CN103995823A - Information recommending method based on social network - Google Patents

Abstract

The invention discloses an information recommending method based on a social network. The information recommending method includes the following steps that first, trust degree and similarity between users are calculated, and a user relation matrix is constructed through weighted values; second, the users are clustered through a community discovering algorithm, and then a closest neighbor set of the users is formed; third, scores are predicted, and a recommending list is generated. The information recommending method based on the social network can achieve the following advantages that first, the cold start problem is solved: trust degree is introduced into the method, if enough neighbors cannot be obtained according to the common grading articles in the recommending process, trustable friends can serve as the start point of prediction, and thus the cold start problem can be relieved, and user coverage can be improved; real time performance is improved: community division is performed on the user network through the community discovering algorithm commonly used in social network analysis, in other words, same user interests are clustered, and thus the time for finding the neighbor set of the users is greatly shortened, and the real time performance of the information recommending response is improved.

Description

Information recommendation method based on social network

Technical Field

The invention relates to the technical field of network information, in particular to an information recommendation method based on a social network.

Background

The rapid development of the internet and the continuously increasing information resources lead to the rapid increase of the information index, and the information service field faces the problem that the information resources are rich, but the information with useful value is difficult to obtain, thereby bringing great information burden to people. On one hand, the phenomenon of information overload (information overload) caused by a large amount of data resources on the network occurs; on the other hand, the user cannot acquire the information resources required by the user. A recommendation system (recommendation systems) is an important method for serving in an 'information push' mode, is a main means for solving the problem of information overload, can actively push information which is possibly needed but difficult to obtain to a user on the basis of analyzing and predicting the user requirement by taking the user as a center, and recommends information resources with more utilization value for the user according to behavior characteristics of the user in different environment occasions.

With the development of the internet there is also a rapid expansion of social networks. Social networks connect people who have the same hobbies, even people who are not acquainted with each other, through the internet, thereby forming a group with a certain characteristic. The social network is a platform which can communicate and exchange with each other and participate in interaction, and is expanded from research departments, schools, governments and business application platforms into a tool for human social communication. Because the internet is virtual, people hide their true identities in the network in many ways, which not only brings a lot of false information, but also makes the trust level between people lower and lower, and communication becomes more difficult. The social network adopts real information registration, so that the identity authenticity and the behavior credibility of network users are enhanced, the information safety and the reliability, the regionality and the real-time performance of user interaction in the system are greatly guaranteed, people can communicate with other people more securely and more easily, and brand new user experience is brought. Through social networks, they may actively publish their features and preferences, actively provide and annotate various resources (such as pictures, videos) or share their knowledge. For example, a user may share a book through a bean, perform social networking and share photos through Facebook, send a microblog through Twitter, publish photos through Flickr, upload videos through YouTube, and so on. Increasingly popular social networks silently change people's lifestyle and orientation of value.

Currently, frequently used recommendation methods include the following:

1) the recommendation method based on the association rule generates the association rule according to the transaction data of the user, and proposes suggestions in combination with the current purchasing behavior of the user, wherein shopping cart analysis is the most typical application of the association rule. The recommendation method based on the association rule has stronger universality and can be applied to various fields, but the extraction of the association rule is difficult, the consumed time is long, and the system is difficult to manage as the number of the association rules is increased continuously.

2) The recommendation method based on the content mainly focuses on the content analysis of information resource items and the construction of a user preference model, and the recommendation function is realized by comparing the similarity of resources and user preferences. Although the recommendation technology based on the content has an intuitive result, simple calculation, quick response time and good interpretability, and can solve the problems of cold start and data sparsity. However, there are still some limitations: the content of the items that can be analyzed is limited, only information that can be represented by a series of feature sets, and multimedia information such as sound, pictures, video, and the like cannot be effectively processed; the user can receive items similar to past favorite recommendations, but cannot find new interested commodities for the user, and the recommendation content is single; quality, style or viewpoint cannot be processed.

3) The collaborative filtering recommendation method is the most successful technology in the current recommendation information system, the basic idea of collaborative filtering is to utilize the similarity between users or items to recommend or predict, the method finds out a group of users with the same preference, and then analyzes the common preference of the users to recommend the target users. The collaborative filtering algorithm has the advantages that the collaborative filtering algorithm does not pay attention to the content of the item, resources are recommended mainly according to the similarity of users or items, and the system can reliably recommend the item only by obtaining enough item evaluation. However, the shortcomings of the collaborative filtering algorithm are also very obvious, namely the problem of cold start, the problem of data sparsity, the problem of expandability and the like.

The development of the social network provides a good channel for personalized recommendation, the invention organically combines trust measurement among users, a social network mode and a personalized recommendation technology, and provides an information recommendation method based on the social network to construct a recommendation system with high efficiency and high accuracy.

