CN105654329A - Integrated recommendation method and apparatus thereof - Google Patents

Abstract

The invention is suitable for the recommendation technology field and provides an integrated recommendation method and an apparatus thereof. The integrated recommendation method comprises the following steps of acquiring data of a plurality of articles; according to an integrated recommendation model trained in advance and the data of the articles, generating recommendation probabilities of the articles, wherein the integrated recommendation model is a model of generating data and marking data based on a user; according to the recommendation probabilities, ordering the articles, and according to sizes of the recommendation probabilities, taking at least one article as a recommended article from high to low. By using the method and the apparatus of the invention, problems that an existing recommended system lacks validity and comprehensiveness and an adverse impact brought by a cold boot problem can not be mitigated are solved; a suitable scope is enlarged and credibility is increased. Therefore, the comprehensiveness and the validity are increased and the adverse impact brought by the cold boot problem is mitigated.

Description

A kind of conformability recommend method and device

Technical field

The invention belongs to recommended technology field, relate in particular to a kind of conformability recommend method and device.

Background technology

On a lot of e-commerce websites, commending system is all installed, by commending system, according to user's informationDemand, interest etc., recommend user by interested user information, product etc., finds institute to reduce userNeed the time of article.

In existing commending system, according to the type of commending system, be divided into content-based commending system andBased on the commending system of user feedback, content-based commending system needs system to divide the content of articleAnalyse, the commending system based on user feedback has various implementation, wherein with based on matrix decompositionMethod the most successful. It is user's marking that method based on matrix decomposition changes into user's Feedback RuleSparse matrix. Then utilize the mathematical method of low-rank matrix decomposition to expand this sparse matrix, to the greatest extentAmount is filled full whole matrix.

Conventionally, existing commending system, has following main shortcoming:

1, lack comprehensive. Because the content of article exists the situation of infeasibility and imperfection, therefore push awayRecommend scope little, recommendation effect is unsatisfactory, in addition, has in a large number user's user in the content of articleGenerated data, and content-based commending system generally can not be processed user generated data, has further limited toThe scope of application of commending system, therefore the scope of application is little, lacks comprehensive.

2, lack validity. Matrix disassembling method can be predicted a user's hobby in future, recommendSystem can not provide any discernible reason and confirm that this is a reasonably recommendation. In other words, matrixThe product decomposing, two low-rank matrixes do not have any physical significance. This can produce the confidence level of commending systemRaw negative effect, that therefore recommends is with a low credibility, lacks validity.

3, cannot alleviate the negative effect that cold start-up problem is brought. Wherein, when a new user just addsWhen commending system, due to any feedback not having about him, therefore commending system can not do in this caseGo out any effective recommendation, this problem is known as cold start-up problem. And existing commending system cannot alleviateThe negative effect that cold start-up problem is brought.

Summary of the invention

The object of the embodiment of the present invention is to provide a kind of conformability recommend method, is intended to solve existing recommendationSystem lacks validity, comprehensive and cannot alleviate the problem of the negative effect that cold start-up problem brings.

The embodiment of the present invention is achieved in that a kind of conformability recommend method, comprising:

Obtain the data of multiple lines;

According to the conformability recommended models of training in advance and the data of described article, generate pushing away of described articleRecommend probability, described conformability recommended models is the model based on user generated data and marking data;

According to described recommendation probability, described article are sorted, by the size of described recommendation probability, from heightTo the low article that are taken to described in one item missing as recommending article.

Another object of the embodiment of the present invention is to provide a kind of conformability recommendation apparatus, comprising:

Data acquisition module, for obtaining the data of multiple lines;

Recommend probability generation module, for according to the conformability recommended models of training in advance and described articleData, generate the recommendation probability of described article, described conformability recommended models be based on user generated data andThe model of marking data;

Recommending module, for according to described recommendation probability, sorts to described article, general by described recommendationThe size of rate, from height to the low article that are taken to described in one item missing as recommending article.

