CN104463615A - Personalized accurate putting method of internet advertisements - Google Patents

Abstract

The invention provides a personalized accurate putting method of internet advertisements. The method comprises the steps that according to a signal of clicking on an advertisement by a user, a feature vector of the user and a feature vector of the advertisement are updated; the feature vector of the user and a feature vector of a web page are updated according to the signal of clicking on the web page by the user. Feature vectors of multiple users, multiple advertisements and multiple documents are updated by using the method multiple times. When one user clicks on one web page, a personalized sorted value of each advertisement in a set of advertisements is calculated according to the feature vector of the user and the feature vector of the web page, and at least one advertisement in the set of advertisements is presented to the user according to the personalized sorted value. The personalized accurate putting method of the internet advertisements can well solve the problem of data sparsity of a recommendation system, achieve personalized accurate putting of the internet advertisements and then improve the click rate of the internet advertisements.

Description

A kind of accurate put-on method of personalization of Internet advertising

Technical field

The present invention relates to internet arena, relate in particular to a kind of accurate put-on method of personalization of Internet advertising.

Background technology

Internet advertising is the main realization mode of network company, and paying per click is the main charge mode of Internet advertising.Improve ad click rate, network company, advertiser and user tripartite can be made to benefit.

Common Internet advertising, comprises pictorial advertisement, mobile flashing type advertisement, pop-up ad and text link advertisement etc.For improving ad click rate and represent rate, advertisement is thrown according to the degree of correlation of the rank of webpage, the bid of advertisement and advertisement and webpage usually by network company.The major defect of existing advertisement putting technology is, cannot carry out advertisement accurately input according to the individualized feature of user.The advertisement that different user is seen under same web page is in the prior art identical.Because user interest hobby varies, if advertisement can be thrown according to the personalized difference of user, ad click rate will be significantly improved and represent rate.

Although existing commending system also can throw in personalized advertisement according to the feature of user, the shortcoming of this technology is the Sparse Problem that cannot overcome under large market demand environment.For collaborative filtering system, it is by analyzing user interest, the user with designated user with similar interests is found in customer group, then these similar users comprehensive are to the click navigation patterns of a certain advertisement, form this designated user to predict the degree of concern of this advertisement, and accordingly to this designated user recommended advertisements.Such as, but find that the advertisement exposure of more than 99% does not have user to click navigation patterns and occurs in actual applications, the advertisement average click-through rate on Facebook is less than 1 ‰.Therefore, when commending system scale is increasing, the number of user and recommended advertisements can significantly increase, and this selects the probability of identical recommended advertisements very low by causing between two users.If account for the ratio of the click relation likely existed to weigh the openness of system with click relation existing between user and recommended advertisements, the degree of rarefication of so existing commending system is very low, and what have even reaches millionth magnitude.Because data are very sparse, make the overwhelming majority based on the commending system of association analysis, as collaborative filtering system etc., desirable effect cannot be reached.

Summary of the invention

In view of above-mentioned prior art Problems existing, the object of the present invention is to provide the accurate put-on method of a kind of personalization of Internet advertising, improve the clicking rate of Internet advertising and represent rate.

According to above-described object, the present invention proposes a kind of accurate put-on method of personalization of Internet advertising, it is characterized in that, described method is included in the server of accessing Internet and performs following steps:

Step S1. obtains and stores user and collects U={1,2 ..., M}, webpage collection D={1,2 ..., N}, set of advertisements A={1,2 ..., P} and feature set K={1,2 ..., L};

Step S2. is that at least one advertisement in described set of advertisements A arranges proper vector initial value;

Step S3. receives the signal that arbitrary user m (m ∈ U) clicks arbitrary advertisement s (s ∈ A), and according to this signal update the proper vector of user m and the proper vector of described advertisement s;

Step S4. receives the signal that arbitrary user m (m ∈ U) clicks arbitrary webpage n (n ∈ D), and according to this signal update the proper vector of user m and the proper vector of described webpage n;

Step S5., according to the described signal in described step S4, calculates the personalized ordering value of each advertisement in described set of advertisements A, and gives described user m according to described personalized ordering value by least one advertisement putting, returns described step S3.

