CN101271559A - Cooperation recommending system based on user partial interest digging - Google Patents

Abstract

The invention discloses a cooperation recommendation system basing on the local interest mining of users. The system firstly mines a project assembly which is similar with the projects of a target user that are not scored. Then, the scoring similarity of the project assembly of the target user and other users is calculated, and hereby the nearest neighbor is calculated. Moreover, the score of the nearest neighbor on the project which is not scored for the target user is used for predicting the score of the project which is not scored for the target user. The project assembly with high score is recommended to the target user. The cooperation recommendation system basing on the local interest mining of the users improves the recommendation precision, increases the satisfaction of the users; the system is applied to IPTV television programs, electronic commerce websites, etc., and can recommend information resources in which the users are interested.

Description

A kind of cooperation recommending system based on user partial interest digging

Technical field

The present invention relates to the personalized service field of network technology, particularly a kind of cooperation recommending system based on user partial interest digging.

Background technology

Personalized recommendation system is widely used in television program recommendations and e-commerce industry, automatically recommend programs of interest information and commodity resource to the user, alleviated " information overload " problem to a certain extent, and online " viewer " can have been changed into " buyer ".Recommending module is a device most crucial in the commending system, and existing recommended technology mainly contains the recommendation of content-based filtration, based on the recommendation of collaborative filtering and based on the combined recommendation of these two kinds of methods.The recommendation efficient and the precision of content-based filtration are low, and can not find new information of interest for the user.Technology based on cooperation is a most successful current recommended technology, and its basic thought is to give oneself other similar user's interest project recommendations of interest-degree.This technological core content is to seek the similar neighbours user of interest-degree for the targeted customer.Suppose user A and B, to the evaluation of estimate of n project resource ( [1,5] interval integer, expression is disliked very much respectively, relatively dislikes, and is general, prefers, and is delithted with.) can be expressed as two vectorial a (a ₁, a ₂, L, a _n) and b (b ₁, b ₂, L, b _n), whether similar standard is to calculate the linear dependence degree of evaluation vector a and b according to related coefficient to weigh two user interests, yet, along with increasing of user's evaluation of estimate, two overall similarities that the user showed sharply reduce, and promptly two all similar possibilities of people's overwhelming majority interest are very little.Therefore, along with sharply increasing of system user number and item number, when this method calculating targeted customer's similar interests neighbours user collected, computation complexity increased, and accuracy reduces, and the extendability of system and recommendation quality descend thereupon.

Summary of the invention

A kind of cooperation recommending system based on user partial interest digging of providing at the deficiencies in the prior art is provided, this system significantly reduces the data volume that participates in calculating, improve system effectiveness, recommend the precision height and have versatility, the recommendation, the e-commerce website shopping that can be used for network TV program are recommended.

The concrete technical scheme that realizes the object of the invention is:

Suppose to have M user and N information project, target of prediction user a is to the interest scores of scoring item not:

(1), through the data pretreatment units, can obtain the attribute matrix Matrix of information project _Attr(N ' S) and user are to the rating matrix Matrix of project _Rating(M ' N).The attribute item of the corresponding project resource of the row of attribute matrix, on behalf of information project, the row of matrix have which attribute.The value Matrix of matrix _Attr(i, j)

0, and 1}, value is that 1 expression project i has attribute j, otherwise does not have this attribute; Scoring Matrix _RatingCapable representative of consumer, row have been represented resource item, the value Matrix of matrix _Rating(i, j)

{ 0,1,2,3,4,5} represents the interest-degree of user i to project j.

Target of prediction user a is to the score value of scoring item k not:

(2), according to the Resource Properties matrix, calculate prediction (target) project k with each project l (0＜l＜N﹠amp; l ¹K) attribute similarity, computing method are as follows:

Here, | attr (k)

Attr (l) | the attribute sum that expression project k and l have,

The attribute number that expression project k has jointly with l.The threshold value of d for being provided with, promptly Gong You attribute outnumbers this threshold value, just thinks that project k is 1 with l attribute similarity.