In embodiments of the present invention, according to the conformability recommended models of training in advance and the data of described article,Generate the recommendation probability of described article, described conformability recommended models is based on user generated data and gradesAccording to model, according to described recommendation probability, described article are sorted, by the size of described recommendation probability,From height to the low article that are taken to described in one item missing as recommending article. By content-based commending system with based on useThe commending system of family feedback is integrated, and utilizes idle a large number of users generated data, alleviates cold start-up problem and bringsNegative effect, and reach the word cloud that each dimension of low-rank matrix is provided to physical significance, solvedExisting commending system lacks validity, comprehensive and cannot alleviate the negative effect that cold start-up problem is broughtProblem, expanded the scope of application, increased confidence level simultaneously, thereby both improved comprehensive and validity,Also alleviated the negative effect that cold start-up problem is brought.

Brief description of the drawings

Fig. 1 is the realization flow figure of a kind of conformability recommend method of providing of the embodiment of the present invention;

Fig. 2 is the realization flow figure of the training conformability recommended models that provides of the embodiment of the present invention;

Fig. 3 is the implementing procedure figure that the embodiment of the present invention provides the step S203 in training conformability recommended models;

Fig. 4 is the system diagram of the integrated recommend method that provides of the present embodiment;

Fig. 5 is the first structured flowchart of the conformability recommendation apparatus that provides of the embodiment of the present invention;

Fig. 6 is conformability recommendation apparatus the second structured flowchart that the embodiment of the present invention provides;

Fig. 7 is conformability recommended models training module in the conformability recommendation apparatus that provides of the embodiment of the present inventionStructured flowchart;

Fig. 8 is the structured flowchart of the conformability recommended models training unit that provides of the embodiment of the present invention;

Fig. 9 is conformability recommendation apparatus the 3rd structured flowchart that the embodiment of the present invention provides.

Detailed description of the invention

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with accompanying drawing and realityExecute example, the present invention is further elaborated. Only should be appreciated that specific embodiment described hereinOnly, in order to explain the present invention, be not intended to limit the present invention.

Embodiment mono-

Fig. 1 is the realization flow figure of a kind of conformability recommend method of providing of the embodiment of the present invention, and details are as follows:

In step S101, obtain the data of multiple lines;

Wherein, obtain the data of multiple lines, can adopt any mode of prior art to obtain, as pass throughThe mode of online is obtained, or the mode importing by local data base is obtained.

In step S102, according to the conformability recommended models of training in advance and the data of described article, rawBecome the recommendation probability of described article, described conformability recommended models is based on user generated data and marking dataModel;

Before step S102, comprising: training conformability recommended models.

In step S103, according to described recommendation probability, described article are sorted, general by described recommendationThe size of rate, from height to the low article that are taken to described in one item missing as recommending article.

Wherein, from height to the low K item article of getting, as recommending article, before output, K conduct is for active userRecommendation article.

The size of K can, for user is from establishing, also can adopt system default.

In embodiments of the present invention, by whole to content-based commending system and the commending system based on user feedbackClose, utilize idle a large number of users generated data, alleviate the negative effect that cold start-up problem is brought, and reachEach dimension of low-rank matrix is provided to the word cloud of physical significance, solved existing commending system and lackedValidity, comprehensive and cannot alleviate the problem of the negative effect that cold start-up problem brings, has expanded and has been suitable forScope has increased confidence level simultaneously, thereby has both improved comprehensive and validity, has also alleviated cold start-up and has askedThe negative effect that topic is brought.

Embodiment bis-

Fig. 2 is the realization flow figure of the training conformability recommended models that provides of the embodiment of the present invention, and details are as follows:

In step S201, obtain user generated data and marking data;

User generated data is that the character property of article is described, and wherein comprised the description of user to article and commentedValency. By the analysis to description, we can train and obtain the attribute of article and user's hobby.

In step S202, described user generated data and marking data are carried out to pretreatment, generate described useThe sparse marking matrix of the lexical matrix of family generated data and described marking data;

In step S203, according to described lexical matrix and sparse marking matrix, the integration that training is set up in advanceProperty recommended models.

In embodiments of the present invention, compared with other commending system, this commending system has utilized user generatedAccording to, can improve the efficiency that entirety is recommended.

