CN106484876A - A kind of based on typical degree and the collaborative filtering recommending method of trust network - Google Patents

Abstract

The present invention provides a kind of based on typical degree and the collaborative filtering recommending method of trust network, the method replaces original sparse rating matrix using dense user typical case's degree matrix and project typical case's degree matrix, and merges improvement of the trust network between user to traditional Collaborative Filtering Recommendation Algorithm.Typical degree matrix of the typical degree matrix and user by using project in Item Sets on the user's collection for liking certain intermediate item collection, alleviate in traditional Collaborative Filtering Recommendation Algorithm due to the few sparse sex chromosome mosaicism of user's score data quantity, fusion users to trust network improves recommendation precision further, while realizing Data Dimensionality Reduction.Recommendation results can fully merge impact of the social trusting relationship of user to similar users interest.

Description

Collaborative filtering recommendation method based on canonical degree and trust network

Technical Field

The invention relates to the technical field of personalized recommendation, in particular to a collaborative filtering recommendation method based on a typical degree and a trust network.

Background

In recent years, with the rapid development of the internet, the information data volume increases, and for internet users, the problem that it is difficult to accurately and efficiently find information of interest becomes a problem that needs to be solved urgently in the development of the internet. According to the user interest characteristics, a solution is established by combining the project characteristics, the constructed user relationship, the potential factors and the like, and a personalized recommendation technology is developed. In current internet website applications, social factors are increasingly fused, social data of users can conveniently share personal interest fields and wider interaction modes, such as attention, comment, forwarding and collection and the like, more information data can be obtained through analysis of the social data, and recommendation precision is improved.

At present, recommendation engines constructed in personalized recommendation technology are widely applied to various fields such as e-commerce, music movies or social networking sites. Two of the commonly used recommendation techniques are: content-based recommendations, collaborative filtering-based recommendations, model-based recommendations. And constructing user preference and the assumption that the user interest is unchanged through historical information data by a content-based recommendation algorithm, and recommending items to the user according to the fitting of the user preference data characteristics and the item characteristics. Collaborative filtering recommendation is widely used in the industry, and by analyzing user interests, designated interest similar users are found in the user population, and the item evaluation of the similar users is integrated to form preference degree prediction of the designated user on the information. Collaborative filtering is particularly useful in product recommendations such as music, movies, and the like. However, in the meantime, the conventional collaborative filtering recommendation technology has many limitations, such as: data sparsity problems, cold start problems, scalability problems. The data sparsity problem is that the system historical user scoring data is few, and the similarity calculation is difficult to have accurate similarity calculation when the similarity calculation between users or projects is carried out; the cold start problem is a problem that cannot be recommended when a new user starts without history scoring data; the expansibility problem is that when a user or a project is newly added, the system needs to be added with a large amount of calculation load. The improvement of the timeliness and the accuracy of recommendation is a research hotspot of a recommendation system.

With the rapid development of the Web2.0 technology, behavior data among users can supplement score data, and a trust network is a type of interactive behavior information of the users and other users in a website and describes mutual trust relations among the users. Since users are more prone to recommendations from trusted parties when recommending items for users that are similar to the users' preferences, trust networks have a key role in user preferences and item recommendations.

Disclosure of Invention

The invention provides a collaborative filtering recommendation method based on a canonical degree and a trust network, which can improve the recommendation precision and simultaneously realize the data dimension reduction.

In order to achieve the technical effects, the technical scheme of the invention is as follows:

a collaborative filtering recommendation method based on a canonical degree and a trust network comprises the following steps:

s1: acquiring history evaluation data of a user on a project and characteristic information data of the project;

s2: expressing the content label of each item into a text vector form, setting the number K of implied topics, and clustering the items by using an LDA topic model to generate a topic distribution theta and a topic-word distribution phi of an item set;

s3: calculating the typical degree of the project in the project set by using the project-subject distribution and the subject word distribution of the project, and constructing a project-project set typical degree matrix;

calculating the typical degree of a user in a user set which likes a certain type of theme and corresponds to the certain type of theme by using the rating data of the user on the project, and constructing a user-user set typical degree matrix;

s4: for a user Ui, calculating the similarity distance between users by using a threshold filtering algorithm, and constructing a nearest neighbor matrix;

s5: calculating the trust degrees of all neighbors of the Ui according to the evaluation of the neighbors of the Ui to the items, and constructing a trust degree matrix of the Ui neighbors of the user;

s6: and replacing the traditional scoring matrix with a typical degree matrix, fusing the weighted average of the trust degrees to calculate the scoring prediction, and generating a recommendation list according to the scoring prediction.

