CN106528643A - Social network based multi-dimension comprehensive recommending method - Google Patents

Abstract

The invention relates to a social network based multi-dimension comprehensive recommending method. The recommending method includes: establishing a social network model and a community model; S2: defining the user compactness, the user entity interest degree, and the entity similarity, wherein the user compactness reflects the compactness of a relation between two users, the user entity interest degree reflects the interest degree of a user in an entity, and the entity similarity reflects the consistent degree of attributes of two entities; and S3: performing multi-dimension entity comprehensive recommending on users in a social network according to the user compactness, the user entity interest degree, and the entity similarity. Compared with the prior art, the method can evaluate the influence of commerce websites of the social network from the user relation dimension, the entity similarity dimension, and the user interest dimension, and can perform relative entity recommending, and has wide actual application prospect.

Description

A kind of various dimensions combined recommendation method based on social networkies

Technical field

The present invention relates to social networkies technical field, more particularly, to a kind of various dimensions combined recommendation based on social networkies Method.

Background technology

In the evolution process that human society is little by little carried out activity from original physical activity to internet virtual space, Interpersonal contacts relation gradually constitutes an interlaced complex network, and social networkies also become the Internet One of major application.In recent years, as people are increasing to the demand for obtaining information and transmission information, and social networkies are Meet the good medium of these demands of people, so that social networkies have obtained rapid development.What is produced therewith is all kinds of Social network sites (such as：QQ, Facebook, MySpace etc.) be able to it is growing, nowadays also become the life of each age level, Work, the main platform of amusement.In such a case, the commercial product recommending based on social networkies is common in ecommerce and online knowledge Extremely it is widely applied in the website such as enjoying, such as：The business web sites such as U.S. group, Taobao, while also gradually causing increasingly The concern of many related researchers.

Often according to some fast sales singly carrying out commercial product recommending in traditional commercial product recommending, other of commodity are have ignored Association attributes, and fail in view of huge social networkies can be also constituted between consumer to carry out various dimensions recommendation.This will Cause recommendation rate low, the DeGrain of recommendation.First, user, calmly can be according to desired commodity when a certain commodity are retrieved Association attributes reach and coincide substantially the lower single operation that can just carry out with the thing desired by heart；Secondly, the social activity constituted by user Network can also be used, and social networkies are constituted, this social network mainly due to reasons such as work, study, blood relationship, amusements Common community users in network probably possess similar hobby enjoyment.Accordingly, it would be desirable to these factors are considered by commending system Enter, improved to reach higher recommendation rate.

In order that recommendation effect becomes more perfect, just arisen at the historic moment based on the commercial product recommending of social networkies.People by The community network of an interlaced complexity is constituted in relations such as work, amusement, study, blood relationships, it is clear that very possible friend Between there are some similar hobby features, when a certain individual in the social networkies have purchased a certain part commodity, at this moment, base Can be according to the association attributes of the commodity (such as in the commercial product recommending of social networkies：Type, price, quality, sales volume) generate a business Product candidate collection, then generates consumer's candidate collection according to the social networkies constituted by the consumer, using related calculation Method some in commodity candidate collection are met into the commercial product recommending of certain condition to consumer's candidate collection in a certain consumer. Obviously, the traditional commercial product recommending system of comparison, has merged social networkies based on the commercial product recommending system of social networkies, has examined That what is considered is more comprehensive, can reach surely better recommendation effect, also improve recommendation rate to a certain extent.The present invention is main A kind of various dimensions combined recommendation method based on social networkies to be proposed for problem above.

The content of the invention

The purpose of the present invention is exactly to provide a kind of based on social networkies to overcome the defect of above-mentioned prior art presence Various dimensions combined recommendation method, respectively from customer relationship dimension, entity similar dimension and user interest dimension, to society Net business web site is handed over to affect to be estimated, so as to the recommendation for preferably carrying out related entities, with wide actual application prospect.

The purpose of the present invention can be achieved through the following technical solutions：

A kind of various dimensions combined recommendation method based on social networkies is comprised the following steps：

S1：Set up social networkies model G=(V, E) and community model C=<CV,CE>, wherein, V is represented in social networkies G User's set, E represents the set of the relation between user, and CV possesses user's set of same interest hobby in representing community C,CE represents the set of the relation between the user for possessing same interest hobby,Possess same interest hobby User composition community C；

S2：User's compactness, user subject interest-degree and entity similarity are defined, user's compactness reflects two use Frequency of interaction aggregative indicator between mutual reliability index, user between the tightness degree of relation between family, with comment stability indicator, user It is relevant with common neighbours and similar community's index, the level of interest that the user subject interest-degree reflection user holds to entity, It is relevant with user interest evidence behavior, the consistent degree of the possessed attribute of two entities of the entity similarity reflection, with entity Type index of similarity, entity price index of similarity, entity quality index of similarity are relevant with entity sales volume index of similarity；

S3：According to user's compactness, user subject interest-degree and entity similarity, carry out to the user in social networkies many The entity integrated recommendation of dimension.