Embodiment tri-

Fig. 3 is the implementing procedure that the embodiment of the present invention provides the step S203 in training conformability recommended modelsFigure, details are as follows:

In step S301, in described lexical matrix and sparse marking matrix, utilize gibbs sampler frameworkSample, generate the hidden variable sampled value of described lexical matrix and sparse marking matrix;

In step S302, according to described hidden variable sampled value, the conformability recommended models that training is set up in advanceParameter.

In the present embodiment, the parameter of the conformability recommended models that training is set up in advance, can improve generationRecommend the accuracy of probability.

Embodiment tetra-

The present embodiment has been described the implementation procedure of setting up conformability recommended models, and details are as follows:

Set up conformability recommended models, described conformability recommended models is:

Wherein, w is the vocabulary that user evaluates article, and x is the scoring of user to article,For article themeParameter, α is the Dirichlet priori parameter that article theme distributes, the β vocabulary distribution Dirichlet that is the themePriori, μ₀For the priori mean parameter of user to theme favorable rating Gaussian distribution,For user likes themeThe prior variance parameter of love degree Gaussian distribution, σ²For the give a mark variance parameter of Gaussian distribution of user.

Wherein, j is article matrix columns, the quantity that M is article, θ_jFor the theme distributed constant of current article j,P(θ_j| α) be the theme distribution of article j, U_jFor article j being beaten to undue user's set, l is article lexical matrixColumns, L_i,jFor article j receives total number of evaluating vocabulary, z is the hidden variable that vocabulary l is corresponding, and K is the themeNumber, w_lFor Evaluation: Current vocabulary, ψ_zThe vocabulary of z of being the theme distributes, and f evaluates hidden variable corresponding to j, x_i,jForThe scoring of user i to article j, μ_i,fFor user's mean parameter that marking distributes to article;

Wherein, formula left side is the recommendation probability of article, and described recommendation probability is proportional to formula right side, formulaRight side Section 1 is that the theme of article distributes, and right side Section 2 is to process the model of user generated data, described inUser generated data comprises the evaluation word of article, and right side Section 3 is to process the model of marking data, described inMarking data comprise the marking of article.

Embodiment five

The present embodiment has been described the training implementing procedure of the parameter of the conformability recommended models of foundation in advance, describes in detailAs follows:

The formula sampling by gibbs, respectively to hidden variable f and z sampling.

The formula of gibbs sampling is as follows:

$P(f_i=j|z,w,f_{\⫬i},x)\∝N\left(x_i\right|{\mathrm{\μ}}_i{\mathrm{\σ}}_i^2+{\mathrm{\σ}}_0^2)\frac{n_{f,\⫬i,j}^v+n_{z,j}^v+\mathrm{\α}}{n_{f,\⫬i,(\·)}+n_{z,(.)}+\mathrm{K\α}}---1)$

$P(z_i=j{|z}_{\⫬i},w,f,x)\∝\frac{n_{\⫬i,j}^{w_i}+\mathrm{\β}}{n_{\⫬i,j}^{(.)}+\left|V\right|\mathrm{\β}}\frac{n_{z,\⫬i,j}^v+n_{f,j}^v+\mathrm{\α}}{n_{z,\⫬i,(.)}+n_{f,(.)}+\mathrm{K\α}}---2)$

Wherein, formula 1) be the formula to f sampling, formula 2) second be the formula to z sampling. TheN (x| μ, σ) in a formula represents taking μ as mean value, and the probability of the normal distribution that σ is standard deviation is closeDegree Distribution Value. The meaning of its dependent variable sees table:

The first step of sampling is to be random number (between 0 to 1) by all μ assignment, by all itsHis parameter equalization, then calculate its dependent variable according to the physical significance of upper table and according to sampling formula to f andZ sampling. F and the z of sampling, can calculate all variablees in sampling formula, based on all variablees after calculatingAgain to f and z sampling, until sampling finishes.