Further, in the step S2:

clustering the items using the LDA topic model produces an item set topic distribution θ and a topic-word distribution Φ, where the item set is represented by item O_iRepresents:

where m represents the number of items, O represents an item, w_i,mRepresenting the typical degree of the item in the item set, and clustering through a topic model to obtain a topic word vector which consists of the attributes and values of the item set and is represented as follows:

where r is a real number from 0 to 1, reflecting the importance of the property in the cluster center,time indicates that the attribute is not contained under a certain topic,is a keyword that indicates that the attribute is a certain topic and can be defined by it;

further, step 2 comprises several substeps, for any item O in the collection of items_i：

1) Probability of occurrence of subject theta_i～Dirichlet(α)；

2) For each word W_inGenerating a topic distribution z_inMult (theta) and words

Further, the step S3 includes calculating the item-item set representativeness and calculating the user-user set representativeness, wherein the calculation process of the item-item set representativeness is as follows:

the item-item set typical degree refers to the combination of the similarity degree of an individual item to a certain item cluster center and the dependency degree of the item to other item cluster centers, and the typical degree of the item in the item set is recorded asThe specific calculation formula is as follows:

wherein,indicating the internal similarity of an item in a set of items, in particularA feature vector representing the cluster class center of the cluster item,the feature vector of each item is represented,representing the similarity of the individual items to the feature vector of the item set center;

wherein,indicating the integration of the external dissimilarity of the item in all sets of items, in particularRepresenting the degree of dissimilarity of individual items to the center of the item, C representing the center of the cluster, N_cIndicates the number of cluster centers, k_jRepresenting a set of items;

the similarity calculation method may adopt a pearson correlation coefficient, a similarity based on euclidean distance, and a modified cosine similarity calculation method, and the similarity calculation formula based on euclidean distance in this embodiment is as follows:

wherein, the generated item set user sets are corresponding relations, n represents the number of the user sets or the item sets,a central feature vector representing the set of items,feature attribute word vectors, v, representing items_j,yAnd v_OyRespectively representing the clustering centers and the individual feature values of the projects to be weighted;

the calculation process of the user-user set typical degree is as follows;

the user set represents the user set which is corresponding to the item set and likes the items, and the typical degree of the user in the user setRelated to the scoring of the user in the project set, including the scoring value P of the user on the project in the project set_gx,rAnd a scoring frequency P in the set of items_gx,fThe typical degree calculation method of the user in the user set is as follows:

in particular

WhereinRepresenting a user U_iIn item set K_xCorresponding user set g_xThe weighted average of the medium scores is,representing a user U_iIn item set K_iCorresponding user set g_xThe score frequency value of (a);

in thatMiddle w_x,yRepresenting the use of items in a user set g_xCorresponding k_xThe degree of representativeness in the set of items is taken as a weight. R_i,yIs the value of the user Ui's rating, R, for the item Oy in the set of items^maxIs the maximum score value, here 5;

whereinRepresenting a user U_iIn user set g_xCorresponding k_xFrequency of scoring in a set of items, where N_x,iRepresenting a user U_iIn item set K_xNumber of user's points, N_iIndicating the total number of items scored at the user; the higher the user scores a set of items and the higher the frequency of scoring, the higher the degree of typicality that the user has in the group of users who like such a set of items;

calculating the similarity between users by using a Euclidean distance-based formula and using a typical degree matrix to replace a traditional item scoring matrix, wherein the similarity is calculated by using similarity calculation based on the Euclidean distance:

where n is the number of user sets, representing users U_iAnd user U_jThe similarity distance of (2).

Further, the specific process of step S4 is as follows:

screening user neighbors by using a similarity threshold filtering method, selecting typical degree similarity among users to obtain a user nearest neighbor matrix:

whereinRepresenting a user U_iGamma is a set filtering neighbor similarity threshold, and when the similarity calculated according to the user typical degree matrix is greater than the set threshold, the neighbor users are added into the target user U_jThe similarity method uses the similarity calculation method based on the Euclidean distance to obtain the user neighbors.

Further, the specific process of step S5 is as follows:

according to the target user U_iThe neighbor user scores the history of the project and calculates U_iThe trust degrees of all neighbors are constructed to form a user U_iThe confidence level of the neighbor comprises two parts, namely a global trust relationship and a local trust relationship, and the specific calculation formula is as follows:

Trust(u_j,u_i)＝ω·gTrust(u_j,u_i)+(1-ω)lTrust(u_j,u_i)

wherein gTrust (u)_j,u_i) Representing user u_jAnd u_iFor global confidence, where lTrust (u)_j,u_i) Representing user u_jAnd u_iThe local trust degree is the adjustment factor omega between 0 and 1, and is used for giving different weights to the global trust degree and the local trust degree to represent different influence degrees, in the social relationship, the higher the mutual frequency between users is, the more the users recommend the prediction success times to other users, and the higher the trust degree is;

the global trust reflects the contribution degree of the existing scoring behavior of the user in the system to other users, and in the traditional collaborative filtering based on the user, if the user carries out accurate recommendation for many times when participating in the recommendation of another user and successfully predicts the scoring of the recommended user to the project, the user has higher global trust degree, namely is more worthy of being recommended and trusted by other users; conversely, when the number of times of successful prediction is less, the user trust degree is lower, and the target user U is subjected to prediction_iIts neighbor U_jTo U_iThe local trust is calculated as follows:

wherein Item (U)_i) For user U_iThe set of items, Item (U), that have been scored_j) For user U_jThe set of items that have already been scored may,represents the proportion of the user common scoring item in the user j, wherein SuccessSet (U)_i,U_j) Indicates the number of correct recommendations as the user U_jAs screened out user U_iWhen the neighbor is close to the user U_jPredictive score and user U_iThe average item score difference is within a certain range, and the specific formula is as follows:

whereinRespectively represent users U_j,U_iThe average value of the scores of the items can reflect the scoring habits of the user, R_j(o) represents user U_jActual rating of item O, P being indicated by neighboring user U_jSupposing that the score of the user I on the item o, E represents an error range value, the setting can be adjusted, and the score adopts a discretization score with the range of 1-5, so that the set score error value is E-1;

RecommandSet(U_i,U_j) Representing a user U_jAnd user U_iIs expressed as:

whereinWhen the user Ui and the user Uj have common scoring items;

the local trust of the user in the system reflects the social trust relationship between a pair of users in the social network, and is the number of interactive actions between the users. The interaction of the calculation of the local trust degree comprises the attention number and the evaluation praise interaction frequency between the attention number and the user of the user, and the specific local trust degree calculation formula is as follows:

wherein in (U)_j) Represented in the user interaction network, user U_jIn degree of (i.e. user U)_jNumber of users of interest, out (U)_i) Expressed in user interaction relation, user U_iOut of service, namely user U_iIs concerned with the number of users, where U_iCij represents user U as target user_iAnd user U_jMax (C) represents the user U_jMaximum number of interactions, when neighboring user U_jThe more the attention is, the more frequently the user interacts with other users, the higher the credibility of the user, and η represents the weight of the adjustment factor between 0 and 1 for the user to adjust the attention information and the interaction information on the calculation of the user credibility.

Further, the specific process of step S6 is as follows:

meanwhile, the typical degree matrix is used for replacing the traditional scoring matrix, the weighted average of the trust degree is fused for calculating the scoring prediction, a recommendation list is generated according to the scoring prediction, the similarity and the trust factor of the typical degree are harmonized, the influence of the two aspects on the weight is synthesized, and the comprehensive weight is finally obtained, wherein the specific formula is as follows:

wherein ω (i, a) represents user i and user a, typsim (i, a) represents the similarity obtained from the canonical degree matrix of the user, and trust (i) represents the confidence of user i;

the harmonic weight obtained according to the harmonic formula of the typical similarity and the trust factor is as follows:

according to the harmonic weights, a recommendation scoring formula is given as follows:

whereinRepresents the average rating of user u for the item, c represents the neighborhood of user u, R_iThe score of the current item for the neighbor user i representing u,representing the grade of the user i to the project, TypSimuser (i, u) representing the similarity of the user i and the user u calculated according to the typical degree matrix, UserTru_iRepresenting the degree of trust, P, of a neighbour user i_uAnd expressing the prediction scores of the recommendation algorithm, and generating Top-K recommendation set results with the highest prediction scores according to the prediction score ranking.

Compared with the prior art, the technical scheme of the invention has the beneficial effects that:

the method uses the dense user typical degree matrix and the project typical degree matrix to replace the original sparse scoring matrix, and combines the improvement of the trust network among the users on the traditional collaborative filtering recommendation algorithm. By using the typical degree matrix of the items in the item set and the typical degree matrix of the users in the user set which likes a certain type of item set, the problem of sparsity caused by small quantity of user scoring data in the traditional collaborative filtering recommendation algorithm is solved, the recommendation precision is further improved by fusing the user trust network, and meanwhile, the data dimension reduction is realized. The recommendation result can fully integrate the influence of the social trust relationship of the user on the similar user interests.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a diagram illustrating a typical relationship between an item-item set and a user set according to the present invention.

Detailed Description

The drawings are for illustrative purposes only and are not to be construed as limiting the patent;

for the purpose of better illustrating the embodiments, certain features of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product;

it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.

The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.

