CN103136694A - Collaborative filtering recommendation method based on search behavior perception - Google Patents

Abstract

The invention discloses a collaborative filtering recommendation method based on search behavior perception. The method includes the following steps: (1) analyzing inquiry behaviors and search behaviors of a user on an e-commerce website; (2) constructing user-product keyword tensor based on search behavior context; (3) constructing a training dataset of a factor machine model and mapping tensor data into vector data; (4) establishing a recommendation method frame based on the factor machine model and utilizing an improved alternate least squares algorithm to carry out parameter estimation; and (5) evaluating the recommendation method based on the factor machine model through an experiment. The collaborative filtering recommendation method based on the search behavior perception utilizes context information ignored by a traditional collaborative filtering recommendation method, solves the problem that a traditional individuation recommendation method cannot provide a recommended result which expresses the intent of the user on a business to business e-commerce website, and is better than a traditional method in accuracy and timeliness of the recommended result.

Description

Collaborative filtering recommending method based on the search behavior perception

Technical field

The present invention relates to personalized recommendation field, internet, particularly relate to a kind of collaborative filtering recommending method based on the search behavior perception.

Background technology

Personalized recommendation system successful Application on the internet is that Internet firm has started new opportunity, particularly e-commerce website in recent years, 30% purchase business is arranged from personalized recommendation system on the B2C e-commerce website of specific area.But commending system is not used widely on the B2B E-commerce website.The role of intermediary is played the part of in the B2B E-commerce website in commercial activity, the buyer is by the supplier of B2B E-commerce website searching target product.In this course, at first the buyer inputs the searching key word relevant to target product, e-commerce website returns to the product a large amount of of the same type from different suppliers, and then the buyer selects the product that satisfies the demands and carries out inquiry to its supplier by the details of browsing product.Commending system on the B2B E-commerce website is intended to for the user provides suitable product candidate, helps the user more effectively to complete above-mentioned commercial activity, thereby improves the user to satisfaction and the dependency degree of website.

Collaborative Filtering Recommendation Algorithm is a kind of proposed algorithm the most frequently used in personalized recommendation system.The interest of collaborative filtering analysis user finds targeted customer's similar users in customer group, and the comprehensively evaluation of these similar users to a certain article, forms at last this targeted customer to the prediction of the fancy grade of special article.Two kinds of main collaborative filtering methods are based on respectively the method for Neighborhood Model and based on the method for enigmatic language justice model (latent factor model).Utilize user's historical behavior data based on the method for Neighborhood Model, calculate the similarity of user's (article) by using the modes such as Pearson came (Pearson) correlativity and included angle cosine, obtain neighbour's set of user's (article), thereby then use the behavioral data relevant to these neighbours to calculate the targeted customer, the scoring of special article is recommended.Its core concept of method based on enigmatic language justice model is by hidden feature (latent factor) contact user interest and article, it decomposes (Matrix Factorization) scheduling algorithm by matrix multiplier and user's rating matrix is decomposed into user's matrix and the article matrix of low-rank, obtains the user to the score in predicting value of article according to the inner product of user characteristics vector sum article characteristics vector.Method based on enigmatic language justice model is widely used in commending system.

On the B2B E-commerce website, utilize the conspiracy relation can be according to user's preference screening product, yet possibly can't embody user's intention fully based on collaborative recommendation results.The collaborative filtering of based on the context perception provides the method that addresses this problem.The context-aware collaborative filtering method can be divided into three types: 1) filter (contextual pre-filtering) method before context, namely select or data configuration by the context driving data; 2) filter (contextual post-filtering) method after context, namely by the context filtering recommendation results; 3) context modeling (contextual modeling) method is about to context fusion in model.In the research of context-aware recommend method, the context modeling method is taken seriously rapidly because it surpasses the superior function of classic method.In actual applications, if use keyword to represent user's search behavior, the model of perception user search behavior comprises the relation between user, product, keyword and three simultaneously so.The collaborative filtering task definition is: according to the value of Partial Elements in the tensor of user, product, keyword formation, the value of disappearance element in the prediction tensor.

The factoring algorithm of tensor has higher computation complexity, thereby causes not being suitable for large-scale recommendation task based on the method for tensor Factorization.Factor machine (Factorization Machine) model is present best context-aware model.Its number of parameters is linear increasing, and each parameter has analytic solution, thereby has effectively solved the problem of computation complexity, and has kept the advantage that collaborative filtering has of context-aware.The perception of usage factor machine model realization of the present invention to the user search behavior, thus can provide more valuable recommendation service for the B2B E-commerce website.