Step S3 is specially：

301：When a certain user buys entityWhen, the initial markers user is originating subscriber O,

302：Entity is obtained according to entity similarityEntity Candidate Recommendation set Represent entity to be recommended Item ' and entityBetween entity similarity, γ represents similarity threshold；

303：Good friend set friend (O) of originating subscriber is obtained according to the social networkies friend relation of originating subscriber；

User's Candidate Recommendation set of originating subscriber is obtained further according to user's compactness closeness(v_{R_user}, O) represent user v to be recommended_{R_user}And the user's compactness between originating subscriber O, α is compactness threshold value；

304：Judge user v to be recommended_{R_user}Treat the user subject interest-degree EI (v of recommended entity item '_{R_user}| Item ' ∈ R_item) whether it is more than interest-degree threshold value beta, if so, then to user v to be recommended_{R_user}Recommend entity to be recommended Item ', if it is not, then do not do recommending；

305：All recommended users to be recommended for crossing entity in step 304 are updated to into originating subscriber, jump procedure 303, terminate after all users in social networkies are traversed recommendation.

User's compactness meets below equation：

In formula, closeness (v_su,v_tu)、cf_L(v_su,v_tu)、mf_L(v_su,v_tu)、if_L(v_su,v_tu) and ncf_L (v_su,v_tu) source user v is represented respectively_suWith targeted customer v_tuBetween user's compactness, comment stability indicator, between user mutually Frequency of interaction aggregative indicator and common neighbours and similar community's index between reliability index, user.

The comment stability indicator meets below equation：

In formula, cf_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween comment stability indicator, COM (v_su,v_tu) represent source user v_suFor targeted customer v_tuThe comment set delivered, | COM (v_su,v_tu) | represent comment set COM(v_su,v_tu) the middle total quantity commented on, com (v_su,v_tu)_iRepresent source user v_suFor targeted customer v_tuI-th comment Numerical value,Represent in comment set COM (v_su,v_tu) in all comments meansigma methodss；

Between the user, mutual reliability index meets below equation：

In formula, mf_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween user between mutual reliability index,For three-number set,Represent source user v_suTo targeted customer v_tuThe meansigma methodss of all evaluations of estimate,Represent source user v_suIn targeted customer v_tuThe information scales forwarded in the information of all issues,Represent source user v_suIn targeted customer v_tuThe information scales of thumb up are carried out in the information of all issues,For three-number set,Represent targeted customer v_tuTo source user v_suThe meansigma methodss of all evaluations of estimate,Represent targeted customer v_tuIn source user v_suThe information scales forwarded in the information of all issues,Represent targeted customer v_tuIn source user v_suThe information scales of thumb up are carried out in the information of all issues；

Between the user, frequency of interaction aggregative indicator meets below equation：

In formula, if_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween user between frequency of interaction comprehensively refer to Mark, | Unit | represent the sum of unit interval piece, l_kRepresent k-th unit interval piece u_kBetween interior two users, frequency of interaction refers to Scale value,K-th unit interval piece u is represented respectively_kInterior source user v_suFor targeted customer v_tuInteraction times, average interaction time length, average interaction time interval,Show k-th unit interval piece u respectively_kIt is endogenous User v_suMaximum interaction times, most long interaction time length for itself all neighbor user, the mutual time interval of most short delivery；

The common neighbours and similar community's index meet below equation：

In formula, ncf_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween common neighbours and similar community refer to Mark,Represent source user v_suThe community at place,Represent targeted customer v_tuThe community at place, i-th common communityIn, source user v_suWith targeted customer v_tuRespective neighbor user set is respectively CC_i(v_su) and CC_i (v_tu), | CC_i| represent CC_iInterior number of users, | CC_i(v_su)∩CC_i(v_tu) | represent CC_i(v_su) and CC_i(v_tu) in common factor Number of users, | CC_i(v_su)∪CC_i(v_tu) | represent CC_i(v_su) and CC_i(v_tu) number of users in union.

The user subject interest-degree meets below equation：

In formula, and EI (v | item) user subject interest-degrees of the user v to entity item is represented, n represents user plane to entity The interest evidence behavior total degree occurred during item, δ_iThe i-th interest evidence row occurred when representing user plane to entity item For the interesting evidence behavior of institute occurred to user's history is classified, D_jJth class interest evidence classification is represented, is directly emerging Interesting evidence, right (D_j) represent jth class interest evidence classification D_jWeight, P_v(D_j| item) represent for user v, work as entity When item occurs, interest evidence classification D_jThe probability of appearance.

There is following relation between the weight of different interest evidence classifications：

In formula, right (D_q) represent q class interest evidence classifications D_qWeight, ID represent user's history occur it is all emerging The set of interesting evidence classification, D_q,D_j∈ ID, | ID | represent the total quantity of interest evidence classification, p (D_q) represent interest evidence classification D_qThe probability occurred in user goes over interest evidence behavior, L (D_q) represent and q class interest evidence classifications D_qWhat is be linked is emerging The quantity of interesting evidence classification, Link (D_j) represent and jth class interest evidence classification D_jThe interest evidence category set being linked, D_q →D_jRepresent D_jIt is by D_qAnd cause, i.e. D_qWith D_jIt is linked.

The entity similarity meets below equation：

In formula, item is used_aAnd item_bTo represent two different entities, sim (item_a,item_b)、sim_type (item_a,item_b)、sim_price(item_a,item_b)、sim_sale(item_a,item_b) difference presentation-entity item_aWith reality Body item_bEntity similarity, entity type index of similarity, entity price index of similarity, entity sales volume index of similarity, sim_quality(item_a) presentation-entity item_aEntity quality index of similarity.

