CN103632290A - Recommendation probability fusion based hybrid recommendation method - Google Patents

Abstract

The invention discloses a recommendation probability fusion based hybrid recommendation method. The method includes the following steps: (1) using a two-dimensional table for representing scoring data of commodities; (2) taking any items in a scored set as unknowns, utilizing a basic recommendation method to obtain a prediction result of each corresponding item, training sets of scores of the scored items and the forecast results of the corresponding items by the aid of a neural network so as to obtain an SFM (score forecast model); (3) utilizing the basic recommendation method to obtain non-scored item forecast results, and utilizing the SFM to obtain a final forecast value of each non-scored item of a set of the non-scored item forecast results; (4) sorting all the non-scored items of a user according to size of each forecast value in a final forecast value set in a descending manner to obtain a non-scored item sorted set, and selecting top N items of the non-scored item sorted set as recommendation results to be recommended to the user. By the method, true situation of user evaluation can be effectively reflected, and precision in personalized recommendation is improved.

Description

A kind of mixing recommend method based on recommending probability fusion

Technical field

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

Background technology

Along with the fast development of ecommerce, information overload phenomenon is further serious.The individual demand how the commodity set based on magnanimity meets user becomes and promotes the major issue that user experiences, improves user satisfaction.Personalized recommendation system is the important means that meets users ' individualized requirement.Personalized recommendation system builds user interest preference model according to the online browse data of user's individuality or purchase data, thereby to user, recommends to meet the commodity of its unique need.Personalized commercial is recommended in the e-commerce websites such as Amazon, store, Jingdone district, Taobao and is widely applied, and has effectively improved the possibility that user buys, and has promoted the experience of user to website service.

Collaborative filtering (Collaborative Filtering) technology is the earliest and the most successfully one of technology of personalized recommendation application, its main thought be its interest preference of user based on thering is similar features also identical this hypothesis build user interest preference model.Although existing research method can be provided fundamental basis and practical advice for the structure of personalized recommendation system, still has many defects:

(1) what recommendation information represented is imperfect.Existing recommend method conventionally represents user or is predicted as a concrete numerical value to the evaluation of commodity, as predictive user thinks that to being evaluated as of commodity scoring is that the possibility of 3 minutes is 100% to user to commodity for 3 minutes.In fact, user to the evaluation of particular commodity conventionally in a kind of uncertain state, as the evaluation to a certain commodity is generally " well ", " all right ", " pretty good ".User is expressed as to user to the nondeterministic statement of commodity evaluation and for commodity, provides the possibility of different scorings, for example commodity are evaluated as to the possibility of 3 minutes, 4 minutes, 5 minutes and are respectively 30%, 40% and 20%, can effectively reflect the truth that user evaluates, to improving the precision of personalized recommendation, there is active influence.Existing method is shown as a concrete numerical value by user's grade form and has ignored the uncertainty that user evaluates, and cannot reflect that user evaluates the truth of commodity, has reduced the precision of commending system.

(2) the fusion problem of recommendation information.The article that collaborative filtering method based on user recommends those and him to have the user of common interest hobby to like by calculating neighbor user, recommendation results focuses on the focus of the microcommunity that reflection is similar with user interest; Project-based collaborative filtering method is recommended those and the similar commodity of commodity that he selected in the past by calculating commodity neighbours to user, and recommendation results focuses on the historical interest of maintaining user.Collaborative filtering based on user, project-based collaborative filtering method produce personalized recommendation result from user and project angle respectively, and the above results is merged and can the recommendation information of various angles be fully utilized, and improve the precision of personalized recommendation.In existing research, lack the Unified frame that different angles recommendation information is merged.For example, a kind of Collaborative Filtering Recommendation Algorithm based on Collaborative Filtering utilizes predicting the outcome as the input of the collaborative filtering method based on user that project-based collaborative filtering method obtains, although fully utilized collaborative filtering method and project-based collaborative filtering method based on user, it does not merge the collaborative filtering result based on user and project-based collaborative filtering result.