Summary of the invention

Goal of the invention: for collaborative filtering can't perception user search behavior problem, the search behavior of analysis user and inquiry behavior, utilize the keyword that uses in the user search behavior as the context of user's inquiry behavior, for the B2B E-commerce website provides a kind of collaborative filtering recommending method based on user search behavior perception.

Technical scheme: the collaborative filtering recommending method based on the search behavior perception comprises the steps:

(101) user's inquiry behavior and search behavior on the analytical electron business web site, the sign of unification user, product and keyword;

(102) structure is based on the contextual user-product of search behavior-keyword tensor, and it uses keyword as the context of inquiry behavior;

(103) structure factor machine model training data set, the mapping of setting up tensor and vector, described mapping changes into one-dimensional vector with three-dimensional tensor;

(104) set up recommend method framework based on factor machine model, described framework learn to obtain the parameter of factor machine model by training dataset and according to the current search behavior as the evaluation of estimate of context-prediction user to product;

(105) experimental evaluation is based on the recommend method of factor machine model.

Wherein, step (101) comprising:

(101-1) user identity in analytical electron business web site journal file carries out the disambiguation operation to user identity, and user identity is mapped to user ID unique in system;

(101-2) type of behavior in identification user behavior record, extract behavior sequence; With the end of inquiry behavior as behavior sequence, will with the beginning of the nearest search behavior of inquiry behavior as behavior sequence, with the medium content of the click behavior between two kinds of behaviors as behavior sequence; Wherein, if the inquiry behavior does not have search behavior before occuring, think that search behavior is empty;

(101-3) resolve related products in behavior sequence, and the keyword that uses in search behavior.

In step (104), be divided into study and predict two processes based on the recommend method framework of factor machine model; In learning process, described framework obtains the parameter of factor machine model according to the training dataset of step (103) structure by study; In forecasting process, the keyword that described framework uses according to given user, product and user uses the model parameter that study obtains to calculate evaluation of estimate.Step (104) adopts improved alternately least-squares algorithm to carry out the parameter estimation of factor machine model, and the mode that adopts is at first to fix and the known quantity of finding the solution cache oblivious, then calculates the analytic solution of finding the solution parameter; Wherein improved alternately least-squares algorithm adopts the mode of calculating in advance known quantity to reduce computation complexity.

Step (105) comprises the preparation data set; Adopt root-mean-square error as evaluation index; Test on data set; Analyze experimental result.

The present invention adopts technique scheme, has following beneficial effect: the present invention uses search behavior on the B2B E-commerce website as the context of its inquiry behavior, makes commending system have perception to user's particular demands.The number of parameters of factor machine model is linear increasing, and each parameter has analytic solution, makes optimized algorithm have lower time complexity.The present invention adopts the context-aware collaborative filtering based on factor machine model, has effectively improved the accuracy of B2B E-commerce recommendation of websites result and ageing, thereby has promoted the user to satisfaction and the dependency degree of e-commerce website.

Description of drawings

Fig. 1 is the method flow diagram of the embodiment of the present invention;

Fig. 2 is the contextual tensor schematic diagram of the search behavior of the embodiment of the present invention;

Fig. 3 is the schematic diagram that the tensor data of the embodiment of the present invention are converted into vector data;

Fig. 4 is the experimental result schematic diagram of the embodiment of the present invention on MovieLens score data collection;

Fig. 5 is the experimental result schematic diagram of the embodiment of the present invention on the hidden feedback data collection of MovieLens.

Embodiment

Below in conjunction with specific embodiment, further illustrate the present invention, should understand these embodiment only is used for explanation the present invention and is not used in and limits the scope of the invention, after having read the present invention, those skilled in the art all fall within the application's claims limited range to the modification of the various equivalent form of values of the present invention.