Compared with prior art, this method can click according to user the individualized feature that ad log and the daily record of user's webpage clicking obtain user, and according to the individualized feature of user and the signal of user's webpage clicking, realizes the accurate input of Internet advertising.The method can solve the Sparse sex chromosome mosaicism of commending system, realizes the accurate input of personalized advertisement, and then improves the clicking rate of Internet advertising and represent rate.

Accompanying drawing explanation

Fig. 1 is the proper vector method for expressing collecting each user in U user;

Fig. 2 is the parameter vector method for expressing of each webpage in webpage collection D;

Fig. 3 is the proper vector method for expressing of each advertisement in set of advertisements A;

Fig. 4 is a kind of accurate put-on method process flow diagram of personalization of Internet advertising;

Fig. 5 is the concrete methods of realizing of step S3 in method described in Fig. 4;

Fig. 6 is the concrete methods of realizing of step S4 in method described in Fig. 4;

Fig. 7 is the concrete methods of realizing of step S5 in method described in Fig. 4.

Embodiment

By reference to the accompanying drawings the inventive method is described in further detail.

First illustrate that user collects the acquisition methods of U, webpage collection D, set of advertisements A and feature set K.The method, in the server of accessing Internet, obtains and stores the webpage collection D that the user be made up of multiple user ID collects U and be made up of multiple banner.Described user ID is the unique identifier of user, as user account number, phone number, Cookie identification code, IP address, Email address or instant communication number.Described banner is the URL address of webpage.The example obtaining multiple user ID be on the internet receive multiple user log-on message and extract user ID wherein, the example obtaining multiple banner is on the internet the network address by sending spider to obtain multiple webpage at the enterprising line search of network.Each advertisement in set of advertisements A describes two parts by advertisement unique identification and advertisement and forms.If described user collects U and contains M element, described webpage collection D contains N number of element, containing P element in described set of advertisements A.

In the server of described accessing Internet, store the feature set K be made up of multiple signature identification.Described multiple feature is the feature that described user collects user in U, and being again the feature of webpage in described webpage collection D, is also the feature of each advertisement in set of advertisements A.User, webpage and advertisement use identical feature set K.If feature that user has " music ", consumer taste music is described, and webpage (or advertisement) has " music " feature, illustrate that webpage (or advertisement) is relevant to musical theme.If described feature set K contains L element.

Introduce the method for expressing of proper vector of user, webpage and advertisement below.Described proper vector method for expressing is similar to the vectorial expression method of vector space model, namely using characteristic item as the base unit of user characteristics or web page characteristics.In this patent, using the set of the degree of correlation of user and each feature as the proper vector of user, using the set of the degree of correlation of webpage and each feature as the proper vector of webpage, using the set of the degree of correlation of advertisement and each feature as the proper vector of advertisement.

Fig. 1 is the proper vector method for expressing collecting each user in U user.The proper vector collecting any one user m in U (m ∈ U) user is set to (uw _m1, uw _m2..., uw _mk..., uw _mL), wherein said uw _mkrepresent the degree of correlation of described user m and feature k (k ∈ K).In addition, the degree of correlation described user being collected each user in U and feature k pools together, and forms a vector, is called kth user's column vector (uw that user collects U _1k, uw _2k..., uw _mk).

Fig. 2 is the proper vector method for expressing of each webpage in webpage collection D.In webpage collection D, the proper vector of any one webpage n (n ∈ D) is set to (aw _n1, dw _n2..., dw _nk..., dw _nL), wherein said dw _nkrepresent the degree of correlation of described webpage n and feature k (k ∈ K).In addition, the degree of correlation of each webpage in described webpage collection D and feature k is pooled together, forms a vector, be called a kth webpage column vector (dw of webpage collection D _1k, dw _2k..., dw _nk).

Fig. 3 is the proper vector method for expressing of each advertisement in set of advertisements A.In set of advertisements A, the proper vector of any one advertisement s (s ∈ A) is set to (aw _s1, aw _s2..., aw _sk..., aw _sL), wherein said aw _skrepresent the degree of correlation of described advertisement s and feature k (k ∈ K).In addition, the degree of correlation of each advertisement in described set of advertisements A and feature k is pooled together, form a vector, be called a kth advertisement column vector (aw of set of advertisements A _1k, aw _2k..., aw _pk).

