CN105740430A - Personalized recommendation method with socialization information fused - Google Patents

Abstract

The present invention provides a personalized recommendation method with socialization information fused. The method comprises the following steps: S1, constructing a user- user trust matrix; S2, constructing a project- project tag similarity matrix; S3, constructing and training a model; and S4, predicting a preference of a user for an unknown project. The method provided by the present invention mainly has the following advantages: 1) a sorting learning method in the information retrieval field is applied to Top- K recommendation, so that the sorting problem in the recommendation system is effectively solved, and the defect that the conventional score-based prediction method can not perform Top-K recommendation is overcome; and 2) socialization information, i.e. user social information and project tag information, is fused into the model based on sorting learning, so that accuracy of recommendation results is improved.

Description

A kind of personalized recommendation method of mosaic society information

Technical field

The present invention relates to personalized recommendation, sequence study and community network field, especially a kind of mosaic society information Personalized recommendation method.

Background technology

Along with developing rapidly of Internet technology particularly ecommerce, in the Internet the growth rate of data considerably beyond The reception speed of the mankind, problem of information overload seems increasingly severe.Us are helped to filter out useful number from mass data According to Information Filtering Technology seem more and more important, personalized recommendation technology just a kind of according to user preference from large-scale data In find the Perfected process of user's data of interest.

At present, the application of personalized recommendation is broadly divided into two big classes.The first kind is score in predicting problem, i.e. by given one The unknown purpose is marked by the history scoring behavior prediction of individual user, and score value i.e. represents user's fancy grade to project.The Two classes are that Top-K recommends problem, and Top-K recommends K the project being devoted to recommend its most probable to like for user.Owing to user is past Coming project above toward concern, therefore compare with score in predicting problem, Top-K provides the user sequence more intuitively Recommendation list, the most practical, this is also the problem that current each big e-commerce website is devoted to solve.

The core of personalized recommendation technology is proposed algorithm, it is presently recommended that algorithm is broadly divided into two big classes, is interior respectively Hold and filter and collaborative filtering.Information filtering recommends method mainly by analyzing user and the content information of project, such as the people of user Mouth statistical information, the description information etc. of project, thus construct the series of features of user and project, eventually through coupling user Recommendation is made with the similarity of project.In contrast to this, collaborative filtering method need not the content of any user or project Information, is a kind of method the most unrelated with field.Collaborative filtering method efficiently utilizes group intelligence, and it is based on such It is assumed that user can like and oneself have the project that same interest user likes, meanwhile, the joint act between user is the most at most Interest between user is the most similar.Collaborative filtering method is broadly divided into two big classes at present, is collaborative filtering based on memory respectively With collaborative filtering based on model.Collaborative filtering method efficiently avoid the problem needing expert's markup information, and It is widely used in various commending system.

Although said method can improve recommendation accuracy rate to a certain extent, but is faced with " number in actual applications According to sparse " and " cold start-up " problem." Sparse " problem refers to that user-project matrix hollow element is too much, has value element mistake Less thus cause the very few problem of availability data；" cold start-up " problem refer to the behavioral data of new user very few cause system without The problem that method analyzes its preference.In recent years, along with the development of online social networks, personalized recommendation side based on socialization's information Method is increasingly paid attention to by industrial quarters and research worker.Socialization's information spinner social network to be included and socialized label, Hao Ma in 2008 et al. proposes SoRec method, and the method using probability matrix to decompose decomposes user-project rating matrix simultaneously Trust matrix with user-user to recommend；Mohsen in 2010 et al. proposes SocMF method, during matrix decomposition The difference of the characteristic vector between constraint user and user friend simultaneously, uses the matrix decomposition trusting conduction based on social networks Method；Le Wu in 2012 et al. proposes NHPMF method, utilizes the label information of user and project, decomposes mould at probability matrix Type adds the label constraint item of access customer and project to carry out model training, and then obtain the potential feature square of user and project Battle array, is predicted user-project preference value.

Above-mentioned recommendation method is to optimize score in predicting accuracy rate as target, and there have in score in predicting problem to be higher accurate Rate, but do not account for the sequencing problem of recommendation list, when carrying out Top-K recommendation, there is certain limitation.Top-K list Recommend to be considered as a sequencing problem, and study of sorting is a kind of method optimizing document ordering in information retrieval field. User-project being inputted the inquiry-document analogizing to information retrieval field to as data, the method for sequence study is the most gradually Applying in personalized recommendation field, similar with tradition sequence learning method, the sequence learning method in personalized recommendation is also led Three major types to be divided into, is Point-wise method, Pair-wise method and List-wise method respectively.Point-wise method It is intended that single project forecast goes out hobby value accurately, tradition collaborative filtering method based on score in predicting broadly falls into Point- Wise method；Nathan etc. proposed Bradley-TerryMF method based on Pair-wise in 2010, and Xin Liu etc. exists Within 2014, proposing RankNet-MF method, method based on Pair-wise utilizes the partial ordering relation instruction that project is liked by user Practice model, to optimize the sorted lists of project；Yue Shi in 2010 et al. proposes ListRank-MF method, uses cross entropy (Cross-Entropy) as loss function, it is proposed that a kind of matrix disassembling method based on List-wise sequence study, purport To the farthest matching of whole recommendation list and optimization.