The entity type index of similarity is calculated based on entity catalog classification tree, meets below equation：

In formula, sim_type (item_a,item_b) presentation-entity item_aWith entity item_bEntity type similarity refer to Mark, c_aPresentation-entity item_aLevel in entity catalog classification tree, c_bPresentation-entity item_bIn entity catalog classification tree Level, distance (item_a,item_b) represent the entity item in entity catalog classification tree_aWith entity item_bBetween path Length；

The entity price index of similarity meets below equation：

In formula, sim_price (item_a,item_b) presentation-entity item_aWith entity item_bEntity price similarity refer to Mark, price_aPresentation-entity item_aPrice, price_bPresentation-entity item_bPrice；

The entity quality index of similarity meets below equation：

In formula, sim_quality (item_a) presentation-entity item_aEntity quality index of similarity, score_aRepresent real Body item_aEntity quality score, score_maxRepresent the highest score in all entity quality scores；

The entity sales volume index of similarity meets below equation：

In formula, sim_sale (item_a,item_b) presentation-entity item_aWith entity item_bEntity sales volume similarity refer to Mark, sale_aPresentation-entity item_aTotal sales volume, sale_bPresentation-entity item_bTotal sales volume.

Compared with prior art, the present invention has advantages below：

1. traditional technology recommended based on single standard is different from, and the present invention carries out social networkies based on various dimensions In entity recommend：From customer relationship dimension, it is standard according to user's compactness, enters between the close user of choice relation Row is recommended；From entity similar dimension, it is standard according to entity similarity, the entity for selecting similarity high is recommended；From User interest dimension is set out, and is standard according to user subject interest-degree, and the entity for selecting user's Interest Measure high is recommended, will Entity is recommended closely to merge with social networkies internal information, it is considered to which ground is more comprehensive, can reach more preferably recommendation effect, have Wide actual application prospect；

2. the recommendation process of the present invention is set up and recommends Rule of judgment buying the originating subscriber of entity as starting, progressively to The proximal subscribers recommended entity of originating subscriber, so as to reach the purpose of the recommended entity of all users, recommended range is wide, will not go out Now omit, reach the purpose of real-time recommendation, and the selection of recommended and recommended entity is accurate.

3., in user's compactness index calculating process proposed by the invention, stable, user's mutual trust journey is mutually commented using user Four aspects such as degree, user mutual frequency, the common neighbours of user and community are calculated, and the compactness of customer relationship is obtained more For comprehensive assessment；

4. in user subject interest-degree calculating process proposed by the invention, using the dependency number of user's direct action evidence Calculate according to carrying out, and the weighing computation method based on association is set up between user interest evidence, it is ensured that make user interest degree meter Calculate accuracy；

5., in the calculating of entity similarity proposed by the invention, four aspects are taken into account：Entity type similarity, entity valency Position similarity, entity quality similarity and entity sales volume similarity, make entity similarity obtain more comprehensive calculating.

6. weight also higher original of the present invention based on the interest evidence class of the higher interest evidence classification of weight its association Reason, the relation existed between the weight for providing different interest evidence classifications, after the initial value of weighted value is given at random, can be according to pass Be final weighted value can be calculated after formula interative computation, compare, the interest evidence that the present invention is used The weight of classification can more accurately react the component shared by interest evidence behavior so that the numerical value of user subject interest-degree It is more accurate.

Description of the drawings

Configuration diagrams of the Fig. 1 for the inventive method；

Recommendation simple process figures of the Fig. 2 for the inventive method.

Specific embodiment

The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with technical solution of the present invention Premised on implemented, give detailed embodiment and specific operating process, but protection scope of the present invention be not limited to Following embodiments.

The present invention provides defined below to social networkies：

Define 1 social networkies model：Social networkies G can turn to a figure with form, G=(V, E), and V represents social networkies In user's set, and use v_iRepresent that user is individual, v_iMiddle subscript i represents the numbering of user, and E is used for describing the relation between user Set；

Define 2 community models：Community C is made up of the user for possessing same interest hobby, is that of social networkies is sub Figure, C=<CV,CE>, it is expressed as when user belongs to community：CV possesses the use of same interest hobby in representing community C Family is gathered, and the relation between community users is expressed as：CE represents the relation between the user for possessing same interest hobby Set；

Define 3 user's compactnesss：Reflect the tightness degree of relation between two users.Calculating the compactness between user (it is designated as：closeness(v_su,v_tu), wherein v_suExpression source user, and v_tuRepresent targeted customer) when, present invention primarily contemplates four Individual aspect：Mutually common neighbours and similar community between frequency of interaction, user between reliability, user between the stability of comment, user；

Define 4 user subject interest-degrees：The entity interest-degree EI (v that user holds to entity_i|item_k) reflection user for A certain entity item_kLevel of interest, item_kThe numbering of middle subscript k presentation-entity, in the present invention, entity interest is by user's Expressed by direct action or other direct testimony of eyewitnesss, for example：The behaviors such as evaluation, browsing time, thumb up, forwarding.

Define 5 entity similarities：Entity similarity represents the consistent degree of the possessed attribute of two entities.Two entities Similarity is designated as sim (item_a,item_b), wherein item_a、item_bRepresent two different entities.The main base of entity similarity Represented four attributes in community set are carried out, and entity attribute is expressed as { type, price, quality, sale }, namely Affiliated type type, price price including entity, quality quality, sales volume sale.

A kind of various dimensions combined recommendation method based on social networkies is comprised the following steps：

S1：Set up social networkies model G=(V, E) and community model C=<CV,CE>, and define user's compactness, user The concept of entity interest-degree and entity similarity；

S2：According to the definition evaluation of user's compactness, user subject interest-degree and entity similarity；

Wherein, user's compactness reflects the tightness degree of relation between two users, between comment stability indicator, user Mutually between reliability index, user, frequency of interaction aggregative indicator is relevant with common neighbours and similar community's index, user subject interest-degree The level of interest that reflection user holds to entity, relevant with user interest evidence behavior, entity similarity reflects two entity institutes Possess the consistent degree of attribute, with entity type index of similarity, entity price index of similarity, entity quality index of similarity It is relevant with entity sales volume index of similarity；

S3：According to user's compactness, user subject interest-degree and entity similarity, carry out to the user in social networkies many The entity integrated recommendation of dimension.