The described degree of correlation is a nonnegative real number value, and it represents the close relation degree of certain feature in user, webpage or advertisement and feature set K.If user, webpage or an advertisement associate with musical features and morely associates with sports feature a little less, we just say that the degree of correlation of this user, webpage or advertisement and musical features is high, low with the degree of correlation of sports feature.In addition when Feature Selection, between some feature, there is correlativity, therefore can be reduced the dimension of feature set K by the correlativity reduced between feature, reduce the load of server stores and calculating, improve efficiency of algorithm.Some feature need not directly be listed in feature set K, because the degree of correlation of these features can by the relatedness computation of one or several further feature in feature set K out.

The following describes the method to set up of proper vector initial value of advertisement, user and webpage.This patent method needs to arrange proper vector initial value to a part of advertisement in set of advertisements A, also can arrange proper vector initial value to a part of user that user collects in U simultaneously, and arrange proper vector initial value to a part of webpage in webpage collection D.The span of the proper vector initial value of advertisement, user or webpage is set to usually for each s ∈ A, m ∈ U, n ∈ D and k ∈ K, has aw _sk∈ [0,1], uw _mk∈ [0,1] and dw _nk∈ [0,1].If the proper vector of advertisement, user or webpage is not set up initial value, proper vector initial value is default is set to null vector for it.Below for advertisement s, user m and webpage n characterization vector initial value method to set up.

Example 1. directly arranges method.Feature sum L=5, feature set K=(science, education, finance and economics, music, physical culture) are such as set, the proper vector (aw of described advertisement s is set _s1, aw _s2, aw _s3, aw _s4, aw _s5) initial value be (0,0.9,0,1,0), namely the degree of correlation of advertisement s and " education " feature is 0.9, is 1, is zero with the degree of correlation of further feature with the degree of correlation of " music " feature.Same method can arrange the proper vector (uw of user m _m1, uw _m2, uw _m3, uw _m4, uw _m5) initial value, and the proper vector (dw of described webpage n is set _n1, dw _n2, dw _n3, dw _n4, dw _n5) initial value.

Example 2. arranges the method for the proper vector initial value of user m (m ∈ U).First a web pages set H={... is submitted to by described user m, r ... } , the proper vector of described webpage r (r ∈ H) is (dw _r1, dw _r2..., dw _rL), then for each feature k ∈ K, uw is set _mk=(σ/x) ∑ _{(r ∈ H)}dw _rk, wherein x is the element number of described set H, and σ is preset constant.Same method, described user m also can collect in U described user and selects one group of user to calculate the proper vector initial value of described user m.

Example 3. arranges the method for the proper vector initial value of webpage.Split catalog is a kind of special web page, and such as portal website generally includes the split catalogs such as news, music, physical culture, finance and economics and science and technology.We suppose that the webpage under same category catalogue has some identical feature, and such as, webpage under physical culture catalogue is all relevant to physical culture.If described webpage n (n ∈ D) is a webpage under split catalog h (h ∈ D), then the proper vector initial value of described webpage n is decided by the proper vector of described split catalog h.Such as each k ∈ K, dw is set _nk=σ dw _hk, wherein σ is preset constant.

Each advertisement, user and webpage are provided with the mark whether its proper vector can be updated, described in be masked as 1 expression advertisement, the proper vector of user or webpage can be updated, described in be masked as 0 representation feature vector and cannot be updated.

Fig. 4 is a kind of accurate put-on method process flow diagram of personalization of Internet advertising.This process flow diagram is included in the server of accessing Internet and performs following steps:

Step S1. obtains and stores user and collects U={1,2 ..., M}, webpage collection D={1,2 ..., N}, set of advertisements A={1,2 ..., P} and feature set K={1,2 ..., L};

Step S2. is that at least one advertisement in described set of advertisements A arranges proper vector initial value;

Step S3. receives the signal that arbitrary user m (m ∈ U) clicks arbitrary advertisement s (s ∈ A), and according to this signal update the proper vector of user m and the proper vector of described advertisement s;

Step S4. receives the signal that arbitrary user m (m ∈ U) clicks arbitrary webpage n (n ∈ D), and according to this signal update the proper vector of user m and the proper vector of described webpage n;

Step S5., according to the described signal in described step S4, calculates the personalized ordering value of each advertisement in described set of advertisements A, and gives described user m according to described personalized ordering value by least one advertisement putting, returns described step S3.