Summary of the invention

The present invention overcomes the weak point that prior art exists, a kind of mixing recommend method based on recommending probability fusion is proposed, not only for merging the recommendation results of different recommend methods generations, provide 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:

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

Step 1, use bivariate table T={U, I, f} represents the score data of commodity;

In described bivariate table T, U={U ₁..., U _u..., U _{| U|}represent that user gathers, I={I ₁..., I _i..., I _{| I|}representing commodity set, f represents the scoring of user to commodity;

Described user gathers in U, | total number that U| is user, U _urepresent u user; In described commodity set I, | total number that I| is commodity, I _irepresent i commodity; Suppose user U _uto commodity I _igrading system S be { S ₁..., S _s..., S _{| S|}, in described grading system S, scoring S _sfor integer and S ₁< ... < S _s< ... < S _{| S|}, S ₁represent the minimum scoring of commodity, S _{| S|}represent the highest scoring of commodity;

In described bivariate table T, by existing scoring set of marking for item of all users

represent,

for described user U _uthe set of scoring,

represent user U _uthe existing scoring of i,

represent user U _utotal number of existing scoring; Order

for described user U _unot scoring set,

represent user U _uscoring of i,

represent user U _utotal number of not scoring, make the described user set of not marking in Arbitrary Term

Step 2, suppose user's set of having marked in Arbitrary Term

for unknown number, by user's set of having marked

in collaborative filtering method and the project-based collaborative filtering method of other utilizations based on user obtain respectively described Arbitrary Term

neighbor user predict the outcome

with neighbours' project forecast result

; Described neighbor user is predicted the outcome

with neighbours' project forecast result

as described Arbitrary Term

the item of scoring predict the outcome

by user's set of having marked

in the set that predicts the outcome with marking that predicts the outcome of all items of scoring

represent;

By the described item set that predicts the outcome of having marked

as the input value of neural network, by the described user set of having marked

as the output valve of described neural network, described neural network is trained, obtain score in predicting model SFM;

Step 3, by user's set of having marked in all described collaborative filtering method based on user and project-based collaborative filtering methods of utilizing obtain the described user set of not marking

in Arbitrary Term

tentative prediction result

by the described user set of not marking

in all predicting the outcome with the set that predicts the outcome of scoring not

represent; By the set that predicts the outcome of described scoring item

as the input value of described score in predicting model SFM, utilize described score in predicting model SFM to obtain a not scoring final predicted value set;

Step 4, by described user U _uall scorings according to the size of each predicted value in a described scoring final predicted value set, carry out a descending sort acquisition ordered set of not marking, described in choosing, the front N item of a scoring ordered set is not recommended described user U as recommendation results _u.

The feature that the present invention is based on the mixing recommend method of recommending probability fusion is also:

Collaborative filtering method based on user in described step 2 is to carry out as follows:

1) by described user U _uto the user of all commodity set of having marked

to the set of scoring of all commodity, utilize respectively constrained Pearson came relatedness metric method to obtain user U with other users _uwith user's similarity set of other users, other users are carried out to descending sort according to the similarity size in the set of described user's similarity and obtain preliminary neighbor user set N _u;

2) by described commodity I _iin the set of marking

in corresponding subscriber's meter be shown scoring user gather R _i;

3) by described preliminary neighborhood N _ugather R with scoring user _ifront k the subscriber's meter occuring simultaneously is shown neighbor user set N _ui;

4) by described neighbor user set N _uiin each user to commodity I _igrade form be shown neighbor user scoring set F _ui;

5) utilize formula (1) 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) be described neighbor user scoring set F _uimiddle scoring S _sthe number of times occurring;

6) by described scoring S _swith the described scoring probability based on user

forming neighbor user predicts the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^U=\{({\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000427982560000022.png,HDA0000427982560000031.png"\; state="\{\{state\}\}">S</figure-callout>}}_1,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_{\left|S\right|}\right))\}---\left(2\right)$

