two user u_iAnd u_jIn same place classification Register number of times probability distribution in time, then use relative entropy to calculate similarity Dynamic gene

(5b) for user in each location category other term frequency-inverse document frequency score TF-IDF_u,c, use the phase of correspondence Seemingly spend Dynamic geneAdjust cosine similarity computing formula, calculate the similarity cos (u between two users_i,u_j)；

(6) social network server is to repeat the above steps (3) between user and the every other user of request recommendation extremely Step (5), obtains the similarity score of this user and other users, chooses n the highest user of score value as recommending knot Fruit is sent to the user that request is recommended, the friend recommendation quantity that the user that n recommends by request is asked.

The present invention compared with prior art has the advantage that

1) due to the fact that the semantic information employing place of registering is classified, recommend data available openness alleviating While, maintain higher recommendation precision.

2) due to the fact that employing term frequency-inverse document frequency model quantifies data of registering, and balances hot topic and registers Place and the preference of user self.

3) due to the fact that behavior distribution in time of registering user is analyzed, and sign between user by calculating The difference being distributed in time to behavior obtains calculating the Dynamic gene of user's similarity, takes full advantage of the geography of the information of registering Information and temporal information, thus ensure that higher recommendation accuracy.

Accompanying drawing explanation

Fig. 1 is the communication system frame diagram that the present invention uses；

Fig. 2 is the flowchart of the present invention；

Fig. 3 is the recommendation accuracy rate comparison diagram of suggested design used by the present invention and the suggested design not considering time factor；

Fig. 4 is the recommendation recall rate comparison diagram of suggested design used by the present invention and the suggested design not considering time factor.

Specific embodiments

The core concept of the present invention is under location-based social networks scene, the semanteme in place of being registered by analysis user Data of registering are carried out classifying to alleviate the problem that data available is the most sparse by information, and comprehensively analyze the heat in place of registering Degree, the preference of individual subscriber and the user behavior regularity of distribution in time of registering, for user is carried out friend recommendation, carries The high degree of accuracy recommended.

With reference to Fig. 2, it is as follows that the present invention realizes step:

Step 1, sets up communication system framework.

With reference to Fig. 1, the communication system that this step is set up includes: user, location facilities, social network server, wherein User all carries out double-direction radio by mobile cellular net or WiFi with location facilities and social network server and is connected；

Described user, comprises application module, DBM and three functional modules of GPS module；Application module is mainly used In generating and sending the information of registering to social network server；GPS module is mainly used in location facilities inquiring position information And the geographical location information of user is provided to application module；DBM is mainly used in storage and management user registers and becomes reconciled Friend's information；

Described location facilities, mainly comprises GPS module, and this GPS module is mainly used in the position enquiring to user and carries out Respond and return the geographical location information of user；

Described social network server, comprises application module and two functional modules of DBM；Application module is mainly used In user's information of registering is analyzed arranging and being the friend recommendation request return recommendation results of user, DBM It is mainly used in storing register data and the semantic information in place of registering of user.

Step 2, gathers information of registering.

User by location facilities obtain self place geographical location information, and by this geographical location information with at that time time Between information be sent to social network server and generate and register information.

Step 3, the information of registering in a large number of storage is pre-processed by social network server.

(3a) for the geographical location information in the information of registering, the position semantic database searching server obtains and registers The semantic information that place is corresponding, is then added up the information of registering of each user by the semantic information classification in its place of registering and signs To number of times；

(3b) for the temporal information in the information of registering, is divided into 5 intervals, respectively 2:00-7:00 the time, 7:00-11:00,11:00-17:00,17:00-22:00,22:00-2:00, counting user u is in place classification (c u) and user is at total number of times sum (c) of registering of this place classification, obtains for the number of times vis that registers on the upper each time interval of c In frequency P of registering of each time interval of place classification c, ((c, v)/sum (c) thus obtain user at certain to user for c, v)=vis Individual place classification is registered behavior distribution in time.