In step S2, (1) calculates user's compactness

1) comment on stability indicator

Comment stability indicator reflects the comment numerical value that source user goes out given in passing all previous comment to targeted customer Fluctuation status.In the present invention, if comment stability value is higher, reflect that comment of the source user to targeted customer is more consistent； Conversely, comment stability value is lower, then the comment numerical value that source user is given to targeted customer is more unstable, that is, differ greatly.This In invention, set COM (v_su,v_tu) representing source user v_suFor targeted customer v_tuThe comment set delivered.It is assumed that in set COM(v_su,v_tu) in the meansigma methodss of all comments beSo, comment of the source user to targeted customer is stable Property index cf_L (v_su,v_tu) numerical value can be calculated with below equation：

In formula, | COM (v_su,v_tu) | represent comment set COM (v_su,v_tu) the middle total quantity commented on, com (v_su,v_tu)_iTable Show source user v_suFor targeted customer v_tuI-th comment numerical value.

For example, if source user v_suTo targeted customer v_tuThe evaluate collection for carrying out is combined into COM (v_su,v_tu)={ com (v_su,v_tu)₁ =0.9, com (v_su,v_tu)₂=1.0, com (v_su,v_tu)₃=0.8 }；Then

2) mutual reliability index between user

Between user, mutual reliability represents two users for the matching degree of degree of belief to each other, i.e., mutual trust degree between user More consistent, then mutual relation compactness is higher.In the present invention, between user, mutual reliability index is embodied by three aspects：Comment is mutual Reliability (reliabilities of comment), mutually forwarding situation (forwarding) and mutually thumb up situation (approving).In the present invention, for source user v_suWith targeted customer v_tuFor, will use between two users of direct correlation The value for stating three aspects is expressed as three-number setWith

In the same manner, can obtainIn three elements implication.

Then source user v_suWith targeted customer v_tuBetween user between mutual reliability index mf_L (v_su,v_tu) computing formula such as Under：

In formula,RepresentIn i-th element,RepresentIn I-th element.

For example, if source user v_suWith targeted customer v_tuThere is the mutual trust value in following tripartite face：

Then：

3) frequency of interaction index between user

In the present invention, frequency of interaction index is calculated using three aspects：The number of times of interaction, interaction in unit interval Average length of time and average interaction time interval.

For source user v_su, it is assumed that k-th unit interval piece u_kThe interior interaction times with its all neighbor user Maximum isIn its past all intersection record, the most long interaction time length that continues isThe mutual time interval of most short delivery isIn k-th unit interval piece u_kIt is interior, source user v_su For targeted customer v_tuInteraction times areSource user v_suIt is past for targeted customer v at which_tuIt is all In interaction, average length of time isAverage interaction time at intervals of

Then source user v_suWith targeted customer v_tuBetween user between frequency of interaction aggregative indicator if_L (v_su,v_tu) meet with Lower formula：

In formula, | Unit | represents the sum of unit interval piece in the past, l_kRepresent k-th unit interval piece u_kInterior two users Between frequency of interaction desired value.

4) common neighbours and similar community's index.

In the present invention, it is considered to which, with the growth of community's scale, user's cohesion can also be decreased.Community's scale is less, User mutual is more frequent, and cohesion is higher；Conversely, community's scale is bigger, user mutual is fewer, and cohesion is lower.Similarly, altogether The intimacy between two users be also reflects with the quantity of neighbours, that is to say, that there are more common neighbours, between user Cohesion is higher.

AssumeWithSource user v is represented respectively_suAnd targeted customer v_tuThe community at place.In each common communityIn (CC_iMiddle subscript i represents the numbering of common community), source user v_suWith targeted customer v_tuHave theirs Neighborhood, respectively CC_i(v_su) and CC_i(v_tu).In the present invention, | CC_i| represent in community CC_iIn number of members, | CC_i (v_su)∩CC_i(v_tu) | represent in set CC_i(v_su)∩CC_i(v_tu) in neighbours' number, | CC_i(v_su)∪CC_i(v_tu) | represent Set CC_i(v_su)∪CC_i(v_tu) in neighbours' number.Source user v_suWith targeted customer v_tuBetween common neighbours and similar society Area index ncf_L (v_su,v_tu) calculation expression it is as follows：

Finally, four indexs according to more than can draw calculating source user v_suWith targeted customer v_tuBetween user it is tight Degree closeness (v_su,v_tu), expression formula is as follows：

(2) user is calculated for the user subject interest-degree of entity

User subject interest-degree is estimated by user's direct action or other directly testimony of eyewitnesss.The present invention is using straight Connect interest evidence (behavior) computational entity interest-degree.User's direct interest evidence includes：Forwarding (forwarding), thumb up (approving) (following) and comment (comments) etc. are paid close attention to,.User subject interest-degree Computing Principle in the present invention It is as follows：

