CN106354886A - Method for screening nearest neighbor by using potential neighbor relation graph in recommendation system - Google Patents

Abstract

The invention discloses a method for screening a nearest neighbor by using a potential neighbor relation graph in a recommendation system. The method comprises the following steps: (1) generating an object cluster set C with a redundant feature; (2) constructing the potential neighbor relation graph corresponding to the cluster set C; (3) quantifying weight of each edge in the potential neighbor relation graph, wherein the weight of the edge represents the possibility that two adjacent objects of the edge become the nearest neighbors; (4) clipping a potential neighbor relation to weed out redundant comparisons; (5) screening the nearest neighbor of a target according to the clipped potential neighbor relation graph. According to the screening method, on the premise of guaranteeing the recommendation precision, a recommendation based on a complete large-scale data set is mapped to a recommendation of a data set with a relatively small scale, so that the scale of the recommendation system is reduced, and high efficiency of the recommendation method is guaranteed.

Description

The method that potential neighborhood figure screens nearest-neighbors is utilized in commending system

Technical field

The present invention relates to recommended technology field, potential neighborhood figure is utilized to screen arest neighbors particularly in commending system The method occupying.

Background technology

Currently ever-increasing data volume makes user need to expend a great deal of time the valuable information that just can find.Collaborative Filtration is considered as one of the effective technology solving information overload problem, is widely used to film, music, books, travelling, new News etc. recommends field.The main thinking of collaborative filtering method is to predict targeted customer to item the hobby of project according to similar users Purpose viewpoint, or according to user, the suggestion of similar terms is implemented to recommend for destination item.Therefore, collaborative filtering needs to solve A key issue be: the how effectively nearest-neighbors of selection target user or project.But commending system has millions of Meter user and project, the still sustainable growth of its scale, search for nearest-neighbors in mass data with traditional method and be difficult to ensure that in conjunction Accurate recommendation was provided in the reason time.

The core of collaborative filtering method is to need all items in commending system (or user) are compared two-by-two, passes through Calculate similarity each other, select nearest-neighbors.The number of times comparing two-by-two is more, and the efficiency of operation is lower, but its search Higher to the probability of nearest-neighbors；Otherwise the number of times comparing two-by-two is fewer, its operational efficiency is higher, but misses nearest-neighbors Probability is also higher, thus having influence on recommendation precision.In commending system, data internal association is in close relations and complicated in addition, valency Value Density Distribution is extremely unbalanced, and the calculating to mass data in commending system can not depend on to complete as Small Sample Database collection The statistical analysiss of office data and iterative calculation, need heuristic data reduction method on demand.

Content of the invention

The technical problem to be solved is to overcome the deficiencies in the prior art and provide in commending system using latent The method screening nearest-neighbors in neighborhood figure, the inventive method, to reject redundancy and unnecessary comparison, had both ensured recommendation effect Rate, and do not sacrifice recommendation precision it is ensured that providing accurate recommendation within reasonable time.

The present invention is to solve above-mentioned technical problem to employ the following technical solutions:

According to the method utilizing potential neighborhood figure to screen nearest-neighbors in commending system proposed by the present invention, including Following steps:

Step 1, set i ∈ o, o is the object set needing to screen nearest-neighbors, and i is object, using Fuzzy clustering techniques foundation The characteristic vector of object, object i is assigned in multiple clusters by probability set in advance, thus produces the cluster containing k object cluster Set c；

Step 2, structure gathering close the corresponding potential neighborhood figure g of c_c={ v_c,e_c, wherein, v_cIt is vertex set, e_cIt is Undirected line set；Specific as follows:

If object i and object j simultaneously appears in the same cluster c that gathering closes c, object i and j be called co-occurrence to and remember For ＜ i, j ＞；For each pair co-occurrence in gathering conjunction c to ＜ i, j ＞, first by corresponding for object i and j vertex v_iAnd v_jIt is added to figure g_cIn, if not there is nonoriented edge between two object i and j, use side e_i,jConnect vertex v_iAnd v_j；Wherein, scheme g_cIn every Side e_i,jRepresent a potential neighborhood, side e_i,jTwo adjacent vertex v_iAnd v_jCorresponding object i and j is referred to as contiguous object, J ∈ o, c ∈ c, v_i∈v_c, v_j∈v_c, e_i,j∈e_c；

Step 3, quantization figure g_cThe weight of middle each edge；

Step 4, to figure g_cCarry out cutting, delete potential neighborhood figure g_cThe weight on middle side is less than w_minSide, remaining side Constitute a new figure g_c'；Wherein, w_minMinimal weight threshold value for setting；

Step 5, choose object i as target, using the potential neighborhood figure g after cutting_c' screening target arest neighbors Occupy, for g_c' in figure target i every adjacent side e_i,j, relatively and calculate utility vector rⁱWith r^jBetween similarity, Ran Houyi Screen its nearest-neighbors according to neighbour's alternative condition in all of its neighbor object of target i；Wherein, rⁱRepresent object i effectiveness to Amount, r^jRepresent the utility vector of object j.