Step 4, the social network server each user u of calculating divides for the term frequency-inverse document frequency of each place classification c Value TF-IDF_u,c。

(4a) social network server calculates user u and registers under classification c of place the word frequency score value TF of information_u,c:

${\mathrm{TF}}_{u,c}=\frac{\left|\right\{u.v_i:v_i.D=c\left\}\right|}{|u.V|},$

Wherein, | { u.v_i:v_i.D=c} | being the number of times registered under classification c of place of user u, | u.V | is that user u signs in all places The total degree arrived；

(4b) social network server calculates the inverse document frequency score value IDF of each place classification c_c:

${\mathrm{IDF}}_c=log\frac{\left|U\right|}{\left|\right\{u_j:c\∈u_j.D\left\}\right|},$

Wherein, | U | represents the quantity of all users in server, | { u_j:c∈u_j.D} | represent the user once registered in place classification c Quantity；

(4c) social network server is by user u word frequency score value TF under classification c of place_u,cInverse with place classification c Document frequency score value IDF_cIt is multiplied, obtains its term frequency-inverse document frequency score under this place classification TF-IDF_{U, c}=TF_{U, c}×IDF_c。

Step 5, social network server calculates the similarity between user.

(5a) social network server calculates i-th user u respectively_iWith jth user u_jRelative entropy in place classification c D(P_c||Q_c) and jth user u_jWith i-th user u_iRelative entropy D (Q in place classification c_c||P_c), it calculates public affairs Formula is as follows:

Wherein, P_cFor i-th user u_iAt the probability distribution registered in time of place classification c, Q_cFor jth user u_j? The probability distribution registered in time of place classification c, i=1,2,3,4,5 respectively corresponding steps (3b) calculate users register with 5 time intervals divided during the probability distribution of time；

(5b) social network server according to the result of (5a) calculate place classification c register information similarity adjustment because of SonAs follows:

$F_{u_i,u_j,c}=\frac{2}{D\left(P_c\right|\left|Q_c\right)+D\left(Q_c\right|\left|P_c\right)+1};$

(5c) user that social network server obtains in step (4c) is in each location category other term frequency-inverse document frequency Score value TF-IDF_u,cOn the basis of, use Dynamic gene to adjust cosine similarity computing formula, calculate between two users Similarity cos (u_i,u_j):

$\cos (u_i,u_j)=\frac{\underset{k=1}{\overset{n}{\Σ}}{I}_k\×J_k\×(F_{\max }-F_{u_i,u_j,k})}{\sqrt{\underset{k=1}{\overset{n}{\Σ}}{\left(I_k\right)}^2}\×\sqrt{\underset{k=1}{\overset{n}{\Σ}}{\left(J_k\right)}^2}},$

Wherein, I_kFor i-th user u_iFor the term frequency-inverse document frequency score of place classification k, J_kFor jth user u_jRight In the term frequency-inverse document frequency score of place classification k, F_maxFor the maximum in all Dynamic gene.

Step 6, social network server makes final friend recommendation.

Between user and every other user that request is recommended by social network server, repeat the above steps (3) is to step (5), obtaining the similarity score of this user and other users, n the user choosing score value the highest sends out as recommendation results The user that the request of giving is recommended, the friend recommendation quantity that the user that n recommends by request is asked.

The effect of the present invention can be further illustrated by following emulation experiment:

1. experiment condition is arranged

Condition 1, obtains, at Foursquare, data set of registering, and chooses the data composition data acquisition system of registering in wherein San Francisco, And choose 100 most users of quantity of wherein registering 50765 register data as the source data tested.

Condition 2, at Intel (R) Celeron (R) G540 processor, the desk-top meter of Windows 7 Ultimate operating system The result of test experiments on calculation machine.

2. experiment content and result

Experiment 1: the experimental data after filtering is divided into training set and test set, first runs the present invention on training set Recommended program, the good friend of the user recommending most like 20 users of user to recommend as request with request that will obtain； On test set, run identical recommended program subsequently, take n most like user respectively and make recommendation, N=1,2,3 ..., 20；Then run on training set and test set and do not consider that the legacy buddy suggested design of time factor obtains Two kinds of accuracys rate recommending method are finally calculated by corresponding buddy list and recommendation results, and computing formula is as follows:

$\mathrm{Pr}ecision=\frac{{\mathrm{\Σ}}_{u\∈U}|R\left(u\right)\∩T\left(u\right)|}{\left|T\left(u\right)\right|}$

Wherein R (u) represents the buddy list calculated from training set, and T (u) represents that the friend recommendation made by test set is arranged Table, U represents all of user in data set.

