CN103632290B - A kind of based on the mixing recommendation method recommending probability fusion - Google Patents

Abstract

The invention discloses and a kind of recommend method based on the mixing recommending probability fusion, the method comprises the steps: 1) score data of commodity is represented with bivariate table；2) to the Arbitrary Term marked in set as unknown number, utilize basis recommendation method to obtain predicting the outcome of respective items, utilize neutral net to be trained obtaining score in predicting model SFM the set predicted the outcome of mark item rating and respective items；3) utilize basis recommendation method item of not marked to predict the outcome, the set that item of not marking predicts the outcome utilizes SFM obtain the final predictive value of item of not marking；4) according to the size of each predictive value in item of not marking final predictive value set, all items of not marking of user being carried out descending and obtain an ordered set of not marking, the front N item choosing an ordered set of not marking recommends user as recommendation results.The present invention can effectively reflect the truth that user evaluates, and improves the precision of personalized recommendation.

Description

A kind of based on the mixing recommendation method recommending probability fusion

Technical field

The invention belongs to e-commerce field, specifically a kind of based on the mixing recommendation method recommending probability fusion.

Background technology

Along with the fast development of ecommerce, information overload phenomenon is further serious.How to meet the individual demand of user based on the commodity set of magnanimity and become the major issue promoting Consumer's Experience, improving user satisfaction.Personalized recommendation system is the important means meeting users ' individualized requirement.Personalized recommendation system is according to the individual online browse data of user or buys data construct user interest preference model, thus recommend to meet the commodity of its unique need to user.Individual commodity recommendation is widely applied at e-commerce websites such as Amazon, store, Jingdone district, Taobaos, is effectively increased the probability that user buys, improves user's experience to website service.

Collaborative filtering (CollaborativeFiltering) technology is personalized recommendation application the earliest and the most successfully one of technology, and its main thought is based on its interest preference of user also this hypothesis identical with similar features and builds user interest preference model.Although the structure that existing research method can be personalized recommendation system is provided fundamental basis and practical advice, but yet suffers from many defects:

(1) what recommendation information represented is imperfect.Evaluating of commodity is generally represented or is predicted as a concrete numerical value by existing recommendation method by user, such as prediction user being evaluated as 3 points and namely think that the probability that commodity scoring is 3 points is 100% by user commodity.It practice, the evaluation of particular commodity is generally in a kind of uncertain state by user, as the evaluation of a certain commodity being generally " well ", " all right ", " pretty good ".The nondeterministic statement of commodity evaluation is expressed as by user user and provides the probability of different scoring for commodity, such as commodity are evaluated as the probability of 3 points, 4 points, 5 points respectively 30%, 40% and 20%, can effectively reflect the truth that user evaluates, the precision improving personalized recommendation is had active influence.User's grade form is shown as a concrete numerical value and have ignored the uncertainty that user evaluates by existing method, it is impossible to reflection user evaluates the truth of commodity, reduces the precision of commending system.

(2) the fusion problem of recommendation information.Collaborative filtering method based on user recommends, by calculating neighbor user, the article that the user that those and he have common interest to like likes, it is recommended that result focuses on the focus reflecting the small group similar with user interest；Project-based collaborative filtering method recommends, to user, the commodity that those are similar with the commodity that he selected in the past by calculating commodity neighbours, it is recommended that result focuses on the historical interest maintaining user.Produce personalized recommendation result from user and item perspective respectively based on the collaborative filtering of user, project-based collaborative filtering method, carry out the above results merging and the recommendation information of various angles can be comprehensively utilized, improve the precision of personalized recommendation.Existing research lacks the Unified frame that different angles recommendation information is merged.Such as, a kind of Collaborative Filtering Recommendation Algorithm based on Collaborative Filtering utilizes the input as the collaborative filtering method based on user that predicts the outcome that project-based collaborative filtering method obtains, although fully utilizing the collaborative filtering method based on user and project-based collaborative filtering method, but collaborative filtering result and the project-based collaborative filtering result based on user is not merged by it.

Summary of the invention

The present invention is the weak point overcoming prior art to exist, propose a kind of based on the mixing recommendation method recommending probability fusion, it is not only the recommendation results merging different recommendation method generation and provides unified framework, and can effectively reflect the truth that user evaluates, improve the precision of personalized recommendation.