The method of Point-wise remains the model towards score in predicting, does not accounts for the characteristic of sequence；Pair-wise Method need to consider the partial ordering relation between all items, the complexity of model training is too high；Although ListRank-MF method Take into account the sequence optimizing whole recommendation list, the sequencing problem of project can be solved to a certain extent, but due at model In the information that incorporates very little, do not account for user and the impact of project socialization information, be still difficult to avoid that " Sparse Property " and " cold start-up " problem, there is significant limitation in actual applications.

Summary of the invention

It is an object of the invention to provide a kind of closer to people's thinking habit, there is the mosaic societyization letter of relatively high-accuracy The personalized recommendation method of breath.

The present invention solves the technical scheme that prior art problem is used: the personalized recommendation of a kind of mosaic society information Method, comprises the following steps:

S1, structure user-user trust matrix:

Oriented degree of belief between a1, acquisition user: in the known data base containing socialization's information, according to described data base In oriented degree of belief, oriented degree of belief between described user between concern Relation acquisition user between user in the social networks that comprises Acquisition methods is as follows:

$t_{uk}=\sqrt{\frac{d^{-}\left(v_k\right)}{d^{+}\left(v_u\right)+d^{-}\left(v_k\right)}}$

Wherein, t_ukRepresent in data base's social networks oriented degree of belief, d between the user u user to user k^-(v_k) represent The quantity that user k is concerned, d⁺(v_u) represent that user u pays close attention to the quantity of user；

A2, normalized: then to degree of belief t oriented between user_ukDo normalized, obtain t_ukUser between have To degree of belief initialization value T_uk, and built user-user trust matrix by oriented degree of belief initialization value between described user；

S2, structure project-item label similarity matrix:

B1, the weight vectors of acquisition project: the item label information in acquisition database uses tf*idf weight to be each Each label labelling weight in project, and the element of the weight vectors of each project is constituted with tf*idf weighted value, project The dimension of label weight vectors is the quantity of label in data base, and the label that the label weight vectors of project represents project is special Levying, concrete grammar is as follows:

Wherein, N is the number of project, w in data base_jtThe weighted value of the label t in expression project j, (j t) represents item to tf Mesh j is marked with the number of times of label t, does not has during visible marking's number of times to be designated as 1, and df (t) represents the project number that label t is labeled, Markd label weight is not had automatically to be designated as 0；

Label similarity between b2, calculating project:

By calculating the label similarity between the cosine similarity acquisition project of the weight vectors of two projects, its computing formula As follows:

$sim(j_1,j_2)=cos({\stackrel{\→}{j}}_1,{\stackrel{\→}{j}}_2)=\frac{{\stackrel{\→}{j}}_1\·{\stackrel{\→}{j}}_2}{\left|\right|{\stackrel{\→}{j}}_1|{|}_F*|\left|{\stackrel{\→}{j}}_2\right|{|}_F}$

Wherein, sim (j₁,j₂) it is project j₁With project j₂Between label similarity,For the item obtained in step b1 Mesh j₁With project j₂Weight vectors；

B3, the label similarity chosen between k nearest neighbor normalization project: according to the label similarity options between project Purpose k nearest neighbor, is normalized the label similarity between the project of k nearest neighbor, obtains between the k nearest neighbor project after normalization Label similarity, and the label similarity between the project outside k nearest neighbor is set to 0, obtain initialized project-item label phase Seemingly spend matrix；

S3, structure and training pattern:

C1, collection training dataset: randomly draw the 80% conduct training number of the known data base containing socialization's information According to collection；

C2, structure user-project rating matrix:

Training data is concentrated each user comprised mark for each project and gives the score value of 1-5 and with these points It is worth and builds user-project rating matrix as matrix element；

C3, calculate user for the first probability of certain score value:

Calculating user i for the scoring of project j by below equation is R_ijThe first probability Wherein,For increasing function and the most satisfied for all xParameter D is that user i commented The number of entry divided；

C4, build and train mosaic society's information list ordering study recommended models, including:

D1, the loss function L (U, V) of structure mosaic society information:

$\begin{array}{c}L(U,V)=\underset{i=1}{\overset{M}{\Σ}}\{-\underset{j=1}{\overset{N}{\Σ}}{I}_{ij}{P}_{l_i}\left(R_{ij}\right)\log \left(P_{l_i}\right(g\left(U_i^T{V}_j\right)\left)\right)\}+\frac{{\mathrm{\λ}}_u}{2}\underset{i=1}{\overset{M}{\Σ}}\left(\right||U_i-\underset{j\∈N_i}{\Σ}{T}_{ij}{U}_j|{|}_F^2)\\ +\frac{{\mathrm{\λ}}_v}{2}\underset{j=1}{\overset{N}{\Σ}}\left(\right||V_j-\underset{k\∈N_p}{\Σ}s(j,k\left)V_k\right|{|}_F^2)+\frac{\mathrm{\λ}}{2}\left(\right|\left|U\right|{|}_F^2+\left|\right|V\left|{|}_F^2\right)\end{array}$

Wherein,R_maxHighest score for scoring；I_ijFor indicator function, M is number of users, and N is item Mesh quantity, if project j is had scoring to record by user i, value is 1, and otherwise value is 0；U_iAnd V_jIt is that composition user is potential respectively Eigenmatrix and the row vector of project potential eigenmatrix U, V, and the dimension of U and V is less than number of users and the quantity of project；WithRepresent user social contact relation penalty coefficient and item label penalty coefficient respectively；||U||²With | | V | |²Be respectively U and Two norms of V；T_ijRepresenting user i and the degree of belief of user j, (j k) represents project j and the label phase reliability of project k to S；

