CN105302873A - Collaborative filtering optimization method based on condition restricted Boltzmann machine - Google Patents

Abstract

The invention discloses a collaborative filtering optimization method based on a condition restricted Boltzmann machine. In the improved condition restricted Boltzmann machine, item category information is fused to serve as a condition layer, and recommendation accuracy is improved in a personalized recommendation system. The collaborative filtering optimization method has the characteristics that modeling is carried out by user-item grading information and item category information, different influences on user interest preference and forecast grading by the user-item grading information and the item category information are considered, and the user-item grading information and the item category information are applied to the calculation of the improved condition restricted Boltzmann machine. Since the influences on user interest preference and forecast grading by the user-item grading information and the item category information are simultaneously considered, the method weakens the restriction of a recommendation system by a single data source and improves recommendation accuracy, and an experiment result indicates that the recommendation accuracy of the method is obviously higher than the recommendation accuracy of a restricted Boltzmann machine method which only adopts the user-item grading information.

Description

A kind of collaborative filtering optimization method based on the limited Boltzmann machine of condition

Technical field

The present invention relates to a kind of collaborative filtering optimization method based on the limited Boltzmann machine of condition, be specifically related to a kind ofly consider user-project score information and project category information jointly on the impact of user interest preference and final prediction scoring, and be applied to the limited Boltzmann machine method of improvement, thus to the method that the recommendation accuracy of commending system improves, be applicable to Collaborative Filtering Recommendation System, belong to the technical field of commending system research.

Background technology

The object of commending system be abundant digging user interest preference, help user to find oneself interested thing.Recent two decades comes, and commending system obtains extensive research, and is successfully applied to various internets commercial system.But how to recommend more accurately for user generates, be one of focus of commending system area research always.

Collaborative filtering is the algorithm be most widely used in commending system, and the very severe problem of traditional collaborative filtering one of being faced with is exactly be difficult to extensive, the high openness data set of process, and the innovatory algorithm therefore based on collaborative filtering emerges in an endless stream.Svd model SVD is a kind of dimensionality reduction technology, has been successfully applied in Collaborative Filtering Recommendation System, effectively can solve the Sparse sex chromosome mosaicism that collaborative filtering faces.Limited Boltzmann machine RBM is also that a kind of effective ways can realize Data Dimensionality Reduction function, limited Boltzmann machine RBM be one have double-layer structure, symmetrical connect and entirely connect without self feed back, interlayer, connectionless stochastic neural net model in layer.Salakhutdinov and Hinton has utilized limited Boltzmann machine RBM to carry out modeling to extensive high openness data, and has successfully been applied to collaborative filtering, improves and recommends accuracy.Georgiev and Nakov proposes a kind of non-IID model framework based on limited Boltzmann machine RBM, carries out modeling to the relevance of user-user, project-project simultaneously.Along with the appearance of the fast learning algorithm (to sdpecific dispersion algorithm CD) of RBM, machine learning has started the upsurge that research RBM is theoretical and apply.

In the existing collaborative filtering method based on limited Boltzmann machine model, user-project score data is all only utilized to carry out modeling, the classification information of project is not all fully used, and does not have correlation technique project category information to be combined to be applied to limited Boltzmann machine model to improve recommendation accuracy.

Summary of the invention

The technical problem to be solved in the present invention is: the limitation overcoming prior art, the Sparse sex chromosome mosaicism that effective solution Collaborative Filtering Recommendation System faces, a kind of collaborative filtering optimization method based on the limited Boltzmann machine of condition is provided, the method convergence project classification information extracts the deeper preference profiles of user, improves the accuracy of collaborative filtering recommending method.

The present invention solves the problems of the technologies described above the technical scheme of employing: a kind of collaborative filtering optimization method based on the limited Boltzmann machine of condition, incorporate project category information to optimize in the limited Boltzmann machine of condition improved, improve the recommendation accuracy of Collaborative Filtering Recommendation System.Specific implementation process is as follows:

First, method design is carried out according to user-project score data and project category data characteristics in the basis of limited Boltzmann machine RBM, convergence project category feature is as condition layer, the method IC-CRBMF (collaborative filtering recommending method based on the limited Boltzmann machine of condition of convergence project classification information) proposed, as shown in Figure 2.

