CN104021230A - Collaborative filtering method based on community discovery - Google Patents

Abstract

The invention discloses a collaborative filtering method based on community discovery. The collaborative filtering method comprises the following steps: 1, converting a user-project network into an adjacent matrix form, wherein if the user scores a project, corresponding elements corresponding to the matrix have values, namely, score data; 2, constructing a user-user matrix, wherein elements in the user-user matrix are similarity of a user and a user, and a calculating method of the similarity adopts a novel Pearson-based similarity calculating method; 3, discovering a community structure through a community discovery method based on the user-user matrix, and performing prediction filling on partial deficient scores in the user-project matrix; 4, calculating similarity relations between a target user and a community and between the user and the user to construct a nearest neighbour candidate set, and completing recommending. According to the method, the problem of cold boot in a conventional filtering algorithm is effectively solved, the precision of algorithm recommendation is effectively improved through improving a similarity calculating formula and performing prediction filling on the deficient scores, and a better performance is achieved on the aspect of an average absolute error.

Description

A kind of collaborative filtering method based on community discovery

Technical field

The invention belongs to communication technical field, relate to a kind of collaborative filtering method based on community discovery.

Background technology

At present, in actual life, the quantity of information that people receive every day is increasing, people enjoyment information easily simultaneously also slowly by various unwanted information institute around.Numerous and diverse people of making of information have no idea to find in the short period of time the information oneself needing, and personalized recommendation service produces therefrom.Personalized recommendation service has not only been removed user from and in shiploads of merchandise, has been found the worry of admiring commodity, also to user, has brought shopping better to experience.

Except Collaborative Filtering Recommendation Algorithm, also have some algorithms based on collaborative filtering also can obtain reasonable division effect, for example, the collaborative filtering recommending method based on taxonomy of goods and user's classification that the people such as Shi Rongjie, Wang Shoujun propose; The collaborative filtering recommending method based on provincial characteristics that the people such as Li Li, Wei Baojun proposes; A kind of method Collaborative Filtering Recommendation System being optimized by polymerization that the people such as Rosin, Ou Yangyuanxin proposes etc.In addition, the socialization collaborative filtering recommending method based on trusting that the people such as Yang Bo, Lei Yu proposes; A kind of personalized collaborative filtering recommending method based on extension feature vector that the people such as Fan Bo, Su Hongyi proposes etc. has all been done good elaboration to the deficiency of collaborative filtering and improvement, but still exists some shortcomings about the research of collaborative filtering.

Some collaborative filtering related algorithms based on above-mentioned, although solved and had some problems in recommendation filter algorithm, but choose and the aspect such as user's similarity calculating at arest neighbors Candidate Set, still exist some shortcomings, the present invention proposes a kind of collaborative filtering based on community discovery, be intended to better improve the recommendation accuracy of proposed algorithm, the present invention concentrates while testing in the classical MovieLens test data of reality, algorithm performance stability and high efficiency, accuracy is high, algorithm is had very important significance and wide application prospect for follow-up collaborative filtering recommending.

Summary of the invention

In view of this, the object of the present invention is to provide a kind of collaborative filtering method based on community discovery, the method is by calculating the similarity between user and user, user-project network is converted to user-user network, then community discovery algorithm is applicable to user-user network and obtains corresponding community structure.According to the community structure obtaining, lack the prediction filling of scoring and choosing of arest neighbors Candidate Set, the scoring of last predictive user to project.

For achieving the above object, the invention provides following technical scheme:

A collaborative filtering method based on community discovery, comprises the following steps: 1) user-project network is converted to adjacency matrix form, if user marks to project, so corresponding to the respective element of matrix with regard to existence value, i.e. score data; 2) structuring user's-user matrix, the element in user-user matrix is user and user's similarity, the computing method of similarity adopt the novel similarity calculating method based on Pearson; 3) based on user-user matrix, by community discovery method, find community structure, and the excalation scoring in user-project matrix is predicted to filling; 4) calculate between targeted customer and community and the pass of the similarity between user and user series structure arest neighbors Candidate Set, and complete recommendation.