Disclosure of Invention

In order to solve the technical problem, the invention provides an information recommendation method based on a social network, which adopts the following technical scheme:

an information recommendation method based on a social network comprises the following steps:

step 1: calculating the trust and similarity between users, and constructing a user relationship matrix by using a weighted value;

step 2: clustering the users by using a community discovery algorithm to form a user nearest neighbor set;

and step 3: the prediction scores and generates a recommendation list.

Constructing a user-score matrix R_m×n：

m represents the number of users, n represents the number of items, r_ijRepresenting the rating of item j by user i.

In the populated user-rating matrix R_m×nOn the basis, the similarity between users is calculated by using Pearson correlation, and a user-user similarity matrix S is constructed:

s_uvrepresents the degree of similarity between user u and user v, and s_uv∈[0,1]。

Calculating the trust degree, and then calculating a weighted value combining the similarity and the trust:

h(i,j)＝θ×trust(i,j)+(1-θ)×sim(i,j)；

θ represents a weighting parameter, trust (i, j) represents the direct confidence of users i and j, sim (i, j) represents the similarity, and h (i, j) represents a weighted value of the two.

And (3) constructing a user relationship matrix H by using the weighted values:

the direct trust is the combination of the interactive trust, the proportion of common friends among users and the evaluation capability of the users, and the calculation formula is as follows:

<math> <mrow> <mi>trust</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mn>2</mn> </mfrac> <mrow> <mo>(</mo> <mi>fri</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>+</mo> <mi>fre</mi> <mo>)</mo> </mrow> <mo>&times;</mo> <mi>comm</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>;</mo> </mrow> </math>

trust (i, j) represents the direct trust of users i and j, fri (i, j) represents the proportion of common friends between the users, fre represents the evaluation capability of the users, and comm (i, j) represents the interactive trust between the users.

The calculation formula of the interaction trust degree is as follows:

<math> <mrow> <mi>comm</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <msub> <mi>w</mi> <mrow> <mi>i</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <mrow> <mi>&Sigma;</mi> <msub> <mi>w</mi> <mrow> <mi>i</mi> <mrow> <mo>(</mo> <mi>out</mi> <mo>)</mo> </mrow> </mrow> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math>

wherein, w_i,jRepresenting the number of messages sent by user i to user j, Σ w_i(out)The total number of messages sent to surrounding users on behalf of the user i;

the calculation formula of the proportion of common friends among users and the evaluation capability of the users is as follows:

<math> <mrow> <mi>fri</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <msub> <mi>n</mi> <mi>i</mi> </msub> <mo>&cap;</mo> <msub> <mi>n</mi> <mi>j</mi> </msub> </mrow> <msub> <mi>n</mi> <mi>i</mi> </msub> </mfrac> <mo>;</mo> </mrow> </math>

<math> <mrow> <mi>fre</mi> <mo>=</mo> <mfrac> <msub> <mi>k</mi> <mi>i</mi> </msub> <mrow> <mi>&Sigma;</mi> <msub> <mi>k</mi> <mi>i</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math>

where fri (i, j) represents the proportion of common friends among users, fre represents the evaluation ability of users, and n_iAnd n_jRepresenting the numbers of friends of users i and j, n_i∩n_jNumber of common friends on their behalf; k is a radical of_iRepresenting the number of evaluations of the recommended user on the i-type merchandise, Σ k_iRepresenting the number of evaluations of all the commodity categories by the user.

The process of the community discovery method is as follows:

step 1: calculating the degree of each user in the network (the number of edges associated with the vertex), and selecting the user i with the highest degree from the degrees as an initial community C_iAnd initializing the modularity Q to be 0;

step 2: find all and community C_iConnected users and put them into the adjacent user set N;

step 3: calculating each user j in the user set N to the community C_iAnd adding the user having the largest contribution degree to the community C_iPerforming the following steps;

step 4: compute community C_iModularity Q'. If Q' > Q, add user j to community C_iSuccessfully marking the user j, updating the modularity Q ═ Q', and returning to Step2 to continue execution; otherwise, go to Step 5;

step 5: the modularity Q has reached a maximum value, i.e. the current community C_iAchieving the optimal result of the division;

step 6: if there are no unmarked users and all communities in the network have been detected, the process ends; otherwise, selecting the user with the maximum degree from the users without marks as a new initial community C_iReturning to step2 to continue execution.

Contribution q of user to community

$q=\frac{L_{\mathrm{in}}}{l_{\mathrm{in}}+L_{\mathrm{out}}};$

L_in: representing the number of connected edges in the community in the unauthorized network; representing an insider of a community in a rights networkThe sum of the weights on the edges.

L_out: representing the number of external connection edges connected with the community in the unauthorized network; representing the sum of the weights on all the external edges connected to the community in a weighted network.

The larger the contribution q of the user to the community is, the more closely the user and the community are connected.

Definition 2: modularity Q

The modularity Q is an important index for judging whether the current community reaches the optimal degree or not in community division, the larger the modularity Q is, the better the community division effect is, the less the connectivity between the community and the community is, thereby realizing the high cohesion inside the module and the low coupling outside the module.

