CN101853470A - Collaborative filtering method based on socialized label - Google Patents

Abstract

The invention discloses a collaborative filtering method based on socialized labels, which comprises the following steps that: (1) first, a tripartite graph is used to build models for three different node, i.e. a user, an article and the socialized label, and random walk algorithm is applied to each user to recommend top-N personalized articles; (2) in order to solve the problem of the sparsity of socialized labels (i.e. articles are always labeled by only a small number of labels) and noise brought by the subjective factors of the user, the invention provides a method for expanding the labeling of the articles with a lasso logistic regression model, i.e. labels related to the semantics are increased for each article, and the labels with noise are removed; and (3) the weight of the labels in the recommendation process is regulated. The collaborative filtering method based on socialized labels organically integrates the semantic information of the socialized labels to the description of the articles, uses the lasso logistic regression model to expand the labels, solves the sparsity and noise problems of the socialized labels so as to greatly improve the precision and the performance of a personalized recommendation system.

Description

A kind of collaborative filtering method based on socialized label

Technical field

The present invention relates to the personalized recommendation field, relate in particular to a kind of collaborative filtering method based on socialized label.

Background technology

Along with the fast development of network and multimedia technology, the amount of images on the internet is explosive increase.According to statistics, 2008, Google index Web webpage scale reached 1,000,000,000,000, and wherein view data is above tens.Present in the time of magnanimity information, make the user be difficult to therefrom find own interested part on the one hand, the information that also makes a large amount of rare people make inquiries on the other hand becomes " the dark information " in the network, can't be obtained by the general user.Personalized recommendation system utilizes the potential interested object of existing selection course or similarity relation excavation user, and then recommends by setting up the binary relation between user and the information products, and its essence is exactly information filtering.Personalized recommendation system not only has important use and is worth in social economy, and is a problem in science that is worth very much research.In fact, it is to solve information overload problem one of the most effective instrument at present.

Collaborative filtering recommending (collaborative filtering recommendation) technology is one of technology the most successful in the commending system, has been widely used in film and has recommended (Netflix), book recommendation fields such as (Amazon).Directly analyzing content with traditional content-based filtration recommends different, collaborative filtering analysis user interest, in customer group, find similar (interest) user to designated user, comprehensive these similar users are to the evaluation of a certain information, and the formation system is to the fancy grade prediction of this designated user to this information.Yet, though extensive studies and application are arranged, the collaborative filtering technology still is difficult to overcome the sparse property of data problem, promptly in existing commending system, the user often only estimated minimum a part of article, and making like this may be inaccurate based on the similarity between the resulting user of evaluation of user.The problem of sparse property is in particular in following two aspects:

1) cold start problem (cold start) promptly when new user or new article join commending system, owing to lack evaluation information, can not find similar user or similar article for it, thereby can't recommend.

2) similar transitivity problem (Neighbor transitivity), promptly user A has similar preference to B, and user B also has similar preference with C, and A also has similar preference to C so, even A did not estimate identical article with C.Therefore traditional collaborative filtering recommending algorithm can't be excavated potential similar users or similar article.

In order to solve above shortcoming, we propose 1) semantic information of using socialized label to provide excavates the similarity of article, thus alleviate because of lacking the bottleneck that user's score information can't accurate Calculation article similarity; 2) use random walk (Random walk) algorithm to solve similar transitivity problem based on graph model.

Further, because socialized label is produced by the user, these total number of labels amounts are very big, and mistake and noise wherein appear unavoidably, simultaneously, each user is accustomed to using a part of label seldom to mark article, makes the semanteme of article not expressed fully by label.In order to address this problem, we propose a kind of model based on lasso logistic regression and pass through related semantic similar label, make that the semanteme of article can be by more comprehensive and accurate expression.

Summary of the invention

The objective of the invention is to overcome the deficiency of existing recommended technology, a kind of collaborative filtering method based on socialized label is provided.