In the method described in Fig. 4, any one user in described user m representative of consumer collection U, and do not refer in particular to certain user, described webpage n represents any one webpage in webpage collection D, and do not refer in particular to certain webpage, any one advertisement in described advertisement behalf set of advertisements A, and do not refer in particular to certain advertisement.Such as, when certain performs described step S3 to S4 m=1023, n=3428, s=305, and m=33456, n=28477, s=1049 when performing described step S3 to S4 next time.

In an application example of the method described in Fig. 4, described method is also included in and performs after described step S3 to S4 reaches preset times, under each feature k ∈ K, respectively to kth user's column vector (uw _1k, uw _2k..., uw _mk), to a kth webpage column vector (dw _1k, dw _2k..., dw _nk) and to a kth advertisement column vector (aw _1k, aw _2k..., aw _pk) carry out the step of standardization processing.Two following examples are with kth user's column vector (uw _1k, uw _2k..., uw _mk) be example implementation treatment step.Same method, can carry out standardization processing to a kth webpage column vector and a kth advertisement column vector.According to arranging requirement, the proper vector of some advertisement, user and webpage can not make to carry out standardization processing with the following method.

Example 1. couples of users collect kth user's column vector (uw in U _1k, uw _2k..., uw _mk) carry out standardization processing method as follows: each user m ∈ U that can be updated for proper vector, if uw _mk>=U _maxk () then establishes uw _mk=1, otherwise uw is set _mk=uw _mk/ U _max(k), wherein said U _maxk () is kth user's column vector (uw _1k, uw _2k..., uw _mk) in the maximum M of numerical value ₁the mean value of individual component, described M ₁for preset constant, or described U is set _maxk () is preset constant, as U _max(k)=1.02.

Example 2. couples of users collect kth user's column vector (uw in U _1k, uw _2k..., uw _mk) carry out standardization processing method as follows: each user m ∈ U that can be updated for proper vector, if uw _mk>=U _maxk () then establishes uw _mk=1, otherwise uw is set _mk=h _o[uw _mk/ U _max(k)], wherein said U _maxk () is kth user's column vector (uw _1k, uw _2k..., uw _mk) in the maximum M of numerical value ₁the mean value of individual component, described M ₁for preset constant, or described U is set _maxk () is preset constant.Described h ₀x () is increasing function, such as establish h ₀(x)=x ^{u (k)}, described u (k) is the preset constant relevant to feature k, such as, arrange u (k) ∈ [1/5,5].

In an application example of method described in Fig. 4, described step S2 also comprises at least one user collected in U for described user and arranges proper vector initial value, and arranges the step of proper vector initial value at least one webpage in described webpage collection D.

Fig. 5 is the concrete methods of realizing of step S3 in method described in Fig. 4.Comprise the steps:

Step S31. receives the signal that any one user m (m ∈ U) clicks any one advertisement s (s ∈ A);

Step S32., according to described signal, reads the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL), wherein said uw _mkrepresent the degree of correlation of described user m and feature k (k ∈ K);

Step S33., according to described signal, reads the proper vector (aw of described advertisement s _s1, aw _s2..., aw _sk..., aw _sL), wherein said aw _skrepresent the degree of correlation of described advertisement s and feature k (k ∈ K);

Step S34. applies following algorithm, upgrades the proper vector of described user m and described advertisement s,

Uw _mk=uw _mk+ λ ₁(s, m, k) f ₁(aw _sk) (for each )

Aw _sk=aw _sk+ λ ₂(m, s, k) f ₂(uw _mk) (for each )

Wherein said λ ₁(s, m, k) be under described feature k described advertisement s to the influence coefficient of described user m, λ ₁(s, m, k)>=0, described λ ₂(m, s, k) be under described feature k described user m to the influence coefficient of described advertisement s, λ ₂(m, s, k)>=0, f ₁(x) and f ₂x () is non-negative increasing function, described UK _mby the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL) in the maximum P of numerical value _mthe set of the feature composition corresponding to individual component, described AK _sby the proper vector (aw of described advertisement s _s1, aw _s2..., aw _sk..., aw _sL) in the maximum Q of numerical value _sthe set of the feature composition corresponding to individual component, described P _mwith described Q _sfor preset constant.