If the network is an unweighted network, the expression of the modularity Q is as follows:

<math> <mrow> <mi>Q</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <mi>m</mi> </mrow> </mfrac> <munder> <mi>&Sigma;</mi> <mi>ij</mi> </munder> <mrow> <mo>(</mo> <msub> <mi>A</mi> <mi>ij</mi> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>k</mi> <mi>i</mi> </msub> <msub> <mi>k</mi> <mi>j</mi> </msub> </mrow> <mrow> <mn>2</mn> <mi>m</mi> </mrow> </mfrac> <mo>)</mo> </mrow> <mi>&delta;</mi> <mrow> <mo>(</mo> <msub> <mi>C</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>C</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mo>;</mo> </mrow> </math>

wherein m is the total number of edges of the network; k is a radical of_iAnd k_jRespectively representing the number of connecting edges of the user i and the user j; a. the_ijRepresenting a network adjacency matrix, A when a user i and a user j of a random network connection are connected_ijWhen user i and user j are not connected, a is 1_ij＝0；δ(Cⁱ,C_j) For the kronecker function, if user i and user j belong to the same community, δ (C)_i,C_j) 1, if not in the same community, δ (C)_i,C_j)＝0。

In a weighted network, the modularity Q may be defined as follows:

<math> <mrow> <mi>Q</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <mi>W</mi> </mrow> </mfrac> <munder> <mi>&Sigma;</mi> <mi>ij</mi> </munder> <mrow> <mo>(</mo> <msub> <mi>A</mi> <mi>ij</mi> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>w</mi> <mi>i</mi> </msub> <msub> <mi>w</mi> <mi>j</mi> </msub> </mrow> <mrow> <mn>2</mn> <mi>W</mi> </mrow> </mfrac> <mo>)</mo> </mrow> <mi>&delta;</mi> <mrow> <mo>(</mo> <msub> <mi>C</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>C</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mo>;</mo> </mrow> </math>

wherein W represents the sum of the weights of all edges in the network, W_iAnd w_jRepresenting the sum of the weights of the edges connected to user i and user j, respectively.

And predicting the scoring of the candidate items by the target user according to the scoring information of the candidate items by the N nearest neighbors of the target user, selecting the first items with the highest predicted scores as recommendation results, and actively pushing the recommendation results to the target user, namely generating top-N information resource recommendation.

<math> <mrow> <msub> <mi>P</mi> <mrow> <mi>u</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mover> <msub> <mi>R</mi> <mi>u</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>+</mo> <mfrac> <mrow> <munder> <mi>&Sigma;</mi> <mrow> <mi>v</mi> <mo>&Element;</mo> <mi>N</mi> </mrow> </munder> <mi>trust</mi> <mrow> <mo>(</mo> <mi>u</mi> <mo>,</mo> <mi>v</mi> <mo>)</mo> </mrow> <mo>&times;</mo> <mrow> <mo>(</mo> <msub> <mi>R</mi> <mrow> <mi>v</mi> <mo>,</mo> <mi>i</mi> </mrow> </msub> <mo>-</mo> <mover> <msub> <mi>R</mi> <mi>v</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> </mrow> <mrow> <munder> <mi>&Sigma;</mi> <mrow> <mi>v</mi> <mo>&Element;</mo> <mi>N</mi> </mrow> </munder> <mi>trust</mi> <mrow> <mo>(</mo> <mi>u</mi> <mo>,</mo> <mi>v</mi> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>;</mo> </mrow> </math>

Wherein, P_u,iRepresents the predicted rating of item i by user u,andrespectively representing the average scores of the user u and the neighbor user v on the project; r_v,iRepresents the rating of the user v on the item i, trust (u, v) represents the trust degree of the user u on the neighbor user v, and N represents the neighbor user candidate set of the user u.

The direct trust (u, v) may also be the transitive trust_L(A,B)Instead, it is calculated as follows:

trust_L(A,B)＝trust(A,X₁)×trust(X₁,X₂)×…×trust(X_n,B)；

wherein, X_iRepresenting users between users A and B on path L, L (A, B) representing an existing trust path between user A and user B, if there are multiple trust paths L (L) between user A and user B in the trust network₁,L₂,…,L_n) (n is more than or equal to 2), selecting the shortest path in the path L, if k shortest paths exist, calculating the formula as follows:

<math> <mrow> <msub> <mi>trust</mi> <mrow> <mi>L</mi> <mrow> <mo>(</mo> <mi>A</mi> <mo>,</mo> <mi>B</mi> <mo>)</mo> </mrow> </mrow> </msub> <mo>=</mo> <mfrac> <mrow> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>k</mi> </munderover> <msub> <mi>trust</mi> <msub> <mi>L</mi> <mi>i</mi> </msub> </msub> </mrow> <mi>k</mi> </mfrac> <msub> <mi>L</mi> <mi>i</mi> </msub> <mo>&Element;</mo> <mo>{</mo> <mi>min</mi> <mrow> <mo>(</mo> <msub> <mi>L</mi> <mn>1</mn> </msub> <mo>,</mo> <msub> <mi>L</mi> <mn>2</mn> </msub> <mo>,</mo> <mo>&CenterDot;</mo> <mo>&CenterDot;</mo> <mo>&CenterDot;</mo> <mo>,</mo> <msub> <mi>L</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> <mo>}</mo> <mo>;</mo> </mrow> </math>

then, the calculation is continued according to the sequence of the path L from the short to the long, if the calculation result of a certain path jCalculate path j as one of the shortest paths and recalculate trust_L(A,B)And when all paths are calculated, a final indirect trust result is obtained.