In order to achieve the above object, the technical solution adopted in the present invention is:

A kind of feature based on the mixing recommendation method recommending probability fusion of the present invention is to carry out as follows:

Step one, bivariate table T={U, I, f} is used to represent the score data of commodity；

In described bivariate table T, U={U₁,...,U_u,...,U_|U|Represent user's set, I={I₁,...,I_i,...,I_|I|Representing commodity set, f represents user's scoring to commodity；

Described user gathers in U, total number that | U | is user, U_uRepresent the u user；In described commodity set I, total number that | I | is commodity, I_iRepresent i-th commodity；Assume user U_uTo commodity I_iGrading system S be { S₁,...,S_s,...,S_|S|, in described grading system S, mark S_sFor integer and S₁< ... < S_s< ... < S_|S|, S₁Represent the minimum scoring of commodity, S_|S|Represent the highest scoring of commodity；

In described bivariate table T, by the existing scoring item set of marking of all usersRepresent, For described user U_uScoring set, Represent user U_uI-th existing scoring item,Represent user U_uTotal number of existing scoring item；OrderFor described user U_uSet of not marking, Represent user U_uI-th do not mark item,Represent user U_uTotal number of item of not marking, make described user not mark setIn Arbitrary Term

Step 2, assume that user has marked setIn Arbitrary TermFor unknown number, set that user has been markedIn other utilize and obtain described Arbitrary Term respectively based on collaborative filtering method and the project-based collaborative filtering method of userNeighbor user predict the outcomeWith neighbours' project forecast result；Described neighbor user is predicted the outcomeWith neighbours' project forecast resultAs described Arbitrary TermThe item of scoring predict the outcomeUser is marked setIn the item of scoring of all items predict the outcome and predict the outcome set with item of markingRepresent；

Predict the outcome set by described item of having markedAs the input value of neutral net, set that described user has been markedAs the output valve of described neutral net, described neutral net is trained, it is thus achieved that score in predicting model SFM；

Step 3, set that user has been markedIn all items utilize the described collaborative filtering method based on user and project-based collaborative filtering method to obtain described user not mark setIn Arbitrary TermPreliminary forecasting resultDescribed user is not marked setIn predicting the outcome of all items predict the outcome set with item of not markingRepresent；Predict the outcome set by described item of not markingAs the input value of described score in predicting model SFM, described score in predicting model SFM is utilized to obtain the final predictive value set of item of not marking；

Step 4, by described user U_uAll items of not marking carry out descending according to the size of each predictive value in the final predictive value set of described item of not marking and obtain and do not mark an ordered set, the front N item of an ordered set of not marking described in choosing recommends described user U as recommendation results_u。

The present invention lies also in based on the feature of the mixing recommendation method recommending probability fusion:

In described step 2, the collaborative filtering method based on user is to carry out as follows:

1) by described user U_uThe user of all commodity is marked setWith other users, the set of scoring of all commodity is utilized respectively constrained Pearson came relatedness metric method and obtains user U_uWith user's similarity set of other users, by other users according to the similarity size in described user's similarity set carry out descending obtain preliminary neighbor user set N_u；

2) by described commodity I_iIn set of markingIn corresponding user be expressed as scoring user and gather R_i；

3) by described preliminary neighborhood N_uR is gathered with scoring user_iFront k the user occured simultaneously is expressed as neighbor user set N_ui；

4) by described neighbor user set N_uiIn each user to commodity I_iGrade form be shown as neighbor user scoring set F_ui；

5) formula (1) is utilized to obtain the scoring probability based on user

${\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)={\mathrm{Num}}_U\left(S_s\right)/k---\left(1\right)$

In formula (1), Num_U(S_s) gather F for the scoring of described neighbor user_uiMiddle scoring S_sThe number of times occurred；

6) by described scoring S_sWith the described scoring probability based on userConstitute neighbor user to predict the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^U=\{(S_1,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_{\left|S\right|}\right))\}---\left(2\right)$

In formula (2), ${\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)\&Element;\left[\mathrm{0,1}\right].$

Described project-based collaborative filtering method is to carry out as follows:

1) by all users to described commodity I_iThe set and the set of scoring of other commodity is utilized respectively constrained Pearson came relatedness metric method obtains described commodity I of marking_iWith the commodity similarity set of other commodity, by other commodity according to the similarity size in described commodity similarity set carry out descending obtain preliminary neighbours commodity set N_i；