In an application example of method described in Fig. 5, described in $f_1\left({\mathrm{aw}}_{\mathrm{sk}}\right)={\mathrm{\σ}}_1\·{[{\mathrm{aw}}_{\mathrm{sk}}/{\mathrm{\Σ}}_{(k\∈{\mathrm{AK}}_s)}{\mathrm{aw}}_{\mathrm{sk}}]}^{a\left(k\right)},$ Described $f_2\left({\mathrm{uw}}_{\mathrm{mk}}\right)={\mathrm{\σ}}_2\·{[{\mathrm{uw}}_{\mathrm{mk}}/{\mathrm{\Σ}}_{(k\∈{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]}^{u\left(k\right)},$ Wherein, described a (k) and u (k) are the preset constant relevant to feature k, such as, arrange a (k) ∈ [1/5,5] and u (k) ∈ [1/5,5], described σ ₁and σ ₂for preset constant.

In an application example of method described in Fig. 5, described f ₁(aw _sk)=σ ₁[aw _sk] ^{a (k)}, described f ₂(uw _mk)=σ ₂[uw _mk] ^{u (k)}, wherein, described a (k) and described u (k) are the preset constant relevant to feature k, such as, arrange a (k) ∈ [1/5,5], u (k) ∈ [1/5,5], described σ ₁and σ ₂for preset constant.

In an application example of method described in Fig. 5, described f is set ₁(aw _sk) and described f ₂(uw _mk) the upper bound be respectively D ₁and D ₂if, f ₁(aw _sk)>=D ₁then f is set ₁(aw _sk)=D ₁if, f ₂(uw _mk)>=D ₂then f is set ₂(uw _mk)=D ₂, wherein said D ₁and D ₂it is preset constant.To each feature k ∈ K, if dw _nkbe less than preset constant, then get f ₁(aw _sk)=0, if uw _mkbe less than preset constant, then get f ₂(uw _mk)=0.

In an application example of method described in Fig. 5, described influence coefficient λ ₁(s, m, k) and described influence coefficient λ ₂the concrete method to set up of (m, s, k) comprises following example:

λ described in example 1. ₁(s, m, k) and described λ ₂(m, s, k) is the increasing function of the similarity sim (m, s) between the proper vector of described user m and described advertisement s respectively.Such as λ is set ₁(s, m, k)=1+c ₁sim (m, s), λ ₂(m, s, k)=1+c ₂sim (m, s), wherein c ₁and c ₂for preset constant, and

sim(m,s)=[∑ _(k∈K)(uw _mk·aw _sk)]／{[∑ _(k∈K)(uw _rnk) ²] ^1/2·[∑ _(k∈K)(aw _sk) ²] ^1/2}

The implication of this example be described user m and described advertisement s proper vector between similarity higher, the scale-up factor of their " ballots " is each other larger.In addition, when calculating sim (m, s), if aw _sk≤ min_aC _k, then aw is got _sk=0, if uw _mk≤ min_uC _k, then uw is got _mk=0, wherein min_aC _kand min_uC _kthe preset constant relevant to feature k respectively.

λ described in example 2. ₁(s, m, k) is b ₁the increasing function of (k), described λ ₂(m, s, k) is b ₂k the increasing function of (), such as arranges λ ₁(s, m, k)=b ₁(k) and λ is set ₂(m, s, k)=b ₂(k), wherein said b ₁(k) and described b ₂k () is the default normal number relevant to feature k respectively.

λ described in example 3. ₁(s, m, k) and described λ ₂(m, s, k) is the subtraction function that described user m clicks the frequency of described set of advertisements A respectively, as established λ ₁(s, m, k)=1/h ₁[freq (m)], λ ₂(m, s, k)=1/h ₁[freq (m)].Described h ₁x () is increasing function, such as establish described h ₁x () is piecewise function, if x<a ₁then h ₃x ()=1, if x>=a ₁then h ₁(x)=1+a ₂(x-a ₁), if x>=a ₃>a ₁then get λ ₁(s, m, k)=0 and λ ₂(m, s, k)=0, wherein a ₁, a ₂and a ₃for default normal number.Described freq (m) clicks the frequency of the advertisement in described set of advertisements A for described user m, and the described frequency refers to that described user m clicks the number of times of the advertisement in described set of advertisements A in preset period of time.