For an entity item, δ_iThe i-th interest evidence behavior occurred when representing user plane to entity item, it is assumed that For an entity item, user repeats interest evidence behavior n time in the past, then interest evidence behavior set expression be Φ= {δ₁,δ₂,...,δ_i,...}：Meanwhile, to user v before a variety of interest evidence behaviors classify, be designated as：D₁, D₂,...,D_j,...,D_m, total quantitys of the m for interest evidence classification, D_jJth class interest evidence classification is represented, is direct interest card According to type, and P (D_j) represent interest evidence classification D_jThe frequency occurred in all of interest evidence behavior.And P_v(D_j| Item) represent for user v, when entity item occurs, interest evidence classification D_jThe probability of appearance.Each class interest evidence class Other D_jWeight be right (D_j), right (D_j)∈[0,1].So, user v is for the user subject interest-degree of entity item The calculation expression of EI (v | item) is as follows：

In formula, P_v(item|D_j) it is for user v, when interest evidence classification D_jDuring appearance, the general of entity item is corresponded to Rate, i.e. user v in the interesting evidence of past institute, interest evidence classification D_jShow the ratio in some entity item, P (D_j)、P_v(item|D_j) directly can be obtained by numerical statistic, it is observation.

For weight right (D_j) computational methods it is as follows：Assume that interest evidence y is drawn by other interest evidence x Rise, then can be designated as：x→y.The present invention with ID represent user's history occur interesting evidence classification set, ID= {D₁,D₂,...,D_j,...,D_m, with j-th interest evidence classification D_jThe interest evidence set Link (D being associated_j) can represent For：

In formula, D_qRepresent q class interest evidence classifications, D_q,D_j∈ ID, D_qAlso it is and D_jThe interest evidence being linked (contact) Classification (can be positive evidence, or hearsay evidence), i.e. D_jIt is by D_qAnd cause and (use D_q→D_jRepresent).

So, calculate weight right (D_j) equation it is as follows：

Wherein, right (D_q) represent q class interest evidence classifications D_qWeight, | ID | represent interest evidence classification sum Amount, | ID |=m, p (D_q) represent interest evidence classification D_qThe probability occurred in user goes over interest evidence behavior, L (D_q) represent With q class interest evidence classifications D_qThe quantity of the interest evidence classification being linked.

Right (the D_j) computing formula represent weight relationship between related interests class evidence, its operation logic is power Again higher interest evidence classification its association interest evidence class weight it is also higher.The initial value of weighted value can be given at random, But according to final weighted value can be calculated after above-mentioned formula interative computation, the mistake brought directly is set so as to reduce weighted value Difference.

(3) the entity similarity between computational entity

1) entity type index of similarity

The present invention carries out type index of similarity calculating based on entity catalog classification tree.Different entities are carried out point by catalogue Class, forms the entity catalog classification tree of tree, i.e. entity catalogue is represented as one tree model；Meanwhile, present invention introduces Catalogue tree hierachy, makes c_iI-th layer of catalogue in presentation-entity catalog classification tree, the ground floor catalogue of such as entity catalog classification tree is table It is shown as c₁。

Assume there are two different entities, be expressed as item_aAnd item_b.Present invention distance (item_a, item_b) carry out presentation-entity item_aResiding TOC level and entity item_bPath distance between residing TOC level.This It is bright for type index of similarity is calculated as follows：

The 1st, if two entity place TOC levels belong to the identical number of plies, entity item_aWith entity item_bEntity class Type index of similarity sim_type (item_a,item_b) computing formula it is as follows：

In formula, c_aPresentation-entity item_aLevel in entity catalog classification tree, c_bPresentation-entity item_bIn entity catalogue Level in classification tree；

If the 2, catalogue residing for two entities is not in same level, the type similarity of two entities is：

2) entity price index of similarity

In the present invention, entity price index of similarity is calculated as：

In formula, sim_price (item_a,item_b) presentation-entity item_aWith entity item_bEntity price similarity refer to Mark, price_aPresentation-entity item_aPrice, price_bPresentation-entity item_bPrice；

3) entity quality index of similarity

In the present invention, entity quality index of similarity is used to calculate a certain entity and current candidate recommends the matter between set Amount similarity degree, recommended can be required with judging whether the entity meets.Assume the highest score of all evaluations of entity quality It is designated as score_max, current candidate recommends to concentrate the meansigma methodss of all entity quality scores to be designated as score_ave, then entity item_aWith It is current it is all can recommended entity total quality index of similarity sim_quality (item_a) computing formula it is as follows：

In formula, score_aPresentation-entity item_aEntity quality score；If meetingThen should Entity item_aIt is integrated into current candidate recommendation similar in quality index.

4) entity sales volume index of similarity

Entity item_aWith entity item_bEntity sales volume index of similarity sim_sale (item_a,item_b) calculating it is public Formula is as follows：

In formula, sale_aPresentation-entity item_aTotal sales volume, sale_bPresentation-entity item_bTotal sales volume.

Finally, by aforementioned four index can computational entity Similarity value, computing formula is as follows：

If sim is (item_a,item_bDuring) >=α (α is compactness threshold value), then by entity item_aInclude with entity item_b In the entity Candidate Recommendation set for sending.The marginal value of wherein α namely entity similarity.