After gibbs sampling process completes, parameter (modelparameters) that can computation model,Calculate according to the following formula:

${\mathrm{\θ}}_{v,j}=\frac{n_{z,j}^v+n_{f,j}^v+\mathrm{\α}}{n_{z,(.)}^v+n_{z,(.)}^v+\mathrm{K\α}}$

${\mathrm{\ψ}}_{{j,w}_i}=\frac{n_j^{w_i}+\mathrm{\β}}{n_j^{(.)}+\left|V\right|\mathrm{\β}}$

${\mathrm{\μ}}_{u,j}={(\frac{1}{{\mathrm{\σ}}_0^2}+\frac{\left|x_{u,(.)}^j\right|}{{\mathrm{\σ}}^2})}^{-1}(\frac{{\mathrm{\μ}}_0}{{\mathrm{\σ}}_0^2}+\frac{{\mathrm{\Σ}}_m{x}_{u,m}^j}{{\mathrm{\σ}}^2})$

Wherein the meaning of all variablees is the same with the variable in gibbs sampling formula, based on passing samplingF and z, can calculate all variablees.

After once all model variables all calculate, can calculate the marking of any user to article,After need using marking by descending and output before several as recommendation article.

Embodiment six

With reference to figure 4, Fig. 4 is the system diagram of the integrated recommend method that provides of the present embodiment, and details are as follows:

The system of moving integrated recommend method mainly comprises two large divisions. It is raw that the dish in left side is used for describing userBecome data.

Wherein, user generated data is that the character property of article is described, and has wherein comprised user's retouching articleState and evaluate. By to the analysis of describing, can train and obtain the attribute of article and user's hobby.

Wherein, M represents the number of article, and L represents that each user generates the number of words comprising in brief comment.Each word is described as a probability distribution based on hidden variable z. Hidden variable z has determined to change wordResiding implicit classification. The corresponding potential feature classification of each hidden variable z.

Therefore it is unique that the word that, each hidden variable z is corresponding distributes. Each word distribution can becomeFor obtaining the key point of word cloud of each dimension.

The dish on model right side is used for describing user feedback.

The numeric type marking to article fancy grade based on user in the feedback of setting user. One highMark representative of consumer article are had to stronger liking, vice versa. Similarly, introduce hidden variable f and carry out generationTable is beaten the main focus of time sharing user at every turn, and based on this focus, each user has an implicit markDistribution, the explicit mark obtaining is a sampling of this distribution.

The binding site of content-based system and the right side system based on user feedback in left side is distribution theta,This distribution has determined the property distribution under article itself, and hidden variable z and f are based on thetaThe sampling distributing. Like this, by organic to content-based commending system and the commending system based on user feedbackCombine, form conformability commending system.

For ease of explanation, taking practical application as example, details are as follows:

One, data processing

1). user's marking data are organized into a sparse marking matrix, the numeral generation of the capable j row of iThe marking of table user i to article j.

2). arrange user's user generated data. Need to determine model vocabulary in this step, by userUser generated data arranges becomes article lexical matrix. First determine vocabulary. The all of all users are commentedValency integrate, obtain all vocabulary, then by stop-word (often appear at the vocabulary in various documents,Be commonly considered as background vocabulary) filter out. To remain vocabulary by word frequency descending, get a suitable numberWord, as the vocabulary of model.

For example can get 10000, the vocabulary coming after 10000 be abandoned, calculate article word frequency squareWhen battle array, can not consider 10000 later vocabulary. According to the vocabulary of obtaining, can calculate each articleWord frequency list, insert in article lexical matrix. Article lexical matrix i is capable, and j row represent j vocabularyThe number of times occurring in to the evaluation of i article user.

Two, model training

The marking matrix that data processing is obtained and article lexical matrix (can be divided into instruction as the input of modelPractice data and verification msg, wherein training data is as mode input, and verification msg is suitable for findingModel parameter), model parameter (hyper-parameter) is carried out to linearity and find, and use at checking numberFinal argument according to the model parameter of above putting up the best performance as model training. By all data (training datasAnd verification msg) all as training data, and use optimum model parameter, use gibbs sampling method pairModel training.

Three. article are recommended

After model training completes, given any one user, can, based on existing model, calculate userTo the marking of all article, then descending, front K the conduct of output is for active user's recommendation article.

Embodiment seven

Fig. 5 is the first structured flowchart of the conformability recommendation apparatus that provides of the embodiment of the present invention, and this device canRun in server end. For convenience of explanation, only show the part relevant to the present embodiment.