The present invention is by analyzing search behavior, inquiry behavior and the click behavior of user on the B2B E-commerce website, uses keyword that search behavior produces as the context of inquiry behavior, builds user-product-keyword three-dimensional tensor.And then by setting up the tensor data to the mapping of vector data, the training data of structure factor machine model.Then provided the method for parameter estimation of factor machine model, it is improved alternately least square method.At last, compare with baseline algorithm on standard data set.As shown in Figure 1, the method for the embodiment of the present invention comprises the following steps:

Step 101: user's search behavior, click behavior and inquiry behavior on the analytical electron business web site, the identification user identity also produces user's set, and the identification product also produces the product set, and the identification searching key word also produces keyword set.In described set, all elements all have the unique sign of the overall situation.Specifically comprise:

(101-1) user identity in analysis B2B E-commerce web log file file, carry out the disambiguation operation to user identity, user identity is mapped to user ID unique in system, and the generation user gathers U.The behavior sequence of each user within a session cycle is used as a user behavior record, and wherein session week index futures user begins to operate from Website login and leaves the website and no longer operate this cycle.

(101-2) type of behavior in identification user behavior record, extract behavior sequence.With the end of inquiry behavior as behavior sequence, will with the beginning of the nearest search behavior of inquiry behavior as behavior sequence, with the medium content of the click behavior between two kinds of behaviors as behavior sequence.Wherein, if the inquiry behavior does not have search behavior before occuring, think that search behavior is empty.

(101-3) resolve related products in behavior sequence, and the keyword that uses in search behavior.Product is corresponded to Product Identifying unique in system and produces product set I.Keyword is carried out the standardization processing such as stem, synonym merging, correspond to keyword unique in system and identify and produce keyword set Q.

Step 102: structure is based on the contextual user-product of search behavior-keyword tensor T ≡ U * I * Q, as shown in Figure 2.Wherein keyword is as the context of user-product behavior.Defined function

, f (u, i, q)=r represents the value of certain element in tensor.The behavior sequence that analytical procedure 101 obtains when user u has used after keyword q searches for, has carried out inquiry to product i, defines f (q, u, i)=1; After u had used keyword q to search for, i clicked to product, but did not carry out inquiry, defined f (q, u, i)=0; Otherwise the value of f (q, u, i) disappearance.

Step 103: structure factor machine model training data set, the mapping of setting up tensor and vector.Mapping process projects to three dimensions of each element of tensor respectively the different components of vector, i.e. T _uiq→ x ^(uiq), T wherein _uiqAn element of expression tensor, x ^(uiq)Represent an one-dimensional vector.

The concrete operations that tensor are mapped to vector are as follows:

Use function phi (u)=x ^(u), value u is converted into vector x ^(u), wherein

Use function phi (i)=x ⁽ⁱ⁾, value i is converted into vector x ⁽ⁱ⁾, wherein

Use function phi (q)=x ^(q), value q is converted into vector x ^(q), wherein

Attended operation by vector

Three parts are connected into a vector, wherein

The attended operation of two vectors of expression.

Transfer function φ () wherein:

Have following three kinds of forms:

(1) classification territory: class variable C can be mapped to a real-valued vector by an index device with it to each classification.As shown in Figure 3, keyword only has 3 kinds of situations, so if the user is u ₁Use keyword q ₂Search for and inquiry product i ₁, vectorial φ (q ₂)=(0,1,0), if keyword is q ₃, vectorial φ (q ₃)=(0,0,1).

(2) classification set territory: class variable set C can be mapped to a real-valued vector by an index device with it to each context, and all non-zero element values are identical, and summation is 1.User u for example ₁Use keyword { q ₂, q ₃Search for and inquiry product i ₁, vectorial φ ({ q ₂, q ₃)=(0,0.5,0.5).

(3) real-valued territory: C is real number, uses real number value as feature, i.e. φ (C)=C.

Step 104: set up the recommend method framework based on factor machine model, framework is divided into study and predicts two processes.In learning process, this framework obtains the parameter of factor machine model according to the training dataset of step 103 structure by study.In forecasting process, the keyword that this framework uses according to given user, product and user uses the model parameter that study obtains to calculate evaluation of estimate.Embodiment is:

(104-1) model representation.Factor machine model is by the relation between each component of factorization interaction parameter measuring vector, and its model is as shown in formula (1):

$\hat{r}\left(x\right)\≡w_0+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_i{x}_i+\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{n}{\mathrm{\Σ}}}{\hat{w}}_{i,j}{x}_i{x}_j---\left(1\right)$

Wherein Factorization interaction parameter between the component of expression vector,

${\hat{w}}_{i,j}\≡\⟨v_i,v_j\⟩=\underset{f=1}{\overset{k}{\mathrm{\Σ}}}{v}_{i,f}\·v_{j,f}---\left(2\right)$

Wherein, need the model parameter Θ of estimation to be: w ₀∈ R, w ∈ R ⁿ, V ∈ R ^{N * k}By formula (1) w as can be known ₀Global bias, w _iInteraction relation between i component of modeling and evaluation of estimate,

Factorization interaction relation between the modeling component.Wherein x represents training sample, Expression prediction and evaluation value.