Collaborative filtering method based on socialized label comprises the steps:

1) at first use tripartite graph that user, article, these three kinds of different node of socialized label are carried out modeling, and application random walk algorithm is each user-customized recommended top-N article;

2) expansion of using lasso logistic regression model to remove article are carried out label, promptly to each article, increase and its semantic relevant label, and remove the label that those have noise, to solve noise and the sparse property problem that exists in the socialized label, promptly article have only seldom a part of label for labelling usually;

3) regulate the weight of label in recommendation process.

Described by tripartite graph to user, article and label modeling and use the random walk algorithm and be for each user recommends the step of article:

1) use the tripartite graph modeling, then this tripartite graph can be expressed as: G={U, and I, T, E}, wherein U, I, T represent user, article, label respectively, and E represents the relation between them, and then this tripartite graph can be expressed as by following adjacency matrix:

$M=\left(\begin{array}{ccc}0& \mathrm{UI}& 0\\ {\mathrm{UI}}^T& 0& \mathrm{IT}\\ 0& {\mathrm{IT}}^T& 0\end{array}\right)$

Wherein UI represents the rating matrix of User to Item, UI _IjBe user u _iTo article i _jScoring; IT is article and label matrix, IT _JlExpression article i _jBy label t _lThe number of times of mark;

2) on this tripartite graph, use the random walk algorithm to recommend.

Transition matrix in the random walk algorithm is defined as follows:

$S_{\mathrm{ij}}=\frac{M_{\mathrm{ij}}}{\underset{i}{\mathrm{\Σ}}{M}_{\mathrm{ij}}}=\left(\begin{array}{ccc}0& P_1& 0\\ P_2& 0& Q_1\\ 0& Q_2& 0\end{array}\right)$

During the random walk algorithm initialization, each node gives initial value R (0).According to transition matrix, iteration " is walked " to other node subsequently, and until convergence, simultaneously, in order to embody individuation principle, in iterative process, the random walk algorithm moves towards the node that those react user preferences with probability α, and the random walk algorithm patternization is defined as follows:

$\left\{\begin{array}{c}R\left(0\right)=\frac{1}{\left|M\right|}\\ R(t+1)=(1-\mathrm{\α})\mathrm{SR}\left(t\right)+\mathrm{\αp}\end{array}\right.$

Wherein R (t) is the recommendation of random algorithm to each node, and p represents the user preference node, in tripartite graph, comprises user's node, article node, and label node, and it is defined as follows:

p＝[p ^user?p ^item?p ^tag] ^T

Wherein, p ^UserBe user self, p ^ItemBe those article nodes of being marked by the user, p ^TagIt is the used label node of user.

In actual computation, above-mentioned equation can be recommended out following form:

$\left\{\begin{array}{c}R{(t+1)}^{\mathrm{user}}=(1-\mathrm{\α})P_{1}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{user}}\\ R{(t+1)}^{\mathrm{tag}}=(1-\mathrm{\α})Q_{2}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{tag}}\\ R{(t+1)}^{\mathrm{item}}=(1-\mathrm{\α})(P_{2}R{\left(t\right)}^{\mathrm{user}}+Q_{1}R{\left(t\right)}^{\mathrm{tag}})+\mathrm{\α}{p}^{\mathrm{item}}\end{array}\right.$

R (t+1) wherein ^User, R (t+1) ^Tag, R (t+1) ^ItemRepresent respectively that to the user recommendation of label and article is to R (t+1) ^ItemOrdering is selected the top n article of its intermediate value maximum and is recommended the user.