Enter one as of the present invention using the method that potential neighborhood figure screens nearest-neighbors in commending system Step prioritization scheme, adopts following formula to calculate side e in described step 3_i,jWeight e_i,j.weight:

$e_{i,j}.weight=\frac{\left|c_{i,j}\right|}{\left|c_i\right|+\left|c_j\right|-\left|c_{i,j}\right|}\·log\frac{\left|e_c\right|}{d\left(v_i\right)}\·log\frac{\left|e_c\right|}{d\left(v_j\right)}$

Wherein, c_iRepresent that the gathering that object i is subordinate to is closed, c_jRepresent that the gathering that object j is subordinate to is closed, c_i,jShare for object i and j Set,| * | is membership in set *, and d (*) represents the degree of summit *.

Enter one as of the present invention using the method that potential neighborhood figure screens nearest-neighbors in commending system Step prioritization scheme, in described step 5, neighbour's alternative condition refers to choose the front k object composition mesh maximum with the similarity of target Target neighbour collects.

Enter one as of the present invention using the method that potential neighborhood figure screens nearest-neighbors in commending system Step prioritization scheme, k >=1.

Enter one as of the present invention using the method that potential neighborhood figure screens nearest-neighbors in commending system Step prioritization scheme, side e_i,jWeight and graph of a relation g_cMiddle side e_i,jThe cluster that object i with j being adjoined shares is relevant.

The present invention adopts above technical scheme compared with prior art, has following technical effect that

Under the premise of ensureing recommendation precision, less by a scale is mapped to based on the recommendation of complete large-scale dataset The recommendation of data set, and significantly reduce commending system scale it is ensured that recommending the high efficiency of method.Concrete manifestation is in the following areas:

(1) using Fuzzy clustering techniques, object is assigned in multiple clusters by certain probability, this redundancy properties provide A kind of indispensable, reliable, method of avoiding missing nearest-neighbors；

(2) in cluster process, by object assignment high for similarity in same cluster, exclude the object that peels off individually.This It is not only advantageous for improving and recommends precision, decrease the unnecessary comparison that the object that peels off brings simultaneously, improve and recommend efficiency；

(3) in the building process of potential neighborhood figure, each pair co-occurrence is to a line at most corresponding diagram, each edge table Show a kind of potential neighborhood, follow-up step is only compared to the contiguous object of each edge, calculates the similar of them Degree；Therefore, eliminate redundancy and unnecessary co-occurrence pair, decrease the number of times comparing two-by-two in object set, improve the effect of recommendation Rate；

(4) pass through to arrange the weight threshold on appropriate side, reject the side that potential neighbours' weighted connections in figure is less than weight, enter One step eliminates unnecessary comparison.

Brief description

Fig. 1 is the handling process of the present invention.

Specific embodiment

Below in conjunction with the accompanying drawings technical scheme is described in further detail:

If Fig. 1 is the handling process of the present invention, following steps: the object gathering that step (1), generation have redundancy properties is closed c；Step (2), structure gathering close the corresponding potential neighborhood figure of c；Step (3), the potential neighborhood in figure each edge of quantization Weight, while weight represent while two objects being adjoined become the probability of nearest-neighbors；Step (4), to potential neighbours close System carries out cutting, rejects unnecessary comparison；Step (5), screen the arest neighbors of target using the potential neighborhood figure after cutting Occupy.

[embodiment 1]

It is known that user collection u, Item Sets i in a film commending system, each project is a film；f_iRepresent The characteristic vector of project i, for describing film school, prize-winning type, performer and director etc.；rⁱThe effectiveness of expression project i ∈ i to Amount, r_u,i∈rⁱIt is the scoring to project i for the user u ∈ u.Commending system utilizes potential neighborhood figure screening destination item i Neighbour occupies, and then utilizes the scoring to the score in predicting user u of all nearest-neighbors of project i to project i for the targeted customer u.At this In embodiment 1, object set is project set, and the detailed process of the nearest-neighbors of screening programme is as follows:

(1) produce gathering to close.Project i ∈ i is assigned to by certain probability according to item characteristic using Fuzzy clustering techniques In multiple clusters, thus produce project gathering and close c.

(2) build gathering and close the corresponding potential neighborhood figure g of c_c={ v_c,e_c, v_cIt is vertex set, e_cIt is nonoriented edge collection Close.Figure g_cConcrete construction method is as follows: for each pair co-occurrence in c to ＜ i, j ＞, first by project i ∈ i and the corresponding summit of j ∈ i v_iAnd v_jIt is added to figure g_cIn, then use side e_i,j∈e_cConnect vertex v_iAnd v_j.During creating graph of a relation, if finding two There is a nonoriented edge between individual project, then need not increase a line more between them.In figure each edge represents that one dives Neighborhood, Similarity Measure need to be passed through in subsequent step, it is determined whether for nearest-neighbors.