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 _imark set and to the set of scoring of other commodity, utilize respectively constrained Pearson came relatedness metric method to obtain described commodity I _iwith the commodity similarity set of other commodity, other commodity are carried out to descending sort according to the similarity size in the set of described commodity similarity and obtain preliminary neighbours' commodity set N _i;

2) by user U _uin the set of marking in corresponding commodity list be shown scoring commodity set R _u;

3) by described preliminary neighbours' commodity set N _iwith scoring commodity set R _ufront k the commodity list occuring simultaneously is shown neighbours' commodity set N _iu;

4) by user U _uto described neighbours' commodity set N _iuin the grade form of each commodity be shown neighbours' commodity scoring set F _iu;

5) utilize formula (3) 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) be described neighbours' commodity scoring set F _iumiddle scoring S _sthe number of times occurring;

6) by described scoring S _swith described project-based scoring probability

formation project neighbours predict the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^I=\{({\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000427982560000022.png,HDA0000427982560000031.png"\; state="\{\{state\}\}">S</figure-callout>}}_1,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_{\left|S\right|}\right))\}---\left(4\right)$

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

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

1, the present invention is expressed as recommendation probability by the recommendation results that obtained separately based on user and project-based recommend method, utilize neural network that the recommendation results that obtained is separately merged, overcome conventional hybrid recommend method recommendation information and represented incomplete problem, for the fusion of different angles recommendation information provides unified framework, it recommends precision to be obviously better than recommend method and the project-based recommend method based on user.

2, the present invention utilizes scoring S _stwo tuple-sets that form with scoring probability based on user

represent the recommendation information that the recommend method based on user obtains, utilize scoring S _stwo tuple-sets that form with project-based scoring probability

represent the recommendation information that project-based recommend method obtains, and user is shown to a concrete numerical value to the evaluation table of commodity compares, two tuples that scoring and scoring probability form can reflect the truth that user evaluates more really.

3, the collaborative filtering method of utilization of the present invention based on user and project-based collaborative filtering method obtain marking and predict the outcome, the item of marking is predicted the outcome to set as input, using the set of marking as output, utilize the distinctive capability of fitting of neural network to obtain score in predicting model SFM, guaranteed the robustness of score in predicting model.

4, the present invention utilizes collaborative filtering method and project-based collaborative filtering method based on user to obtain tentative prediction result the set of not marking, input using tentative prediction result as a score in predicting model SFM final predicted value set that obtains not marking, and then obtain final recommendation results, compare with traditional recommend method, the present invention can carry out effective integration to the recommendation results of different recommend methods, has improved the precision of personalized recommendation.

5, the present invention carries mixing recommend method the result of the collaborative filtering method based on user and project-based collaborative filtering method has been carried out merging the diversity that is conducive to improve recommendation results, the actual preferences that more meets user, has overcome the shortcoming of using merely the collaborative filtering method based on user in prior art or using merely project-based collaborative filtering method.

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 clothes and mobile phone, film and music and the service products such as arrangement of spending a holiday, can use at platforms such as the webpage of computer and mobile phone and App, have wide range of applications.

Accompanying drawing explanation

Fig. 1 is 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.

Embodiment

The present invention utilizes bivariate table to represent the score data of commodity, to mark set in Arbitrary Term as unknown number, utilize basic recommend method to obtain predicting the outcome of respective items, utilize neural network to train the set predicting the outcome of mark item rating and respective items and obtain score in predicting model SFM.The set that the not scoring item obtaining by basic recommend method is predicted the outcome utilizes score in predicting model SFM to obtain the not final predicted value of scoring item.Finally, on standard data set, compare with basic algorithm.As shown in Figure 1, the method for the embodiment of the present invention comprises the following steps:

Step 1, use bivariate table T={U, I, f} represents the score data of commodity, specifically comprises:

As table 1, U={U ₁..., U _u..., U _{| U|}represent that user gathers, I={I ₁..., I _i..., I _{| I|}representing commodity set, f represents the scoring of user to commodity;

User gathers in U, | total number that U| is user, U _urepresent u user; In commodity set I, | total number that I| is commodity, I _irepresent i commodity; Suppose user U _uto commodity I _igrading system S be { S ₁..., S _s..., S _{| S|}, in grading system S, scoring S _sfor integer and S ₁< ... < S _s< ... < S _{| S|}, S ₁represent the minimum scoring of commodity, S _{| S|}represent the highest scoring of commodity;

Table 1

In bivariate table T, by existing scoring set of marking for item of all users represent,

for the set of scoring of user Uu,

represent user U _uthe existing scoring of i,

represent user U _utotal number of existing scoring; Order for user U _unot scoring set,

represent user U _uscoring of i,

represent user U _utotal number of not scoring, make user's set of not marking

in Arbitrary Term

Step 2, utilize basic recommend method to obtain predicting the outcome of respective items, utilize neural network to train the predicting the outcome of item rating and respective items of marking and obtain score in predicting model SFM.Concrete steps comprise:

1) suppose user's set of having marked in Arbitrary Term

for unknown number, by user's set of having marked

in the collaborative filtering method of other utilizations based on the user neighbor user that obtains Arbitrary Term predict the outcome

other fingers are in the set of marking ${\hat{F}}_u=\{{\hat{f}}_{\mathrm{<figure-callout\; id="1"\; label="u"\; filenames="HDA0000427982560000022.png,HDA0000427982560000031.png"\; state="\{\{state\}\}">u</figure-callout>}1},...,{\hat{f}}_{\mathrm{ui}},...,{\hat{f}}_{u\left|{\hat{F}}_u\right|}\}$ In, except

all items in addition;

1.1) by user U _uto all commodity I={I ₁..., I _i..., I _{| I|}user's set of having marked

to the set of scoring of all commodity, utilize respectively constrained Pearson came relatedness metric method to obtain user U with other users _uwith user's similarity set of other users, other users are carried out to descending sort according to the similarity size in the set of user's similarity and obtain preliminary neighbor user set N _u, other users refer to user and gather U={U ₁..., U _u..., U _{| U|}in except user U _uall users in addition, other users refer in the set of marking the set of scoring of all commodity

in except user U _uthe set of scoring the set of marking of all users in addition; User's method for measuring similarity comprises three kinds of cosine similarity, Pearson came correlativity and constrained Pearson came correlativitys, for more different method for measuring similarity the present invention is directed to and designed 5 groups of experiments for the impact of the inventive method precision of prediction, experimental result as shown in Figure 2.In figure, horizontal ordinate represents data set, ordinate represents to test predicated error, in figure curve we can find out, for the standard data set in the present invention, the error of the method for measuring similarity based on constrained Pearson came correlativity is all lower than cosine method for measuring similarity and Pearson's relativity measurement method, thereby contributes to improve the precision of prediction of the inventive method.

1.2) by commodity I _iin the set of marking

in corresponding subscriber's meter be shown scoring user gather R _i;

1.3) by preliminary neighbor user set N _ugather R with scoring user _ifront k the subscriber's meter occuring simultaneously is shown neighbor user set N _ui; The selection of k is the key factor of the collaborative filtering recommending method effect of impact based on user, and in order to verify the impact of neighbor user number on the inventive method precision of prediction, the present invention has designed 7 groups of experiments, k neighbor user, k=10,20 are chosen respectively in every group of experiment,, 80; And calculate precision of prediction.As shown in Figure 3, in figure, horizontal ordinate represents neighbor user number to experimental result, and ordinate represents to test predicated error.When neighbor user, count k hour, prediction probability is accurate not, has reduced the accuracy of prediction; When neighbor user, count k when larger, the similarity between user is not high, also can reduce the accuracy of algorithm predicts.Therefore,, for standard data set that this experiment adopts, when being chosen between [30,70], neighbor user number can obtain good prediction effect; For other data sets, best neighbor number of users definite depends on concrete data cases.