2) by user U_uIn set of markingIn corresponding commodity list be shown as scoring commodity set R_u；

3) by described preliminary neighbours commodity set N_iWith scoring commodity set R_uFront k the commodity list occured simultaneously is shown as neighbours commodity set N_iu；

4) by user U_uTo described neighbours commodity set N_iuIn the grade form of each commodity be shown as neighbours' commodity scoring set F_iu；

5) formula (3) is utilized to obtain project-based scoring probability

${\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)={\mathrm{Num}}_I\left(S_s\right)/k---\left(3\right)$

In formula (3), Num_I(S_s) gather F for the scoring of described neighbours' commodity_iuMiddle scoring S_sThe number of times occurred；

6) by described scoring S_sWith described project-based scoring probabilityComposition project neighbours predict the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^I=\{(S_1,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_{\left|S\right|}\right))\}---\left(4\right)$

In formula (4), ${\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)\&Element;\left[\mathrm{0,1}\right].$

Compared with the prior art, beneficial effects of the present invention is embodied in:

1, the recommendation results each obtained based on user and project-based recommendation method is expressed as recommending probability by the present invention, neutral net is utilized respective obtained recommendation results to be merged, overcoming conventional hybrid recommends method recommendation information to represent incomplete problem, fusion for different angles recommendation information provides unified framework, and it recommends precision to be substantially better than the recommendation method based on user and project-based recommendation method.

2, the present invention utilizes scoring S_sWith scoring two tuple-sets that constitute of probability based on userRepresent the recommendation information that the recommendation method based on user obtains, utilize scoring S_sTwo tuple-sets constituted with project-based scoring probabilityRepresenting the recommendation information that project-based recommendation method obtains, compared with the evaluation table of commodity being shown as a concrete numerical value with by user, two tuples that scoring and scoring probability are constituted can reflect the truth that user evaluates more really.

3, the present invention utilizes and predicts the outcome based on the collaborative filtering method of user and project-based collaborative filtering method item of having been marked, item of marking is predicted the outcome set as inputting, using set of marking as exporting, the distinctive capability of fitting of neutral net is utilized to obtain score in predicting model SFM, it is ensured that the robustness of score in predicting model.

4, set of not marking is utilized collaborative filtering method and project-based collaborative filtering method based on user to obtain preliminary forecasting result by the present invention, preliminary forecasting result is not marked the final predictive value set of item as the input of score in predicting model SFM, and then obtain final recommendation results, compared with conventional recommendation method, difference can be recommended the recommendation results of method to carry out effective integration by the present invention, improves the precision of personalized recommendation.

5, the present invention carried mixing recommendation method by based on the collaborative filtering method of user and the result of project-based collaborative filtering method carried out merge be conducive to improve recommendation results multiformity, more conform to the actual preferences of user, overcome the simple shortcoming using the collaborative filtering method based on user or the simple project-based collaborative filtering method of use in prior art.

6, the present invention can be used for the personalized recommendation system of digital product, the travelling routes such as the entity products such as clothing and mobile phone, film and music and the service products such as arrangement of spending a holiday, it is possible to the platform such as webpage and App at computer and mobile phone uses, and has wide range of applications.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the present invention；

Fig. 2 is the sensitivity experiments result of method for measuring similarity；

Fig. 3 is neighbor user (commodity) number sensitivity experiments result；

Fig. 4 is forecasting accuracy experimental result；

Fig. 5 is prediction training pattern sensitivity experiments result.

Detailed description of the invention

The present invention utilizes bivariate table to represent the score data of commodity, to the Arbitrary Term marked in set as unknown number, utilize basis recommendation method to obtain predicting the outcome of respective items, utilize neutral net to be trained obtaining score in predicting model SFM the set predicted the outcome of mark item rating and respective items.Score in predicting model SFM is utilized to obtain the final predictive value of item of not marking the set that the item of not marking obtained by basis recommendation method is predicted the outcome.Finally, standard data set compares with basic algorithm.As it is shown in figure 1, the method for the embodiment of the present invention comprises the following steps:

Step one, use bivariate table T={U, I, f} to represent the score data of commodity, specifically include:

Such as table 1, U={U₁,...,U_u,...,U_|U|Represent user's set, I={I₁,...,I_i,...,I_|I|Representing commodity set, f represents user's scoring to commodity；

User gathers in U, total number that | U | is user, U_uRepresent the u user；In commodity set I, total number that | I | is commodity, I_iRepresent i-th commodity；Assume user U_uTo commodity I_iGrading system S be { S₁,...,S_s,...,S_|S|, in grading system S, mark S_sFor integer and S₁< ... < S_s< ... < S_|S|, S₁Represent the minimum scoring of commodity, S_|S|Represent the highest scoring of commodity；

Table 1

In bivariate table T, by the existing scoring item set of marking of all usersRepresent, For the scoring set of user Uu, Represent user U_uI-th existing scoring item,Represent user U_uTotal number of existing scoring item；OrderFor user U_uSet of not marking, Represent user U_uI-th do not mark item,Represent user U_uTotal number of item of not marking, make user not mark setIn Arbitrary Term

Step 2, utilize basis recommendation method to obtain predicting the outcome of respective items, will mark item rating and predicting the outcome of respective items utilizes neutral net to be trained obtaining score in predicting model SFM.Concrete steps include:

1) assume that user has marked setIn Arbitrary TermFor unknown number, set that user has been markedIn other utilize based on user collaborative filtering method obtain Arbitrary Term neighbor user predict the outcomeOther refer in set of marking ${\hat{F}}_u=\{{\hat{f}}_{u1},...,{\hat{f}}_{\mathrm{ui}},...,{\hat{f}}_{u\left|{\hat{F}}_u\right|}\}$ In, exceptAll items in addition；

1.1) by user U_uTo all commodity I={I₁,...,I_i,...,I_|I|User marked setWith other users, the set of scoring of all commodity is utilized respectively constrained Pearson came relatedness metric method and obtains user U_uWith user's similarity set of other users, by other users according to the similarity size in user's similarity set carry out descending obtain preliminary neighbor user set N_u, other users refer to and gather U={U user₁,...,U_u,...,U_|U|In except user U_uAll users in addition, the set of scoring of all commodity is referred in set of marking by other usersIn except user U_uScoring setAll users in addition mark set；The method for measuring similarity of user includes cosine similarity, Pearson came dependency and constrained Pearson came dependency three kinds, devising 5 groups of experiments in order to relatively more different method for measuring similarity the present invention is directed to for the impact of the inventive method precision of prediction, experimental result is as shown in Figure 2.In figure, abscissa represents data set, vertical coordinate represents experiment forecast error, be we can see that by curve in figure, for the standard data set in the present invention, error based on the method for measuring similarity of constrained Pearson came dependency is below cosine similarity metric method and Pearson correlation measure, thus being favorably improved the precision of prediction of the inventive method.

1.2) by commodity I_iIn set of markingIn corresponding user be expressed as scoring user and gather R_i；

1.3) by preliminary neighbor user set N_uR is gathered with scoring user_iFront k the user occured simultaneously is expressed as neighbor user set N_ui；The selection of k is the key factor affecting the collaborative filtering recommending method effect based on user, and in order to verify the impact on the inventive method precision of prediction of the neighbor user number, the present invention devises 7 groups of experiments, often group is tested and is chosen k neighbor user respectively, k=10, and 20, ..., 80；And calculate precision of prediction.Experimental result as it is shown on figure 3, in figure abscissa represent neighbor user number, vertical coordinate represents experiment forecast error.When neighbor user number k is less, it was predicted that probability is accurate not, reduces the accuracy of prediction；When neighbor user number k is bigger, the similarity between user is not high, also can reduce the accuracy of algorithm predicts.Therefore, for the adopted standard data set of this experiment, neighbor user number can obtain good prediction effect when selecting between [30,70]；For other data sets, the determination of best neighbor number of users depends on concrete data cases.