IC-CRBMF comprises three parts: visible layer V=(V ₁, V ₂..., V _m), for carrying out modeling to m user's (or project) score data, each scoring uses " softmaxunits " to represent, wherein represent when marking as k, then this other value of scoring vector is all 0, and grading system is K; Hidden layer H=(h ₁, h ₂..., h _h), feature extraction is carried out in effect, each unit h _jrepresent with binary numeral; Condition layer F=(f ₁, f ₂..., f _f), for carrying out modeling to project category feature, user's rating matrix or project rating matrix for visible layer, determine the value of condition layer unit, according to being, based on user or based on project, IC-CRBMF method is divided into two kinds: IC-CRBMF_UserBased and IC-CRBMF_ItemBased.

IC-CRBMF service condition polynomial expression probability distribution each row to visible layer rating matrix carry out modeling, use Bernoulli probability distribution to carry out modeling to hidden layer user (or project) feature.Therefore, the conditional probability distribution computing formula of IC-CRBMF is as follows:

$p(v_i^k=1|H,F)=\frac{\exp (b_i^k+{\mathrm{\Σ}}_{j=1}^H{W}_{ij}^k{h}_j+{\mathrm{\Σ}}_{q=1}^F{f}_q{\mathrm{VF}}_{qi})}{{\mathrm{\Σ}}_{k=1}^K\exp (b_i^k+{\mathrm{\Σ}}_{j=1}^H{W}_{ij}^k{h}_j+{\mathrm{\Σ}}_{q=1}^F{f}_q{\mathrm{VF}}_{qi})}$

$p(h_j=1|V,F)=\mathrm{\σ}(b_j+\underset{i=1}{\overset{V}{\Σ}}\underset{k=1}{\overset{K}{\Σ}}{v}_i^k{W}_{ij}^k+\underset{q=1}{\overset{F}{\Σ}}{f}_q{\mathrm{HF}}_{qj})$

Wherein represent that visible layer i-th scoring is the binary numeral of k; h _jrepresent the binary numeral of a hidden layer jth unit; f _qthe eigenwert of expression condition layer q unit; represent that visible layer i-th scoring is the unit biasing of k; b _jrepresent the biased of a hidden layer jth unit; represent the connection weight between visible layer and hidden layer; VF _qirepresent the connection weight of visible layer and condition layer; HF _qirepresent the connection weight of hidden layer and condition layer; it is activation function.

The marginal probability distribution of IC-CRBMF, visible layer V:

$p\left(V\right)=\underset{H}{\Σ}\underset{F}{\Σ}\frac{\exp (-E(V,H,F))}{{\mathrm{\Σ}}_{V^{\′},H^{\′},F^{\′}}\exp (-E(V^{\′},H^{\′},F^{\′}\left)\right)}$

The energy function of IC_CRBMF:

$E\left(V,H,F\right)=-\underset{i=1}{\overset{V}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{v_i}^k{b_i}^k-\underset{i=1}{\overset{V}{\mathrm{\Σ}}}\underset{j=1}{\overset{H}{\mathrm{\Σ}}}\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{v_i}^k{W_{ij}}^k{h}_j-\underset{j=1}{\overset{H}{\mathrm{\Σ}}}{h}_j{b}_j-\underset{i=1}{\overset{V}{\mathrm{\Σ}}}\underset{q=1}{\overset{F}{\mathrm{\Σ}}}{f}_q{\mathrm{VF}}_{qi}-\underset{j=1}{\overset{H}{\mathrm{\Σ}}}\underset{q=1}{\overset{F}{\mathrm{\Σ}}}{f}_f{\mathrm{HF}}_{qj}$

Utilize the gradient of maximal possibility estimation calculating target function (marginal probability distribution function), finally use gradient rise method iteration undated parameter, finally obtain the optimum value of all parameters of IC-CRBMF.