(104-2) optimization aim.The definition loss function is:

$L(r,\hat{r})=\underset{(x,r)\∈S}{\mathrm{\Σ}}{(\hat{r}\left(x\right)-r)}^2---\left(3\right)$

Wherein training data of (x, r) expression with and corresponding evaluation of estimate, S represents the training data set.According to the defined loss function of formula (3) target that is optimized, increase regularization term after optimization aim as shown in formula (4):

$\mathrm{RLS}-\mathrm{OPT}=\underset{(x,r)\∈S}{\mathrm{\Σ}}{(\hat{r}\left(x\right)-r)}^2+\underset{\mathrm{\θ}\∈\mathrm{\Θ}}{\mathrm{\Σ}}{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}{\mathrm{\θ}}^2---\left(4\right)$

Wherein

Regularization term, λ _(θ)The regularization factor corresponding to parameter θ.

(104-3) use the alternately parameter of least square method estimation factor machine model, i.e. w ₀∈ R, w ∈ R ⁿ, V ∈ R ^{N * k}In the process of calculating, represent model parameter with θ, for the mode that all parameter θ ∈ Θ adopt combined optimization, namely alternately calculate the optimal value of each parameter, until satisfy end condition.

Alternately the computation complexity of least-squares algorithm is usually higher, but can effectively reduce computation complexity in the situation that can access analytic solution by pre-computation methods.Adopt when alternately least-squares algorithm is estimated the parameter of factor machine model, factor machine model is in the situation that fixing parameter with finding the solution cache oblivious is linear model as can be known, model can be divided into the constant part of finding the solution cache oblivious and with the linear function part of finding the solution parameter correlation, can obtain fast the analytic solution of parameter by the invariant of precomputation model.Theorem one and theorem two have provided respectively the linear model proof of factor machine model and the process of optimization aim being found the solution the Parameter analysis of electrochemical solution.

Theorem one: for each model parameter θ ∈ Θ, its factor machine is all a linear model, that is:

$\hat{r}\left(x\right|\mathrm{\θ})=g_{\left(\mathrm{\θ}\right)}\left(x\right)+\mathrm{\θ}{h}_{\left(\mathrm{\θ}\right)}\left(x\right)---\left(5\right)$

Function g wherein, the form of h is relevant to θ, but concrete value has nothing to do with θ.

Proof procedure is as follows:

Can obtain parameter w by formula (1) and (2) ₀, w _l, v _l,fConcrete form, as follows:

w ₀:

w _l:

v _l,f:

Can find out from above formula, for each parameter θ ∈ Θ, its factor machine model can be described as relatively independent two parts: constant component g _(θ)(x) and the linear function part θ h relevant to θ _(θ)(x), this like factor machine model can be described as the linear function expression formula formally:

Theorem two: can obtain its analytic solution for each the parameter θ in optimization aim,

$\mathrm{\θ}=-\frac{{\mathrm{\Σ}}_{(x,r)\∈S}(g_{\left(\mathrm{\θ}\right)}\left(x\right)-r)h_{\left(\mathrm{\θ}\right)}\left(x\right)}{{\mathrm{\Σ}}_{(x,r)\∈S}{h}_{\left(\mathrm{\θ}\right)}^2+{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}}---\left(6\right)$

Its proof procedure is as follows:

$\mathrm{RLS}-\mathrm{OPT}=\underset{(x,r)\∈S}{\mathrm{\Σ}}{(\hat{r}\left(x\right)-r)}^2+\underset{\mathrm{\θ}\∈\mathrm{\Θ}}{\mathrm{\Σ}}{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}{\mathrm{\θ}}^2$

Its parameter is asked local derviation, and we can obtain:

$\frac{\∂}{\∂\mathrm{\θ}}\mathrm{RLS}-\mathrm{OPT}=\underset{(x,r)\∈S}{\mathrm{\Σ}}2(\hat{r}\left(x\right)-r)h_{\left(\mathrm{\θ}\right)}\left(x\right)+2{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}\mathrm{\θ}$