As shown in Fig. 2 the concrete recommendation process of step S3 is：

1) when a certain user buys entityWhen, the initial markers user is originating subscriber O, good according to its social networks Friendly relation show that its good friend's set is designated as friend (O), and the good friend in the set meets condition " v_i∈friend(O)∩ closeness(v_i,O)≥α∩v_iEntity was not bought" when follow the steps below：

1. for each entity item present in all social networkies_k(item_kIn subscript k presentation-entity set ITEM K-th entity), if meetThen by item_kIt is put into set R_item, set R_ Item interior elements are represented with tiem '；The collection of all entity compositions for meeting above-mentioned condition is collectively referred to as entity Candidate Recommendation set R_ (originating subscriber O is bought item)。

2) for all v_i∈ friend (O), carry out following work：

1. closeness (v are calculated_i, O) value, and judge whether to meet condition closeness (v_i, O) and more than threshold alpha. If it is satisfied, by v_iIt is put in user Candidate Recommendation set R_user, set R_user interior element v_{R_user}Represent；It is unsatisfactory for When, then directly ignore v_i。

2. an entity tiem ' to be recommended is randomly choosed from R_item, calculate EI (v_{R_user}| tiem ' ∈ R_item) Value, and judge EI (v_{R_user}| tiem ' ∈ R_item) whether meet more than threshold value beta.If meeting, to user v_{R_user}Recommend real Body tiem ', and entity tiem ' is deleted from R_item；Otherwise do not do recommendation and directly delete tiem '.

3. repeat step 2) in 1. 2., the entity in R_item for sky；

4. repeat it is above-mentioned 1. 2. 3. step until all recommended mistakes per family of R_user.

3) all recommended users for crossing entity in R_user are respectively directed to all in R_user as originating subscriber Originating subscriber, by execution step 1) in 1. produced entity Candidate Recommendation set R_item continue the friend to new originating subscriber Friend recommended, and repeat step 2) in 1. 2. 3.；All users (node) in the social networkies are traversed Recommend.

Fig. 1 is the framework of proposed algorithm of the present invention.1. arrow in Fig. 1 to represent and produce one based on the set of user's Candidate Recommendation Individual social networkies；2. arrow produces new user's Candidate Recommendation set based on the social networkies for being generated；3. arrow represents Multiple user's Candidate Recommendation collection are combined into a large-scale new social networkies；4. arrow represents that proposed algorithm with social networkies is Set up on mutual basis；5. arrow to represent and produce final user candidate based on the large-scale social networkies for being generated Recommend set；6. arrow represents that the entity in the entity candidate collection for generating proposed algorithm recommends user's Candidate Recommendation collection User in conjunction；7. arrow to represent and obtain entity candidate collection based on proposed algorithm.Fig. 1 can be seen that proposed algorithm is core, Which carries out the combined recommendation of three dimensions based on social networkies to entity, and these three dimensions are respectively：Entity similarity, user are tight Density, user are for the interest-degree of entity.The thought of proposed algorithm：First, computational entity item_kWith entityIt is similar Degree, if the value of the similarity is more than threshold gamma, by entity item_kIt is put in entity candidate collection R_item, otherwise then refuses To be put into；Secondly, calculate user v_iWith the compactness of user O, if the value of the compactness be more than threshold alpha, by the user v_iIt is put into In user's candidate collection R_user；Moreover, calculate the user v in user's candidate collection R_user_{R_user}For entity candidate collection The interest-degree of entity tiem ' in R_item, if the value of the interest-degree is more than threshold value beta, by entity item_kRecommend user v_{R_user}, conversely, not recommended.

Therefore, recommendation method of the present invention can user buy an entity after, in time along social networkies, around The entity of purchase and the user of entity is bought and has initiated combined recommendation, combined recommendation has been related to three dimensions：User's compactness index The compactness of customer relationship is made to obtain more comprehensive assessment, user subject interest-degree ensures to make user interest degree calculate accurately Property, entity similarity makes entity similarity degree obtain more comprehensive calculating, realizes to social network business web sites such as Taobao, U.S. groups The assessment of impact, preferably carries out recommendation of related entities etc., the prospect with practical application.

Claims (8)

1. a kind of various dimensions combined recommendation method based on social networkies, it is characterised in that comprise the following steps：

S1：Set up social networkies model G=(V, E) and community model C=<CV,CE>, wherein, V represents the use in social networkies G Family is gathered, and E represents the set of the relation between user, and CV possesses user's set of same interest hobby in representing community C,CE represents the set of the relation between the user for possessing same interest hobby,

S2：Define user's compactness, user subject interest-degree and entity similarity, user's compactness reflect two users it Between relation tightness degree, with comment stability indicator, mutual frequency of interaction aggregative indicator and altogether between reliability index, user between user It is relevant with neighbours and similar community's index, the level of interest that user subject interest-degree reflection user holds to entity, with Family interest evidence behavior is relevant, the consistent degree of the possessed attribute of two entities of the entity similarity reflection, with entity type Index of similarity, entity price index of similarity, entity quality index of similarity are relevant with entity sales volume index of similarity；

S3：According to user's compactness, user subject interest-degree and entity similarity, various dimensions are carried out to the user in social networkies Entity integrated recommendation.

2. a kind of various dimensions combined recommendation method based on social networkies according to claim 1, it is characterised in that described Step S3 is specially：

301：When a certain user buys entityWhen, the initial markers user is originating subscriber O,

302：Entity is obtained according to entity similarityEntity Candidate Recommendation set Represent entity to be recommended Item ' and entityBetween entity similarity, γ represents similarity threshold；

303：Good friend set friend (O) of originating subscriber is obtained according to the social networkies friend relation of originating subscriber；

User's Candidate Recommendation set of originating subscriber is obtained further according to user's compactness closeness(v_{R_user}, O) represent user v to be recommended_{R_user}And the user's compactness between originating subscriber O, α is compactness threshold value；

304：Judge user v to be recommended_{R_user}Treat the user subject interest-degree EI (v of recommended entity item '_{R_user}|item′∈ R_item) whether it is more than interest-degree threshold value beta, if so, then to user v to be recommended_{R_user}Recommend entity item ' to be recommended, if it is not, Then do not do and recommend；

305：All recommended users to be recommended for crossing entity in step 304 are updated to into originating subscriber, jump procedure 303, directly All users in social networkies terminate after being traversed recommendation.

