CN104239399A - Method for recommending potential friends in social network - Google Patents

Abstract

The invention relates to a method for recommending potential friends in a social network and realizes the recommendation for the potential friends through two layers of moulds. The method comprises the flowing steps of in the first layer, firstly carrying out subject classification on the articles of all users, constructing a user attention vector, and calculating the cosine similarity of the users to obtain a rough potential-friend recommending list; in the second layer, taking the interest changes with time of the users into account, taking change factors into the calculation of the similarity, and then further detailing the friend recommendation on the basis of the rough recommending list in the first layer. Because the characteristic that the interests of the users are reflected by the articles is utilized, by starting from the interest similarity and bypassing the limitation of an existing social network on the basis of the original friend relationship of the users, the friend recommending method based on the interest similarity is realized, the friend recommending range is expanded, and the recommendation of the friends is further more accurately and effectively realized within a large range.

Description

Potential friend recommendation method in social networks

Technical field

The present invention relates to computer program field, particularly relate to a kind of potential friend recommendation method in social networks.

Background technology

The rise of Web2.0 allows increasing network user participate in social networks, and they hanker after carrying out resource sharing and information exchange, is got more and more by the mutual interchange of online social networks.The research producing content for user not only can help businessman to understand the hobby trend of all kinds of different user group to commodity, also can improve multiple network service, has great importance for lifting Consumer's Experience.

Because the network user has been not content with the social circle of the circle of friends formation in reality gradually, in social networks, the strange user having common interest hobby is found to become the demand of numerous user.In social networks, the friend-making circle extended one's service is one of major way of social networks development, effectively finds interested contact person can give user and makes friends or professional help.But the friend recommendation function that existing system provides is not very good.Therefore, propose a kind of effective friend recommendation mechanism and there is very strong realistic meaning.

Find by literature search, Chen W and Fong S at article " Social network collaborating filtering framework and online trust factors:A case study on Facebook " (" by research Facebook data, social networks collaborative filtering filtering model in conjunction with trust factor ") (2010fifth international conference on ICDIM, the recommend method that trust factor combines with collaborative filtering model is proposed IEEE2010:266-273), be specially: first, the method of data mining (decision tree and correlation rule) is utilized to analyze user data, show that in doings, each key element is on the impact of users to trust degree.Then these degree of belief influence factors quantized, the part as collaborative filtering model inputs.Finally, collaborative filtering recommending is carried out in conjunction with user characteristics.But there is certain difficulty when quantizing the weight of each key element of degree of belief in the method, well can not determine the weight of each key element, this certainly will bring certain impact to final recommendation results.

Find by literature search, the people such as Nitai B.Silva and Ing-Ren Tsang propose to utilize network topology structure to carry out commending friends at article " A graph-based friend recommendation system using genetic algorithm " (" utilizing the friend recommendation system based on graph structure of genetic algorithm ") (2010IEEE congress on Evolutionary Computation, 2010:1-7).Be specially: first, be partitioned into the local subgraph in active user's certain distance, then utilize the network structure feature of the good friend of Analysis of Genetic Algorithms user and the good friend of good friend, filter out and become the larger nodes recommendations of good friend's probability to user.But this commending system is a kind of acquaintance's recommendation mechanisms, not very remarkable to the contribution of user development friend-making circle.

Summary of the invention

The technical problem to be solved in the present invention how to expand the scope of friend recommendation, so on a large scale in realize the recommendation of good friend more accurately and effectively.

In order to solve this technical matters, the invention provides the potential friend recommendation method in two kinds of similar social networks, wherein a kind of method comprises the steps:

S01: obtain from social networking service device all articles that every user delivered in a year, and carry out stored record;

S02: add up the article sum of every user under each theme;

After completing steps S02, carry out step S03 and S04 respectively, after completing steps S04, implementation step S05, wherein:

S03: utilize the cosine law according to the rough similarity between following formulae discovery user:

${\mathrm{similarity}}_{(i,j)}=\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,t)}\×n_{(j,t)}}{\sqrt{{n_{(i,1)}}^2+{n_{(i,2)}}^2+...+{n_{(i,k)}}^2}\sqrt{{n_{(j,1)}}^2+{n_{(j,2)}}^2+...+{n_{(j,k)}}^2}}$

Wherein, similarity _{(i, j)}for the rough similarity of user i and user j, the quantity that k is the theme, n _{i, t}represent the article number that user i delivers under theme t, n _{j, t}represent the article number that user j delivers under theme t;

For a specific user, to other users according to similarity similarity rough between itself and this specific user _{(i, j)}size arrange, select the potential user rough recommendation tables of the highest x of a similarity user as this specific user;

S04: the article number delivered under k theme every month according to each user, according to the difference dif between the following formulae discovery m month and the m+1 month _i[m]:

${\mathrm{dif}}_i\left[m\right]=1-\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,m+1,t)}\×n_{(j,m,t)}}{\sqrt{n_{(i,m+\mathrm{1,1})}^2+n_{(i,m+\mathrm{1,2})}^2+...+n_{(i,m+1,k)}^2}\×\sqrt{n_{(i,m,1)}^2+n_{(i,m,2)}^2+...+n_{(i,m,k)}^2}}$

And then according to the interests change susceptibility λ of each user in 1 year in a following formulae discovery x user _i:

${\mathrm{\λ}}_i=\frac{\underset{i=1}{\overset{11}{\mathrm{\Σ}}}{\mathrm{dif}}_i\left[m\right]}{11}$

S05: that adds up article delivers the time, by the similarity sim of two users in following formulae discovery theme t _t(i, j):

${\mathrm{sim}}_t(i,j)=\frac{{\mathrm{\Σ}}_{b=1}^{\left|n_{(i,t)}\right|}{\mathrm{\Σ}}_{a=1}^{\left|n_{(j,t)}\right|}{e}^{-\mathrm{\λ}(\mathrm{time}\left(a\right)-\mathrm{time}\left(b\right))}}{\left|n_{(i,t)}\right|\×\left|n_{(j,t)}\right|}$

Wherein, n _{(i, t)}with n _{(j, t)}be respectively user i in theme t, article that user j has separately sum, λ is sensitive factor, and described sensitive factor λ is obtained by following formulae discovery:

$\mathrm{\λ}=\frac{({\mathrm{\λ}}_i+{\mathrm{\λ}}_j)}{2}$

Finally, refining similar degree sim (i, j) between two users is obtained by following formula

$\mathrm{sim}(i,j)=\frac{{\mathrm{\Σ}}_{t=1}^{k-1}{\mathrm{sim}}_t(i,j)*(n_{(i,t)}+n_{(j,t)})}{{\mathrm{\Σ}}_{t=1}^{k-1}(n_{(i,t)}+n_{(j,t)})}$

After completing S03 and S05, implement following steps:

S06: for this specific user, by x user in its rough recommendation list according to the refining similar degree sim (i between itself and this specific user, j) size arranges, and obtains final recommendation list, thus to the recommendation list that this specific user provides this final.

Potential friend recommendation method in another kind of social networks, comprises the steps:

S01: obtain from social networking service device all articles that every user delivered in a year, and carry out stored record;

S02: add up the article sum of every user under each theme;

S03: utilize the cosine law according to the rough similarity between following formulae discovery user:

${\mathrm{similarity}}_{(i,j)}=\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,t)}\×n_{(j,t)}}{\sqrt{{n_{(i,1)}}^2+{n_{(i,2)}}^2+...+{n_{(i,k)}}^2}\sqrt{{n_{(j,1)}}^2+{n_{(j,2)}}^2+...+{n_{(j,k)}}^2}}$

Wherein, similarity _{(i, j)}for the rough similarity of user i and user j, the quantity that k is the theme, n _{i, t}represent the article number that user i delivers under theme t, n _{j, t}represent the article number that user j delivers under theme t;

For a specific user, to other users according to similarity similarity rough between itself and this specific user _{(i, j)}size arrange, select the potential user rough recommendation tables of the highest x of a similarity user as this specific user;

S04: the article number delivered under k theme every month according to each user in this specific user and the rough recommendation tables of potential user, according to the difference dif between the following formulae discovery m month and the m+1 month _i[m]:

${\mathrm{dif}}_i\left[m\right]=1-\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,m+1,t)}\×n_{(j,m,t)}}{\sqrt{n_{(i,m+\mathrm{1,1})}^2+n_{(i,m+\mathrm{1,2})}^2+...+n_{(i,m+1,k)}^2}\×\sqrt{n_{(i,m,1)}^2+n_{(i,m,2)}^2+...+n_{(i,m,k)}^2}}$

And then according to the wherein interests change susceptibility λ of each user in 1 year in a following formulae discovery x user _i:

${\mathrm{\λ}}_i=\frac{\underset{i=1}{\overset{11}{\mathrm{\Σ}}}{\mathrm{dif}}_i\left[m\right]}{11}$

S05: that adds up each user's article in this specific user and the rough recommendation tables of potential user delivers the time, by the similarity sim of two users in following formulae discovery theme t _t(i, j):

${\mathrm{sim}}_t(i,j)=\frac{{\mathrm{\Σ}}_{b=1}^{\left|n_{(i,t)}\right|}{\mathrm{\Σ}}_{a=1}^{\left|n_{(j,t)}\right|}{e}^{-\mathrm{\λ}(\mathrm{time}\left(a\right)-\mathrm{time}\left(b\right))}}{\left|n_{(i,t)}\right|\×\left|n_{(j,t)}\right|}$