The present invention's advantage is compared with prior art:

(1), based on the collaborative filtering of limited Boltzmann machine only make use of user-project score information, and the information not making full use of other carrys out the interest preference of digging user.The present invention has incorporated abundant project category information in the collaborative filtering method based on the limited Boltzmann machine of condition, as condition layer characteristic information, fully take into account the impact that project category information is marked on user interest and prediction, deeper extraction user (project) feature, improves the recommendation accuracy of collaborative filtering system.

(2), as can be seen from experimental result, introduce the score in predicting precision that project category information can improve the collaborative filtering recommending method based on limited Boltzmann machine really, and the weighted array method HybridIC-CRBMF (see Fig. 3) of IC-CRBMF_UserBased and IC-CRBMF_ItemBased also obtains the accuracy more excellent than single model.

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is the limited Boltzmann machine frame diagram of condition of convergence project classification information.

Fig. 3 is based on user and project-based mixing condition Boltzmann machine frame diagram.

Fig. 4 is MAE and the RMSE value of each model.

Embodiment

The present invention is further illustrated below in conjunction with the drawings and specific embodiments.

A kind of collaborative filtering optimization method based on the limited Boltzmann machine of condition, convergence project classification information, a kind of limited Boltzmann machine method IC-CRBMF of improvement is proposed, consider the impact that project category information is marked on user interest preference and prediction, and incorporate the limited Boltzmann machine model of improvement, thus improve the recommendation accuracy of collaborative filtering recommending method.

Described method IC-CRBMF, considers the impact of project category information on user interest preference, prediction scoring, using the condition layer of project category feature as model.According to the different expression form of visible layer, IC-CRBMF is divided into again IC-CRBMF_UserBased based on user and project-based IC-CRBMF_ItemBased two methods.

Described method IC-CRBMF_UserBased, visible layer carries out modeling according to the score information of each user, and condition layer is that the category feature information of all items of marking according to this user carries out modeling.

Described method IC-CRBMF_UserBased, each user is that a limited Boltzmann machine RBM trains example, the visible layer unit of varying number can be had, but all limited Boltzmann machine RBM have the hidden layer unit of equal number, i.e. all user's Share interlinkage weights and offset parameter.

Described method IC-CRBMF_ItemBased, visible layer carries out modeling according to the score information of each project, and condition layer carries out modeling according to the category feature information of this project.

Described method IC-CRBMF_ItemBased, each project is that a limited Boltzmann machine RBM trains example, the visible layer unit of varying number can be had, but all limited Boltzmann machine RBM have the hidden layer unit of equal number, i.e. all items Share interlinkage weight and offset parameter.

Described method IC-CRBMF_UserBased and IC-CRBMF_ItemBased, most important difference is that the scoring of visible layer represents the expression with the project category feature of condition layer.

Method realizes being divided into two stages, and first stage is the method design stage, mainly carries out modelling according to the feature of limited Boltzmann machine model and the feature of data set.IC-CRBMF mainly comprises three part: visible layer V, hidden layer H and condition layer F.

As shown in Figure 1, mainly comprise the steps:

Steps A 1), utilize condition polynomial expression probability distribution to visible layer each row scoring vector carry out modeling, the unit of visible layer the probability be activated is:

$p(v_i^k=1|H,F)=\frac{\exp (b_i^k+{\mathrm{\Σ}}_{j=1}^H{W}_{ij}^k{h}_j+{\mathrm{\Σ}}_{q=1}^F{f}_q{\mathrm{VF}}_{qi})}{{\mathrm{\Σ}}_{k=1}^K\exp (b_i^k+{\mathrm{\Σ}}_{j=1}^H{W}_{ij}^k{h}_j+{\mathrm{\Σ}}_{q=1}^F{f}_q{\mathrm{VF}}_{qi})}$

Wherein represent that visible layer i-th scoring is the binary numeral of k; h _jrepresent the binary numeral of a hidden layer jth unit; f _qthe eigenwert of expression condition layer q unit; represent that visible layer i-th scoring is the unit biasing of k; b _jrepresent the biased of a hidden layer jth unit; represent the connection weight between visible layer and hidden layer; VF _qirepresent the connection weight of visible layer and condition layer.