3. a kind of various dimensions combined recommendation method based on social networkies according to claim 1, it is characterised in that described User's compactness meets below equation：

$closeness(v_{su},v_{tu})=\frac{1}{4}\&lsqb;cf\_L(v_{su},v_{tu})+mf\_L(v_{su},v_{tu})+if\_L(v_{su},v_{tu})+ncf\_L(v_{su},v_{tu})\&rsqb;$

In formula, closeness (v_su,v_tu)、cf_L(v_su,v_tu)、mf_L(v_su,v_tu)、if_L(v_su,v_tu) and ncf_L (v_su, v_tu) source user v is represented respectively_suWith targeted customer v_tuBetween user's compactness, comment stability indicator, mutual reliability between user Frequency of interaction aggregative indicator and common neighbours and similar community's index between index, user.

4. a kind of various dimensions combined recommendation method based on social networkies according to claim 1, it is characterised in that described Comment stability indicator meets below equation：

In formula, cf_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween comment stability indicator, COM (v_su,v_tu) Represent source user v_suFor targeted customer v_tuThe comment set delivered, | COM (v_su,v_tu) | represent comment set COM (v_su, v_tu) the middle total quantity commented on, com (v_su,v_tu)_iRepresent source user v_suFor targeted customer v_tuI-th comment numerical value,Represent in comment set COM (v_su,v_tu) in all comments meansigma methodss；

Between the user, mutual reliability index meets below equation：

$mf\_L(v_{su},v_{tu})=\frac{\underset{i=1}{\overset{3}{\&Sigma;}}(V\_L_{v_{su},v_{tu}}\&lsqb;i\&rsqb;\&times;V\_L_{v_{tu},v_{su}}\&lsqb;i\&rsqb;)}{\sqrt{\underset{i=1}{\overset{3}{\&Sigma;}}{(V\_L_{v_{su},v_{tu}}\&lsqb;i\&rsqb;)}^2}\&times;\sqrt{\underset{i=1}{\overset{3}{\&Sigma;}}{(V\_L_{v_{tu},v_{su}}\&lsqb;i\&rsqb;)}^2}}$

In formula, mf_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween user between mutual reliability index, For three-number set,Represent source user v_suTo targeted customer v_tuThe meansigma methodss of all evaluations of estimate, Represent source user v_suIn targeted customer v_tuThe information scales forwarded in the information of all issues,Expression source User v_suIn targeted customer v_tuThe information scales of thumb up are carried out in the information of all issues,For three-number set,Represent targeted customer v_tuTo source user v_suThe meansigma methodss of all evaluations of estimate,Represent that target is used Family v_tuIn source user v_suThe information scales forwarded in the information of all issues,Represent targeted customer v_tu Source user v_suThe information scales of thumb up are carried out in the information of all issues；

Between the user, frequency of interaction aggregative indicator meets below equation：

$if\_L(v_{su},v_{tu})=\frac{\underset{k=1}{\overset{\left|Unit\right|}{\&Sigma;}}\left(l_k\right)}{\left|Unit\right|}$

$l_k=\frac{1}{3}\&lsqb;\frac{num{(v_{su},v_{tu})}_{u_k}}{max{\&lsqb;num\left(v_{su}\right)\&rsqb;}_{u_k}}+\frac{len{(v_{su},v_{tu})}_{u_k}}{max{\&lsqb;len\left(v_{su}\right)\&rsqb;}_{u_k}}+\frac{min{\&lsqb;\mathrm{int}(v_{su},v_{tu})\&rsqb;}_{u_k}}{\mathrm{int}{(v_{su},v_{tu})}_{u_k}}\&rsqb;$

In formula, if_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween user between frequency of interaction comprehensively refer to Mark, | Unit | represent the sum of unit interval piece, l_kRepresent k-th unit interval piece u_kInteraction frequency between interior two users Rate desired value,K-th unit interval is represented respectively Piece u_kInterior source user v_suFor targeted customer v_tuInteraction times, average interaction time length, average interaction time interval,Show k-th unit interval piece u respectively_kEndogenous use Family v_suMaximum interaction times, most long interaction time length for itself all neighbor user, the mutual time interval of most short delivery；

The common neighbours and similar community's index meet below equation：

$ncf\_L(v_{su},v_{tu})=\frac{\underset{{\mathrm{CC}}_i\&Element;C_{v_{su}}\&cap;C_{v_{tu}}}{\&Sigma;}\&lsqb;(1/{\log }_2|{\mathrm{CC}}_i\left|\right)\&times;\frac{|{\mathrm{CC}}_i\left(v_{su}\right)\&cap;{\mathrm{CC}}_i\left(v_{tu}\right)|}{|{\mathrm{CC}}_i\left(v_{su}\right)\&cup;{\mathrm{CC}}_i\left(v_{tu}\right)|}\&rsqb;}{\underset{{\mathrm{CC}}_i\&Element;C_{v_{su}}\&cap;C_{v_{tu}}}{\&Sigma;}(1/{\log }_2|{\mathrm{CC}}_i\left|\right)}$

In formula, ncf_L (v_su,v_tu) represent source user v_suWith targeted customer v_tuBetween common neighbours and similar community's index,Represent source user v_suThe community at place,Represent targeted customer v_tuThe community at place, i-th common communityIn, source user v_suWith targeted customer v_tuRespective neighbor user set is respectively CC_i(v_su) and CC_i (v_tu), | CC_i| represent CC_iInterior number of users, | CC_i(v_su)∩CC_i(v_tu) | represent CC_i(v_su) and CC_i(v_tu) in common factor Number of users, | CC_i(v_su)∪CC_i(v_tu) | represent CC_i(v_su) and CC_i(v_tu) number of users in union.