1.4) by neighbor user set N_uiIn each user to commodity I_iGrade form be shown as neighbor user scoring set F_ui；

1.5) formula (1) is utilized to obtain the scoring probability based on user, the user U that namely obtains with the collaborative filtering method based on user_uTo commodity I_iGrading system be S_sProbability:

${\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)={\mathrm{Num}}_U\left(S_s\right)/k---\left(1\right)$

In formula (1), Num_U(S_s) gather F for the scoring of described neighbor user_uiMiddle scoring S_sThe number of times occurred；

1.6) will mark S_sWith the scoring probability based on userConstitute neighbor user to predict the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^U=\{(S_1,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_{\left|S\right|}\right))\}---\left(2\right)$

In formula (2), ${\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)\&Element;\left[\mathrm{0,1}\right].$

2) assume that user has marked setIn Arbitrary TermFor unknown number, set that user has been markedIn other utilize project-based collaborative filtering method obtain Arbitrary Term project neighbours predict the outcome, other refer in set of markingIn, except i-th existing scoring itemAll items in addition；

2.1) user is gathered U to commodity I_iThe set and the set of scoring of other commodity is utilized respectively constrained Pearson came relatedness metric method obtains commodity I of marking_iWith the commodity similarity set of other commodity, by other commodity according to the similarity size in commodity similarity set carry out descending obtain preliminary commodity neighborhood N_i；Other commodity refer at commodity set I={I₁,...,I_i,...,I_|I|In except commodity I_iAll commodity in addition.

2.2) by user U_uIn set of markingIn corresponding commodity list be shown as scoring commodity set R_u；

2.3) by preliminary commodity neighborhood N_iWith scoring commodity set R_uFront k the commodity list occured simultaneously is shown as commodity neighborhood N_iu；The selection of k is the key factor affecting project-based collaborative filtering recommending method effect, and in order to verify the impact on the inventive method precision of prediction of neighbours' item number, the present invention devises 7 groups of experiments, often k neighbours' project is chosen in group experiment respectively, k=10, and 20, ..., 80；And calculate precision of prediction.Experimental result as it is shown on figure 3, in figure abscissa represent neighbours' item number, vertical coordinate represents experiment forecast error.When neighbours item number k is less, it was predicted that probability is accurate not, reduces the accuracy of prediction；When neighbours item number k is bigger, the similarity between project is not high, also can reduce the accuracy of algorithm predicts.Therefore, for the adopted standard data set of this experiment, neighbor user number can obtain good prediction effect when selecting between [30,70]；For other data sets, the determination of best neighbor number of users depends on concrete data cases.

2.4) by user U_uTo commodity neighborhood N_iuIn the grade form of each commodity be shown as neighbours' commodity scoring set F_iu；

2.5) formula (3) is utilized to obtain project-based scoring probability, the user U that namely obtains with project-based collaborative filtering method_uTo commodity I_iGrading system be S_sProbability:

${\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)={\mathrm{Num}}_I\left(S_s\right)/k---\left(3\right)$

In formula (3), Num_I(S_s) gather F for the scoring of neighbours' commodity_iuMiddle scoring S_sThe number of times occurred；

2.6) will mark S_sWith project-based scoring probabilityComposition project neighbours predict the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^I=\{(S_1,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_{\left|S\right|}\right))\}---\left(4\right)$

In formula (4),

3) neighbor user is predicted the outcomeWith neighbours' project forecast resultAs Arbitrary TermThe item of scoring predict the outcomeSpecifically, the item of scoring of Arbitrary Term predict the outcome into:

${\hat{f}}_{\mathrm{ui}}^0=\{(S_1,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_{\left|S\right|}\right)),(S_1,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_{\left|S\right|}\right))\}$

All items in set of marking are performed step 1) and step 2), by set of markingIn predicting the outcome of all items predict the outcome set with item of markingRepresent；

4) predict the outcome set by item of markingAs the input value of neutral net, by set of markingAs the output valve of neutral net, neutral net is trained, it is thus achieved that score in predicting model SFM.In the present embodiment, the neutral net adopted refers to radial base neural net.Radial base neural net has stronger input and output mapping function, has the characteristic that unique the best is approached and learning process fast convergence rate.

Step 3, set that user has been markedIn all itemsThe collaborative filtering method based on user and project-based collaborative filtering method is utilized to obtain set of not markingIn Arbitrary TermPreliminary forecasting resultTo not mark setIn all itemsPredict the outcome and predict the outcome set with item of not markingRepresent；Item of not marking predicts the outcome setAs the input value of score in predicting model SFM, score in predicting model SFM is utilized to export the final predictive value set of item of not marking；

Step 4, by user U_uAll items of not markingCarrying out descending according to the size of each predictive value in item of not marking final predictive value set and obtain an ordered set of not marking, the front N item choosing an ordered set of not marking recommends user U as recommendation results_u, N represents recommendation number, can according to specifically recommending scene settings.