(3), according to the user to the project rating matrix, calculate prediction project k with sundry item l (0＜l＜N﹠amp; l ¹K) similarity, computing method are as follows:

Here, r _{U, k}, r _{U, l}Represent the scoring of user u respectively to project k, l.r _k, r _lThe average score of representing project k, l respectively. Expression all has user's set of scoring to project k, l.

(4), calculate the size of prediction project, computing method are as follows with the sundry item similarity according to formula (1), (2):

Sim(k，l)＝(1-l)Sim _attr(k，l)+lSim _rating(k，l)(3)

$l\hat{I}\left[\mathrm{0,1}\right]$

Parameter p is set, calculates with preceding p the highest project set ISet of the similarity of prediction project k.

(5), based on rating matrix Matrix _Rating, calculate targeted customer a with the similarity of other user u on partial interest collection ISet, computing formula is as follows:

Here, r _{A, i}Be element in the rating matrix, expression user a is to the score value of project i, r _aThe average score of expression user a.

(6), a parameter q is set, calculate preceding q the most similar user on partial interest collection ISet according to formula (3) with user a, gather USet as the neighbours of user a.

(7), gather USet, predictive user a is to the score value of project k according to the neighbours of a:

(8), repeating step (2)-(7), measurable targeted customer a is to all not scorings of scoring item, then the highest Item Sets of scoring recommended targeted customer a.

Compare with background technology, the present invention has following advantage:

1, at the not scoring item that will predict, user's rating matrix is carried out partial interest filter, significantly reduce the data volume that participates in calculating, improve system effectiveness.

2,, calculate the partial interest similarity between the user, the accuracy height based on local scoring item.

3, the degree of accuracy of recommend method improves.

4, the inventive method has versatility, and the recommendation, the e-commerce website shopping that can be used for network TV program are recommended.

Description of drawings

Fig. 1 is a structural representation of the present invention

Fig. 2 is recommending module apparatus structure synoptic diagram among the present invention

Fig. 3 is a process flow diagram of the present invention

Fig. 4 is for calculating destination item neighbour's process flow diagram

Fig. 5 is for calculating targeted customer neighbour's process flow diagram

Specific embodiments

Now describe in detail in conjunction with the accompanying drawings:

Consult Fig. 1, device 110 is to the raw data pre-service, obtains recommending required structural data and storage.

Device 120 is the attribute feature vectors that extract resource item 110 from installing, and makes up the attribute matrix and the storage of project.For example the attribute for the IPTV program can be divided into following 25 kinds: action, risk, animation, comedy, science fiction, love, war, terror, biography, animation, crime, record, the story of a play or opera, family, illusion, black, history, music, song and dance, mystery, short-movie, terrible, western, variety, TV play.Suppose that the corresponding above-mentioned attribute vector of certain TV programme Movie1 and Movie2 is followed successively by (1 expression has this attribute, and 0 expression does not have this attribute):

a(1，1，0，0，1，1，0，0，0，0，0，0，1，0，0，0，0，0，0，1，0，1，0，0，0)

b(1，0，0，1，1，1，0，0，0，0，1，0，1，1，0，0，0，0，0，1，0，1，0，0，0)

Device 130 is from installing 110 the user that the extracts interest scores data to project, obtain the scoring vector of user to all items, making up user-project rating matrix and storage.For example user a, b are followed successively by (1: dislike very much 2: relatively dislike 3: general, 4: prefer 5: be delithted with) to n item resource item scoring vector:

R ₁(1，3，5，4，5，5，1，…，4，5，5)

R ₂(5，2，1，4，5，5，5，…，1，2，5)

Device 140 its inputs are: item attribute matrix and user-project rating matrix.Output is: recommend each user's interest bulleted list.