The expansion that described logical lasso logistic regression model removes article are carried out label, thus the step that solves sparse property of label and noise problem is:

1) for each label b, the mark training sample, promptly to each article a, if it is by this label for labelling mistake, then it is positive sample, that is: y _Ab=1, otherwise become negative sample, y _Ab=-1;

2) following maximum likelihood function is found the solution in training in whole sample space, obtains regression coefficient

:

Maximum likelihood function wherein

For:

$L({\mathrm{\β}}_{b^{*}},{\mathrm{\γ}}_b)=-\underset{a=1}{\overset{m}{\mathrm{\Σ}}}1+\exp (-y_{\mathrm{ab}}({\mathrm{\β}}_{b^{*}}^{T}I{T}_{a^{*}}+{\mathrm{\γ}}_b))$

Regression coefficient β _BlBe the degree of correlation semantically of label l and label b, β _Bl＞0 expression positive correlation, β _Bl＜0 expression negative correlation, and β _Bl=0 expression is uncorrelated;

3) for article a, if marked, then be used for describing the semanteme of article with the maximally related K of a label b label by label b, thus the label of expansion article, and K the label least relevant with label b is used for removing the noise of describing in the article tag.

The weight step of described adjusting label in recommendation process is: in the random walk algorithm, to user's weight different with the increase of socialized label item of marking, be respectively δ and 2-δ, then this random walk algorithm is as follows:

$\left\{\begin{array}{c}R{(t+1)}^{\mathrm{user}}=(1-\mathrm{\α})P_{1}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{user}}\\ R{(t+1)}^{\mathrm{tag}}=(1-\mathrm{\α})Q_{2}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{tag}}\\ R{(t+1)}^{\mathrm{item}}=(1-\mathrm{\α})({\mathrm{\δP}}_{2}R{\left(t\right)}^{\mathrm{user}}+{(2-\mathrm{\δ})Q}_{1}R{\left(t\right)}^{\mathrm{tag}})+\mathrm{\α}{p}^{\mathrm{item}}\end{array}\right.$

R (t+1) wherein ^User, R (t+1) ^Tag, R (t+1) ^ItemRepresent to the user recommendation of label and article, p respectively ^UserBe user self, p ^ItemBe those article nodes of being marked by the user, p ^TagIt is the used label node of user.

Collaborative filtering method based on socialized label proposed by the invention organically combines the semantic information that socialized label is described for article, and the expansion of using lasso logistic regression model to remove article are carried out label, solve the sparse property and the noise problem of socialized label, thereby significantly promote the degree of accuracy and the performance of personalized recommendation system.

Description of drawings

Fig. 1 is the present invention and the curve map of baseline collaborative filtering on " degree of accuracy ";

Fig. 2 is the present invention and the curve map of baseline collaborative filtering on " recall rate ";

Fig. 3 is the present invention and the histogram of baseline collaborative filtering on Macro DOA.

Embodiment

Collaborative filtering method based on socialized label comprises the steps:

1) at first use tripartite graph that user, article, these three kinds of different node of socialized label are carried out modeling, and application random walk algorithm is each user-customized recommended top-N article;

2) expansion of using lasso logistic regression model to remove article are carried out label, promptly to each article, increase and its semantic relevant label, and remove the label that those have noise, to solve noise and the sparse property problem that exists in the socialized label, promptly article have only seldom a part of label for labelling usually;

3) regulate the weight of label in recommendation process.

Described by tripartite graph to user, article and label modeling and use the random walk algorithm and be for each user recommends the step of article:

1) use the tripartite graph modeling, then this tripartite graph can be expressed as: G={U, and I, T, E}, wherein U, I, T represent user, article, label respectively, and E represents the relation between them, and then this tripartite graph can be expressed as by following adjacency matrix:

$M=\left(\begin{array}{ccc}0& \mathrm{UI}& 0\\ {\mathrm{UI}}^T& 0& \mathrm{IT}\\ 0& {\mathrm{IT}}^T& 0\end{array}\right)$

Wherein UI represents the rating matrix of User to Item, UI _IjBe user u _iTo article i _jScoring; IT is article and label matrix, IT _JlExpression article i _jBy label t _lThe number of times of mark;

2) on this tripartite graph, use the random walk algorithm to recommend.