Carry out experimental demonstration for the inventive method, specifically include:

1) standard data set is prepared

The present invention uses MovieLens data set to verify that probability fusion recommends the effectiveness of method as standard data set, and MovieLens data set is widely used personalized recommendation data set.In MovieLens data set, the film that oneself has been seen by user is marked, and score value is 1 to 5 points, and data set includes 943 isolated users, 1682 films, 100000 scorings.Training set and test set adopt the rule of 80%/20% to split, and namely randomly choose 80000 scorings as training set, and 20000 scorings are as test set.

2) evaluation index

Adopt mean absolute error (MAE) as the evaluation index of the present embodiment.Mean absolute deviation MAE is by calculating the accuracy of the deviation measurement prediction between user's scoring and the final predictive value of respective items actual in test set, and MAE is more little, it is recommended that quality is more high.If actual user marks, set is { p₁,...,p_l,...,p_n, corresponding predictive value set expression is { q₁,...,q_l,...,q_n, then mean absolute error is defined as formula (5):

$\mathrm{MAE}=\frac{{\mathrm{\&Sigma;}}_{l=1}^n|q_l-p_l|}{n}---\left(5\right)$

3) test on standard data set

In order to verify the effectiveness of institute of the present invention extracting method, herein it is modeled on 5 group data sets of MovieLens data set and predicts, and will predict the outcome and truly mark and compare.As shown in Figure 4, in figure, abscissa represents data set sequence number to experimental result, and vertical coordinate represents experiment forecast error.With the collaborative filtering method based on user and based on compared with the collaborative filtering method of product, the forecast error of the method for the present invention is below the method based on user and project-based method, thus all can obtain more excellent predictablity rate on each data set.

For verifying the robustness of institute of the present invention extracting method, herein by the hidden layer neuron quantity and the hidden layer learning function that change radial base neural net, separately design 7 groups of experiments and be verified.Experimental result as it is shown in figure 5, in figure abscissa represent experiment sequence number, vertical coordinate represents experiment forecast error.As seen from Figure 5, changing hidden layer neuron quantity and the hidden layer learning function of radial base neural net, the forecast error of the inventive method there will be certain change；But, under rational hidden layer neuron quantity and hidden layer learning function are arranged, the forecast error of context of methods is consistently lower than the method based on user and project-based method, thus the present invention is better than the collaborative filtering method based on user and the collaborative filtering recommending method based on product.

Claims (1)

1. recommend a method based on the mixing recommending probability fusion, it is characterized in that carrying out as follows:

Step one, bivariate table T={U, I, f} is used to represent the score data of commodity；

In described bivariate table T, U={U₁,...,U_u,...,U_|U|Represent user's set, I={I₁,...,I_i,...,I_|I|Representing commodity set, f represents user's scoring to commodity；

Described user gathers in U, total number that | U | is user, U_uRepresent the u user；In described commodity set I, total number that | I | is commodity, I_iRepresent i-th commodity；Assume user U_uTo commodity I_iGrading system S be { S₁,...,S_s,...,S_|S|, in described grading system S, mark S_sFor integer and S₁< ... < S_s< ... < S_|S|, S₁Represent the minimum scoring of commodity, S_|S|Represent the highest scoring of commodity；

In described bivariate table T, by the existing scoring item set of marking of all usersRepresent, For described user U_uScoring set, Represent user U_uI-th existing scoring item,Represent user U_uTotal number of existing scoring item；OrderFor described user U_uSet of not marking, Represent user U_uI-th do not mark item,Represent user U_uTotal number of item of not marking, make described user not mark setIn Arbitrary Term

Step 2, assume that user has marked setIn Arbitrary TermFor unknown number, set that user has been markedIn other utilize and obtain described Arbitrary Term respectively based on collaborative filtering method and the project-based collaborative filtering method of userNeighbor user predict the outcomeWith neighbours' project forecast resultDescribed neighbor user is predicted the outcomeWith neighbours' project forecast resultAs described Arbitrary TermThe item of scoring predict the outcomeUser is marked setIn the item of scoring of all items predict the outcome and predict the outcome set with item of markingRepresent；