The information recommendation method based on the social network, provided by the invention, can achieve the following beneficial effects:

(1) the cold start problem is solved. The cold start problem means that when a new user joins the recommendation system, since no record is made on the historical behavior or score of the new user, effective information recommendation cannot be performed on the new user. The method introduces the trust degree, and if enough neighbors cannot be obtained according to the common scoring articles during recommendation, the trusted friends can be used as the starting points of prediction, so that the cold start problem can be relieved, and the user coverage degree can be improved.

(2) The real-time performance is improved. In the method, the community discovery algorithm commonly used in social network analysis is adopted to perform community division on the user network, namely the same user interest clustering is performed, so that the time is greatly shortened when a user neighbor set is searched, and the response real-time performance of information recommendation is improved.

Drawings

FIG. 1 is a flow chart of a social network based recommendation method.

FIG. 2 is a diagram of interaction between users.

FIG. 3 is a graph of computation of transitive belief for multiple paths.

FIG. 4 is a flow diagram of community discovery.

Detailed Description

The invention provides an information recommendation method based on a social network. Firstly, calculating the trust and similarity between users, and constructing a user relationship matrix by using a weighted value; secondly, clustering the users by using a community discovery algorithm to form a user nearest neighbor set; finally, the scores are predicted and a recommendation list is generated.

As shown in fig. 1, collecting related information of users, items, scores, social networks and the like, calculating trust and similarity between the users according to the information, and constructing a user relationship matrix by using weighted values; secondly, dividing the users by using a community discovery algorithm to form a user nearest neighbor set; finally, the scores are predicted and a recommendation list is generated.

1. Direct degree of trust

Direct trust is a quantification of the degree of trust one user has on another user. The trust is introduced into the personalized recommendation system, and the problem of quantification of the trust relationship is solved to enable the trust relationship to become data which can be calculated or subjected to related operations. Combining multiple properties of trust and the relevant characteristics of the actual social network, the computational problem of direct trust is considered herein from the following aspects.

1) Degree of mutual trust

The interaction degree refers to the interaction degree among users in the social network relationship, such as the interaction times within the period time of the users, such as conversation, sending short messages, mails, microblog concerns, SNS messages and the like. An interaction graph among users is shown in fig. 2, wherein each node in the graph represents a user in a social network, edges among the users represent interaction senders to receivers by directional arrows, and weighted values on the edges represent information sum of various interaction types. Through an interaction graph among users, the interaction trust degree among the users can be defined as follows:

<math> <mrow> <mi>comm</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <msub> <mi>w</mi> <mrow> <mi>i</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <mrow> <mi>&Sigma;</mi> <msub> <mi>w</mi> <mrow> <mi>i</mi> <mrow> <mo>(</mo> <mi>out</mi> <mo>)</mo> </mrow> </mrow> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math>

wherein, w_i,jRepresenting the number of messages sent by user i to user j, Σ w_i(out)Representing the total number of messages sent by user i to the surrounding users.

2) Recommending confidence

The recommendation trust degree comprises the proportion of common friends among users and the evaluation capability of the users. The user always trusts the user who bought or used the commodity more, and if the background information of the user, such as profession, interest and the like, has a certain relation with the commodity, the background information of the user enables the user to have certain evaluation capacity for the commodity, and the capacity can also be a reference factor of the credibility of the user for recommending the commodity. Meanwhile, the number of common friends between the users can also reflect the relationship between the users and the friends. fri (i, j) represents the proportion of common friends among users, and fre represents the calculation of the evaluation capability of the users as follows:

<math> <mrow> <mi>fri</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <msub> <mi>n</mi> <mi>i</mi> </msub> <mo>&cap;</mo> <msub> <mi>n</mi> <mi>j</mi> </msub> </mrow> <msub> <mi>n</mi> <mi>i</mi> </msub> </mfrac> <mo>;</mo> </mrow> </math>

<math> <mrow> <mi>fre</mi> <mo>=</mo> <mfrac> <msub> <mi>k</mi> <mi>i</mi> </msub> <mrow> <mi>&Sigma;</mi> <msub> <mi>k</mi> <mi>i</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math>

wherein n is_iAnd n_jRepresenting the numbers of friends of users i and j, n_i∩n_jNumber of common friends on their behalf; k is a radical of_iRepresenting the number of evaluations of the recommended user on the i-type merchandise, Σ k_iRepresenting the number of evaluations of all the commodity categories by the user.