5. a kind of various dimensions combined recommendation method based on social networkies according to claim 1, it is characterised in that described User subject interest-degree meets below equation：

In formula, and EI (v | item) user subject interest-degrees of the user v to entity item is represented, n represents user plane to entity item When the interest evidence behavior total degree that occurs, δ_iThe i-th interest evidence behavior occurred when representing user plane to entity item, it is right The interesting evidence behavior of institute that user's history occurs is classified, D_jJth class interest evidence classification is represented, is direct interest card According to right (D_j) represent jth class interest evidence classification D_jWeight, P_v(D_j| item) represent for user v, as entity item During appearance, interest evidence classification D_jThe probability of appearance.

6. a kind of various dimensions combined recommendation method based on social networkies according to claim 5, it is characterised in that different There is following relation between the weight of interest evidence classification：

$right\left(D_j\right)=\frac{1-p\left(D_j\right)}{\left|ID\right|}+p\left(D_j\right)\&times;\underset{D_q\&Element;Link\left(D_j\right)}{\&Sigma;}\frac{right\left(D_q\right)}{L\left(D_q\right)}$

In formula, right (D_q) represent q class interest evidence classifications D_qWeight, ID represent user's history occur it is be interested in demonstrate,prove According to the set of classification, D_q,D_j∈ ID, | ID | represent the total quantity of interest evidence classification, p (D_q) represent interest evidence classification D_q The probability occurred in user's past interest evidence behavior, L (D_q) represent and q class interest evidence classifications D_qThe interest card being linked According to the quantity of classification, Link (D_j) represent and jth class interest evidence classification D_jThe interest evidence category set being linked, D_q→D_j Represent D_jIt is by D_qAnd cause, i.e. D_qWith D_jIt is linked.

7. a kind of various dimensions combined recommendation method based on social networkies according to claim 1, it is characterised in that described Entity similarity meets below equation：

$sim({\mathrm{item}}_a,{\mathrm{item}}_b)=\frac{1}{4}\&times;\left[\begin{array}{c}sim\_type({\mathrm{item}}_a,{\mathrm{item}}_b)+sim\_price({\mathrm{item}}_a,{\mathrm{item}}_b)\\ +sim\_quality\left({\mathrm{item}}_a\right)+sim\_sale({\mathrm{item}}_a,{\mathrm{item}}_b)\end{array}\right]$

In formula, item is used_aAnd item_bTo represent two different entities, sim (item_a,item_b)、sim_type(item_a, item_b)、sim_price(item_a,item_b)、sim_sale(item_a,item_b) difference presentation-entity item_aWith entity item_b Entity similarity, entity type index of similarity, entity price index of similarity, entity sales volume index of similarity, sim_ quality(item_a) presentation-entity item_aEntity quality index of similarity.

8. a kind of various dimensions combined recommendation method based on social networkies according to claim 1, it is characterised in that described Entity type index of similarity is calculated based on entity catalog classification tree, meets below equation：

$sim\_type({\mathrm{item}}_a,{\mathrm{item}}_b)=\left\{\begin{array}{cc}\sqrt[c_a]{\frac{1}{1+dis\tan ce({\mathrm{item}}_a,{\mathrm{item}}_b)}}& a_c=c_b\\ \sqrt[\frac{c_a+c_b}{2}]{\frac{1}{1+dis\tan ce({\mathrm{item}}_a,{\mathrm{item}}_b)}}& a_c\&NotEqual;c_b\end{array}\right.$

In formula, sim_type (item_a,item_b) presentation-entity item_aWith entity item_bEntity type index of similarity, c_aTable Show entity item_aLevel in entity catalog classification tree, c_bPresentation-entity item_bLevel in entity catalog classification tree, distance(item_a,item_b) represent the entity item in entity catalog classification tree_aWith entity item_bBetween path；

The entity price index of similarity meets below equation：

$sim\_price({\mathrm{item}}_a,{\mathrm{item}}_b)=1-\frac{\left|{\mathrm{price}}_a-{\mathrm{price}}_b\right|}{{\mathrm{price}}_a}$

In formula, sim_price (item_a,item_b) presentation-entity item_aWith entity item_bEntity price index of similarity, price_aPresentation-entity item_aPrice, price_bPresentation-entity item_bPrice；

The entity quality index of similarity meets below equation：

$sim\_quality\left({\mathrm{item}}_a\right)=\frac{{\mathrm{score}}_a}{{\mathrm{score}}_{max}}$

In formula, sim_quality (item_a) presentation-entity item_aEntity quality index of similarity, score_aPresentation-entity item_aEntity quality score, score_maxRepresent the highest score in all entity quality scores；

The entity sales volume index of similarity meets below equation：

$sim\_sale({\mathrm{item}}_a,{\mathrm{item}}_b)=\frac{\left|{\mathrm{sale}}_b-{\mathrm{sale}}_a\right|}{{\mathrm{sale}}_a}$

In formula, sim_sale (item_a,item_b) presentation-entity item_aWith entity item_bEntity sales volume index of similarity, sale_aPresentation-entity item_aTotal sales volume, sale_bPresentation-entity item_bTotal sales volume.