Wherein, n _{(i, t)}with n _{(j, t)}be respectively user i in theme t, article that user j has separately sum, λ is sensitive factor, and described sensitive factor λ is obtained by following formulae discovery:

$\mathrm{\λ}=\frac{({\mathrm{\λ}}_i+{\mathrm{\λ}}_j)}{2}$

Finally, refining similar degree sim (i, j) between two users is obtained by following formula

$\mathrm{sim}(i,j)=\frac{{\mathrm{\Σ}}_{t=1}^{k-1}{\mathrm{sim}}_t(i,j)*(n_{(i,t)}+n_{(j,t)})}{{\mathrm{\Σ}}_{t=1}^{k-1}(n_{(i,t)}+n_{(j,t)})}$

S06: for this specific user, by x user in its rough recommendation list according to the refining similar degree sim (i between itself and this specific user, j) size arranges, and obtains final recommendation list, thus to the recommendation list that this specific user provides this final.

In described step S01, all articles that every user delivered in a year store in the following manner:

Blog(user _i)＝{b ₁，b ₂，......，b _q}

Wherein, Blog (user _i) represent all article set of user i, b _qrefer to Blog (user _i) in q section article.

Also comprise the steps: between described step S01 and S02

For the article of every user, based on the tag along sort preset, utilize many labels subject classification model to classify, after classification, every section of article obtains three labels, the corresponding theme of each label.

Many labels subject classification model comprises ground floor model and second layer model;

Described ground floor model, for being that every section of article carries out word segmentation processing by participle instrument, filtered the effective word obtaining every section of article with the filtration of stop-word subsequently by part of speech, use term frequency-inverse document frequency to calculate the weight of each word afterwards, thus obtain the keyword of every section of article;

Described second layer model, for the keyword obtained in described ground floor model and label being carried out semantic matches according to semantic knowledge-base, afterwards, the label obtained the keyword of every section of article uses the method for multi-source evidence fusion to calculate the final label of every section of article.

The method of described multi-source evidence fusion have employed Dempster composition rule.

Described participle instrument is ICTCLAS participle instrument.

Described semantic knowledge-base is " knowing net " and " Chinese thesaurus expansion version ".

In described step S03, first according to the article sum under each theme of user, obtain the article theme attention rate vector of user:

user _i＝((topic ₁，n _i，1)，(topic ₂，n _i，2)，......，(topic _k，n _i，k))

Wherein, user _irepresent user i, topic _trepresent t theme, the quantity that k is the theme;

And then carry out rough similarity similarity according to this theme concern vector _{(i, j)}.

In described step S04, first according to the article sum under user's every month each theme, to obtain in 1 year theme in each month and pay close attention to distribution table:

user _i[m]＝{(topic ₁，n _i，m，1)，(topic ₂，n _i，m，2)，......，(topic _k，n _i，m，k)}

Wherein, m represents month, n _{i, m, t}represent the article number that user i delivers under the m month, theme t, the quantity that k is the theme.

The present invention realizes the recommendation to potential good friend by two-layer model.First ground floor is carry out subject classification to the article of each user, constructs user's attention rate vector, calculates cosine (cosine) similarity of user, obtain rough potential friend recommendation list.User's interests change is in time counted consideration by the second layer, is considered by changing factor in Similarity Measure, then based on the further refinement friend recommendation of alligatoring recommendation tables of ground floor.Owing to make use of this feature of article reflection user interest, from similar interests degree, walk around the restriction of existing social networks based on the original friend relation of user, achieve the friend recommendation method based on Interest Similarity, expanded the scope of friend recommendation, so on a large scale in realize the recommendation of good friend more accurately and effectively.

Accompanying drawing explanation

Fig. 1 is potential friend recommendation illustraton of model in the embodiment of the present invention 1;

Fig. 2 is keyword semantic matches and blog classification process figure in the embodiment of the present invention 1;

Fig. 3 is the rough recommendation figure based on total Interest Similarity in the embodiment of the present invention 1;

Fig. 4 is the refinement recommendation figure counting the article time in the embodiment of the present invention 1.

Embodiment

Below with reference to Fig. 1 to Fig. 4, by two embodiments, potential friend recommendation method in social networks provided by the invention is described in detail, it is optional two embodiments of the present invention, can think, those skilled in the art, in the scope not changing the present invention's spirit and content, can modify to it and polish.

Embodiment 1

Please refer to Fig. 1, present embodiments provide a kind of potential friend recommendation method in social networks, the present embodiment utilizes the open data platform of Renren Network to obtain data set, and process according to flow process shown in Fig. 1 of the present invention these data, the method comprises the steps:

S01: obtain all articles that every user delivered in a year from social networking service device, and carry out stored record, in the present embodiment, namely first 5314 good friends are obtained from Renren Network api interface, share article and 138901 sections of original articles for 17956 sections that excavate them altogether, article wherein, also can be described as blog article; In described step S01, all articles that every user delivered in a year store in the following manner:

Blog(user _i)＝{b ₁，b ₂，......，b _q}

Wherein, Blog (user _i) represent all article set of user i, b _qrefer to Blog (user _i) in q section article.