Further, in step 2) in, utilize user-project network structuring user's-user network, element in the corresponding user-user of user-user network adjacency matrix is the similarity between user, by new calculating formula of similarity called after Pearsimilarity, its concrete computing method are as follows:

${\mathrm{sim}}_{\mathrm{senti}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})=(1-\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{i}t}-r_{u_{j}t})}^2/\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{C_{\max }}^2})$

${\mathrm{sim}}_{\mathrm{CF}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})=\frac{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})(r_{u_{j}t}-\stackrel{\‾}{r_{u_j}})}{\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})}^2}\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{j}t}-\stackrel{\‾}{r_{u_j}})}^2}}$

$\mathrm{Pearsimilarity}(\stackrel{\→}{u_i},{\stackrel{\→}{u}}_j)=(1-\mathrm{\α})\×{\mathrm{sim}}_{\mathrm{CF}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})+\mathrm{\α}\×{\mathrm{sim}}_{\mathrm{senti}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})$

Wherein, what represent is the scoring vector of user i and user j, I _uiwith I _ujrepresent respectively the set of user i and user j scoring item, with represent user i and the scoring situation of user j to project t, α, for controlling parameter, controls the requirement of similarity formula to data susceptibility.

Further, in step 3) in, the disappearance scoring in user-project matrix is predicted to filling, when predicting filling, consider the factor aspect two: the relevance of marking between commodity, the impact of community member on project forecast scoring; Concrete score in predicting formula is as follows:

$\mathrm{user}\_\mathrm{rating}\left(x\right)=\stackrel{\‾}{r_u}+\underset{m\∈C_u\left(x\right)}{\mathrm{\Σ}}\frac{r_{\mathrm{mx}}-\stackrel{\‾}{r_m}}{\left|C_u\left(x\right)\right|}$

$\mathrm{Item}\_\mathrm{rating}\left(x\right)=\frac{\underset{y\∈\mathrm{Neigh}\left(I_x\right)}{\mathrm{\Σ}}{r}_y\×{\mathrm{sim}}_{\mathrm{item}}(x,y)}{\underset{y\∈\mathrm{Neigh}\left(I_x\right)}{\mathrm{\Σ}}{\mathrm{sim}}_{\mathrm{item}}(x,y)}$

rc _ux＝(1-β)×user_rating(x)+β×Item_rating(x)

the mean value that represents user u scoring, C _u(x) represent the set that the affiliated community of user u forms the user of project x scoring, | C _u(x) | represent the number of user in set, r _mxrepresent the scoring of user to project x in the affiliated community of user u, the average of its scoring, r _ythe score value that represents user u commodity y, Neigh (I _x) represent and the set of project x similar terms, wherein β is Controlling parameters.

Further, in step 4) in, according to the similarity between targeted customer and intercommunal similarity and targeted customer and other users, select arest neighbors Candidate Set, wherein targeted customer and intercommunal calculating formula of similarity are as follows:

$\mathrm{sim}\_\mathrm{corr}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=\frac{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})(r_{c_{j}t}-\stackrel{\‾}{r_{c_j}})}{\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})}^2}\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{c_{j}t}-\stackrel{\‾}{r_{c_j}})}^2}}$

$\mathrm{sim}\_\mathrm{dsenti}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=1-\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{u_{i}t}-r_{c_{j}t})}^2/\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{C_{\max }}^2}$

$\mathrm{sim}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=(1-\mathrm{\α})\×\mathrm{sim}\_\mathrm{corr}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})+\mathrm{\α}\×\mathrm{sim}\_\mathrm{dsenti}({\stackrel{\→}{u}}_i,\stackrel{\→}{C_j})$

Suppose that algorithm is divided into (C by user-user social networks ₁, C ₂, C ₃c _ic _t), t is community's number, although targeted customer belong to some communities this not represent that this targeted customer and other communities have no associated, so first calculate targeted customer and intercommunal similarity; The C of community _tmean value to commodity scoring is designated as centroid vector wherein represent the C of community _tin the average score of user to project j, represent respectively user i and the C of community _jscoring vector, for user u _ithe project set of evaluating, represent the C of community _jin all users institute scoring item destination aggregation (mda), represent user u _iscoring to project t, represent user u _ithe average of scoring, represent the C of community _jin the mean value of user to project t scoring, and for centroid vector in the average of each component, α is data sensitive degree Controlling parameters.

Further, according to step 4) selected arest neighbors Candidate Set, calculating the scoring of targeted customer to specific project, concrete score in predicting formula is as follows:

$r_u\left(x\right)=\stackrel{\‾}{r_u}+\frac{\underset{u^{\′}\∈\mathrm{Neigh}\left(u\right)}{\mathrm{\Σ}}[\mathrm{sim}(u,u^{\′})\×(r_{u^{\′}x}-\stackrel{\‾}{r_{u^{\′}}})]}{\underset{u^{\′}\∈\mathrm{Neigh}\left(u\right)}{\mathrm{\Σ}}\left|\mathrm{sim}(u,u^{\′})\right|}$

K the set that arest neighbors user forms of Neigh (u), sim (u, u') represents the similarity of user u and user u'; If there is score value to x in user u', so r _u'xrepresent the scoring of user u' to x, if user u' does not exist score value, so r to x _u'x=rc _ux, rc _uxrepresent prediction score value.