Transition matrix in the random walk algorithm is defined as follows:

$S_{\mathrm{ij}}=\frac{M_{\mathrm{ij}}}{\underset{i}{\mathrm{\Σ}}{M}_{\mathrm{ij}}}=\left(\begin{array}{ccc}0& P_1& 0\\ P_2& 0& Q_1\\ 0& Q_2& 0\end{array}\right)$

During the random walk algorithm initialization, each node gives initial value R (0).According to transition matrix, iteration " is walked " to other node subsequently, and until convergence, simultaneously, in order to embody individuation principle, in iterative process, the random walk algorithm moves towards the node that those react user preferences with probability α, and the random walk algorithm patternization is defined as follows:

$\left\{\begin{array}{c}R\left(0\right)=\frac{1}{\left|M\right|}\\ R(t+1)=(1-\mathrm{\α})\mathrm{SR}\left(t\right)+\mathrm{\αp}\end{array}\right.$

Wherein R (t) is the recommendation of random algorithm to each node, and p represents the user preference node, in tripartite graph, comprises user's node, article node, and label node, and it is defined as follows:

p＝[p ^user?p ^item?p ^tag] ^T

Wherein, p ^UserBe user self, p ^ItemBe those article nodes of being marked by the user, p ^TagIt is the used label node of user.

In actual computation, above-mentioned equation can be recommended out following form:

$\left\{\begin{array}{c}R{(t+1)}^{\mathrm{user}}=(1-\mathrm{\α})P_{1}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{user}}\\ R{(t+1)}^{\mathrm{tag}}=(1-\mathrm{\α})Q_{2}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{tag}}\\ R{(t+1)}^{\mathrm{item}}=(1-\mathrm{\α})(P_{2}R{\left(t\right)}^{\mathrm{user}}+Q_{1}R{\left(t\right)}^{\mathrm{tag}})+\mathrm{\α}{p}^{\mathrm{item}}\end{array}\right.$

R (t+1) wherein ^User, R (t+1) ^Tag, R (t+1) ^ItemRepresent respectively that to the user recommendation of label and article is to R (t+1) ^ItemOrdering is selected the top n article of its intermediate value maximum and is recommended the user.According to thumb rule, α value 0.15; The user on average needs iteration to reach convergence state 35 times.

The expansion that described logical lasso logistic regression model removes article are carried out label, thus the step that solves sparse property of label and noise problem is:

1) for each label b, the mark training sample, promptly to each article a, if it is by this label for labelling mistake, then it is positive sample, that is: y _Ab=1, otherwise become negative sample, y _Ab=-1;

2) following maximum likelihood function is found the solution in training in whole sample space, obtains regression coefficient

:

Maximum likelihood function wherein

For:

$L({\mathrm{\β}}_{b^{*}},{\mathrm{\γ}}_b)=-\underset{a=1}{\overset{m}{\mathrm{\Σ}}}1+\exp (-y_{\mathrm{ab}}({\mathrm{\β}}_{b^{*}}^{T}I{T}_{a^{*}}+{\mathrm{\γ}}_b))$

Regression coefficient β _BlBe the degree of correlation semantically of label l and label b, β _Bl＞0 expression positive correlation, β _Bl＜0 expression negative correlation, and β _Bl=0 expression is uncorrelated;

Formula Derivation as follows:

Each article a is expressed as the vector (z is a total number of labels) of a z dimension:

IT wherein _AlIt is label l is used for marking article a by the user total degree.Then article a can be represented as:

$y_{\mathrm{ab}}=\mathrm{\Θ}\left({{\mathrm{\β}}_{b^{*}}}^T{\mathrm{IT}}_{a^{*}}\right)=\mathrm{\Θ}\left(\underset{l=1}{\overset{z}{\mathrm{\Σ}}}{\mathrm{\β}}_{\mathrm{bl}}{\mathrm{IT}}_{\mathrm{al}}\right)$

Wherein

And

Be regression coefficient.