Theme belonging to article can to refer in social networks article just selected at the beginning of writing belonging to theme, so in the case, theme is known, be stored in the fix information in social networking service device, this method just transfers use; Certainly, affiliated theme also can be determined as described in the embodiment, and in order to determine the theme of article, the present embodiment, after described step S01, before S02, also comprises the steps:

For the article of every user, based on the tag along sort preset, utilize many labels subject classification model to classify, after classification, every section of article obtains three labels, the corresponding theme of each label.

Specifically, in the present embodiment, first, for the article of every user, based on Sina's 41 tag along sorts (see table 1), many labels subject classification model is utilized to classify.After classification, every section of article obtains three labels:

b _q＝{label _(q,1),label _(q,2),label _(q,3)}

Described many labels subject classification model, refers to the theme variation based on blog content and the feature such as dispersed, have employed the blog subject classification model of multi-class, many labelings.This model is mainly divided into two-layer:

Ground floor model adopts vector space model (Vector space model, VSM) to carry out text representation.Based on the concrete feature of Chinese, it is that every section of article carries out word segmentation processing that model have employed ICTCLAS participle instrument, the effective word obtaining every section of article with the filtration of stop-word is filtered subsequently by part of speech, TF-IDF (Term frequency – Inverse document frequency, term frequency-inverse document frequency) is used to calculate the weight of each word and obtain the keyword of every section of article afterwards.

Furthermore, utilize vector space model (Vector space model, VSM) to carry out text representation for the article of every user to be:

B(q)＝{(KeyWord ₁，w ₁)，(KeyWord ₂，w ₂)，......，(KeyWord _i，w _i)}

Wherein, B (q) represents effective entry set of article q, w _irepresent KeyWord _iweighted value.

Second layer model uses two kinds of semantic knowledge-bases: " knowing net " and " Chinese thesaurus expansion version ", and the tag along sort of determining of the keyword obtained in ground floor model described in the and sina blog is carried out semantic matches.Afterwards, the tag along sort obtained the keyword of every section of article uses the method for multi-source evidence fusion: Dempster composition rule, calculates the class label that every section of article is final.Semantic matches and labeling process flow diagram are as shown in Fig. 2.

The theme of article can carry out subsequent step after determining.In the present embodiment, please refer to Fig. 1, after article theme is determined, carry out step S03 and S04 respectively, there is no precedence relationship between step 03 and step 04, after completing steps S04, implementation step S05, as long as implement S06 again after completing S03 and S05 respectively.S04 in the present embodiment calculates the interests change susceptibility of each user, and then carries out the calculating of refining similar degree.

About step S03: the rough process flow diagram recommended as shown in Figure 3.

Wherein, utilize the cosine law according to the rough similarity between following formulae discovery user:

${\mathrm{similarity}}_{(i,j)}=\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,t)}\×n_{(j,t)}}{\sqrt{{n_{(i,1)}}^2+{n_{(i,2)}}^2+...+{n_{(i,k)}}^2}\sqrt{{n_{(j,1)}}^2+{n_{(j,2)}}^2+...+{n_{(j,k)}}^2}}$

Wherein, similarity _{(i, j)}for the rough similarity of user i and user j, the quantity that k is the theme, n _{i, t}represent the article number that user i delivers under theme t, n _{j, t}represent the article number that user j delivers under theme t;

For a specific user, to other users according to similarity similarity rough between itself and this specific user _{(i, j)}size arrange, select the potential user rough recommendation tables of the highest x of a similarity user as this specific user;

Specifically, in described step S03, first according to the article sum under each theme of user, obtain the article theme attention rate vector of user:

user _i＝((topic ₁，n _i，1)，(topic ₂，n _i，2)，......，(topic _k，n _i，k))

Wherein, user _irepresent user i, topic _trepresent t theme, the quantity that k is the theme;

And then carry out rough similarity similarity according to this theme concern vector _{(i, j)}.

About step S04: the article number delivered under k theme every month according to each user, according to the difference dif between the following formulae discovery m month and the m+1 month _i[m]:

${\mathrm{dif}}_i\left[m\right]=1-\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,m+1,t)}\×n_{(j,m,t)}}{\sqrt{n_{(i,m+\mathrm{1,1})}^2+n_{(i,m+\mathrm{1,2})}^2+...+n_{(i,m+1,k)}^2}\×\sqrt{n_{(i,m,1)}^2+n_{(i,m,2)}^2+...+n_{(i,m,k)}^2}}$

And then according to the interests change susceptibility λ of each user in 1 year in a following formulae discovery x user _i, namely obtain the interests change susceptibility λ of user in 1 year by the mean value of 11 differences again _i:

${\mathrm{\λ}}_i=\frac{\underset{i=1}{\overset{11}{\mathrm{\Σ}}}{\mathrm{dif}}_i\left[m\right]}{11}$

In described step S04, article sum first under each theme foundation user's every month, to obtain in 1 year theme in each month and pay close attention to distribution table, specifically, it is using the moon as node, the interest distribution of counting user every month, builds the theme of user in 1 year in each month m and pays close attention to distribution table:

user _i[m]＝{(topic ₁，n _i，m，1)，(topic ₂，n _i，m，2)，......，(topic _k，n _i，m，k)}

Wherein, m represents month, n _{i, m, t}represent the article number that user i delivers under the m month, theme t, the quantity that k is the theme.

Connection with step S05: that adds up article delivers the time, by the similarity sim of two users in following formulae discovery theme t _t(i, j):

${\mathrm{sim}}_t(i,j)=\frac{{\mathrm{\Σ}}_{b=1}^{\left|n_{(i,t)}\right|}{\mathrm{\Σ}}_{a=1}^{\left|n_{(j,t)}\right|}{e}^{-\mathrm{\λ}(\mathrm{time}\left(a\right)-\mathrm{time}\left(b\right))}}{\left|n_{(i,t)}\right|\×\left|n_{(j,t)}\right|}$

Wherein, n _{(i, t)}with n _{(j, t)}be respectively user i in theme t, article that user j has separately sum, λ is sensitive factor, and described sensitive factor λ is obtained by following formulae discovery:

$\mathrm{\λ}=\frac{({\mathrm{\λ}}_i+{\mathrm{\λ}}_j)}{2}$

Finally, refining similar degree sim (i, j) between two users is obtained by following formula

$\mathrm{sim}(i,j)=\frac{{\mathrm{\Σ}}_{t=1}^{k-1}{\mathrm{sim}}_t(i,j)*(n_{(i,t)}+n_{(j,t)})}{{\mathrm{\Σ}}_{t=1}^{k-1}(n_{(i,t)}+n_{(j,t)})}$

About step S05: obtaining the interests change susceptibility λ of every user self _iafterwards, get back in each theme class further, the joining day factor calculates.

To any theme j, find each self-corresponding article collection of user i and user j under this theme: Blog (user _i, topic _t) and Blog (user _j, topic _t).That traces back to every section of article in theme delivers time time (b).

Counting user article deliver the time, by the similarity sim of two users in following formulae discovery theme t _t(i, j):

${\mathrm{sim}}_t(i,j)=\frac{{\mathrm{\Σ}}_{b=1}^{\left|n_{(i,t)}\right|}{\mathrm{\Σ}}_{a=1}^{\left|n_{(j,t)}\right|}{e}^{-\mathrm{\λ}(\mathrm{time}\left(a\right)-\mathrm{time}\left(b\right))}}{\left|n_{(i,t)}\right|\×\left|n_{(j,t)}\right|}$

Wherein, n _{(i, t)}with n _{(j, t)}be respectively user i in theme t, article that user j has separately sum, λ is sensitive factor, and λ is sensitive factor, if change λ → 0, then similarity does not affect by lead time substantially, if but λ → 1, then similarity can in time gap increase and greatly weaken.Interests change susceptibility due to user self truly reflects the susceptibility of user for time difference, and this model employs the interests change susceptibility λ of user i and j self _iwith λ _jcalculate sensitive factor.Described sensitive factor λ is obtained by following formulae discovery:

$\mathrm{\λ}=\frac{({\mathrm{\λ}}_i+{\mathrm{\λ}}_j)}{2}$

Finally, refining similar degree sim (i, j) between two users is obtained by following formula

$\mathrm{sim}(i,j)=\frac{{\mathrm{\Σ}}_{t=1}^{k-1}{\mathrm{sim}}_t(i,j)*(n_{(i,t)}+n_{(j,t)})}{{\mathrm{\Σ}}_{t=1}^{k-1}(n_{(i,t)}+n_{(j,t)})}$

After completing S03 and S05, implement following steps:

S06: for this specific user, by x user in its rough recommendation list according to the refining similar degree sim (i between itself and this specific user, j) size arranges, and obtains final recommendation list, thus to the recommendation list that this specific user provides this final.Its process can see Fig. 4.

The list of table 1 theme label

Embodiment 2

The present embodiment 1 is only with the difference of embodiment 1, in the present embodiment, S3, S4, S5 has sequencing, namely implement successively, so, the interests change susceptibility of x user in rough list and specific that user is only calculated in S4, the refining similar degree of x user in that user specific and rough list is only calculated in S5, and unlike in embodiment 1, calculate interests change susceptibility and the refining similar degree of all users, information needed can be transferred in all results in embodiment 1, only need calculate once, just can meet the friend recommendation of all users, embodiment 2 carries out the calculating of S4 and S5 respectively, namely need to calculate respectively.Except above difference, embodiment concrete in each step is all similar to embodiment 1.