Predict the outcome set by described item of having markedAs the input value of neutral net, set that described user has been markedAs the output valve of described neutral net, described neutral net is trained, it is thus achieved that score in predicting model SFM；

Step 3, set that user has been markedIn all items utilize the described collaborative filtering method based on user and project-based collaborative filtering method to obtain described user not mark setIn Arbitrary TermPreliminary forecasting resultDescribed user is not marked setIn predicting the outcome of all items predict the outcome set with item of not markingRepresent；Predict the outcome set by described item of not markingAs the input value of described score in predicting model SFM, described score in predicting model SFM is utilized to obtain the final predictive value set of item of not marking；

Step 4, by described user U_uAll items of not marking carry out descending according to the size of each predictive value in the final predictive value set of described item of not marking and obtain and do not mark an ordered set, the front N item of an ordered set of not marking described in choosing recommends described user U as recommendation results_u；

In described step 2, the collaborative filtering method based on user is to carry out as follows:

1) by described user U_uThe user of all commodity is marked setWith other users, the set of scoring of all commodity is utilized respectively constrained Pearson came relatedness metric method and obtains user U_uWith user's similarity set of other users, by other users according to the similarity size in described user's similarity set carry out descending obtain preliminary neighbor user set N_u；

2) by described commodity I_iIn set of markingIn corresponding user be expressed as scoring user and gather R_i；

3) by described preliminary neighborhood N_uR is gathered with scoring user_iFront k the user occured simultaneously is expressed as neighbor user set N_ui；

4) by described neighbor user set N_uiIn each user to commodity I_iGrade form be shown as neighbor user scoring set F_ui；

5) formula (1) is utilized to obtain the scoring probability based on user

${\mathrm{Pr}}_{ui}^U\left(S_s\right)={\mathrm{Num}}_U\left(S_s\right)/k---\left(1\right)$

In formula (1), Num_U(S_s) gather F for the scoring of described neighbor user_uiMiddle scoring S_sThe number of times occurred；

6) by described scoring S_sWith the described scoring probability based on userConstitute neighbor user to predict the outcome

${\mathrm{Pr}}_{ui}^U=\{(S_1,{\mathrm{Pr}}_{ui}^U(S_1\left)\right),...,(S_s,{\mathrm{Pr}}_{ui}^U(S_s\left)\right),...,(S_{\left|S\right|},{\mathrm{Pr}}_{ui}^U(S_{\left|S\right|}\left)\right)\}---\left(2\right)$

In formula (2), ${\mathrm{Pr}}_{ui}^U\left(S_s\right)\&Element;\&lsqb;0,1\&rsqb;;$

Described project-based collaborative filtering method is to carry out as follows:

1) by all users to described commodity I_iThe set and the set of scoring of other commodity is utilized respectively constrained Pearson came relatedness metric method obtains described commodity I of marking_iWith the commodity similarity set of other commodity, by other commodity according to the similarity size in described commodity similarity set carry out descending obtain preliminary neighbours commodity set N_i；

2) by user U_uIn set of markingIn corresponding commodity list be shown as scoring commodity set R_u；

3) by described preliminary neighbours commodity set N_iWith scoring commodity set R_uFront k the commodity list occured simultaneously is shown as neighbours commodity set N_iu；

4) by user U_uTo described neighbours commodity set N_iuIn the grade form of each commodity be shown as neighbours' commodity scoring set F_iu；

5) formula (3) is utilized to obtain project-based scoring probability

${\mathrm{Pr}}_{ui}^I\left(S_s\right)={\mathrm{Num}}_I\left(S_s\right)/k---\left(3\right)$

In formula (3), Num_I(S_s) gather F for the scoring of described neighbours' commodity_iuMiddle scoring S_sThe number of times occurred；

6) by described scoring S_sWith described project-based scoring probabilityComposition project neighbours predict the outcome

${\mathrm{Pr}}_{ui}^I=\{(S_1,{\mathrm{Pr}}_{ui}^I(S_1\left)\right),...,(S_s,{\mathrm{Pr}}_{ui}^I(S_s\left)\right),...,(S_{\left|S\right|},{\mathrm{Pr}}_{ui}^I(S_{\left|S\right|}\left)\right)\}---\left(4\right)$

In formula (4), ${\mathrm{Pr}}_{ui}^I\left(S_s\right)\&Element;\&lsqb;0,1\&rsqb;.$