Example 1

As shown in fig. 1, a collaborative filtering recommendation method based on a canonical degree and a trust network includes the following steps:

step 1: and traversing all users in the current network system, and acquiring user and historical scoring data, items and user characteristic information.

Step 2: preprocessing data: expressing the content label of each item into a form of a characteristic text vector, setting the number K of implicit topics, clustering the items by using an LDA topic model to generate a topic distribution theta and a topic-word distribution phi of an item set, wherein the item set is composed of items O_iRepresents:

where m represents the number of items, O represents an item, w_i,mIndicating the degree of representativeness of the item in the set of items.

Through topic model clustering, the obtained topic word vector (cluster center feature vector) is composed of item set attributes and values and is represented as follows:

where r is a real number from 0 to 1, reflecting the importance of the attribute in the cluster center.Time indicates that the attribute is not contained under a certain topic,is a keyword that indicates that the attribute is a certain topic and can be defined by it.

Specifically, step 2 comprises the following substeps, for any item O in the collection of items_i：

(2-1) topic occurrence probability θ_i～Dirichlet(α)

(2-2) for each word W_inGenerating a topic distribution z_inMult (theta) and words

The specific implementation taking the movie scoring project data as an example is to extract a keyword label according to movie content expression information, and construct a topic-word vector as a feature vector of a clustering center through topic model clustering.

In step 3, the method specifically comprises two parts of item-item set typical degree calculation and user-user set typical degree calculation (as shown in fig. 2). And constructing the item-item set typical degree and the user-user set typical degree by using a Euclidean distance-based formula.

(3-1) calculation of item-item set representativeness

The item-item set representativeness refers to the combination of the degree of similarity of an individual item to a certain item cluster center and the degree of dependency of the item to other item cluster centers, which are respectively called: internal similarity and external dissimilarity characterize an item in a set of items asThe specific calculation formula is as follows:

wherein

Indicating the internal similarity of an item in a set of items, in particularA feature vector representing the cluster class center of the cluster item,the feature vector of each item is represented,representing the similarity of the individual items to the central feature vector of the item set.

Wherein

Representing the external dissimilarity of the composite item in all sets of itemsIn particularRepresenting the degree of dissimilarity of individual items to the center of the item, C representing the center of the cluster, N_cIndicates the number of cluster centers, k_jRepresenting a set of items.

The similarity calculation method may adopt a pearson correlation coefficient, a similarity based on euclidean distance, and a modified cosine similarity calculation method, and the similarity calculation formula based on euclidean distance in this embodiment is as follows:

wherein, the generated item set user sets are corresponding relations, n represents the number of the user sets or the item sets,a central feature vector representing the set of items,feature attribute word vectors, v, representing items_j,yAndand respectively representing the cluster center and the item individual characteristic values to be weighted.

In particular in movie data, feature vectors such as the item Se7en (seven guilt) can be represented as attribute features according to the cluster centerAccording to the project typical degree calculation formula, the project feature vector and the typical degree of each theme can be obtained. For example, tamanik number is both an love movie and a disaster movie, but is more typical in the love movie category than in the disaster movie category. The calculation formula can obtain a typical degree matrix of the item-item set.

(3-2) calculation of user-user set representativeness

The user set represents the user set which is corresponding to the item set and likes the items, and the typical degree of the user in the user set is

The scores of the users in the project set are related, and the scores P of the users on the projects in the project set are included_gx,rAnd a scoring frequency P in the set of items_gx,fThe typical degree calculation method of the user in the user set is as follows:

in particular

WhereinRepresenting a user U_iIn item set K_xCorresponding user set g_xThe weighted average of the medium scores is,representing a user U_iIn item set K_iCorresponding user set g_xThe score frequency value of (1).

In thatMiddle w_x,yRepresenting the use of items in a user set g_xCorresponding k_xThe degree of representativeness in the set of items is taken as a weight. R_i,yOf the user Ui to the item Oy in the set of itemsValue of credit, R^maxIs the maximum score value, here 5.

WhereinRepresenting a user U_iIn user set g_xCorresponding k_xFrequency of scoring in a set of items. Wherein N is_x,iRepresenting a user U_iIn item set K_xNumber of user's points, N_iIndicating the total number of items scored at the user.

Thus, the higher the user scores a set of items and the higher the frequency of scoring, the higher the degree of typability that the user has in a group of users who like such a set of items.

And (3-3) calculating the similarity between the users by using a Euclidean distance-based formula and using a typical degree matrix instead of the traditional item scoring matrix.