Then make that derivative is 0, can obtain formula as follows:

$\underset{(x,r)\∈S}{\mathrm{\Σ}}2(\hat{r}\left(x\right)-r)h_{\left(\mathrm{\θ}\right)}\left(x\right)+2{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}\mathrm{\θ}=0$

With formula (5) substitution wherein, can obtain following equation:

$\underset{(x,r)\∈S}{\mathrm{\Σ}}(g_{\left(\mathrm{\θ}\right)}\left(x\right)+\mathrm{\θ}{h}_{\left(\mathrm{\θ}\right)}\left(x\right)-r)h_{\left(\mathrm{\θ}\right)}\left(x\right)+{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}\mathrm{\θ}=0$

Expansion obtains following equation:

$\underset{(x,r)\∈S}{\mathrm{\Σ}}(g_{\left(\mathrm{\θ}\right)}\left(x\right)-r)h_{\mathrm{\θ}}\left(x\right)+\underset{(x,r)\∈S}{\mathrm{\Σ}}\mathrm{\θ}{h}_{\left(\mathrm{\θ}\right)}^2\left(x\right)+{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}\mathrm{\θ}=0$

Abbreviation can get:

$\mathrm{\θ}=-\frac{{\mathrm{\Σ}}_{(x,r)\∈S}(g_{\left(\mathrm{\θ}\right)}\left(x\right)-r)h_{\left(\mathrm{\θ}\right)}\left(x\right)}{{\mathrm{\Σ}}_{(x,r)\∈S}{h}_{\left(\mathrm{\θ}\right)}^2\left(x\right)+{\mathrm{\λ}}_{\left(\mathrm{\θ}\right)}}$

Can find out from above-mentioned formula (6), in factor machine model, can calculate analytic solution for each parameter θ, thereby can adopt improved alternately least-squares algorithm to carry out the calculating of parameter.

Improved alternately least-squares algorithm:

The present embodiment is the little parameter of iterative computation reciprocation at first, then calculates the relatively large parameter of reciprocation, i.e. iterative computation w successively ₀, w _l, v _l,fComputing formula (6) by θ can find out, the calculating of the optimal value of θ depends primarily on the calculating of g and these two functional values of h.By theorem one as can be known function g and h and parameter θ irrelevant, therefore can adopt in advance the mode of the value of computing function g and h to avoid double counting in iterative process, thereby reduce computation complexity.

For function g, definition:

G is arranged _(θ)(x)=e (x, r| Θ)-θ h _(θ)θ.At first calculate

θ becomes θ when parameter ^*The time, then upgrade e, more new formula is as follows:

e(x,r|Θ ^*)＝e(x,r|Θ)+(θ ^*-θ)h _(θ)(x) (7)

For function h, definition:

Have:

$h_{\left(v_{l,f}\right)}\left(x\right)=x_l\underset{i=1,i\≠l}{\mathrm{\Σ}}{v}_{i,f}{x}_i$

$=x_l\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{v}_{i,f}{x}_i-x_l^2{v}_{l,f}$

$=x_{l}q(x,f|\mathrm{\Θ})-x_l^2{v}_{l,f}$

Due to for parameter w ₀And w _l, the h that it is corresponding _(θ)(x) can calculate in the time at constant, and for parameter v _i,f, the parameter that it is corresponding

Comprise a circulation, can adopt the mode of calculating in advance known quantity to complete the calculating of parameter at the constant time complexity this moment.By observing

Can find out that q and l are separate, therefore can calculate in advance q, as parameter v _l,fBecome

Shi Gengxin q, more new formula is as follows:

$q(x,f|{\mathrm{\Θ}}^{*})=q(x,f|\mathrm{\Θ})+(v_{l,f}^{*}-v_{l,f})x_l---\left(8\right)$

The above-mentioned improved least-squares algorithm false code that replaces is as follows:

Step 105: experimental evaluation specifically comprises based on the recommend method of factor machine model:

(105-1) prepare data set;

A. use MovieLens1M(http: //www.grouplens.org/node/73) data set is as the experimental data collection of the present embodiment.The MovieLens1M data set comprises 6040 users and 3900 films, 1000209 user's evaluating datas (span 1～5) altogether, and every corresponding several keywords of film represent the theme of film.The present embodiment is with the product in the commercial affairs of film simulation electronic, with the theme simulation searching key word of film.