We pay close attention to following probability model:

$p(y_{\mathrm{ab}}=+1|{\mathrm{\β}}_{b^{*}},{\mathrm{IT}}_{a^{*}})=\frac{1}{1+\exp (-{\mathrm{\β}}_{b^{*}}^T{\mathrm{IT}}_{a^{*}}-{\mathrm{\γ}}_b)}$

γ wherein _bBe intercept.A kind of right

Method of estimation can transform finding the solution to following maximum likelihood problem:

Likelihood function wherein

Be defined as follows:

$L({\mathrm{\β}}_{b^{*}},{\mathrm{\γ}}_b)=-\underset{a=1}{\overset{m}{\mathrm{\Σ}}}1+\exp (-y_{\mathrm{ab}}({\mathrm{\β}}_{b^{*}}^{T}I{T}_{a^{*}}+{\mathrm{\γ}}_b))$

In solution procedure, in order to prevent " cross and adapt to " (over-fitting) problem, we introduce l ₁The penalty term of-norm (such as: make

Satisfy the Laplace prior distribution) thus the sparse expression characteristics that assurance is found the solution.Therefore the maximum likelihood problem is converted into finding the solution of following equation:

3) for article a, if marked, then be used for describing the semanteme of article with the maximally related K of a label b label by label b, thus the label of expansion article, and K the label least relevant with label b is used for removing the noise of describing in the article tag.

Form 1 has provided and has used lasso logistic regression model that label is carried out a related object lesson: given label war, go out maximally related label by this model training, then be noted as the film of war, also may be by following label for labelling, as Vietnam (Vietnam War), the Jacket (about the film of Gulfwar), WW2 (World War II), no happy end (may be used for describing the final result of war) or the like.By this method, article are semantically being described more fully, thus can the better utilization socialized label.

Form 1 use that lasso logistic regression model training goes out with " war " maximally related label

??Tag?id	??β	??Tag
			??2270	??12.08	??Vietnam
??3930	??10.24	??the?Jacket
			??3411	??9.96	??SNL?alums
??375	??9.81	??no?happy?end
			??2298	??8.67	??WW2

The weight step of described adjusting label in recommendation process is: in the random walk algorithm, to user's weight different with the increase of socialized label item of marking, be respectively δ and 2-δ, then this random walk algorithm is as follows:

$\left\{\begin{array}{c}R{(t+1)}^{\mathrm{user}}=(1-\mathrm{\α})P_{1}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{user}}\\ R{(t+1)}^{\mathrm{tag}}=(1-\mathrm{\α})Q_{2}R{\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{tag}}\\ R{(t+1)}^{\mathrm{item}}=(1-\mathrm{\α})(\mathrm{\δ}{P}_{2}R{\left(t\right)}^{\mathrm{user}}+{(2-\mathrm{\δ})Q}_{1}R{\left(t\right)}^{\mathrm{tag}})+\mathrm{\α}{p}^{\mathrm{item}}\end{array}\right.$

R (t+1) wherein ^User, R (t+1) ^Tag, R (t+1) ^ItemRepresent to the user recommendation of label and article, p respectively ^UserBe user self, p ^ItemBe those article nodes of being marked by the user, p ^TagIt is the used label node of user.

Experimental result finds that the effect of δ=0.4 o'clock recommendation is best.

Embodiment

The experiment data set take from MovieLens ( Http:// www.grouplens.org/node/73) middle 10M data, we therefrom select 861 users, and 5003 article and 6147 labels are so that each user can mark simultaneously at least and mark 3 different article.For this part data set, we in 80%～20% ratio random division 5 times, generate 5 groups of different training sets and test set with it.Final experimental result is the mean value of 5 groups of experimental results.