Combining the two trust degrees to obtain the direct trust degree trust (i, j):

<math> <mrow> <mi>trust</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mn>2</mn> </mfrac> <mrow> <mo>(</mo> <mi>fri</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>+</mo> <mi>fre</mi> <mo>)</mo> </mrow> <mo>&times;</mo> <mi>comm</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>;</mo> </mrow> </math>

2. transitive confidence computation

The principle that direct trust is preferred to indirect trust is followed in the calculation of the degree of trust, and L (A, B) represents the existing trust path between the user A and the user B. If a unique trust path L (A, X) exists between user A and user B in the trust network₁,X₂,…,X_nB), then the transitive confidence between users a and B is the product of all direct trusts on the trust path L, and the expression is:

trust_L(A,B)＝trust(A,X₁)×trust(X₁,X₂)×…×trust(X_n,B)；

wherein, X_iRepresenting the users between users a and B on path L.

If there are multiple trust paths L (L) between user A and user B in the trust network₁,L₂,…,L_n) (n is more than or equal to 2), selecting the shortest path in the path L, if k shortest paths exist, calculating the formula as follows:

<math> <mrow> <msub> <mi>trust</mi> <mrow> <mi>L</mi> <mrow> <mo>(</mo> <mi>A</mi> <mo>,</mo> <mi>B</mi> <mo>)</mo> </mrow> </mrow> </msub> <mo>=</mo> <mfrac> <mrow> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>k</mi> </munderover> <msub> <mi>trust</mi> <msub> <mi>L</mi> <mi>i</mi> </msub> </msub> </mrow> <mi>k</mi> </mfrac> <msub> <mi>L</mi> <mi>i</mi> </msub> <mo>&Element;</mo> <mo>{</mo> <mi>min</mi> <mrow> <mo>(</mo> <msub> <mi>L</mi> <mn>1</mn> </msub> <mo>,</mo> <msub> <mi>L</mi> <mn>2</mn> </msub> <mo>,</mo> <mo>&CenterDot;</mo> <mo>&CenterDot;</mo> <mo>&CenterDot;</mo> <mo>,</mo> <msub> <mi>L</mi> <mi>n</mi> </msub> <mo>)</mo> </mrow> <mo>}</mo> <mo>;</mo> </mrow> </math>

then, the calculation is continued according to the sequence of the path L from the short to the long, if the calculation result of a certain path jCalculate path j as one of the shortest paths and recalculate trust_L(A,B)And when all paths (if the links are too long and can be ignored) are calculated, a final indirect trust result is obtained. The calculation method effectively avoids misjudgment caused by the fact that the trust degree of the shortest trust path is too low.

As shown in FIG. 3, there are multiple trusted paths from user A to user B, and the shortest path L (A, X) should be calculated first₃B) and L (A, X)₄Trust of B)_L(A,B)The value of the one or more of,but the other path L (A, X)₁,X₂Degree of transitive Trust of B) <math> <mrow> <msub> <mi>trust</mi> <mrow> <mi>L</mi> <mrow> <mo>(</mo> <mi>A</mi> <mo>,</mo> <msub> <mi>X</mi> <mn>1</mn> </msub> <mo>,</mo> <msub> <mi>X</mi> <mn>2</mn> </msub> <mo>,</mo> <mi>B</mi> <mo>)</mo> </mrow> </mrow> </msub> <mo>=</mo> <mn>0.8</mn> <mo>&times;</mo> <mn>0.8</mn> <mo>&times;</mo> <mn>0.9</mn> <mo>=</mo> <mn>0.576</mn> <mo>></mo> <msub> <mi>trust</mi> <mrow> <mi>L</mi> <mrow> <mo>(</mo> <mtext>A,B</mtext> <mo>)</mo> </mrow> </mrow> </msub> <mo>=</mo> <mn>0.48</mn> <mo>,</mo> </mrow> </math> So path L (A, X)₁,X₂And B) should also be added to the shortest path set to obtain the final transfer trust degree of

3. Community discovery

In a real network, there are often some users that are very closely related to other users (i.e., the point in the network where the user is very large, we refer to as the "central user"). The central user is taken as a starting point of an initial community, then, the adjacent users with the largest contribution degree to the community are continuously added into the community (if a plurality of users have larger contribution degrees to the community, the users are added into the community), and a stable community is formed when the global contribution degree reaches the maximum.

For a clearer understanding of the clustering process, the degree of contribution and the degree of modularity to the community are defined as follows:

definition 1: contribution q of user to community

$q=\frac{L_{\mathrm{in}}}{l_{\mathrm{in}}+L_{\mathrm{out}}};$

L_in: representing the number of connected edges in the community in the unauthorized network; representing the weight sum of all edges inside a community in a weighted networkAnd (c).