Specifically, this enforcement provides a kind of potential friend recommendation method in social networks, comprises the steps:

S01: obtain from social networking service device all articles that every user delivered in a year, and carry out stored record;

S02: add up the article sum of every user under each theme;

S03: utilize the cosine law according to the rough similarity between following formulae discovery user:

${\mathrm{similarity}}_{(i,j)}=\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,t)}\×n_{(j,t)}}{\sqrt{{n_{(i,1)}}^2+{n_{(i,2)}}^2+...+{n_{(i,k)}}^2}\sqrt{{n_{(j,1)}}^2+{n_{(j,2)}}^2+...+{n_{(j,k)}}^2}}$

Wherein, similarity _{(i, j)}for the rough similarity of user i and user j, the quantity that k is the theme, n _{i, t}represent the article number that user i delivers under theme t, n _{j, t}represent the article number that user j delivers under theme t;

For a specific user, to other users according to similarity similarity rough between itself and this specific user _{(i, j)}size arrange, select the potential user rough recommendation tables of the highest x of a similarity user as this specific user;

S04: the article number delivered under k theme every month according to each user in this specific user and the rough recommendation tables of potential user, according to the difference dif between the following formulae discovery m month and the m+1 month _i[m]:

${\mathrm{dif}}_i\left[m\right]=1-\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,m+1,t)}\×n_{(j,m,t)}}{\sqrt{n_{(i,m+\mathrm{1,1})}^2+n_{(i,m+\mathrm{1,2})}^2+...+n_{(i,m+1,k)}^2}\×\sqrt{n_{(i,m,1)}^2+n_{(i,m,2)}^2+...+n_{(i,m,k)}^2}}$

And then according to the wherein interests change susceptibility λ of each user in 1 year in a following formulae discovery x user _i:

${\mathrm{\λ}}_i=\frac{\underset{i=1}{\overset{11}{\mathrm{\Σ}}}{\mathrm{dif}}_i\left[m\right]}{11}$

S05: that adds up each user's article in this specific user and the rough recommendation tables of potential user delivers the time, by the similarity sim of two users in following formulae discovery theme t _t(i, j):

${\mathrm{sim}}_t(i,j)=\frac{{\mathrm{\Σ}}_{b=1}^{\left|n_{(i,t)}\right|}{\mathrm{\Σ}}_{a=1}^{\left|n_{(j,t)}\right|}{e}^{-\mathrm{\λ}(\mathrm{time}\left(a\right)-\mathrm{time}\left(b\right))}}{\left|n_{(i,t)}\right|\×\left|n_{(j,t)}\right|}$

Wherein, n _{(i, t)}with n _{(j, t)}be respectively user i in theme t, article that user j has separately sum, λ is sensitive factor, and described sensitive factor λ is obtained by following formulae discovery:

$\mathrm{\λ}=\frac{({\mathrm{\λ}}_i+{\mathrm{\λ}}_j)}{2}$

Finally, refining similar degree sim (i, j) between two users is obtained by following formula

$\mathrm{sim}(i,j)=\frac{{\mathrm{\Σ}}_{t=1}^{k-1}{\mathrm{sim}}_t(i,j)*(n_{(i,t)}+n_{(j,t)})}{{\mathrm{\Σ}}_{t=1}^{k-1}(n_{(i,t)}+n_{(j,t)})}$

S06: for this specific user, by x user in its rough recommendation list according to the refining similar degree sim (i between itself and this specific user, j) size arranges, and obtains final recommendation list, thus to the recommendation list that this specific user provides this final.

In the present embodiment, use everybody network data, the calculating of method flow of the present invention and user's similarity is illustrated, the results show, when not considering sequence, the rough recommendation list that this model provides can obtain the effect up to 67%, and thinning process is by the rank precision improvement of recommendation list 17%.Thus demonstrate validity and the accuracy of method.

In sum, the present invention is by the recommendation of two-layer model realization to potential good friend.First ground floor is carry out subject classification to the article of each user, constructs user's attention rate vector, calculates cosine (cosine) similarity of user, obtain rough potential friend recommendation list.User's interests change is in time counted consideration by the second layer, is considered by changing factor in Similarity Measure, then based on the further refinement friend recommendation of alligatoring recommendation tables of ground floor.Owing to make use of this feature of article reflection user interest, from similar interests degree, walk around the restriction of existing social networks based on the original friend relation of user, achieve the friend recommendation method based on Interest Similarity, expanded the scope of friend recommendation, so on a large scale in realize the recommendation of good friend more accurately and effectively.