D2, the list ordering study recommended models of establishment mosaic society information:

User potential eigenmatrix U and the gradient of project potential eigenmatrix V, and profit is obtained from loss function L (U, V) With gradient descent method, loss function is trained；Wherein, user potential eigenmatrix U and the ladder of project potential eigenmatrix V The acquisition methods of degree is as follows:

$\begin{array}{c}\frac{\∂L}{\∂U_i}=\underset{j=1}{\overset{N}{\Σ}}{I}_{ij}(\frac{\exp \left(g\right({U_i}^T{V}_j\left)\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(g\right(U_i^T{V}_k\left)\right)}-\frac{\exp \left(R_{ij}\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(R_{ik}\right)})g^{\′}\left(U_i^T{V}_j\right)V_j\\ +{\mathrm{\λ}}_u(U_i-\underset{j\∈N_i}{\Σ}{T}_{ij}{U}_j)-{\mathrm{\λ}}_u\underset{i\∈N_k}{\Σ}{T}_{ki}(U_k-\underset{j\∈N_k}{\Σ}{T}_{kj}{U}_j)+{\mathrm{\λU}}_i\end{array}$

$\begin{array}{c}\frac{\∂L}{\∂V_j}=\underset{i=1}{\overset{M}{\Σ}}{I}_{ij}(\frac{\exp \left(g\right({U_i}^T{V}_j\left)\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(g\right(U_i^T{V}_k\left)\right)}-\frac{\exp \left(R_{ij}\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(R_{ik}\right)})g^{\′}\left(U_i^T{V}_j\right)U_i\\ +{\mathrm{\λ}}_v(V_j-\underset{k\∈N_j}{\Σ}s(j,k)V_k)-{\mathrm{\λ}}_v\underset{j\∈N_l}{\Σ}s(l,j)(V_l-\underset{k\∈N_l}{\Σ}s(l,k\left)V_k\right)+{\mathrm{\λV}}_j\end{array}$

S4, prediction user are for the unknown purpose preference:

Prediction user's preference to project as follows:

R′_ij=U_i ^TV_j

Wherein R '_ijIt is user i preference that project j is predicted, U_iAnd V_jIt is user potential eigenmatrix U and item respectively Row vector in mesh potential eigenmatrix V.

The method that degree of belief oriented between user is done in step a2 normalized is:

Wherein, T_ukFor t_ukUser between oriented degree of belief initialization value, N_uThe all users set paid close attention to by user u.

In step b3, the label similarity between the project of described k nearest neighbor is normalized method as follows:

$s(j_1,j_2)=\frac{sim(j_1,j_2)}{\underset{v\∈N_j}{\Σ}sim(j_1,v)}$

Wherein N_jExpression project j₁K nearest neighbor set.

The method utilizing gradient descent method to be trained loss function is as follows:

E1., parameter d, λ are set_u,λ_v, λ, learning rate η and maximum iteration time maxIter；

E2. it is uniformly distributed eigenmatrix potential to user according to [0,1] and project potential eigenmatrix U, V are carried out initially Change, initialize error current and last error is all infinitely great, i.e. currError=preError=∞, iterations IterCount=0；

Renewal user potential eigenmatrix U and project potential eigenmatrix V the most as follows:

$U_i\←U_i-\mathrm{\η}\frac{\∂L}{\∂U_i},i=1,2,...,M$

$V_j\←V_j-\mathrm{\η}\frac{\∂L}{\∂V_j},j=1,2,...,N$

E4. iterations adds 1, calculates error current according to the loss function L (U, V) in step d1, if error current More than last error or iterations more than maximum iteration time, the most then meet the condition of convergence to step volume e5；Otherwise, in order First-order error is error current, continues e3；

E5. exporting the user potential eigenmatrix U restrained and project potential eigenmatrix V, training process terminates.

In step c3, function

K in step b3 is 10.

Described socialization information includes social network and socialized label.

In step d1, highest score R of scoring_maxIt is 5.

The beneficial effects of the present invention is: the present invention, compared with existing collaborative filtering method, has a following obvious advantage:

1) traditional probability matrix decomposition method with minimization score in predicting error as target, and the method for the invention with Matching whole entry sorting list is target.

2) present invention simultaneously using socialization's information of user and project i.e. users to trust degree and item label similarity as The bound term of matrix decomposition, makes the characteristic vector with the user of bigger degree of belief the most similar, meanwhile, has bigger label The characteristic vector of the project of similarity as close possible to, be more nearly the thinking habit of the mankind.

3, in stand-alone environment (CPU is double-core 3.0GHz, inside saves as 4G), at d=5, λ=0.001, λ_u=0.05, λ_v= 0.001, during item nearest neighbor number k=10, as a example by NDCG@1, method disclosed by the invention and Random, UserAvg, RankNet-MF, Bradley-TerryMF, ListRank-MF compare and are respectively increased 16.6%, and 6.8%, 12.4%, 12.5%, 0.9%.

Accompanying drawing explanation

Fig. 1 is the overview flow chart of the present invention.

Fig. 2 is users to trust matrix calculus flow chart.

Fig. 3 is item label similarity matrix calculation flow chart.

Fig. 4 is to utilize gradient descent method that the list ordering of mosaic society's information is learnt the training flow process of recommended models Figure.