1.4) by neighbor user set N _uiin each user to commodity I _igrade form be shown neighbor user scoring set F _ui;

1.5) utilize formula (1) to obtain the scoring probability based on user

, the user U that uses the collaborative filtering method based on user to obtain _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) be described neighbor user scoring set F _uimiddle scoring S _sthe number of times occurring;

1.6) S that will mark _swith the scoring probability based on user

forming neighbor user predicts the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^U=\{({\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000427982560000022.png,HDA0000427982560000031.png"\; state="\{\{state\}\}">S</figure-callout>}}_1,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_{\left|S\right|}\right))\}---\left(2\right)$

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

2) suppose user's set of having marked

in Arbitrary Term

for unknown number, by user's set of having marked

in other project neighbours that utilize project-based collaborative filtering method to obtain Arbitrary Term predict the outcome

, other fingers are in the set of marking

in, except i existing scoring

all items in addition;

2.1) user is gathered to U to commodity I _ithe set and utilize respectively constrained Pearson came relatedness metric method to obtain commodity I to the set of scoring of other commodity of marking _iwith the commodity similarity set of other commodity, other commodity are carried out to descending sort according to the similarity size in the set of commodity similarity and obtain preliminary commodity neighborhood N _i; Other commodity refer at commodity set I={I ₁..., I _i..., I _{| I|}in except commodity I _iall commodity in addition.

2.2) by user U _uin the set of marking in corresponding commodity list be shown scoring commodity set R _u;

2.3) by preliminary commodity neighborhood N _iwith scoring commodity set R _ufront k the commodity list occuring simultaneously is shown commodity neighborhood N _iu; The selection of k is the key factor of the project-based collaborative filtering recommending method effect of impact, and in order to verify the impact of neighbours' item number on the inventive method precision of prediction, the present invention has designed 7 groups of experiments, k neighbours' project, k=10,20 are chosen respectively in every group of experiment,, 80; And calculate precision of prediction.As shown in Figure 3, in figure, horizontal ordinate represents neighbours' item number to experimental result, and ordinate represents to test predicated error.When neighbours' item number k hour, prediction probability is accurate not, has reduced the accuracy of prediction; When neighbours' item number k is larger, the similarity between project is not high, also can reduce the accuracy of algorithm predicts.Therefore,, for standard data set that this experiment adopts, when being chosen between [30,70], neighbor user number can obtain good prediction effect; For other data sets, best neighbor number of users definite depends on concrete data cases.

2.4) by user U _uto commodity neighborhood N _iuin the grade form of each commodity be shown neighbours' commodity scoring set F _iu;

2.5) utilize formula (3) to obtain project-based scoring probability , the user U obtaining 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) be neighbours' commodity scoring set F _iumiddle scoring S _sthe number of times occurring;

2.6) S that will mark _swith project-based scoring probability

formation project neighbours predict the outcome

${\mathrm{Pr}}_{\mathrm{ui}}^I=\{({\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000427982560000022.png,HDA0000427982560000031.png"\; state="\{\{state\}\}">S</figure-callout>}}_1,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_{\left|S\right|}\right))\}---\left(4\right)$

In formula (4),

3) neighbor user is predicted the outcome

with neighbours' project forecast result

as Arbitrary Term

the item of scoring predict the outcome

particularly, the item of scoring of Arbitrary Term predict the outcome into:

${\hat{f}}_{\mathrm{ui}}^0=\{({\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000427982560000022.png,HDA0000427982560000031.png"\; state="\{\{state\}\}">S</figure-callout>}}_1,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_{\left|S\right|}\right)),({\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000427982560000022.png,HDA0000427982560000031.png"\; state="\{\{state\}\}">S</figure-callout>}}_1,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_1\right)),...,(S_s,{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)),...,(S_{\left|S\right|},{\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_{\left|S\right|}\right))\}$