The similarity calculation uses similarity calculation based on Euclidean distance:

where n is the number of user sets, representing users U_iAnd user U_jDistance of similarity of

Specifically, in step 4, the selection of neighbors is a key factor influencing the accuracy of the recommendation effect, and a dense user-user set typical degree matrix is used instead of a sparse user item scoring matrix, wherein the user typical degree matrix refers to a graph whose data is that dense data users have different degrees of typical degrees in each user set. And the original user scoring matrix is very sparse, a similarity threshold filtering method is used for screening user neighbors on the typical degree matrix according to a reference graph, and the typical degree similarity between the users is selected to obtain the user nearest neighbor matrix.

WhereinRepresenting a user U_iGamma is a set filtering neighbor similarity threshold, and when the similarity calculated according to the user typical degree matrix is greater than the set threshold, the neighbor users are added into the target user U_jThe similarity method uses the similarity calculation method based on the Euclidean distance to obtain the user neighbors.

Specifically, in step 5, according to the history scores of the neighbor users of the target user Ui on the items, the trust degrees of all the neighbors of the Ui are calculated, and a trust degree matrix of the neighbors of the user Ui is constructed. The trust degree of the neighbor comprises two parts, namely a global trust relationship and a local trust relationship. The specific calculation formula is as follows:

Trust(u_j,u_i)＝ω·gTrust(u_j,u_i)+(1-ω)lTrust(u_j,u_i)

wherein gTrust (u)_j,u_i) Representing user u_jAnd u_iFor global confidence, where lTrust (u)_j,u_i) Representing user u_jAnd u_iAnd omega is an adjusting factor between 0 and 1 for the local confidence, and is used for giving different weights to the global confidence and the local confidence to represent different influence degrees of the global confidence and the local confidence. In the social relationship, the higher the interaction frequency among the users is, the more the users successfully predict the recommendation of other users, and the higher the trust degree is.

The global trust reflects the contribution degree of the existing scoring behavior of the user in the system to other users, and in the traditional collaborative filtering based on the user, if the user carries out accurate recommendation for many times when participating in the recommendation of another user and successfully predicts the scoring of the recommended user to the item, the user has higher global trust degree, namely is more worthy of being recommended and trusted by other users. Conversely, the less the number of times the prediction succeeds, the lower the user confidence. For the target user Ui, the local trust of the neighbor Uj of the target user Ui to the Ui is calculated as follows:

wherein Item (U)_i) For user U_iThe set of items, Item (U), that have been scored_j) For user U_jThe set of items that have already been scored may,represents the proportion of the user common scoring item in the user j, wherein SuccessSet (U)_i,U_j) The number of correctly recommended users is represented, when the user Uj is adjacent to the screened user Ui, the difference between the predicted score of the user Uj and the average project score of the user Ui is within a certain range, and the specific formula is as follows:

whereinRespectively represents that the average value of the user Uj and Ui for the item can reflect the scoring habit of the user, R_j(o) represents user U_jThe actual rating of item O. P represents the score of the item o inferred by the neighboring user Uj by the user I. E represents an error range value, and the setting can be adjusted, and since the score is a discretized score in the range of 1 to 5, the score error value is set to E equal to 1.

RecommandSet(U_i,U_j) Representing a user U_jAnd user U_iIs provided with the number of common score items. Expressed as:

whereinIndicating when user Ui has a common scoring item with user Uj.

The local trust of the user in the system reflects the social trust relationship between a pair of users in the social network, and is the number of interactive actions between the users. The interaction of the calculation of the local credibility comprises the attention number and the evaluation praise interaction frequency between the attention number and the user. The specific local confidence calculation formula is as follows:

wherein in (U)_j) Represented in the user interaction network, user U_jIn degree of (i.e. user U)_jNumber of users of interest, out (U)_i) Expressed in user interaction relation, user U_iOut of service, namely user U_iIs concerned with the number of users, where U_iCij represents the number of interactions between the user Ui and the user Uj, max (C) represents the maximum number of interactions between the user Uj, when the number of concerned users Uj is more, the more frequently the users interact with other users, the higher the credibility of the users is, η represents that the adjustment factor is between 0 and 1, and the weight of the users for adjusting the concerned information and the interactive information on the calculation of the user credibility is.

Step 6: meanwhile, a typical degree matrix is used for replacing a traditional scoring matrix, a weighted average calculation scoring prediction of the trust degree is fused, a recommendation list is generated according to the scoring prediction, the similarity and the trust factor of the typical degree are harmonized, the influence of the two aspects on the weight is integrated, and the comprehensive weight is finally obtained.

Where ω (i, a) represents user i and user a, typsim (i, a) represents the similarity obtained from the user's canonical degree matrix, and trust (i) represents the confidence of user i.