Detailed description of the invention

Below in conjunction with the drawings and the specific embodiments, the present invention will be described:

Fig. 1 is the overview flow chart of the personalized recommendation method of a kind of mosaic society of present invention information.A kind of fusion society The personalized recommendation method of information can be changed, comprise the following steps:

S1, structure user-user trust matrix:

Oriented degree of belief between a1, acquisition user: in the known data base containing socialization's information, according in this data base Oriented degree of belief between concern Relation acquisition user between user in the social networks comprised.Wherein, socialization's information includes society Changing network and socialized label, socialization's information is mainly derived from online social networks, such as Semen Sojae Preparatum net etc.；Concern between user is closed If owner obtains from social networks, and this concern and the social networks being concerned relation are by user in social networks Actively close between the user for comprising in the social networks of the known data base containing socialization's information as shown in table 1 below of statement Note relation table, is that the value of 1 represents this row user these row user is had concern relation in table；

Relation table is paid close attention between table 1 user

User	1	2	3	4
					1		1	1	1
2			1
					3	1			1
4		1	1

Assume in social networks, co-exist in M user, if t_ukRepresent the user u degree of belief to user k, then t_ukThe biggest Represent that user k is the biggest to the power of influence of user's u interest；Otherwise, user k is the least to the power of influence of user u.Meanwhile, if user is u Pay close attention to more multi-user, then t_ukShould be along with minimizing；If user k is paid close attention to by more multi-user, then t_ukShould increase.More than based on Analyzing, between the user of the present invention, the acquisition methods of oriented degree of belief is as follows:

$t_{uk}=\sqrt{\frac{d^{-}\left(v_k\right)}{d^{+}\left(v_u\right)+d^{-}\left(v_k\right)}}$

Wherein, t_ukRepresent in data base's social networks oriented degree of belief, d between the user u user to user k^-(v_k) represent The quantity that user k is concerned, d⁺(v_u) represent that user u pays close attention to the quantity of user；

A2, normalized: then to degree of belief t oriented between user_ukDo normalized, obtain t_ukUser between have To degree of belief initialization value T_uk, it may be assumed that

Wherein, T_ukFor t_ukUser between oriented degree of belief initialization value, N_uThe institute paid close attention to by user u User is had to gather.

Oriented degree of belief initialization value T between described user_ukBuild user-user as matrix element and trust matrix P ∈ R^{M ×M}。

Obtain believing with the user-user shown in table 2 by the method that table 1 trusts matrix according to step S1, structure user-user The user-user appointing the element of degree matrix table to constitute trusts matrix, and the element in table 2 represents this row user letter to these row user The relation of appointing, 0 represents do not have trusting relationship, and trusting relationship is oriented, and has made normalized.

Table 2 user-user trusts matrix element table

S2, structure project-item label similarity matrix: idiographic flow is as it is shown on figure 3, comprise the following steps:

B1, the weight vectors of acquisition project: assume that one has N number of project, L label, if label occurrence number is the most, then This label is the most important, and the project of label for labelling is the most simultaneously, then its discrimination is the lowest, therefore weight w of label t in project i_itAdopt Each label labelling weight for each project is carried out by tf*idf weight.

Item label information in acquisition database to use tf*idf weight be each label labelling in each project Weight, and the element of the weight vectors of each project is constituted with tf*idf weighted value, the dimension of the label weight vectors of project is i.e. For the quantity of label in data base, the label weight vectors of project represents the label characteristics of project, and concrete grammar is as follows:

Wherein, N is the number of project, w in data base_jtThe weighted value of the label t in expression project j, (j t) represents item to tf Mesh j is marked with the number of times of label t, does not has during visible marking's number of times to be designated as 1, and df (t) represents the project number that label t is labeled, Markd label weight is not had automatically to be designated as 0；

So far, each project can be expressed as the weight vectors of L dimension.

Label similarity between b2, calculating project:

By calculating the label similarity between the cosine similarity acquisition project of the weight vectors of two projects, its computing formula As follows:

$sim(j_1,j_2)=cos({\stackrel{\→}{j}}_1,{\stackrel{\→}{j}}_2)=\frac{{\stackrel{\→}{j}}_1\·{\stackrel{\→}{j}}_2}{\left|\right|{\stackrel{\→}{j}}_1|{|}_F*|\left|{\stackrel{\→}{j}}_2\right|{|}_F}$

Wherein, sim (j₁,j₂) it is project j₁With project j₂Between label similarity,For the item obtained in step b1 Mesh j₁With project j₂Weight vectors；

B3, the label similarity chosen between k nearest neighbor normalization project: according to the label similarity options between project Purpose k nearest neighbor, the value of K is preferably 10 herein.Label similarity between the project of k nearest neighbor is normalized, obtains normalizing The label similarity between k nearest neighbor project after change, and the label similarity between the project outside k nearest neighbor is set to 0, at the beginning of obtaining The project of beginningization-item label similarity matrix Q ∈ R^N×N；Wherein the label similarity between the project of k nearest neighbor is normalized Method is as follows:

$s(j_1,j_2)=\frac{sim(j_1,j_2)}{\underset{v\∈N_j}{\Σ}sim(j_1,v)}$

Wherein N_jExpression project j₁K nearest neighbor set.

When neighbour's number k is set to 2, obtain the project shown in table 3-item label similarity matrix list of elements, unit in table 3 Element represents the similarity of this row project and this list of items, and similarity has made normalized.