To all execution steps 1 in the set of marking) and step 2), by the set of marking

in all set that predict the outcome with marking that predict the outcome

represent;

4) by the item set that predicts the outcome of marking

as the input value of neural network, by the set of marking

as the output valve of neural network, neural network is trained, obtain score in predicting model SFM.In the present embodiment, the neural network adopting 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, by user's set of having marked

in all

collaborative filtering method and the project-based collaborative filtering method of utilization based on user obtains not scoring set in Arbitrary Term

tentative prediction result

the set of not marking

in all

predict the outcome with the set that predicts the outcome of scoring item not

represent; The item set that predicts the outcome of not marking

as the input value of score in predicting model SFM, utilize a not scoring final predicted value set of score in predicting model SFM output;

Step 4, by user U _uall scoring

according to the size of each predicted value in a scoring final predicted value set not, carry out a descending sort acquisition ordered set of not marking, choose the not front N item of a scoring ordered set and recommend user U as recommendation results _u, N represents to recommend number, can set according to concrete recommendation scene.

For the inventive method, carry out experimental demonstration, specifically comprise:

1) prepare standard data set

The present invention uses MovieLens data set as the validity of standard data set checking probability fusion recommend method, and MovieLens data set is widely used personalized recommendation data set.In MovieLens data centralization, the film that user has seen oneself is marked, and score value is 1 to 5 minute, and data set comprises 943 isolated users, 1682 films, 100000 scorings.The rule of training set and test set employing 80%/20% is cut apart, and selects at random 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 user's scoring and the deviation measurement prediction between the final predicted value of respective items actual in test set, and MAE is less, recommends quality higher.If actual user marks, set is { p ₁..., p _l..., p _n, corresponding predicted value set expression is { q ₁..., q _l..., q _n, 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) on standard data set, test

In order to verify the validity of institute of the present invention extracting method, herein on 5 group data sets of MovieLens data set, carry out model and forecast, and will predict the outcome and compare with true scoring.As shown in Figure 4, in figure, horizontal ordinate represents data set sequence number to experimental result, and ordinate represents to test predicated error.Compare with the collaborative filtering method based on product with the collaborative filtering method based on user, the predicated error of method of the present invention is all lower than the method based on user and project-based method, thereby all can obtain more excellent predictablity rate on each data set.

Robustness for checking institute of the present invention extracting method, by changing hidden layer neuron quantity and the hidden layer learning function of radial base neural net, has designed respectively 7 groups and has verified herein.As shown in Figure 5, in figure, horizontal ordinate represents to test sequence number to experimental result, and ordinate represents to test predicated error.As seen from Figure 5, change hidden layer neuron quantity and the hidden layer learning function of radial base neural net, the predicated error of the inventive method there will be certain variation; But, under reasonably hidden layer neuron quantity and hidden layer learning function arrange, the predicated error of this paper method is all the time lower than the method based on user and project-based method, thereby the present invention is better than collaborative filtering method and the collaborative filtering recommending method based on product based on user.

Claims (3)

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

Step 1, use bivariate table T={U, I, f} represents the score data of commodity;

In described bivariate table T, U={U ₁..., U _u..., U _{| U|}represent that user gathers, I={I ₁..., I _i..., I _{| I|}representing commodity set, f represents the scoring of user to commodity;

Described user gathers in U, | total number that U| is user, U _urepresent u user; In described commodity set I, | total number that I| is commodity, I _irepresent i commodity; Suppose user U _uto commodity I _igrading system S be { S ₁..., S _s..., S _{| S|}, in described grading system S, scoring S _sfor integer and S ₁< ... < S _s< ... < S _{| S|}, S ₁represent the minimum scoring of commodity, S _{| S|}represent the highest scoring of commodity;

In described bivariate table T, by existing scoring set of marking for item of all users

represent,

for described user U _uthe set of scoring,

represent user U _uthe existing scoring of i,

represent user U _utotal number of existing scoring; Order

for described user U _unot scoring set,

represent user U _uscoring of i,

represent user U _utotal number of not scoring, make the described user set of not marking in Arbitrary Term