The situation of change that relatively the two accuracy rate increases with commending friends quantity, result is as shown in Figure 3.

It can be seen from figure 3 that after the good friend's quantity recommended is more than 5, the friend recommendation scheme used by the present invention accurate Rate is above not considering traditional friend recommendation scheme of time factor.Reality use scene under, it is recommended that good friend's number Amount is usually above 5, it can be said that the accuracy rate of the friend recommendation scheme used by the present invention is substantially better than traditional not examining Consider the suggested design of time factor.

Experiment 2: the experimental data after filtering is divided into training set and test set, first runs the present invention on training set Recommended program, the good friend of the user recommending most like 20 users of user to recommend as request with request that will obtain； On test set, run identical recommended program subsequently, take n most like user respectively and make recommendation, N=1,2,3 ..., 20；Then run on training set and test set and do not consider that the legacy buddy suggested design of time factor obtains Two kinds of recall rates recommending method are finally calculated by corresponding buddy list and recommendation results, and computing formula is as follows:

$\mathrm{Re}call=\frac{{\mathrm{\Σ}}_{u\∈U}|R\left(u\right)\∩T\left(u\right)|}{\left|R\left(u\right)\right|}$

Wherein R (u) represents the buddy list calculated from training set, and T (u) represents that the friend recommendation made by test set is arranged Table, U represents all of user in data set.

The situation of change that relatively the two recall rate increases with commending friends quantity, result is as shown in Figure 4.

As seen from Figure 4, when the good friend's negligible amounts recommended, the friend recommendation scheme used by the present invention is not examined with traditional The recall rate difference of the friend recommendation scheme considering time factor is little, along with the increase of commending friends quantity, used by the present invention Scheme recall rate gradually exceed the friend recommendation scheme of traditional time that do not considers.Also due in the use scene of reality Under, it is recommended that good friend's quantity the most more, it can be said that the recall rate of the friend recommendation scheme used by the present invention is the most excellent In traditional suggested design not considering time factor.

In sum, the present invention is superior to traditional friend recommendation system in the accuracy rate of recommendation results in terms of recall rate, Can be that user based on position social networks provides quality higher friend recommendation service.

Claims (6)

1. a friend recommendation method based on position Yu time in the social networks of position, including:

(1) a communication system framework being made up of user, location facilities, social network server is set up, wherein:

User, for communicating with location facilities and social network server by mobile cellular network or WiFi；

Location facilities, realizes location for collaborative user's GPS module；

Social network server, for storing the positional information of user, and carries out friend recommendation to user on this basis；

(2) longitude and latitude of position and current temporal information are sent to social network server by user, generate information of registering；

(3) the storage information of registering is pre-processed by social network server:

(3a) for the longitude and latitude in the information of registering, the position semantic database searching server obtains corresponding with place of registering Semantic information, then the information of registering of each user is added up by the semantic information classification in its place of registering and registers number of times；

(3b) for the temporal information in the information of registering, on the basis of step (3a), each user is added up each Behavior of registering under some classification distribution in time；

(4) according to user at the number of times of registering of each place classification, social network server calculates each user u for each Term frequency-inverse document frequency score TF-IDF of some classification c_u,c；

(4a) according to user u number of times of registering under the classification of each place, social network server calculates user u each The word frequency score value TF of information of registering under some classification c_u,c；

(4b) according to all users number of times of registering under the classification of each place, social network server calculates each place classification The inverse document frequency score value IDF of c_c；

(4c) social network server is by user u word frequency score value TF under classification c of place_u,cInverse document with place classification c Frequency score IDF_cIt is multiplied, obtains its term frequency-inverse document frequency score under this place classification TF-IDF_u,c=TF_u,c×IDF_c；