With reference to Fig. 5, this conformability recommendation apparatus, comprising:

Data acquisition module 51, for obtaining the data of multiple lines;

Recommend probability generation module 52, for according to the conformability recommended models of training in advance and described articleData, generate the recommendation probability of described article, described conformability recommended models is based on user generated dataModel with marking data;

Recommending module 53, for according to described recommendation probability, sorts to described article, by described recommendationThe size of probability, from height to the low article that are taken to described in one item missing as recommending article.

In a kind of implementation of the present embodiment, with reference to figure 6, Fig. 6 is the integration that the embodiment of the present invention providesProperty recommendation apparatus the second structured flowchart, this conformability recommendation apparatus, also comprises:

Conformability recommended models training module 54, for training conformability recommended models.

In a kind of implementation of the present embodiment, with reference to figure 7, Fig. 7 be the embodiment of the present invention provide wholeThe structured flowchart of conformability recommended models training module 54 in closing property recommendation apparatus, described conformability recommended modelsTraining module 54, comprising:

Acquiring unit 541, for obtaining user generated data and marking data;

Pretreatment unit 542, for described user generated data and marking data are carried out to pretreatment, generatesThe sparse marking matrix of the lexical matrix of described user generated data and described marking data;

Conformability recommended models training unit 543, for according to described lexical matrix and sparse marking matrix,The conformability recommended models that training is set up in advance.

In a kind of implementation of the present embodiment, with reference to figure 8, Fig. 8 is the integration that the embodiment of the present invention providesThe structured flowchart of property recommended models training unit 543, this conformability recommended models training unit 543, comprising:

Sampling subelement 5431, at described lexical matrix and sparse marking matrix, utilizes gibbs to adoptSample framework is sampled, and generates the hidden variable sampled value of described lexical matrix and sparse marking matrix;

Training subelement 5432, for according to described hidden variable sampled value, trains the conformability of setting up in advance to push awayRecommend the parameter of model.

In a kind of implementation of the present embodiment, with reference to figure 9, Fig. 9 is the integration that the embodiment of the present invention providesProperty recommendation apparatus the 3rd structured flowchart, this conformability recommendation apparatus, also comprises:

Conformability recommended models is set up module 55, and for setting up conformability recommended models, described conformability is recommendedModel is:

Wherein, w is the vocabulary that user evaluates article, and x is the scoring of user to article,For article themeParameter, α is the Dirichlet priori parameter that article theme distributes, the β vocabulary distribution Dirichlet that is the themePriori, μ₀For the priori mean parameter of user to theme favorable rating Gaussian distribution,For user likes themeThe prior variance parameter of love degree Gaussian distribution, σ²For the give a mark variance parameter of Gaussian distribution of user.

Wherein, j is article matrix columns, the quantity that M is article, θ_jFor the theme distributed constant of current article j,P(θ_j| α) be the theme distribution of article j, U_jFor article j being beaten to undue user's set, l is article lexical matrixColumns, L_i,jFor article j receives total number of evaluating vocabulary, z is the hidden variable that vocabulary l is corresponding, and K is the themeNumber, w_lFor Evaluation: Current vocabulary, ψ_zThe vocabulary of z of being the theme distributes, and f evaluates hidden variable corresponding to j, x_i,jForThe scoring of user i to article j, μ_i,fFor user's mean parameter that marking distributes to article;

Wherein, formula left side is the recommendation probability of article, and described recommendation probability is proportional to formula right side, formulaRight side Section 1 is that the theme of article distributes, and right side Section 2 is to process the model of user generated data, described inUser generated data comprises the evaluation word of article, and right side Section 3 is to process the model of marking data, described inMarking data comprise the marking of article.

The device that the embodiment of the present invention provides can be applied in the embodiment of the method for aforementioned correspondence, details referring toThe description of above-described embodiment, does not repeat them here.

Through the above description of the embodiments, those skilled in the art can be well understood to thisThe bright mode that can add essential common hardware by software realizes. Described program can be stored in and can readIn storage medium, described storage medium, as random access memory, flash memory, read-only storage, able to programmeMemory read, electrically erasable programmable memory, register etc. This storage medium is positioned at memory, processesInformation in device read memory, in conjunction with the method described in each embodiment of its hardware implement the present invention.