The harmonic weight obtained according to the harmonic formula of the typical similarity and the trust factor is as follows:

according to the harmonic weights, a recommendation scoring formula is given as follows:

whereinRepresents the average rating of user u for the item, c represents the neighborhood of user u, R_iThe score of the current item for the neighbor user i representing u,and the value of the user i on the project is shown, and the TypSimuser (i, u) shows the similarity of the user i and the user u calculated according to the typical degree matrix. UserTru_iRepresenting the confidence level of the neighbor user i. P_uAnd expressing the prediction scores of the recommendation algorithm, and generating Top-K recommendation set results with the highest prediction scores according to the prediction score ranking.

The same or similar reference numerals correspond to the same or similar parts;

the positional relationships depicted in the drawings are for illustrative purposes only and are not to be construed as limiting the present patent;

it should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.

Claims (6)

1. A collaborative filtering recommendation method based on a canonical degree and a trust network is characterized by comprising the following steps:

s1: acquiring history evaluation data of a user on a project and characteristic information data of the project;

s2: expressing the content label of each item into a text vector form, setting the number K of implied topics, and clustering the items by using an LDA topic model to generate a topic distribution theta and a topic-word distribution phi of an item set;

s3: calculating the typical degree of the project in the project set by using the project-subject distribution and the subject word distribution of the project, and constructing a project-project set typical degree matrix;

calculating the typical degree of a user in a user set which likes a certain type of theme and corresponds to the certain type of theme by using the rating data of the user on the project, and constructing a user-user set typical degree matrix;

s4: for a user Ui, calculating the similarity distance between users by using a threshold filtering algorithm, and constructing a nearest neighbor matrix;

s5: calculating the trust degrees of all neighbors of the Ui according to the evaluation of the neighbors of the Ui to the items, and constructing a trust degree matrix of the Ui neighbors of the user;

s6: and replacing the traditional scoring matrix with a typical degree matrix, fusing the weighted average of the trust degrees to calculate the scoring prediction, and generating a recommendation list according to the scoring prediction.

2. The collaborative filtering recommendation method based on representativeness and trust network according to claim 1, wherein in said step S2:

clustering the items using the LDA topic model produces an item set topic distribution θ and a topic-word distribution Φ, where the item set is represented by item O_iRepresents:

$k_i=\{O_1^{w_{i,1}},O_2^{w_{i,2}}...O_m^{w_{i,m}}\}$

wherein m representsNumber of items, O represents an item, w_i,mRepresenting the typical degree of the item in the item set, and clustering through a topic model to obtain a topic word vector which consists of the attributes and values of the item set and is represented as follows:

$\stackrel{\→}{t_{k_j}}=(p_{k_j,1}:r_{k_j,1},p_{k_j,2}:r_{k_j,2}...p_{k_j,m}:r_{k_j,m})$

where r is a real number from 0 to 1, reflecting the importance of the property in the cluster center,time indicates that the attribute is not contained under a certain topic,is a keyword that indicates that the attribute is a certain topic and can be defined by it;

further, step 2 comprises several substeps, for any item O in the collection of items_i：

1) Probability of occurrence of subject theta_i～Dirichlet(α)；

2) For each word W_inGenerating a topic distribution z_inMult (theta) and words

3. The collaborative filtering recommendation method based on the representativeness and the trust network of claim 2, wherein the step S3 comprises a calculation of item-item set representativeness and a calculation of user-user set representativeness, wherein the calculation of the item-item set representativeness is as follows:

the item-item set typical degree refers to the combination of the similarity degree of an individual item to a certain item cluster center and the dependency degree of the item to other item cluster centers, and the typical degree of the item in the item set is recorded asThe specific calculation formula is as follows:

$T_{k_x}\left(O_i\right)=\mathrm{\α}(\stackrel{\→}{t_{k_j}},\stackrel{\→}{p_{O_y}})=\mathrm{\β}(\stackrel{\→}{t_{k_j}},\stackrel{\→}{p_{O_y}})\·\mathrm{\δ}(\stackrel{\→}{t_{k_j}},\stackrel{\→}{p_{O_y}})$

wherein,indicating the internal similarity of an item in a set of items, in particularA feature vector representing the cluster class center of the cluster item,the feature vector of each item is represented,representing the similarity of the individual items to the feature vector of the item set center;

wherein,indicating the integration of the external dissimilarity of the item in all sets of items, in particularRepresenting the degree of dissimilarity of individual items to the center of the item, C representing the center of the cluster, N_cIndicates the number of cluster centers, k_jRepresenting a set of items;

the similarity calculation method may adopt a pearson correlation coefficient, a similarity based on euclidean distance, and a modified cosine similarity calculation method, and the similarity calculation formula based on euclidean distance in this embodiment is as follows:

wherein, the generated item set user sets are corresponding relations, n represents the number of the user sets or the item sets,a central feature vector representing the set of items,feature attribute word vectors, v, representing items_j,yAnd v_OyRespectively representing the clustering centers and the individual feature values of the projects to be weighted;

the calculation process of the user-user set typical degree is as follows;

the user set represents a user set which likes the items and corresponds to the item set, and the typical degree of the user in the user set is related to the rating of the user in the item set, including the rating value P of the user on the items in the item set_gx,rAnd a scoring frequency P in the set of items_gx,fThe typical degree calculation method of the user in the user set is as follows:

${\mathrm{\τ}}_{g_x}\left(U_i\right)=\frac{S_{gx,r}^i+S_{gx,f}^i}{2}$

in particular

$S_{gx,r}^i=\frac{{\mathrm{\Σ}}_{y=1}^n{w}_{x,y}\·R_{i,y}}{n\·R^{max}}$

$S_{gx,f}^i=\frac{N_{x,i}}{N_i}=\frac{N_{x,i}}{{\mathrm{\Σ}}_{y=1}^n{N}_{y,i}}$

WhereinRepresenting a user U_iIn item set K_xCorresponding user set g_xWeighted average of medium scores, Sⁱ _gx,fRepresenting a user U_iIn item set K_iCorresponding user set g_xThe score frequency value of (a);

in thatMiddle w_x,yRepresenting the use of items in a user set g_xCorresponding k_xThe degree of representativeness in the set of items is taken as a weight. R_i,yIs the value of the user Ui's rating, R, for the item Oy in the set of items^maxIs the maximum score value, here 5;

whereinRepresenting a user U_iIn user set g_xCorresponding k_xFrequency of scoring in a set of items, where N_x,iRepresenting a user U_iIn item set K_xNumber of user's points, N_iIndicating the total number of items scored at the user; the higher the user scores a set of items and the higher the frequency of scoring, the higher the degree of typicality that the user has in the group of users who like such a set of items;

calculating the similarity between users by using a Euclidean distance-based formula and using a typical degree matrix to replace a traditional item scoring matrix, wherein the similarity is calculated by using similarity calculation based on the Euclidean distance:

$Sim(U_i,U_j)=\exp (-\sqrt{\underset{y=1}{\overset{n}{\Σ}}|u_{i,y}-u_{j,y}{|}^2})$

where n is the number of user sets, representing users U_iAnd user U_jThe similarity distance of (2).

4. The collaborative filtering recommendation method based on the representativeness and trust network of claim 3, wherein the specific process of the step S4 is as follows:

screening user neighbors by using a similarity threshold filtering method, selecting typical degree similarity among users to obtain a user nearest neighbor matrix:

${\stackrel{\→}{N}}_i=\left\{U_i\right|Sim(U_i,U_j)\≥\mathrm{\γ}\}$

whereinRepresenting a user U_iGamma is a set filtering neighbor similarity threshold, and when the similarity calculated according to the user typical degree matrix is greater than the set threshold, the neighbor users are added into the target user U_jThe similarity method uses the similarity calculation method based on the Euclidean distance to obtain the user neighbors.

5. The collaborative filtering recommendation method based on representativeness and trust network according to claim 4, wherein the specific process of the step S5 is as follows:

according to the target user U_iThe neighbor user scores the history of the project and calculates U_iThe trust degrees of all neighbors are constructed to form a user U_iA matrix of confidence levels of neighbors, the confidence level of a neighbor comprising two parts, global confidenceThe specific calculation formula of the arbitrary relation and the local trust relation is as follows:

Trust(u_j,u_i)＝ω·gTrust(u_j,u_i)+(1-ω)lTrust(u_j,u_i)

wherein gTrust (u)_j,u_i) Representing user u_jAnd u_iFor global confidence, where lTrust (u)_j,u_i) Representing user u_jAnd u_iThe local trust degree is the adjustment factor omega between 0 and 1, and is used for giving different weights to the global trust degree and the local trust degree to represent different influence degrees, in the social relationship, the higher the mutual frequency between users is, the more the users recommend the prediction success times to other users, and the higher the trust degree is;

the global trust reflects the contribution degree of the existing scoring behavior of the user in the system to other users, and in the traditional collaborative filtering based on the user, if the user carries out accurate recommendation for many times when participating in the recommendation of another user and successfully predicts the scoring of the recommended user to the project, the user has higher global trust degree, namely is more worthy of being recommended and trusted by other users; conversely, when the number of times of successful prediction is less, the user trust degree is lower, and the target user U is subjected to prediction_iIts neighbor U_jTo U_iThe local trust is calculated as follows:

${\mathrm{gTrust}}_{U_j\→U_i}=\frac{|Item\left(U_j\right)\∩Item\left(U_i\right)|}{\left|Item\left(U_j\right)\right|}\frac{\left|SuccessSet(U_j,U_i)\right|}{\left|\mathrm{Re}commandSet(U_j,U_i)\right|}$

wherein Item (U)_i) For user U_iThe set of items, Item (U), that have been scored_j) For user U_jThe set of items that have already been scored may,represents the proportion of the user common scoring item in the user j, wherein SuccessSet (U)_i,U_j) Indicates the number of correct recommendations as the user U_jAs users screened outU_iWhen the neighbor is close to the user U_jPredictive score and user U_iThe average item score difference is within a certain range, and the specific formula is as follows:

$\left\{\begin{array}{c}P=\stackrel{\‾}{R_i}+\left(R_j\right(o)-\stackrel{\‾}{R_j})*TypSimUser\\ |P-{\stackrel{\‾}{R}}_i|\≤e\end{array}\right.$

whereinRespectively represent users U_j,U_iThe average value of the scores of the items can reflect the scoring habits of the user, R_j(o) represents user U_jActual rating of item O, P being indicated by neighboring user U_jSupposing that the user I scores the item o, and E represents an error range value, the setting can be adjusted, and the score adopts the range of1-5, thus setting the score error value as e-1;

RecommandSet(U_i,U_j) Representing a user U_jAnd user U_iIs expressed as:

$\mathrm{Re}commandSet(U_i,U_j)=\{(U_i,U_j):P_{U_i,U_j}\left(i\right)!=null\}$

whereinWhen the user Ui and the user Uj have common scoring items;

the local trust of the user in the system reflects the social trust relationship between a pair of users in the social network, and is the number of interactive actions between the users. The interaction of the calculation of the local trust degree comprises the attention number and the evaluation praise interaction frequency between the attention number and the user of the user, and the specific local trust degree calculation formula is as follows:

$lTrust(U_i,U_j)=\mathrm{\η}\sqrt{\frac{in\left(U_j\right)}{in\left(U_j\right)+out\left(U_i\right)}}+(1-\mathrm{\η})\frac{C_{i,j}}{max\left(C\right)}$

wherein in (U)_j) Represented in the user interaction network, user U_jIn degree of (i.e. user U)_jNumber of users of interest, out (U)_i) Expressed in user interaction relation, user U_iOut of service, namely user U_iIs concerned with the number of users, where U_iCij represents user U as target user_iAnd user U_jMax (C) represents the user U_jMaximum number of interactionsNumber, when the neighboring user U_jThe more the attention is, the more frequently the user interacts with other users, the higher the credibility of the user, and η represents the weight of the adjustment factor between 0 and 1 for the user to adjust the attention information and the interaction information on the calculation of the user credibility.

6. The collaborative filtering recommendation method based on representativeness and trust network according to claim 5, wherein the specific process of the step S6 is as follows:

meanwhile, the typical degree matrix is used for replacing the traditional scoring matrix, the weighted average of the trust degree is fused for calculating the scoring prediction, a recommendation list is generated according to the scoring prediction, the similarity and the trust factor of the typical degree are harmonized, the influence of the two aspects on the weight is synthesized, and the comprehensive weight is finally obtained, wherein the specific formula is as follows:

$\frac{1}{\mathrm{\ω}(i,a)}=\frac{2}{1/TypSimUser(i,a)+1/Trust\left(i\right)}$

wherein ω (i, a) represents user i and user a, typsim (i, a) represents the similarity obtained from the canonical degree matrix of the user, and trust (i) represents the confidence of user i;

the harmonic weight obtained according to the harmonic formula of the typical similarity and the trust factor is as follows:

$\mathrm{\ω}(i,a)=\frac{2TypSimUser(i,a)\·Trust\left(i\right)}{TypSimUser(i,a)+Trust\left(i\right)}$

according to the harmonic weights, a recommendation scoring formula is given as follows:

$P_u=\stackrel{\‾}{R_u}+\frac{\underset{i=1}{\overset{c}{\Σ}}(R_i-\stackrel{\‾}{R_i})*\mathrm{\ω}(i,a)}{\underset{i=1}{\overset{c}{\Σ}}\left|\mathrm{\ω}(i,a)\right|}$

whereinRepresents the average rating of user u for the item, c represents the neighborhood of user u, R_iThe score of the current item for the neighbor user i representing u,representing the grade of the user i to the project, TypSimuser (i, u) representing the similarity of the user i and the user u calculated according to the typical degree matrix, UserTru_iRepresenting the degree of trust, P, of a neighbour user i_uAnd expressing the prediction scores of the recommendation algorithm, and generating Top-K recommendation set results with the highest prediction scores according to the prediction score ranking.