Device 150 represents the bulleted lists of recommending to the user from the interface, and the feedback information of recording user upgrades user interest, i.e. user interest scores matrix in the updating device 130.

Consult Fig. 2, recommending module device 140 of the present invention comprises interest scores filtration unit 210, targeted customer neighbour generation device 220 and cooperation recommending device 230.

Interest scores filtration unit 210, it is based on the project resource attribute matrix, at not scoring item for targeted customer prediction, filter out and this dissimilar project of scoring item not, calculate with not similar neighbour's project (partial interest project) set of scoring item.

Targeted customer neighbour generation device 220 is based on the neighbour who the scoring of installing the similar neighbour's project of the 210 same not scoring items that produce is calculated the targeted customer.

Cooperation recommending device 230, the neighbour who is based on the targeted customer of device 220 generations calculates the targeted customer to the not prediction scoring of scoring item, and produces recommendation list according to the height of prediction scoring.

Consult Fig. 3, be recommended flowsheet figure of the present invention.

(1), at first, step 310 is set up resource item attribute matrix Matrix _Attr(N ' S), this is to be finished by said apparatus 120.Step 320 is set up user-project rating matrix Matrix _Rating(M ' N), this is to be finished by said apparatus 130.

(2), secondly, step 330 is calculated neighbour's Item Sets of the not scoring item k that will predict.Concrete calculation process is consulted shown in Figure 4, among Fig. 4:

1., step 410 and 420 attribute vector of extracting objects project and sundry item from the item attribute matrix respectively, for example the destination item attribute vector is a, b is other a certain purpose attribute vectors:

a(1，1，0，0，1，1，0，0，0，0，0，0，1，0，0，0，0，0，0，1，0，0，0，0，0)

b(1，0，0，1，0，1，0，0，0，0，1，0，1，1，0，0，0，0，0，1，0，0，1，0，0)

2., step 430 is calculated the attribute similarity of destination item with sundry item according to formula (1), parameter in the formula (1) $d=\frac{10}{2}=5:$

${\mathrm{Sim}}_{\mathrm{attr}}(\stackrel{\‾}{a},\stackrel{\‾}{b})=\frac{4}{10}=0.4$

Calculate the scoring similarity of destination item according to formula (2) with sundry item.

3., step 440 calculates with the destination item similarity according to formula (3), and chooses that (the set ISet of individual project of 1＜p＜N) is the destination item neighbour with the most similar p of destination item.

(3), then, step 340 is calculated targeted customer's neighbour user, and idiographic flow is consulted Fig. 5, among Fig. 5:

1., the destination item neighbour that obtained by step 330,320 successively of step 510,520 gathers ISet and the user rating matrix Matrix to project _Rating(M ' N).

2., step 530: rating matrix Matrix _Rating(M ' N) is the scoring of user to all projects, extracts the score data of project among the pair set ISet wherein here, obtains Matrix _Rating(the sub-rating matrix of M ' N).

3., step 540: based on the submatrix that step 530 obtains, calculate the similarity of other users, calculate according to formula (4) with the scoring of targeted customer on ISet.Hypothetical target project set ISet={i ₁, i ₂, i ₃, i ₄, i ₅, i ₆, targeted customer a and certain user b correspondence scoring vector on this ISet is:

R ₁(5，4，5，4，5，1)

R ₂(5，4，4，5，5，2)

The average score of user a and b is followed successively by 4,4, then by calculating formula of similarity (4), obtains

The rest may be inferred, can calculate the similarity S set Set={s of targeted customer a with every other user ₁, s ₂, s ₃, L, s _M-1}

4., step 550: Set sorts from big to small to the similarity S set, choose the similarity maximum preceding q (the individual user of 1＜q＜M) is as the arest neighbors of targeted customer a:

USet＝{u ₁，u ₂，u ₃，L，u _p-1}

(4), then, step 350 target of prediction user a is according to neighbour user the scoring of a to project k to be predicted in the scoring of project k to the not scoring of scoring item k: the neighbour who supposes to obtain targeted customer a is USet={u ₁, u ₂, u ₃, u ₄, u ₅, u ₆, and be followed successively by based on the similarity that the project neighbour collects ISet with user a: 1,0.8,0.6,0.9,1,0.8}; Neighbour USet is followed successively by the scoring of scoring item k not: and 5,5,4,4,5,4}; Neighbour USet average score separately is followed successively by: { 3,4,4,4,4,3, }, the average score of targeted customer a is 4; Then according to formula (3) target of prediction user a to being calculated as follows of scoring item k not:

$=4+\frac{1\×(5-3)+0.8(5-4)+0.6\×(4-4)+0.9\×(4-4)+1\×(5-4)+0.8\×(4-3)}{1+0.8+0.6+0.9+1+0.8}$

$=4.90$

The rest may be inferred, and measurable calculating targeted customer a is to other score values of scoring item not.

(5), last, step 360 produces recommendation list: suppose that here step 350 dopes targeted customer a to scoring item { k not ₁, k ₂, k ₃, k ₄, k ₅, k ₆Predicted value be followed successively by: { 3.1,2.5,4.9,4.6,5,2.5} then according to the height of predicted value, screens interested project and recommends, and can obtain { k ₃: 4.9, k ₄: 4.6, k ₅: 5}, the regular { k that turns to rounds up ₃: 5, k ₄: 5, k ₅: 5}, expression { k ₃: be delithted with k ₄: be delithted with k ₅: be delithted with, therefore project { k ₃, k ₄, k ₅Recommend targeted customer a.

The rest may be inferred, repeats flow process 330 to flow process 360, can be each user in predicting to the interest-degree of scoring item not and recommend interested project resource.

The present invention can be applied in the IPTV television program recommendation system, the TV programme historical record data of utilizing the user to watch, can obtain the interest-degree scoring of user to some program, based on these score data in conjunction with this recommend method, can predictive user like the TV programme resource seen in the future, recommend the user by the form of menu.The present invention also can be applied to e-commerce website and recommend, as online shopping site, historical record by the recording user purchase, can obtain the user to buying the interest-degree scoring of commodity, utilize these score data and in conjunction with this recommend method, can recommend the high commodity of interest-degree to the user, saving the user seeks the time of oneself liking commodity on the one hand, can improve the benefit of shopping website on the one hand again.

Claims (2)

1, a kind of cooperation recommending system based on user partial interest digging is characterized in that it comprises:

A data pretreatment unit, be used for the Collection and analysis project resource information characteristics, user browse record and user hobby scoring record to project resource;

An item attribute memory storage is used to construct the attribute matrix of project resource, and matrix column is all Property Names of project resource, the attribute record of each project of behavior of matrix;

A user interest storage matrix is used for the scoring of recording user to project, and matrix column is all items resource name, and each user of the behavior of matrix is to the scoring of project resource;

A recommending module device comes the interested not scoring item of predictive user according to the existing rating matrix of user, produces recommendation list according to the prediction interest level;

A user interaction means is used for the feedback information of recording user, upgrades user's interest.

2, commending system according to claim 1 is characterized in that described recommending module device comprises:

A, interest scores filtration unit, be based on project resource attribute matrix and user's rating matrix, at for the not scoring item of targeted customer prediction, filter out and this dissimilar project of scoring item not, calculate with the similar neighbour's project set of scoring item not;

D, targeted customer neighbour generation device are based on the neighbour who the scoring of the similar neighbour's project of the same not scoring item of interest scores filtration unit generation is calculated the targeted customer;

C, cooperation recommending device, be based on the resulting targeted customer's of targeted customer neighbour generation device neighbour and the scoring of scoring item is not calculated the targeted customer to the not prediction scoring of scoring item, and produce recommendation list according to the height of the prediction scoring of calculating.

Images