Beneficial effect of the present invention is: the collaborative filtering method based on community discovery provided by the invention, effectively solved the cold start-up problem in traditional filtering algorithm, meanwhile, by improving calculating formula of similarity and the prediction of disappearance scoring being filled and also effectively raised the precision that algorithm is recommended; In addition, this method is compared with traditional collaborative filtering, has better performance in mean absolute error.

Accompanying drawing explanation

In order to make object of the present invention, technical scheme and beneficial effect clearer, the invention provides following accompanying drawing and describe:

Fig. 1 is the macro flow chart of the method for the invention.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described in detail.

Fig. 1 is the macro flow chart of the method for the invention, as shown in the figure, collaborative filtering method based on community discovery of the present invention comprises following four steps: 1) user-project network is converted to adjacency matrix form, if user marks to project, so corresponding to the respective element of matrix with regard to existence value, i.e. score data; 2) structuring user's-user matrix, the element in user-user matrix is user and user's similarity, the computing method of similarity adopt the novel similarity calculating method based on Pearson; 3) based on user-user matrix, by community discovery method, find community structure, and the excalation scoring in user-project matrix is predicted to filling; 4) calculate between targeted customer and community and the pass of the similarity between user and user series structure arest neighbors Candidate Set, and complete recommendation.

In the present embodiment, specifically, in step 2) in, utilize user-project network structuring user's-user network, element in the corresponding user-user of user-user network adjacency matrix is the similarity between user, by new calculating formula of similarity called after Pearsimilarity, its concrete computing method are as follows:

${\mathrm{sim}}_{\mathrm{senti}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})=(1-\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{i}t}-r_{u_{j}t})}^2/\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{C_{\max }}^2})$

${\mathrm{sim}}_{\mathrm{CF}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})=\frac{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})(r_{u_{j}t}-\stackrel{\‾}{r_{u_j}})}{\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})}^2}\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{j}t}-\stackrel{\‾}{r_{u_j}})}^2}}$

$\mathrm{Pearsimilarity}(\stackrel{\→}{u_i},{\stackrel{\→}{u}}_j)=(1-\mathrm{\α})\×{\mathrm{sim}}_{\mathrm{CF}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})+\mathrm{\α}\×{\mathrm{sim}}_{\mathrm{senti}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})$

Wherein, what represent is the scoring vector of user i and user j, I _uiwith I _ujrepresent respectively the set of user i and user j scoring item, with represent user i and the scoring situation of user j to project t, α, for controlling parameter, controls the requirement of similarity formula to data susceptibility.

In step 3) in, the disappearance scoring in user-project matrix is predicted to filling, when predicting filling, consider the factor aspect two: the relevance of marking between commodity, the impact of community member on project forecast scoring; Concrete score in predicting formula is as follows:

$\mathrm{user}\_\mathrm{rating}\left(x\right)=\stackrel{\‾}{r_u}+\underset{m\∈C_u\left(x\right)}{\mathrm{\Σ}}\frac{r_{\mathrm{mx}}-\stackrel{\‾}{r_m}}{\left|C_u\left(x\right)\right|}$

$\mathrm{Item}\_\mathrm{rating}\left(x\right)=\frac{\underset{y\∈\mathrm{Neigh}\left(I_x\right)}{\mathrm{\Σ}}{r}_y\×{\mathrm{sim}}_{\mathrm{item}}(x,y)}{\underset{y\∈\mathrm{Neigh}\left(I_x\right)}{\mathrm{\Σ}}{\mathrm{sim}}_{\mathrm{item}}(x,y)}$

rc _ux＝(1-β)×user_rating(x)+β×Item_rating(x)

the mean value that represents user u scoring, C _u(x) represent the set that the affiliated community of user u forms the user of project x scoring, | C _u(x) | represent the number of user in set, r _mxrepresent the scoring of user to project x in the affiliated community of user u, the average of its scoring, r _ythe score value that represents user u commodity y, Neigh (I _x) represent and the set of project x similar terms, wherein β is Controlling parameters.