Steps A 2), utilize Bernoulli probability distribute modeling is carried out to the proper vector of hidden layer, the unit h of hidden layer _jthe probability be activated:

$p(h_j=1|V,F)=\mathrm{\σ}(b_j+\underset{i=1}{\overset{V}{\Σ}}\underset{k=1}{\overset{K}{\Σ}}{v}_i^k{W}_{ij}^k+\underset{q=1}{\overset{F}{\Σ}}{f}_q{\mathrm{HF}}_{qj})$

Wherein represent that visible layer i-th scoring is the binary numeral of k; f _qthe eigenwert of expression condition layer q unit; represent that visible layer i-th scoring is the unit biasing of k; b _jrepresent the biased of a hidden layer jth unit; represent the connection weight between visible layer and hidden layer; HF _qirepresent the connection weight of hidden layer and condition layer; it is activation function.

Second stage is learning phase, mainly tectonic model parameter obtain best parameter value for predicting scoring.Comprise the steps:

Step B1), parameter initialization;

From the method design stage, the major parameter of IC-CRBMF: the connection weight of visible layer and hidden layer the connection weight VF of visible layer and condition layer _qi, hidden layer and condition layer connection weight HF _qiall use average to be 0, standard deviation be 0.01 normal distribution carry out initialization; Visible layer unit biasing hidden layer unit biasing b _jbe initialized as complete zero.

Step B2), perform stochastic gradient rise method and carry out parameter renewal, upgraded the optimum value obtaining model parameter by continuous iteration, the gradient formula of parameter is as follows:

${\mathrm{\&Delta;W}}_{ij}^k=\frac{\&part;\log p\left(V\right)}{\&part;W_{ij}^k}=<v_i^k{h}_j{>}_{data}-<v_i^k{h}_j{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;b}}_i^k=\frac{\&part;\log p\left(V\right)}{\&part;b_{ij}^k}=<v_i^k{>}_{data}-<v_i^k{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;b}}_j=\frac{\&part;\log p\left(V\right)}{\&part;b_j}=<h_j{>}_{data}-<h_j{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;VF}}_{qi}=\frac{\&part;\log p\left(V\right)}{\&part;{\mathrm{VF}}_{qi}}=<v_i^k{f}_q{>}_{data}-<v_i^k{f}_q{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;HF}}_{qj}=\frac{\&part;\log p\left(V\right)}{\&part;{\mathrm{HF}}_{qj}}=<h_j{f}_q{>}_{data}-<h_j{f}_q{>}_{\mathrm{mod}el}$

Wherein <> _datarepresent the expectation defined by training set; <> _modelrepresent the expectation defined by model IC-CRBMF.

Due to <> _modelbe difficult to calculate, use the expectation of fast learning algorithm-come sdpecific dispersion algorithm CD parameter in approximate treatment IC-CRBMF, therefore the gradient formula of parameter is amended as follows:

${\mathrm{\&Delta;W}}_j^k=<v_i^k{h}_j{>}_{data}-<v_i^k{h}_j{>}_T$

${\mathrm{\&Delta;b}}_i^k=<v_i^k{>}_{data}-<v_i^k{>}_T$

△b _j＝<h _j> _data-<h _j> _T

${\mathrm{\&Delta;VF}}_{qi}=<v_i^k{f}_q{>}_{data}-<v_i^k{f}_q{>}_T$

△HF _qj＝<h _jf _q> _data-<h _jf _q> _T

Wherein <> _trepresent that performing T Gibbs samples the expectation value obtained.