B. analyze the evaluating data of MovieLens1M, get evaluation of estimate greater than 3 customer behavior modeling inquiry behavior, get evaluation of estimate and be not more than 3 customer behavior modeling and click behavior.Every film all is associated with one or more themes, and with the keyword of theme simulation search behavior, thereby each user behavior has corresponding context.

C. select at random all behavioral datas of MovieLens1M data centralization 80% as training dataset, remaining part is as test data set.

(105-2) evaluation index;

Adopt root-mean-square error (RMSE) as the evaluation index of the present embodiment.After algorithm is completed iteration at every turn, calculate the root-mean-square error of "current" model on test set.

(105-3) test on data set;

A. relatively based on the collaborative filtering method of user search behavior perception and experimental result based on the collaborative filtering method of SVD.The present embodiment uses stochastic gradient descent algorithm and improved alternately least-squares algorithm to train, and the SVD model adopts stochastic gradient descent algorithm to train.Wherein, the hidden factor dimension of factor machine model is made as 20; The learning rate of stochastic gradient descent algorithm is made as 0.002, and the regularization factor is made as 0.01.

B. compare the factor machine model (context-aware FM) of context-aware and the experimental result of context-free factor machine model (context-free FM).Wherein, the hidden factor dimension of factor machine model is made as 20.

(105-4) analyze experimental result;

A. by Fig. 4 and Fig. 5 as can be known, under context-free prerequisite, all be better than collaborative filtering method based on SVD based on the collaborative filtering method of user search behavior perception on score data collection and hidden feedback data collection, therefore improved the correctness of system.

B. as shown in Figure 4, on the score data collection, root-mean-square error after the factor machine model of context-aware and the convergence of context-free factor machine model is close, but therefore the speed of convergence of the factor machine model of context-aware has improved the actual effect of training speed and system effectively obviously faster than context-free factor machine model.

Claims (5)

1. based on the collaborative filtering recommending method of search behavior perception, it is characterized in that, comprise the steps:

(101) user's inquiry behavior and search behavior on the analytical electron business web site, the sign of unification user, product and keyword;

(102) structure is based on the contextual user-product of search behavior-keyword tensor, and it uses keyword as the context of inquiry behavior;

(103) structure factor machine model training data set, the mapping of setting up tensor and vector, described mapping changes into one-dimensional vector with three-dimensional tensor;

(104) set up recommend method framework based on factor machine model, described framework learns to obtain the parameter of factor machine model by training dataset, and according to the current search behavior as the evaluation of estimate of context-prediction user to product;

(105) experimental evaluation is based on the recommend method of factor machine model.

2. the collaborative filtering recommending method based on the search behavior perception according to claim 1, it is characterized in that: described step (101) comprising:

(101-1) user identity in analytical electron business web site journal file carries out the disambiguation operation to user identity, and user identity is mapped to user ID unique in system;

(101-2) type of behavior in identification user behavior record, extract behavior sequence; With the end of inquiry behavior as behavior sequence, will with the beginning of the nearest search behavior of inquiry behavior as behavior sequence, with the medium content of the click behavior between two kinds of behaviors as behavior sequence; Wherein, if the inquiry behavior does not have search behavior before occuring, think that search behavior is empty;

(101-3) resolve related products in behavior sequence, and the keyword that uses in search behavior.

3. the collaborative filtering recommending method based on the search behavior perception according to claim 1 is characterized in that: in described step (104), described recommend method framework is divided into study and predicts two processes; In learning process, described recommend method framework obtains the parameter of factor machine model according to the training dataset of step (103) structure by study; In forecasting process, the keyword that described framework uses according to given user, product and user uses the model parameter that study obtains to calculate evaluation of estimate.

4. according to claim 1 or 3 described collaborative filtering recommending methods based on the search behavior perception, it is characterized in that: described step (104) adopts improved alternately least-squares algorithm to carry out the parameter estimation of factor machine model, and the mode that adopts is at first to fix and the known quantity of finding the solution cache oblivious, then calculates the analytic solution of finding the solution parameter.

5. the collaborative filtering recommending method based on the search behavior perception according to claim 1, it is characterized in that: described step (105) comprises the preparation data set; Adopt root-mean-square error as evaluation index; Test on data set; Analyze experimental result.

Images