In step 4) in, according to the similarity between targeted customer and intercommunal similarity and targeted customer and other users, select arest neighbors Candidate Set, wherein targeted customer and intercommunal calculating formula of similarity are as follows:

$\mathrm{sim}\_\mathrm{corr}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=\frac{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})(r_{c_{j}t}-\stackrel{\‾}{r_{c_j}})}{\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})}^2}\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{c_{j}t}-\stackrel{\‾}{r_{c_j}})}^2}}$

$\mathrm{sim}\_\mathrm{dsenti}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=1-\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{u_{i}t}-r_{c_{j}t})}^2/\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{C_{\max }}^2}$

$\mathrm{sim}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=(1-\mathrm{\α})\×\mathrm{sim}\_\mathrm{corr}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})+\mathrm{\α}\×\mathrm{sim}\_\mathrm{dsenti}({\stackrel{\→}{u}}_i,\stackrel{\→}{C_j})$

Suppose that algorithm is divided into (C by user-user social networks ₁, C ₂, C ₃c _ic _t), t is community's number, although targeted customer belong to some communities this not represent that this targeted customer and other communities have no associated, so first calculate targeted customer and intercommunal similarity; The C of community _tmean value to commodity scoring is designated as centroid vector wherein represent the C of community _tin the average score of user to project j, represent respectively user i and the C of community _jscoring vector, the project set of evaluating for user ui, represent all users institute scoring item destination aggregation (mda) in the Cj of community, represent user u _iscoring to project t, represent user u _ithe average of scoring, represent the C of community _jin the mean value of user to project t scoring, and for centroid vector in the average of each component, α is data sensitive degree Controlling parameters.

According to step 4) selected arest neighbors Candidate Set, calculate the scoring of targeted customer to specific project, concrete score in predicting formula is as follows:

$r_u\left(x\right)=\stackrel{\‾}{r_u}+\frac{\underset{u^{\′}\∈\mathrm{Neigh}\left(u\right)}{\mathrm{\Σ}}[\mathrm{sim}(u,u^{\′})\×(r_{u^{\′}x}-\stackrel{\‾}{r_{u^{\′}}})]}{\underset{u^{\′}\∈\mathrm{Neigh}\left(u\right)}{\mathrm{\Σ}}\left|\mathrm{sim}(u,u^{\′})\right|}$

K the set that arest neighbors user forms of Neigh (u), sim (u, u') represents the similarity of user u and user u'; If there is score value to x in user u', so r _u'xrepresent the scoring of user u' to x, if user u' does not exist score value, so r to x _u'x=rc _ux, rc _uxrepresent prediction score value.

Finally explanation is, above preferred embodiment is only unrestricted in order to technical scheme of the present invention to be described, although the present invention is described in detail by above preferred embodiment, but those skilled in the art are to be understood that, can to it, make various changes in the form and details, and not depart from the claims in the present invention book limited range.

Claims (5)

1. the collaborative filtering method based on community discovery, is characterized in that: comprise the following steps:

1) user-project network is converted to adjacency matrix form, if user marks to project, so corresponding to the respective element of matrix with regard to existence value, i.e. score data;

2) structuring user's-user matrix, the element in user-user matrix is user and user's similarity, the computing method of similarity adopt the novel similarity calculating method based on Pearson;

3) based on user-user matrix, by community discovery method, find community structure, and the excalation scoring in user-project matrix is predicted to filling;

4) calculate between targeted customer and community and the pass of the similarity between user and user series structure arest neighbors Candidate Set, and complete recommendation.

2. a kind of collaborative filtering method based on community discovery according to claim 1, it is characterized in that: in step 2) in, utilize user-project network structuring user's-user network, element in the corresponding user-user of user-user network adjacency matrix is the similarity between user, by new calculating formula of similarity called after Pearsimilarity, its concrete computing method are as follows:

${\mathrm{sim}}_{\mathrm{senti}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})=(1-\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{i}t}-r_{u_{j}t})}^2/\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{C_{\max }}^2})$

${\mathrm{sim}}_{\mathrm{CF}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})=\frac{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})(r_{u_{j}t}-\stackrel{\‾}{r_{u_j}})}{\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})}^2}\sqrt{\underset{t\∈I_{\mathrm{ui}}\∩I_{\mathrm{uj}}}{\mathrm{\Σ}}{(r_{u_{j}t}-\stackrel{\‾}{r_{u_j}})}^2}}$

$\mathrm{Pearsimilarity}(\stackrel{\→}{u_i},{\stackrel{\→}{u}}_j)=(1-\mathrm{\α})\×{\mathrm{sim}}_{\mathrm{CF}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})+\mathrm{\α}\×{\mathrm{sim}}_{\mathrm{senti}}(\stackrel{\→}{u_i},\stackrel{\→}{u_j})$

Wherein, what represent is the scoring vector of user i and user j, I _uiwith I _ujrepresent respectively the set of user i and user j scoring item, with represent user i and the scoring situation of user j to project t, α, for controlling parameter, controls the requirement of similarity formula to data susceptibility.