λ described in example 4. ₂(m, s, k) is the subtraction function of the clicked frequency of described advertisement s, as established λ ₂(m, s, k)=1/h ₂[freq1 (s)], wherein said h ₂x () is increasing function, described h ₂the method to set up of (x) and described h ₁x () is similar.Described freq1 (s) is the clicked frequency of described advertisement s, and the described frequency refers to the number of times that described advertisement s is clicked in preset period of time.

λ described in example 5. ₁(s, m, k) is increasing function, described λ ₂(m, s, k) is $\left\{\right[{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]/[{\mathrm{\&Sigma;}}_{(k\&Element;K)}{\mathrm{uw}}_{\mathrm{mk}}\left]\right\}$ Increasing function.Such as arrange ${\mathrm{tmp}}_1=\left\{\right[{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{AK}}_s)}{\mathrm{aw}}_{\mathrm{sk}}]/[{\mathrm{\&Sigma;}}_{(k}$ $\mathrm{aw}_{\mathrm{sk}}^{\&Element;K)}\left]\right\}$ , then λ ₁(s, m, k)=1+ σ tmp ₁, such as arrange ${\mathrm{tmp}}_2=\left\{\right[{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]/[{\mathrm{\&Sigma;}}_{(k\&Element;K)}{\mathrm{uw}}_{\mathrm{mk}}\left]\right\},$ Then λ ₂(m, s, k)=1+ σ tmp ₂, wherein said σ is preset constant.

λ described in example 6. ₁(s, m, k)=a ₁(s) u ₂(m), described λ ₂(m, s, k)=u ₁(m) a ₂(s), wherein a ₁s () represents the proper vector whether proper vector of advertisement s may be used for upgrading user and collect user in U, u ₂m () represents whether the proper vector of user m can be upgraded by the proper vector of advertisement in set of advertisements A, u ₁m () represents whether the proper vector of user m may be used for upgrading the proper vector of advertisement in set of advertisements A, a ₂s () represents whether the proper vector of advertisement s can be collected the proper vector of user in U by user and upgrade.U ₁(m), u ₂(m), a ₁(s) and a ₂s () is parameter preset, their value is 0 or 1.1 representative be, 0 represent no.

λ described in example 7. ₁(s, m, k)=a (x _s), described λ ₂(m, s, k)=b (y _m), wherein said x _sfor the significance level score value of described advertisement s, the clicked number of times of such as described advertisement s or bid, described y _mfor the significance level score value of described user m, the such as bean vermicelli quantity of described user m, described a (x) and b (y) are increasing function, described a (x) is mapped to a pre-set interval [1 x, d], described b (y) is mapped to a pre-set interval [1, e] y, and wherein d and e is preset constant.

Example 8. uses the combination of at least two kinds of methods in each method described in above-mentioned example 1 to example 7, generates described λ ₁(s, m, k) and λ ₂(m, s, k), such as

λ ₁(s,m,k)=[1+c ₁·sim(m,s)]·[a ₁(s)·u ₂(m)].b ₁(k)·a(x _s)

λ ₂(m,s，k)=[1+c ₂·sim(m,s)]·[u ₁(m)·a ₂(s)]·b ₂(k)·b(y _m)

Fig. 6 is the concrete methods of realizing of step S4 in method described in Fig. 4.Comprise the steps:

Step S41. receives the signal that any one user m (m ∈ U) clicks any one webpage n (n ∈ D);

Step S42., according to described signal, reads the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL), wherein said uw _mkrepresent the degree of correlation of described user m and feature k (k ∈ K);

Step S43., according to described signal, reads the proper vector (dw of described webpage n _n1, dw _n2..., dw _nk..., dw _nL), wherein said dw _nkrepresent the degree of correlation of described webpage n and feature k (k ∈ K);

Step S44. applies following algorithm, upgrades the proper vector of described user m and described webpage n,

Uw _mk=uw _mk+ λ ₃(n, m, k) f ₃(dw _nk) (for each )

Dw _nk=dw _nk+ λ ₄(m, n, k) f ₄(uw _mk) (for each )

Wherein, described λ ₃(n, m, k) be under described feature k described webpage n to the influence coefficient of described user m, and λ ₃(n, m, k)>=0, described λ ₄(m, n, k) be under described feature k described user m to the influence coefficient of described webpage n, and λ ₄(m, n, k)>=0, f ₃(x) and f ₄x () is non-negative increasing function, described UK _mby the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL) in the maximum P of numerical value _mthe set of the feature composition corresponding to individual component, described DK _nby the proper vector (dw of described webpage n _n1, dw _n2..., dw _nk..., dw _nL) in the maximum R of numerical value _nthe set of the feature composition corresponding to individual component, described P _mwith described R _nfor preset constant.