Table 3 projects-item label similarity matrix list of elements

S3, structure training pattern: as shown in Figure 4, comprise the following steps:

C1, collection training dataset: randomly draw the 80% conduct training number of the known data base containing socialization's information According to collection；

C2, structure user-project rating matrix

Each user training data concentration comprised is for the scoring R of each project_ijGive 1 to R_maxScore value, generally Select the score value of 1-5 and build user-project rating matrix using these score values as matrix element；User as shown in table 4 below- Project grade form, is the part data set of extraction from Epinions data set.

Table 4 users-project grade form

	1	2	3	4	5	6
							1	2		4		4
2		3	5			5
							3	3		1	4
4		3		5		2

C3, calculate user for the first probability of certain score value

Assuming that one has M user, N number of project, R is the matrix of a M × N, R_ijRepresent the user i scoring to project j, R_ijIt is typically one from 1 to R_max, calculating user i for the scoring of project j by below equation is R_ijThe first probabilityWherein,For increasing function and the most satisfied for all xSelect functionParameter D is that user i comments the undue number of entry；

The first probabilityIt is the most the first general that the project that illustrates is discharged to primary probit in given sorted lists Rate.Obviously, score value R_ijThe biggest, then user is the biggest to the fancy grade of this project, and correspondingly the first probit is the highest, and more having can First can be discharged in the ranking.

In theory of information, generally weigh a probability distribution with cross entropy (Cross-Entropy) and given probability divides The similarity degree of cloth, cross entropy is the least, shows that two probability distribution are the most similar, especially, if two probability distribution complete Cause, then cross entropy is minimum.It is likewise possible to weigh the first probability distribution of prediction project sorted lists and known with cross entropy The similarity degree of the first probability distribution of entry sorting list.

C4, build and train mosaic society's information list ordering study recommended models

Simultaneously take account of the interest trusting each other between user the most similar, between the user that degree of belief is the biggest simultaneously The similarity of feature also tends to the biggest, and the power of influence between user also can be the biggest；On the other hand, the item label shown in table 5 As an important dimension of described project feature, there is the feature of short and small refine, can largely reflect a project Feature, therefore between project, characteristic vector between the highest then project of label similarity should be the most similar.

Table 5 item label table

	1	2	3	4	5	6	7
								1		1		1			1
2	1	1	1
								3			1	1		1
4	1				1		1
								5				1	1	1	1
6	1	1	1			1

Analyze based on above, by adding users to trust degree and the punishment of label similarity in existing loss function , obtain the list ordering study recommended models of mosaic society's information as follows.

D1, the loss function L (U, V) of structure mosaic society information, the list ordering study of mosaic society's information pushes away Recommend model to build based on this loss function

$\begin{array}{c}L(U,V)=\underset{i=1}{\overset{M}{\Σ}}\{-\underset{j=1}{\overset{N}{\Σ}}{I}_{ij}{P}_{l_i}\left(R_{ij}\right)\log \left(P_{l_i}\right(g\left(U_i^T{V}_j\right)\left)\right)\}+\frac{{\mathrm{\λ}}_u}{2}\underset{i=1}{\overset{M}{\Σ}}\left(\right||U_i-\underset{j\∈N_i}{\Σ}{T}_{ij}{U}_j|{|}_F^2)\\ +\frac{{\mathrm{\λ}}_v}{2}\underset{j=1}{\overset{N}{\Σ}}\left(\right||V_j-\underset{k\∈N_p}{\Σ}s(j,k\left)V_k\right|{|}_F^2)+\frac{\mathrm{\λ}}{2}\left(\right|\left|U\right|{|}_F^2+\left|\right|V\left|{|}_F^2\right)\end{array}---\left(1\right)$

Wherein,R_maxFor scoring highest score, generally 5；I_ijFor indicator function, M is number of users Amount, N is the number of entry, if project j is had scoring to record by user i, value is 1, and otherwise value is 0；U_iAnd V_jIt is user respectively Potential eigenmatrix and i-th row vector of project potential eigenmatrix U, V and jth row vector, and the dimension of U and V is the least In number of users and the quantity of project；WithRepresent that user social contact relation penalty coefficient and item label punishment are respectively Number；||U||²With | | V | |²It is respectively two norms of U and V；T_ijRepresenting user i and the degree of belief of user j, (j k) represents project to S The label phase reliability of j and project k；

D2, the list ordering study recommended models of establishment mosaic society information

User potential eigenmatrix U and the gradient of project potential eigenmatrix V, and profit is obtained from loss function L (U, V) With gradient descent method, loss function is trained；Wherein, user potential eigenmatrix U and the ladder of project potential eigenmatrix V The acquisition methods of degree is as follows:

$\begin{array}{c}\frac{\∂L}{\∂U_i}=\underset{j=1}{\overset{N}{\Σ}}{I}_{ij}(\frac{\exp \left(g\right({U_i}^T{V}_j\left)\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(g\right({U_i}^T{V}_k\left)\right)}-\frac{\exp \left(R_{ij}\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(R_{ik}\right)})g^{\′}\left(U_i^T{V}_j\right)V_j\\ +{\mathrm{\λ}}_u(U_i-\underset{j\∈N_i}{\Σ}{T}_{ij}{U}_j)-{\mathrm{\λ}}_u\underset{i\∈N_k}{\Σ}{T}_{ki}(U_k-\underset{j\∈N_k}{\Σ}{T}_{kj}{U}_j)+{\mathrm{\λU}}_i\end{array}---\left(2\right)$