L_out: representing the number of external connection edges connected with the community in the unauthorized network; representing the sum of the weights on all the external edges connected to the community in a weighted network.

The non-weighted network means that there is no weight value on the edge, and the weighted network means that there is a weight value on the edge.

The larger the contribution q of the user to the community is, the more closely the user and the community are connected.

Definition 2: modularity Q

The modularity Q is an important index for judging whether the current community reaches the optimal degree or not in community division, the larger the modularity Q is, the better the community division effect is, the less the connectivity between the community and the community is, thereby realizing the high cohesion inside the module and the low coupling outside the module.

If the network is an unweighted network, the expression of the modularity Q is as follows:

<math> <mrow> <mi>Q</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <mi>m</mi> </mrow> </mfrac> <munder> <mi>&Sigma;</mi> <mi>ij</mi> </munder> <mrow> <mo>(</mo> <msub> <mi>A</mi> <mi>ij</mi> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>k</mi> <mi>i</mi> </msub> <msub> <mi>k</mi> <mi>j</mi> </msub> </mrow> <mrow> <mn>2</mn> <mi>m</mi> </mrow> </mfrac> <mo>)</mo> </mrow> <mi>&delta;</mi> <mrow> <mo>(</mo> <msub> <mi>C</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>C</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mo>;</mo> </mrow> </math>

wherein m is the total number of edges of the network; k is a radical of_iAnd k_jRespectively representing the number of connecting edges of the user i and the user j; a. the_ijRepresenting a network adjacency matrix, A when a user i and a user j of a random network connection are connected_ijWhen user i and user j are not connected, a is 1_ij＝0；δ(C_i,C_j) For the kronecker function, if user i and user j belong to the same community, δ (C)_i,C_j) 1, if not in the same community, δ (C)_i,C_j)＝0。

In a weighted network, the modularity Q may be defined as follows:

<math> <mrow> <mi>Q</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mn>2</mn> <mi>W</mi> </mrow> </mfrac> <munder> <mi>&Sigma;</mi> <mi>ij</mi> </munder> <mrow> <mo>(</mo> <msub> <mi>A</mi> <mi>ij</mi> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>w</mi> <mi>i</mi> </msub> <msub> <mi>w</mi> <mi>j</mi> </msub> </mrow> <mrow> <mn>2</mn> <mi>W</mi> </mrow> </mfrac> <mo>)</mo> </mrow> <mi>&delta;</mi> <mrow> <mo>(</mo> <msub> <mi>C</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>C</mi> <mi>j</mi> </msub> <mo>)</mo> </mrow> <mo>;</mo> </mrow> </math>

wherein W represents the sum of the weights of all edges in the network, W_iAnd w_jRepresenting the sum of the weights of the edges connected to user i and user j, respectively.

The community discovery algorithm is shown in fig. 4:

step 1: calculating the degree of each user in the network (the number of edges associated with the vertex), and selecting the user i with the highest degree from the degrees as an initial community C_iAnd initializing the modularity Q to be 0;

step 2: find all and community C_iConnected users and put them into the adjacent user set N;

step 3: calculating each user j in the user set N to the community C_iAnd adding the user having the largest contribution degree to the community C_iPerforming the following steps;

step 4: compute community C_iModularity Q'. If Q' > Q, add user j to community C_iSuccessfully marking the user j, updating the modularity Q ═ Q', and returning to Step2 to continue execution; otherwise, go to Step 5;

step 5: the modularity Q has reached a maximum value, i.e. the current community C_iAchieving the optimal result of the division;

step 6: if there are no unmarked users and all communities in the network have been detected, the process ends; otherwise, selecting the user with the maximum degree from the users without marks as a new initial community C_iReturning to step2 to continue execution.

4. Implementation of recommendation method

1) Building a user relationship matrix

Firstly, acquiring user information, project information and rating information, and representing the relationship between a user and a project in a matrix mode to form a user-rating matrix R_m×nHere, the item information refers to a product, a movie, or the like recommended to the user. Where m represents the number of users, n represents the number of items, r_ijRepresenting the rating of item j by user i. To reduce sparsity of the user-scoring matrix, the unscored r may be used_ijSet to 0 or the mean of all the users i that have been scored or the mean of the items j that have been scored.

Second, in the populated user-score matrix R_m×nOn the basis, the similarity is calculated by using Pearson correlation, wherein the Pearson correlation is a method for calculating the similarity, if the item set jointly scored by the user I and the user j is set as I_ijThen the similarity sim (i, j) measured by Pearson correlation is:

<math> <mrow> <mi>sim</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>,</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <munder> <mi>&Sigma;</mi> <msub> <mrow> <mi>c</mi> <mo>&Element;</mo> <mi>I</mi> </mrow> <mi>ij</mi> </msub> </munder> <mrow> <mo>(</mo> <msub> <mi>R</mi> <mi>ic</mi> </msub> <mo>-</mo> <mover> <msub> <mi>R</mi> <mi>c</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> <mrow> <mo>(</mo> <msub> <mi>R</mi> <mi>jc</mi> </msub> <mo>-</mo> <mover> <msub> <mi>R</mi> <mi>j</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> </mrow> <msqrt> <munder> <mi>&Sigma;</mi> <msub> <mrow> <mi>c</mi> <mo>&Element;</mo> <mi>I</mi> </mrow> <mi>ij</mi> </msub> </munder> <msup> <mrow> <mo>(</mo> <msub> <mi>R</mi> <mi>ic</mi> </msub> <mo>-</mo> <mover> <msub> <mi>R</mi> <mi>i</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>&times;</mo> <munder> <mi>&Sigma;</mi> <msub> <mrow> <mi>c</mi> <mo>&Element;</mo> <mi>I</mi> </mrow> <mi>ij</mi> </msub> </munder> <msup> <mrow> <mo>(</mo> <msub> <mi>R</mi> <mi>jc</mi> </msub> <mo>-</mo> <mover> <msub> <mi>R</mi> <mi>j</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> </mfrac> <mo>;</mo> </mrow> </math>

wherein R is_icAnd R_jcRepresents the rating of item c by both users;andrespectively representing users I and j in I_ijAverage score above.

On the basis of the above, a user-user similarity matrix S is constructed. Wherein s is_uvRepresents the degree of similarity between user u and user v, and s_uv∈[0,1]。

A user-user similarity relationship matrix S:

calculating the trust degree, and then calculating a weighted value combining the similarity and the trust:

h(i,j)＝θ×trust(i,j)+(1-θ)×sim(i,j)；

θ represents a weighting parameter, trust (i, j) represents the confidence of users i and j, sim (i, j) represents the similarity, and h (i, j) represents a weighting value of the two.

Constructing a user relationship matrix H:

2) clustering users with the same interest to form a neighbor set

Clustering the users according to the user relationship matrix by using the community discovery algorithm described above, and dividing the users into a plurality of communities { C₁,C₂,…,C_nAnd users in each community have the same interest and hobbies, and a neighbor set is formed. If a new user joins, the user can be reclassified by updating the network.

3) Predictive scoring to generate recommendations

And predicting the scoring of the candidate items by the target user according to the scoring information of the candidate items by the N nearest neighbors of the target user, selecting the first items with the highest predicted scores as recommendation results, and actively pushing the recommendation results to the target user, namely generating top-N information resource recommendation.

<math> <mrow> <msub> <mi>P</mi> <mrow> <mi>u</mi> <mo>,</mo> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mover> <msub> <mi>R</mi> <mi>u</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>+</mo> <mfrac> <mrow> <munder> <mi>&Sigma;</mi> <mrow> <mi>v</mi> <mo>&Element;</mo> <mi>N</mi> </mrow> </munder> <mi>trust</mi> <mrow> <mo>(</mo> <mi>u</mi> <mo>,</mo> <mi>v</mi> <mo>)</mo> </mrow> <mo>&times;</mo> <mrow> <mo>(</mo> <msub> <mi>R</mi> <mrow> <mi>v</mi> <mo>,</mo> <mi>i</mi> </mrow> </msub> <mo>-</mo> <mover> <msub> <mi>R</mi> <mi>v</mi> </msub> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> </mrow> <mrow> <munder> <mi>&Sigma;</mi> <mrow> <mi>v</mi> <mo>&Element;</mo> <mi>N</mi> </mrow> </munder> <mi>trust</mi> <mrow> <mo>(</mo> <mi>u</mi> <mo>,</mo> <mi>v</mi> <mo>)</mo> </mrow> </mrow> </mfrac> <mo>;</mo> </mrow> </math>

Wherein,andrespectively representing the average scores of the user u and the neighbor user v on the project; trust (u, v) represents the degree of trust of user u for neighbor user v.

The direct trust (u, v) may also be the transitive trust_L(A,B)Instead, the calculation method is as described above.

Claims (8)

1. An information recommendation method based on a social network comprises the following steps:

step 1: calculating the trust and similarity between users, and constructing a user relationship matrix by using a weighted value;

step 2: clustering the users by using a community discovery algorithm to form a user nearest neighbor set;

and step 3: the prediction scores and generates a recommendation list.

2. The information recommendation method based on social network as claimed in claim 1, wherein the user relationship matrix is constructed in step1 by the following method:

constructing a user-score matrix R_m×n：

m represents the number of users, n represents the number of items, r_ijA score for item j on behalf of user i;

in the populated user-rating matrix R_m×nOn the basis, the similarity between users is calculated by using Pearson correlation, and a user-user similarity matrix S is constructed:

s_uvrepresents the degree of similarity between user u and user v, and s_uv∈[0,1]；

Calculating the trust degree, and then calculating a weighted value combining the similarity and the trust:

h(i,j)＝θ×trust(i,j)+(1-θ)×sim(i,j)；

theta represents a weighting parameter, trust (i, j) represents the direct trust of users i and j, sim (i, j) represents similarity, and h (i, j) represents a weighted value of the two;

and (3) constructing a user relationship matrix H by using the weighted values:

3. the information recommendation method based on the social network as claimed in claim 2, wherein the direct trust is a combination of the interactive trust, the proportion of common friends among users, and the evaluation ability of the users, and the calculation formula is as follows:

trust (i, j) represents the direct trust of users i and j, fri (i, j) represents the proportion of common friends between the users, fre represents the evaluation capability of the users, and comm (i, j) represents the interactive trust between the users;

the calculation formula of the interaction trust degree is as follows:

wherein, w_i,jRepresenting the number of messages sent by user i to user j, Σ w_i(out)The total number of messages sent to surrounding users on behalf of the user i;

the calculation formula of the proportion of common friends among users and the evaluation capability of the users is as follows:

where fri (i, j) represents the proportion of common friends among users, fre represents the evaluation ability of users, and n_iAnd n_jRepresenting the numbers of friends of users i and j, n_i∩n_jNumber of common friends on their behalf; k is a radical of_iRepresenting the number of evaluations of the recommended user on the i-type merchandise, Σ k_iRepresenting the number of evaluations of all the commodity categories by the user.

4. The method of claim 1, wherein the community discovery method in step2 comprises the following steps:

step 1: calculating the degree of each user in the network (the number of edges associated with the vertex), and selecting the user i with the highest degree from the degrees as an initial community C_iAnd initializing the modularity Q to be 0;

step 2: find all and community C_iConnected users and put them into the adjacent user set N;

step 3: calculating each user j in the user set N to the community C_iAnd adding the user having the largest contribution degree to the community C_iPerforming the following steps;

step 4: compute community C_iIf Q' is greater than Q, adding user j to community C_iSuccessfully marking the user j, updating the modularity Q ═ Q', and returning to Step2 to continue execution; otherwise, go to Step 5;

step 5: the modularity Q has reached a maximum value, i.e. the current community C_iAchieving the optimal result of the division;

step 6: if there are no unmarked users and all communities in the network have been detected, the process ends; otherwise, selecting the user with the maximum degree from the users without marks as a new initial community C_iReturning to step2 to continue execution.

5. The information recommendation method based on the social network as claimed in claim 4, wherein the contribution q of the user to the community is calculated as follows:

L_in: representing the number of connected edges in the community in the unauthorized network; representing the total weight of all edges in the community in the weighted network;

L_out: representing the number of external connection edges connected with the community in the unauthorized network; representing outside connections to a community in a privileged networkThe sum of the weights of all edges is calculated;

the larger the contribution q of the user to the community is, the more closely the user and the community are connected.

6. The information recommendation method based on social network as claimed in claim 4, wherein the modularity Q is expressed as follows:

if the network is an unweighted network, the expression of the modularity Q is as follows:

wherein m is the total number of edges of the network; k is a radical of_iAnd k_jRespectively representing the number of connecting edges of the user i and the user j; a. the_ijRepresenting a network adjacency matrix, A when a user i and a user j of a random network connection are connected_ijWhen user i and user j are not connected, a is 1_ij＝0；δ(C_i,C_j) For the kronecker function, if user i and user j belong to the same community, δ (C)_i,C_j) 1, if not in the same community, δ (C)_i,C_j)＝0；

In a weighted network, the modularity Q may be defined as follows:

wherein W represents the sum of the weights of all edges in the network, W_iAnd w_jRepresenting the sum of the weights of the edges connected to user i and user j, respectively.

7. The information recommendation method based on the social network as claimed in claim 1, wherein the scoring recommendation method in step3 is as follows: predicting the scoring of the candidate items by the target user according to the scoring information of the candidate items by the N nearest neighbors of the target user, selecting the first items with the highest predicted scores as recommendation results and actively pushing the recommendation results to the target user, namely generating top-N information resource recommendation:

wherein, P_u,iRepresents the predicted rating of item i by user u,andrespectively representing the average scores of the user u and the neighbor user v on the project; r_v,iRepresents the rating of the user v on the item i, trust (u, v) represents the trust degree of the user u on the neighbor user v, and N represents the neighbor user candidate set of the user u.

8. The method of claim 7, wherein the direct trust (u, v) is also the transitive trust (Trust)_L(A,B)Instead, it is calculated as follows:

trust_L(A,B)＝trust(A,X₁)×trust(X₁,X₂)×…×trust(X_n,B)；

wherein, X_iRepresenting users between users A and B on path L, L (A, B) representing an existing trust path between user A and user B, if there are multiple trust paths L (L) between user A and user B in the trust network₁,L₂,…,L_n) (n is more than or equal to 2), selecting the shortest path in the path L, if k shortest paths exist, calculating the formula as follows:

then, the calculation is continued according to the sequence of the path L from the short to the long, if the calculation result of a certain path jCalculate path j as one of the shortest paths and recalculate trust_L(A,B)And when all paths are calculated, a final indirect trust result is obtained.