The above be only the specific embodiment of the present invention, but protection scope of the present invention is not limited toThis, any be familiar with those skilled in the art the present invention disclose technical scope in, can expect easilyVariation or replacement, within all should being encompassed in protection scope of the present invention. Therefore, protection scope of the present inventionShould be as the criterion with the protection domain of claim.

Claims (10)

1. a conformability recommend method, is characterized in that, comprising:

Obtain the data of multiple lines;

According to the conformability recommended models of training in advance and the data of described article, generate pushing away of described articleRecommend probability, described conformability recommended models is the model based on user generated data and marking data;

According to described recommendation probability, described article are sorted, by the size of described recommendation probability, from heightTo the low article that are taken to described in one item missing as recommending article.

2. conformability recommend method according to claim 1, is characterized in that, in described basis in advanceBefore the conformability recommended models of training and the data of described article, comprising:

Training conformability recommended models.

3. conformability recommend method according to claim 2, is characterized in that, described training conformabilityRecommended models, is specially:

Obtain user generated data and marking data;

Described user generated data and marking data are carried out to pretreatment, generate the word of described user generated dataThe sparse marking matrix of remittance matrix and described marking data;

According to described lexical matrix and sparse marking matrix, the conformability recommended models that training is set up in advance.

4. conformability recommend method as claimed in claim 3, is characterized in that, described according to described vocabularyMatrix and sparse marking matrix, the conformability recommended models that training is set up in advance, is specially:

In described lexical matrix and sparse marking matrix, utilize gibbs sampler framework sampling, described in generationThe hidden variable sampled value of lexical matrix and sparse marking matrix;

According to described hidden variable sampled value, the parameter of the conformability recommended models that training is set up in advance.

5. conformability recommend method as claimed in claim 2, is characterized in that, in described training conformabilityBefore recommended models, comprising:

Set up conformability recommended models, described conformability recommended models is:

$P(w,x|\mathrm{\Θ};\mathrm{\α},\mathrm{\β},{\mathrm{\μ}}_0,{\mathrm{\σ}}_0^2,{\mathrm{\σ}}^2)$

$\∝\underset{j=1}{\overset{M}{\mathrm{\Π}}}P\left({\mathrm{\θ}}_j\right|\mathrm{\α})\underset{j\∈U_j}{\mathrm{\Π}}\left(\underset{l=1}{\overset{L_{i,j}}{\mathrm{\Π}}}\underset{z=1}{\overset{K}{\mathrm{\Σ}}}P\left(z\right|{\mathrm{\θ}}_j)P\left(w_1\right|{\mathrm{\ψ}}_2)\right)\left(\underset{f=1}{\overset{K}{\mathrm{\Σ}}}P\left(f\right|{\mathrm{\θ}}_j)P\left(x_{i,j}\right|{\mathrm{\μ}}_{i,f},{\mathrm{\σ}}^2)\right)$

Wherein, w is the vocabulary that user evaluates article, and x is the scoring of user to article, and Θ is article themeParameter, α is the Dirichlet priori parameter that article theme distributes, the β vocabulary distribution Dirichlet that is the themePriori, μ₀For the priori mean parameter of user to theme favorable rating Gaussian distribution,For user likes themeThe prior variance parameter of love degree Gaussian distribution, σ²For the give a mark variance parameter of Gaussian distribution of user.

Wherein, j is article matrix columns, the quantity that M is article, θ_jFor the theme distributed constant of current article j,P(θ_j| α) be the theme distribution of article j, U_jFor article j being beaten to undue user's set, l is article lexical matrixColumns, L_i,jFor article j receives total number of evaluating vocabulary, z is the hidden variable that vocabulary j is corresponding, and K is the themeNumber, w_lFor Evaluation: Current vocabulary, z_zThe vocabulary of z of being the theme distributes, and f evaluates hidden variable corresponding to j, x_i,jForThe scoring of user i to article j, μ_i,fFor user's mean parameter that marking distributes to article;

Wherein, formula left side is the recommendation probability of article, and described recommendation probability is proportional to formula right side, formulaRight side Section 1 is that the theme of article distributes, and right side Section 2 is to process the model of user generated data, described inUser generated data comprises the evaluation word of article, and right side Section 3 is to process the model of marking data, described inMarking data comprise the marking of article.