(5) similarity between social network server calculating user:

(5a) use user to register under the classification of each place behavior distribution in time, calculate two users in same place Register under classification the relative entropy D (P Q) being distributed, and wherein P and Q is two user u_iAnd u_jRegistering of same place classification Number of times probability distribution in time, then uses relative entropy to calculate similarity Dynamic gene

(5b) for user in each location category other term frequency-inverse document frequency score TF-IDF_u,c, use the similar of correspondence Degree Dynamic geneAdjust cosine similarity computing formula, calculate the similarity cos (u between two users_i,u_j)；

(6) repeat the above steps (3) extremely step between user and the every other user that request is recommended by social network server Suddenly (5), obtaining the similarity score of this user and other users, n the user choosing score value the highest sends out as recommendation results The user that the request of giving is recommended, the friend recommendation quantity that the user that n recommends by request is asked.

Method the most according to claim 1, wherein the behavior of registering of step (3b) counting user is when being distributed in time, It is that the time is divided into by one day 5 intervals, respectively 2:00-7:00,7:00-11:00,11:00-17:00,17: 00-22:00,22:00-2:00, counting user u in classification c of place the number of times vis that registers on each time interval (c, u) With user at total number of times sum (c) of registering of this place classification, obtain the user's registering frequency at each time interval of place classification c ((c, v)/sum (c) obtain user and register behavior distribution in time in certain place classification rate P for c, v)=vis.

Method the most according to claim 1, wherein in step (4a), social network server calculates user u each The word frequency score value TF of information of registering under classification c of place_u,c, it is to be calculated by equation below:

${\mathrm{TF}}_{u,c}=\frac{\left|\right\{u.v_i:v_i.D=c\left\}\right|}{|u.V|},$

Wherein, | { u.v_i:v_i.D=c} | being the number of times registered under classification c of place of user u, | u.V | is that user u registers in all places Total degree.

Method the most according to claim 1, wherein the inverse document frequency score value IDF of step (4b) place classification c_cPass through Following formula calculates:

${\mathrm{IDF}}_c=log\frac{\left|U\right|}{\left|\right\{u_j:c\∈u_j.D\left\}\right|},$

Wherein, | U | represents the quantity of all users in server, | { u_j:c∈u_j.D} | the user's that expression was once registered in place classification c Quantity.

Method the most according to claim 1, wherein calculates two user u in step (5a)_iAnd u_jFor place classification c Dynamic gene during information similarity of registeringCarry out as follows:

(5a1) i-th user u is calculated respectively_iWith jth user u_jRelative entropy D (P in place classification c_c||Q_c) and jth Individual user u_jWith i-th user u_iRelative entropy D (Q in place classification c_c||P_c):

Wherein, P_cFor i-th user u_iAt the probability distribution registered in time of place classification c, Q_cFor jth user u_jIn place The probability distribution registered in time of classification c, i=1,2,3,4,5 the most corresponding steps (3b) calculate user and register in time 5 time intervals divided during probability distribution；

(5a2) Dynamic gene of information similarity of registering according to result calculating place classification c of (5a1)As follows:

$F_{u_i,u_j,c}=\frac{2}{D\left(P_c\right|\left|Q_c\right)+D\left(Q_c\right|\left|P_c\right)+1}.$

Method the most according to claim 1, wherein step (5b) utilizes between cosine similarity two users of calculating similar Degree cos (u_i,u_j), calculated by following formula:

$cos(u_i,u_j)=\frac{\underset{k=1}{\overset{n}{\Σ}}{I}_k\×J_k\×(F_{max}-F_{u_i,u_j,k})}{\sqrt{\underset{k=1}{\overset{n}{\Σ}}{\left(I_k\right)}^2}\×\sqrt{\underset{k=1}{\overset{n}{\Σ}}{\left(J_k\right)}^2}},$

Wherein, I_kFor i-th user u_iFor the term frequency-inverse document frequency score of place classification k, J_kFor jth user u_jFor The term frequency-inverse document frequency score of place classification k, F_maxFor the maximum in all Dynamic gene.