For the validity of algorithm proposed by the invention is described, we also will use three kinds of traditional collaborative filterings to do the contrast experiment, be respectively: based on user's collaborative filtering (U-CF), based on the collaborative filtering (I-CF) of article and at a kind of random walk algorithm (ItemRank) that does not use social label.We experimentize on each step to method proposed by the invention, that is: 1) and TGRW performing step 1, only on tripartite graph, finish the random walk algorithm; TGRW (L) performing step 1 and step 2 promptly use lassologistic regression model that article are carried out finishing the random walk algorithm on after the label expansion three ones.TGRW (LW) is the final method of using of the present invention, promptly on the basis of TGRW (L), the weight of label and scoring is adjusted.

We adopt three kinds of experience evaluating methods of information retrieval field respectively: rate of precision, recall rate and MacroDOA. evaluate and test the performance of algorithm.

Fig. 1 is method and the curve map of baseline collaborative filtering on " degree of accuracy " proposed by the invention.Fig. 2 is method and the curve map of baseline collaborative filtering on " recall rate " proposed by the invention.Fig. 3 is method and the histogram of baseline collaborative filtering on Macro DOA proposed by the invention.

The experimental result demonstration, 1, the TGRW algorithm has the obvious lifting on the performance with respect to traditional collaborative filtering, proves the introducing of socialized label, can significantly promote the efficient of proposed algorithm.2, TGRW (L) is with respect to TGRW, and performance has significant lifting, thereby proof uses lasso logistic regression model to the article extension tag, thereby alleviate the sparse property problem of socialized label, but use the socialized label better application in the middle of recommendation process.3, adjust label and the weight of scoring in recommendation process, make TGRW (LW) algorithm (when recommending N=100) on the accurate rate 3.7% lifting be arranged again than TGRW (L) algorithm.

Claims (4)

1. the collaborative filtering method based on socialized label is characterized in that comprising the steps:

1) at first use tripartite graph that user, article, these three kinds of different node of socialized label are carried out modeling, and application random walk algorithm is each user-customized recommended top-N article;

2) expansion of using lasso logistic regression model to remove article are carried out label, promptly to each article, increase and its semantic relevant label, and remove the label that those have noise, to solve noise and the sparse property problem that exists in the socialized label, promptly article are usually only by seldom a part of label for labelling;

3) regulate the weight of label in recommendation process.

2. a kind of collaborative filtering method based on socialized label according to claim 1 is characterized in that, described by tripartite graph to user, article and label modeling and use the random walk algorithm and be for each user recommends the step of article:

1) use the tripartite graph modeling, then this tripartite graph can be expressed as: G={U, and I, T, E}, wherein U, I, T represent user, article, label respectively, and E represents the relation between them, and then this tripartite graph can be expressed as by following adjacency matrix:

$M=\left(\begin{array}{ccc}0& \mathrm{UI}& 0\\ {\mathrm{UI}}^T& 0& \mathrm{IT}\\ 0& {\mathrm{IT}}^T& 0\end{array}\right)$

Wherein UI represents the rating matrix of User to Item, UI _IjBe user u _iTo article i _jScoring; IT is article and label matrix, IT _JlExpression article i _jBy label t _lThe number of times of mark;

2) on this tripartite graph, use the random walk algorithm to recommend.

Transition matrix in the random walk algorithm is defined as follows:

$S_{\mathrm{ij}}=\frac{M_{\mathrm{ij}}}{\underset{i}{\mathrm{\Σ}}{M}_{\mathrm{ij}}}=\left(\begin{array}{ccc}0& P_1& 0\\ P_2& 0& Q_1\\ 0& Q_2& 0\end{array}\right)$

During the random walk algorithm initialization, each node gives initial value R (0).According to transition matrix, iteration " is walked " to other node subsequently, and until convergence, simultaneously, in order to embody individuation principle, in iterative process, the random walk algorithm moves towards the node that those react user preferences with probability α, and the random walk algorithm patternization is defined as follows:

$\left\{\begin{array}{c}R\left(0\right)=\frac{1}{\left|M\right|}\\ R(t+1)=(1-\mathrm{\α})\mathrm{SR}\left(t\right)+\mathrm{\αp}\end{array}\right.$

Wherein R (t) is the recommendation of random algorithm to each node, and p represents the user preference node, in tripartite graph, comprises user's node, article node, and label node, and it is defined as follows:

p＝[p ^user?p ^item?p ^tag] ^T

Wherein, p ^UserBe user self, p ^ItemBe those article nodes of being marked by the user, p ^TagIt is the used label node of user.