(3) quantify figure g_cEach edge weight.Graph of a relation g_cAfter finishing, for screening neighbor relationships, requirement figure further Middle each edge.Weighing computation method is as follows: knownRepresent that the gathering that project i, j is subordinate to is closed respectively, c_i,j=c_i ∩c_jIt is referred to as the set that project i and j share.The cluster shared as project i and j more (i.e. | c_i,j| value is bigger), they become each other The probability of nearest-neighbors is bigger.Except considering | c_i,j|, side right escheat need to consider the sum of the cluster that its adjacent project is subordinate to, And the degree of the adjacent vertex on side.When the cluster that the adjacent project on certain side is subordinate to is fewer, the weight on this side should be higher；When certain side Summit degree less, the weight on this side should be higher.Calculate side e with following formula_i,j∈e_cWeight

$e_{i,j}.weight=\frac{\left|c_{i,j}\right|}{\left|c_i\right|+\left|c_j\right|-\left|c_{i,j}\right|}\·log\frac{\left|e_c\right|}{d\left(v_i\right)}\·log\frac{\left|e_c\right|}{d\left(v_j\right)}$

Wherein, d (*) represents the degree of summit *.

(4) the unnecessary comparison of cutting.Minimal weight threshold value w according to setting_minDelete figure g_cThe weight on middle side is less than w_min's Side, remaining side constitutes the figure g after cutting_c'.

(5) project i of choosing, as target, screens the nearest-neighbors of target i.For g_c'Every adjacent side of in figure target i e_i,j, the utility vector r of comparison object i and j project i in user-project utility matrixⁱUtility vector r with project j^jBetween Similarity, calculate the similarity between them；Then screen in all of its neighbor project of target i according to neighbour's alternative condition Its nearest-neighbors, in this example 1, chooses neighbour's collection that the front k project maximum with the similarity of destination item constitutes target. Finally utilize targeted customer u that project neighbour is concentrated with the scoring to destination item for the score in predicting targeted customer u of each neighbour.

[embodiment 2]

It is known that user collection u, Item Sets i in a film commending system, each project is a film；f_qRepresent The characteristic vector of user q, for describing the sex of user, age, occupation etc.；r_qRepresent the utility vector of user q ∈ u, r_q,m∈ r_uIt is the scoring to project m ∈ i for the user q.Commending system utilizes potential neighborhood figure to screen the nearest-neighbors of targeted customer q, Then utilize all nearest-neighbors scoring to project m to the score in predicting user q of project m of targeted customer q.In the present embodiment In 2, object set is gathered for user, and the detailed process of the nearest-neighbors of screening user is as follows:

(1) produce gathering to close.User u ∈ u is assigned to by certain probability according to user characteristicses using Fuzzy clustering techniques In multiple clusters, thus produce user's gathering and close c.

(2) build gathering and close the corresponding potential neighborhood figure g of c_c={ v_c,e_c, v_cIt is vertex set, e_cIt is nonoriented edge collection Close.Figure g_cConcrete construction method is as follows: for each pair co-occurrence in c to ＜ q, h ＞, first by user q ∈ u and the corresponding summit of h ∈ u v_qWith_vhIt is added to figure g_cIn, then use side e_q,h∈e_cConnect vertex v_qWith_vh.During creating graph of a relation, if finding two There is a nonoriented edge between user, then need not increase a line more between them.In figure each edge represents that one is potential Neighborhood, Similarity Measure need to be passed through in subsequent step, it is determined whether for nearest-neighbors.

(3) quantify figure g_cEach edge weight.Graph of a relation g_cAfter finishing, for screening neighbor relationships, requirement figure further Middle each edge.Weighing computation method is as follows: knownRepresent that the gathering that user q, h are subordinate to is closed respectively, c_q,h=c_q ∩c_hThe set that referred to as user q and h shares.When user q and h share cluster more (i.e. | c_q,h| value is bigger), they become each other The probability of nearest-neighbors is bigger.Except considering | c_q,h|, side right escheat need to consider the sum of the cluster that its adjacent user is subordinate to, And the degree of the adjacent vertex on side.When the cluster that the adjacent user on certain side is subordinate to is fewer, the weight on this side should be higher；When certain side Summit degree less, the weight on this side should be higher.Therefore, available following formula calculates side e_q,hWeight:

$weight(q,h)=\frac{\left|c_{q,h}\right|}{\left|c_q\right|+\left|c_h\right|-\left|c_{q,h}\right|}\·\log \frac{\left|e_c\right|}{d\left(v_q\right)}\·\log \frac{\left|e_c\right|}{d\left(v_h\right)},$

Wherein, d (*) represents the degree of summit *.