6. a conformability recommendation apparatus, is characterized in that, comprising:

Data acquisition module, for obtaining the data of multiple lines;

Recommend probability generation module, for according to the conformability recommended models of training in advance and described articleData, generate the recommendation probability of described article, described conformability recommended models be based on user generated data andThe model of marking data;

Recommending module, for according to described recommendation probability, sorts to described article, general by described recommendationThe size of rate, from height to the low article that are taken to described in one item missing as recommending article.

7. conformability recommendation apparatus as claimed in claim 6, is characterized in that, described conformability is recommended dressPut, also comprise:

Conformability recommended models training module, for training conformability recommended models.

8. conformability recommendation apparatus as claimed in claim 7, is characterized in that, described conformability is recommended mouldType training module, comprising:

Acquiring unit, for obtaining user generated data and marking data;

Pretreatment unit, for described user generated data and marking data are carried out to pretreatment, described in generationThe sparse marking matrix of the lexical matrix of user generated data and described marking data;

Conformability recommended models training unit, for according to described lexical matrix and sparse marking matrix, trainsThe conformability recommended models of setting up in advance.

9. conformability recommendation apparatus as claimed in claim 8, is characterized in that, described conformability is recommended mouldType training unit comprises:

Sampling subelement, at described lexical matrix and sparse marking matrix, utilizes gibbs sampler frameFrame is sampled, and generates the hidden variable sampled value of described lexical matrix and sparse marking matrix;

Training subelement, for according to described hidden variable sampled value, trains the conformability of setting up in advance to recommend mouldThe parameter of type.

10. conformability recommendation apparatus as claimed in claim 7, is characterized in that, described conformability is recommendedDevice also comprises:

Conformability recommended models is set up module, and for setting up conformability recommended models, described conformability is recommended mouldType is:

$P(w,x|\mathrm{\Θ};\mathrm{\α},\mathrm{\β},{\mathrm{\μ}}_0,{\mathrm{\σ}}_0^2,{\mathrm{\σ}}^2)$

$\∝\underset{j=1}{\overset{M}{\mathrm{\Π}}}P\left({\mathrm{\θ}}_j\right|\mathrm{\α})\underset{j\∈U_j}{\mathrm{\Π}}\left(\underset{l=1}{\overset{L_{i,j}}{\mathrm{\Π}}}\underset{z=1}{\overset{K}{\mathrm{\Σ}}}P\left(z\right|{\mathrm{\θ}}_j)P\left(w_1\right|{\mathrm{\ψ}}_2)\right)\left(\underset{f=1}{\overset{K}{\mathrm{\Σ}}}P\left(f\right|{\mathrm{\θ}}_j)P\left(x_{i,j}\right|{\mathrm{\μ}}_{i,f},{\mathrm{\σ}}^2)\right)$

Wherein, w is the vocabulary that user evaluates article, and x is the scoring of user to article, and Θ is article themeParameter, α is the Dirichlet priori parameter that article theme distributes, the β vocabulary distribution Dirichlet that is the themePriori, μ₀For the priori mean parameter of user to theme favorable rating Gaussian distribution,For user likes themeThe prior variance parameter of love degree Gaussian distribution, σ²For the give a mark variance parameter of Gaussian distribution of user.

Wherein, j is article matrix columns, the quantity that M is article, θ_jFor the theme distributed constant of current article j,P(θ_j| α) be the theme distribution of article j, U_jFor article j being beaten to undue user's set, l is article lexical matrixColumns, L_i,jFor article j receives total number of evaluating vocabulary, z is the hidden variable that vocabulary l is corresponding, and K is the themeNumber, w_lFor Evaluation: Current vocabulary, z_zThe vocabulary of z of being the theme distributes, and f evaluates hidden variable corresponding to j, x_i,jForThe scoring of user i to article j, μ_i,fFor user's mean parameter that marking distributes to article;

Wherein, formula left side is the recommendation probability of article, and described recommendation probability is proportional to formula right side, formulaRight side Section 1 is that the theme of article distributes, and right side Section 2 is to process the model of user generated data, described inUser generated data comprises the evaluation word of article, and right side Section 3 is to process the model of marking data, described inMarking data comprise the marking of article.