Step 2, suppose user's set of having marked

in Arbitrary Term

for unknown number, by user's set of having marked

in collaborative filtering method and the project-based collaborative filtering method of other utilizations based on user obtain respectively described Arbitrary Term

neighbor user predict the outcome

with neighbours' project forecast result

described neighbor user is predicted the outcome

with neighbours' project forecast result

as described Arbitrary Term the item of scoring predict the outcome

by user's set of having marked

in the set that predicts the outcome with marking that predicts the outcome of all items of scoring

represent;

By the described item set that predicts the outcome of having marked

as the input value of neural network, by the described user set of having marked

as the output valve of described neural network, described neural network is trained, obtain score in predicting model SFM;

Step 3, by user's set of having marked

in all described collaborative filtering method based on user and project-based collaborative filtering methods of utilizing obtain the described user set of not marking in Arbitrary Term

tentative prediction result

by the described user set of not marking

in all predicting the outcome with the set that predicts the outcome of scoring not represent; By the set that predicts the outcome of described scoring item

as the input value of described score in predicting model SFM, utilize described score in predicting model SFM to obtain a not scoring final predicted value set;

Step 4, by described user U _uall scorings according to the size of each predicted value in a described scoring final predicted value set, carry out a descending sort acquisition ordered set of not marking, described in choosing, the front N item of a scoring ordered set is not recommended described user U as recommendation results _u.

2. the mixing recommend method based on recommending probability fusion according to claim 1, is characterized in that: the collaborative filtering method based on user in described step 2 is to carry out as follows:

1) by described user U _uto the user of all commodity set of having marked

to the set of scoring of all commodity, utilize respectively constrained Pearson came relatedness metric method to obtain user U with other users _uwith user's similarity set of other users, other users are carried out to descending sort according to the similarity size in the set of described user's similarity and obtain preliminary neighbor user set N _u;

2) by described commodity I _iin the set of marking in corresponding subscriber's meter be shown scoring user gather R _i;

3) by described preliminary neighborhood N _ugather R with scoring user _ifront k the subscriber's meter occuring simultaneously is shown neighbor user set N _ui;

4) by described neighbor user set N _uiin each user to commodity I _igrade form be shown neighbor user scoring set F _ui;

5) utilize formula (1) 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) be described neighbor user scoring set F _uimiddle scoring S _sthe number of times occurring;

6) by described scoring S _swith the described scoring probability based on user

forming neighbor user predicts 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), ${\mathrm{Pr}}_{\mathrm{ui}}^U\left(S_s\right)\&Element;\left[\mathrm{0,1}\right].$

3. the mixing recommend method based on recommending probability fusion according to claim 1, is characterized in that: described project-based collaborative filtering method is to carry out as follows:

1) by all users to described commodity I _imark set and to the set of scoring of other commodity, utilize respectively constrained Pearson came relatedness metric method to obtain described commodity I _iwith the commodity similarity set of other commodity, other commodity are carried out to descending sort according to the similarity size in the set of described commodity similarity and obtain preliminary neighbours' commodity set N _i;

2) by user U _uin the set of marking

in corresponding commodity list be shown scoring commodity set R _u;

3) by described preliminary neighbours' commodity set N _iwith scoring commodity set R _ufront k the commodity list occuring simultaneously is shown neighbours' commodity set N _iu;

4) by user U _uto described neighbours' commodity set N _iuin the grade form of each commodity be shown neighbours' commodity scoring set F _iu;

5) utilize formula (3) 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) be described neighbours' commodity scoring set F _iumiddle scoring S _sthe number of times occurring;

6) by described scoring S _swith described project-based scoring probability formation 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), ${\mathrm{Pr}}_{\mathrm{ui}}^I\left(S_s\right)\&Element;\left[\mathrm{0,1}\right].$

Images