In actual experiment, the selection of learning rate is most important, and the excessive algorithm the convergence speed of learning rate causes soon algorithm unstable, can be unstable to avoid the algorithm but speed of convergence is too slow although learning rate is too small.In order to overcome this contradiction, also momentum term is introduced in the process of undated parameter, the amendment direction of t subparameter value is not exclusively determined by the likelihood gradient direction of current sample, and revise direction and t subgradient direction by t-1 subparameter and jointly determine, local best points can be converged to too early to avoid the algorithm like this.Parameter more new formula is:

$W_{ij}^k\left(t\right)=\mathrm{\&kappa;}\&CenterDot;W_{ij}^k(t-1)+\mathrm{\&epsiv;}\&CenterDot;{\mathrm{\&Delta;W}}_{ij}^k$

$b_i^k\left(t\right)=\mathrm{\&kappa;}\&CenterDot;b_i^k(t-1)+\mathrm{\&epsiv;}\&CenterDot;{\mathrm{\&Delta;b}}_i^k$

b _j(t)＝κ·b _j(t-1)+ε·△b _j

VF _qi(t)＝κ·VF _qi(t-1)+ε·△VF _qi

HF _qj(t)＝κ·HF _qj(t-1)+ε·△HF _qj

Wherein κ is the learning rate of momentum term, and ε is the learning rate of parameter gradients, and two parameters are all the constants arranged in advance.

Stochastic gradient rise method is used to need to carry out several times iterative computation; During each iteration, travel through a training dataset; Often obtain user's scoring (or project scoring), computation model parameter: likelihood gradient: then all parameters of IC-CRBMF are upgraded according to above-mentioned formula.When meeting the requirements of precision of prediction, iteration stopping, thus the optimum value obtaining model parameter.

Step B3), prediction scoring, the parameter value according to the best carries out score in predicting;

$R=\underset{k=1}{\overset{K}{\&Sigma;}}k\&CenterDot;p(v_i^k=1|H,F)$

According to being, based on user with based on project, method C-CRBMF is divided into IC-CRBMF_UserBased and IC-CRBMF_ItemBased two kinds.The final prediction scoring of the weighted array method HybridIC-CRBMF of method IC-CRBMF_UserBased and IC-CRBMF_ItemBased is the weighted array of two method prediction scorings, and computing formula is as follows:

R _Hybrid＝β·R _{IC-CRBMF_ItemBased}+(1-β)·R _{IC-CRBMF_UserBased}

Wherein β represents combining weights, and β span is: 0≤β≤1.

As in Fig. 2, V, W, F represent visible layer, hidden layer, condition layer respectively; represent that visible layer i-th scoring is the binary numeral of k; h _jrepresent a hidden layer jth unit binary numeral (j=1 ..., H); f _qthe eigenwert of expression condition layer q unit; W represents the connection weight between visible layer and hidden layer; VF represents the connection weight of visible layer and condition layer; HF represents the connection weight of hidden layer and condition layer; RBM represents limited Boltzmann machine; IC-CRBMF represents the limited Boltzmann machine of the condition of convergence project classification information; MissingRating represents disappearance scoring; ItemGenreFeatureVector represents project category proper vector.

In Fig. 3, V, W, F represent visible layer, hidden layer, condition layer respectively; represent that visible layer i-th scoring is the binary numeral of k; h _jrepresent a hidden layer jth unit binary numeral (j=1 ..., H); f _qthe eigenwert of expression condition layer q unit; W represents the connection weight between visible layer and hidden layer; VF represents the connection weight of visible layer and condition layer; HF represents the connection weight of hidden layer and condition layer; RBM represents limited Boltzmann machine; IC-CRBMF represents the limited Boltzmann machine of the condition of convergence project classification information; MissingRating represents disappearance scoring; ItemGenreFeatureVector represents project category proper vector; Model represents input method; It is project-based, method is based on user that Item_Based, User_Based represent method respectively; R _{iC-CRBMF_UserBased}represent the prediction scoring of IC-CRBMF_UserBased; R _{iC-CRBMF_ItemBased}represent the prediction scoring of IC-CRBMF_ItemBased method; β represents weight, and span is 0≤β≤1; ∑ represents summation operation; HybridIC-CRBMF represents R _{iC-CRBMF_ItemBased}and R _{iC-CRBMF_ItemBased}weighted array method; R _{hybrid_IC-CRBMF}represent the prediction scoring of HybridIC-CRBMF method.