3. a kind of collaborative filtering method based on community discovery according to claim 1, it is characterized in that: in step 3) in, disappearance scoring in user-project matrix is predicted to filling, when predicting filling, consider the factor aspect two: the relevance of marking between commodity, the impact of community member on project forecast scoring; Concrete score in predicting formula is as follows:

$\mathrm{user}\_\mathrm{rating}\left(x\right)=\stackrel{\‾}{r_u}+\underset{m\∈C_u\left(x\right)}{\mathrm{\Σ}}\frac{r_{\mathrm{mx}}-\stackrel{\‾}{r_m}}{\left|C_u\left(x\right)\right|}$

$\mathrm{Item}\_\mathrm{rating}\left(x\right)=\frac{\underset{y\∈\mathrm{Neigh}\left(I_x\right)}{\mathrm{\Σ}}{r}_y\×{\mathrm{sim}}_{\mathrm{item}}(x,y)}{\underset{y\∈\mathrm{Neigh}\left(I_x\right)}{\mathrm{\Σ}}{\mathrm{sim}}_{\mathrm{item}}(x,y)}$

rc _ux＝(1-β)×user_rating(x)+β×Item_rating(x)

the mean value that represents user u scoring, C _u(x) represent the set that the affiliated community of user u forms the user of project x scoring, | C _u(x) | represent the number of user in set, r _mxrepresent the scoring of user to project x in the affiliated community of user u, the average of its scoring, r _ythe score value that represents user u commodity y, Neigh (I _x) represent and the set of project x similar terms, wherein β is Controlling parameters.

4. a kind of collaborative filtering method based on community discovery according to claim 1, it is characterized in that: in step 4) in, according to the similarity between targeted customer and intercommunal similarity and targeted customer and other users, select arest neighbors Candidate Set, wherein targeted customer and intercommunal calculating formula of similarity are as follows:

$\mathrm{sim}\_\mathrm{corr}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=\frac{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})(r_{c_{j}t}-\stackrel{\‾}{r_{c_j}})}{\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{u_{i}t}-\stackrel{\‾}{r_{u_i}})}^2}\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{c_{j}t}-\stackrel{\‾}{r_{c_j}})}^2}}$

$\mathrm{sim}\_\mathrm{dsenti}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=1-\sqrt{\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{(r_{u_{i}t}-r_{c_{j}t})}^2/\underset{t\∈I_{u_i}\∩I_{C_j}}{\mathrm{\Σ}}{C_{\max }}^2}$

$\mathrm{sim}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})=(1-\mathrm{\α})\×\mathrm{sim}\_\mathrm{corr}(\stackrel{\→}{u_i},\stackrel{\→}{C_j})+\mathrm{\α}\×\mathrm{sim}\_\mathrm{dsenti}({\stackrel{\→}{u}}_i,\stackrel{\→}{C_j})$

Suppose that algorithm is divided into (C by user-user social networks ₁, C ₂, C ₃c _ic _t), t is community's number, although targeted customer belong to some communities this not represent that this targeted customer and other communities have no associated, so first calculate targeted customer and intercommunal similarity; The C of community _tmean value to commodity scoring is designated as centroid vector wherein represent the C of community _tin the average score of user to project j, represent respectively user i and the C of community _jscoring vector, for user u _ithe project set of evaluating, represent the C of community _jin all users institute scoring item destination aggregation (mda), represent user u _iscoring to project t, represent user u _ithe average of scoring, represent the C of community _jin the mean value of user to project t scoring, and for centroid vector in the average of each component, α is data sensitive degree Controlling parameters.

5. a kind of collaborative filtering method based on community discovery according to claim 4, is characterized in that: according to step 4) selected arest neighbors Candidate Set, calculate the scoring of targeted customer to specific project, concrete score in predicting formula is as follows:

$r_u\left(x\right)=\stackrel{\‾}{r_u}+\frac{\underset{u^{\′}\∈\mathrm{Neigh}\left(u\right)}{\mathrm{\Σ}}[\mathrm{sim}(u,u^{\′})\×(r_{u^{\′}x}-\stackrel{\‾}{r_{u^{\′}}})]}{\underset{u^{\′}\∈\mathrm{Neigh}\left(u\right)}{\mathrm{\Σ}}\left|\mathrm{sim}(u,u^{\′})\right|}$

K the set that arest neighbors user forms of Neigh (u), sim (u, u') represents the similarity of user u and user u'; If there is score value to x in user u', so r _u'xrepresent the scoring of user u' to x, if user u' does not exist score value, so r to x _u'x=rc _ux, rc _uxrepresent prediction score value.