In actual computation, above-mentioned equation can be recommended out following form:

$\left\{\begin{array}{c}{R(t+1)}^{\mathrm{user}}=(1-\mathrm{\α})P_1{R\left(t\right)}^{\mathrm{item}}+{\mathrm{\αp}}^{\mathrm{user}}\\ {R(t+1)}^{\mathrm{tag}}=(1-\mathrm{\α})Q_2{R\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{tag}}\\ {R(t+1)}^{\mathrm{item}}=(1-\mathrm{\α})(P_2{R\left(t\right)}^{\mathrm{user}}+Q_1{R\left(t\right)}^{\mathrm{tag}})+\mathrm{\α}{p}^{\mathrm{item}}\end{array}\right.$

R (t+1) wherein ^User, R (t+1) ^Tag, R (t+1) ^ItemRepresent recommendation respectively, to R (t+1) to user, label and article ^ItemOrdering is selected the top n article of its intermediate value maximum and is recommended the user.

3. a kind of collaborative filtering method according to claim 1 based on socialized label, it is characterized in that, the expansion that described logical lasso logistic regression model removes article are carried out label, thus the step that solves sparse property of label and noise problem is:

1) for each label b, the mark training sample, promptly to each article a, if it is by this label for labelling mistake, then it is positive sample, that is: y _Ab=1, otherwise become negative sample, y _Ab=-1;

2) following maximum likelihood function is found the solution in training in whole sample space, obtains regression coefficient

Maximum likelihood function wherein

For:

$L({\mathrm{\β}}_{b^{*}},{\mathrm{\γ}}_b)=-\underset{a=1}{\overset{m}{\mathrm{\Σ}}}1+\exp (-y_{\mathrm{ab}}({\mathrm{\β}}_{b^{*}}^T{\mathrm{IT}}_{a^{*}}+{\mathrm{\γ}}_b))$

Regression coefficient β _BlBe the degree of correlation semantically of label l and label b, β _Bl＞0 expression positive correlation, β _Bl＜0 expression negative correlation, and β _Bl=0 expression is uncorrelated;

3) for article a, if marked, then be used for describing the semanteme of article with the maximally related K of a label b label by label b, thus the label of expansion article, and K the label least relevant with label b is used for removing the noise of describing in the article tag.

4. a kind of collaborative filtering method according to claim 1 based on socialized label, it is characterized in that, the weight step of described adjusting label in recommendation process is: in the random walk algorithm, to user's weight different of marking with the increase of socialized label item, be respectively δ and 2-δ, then this random walk algorithm is as follows:

$\left\{\begin{array}{c}{R(t+1)}^{\mathrm{user}}=(1-\mathrm{\α})P_1{R\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{user}}\\ {R(t+1)}^{\mathrm{tag}}=(1-\mathrm{\α})Q_2{R\left(t\right)}^{\mathrm{item}}+\mathrm{\α}{p}^{\mathrm{tag}}\\ {R(t+1)}^{\mathrm{item}}=(1-\mathrm{\α})(\mathrm{\δ}{P}_2{R\left(t\right)}^{\mathrm{user}}+(2-\mathrm{\δ})Q_1{R\left(t\right)}^{\mathrm{tag}})+{\mathrm{\αp}}^{\mathrm{item}}\end{array}\right.$

R (t+1) wherein ^User, R (t+1) ^Tag, R (t+1) ^ItemRepresent to the user recommendation of label and article, p respectively ^UserBe user self, p ^ItemBe those article nodes of being marked by the user, p ^TagIt is the used label node of user.