$\begin{array}{c}\frac{\∂L}{\∂V_j}=\underset{i=1}{\overset{M}{\Σ}}{I}_{ij}(\frac{\exp \left(g\right({U_i}^T{V}_j\left)\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(g\right(U_i^T{V}_k\left)\right)}-\frac{\exp \left(R_{ij}\right)}{\underset{k=1}{\overset{N}{\Σ}}{I}_{ik}\exp \left(R_{ik}\right)})g^{\′}\left(U_i^T{V}_j\right)U_i\\ +{\mathrm{\λ}}_v(V_j-\underset{k\∈N_j}{\Σ}s(j,k)V_k)-{\mathrm{\λ}}_v\underset{j\∈N_l}{\Σ}s(l,j)(V_l-\underset{k\∈N_l}{\Σ}s(l,k\left)V_k\right)+{\mathrm{\λV}}_j\end{array}---\left(3\right)$

The method utilizing gradient descent method to be trained loss function is as follows:

User-user is trusted matrix, project-item label similarity matrix and user-project rating matrix as defeated Enter,

E1., parameter d, λ are set_u,λ_v, λ, learning rate η and maximum iteration time maxIter, i.e. initiation parameter；

E2. it is uniformly distributed eigenmatrix potential to user according to [0,1] and project potential eigenmatrix U, V are carried out initially Change, initialize error current and last error is all infinitely great, i.e. currError=preError=∞, iterations IterCount=0；

E3. gradient is calculated, as follows by U according to formula (2) and (3)_iAnd V_jRenewal realize to user dive Renewal at eigenmatrix U and project potential eigenmatrix V:

$U_i\←U_i-\mathrm{\η}\frac{\∂L}{\∂U_i},i=1,2,...,M---\left(4\right)$

$V_j\←V_j-\mathrm{\η}\frac{\∂L}{\∂V_j},j=1,2,...,N---\left(5\right)$

E4. iterations adds 1, calculates error current currError according to formula (1).If judging, error current is more than Last error i.e. currError ＞ preError or iterations are more than maximum iteration time i.e. iterCount ＞ maxIter Then meet the condition of convergence to step 5；Otherwise, making preError is currError, continues e3；

E5. exporting the user potential eigenmatrix U restrained and project potential eigenmatrix V, training process terminates.

Obtain the potential eigenmatrix of user and the project potential eigenmatrix list of elements is distinguished the most as shown in table 6 and table 7, Qi Zhongcan Number is set to d=5, λ=0.001, λ_u=0.05, λ_v=0.001, η=0.01, maximum iteration time is 500.In table 6-7 each Row represents characteristic vector U of this row user (project)_i(V_j)。

The table 6 user potential eigenmatrix list of elements

	1	2	3	4	5
						1	0.557	0.097	0.55	0.516	-0.074
2	0.58	-0.402	0.55	0.364	0.186
						3	-0.051	-0.561	0.456	0.585	0.036
4	-0.235	-0.489	0.403	0.68	-0.048

The table 7 project potential eigenmatrix list of elements

	1	2	3	4	5
						1	-0.88	-1.26	-0.32	-0.19	0.43
2	-1.04	0.74	-0.6	0.025	-0.586
						3	1.295	0.903	-0.013	-0.421	0.151
4	-0.262	-1.16	0.691	1.076	0.074
						5	0.286	0.455	0.315	0.362	-0.228
6	1.04	0.322	-0.025	-0.705	0.223

S4, prediction user are for the unknown purpose preference

Prediction user's preference to project as follows:

R′_ij=U_i ^TV_j

Wherein R '_ijIt is user i preference that project j is predicted, U_iAnd V_iIt is by the row of mosaic society's information respectively I-th row vector of the list sorting study recommended models user potential eigenmatrix U that obtains of training and project potential eigenmatrix V and Jth row vector.

As a example by user 1, project 2 that it is not marked by available user 1, project 4, the preference value of project 6 be respectively- 0.782,0.67,0.22, it is therefore project 4, project 6, project 2 according to the recommended project list of preference value descending user 1, In like manner, the project interested of other user can be recommended the bulleted list obtaining recommending by similar method.So far, this The method of bright proposition completes a recommended flowsheet.

Contrast on effect:

Apply the present invention to truthful data collection Baidu film and recommend contest data set, be called for short BaiduMovie data set. BaiduMovie data set by company of Baidu disclosed in the film commending system algorithm creative contest that in May, 2013 holds, This data set mainly has following information: user-film scoring record, user pay close attention to relation, film label information.Data set wraps Containing 9722 users, 1256998 scoring records to 7889 projects, the density of score data is 1.64%, and these are used simultaneously Having 7898 concern relations between family, the density paying close attention to relation is 0.0083%, have 1121 labels, average each project have by Marked 10 labels.