In Fig. 4:

CFModel: collaborative filtering method;

BasicUserBasedCF: traditional collaborative filtering based on user;

BasicItemBasedCF: traditional project-based collaborative filtering;

SVD: svd;

AutoEncode: autocoder;

Stack_AutoEncode: based on the scrambler of stack;

RBM_UserBased: based on the limited Boltzmann machine of user;

RBM_ItemBased: project-based limited Boltzmann machine;

The weighted array method of HybridRBMModel:RBM_UserBased and RBM_ItemBased;

IC-CRBMF_UserBased: the limited Boltzmann machine of condition of convergence project classification information;

IC-CRBMF_ItemBased: the limited Boltzmann machine of condition of convergence project classification information;

The weighted array method of HybridIC-CRBMFModel:IC-CRBMF_UserBased and IC-CRBMF_ItemBased.

MAE: mean absolute error;

RMSE: root square error;

$MAE=\frac{1}{n}\underset{j=1}{\overset{n}{\&Sigma;}}|R_{predict\_j}-R_{real\_j}|$

$RMSE=\sqrt{\frac{1}{n}\underset{j=1}{\overset{n}{\&Sigma;}}{(R_{predict\_j}-R_{real\_j})}^2}$

N represents the number of data sample; R _{predict_j}represent a jth prediction scoring; R _{real_j}represent a jth actual scoring.

Claims (4)

1. based on a collaborative filtering optimization method for the limited Boltzmann machine of condition, it is characterized in that: described method is divided into two stages:

First stage is the design phase, carry out modelling according to the feature of the limited Boltzmann machine model of condition and the feature of data set, model comprises visible layer V, i.e. score data input layer, hidden layer H, i.e. feature extraction layer and condition layer F, i.e. condition data input layer;

Performing step is as follows:

Steps A 1, utilize condition polynomial expression probability distribution to visible layer v each row scoring vector carry out modeling, the unit of visible layer the probability be activated is:

$p(v_i^k=1|H,F)=\frac{\exp (b_i^k+{\mathrm{\&Sigma;}}_{j=1}^H{W}_{ij}^k{h}_j+{\mathrm{\&Sigma;}}_{q=1}^F{f}_q{\mathrm{VF}}_{qi})}{{\mathrm{\&Sigma;}}_{k=1}^K\exp (b_i^k+{\mathrm{\&Sigma;}}_{j=1}^H{W}_{ij}^k{h}_j+{\mathrm{\&Sigma;}}_{q=1}^F{f}_q{\mathrm{VF}}_{qi})}$

Wherein represent that visible layer i-th scoring is the binary numeral of k; h _jrepresent the binary numeral of a hidden layer jth unit; f _qthe eigenwert of expression condition layer q unit; represent that visible layer i-th scoring is the unit biasing of k; b _jrepresent the biased of a hidden layer jth unit; represent the connection weight between visible layer and hidden layer; VF _qirepresent the connection weight of visible layer and condition layer;

Steps A 2, utilize Bernoulli probability distribute modeling is carried out to the proper vector of hidden layer H, the unit h of hidden layer _jthe probability be activated:

$p(h_j=1|V,F)=\mathrm{\&sigma;}(b_j+\underset{i=1}{\overset{V}{\&Sigma;}}\underset{k=1}{\overset{K}{\&Sigma;}}{v}_i^k{W}_{ij}^k+\underset{q=1}{\overset{F}{\&Sigma;}}{f}_q{\mathrm{HF}}_{qj})$

Wherein represent that visible layer i-th scoring is the binary numeral of k; f _qthe eigenwert of expression condition layer q unit; represent that visible layer i-th scoring is the unit biasing of k; b _jrepresent the biased of a hidden layer jth unit; represent the connection weight between visible layer and hidden layer; HF _qirepresent the connection weight of hidden layer and condition layer; activation function, wherein $x=b_j+\underset{i=1}{\overset{V}{\&Sigma;}}\underset{k}{\overset{K}{\&Sigma;}}{v}_i^k{W}_{ij}^k+\underset{q=1}{\overset{F}{\&Sigma;}}{f}_q{\mathrm{HF}}_{qj};$

Second stage is learning phase, and tectonic model parameter also obtains best parameter value, for predicting scoring, comprises the steps:

Step B1, parameter initialization

Parameter is the connection weight of visible layer and hidden layer the connection weight VF of visible layer and condition layer _qi, hidden layer and condition layer connection weight HF _qiall use average to be 0, standard deviation be 0.01 normal distribution carry out initialization; Visible layer unit biasing hidden layer unit biasing b _jbe initialized as complete zero;

Step B2, employing stochastic gradient rise method carry out parameter renewal, and upgraded the optimum value obtaining model parameter by continuous iteration, the gradient formula of parameter is as follows:

${\mathrm{\&Delta;W}}_{ij}^k=\frac{\&part;logp\left(V\right)}{\&part;W_{ij}^k}=<v_i^k{h}_j{>}_{data}-<v_i^k{h}_j{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;b}}_i^k=\frac{\&part;\log p\left(V\right)}{\&part;b_i^k}=<v_i^k{>}_{data}-<v_i^k{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;b}}_j=\frac{\&part;\log p\left(V\right)}{\&part;b_j}=<h_j{>}_{data}-<h_j{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;VF}}_{qi}=\frac{\&part;\log p\left(V\right)}{\&part;{\mathrm{VF}}_{qi}}=<v_i^k{f}_q{>}_{data}-<v_i^k{f}_q{>}_{\mathrm{mod}el}$

${\mathrm{\&Delta;HF}}_{qj}=\frac{\&part;\log p\left(V\right)}{\&part;{\mathrm{HF}}_{qj}}=<h_j{f}_q{>}_{data}-<h_j{f}_q{>}_{\mathrm{mod}el}$

Wherein <> _datarepresent the expectation defined by training set; <> _modelrepresent the expectation defined by model IC-CRBMF;

Step B3, prediction scoring, the parameter value according to the best carries out score in predicting;

$R=\underset{k=1}{\overset{K}{\&Sigma;}}k\&CenterDot;p(v_i^k=1|H,F)$

IC-CRBMF: the collaborative filtering recommending method based on the limited Boltzmann machine of condition representing convergence project classification information;

According to based on user with based on project, IC-CRBMF is divided into the IC-CRBMF_UserBased based on user and project-based IC-CRBMF_ItemBased two kinds, then obtain final predicting the outcome by the combination of HybridIC-CRBMF mixed weighting, be calculated as follows:

R _Hybrid＝β·R _{IC-CRBMF_ItemBased}+(1-β)·R _{IC-CRBMF_UserBased}

Wherein β represents combining weights.

2. the collaborative filtering optimization method based on the limited Boltzmann machine of condition according to claim 1, is characterized in that: in described step B2, use stochastic gradient rise method to need to carry out several times iterative computation; During each iteration, travel through a training dataset; Often obtain user's scoring or project scoring, computation model parameter: likelihood gradient: then according to the parameter of step B2 more new formula upgrade all parameters of IC-CRBMF, when meeting the requirements of precision of prediction, iteration stopping, thus the optimum value obtaining model parameter.

3. the collaborative filtering optimization method based on the limited Boltzmann machine of condition according to claim 1, is characterized in that: in described step B2, <> _modelbe difficult to calculate, use fast learning algorithm namely to sdpecific dispersion algorithm CD, by the expectation of parameter in approximate treatment IC-CRBMF (collaborative filtering recommending method based on the limited Boltzmann machine of condition of convergence project classification information), the gradient formula of parameter is amended as follows:

${\mathrm{\&Delta;W}}_{ij}^k=<v_i^k{h}_j{>}_{data}-<v_i^k{h}_j{>}_T$

${\mathrm{\&Delta;b}}_i^k=<v_i^k{>}_{data}-<v_i^k{>}_T$

△b _j＝<h _j> _data-<h _j> _T

${\mathrm{\&Delta;VF}}_{qi}=<v_i^k{f}_q{>}_{data}-<v_i^k{f}_q{>}_T$

△HF _qj＝<h _jf _q> _data-<h _jf _q> _T

Wherein <> _trepresent that performing T Gibbs samples the expectation value obtained.

4. the collaborative filtering optimization method based on the limited Boltzmann machine of condition according to claim 1, is characterized in that: the span of described β is: 0≤β≤1.