Claims (10)

1. the potential friend recommendation method in social networks, comprises the steps:

S01: obtain from social networking service device all articles that every user delivered in a year, and carry out stored record;

S02: add up the article sum of every user under each theme;

After completing steps S02, carry out step S03 and S04 respectively, after completing steps S04, implementation step S05, wherein:

S03: utilize the cosine law according to the rough similarity between following formulae discovery user:

${\mathrm{similarity}}_{(i,j)}=\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,t)}\×n_{(j,t)}}{\sqrt{{n_{(i,1)}}^2+{n_{(i,2)}}^2+...+{n_{(i,k)}}^2}\sqrt{{n_{(j,1)}}^2+{n_{(j,2)}}^2+...+{n_{(j,k)}}^2}}$

Wherein, similarity _{(i, j)}for the rough similarity of user i and user j, the quantity that k is the theme, n _{i, t}represent the article number that user i delivers under theme t, n _{j, t}represent the article number that user j delivers under theme t;

For a specific user, to other users according to similarity similarity rough between itself and this specific user _{(i, j)}size arrange, select the potential user rough recommendation tables of the highest x of a similarity user as this specific user;

S04: the article number delivered under k theme every month according to each user, according to the difference dif between the following formulae discovery m month and the m+1 month _i[m]:

${\mathrm{dif}}_i\left[m\right]=1-\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,m+1,t)}\×n_{(j,m,t)}}{\sqrt{n_{(i,m+\mathrm{1,1})}^2+n_{(i,m+\mathrm{1,2})}^2+...+n_{(i,m+1,k)}^2}\×\sqrt{n_{(i,m,1)}^2+n_{(i,m,2)}^2+...+n_{(i,m,k)}^2}}$

And then according to the interests change susceptibility λ of each user in 1 year in a following formulae discovery x user _i:

${\mathrm{\λ}}_i=\frac{\underset{i=1}{\overset{11}{\mathrm{\Σ}}}{\mathrm{dif}}_i\left[m\right]}{11}$

S05: that adds up article delivers the time, by the similarity sim of two users in following formulae discovery theme t _t(i, j):

${\mathrm{sim}}_t(i,j)=\frac{{\mathrm{\Σ}}_{b=1}^{\left|n_{(i,t)}\right|}{\mathrm{\Σ}}_{a=1}^{\left|n_{(j,t)}\right|}{e}^{-\mathrm{\λ}(\mathrm{time}\left(a\right)-\mathrm{time}\left(b\right))}}{\left|n_{(i,t)}\right|\×\left|n_{(j,t)}\right|}$

Wherein, n _{(i, t)}with n _{(j, t)}be respectively user i in theme t, article that user j has separately sum, λ is sensitive factor, and described sensitive factor λ is obtained by following formulae discovery:

$\mathrm{\λ}=\frac{({\mathrm{\λ}}_i+{\mathrm{\λ}}_j)}{2}$

Finally, refining similar degree sim (i, j) between two users is obtained by following formula

$\mathrm{sim}(i,j)=\frac{{\mathrm{\Σ}}_{t=1}^{k-1}{\mathrm{sim}}_t(i,j)*(n_{(i,t)}+n_{(j,t)})}{{\mathrm{\Σ}}_{t=1}^{k-1}(n_{(i,t)}+n_{(j,t)})}$

After completing S03 and S05, implement following steps:

S06: for this specific user, by x user in its rough recommendation list according to the refining similar degree sim (i between itself and this specific user, j) size arranges, and obtains final recommendation list, thus to the recommendation list that this specific user provides this final.

2. the potential friend recommendation method in social networks, comprises the steps:

S01: obtain from social networking service device all articles that every user delivered in a year, and carry out stored record;

S02: add up the article sum of every user under each theme;

S03: utilize the cosine law according to the rough similarity between following formulae discovery user:

${\mathrm{similarity}}_{(i,j)}=\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,t)}\×n_{(j,t)}}{\sqrt{{n_{(i,1)}}^2+{n_{(i,2)}}^2+...+{n_{(i,k)}}^2}\sqrt{{n_{(j,1)}}^2+{n_{(j,2)}}^2+...+{n_{(j,k)}}^2}}$

Wherein, similarity _{(i, j)}for the rough similarity of user i and user j, the quantity that k is the theme, n _{i, t}represent the article number that user i delivers under theme t, n _{j, t}represent the article number that user j delivers under theme t;

For a specific user, to other users according to similarity similarity rough between itself and this specific user _{(i, j)}size arrange, select the potential user rough recommendation tables of the highest x of a similarity user as this specific user;

S04: the article number delivered under k theme every month according to each user in this specific user and the rough recommendation tables of potential user, according to the difference dif between the following formulae discovery m month and the m+1 month _i[m]:

${\mathrm{dif}}_i\left[m\right]=1-\frac{\underset{t=1}{\overset{k}{\mathrm{\Σ}}}{n}_{(i,m+1,t)}\×n_{(j,m,t)}}{\sqrt{n_{(i,m+\mathrm{1,1})}^2+n_{(i,m+\mathrm{1,2})}^2+...+n_{(i,m+1,k)}^2}\×\sqrt{n_{(i,m,1)}^2+n_{(i,m,2)}^2+...+n_{(i,m,k)}^2}}$

And then according to the wherein interests change susceptibility λ of each user in 1 year in a following formulae discovery x user _i:

${\mathrm{\λ}}_i=\frac{\underset{i=1}{\overset{11}{\mathrm{\Σ}}}{\mathrm{dif}}_i\left[m\right]}{11}$

S05: that adds up each user's article in this specific user and the rough recommendation tables of potential user delivers the time, by the similarity sim of two users in following formulae discovery theme t _t(i, j):

${\mathrm{sim}}_t(i,j)=\frac{{\mathrm{\Σ}}_{b=1}^{\left|n_{(i,t)}\right|}{\mathrm{\Σ}}_{a=1}^{\left|n_{(j,t)}\right|}{e}^{-\mathrm{\λ}(\mathrm{time}\left(a\right)-\mathrm{time}\left(b\right))}}{\left|n_{(i,t)}\right|\×\left|n_{(j,t)}\right|}$

Wherein, n _{(i, t)}with n _{(j, t)}be respectively user i in theme t, article that user j has separately sum, λ is sensitive factor, and described sensitive factor λ is obtained by following formulae discovery:

$\mathrm{\λ}=\frac{({\mathrm{\λ}}_i+{\mathrm{\λ}}_j)}{2}$

Finally, refining similar degree sim (i, j) between two users is obtained by following formula

$\mathrm{sim}(i,j)=\frac{{\mathrm{\Σ}}_{t=1}^{k-1}{\mathrm{sim}}_t(i,j)*(n_{(i,t)}+n_{(j,t)})}{{\mathrm{\Σ}}_{t=1}^{k-1}(n_{(i,t)}+n_{(j,t)})}$

S06: for this specific user, by x user in its rough recommendation list according to the refining similar degree sim (i between itself and this specific user, j) size arranges, and obtains final recommendation list, thus to the recommendation list that this specific user provides this final.

3. the potential friend recommendation method in social networks as claimed in claim 1 or 2, is characterized in that: in described step S01, and all articles that every user delivered in a year store in the following manner:

Blog(user _i)＝{b ₁，b ₂，......，b _q}

Wherein, Blog (user _i) represent all article set of user i, b _qrefer to Blog (user _i) in q section article.

4. the potential friend recommendation method in social networks as claimed in claim 1 or 2, is characterized in that: also comprise the steps: between described step S01 and S02

For the article of every user, based on the tag along sort preset, utilize many labels subject classification model to classify, after classification, every section of article obtains three labels, the corresponding theme of each label.

5. the potential friend recommendation method in social networks as claimed in claim 4, is characterized in that: described many labels subject classification model comprises ground floor model and second layer model;

Described ground floor model, for being that every section of article carries out word segmentation processing by participle instrument, filtered the effective word obtaining every section of article with the filtration of stop-word subsequently by part of speech, use term frequency-inverse document frequency to calculate the weight of each word afterwards, thus obtain the keyword of every section of article;

Described second layer model, for the keyword obtained in described ground floor model and label being carried out semantic matches according to semantic knowledge-base, afterwards, the label obtained the keyword of every section of article uses the method for multi-source evidence fusion to calculate the final label of every section of article.

6. the potential friend recommendation method in social networks as claimed in claim 5, is characterized in that: the method for described multi-source evidence fusion have employed Dempster composition rule.

7. the potential friend recommendation method in social networks as claimed in claim 5, is characterized in that: described participle instrument is ICTCLAS participle instrument.

8. the potential friend recommendation method in social networks as claimed in claim 5, is characterized in that: described semantic knowledge-base is " knowing net " and " Chinese thesaurus expansion version ".

9. the potential friend recommendation method in social networks as claimed in claim 1 or 2, is characterized in that: in described step S03, first according to the article sum under each theme of user, obtains the article theme attention rate vector of user:

user _i＝((topic ₁，n _i，1)，(topic ₂，n _i，2)，......，(topic _k，n _i，k))

Wherein, user _irepresent user i, topic _trepresent t theme, the quantity that k is the theme;

And then carry out rough similarity similarity according to this theme concern vector _{(i, j)}.

10. the potential friend recommendation method in social networks as claimed in claim 1 or 2, is characterized in that: in described step S04, first according to the article sum under user's every month each theme, to obtain in 1 year theme in each month and pays close attention to distribution table:

user _i[m]＝{(topic ₁，n _i，m，1)，(topic ₂，n _i，m，2)，......，(topic _k，n _i，m，k)}

Wherein, m represents month, n _{i, m, t}represent the article number that user i delivers under the m month, theme t, the quantity that k is the theme.