Apply the method for the invention in BaiduMovie data set, for making experimental result have more cogency, use 5-rolls over cross validation method, and data set is divided into 5 parts at random and averagely, choose successively wherein 1 part as test set, remain 4 Number, according to as training set, is trained 5 models and obtains 5 experimental results, chooses the meansigma methods of 5 results as final experiment Result.Unlike score in predicting problem, the present invention selects NDCG value as the index of evaluation sequence quality, and user gathers Q Before in, the ranking results NDCG computing formula of K project is as follows:

$NDCG(Q,k)=\frac{1}{\left|Q\right|}\underset{u\∈Q}{\Σ}{Z}_u\underset{p=1}{\overset{k}{\Σ}}\frac{2^{R(u,p)}-1}{log(1+p)}$

Wherein Z_uIt is normalization factor so that optimum sequence NDCG value is 1, and (u p) represents in user's u optimal sequencing R The score value of pth position.

Experiment have chosen 5 existing recommendation methods, including recommending method (Random), UserAvg and two at random Method (RankNet-MF and Bradley-TerryMF) based on Pair-wise, a method based on List-wise (ListRank-MF), the experimental result contrast of 6 methods is as shown in table 8.

Table 8 accuracy rate contrast table

Test result indicate that, the method for the present invention recommends accuracy rate to have substantially compared with other recommendation method based on sequence study Raising.

Above content is to combine concrete optimal technical scheme further description made for the present invention, it is impossible to assert Being embodied as of the present invention is confined to these explanations.For general technical staff of the technical field of the invention, On the premise of present inventive concept, it is also possible to make some simple deduction or replace, all should be considered as belonging to the present invention's Protection domain.

Claims (8)

1. the personalized recommendation method of mosaic society's information, it is characterised in that comprise the following steps:

S1, structure user-user trust matrix:

Oriented degree of belief between a1, acquisition user: in the known data base containing socialization's information, wraps according in described data base Oriented degree of belief, the acquisition of oriented degree of belief between described user between concern Relation acquisition user between user in the social networks contained Method is as follows:

$t_{uk}=\sqrt{\frac{d^{-}\left(v_k\right)}{d^{+}\left(v_u\right)+d^{-}\left(v_k\right)}}$

Wherein, t_ukRepresent in data base's social networks oriented degree of belief, d between the user u user to user k^-(v_k) represent user k The quantity being concerned, d⁺(v_u) represent that user u pays close attention to the quantity of user；

A2, normalized: then to degree of belief t oriented between user_ukDo normalized, obtain t_ukUser between oriented trust Degree initialization value T_uk, and built user-user trust matrix by oriented degree of belief initialization value between described user；

S2, structure project-item label similarity matrix:

B1, the weight vectors of acquisition project: the item label information in acquisition database uses tf*idf weight to be each project In each label labelling weight, and constituted the element of the weight vectors of each project with tf*idf weighted value, the label of project The dimension of weight vectors is the quantity of label in data base, and the label weight vectors of project represents the label characteristics of project, Concrete grammar is as follows:Wherein, N is the number of project, w in data base_jtIn expression project j The weighted value of label t, (j t) represents that project j is marked with the number of times of label t, does not has during visible marking's number of times to be designated as 1, df (t) tf Represent the labeled project number of label t, do not have markd label weight to be automatically designated as 0；

Label similarity between b2, calculating project:

By calculating the label similarity between the cosine similarity acquisition project of the weight vectors of two projects, its computing formula is such as Under:

$sim(j_1,j_2)=cos({\stackrel{\→}{j}}_1,{\stackrel{\→}{j}}_2)=\frac{{\stackrel{\→}{j}}_1\·{\stackrel{\→}{j}}_2}{\left|\right|{\stackrel{\→}{j}}_1|{|}_F*|\left|{\stackrel{\→}{j}}_2\right|{|}_F}$

Wherein, sim (j₁,j₂) it is project j₁With project j₂Between label similarity,For project j obtained in step b1₁ With project j₂Weight vectors；

B3, the label similarity chosen between k nearest neighbor normalization project: according to the label similarity options purpose between project K nearest neighbor, is normalized the label similarity between the project of k nearest neighbor, obtains the label between the k nearest neighbor project after normalization Similarity, and the label similarity between the project outside k nearest neighbor is set to 0, obtain initialized project-item label similarity Matrix；

S3, structure and training pattern:

C1, gather training dataset: randomly draw the 80% of the known data base containing socialization's information as training dataset；

C2, structure user-project rating matrix:

Training data is concentrated each user comprised mark for each project give the score value of 1-5 and make with these score values User-project rating matrix is built for matrix element；

C3, calculate user for the first probability of certain score value:

Calculating user i for the scoring of project j by below equation is R_ijThe first probability Wherein,For increasing function and the most satisfied for all xParameter D is that user i commented The number of entry divided；

C4, build and train mosaic society's information list ordering study recommended models, including:

D1, the loss function L (U, V) of structure mosaic society information:

$\begin{array}{c}L(U,V)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\{-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{I}_{ij}{P}_{l_i}\left(R_{ij}\right)\log \left(P_{l_i}\right(g\left(U_i^T{V}_j\right)\left)\right)\}+\frac{{\mathrm{\λ}}_u}{2}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\left(\right||U_i-\underset{j\∈N_i}{\mathrm{\Σ}}{T}_{ij}{U}_j|{|}_F^2)\\ +\frac{{\mathrm{\λ}}_v}{2}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}\left(\right||V_j-\underset{k\∈N_p}{\mathrm{\Σ}}s(j,k\left)V_k\right|{|}_F^2)+\frac{\mathrm{\λ}}{2}\left(\right|\left|U\right|{|}_F^2+\left|\right|V\left|{|}_F^2\right)\end{array}$