(4) the unnecessary comparison of cutting.Minimal weight threshold value w according to setting_minDelete figure g_cThe weight on middle side is less than w_min's Side, remaining side constitutes the figure g after cutting_c'.

(5) choose user q as target, screen the nearest-neighbors of targeted customer q.For g_c' in figure target q every neighbour Edge fit e_q,h, comparison object q and h in user-project utility matrix user q utility vector r_qUtility vector with user h r_hBetween similarity, calculate the similarity between them；Then foundation neighbour's alternative condition is in all of its neighbor user of target q Middle its nearest-neighbors of screening, in this example 2, choose the front k user maximum with the similarity of targeted customer q and constitute target Neighbour collects.The neighbour finally utilizing targeted customer q concentrates each neighbour to the score in predicting targeted customer q of project to destination item Scoring.

Specific embodiments described above, has been carried out to the purpose of the present invention, technical scheme and beneficial effect further Detailed description, be should be understood that and be the foregoing is only specific embodiments of the present invention, be not limited to this Bright scope, any those skilled in the art, that is made on the premise of the design without departing from the present invention and principle is equivalent Change and modification, all should belong to the scope of protection of the invention.

Claims (5)

1. utilize potential neighborhood figure to screen the method for nearest-neighbors it is characterised in that walking below including in commending system Rapid:

Step 1, set i ∈ o, o is the object set needing to screen nearest-neighbors, and i is object, using Fuzzy clustering techniques according to object Characteristic vector, object i is assigned in multiple clusters by probability set in advance, thus produce containing k object cluster gathering conjunction c；

Step 2, structure gathering close the corresponding potential neighborhood figure g of c_c={ v_c,e_c, wherein, v_cIt is vertex set, e_cIt is undirected Line set；Specific as follows:

If object i and object j simultaneously appears in the same cluster c that gathering closes c, object i and j be called co-occurrence to and be designated as ＜ I, j ＞；For each pair co-occurrence in gathering conjunction c to ＜ i, j ＞, first by corresponding for object i and j vertex v_iAnd v_jIt is added to figure g_cIn, If not there is nonoriented edge between two object i and j, use side e_i,jConnect vertex v_iAnd v_j；Wherein, scheme g_cMiddle each edge e_i,j Represent a potential neighborhood, side e_i,jTwo adjacent vertex v_iAnd v_jCorresponding object i and j is referred to as contiguous object, j ∈ o, C ∈ c, v_i∈v_c, v_j∈v_c, e_i,j∈e_c；

Step 3, quantization figure g_cThe weight of middle each edge；

Step 4, to figure g_cCarry out cutting, delete potential neighborhood figure g_cThe weight on middle side is less than w_minSide, remaining side is constituted One new figure g_c'；Wherein, w_minMinimal weight threshold value for setting；

Step 5, choose object i as target, using the potential neighborhood figure g after cutting_c'The nearest-neighbors of screening target, pin To g_c'Every adjacent side e of in figure target i_i,j, relatively and calculate utility vector rⁱWith r^jBetween similarity, then according near Adjacent alternative condition screens its nearest-neighbors in all of its neighbor object of target i；Wherein, rⁱRepresent the utility vector of object i, r^j Represent the utility vector of object j.

2. the method utilizing potential neighborhood figure to screen nearest-neighbors in commending system according to claim 1, its It is characterised by, in described step 3, adopt following formula to calculate side e_i,jWeight e_i,j.weight:

$e_{i,j}.weight=\frac{\left|c_{i,j}\right|}{\left|c_i\right|+\left|c_j\right|-\left|c_{i,j}\right|}\·\log \frac{\left|e_c\right|}{d\left(v_i\right)}\·\log \frac{\left|e_c\right|}{d\left(v_j\right)}$

Wherein, c_iRepresent that the gathering that object i is subordinate to is closed, c_jRepresent that the gathering that object j is subordinate to is closed, c_i,jThe collection shared for object i and j Close,c_i,j=c_i∩c_j, | * | is membership in set *, and d (*) represents the degree of summit *.

3. the method utilizing potential neighborhood figure to screen nearest-neighbors in commending system according to claim 1, its It is characterised by, in described step 5, neighbour's alternative condition refers to choose the front k object composition target maximum with the similarity of target Neighbour collection.

4. the method utilizing potential neighborhood figure to screen nearest-neighbors in commending system according to claim 1, its It is characterised by, k >=1.

5. the method utilizing potential neighborhood figure to screen nearest-neighbors in commending system according to claim 1, its It is characterised by, side e_i,jWeight and graph of a relation g_cMiddle side e_i,jThe cluster that object i with j being adjoined shares is relevant.