Wherein,R_maxHighest score for scoring；I_ijFor indicator function, M is number of users, and N is item number Amount, if project j is had scoring to record by user i, value is 1, and otherwise value is 0；U_iAnd V_jIt is the potential eigenmatrix of user respectively The row vector of eigenmatrix U, V potential with project, and the dimension of U and V is less than number of users and the quantity of project；WithPoint Biao Shi user social contact relation penalty coefficient and item label penalty coefficient；||U||²With | | V | |²It is respectively two norms of U and V； T_ijRepresenting user i and the degree of belief of user j, (j k) represents project j and the label phase reliability of project k to S；

D2, the list ordering study recommended models of establishment mosaic society information:

From loss function L (U, V), obtain user potential eigenmatrix U and the gradient of project potential eigenmatrix V, and utilize ladder Loss function is trained by degree descent method；Wherein, the gradient of user potential eigenmatrix U and project potential eigenmatrix V Acquisition methods is as follows:

$\begin{array}{c}\frac{\∂L}{\∂U_i}=\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{I}_{ij}(\frac{\exp \left(g\right({U_i}^T{V}_j))}{\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{I}_{ik}\exp \left(g\right(U_i^T{V}_k))}-\frac{\exp \left(R_{ij}\right)}{\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{I}_{ik}\exp \left(R_{ik}\right)})g^{\′}\left(U_i^T{V}_j\right)V_j\\ +{\mathrm{\λ}}_u(U_i-\underset{j\∈N_i}{\mathrm{\Σ}}{T}_{ij}{U}_j)-{\mathrm{\λ}}_u\underset{i\∈N_k}{\mathrm{\Σ}}{T}_{ki}(U_k-\underset{j\∈N_k}{\mathrm{\Σ}}{T}_{kj}{U}_j)+{\mathrm{\λU}}_i\end{array}$

$\begin{array}{c}\frac{\∂L}{\∂V_j}=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{I}_{ij}(\frac{\exp \left(g\right({U_i}^T{V}_j\left)\right)}{\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{I}_{ik}\exp \left(g\right(U_i^T{V}_k\left)\right)}-\frac{\exp \left(R_{ij}\right)}{\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{I}_{ik}\exp \left(R_{ik}\right)})g^{\′}\left(U_i^T{V}_j\right)U_i\\ +{\mathrm{\λ}}_v(V_j-\underset{k\∈N_j}{\mathrm{\Σ}}s(j,k\left)V_k\right)-{\mathrm{\λ}}_v\underset{j\∈N_l}{\mathrm{\Σ}}s(l,j)(V_l-\underset{k\∈N_l}{\mathrm{\Σ}}s(l,k\left)V_k\right)+{\mathrm{\λV}}_j\end{array}$

S4, prediction user are for the unknown purpose preference:

Prediction user's preference to project as follows:

R’_ij=U_i ^TV_j

Wherein R'_ijIt is user i preference that project j is predicted, U_iAnd V_jIt is by the list ordering of mosaic society's information respectively Study recommended models trains the user potential eigenmatrix U and i-th row vector of project potential eigenmatrix V and jth row obtained Vector.

The personalized recommendation method of a kind of mosaic society the most according to claim 1 information, it is characterised in that step a2 In degree of belief oriented between user done the method for normalized be:

Wherein, T_ukFor t_ukUser between oriented degree of belief initialization value, N_uThe all users set paid close attention to by user u.

The personalized recommendation method of a kind of mosaic society the most according to claim 1 information, it is characterised in that step b3 In, the label similarity between the project of described k nearest neighbor is normalized method as follows:

$s(j_1,j_2)=\frac{sim(j_1,j_2)}{\underset{v\∈N_j}{\Σ}sim(j_1,v)}$

Wherein N_jExpression project j₁K nearest neighbor set.

The personalized recommendation method of a kind of mosaic society the most according to claim 1 information, it is characterised in that utilize ladder The method that loss function is trained by degree descent method is as follows:

E1., parameter d, λ are set_u,λ_v, λ, learning rate η and maximum iteration time maxIter；

E2. it is uniformly distributed eigenmatrix potential to user according to [0,1] and project potential eigenmatrix U, V initialize, just Beginningization error current and last error are all infinitely great, i.e. currError=preError=∞, iterations IterCount=0；

Renewal user potential eigenmatrix U and project potential eigenmatrix V the most as follows:

$U_i\←U_i-\mathrm{\η}\frac{\∂L}{\∂U_i},i=1,2,...,M$

$V_j\←V_j-\mathrm{\η}\frac{\∂L}{\∂V_j},j=1,2,...,N$

E4. iterations adds 1, calculates error current, if error current is more than according to the loss function L (U, V) in step d1 Last error or iterations more than maximum iteration time, the most then meet the condition of convergence to step volume e5；Otherwise, the order last time Error is error current, continues e3；

E5. exporting the user potential eigenmatrix U restrained and project potential eigenmatrix V, training process terminates.

The personalized recommendation method of a kind of mosaic society the most according to claim 1 information, it is characterised in that step c3 In, function

The personalized recommendation method of a kind of mosaic society the most according to claim 1 information, it is characterised in that step b3 In K be 10.

The personalized recommendation method of a kind of mosaic society the most according to claim 1 information, it is characterised in that described society Information of changing includes social network and socialized label.

The personalized recommendation method of a kind of mosaic society the most according to claim 1 information, it is characterised in that step d1 In, highest score R of scoring_maxIt is 5.