In an application example of method described in Fig. 6, described in $f_3\left({\mathrm{dw}}_{\mathrm{nk}}\right)={\mathrm{\&sigma;}}_3\&CenterDot;{[{\mathrm{dw}}_{\mathrm{nk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{DK}}_n)}{\mathrm{dw}}_{\mathrm{nk}}]}^{d\left(k\right)},$ Described $f_4\left({\mathrm{uw}}_{\mathrm{mk}}\right)={\mathrm{\&sigma;}}_4\&CenterDot;{[{\mathrm{uw}}_{\mathrm{mk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]}^{u\left(k\right)},$ Wherein said d (k) and described u (k) are the preset constant relevant to feature k, such as, establish d (k) ∈ [1/5,5], u (k) ∈ [1/5,5], described σ ₃and σ ₄for preset constant.

In an application example of method described in Fig. 6, described f ₃(dw _nk)=σ ₃[dw _nk] ^{d (k)}, described f ₄(uw _mk)=σ ₄[uw _mk] ^{u (k)}, wherein said d (k) and described u (k) are the preset constant relevant to feature k, such as, arrange d (k) ∈ [1/5,5], u (k) ∈ [1/5,5], described σ ₃and σ ₄for preset constant.

In an application example of method described in Fig. 6, described f is set ₃(dw _nk) and described f ₄(uw _mk) the upper bound be respectively D ₃and D ₄if, f ₃(dw _nk)>=D ₃then f is set ₃(dw _nk)=D ₃if, f ₄(uw _mk)>=D ₄then f is set ₄(uw _mk)=D ₄, wherein said D ₃and D ₄it is preset constant.For each feature k ∈ K, if dw _nkbe less than preset constant, then get f ₃(dw _nk)=0, if uw _mkbe less than preset constant, then get f ₄(uw _mk)=0.

In an application example of method described in Fig. 6, described influence coefficient λ ₃(n, m, k) and described influence coefficient λ ₄the concrete method to set up of (m, n, k), comprises following example:

λ described in example 1. ₃(n, m, k) and described λ ₄(m, n, k) is the increasing function of the similarity sim (m, n) between the proper vector of described user m and described webpage n respectively.Such as λ is set ₃(n, m, k)=1+c ₃sim (m, n), λ ₄(m, n, k)=1+c ₄sim (m, n), wherein c ₃and c ₄for preset constant, and

sim(m，n)=[∑ _(k∈K)(uw _mk·dw _nk)]／{[∑ _(k∈K)(uw _mk) ²] ^1/2·[∑( _k∈K)(dw _nk) ²] ^1/2}

The implication of this example be described user m and described webpage n proper vector between similarity higher, the scale-up factor of their " ballots " is each other larger.In addition, when calculating sim (m, n), if dw _nk≤ mindC _k, then dw is got _nk=0, if uw _mk≤ min_uC _k, then uw is got _mk=0, wherein min_dC _kand min_uC _kthe preset constant relevant to feature k respectively.

λ described in example 2. ₃(n, m, k) is b ₁the increasing function of (k), described λ ₄(m, n, k) is b ₂k the increasing function of (), such as arranges λ ₃(n, m, k)=b ₁(k) and λ is set ₄(m, n, k)=b ₂(k), wherein said b ₁(k) and described b ₂k () is the default normal number relevant to feature k respectively.

λ described in example 3. ₃(n, m, k) and described λ ₄(m, n, k) is the subtraction function that described user m clicks the frequency of described webpage collection D respectively, as established λ ₃(n, m, k)=1/h ₃[freq (m)], λ ₄(m, n, k)=1/h ₃[freq (m)].Described h ₃x () is increasing function, such as establish described h ₃x () is piecewise function, if x<a ₁then h ₃x ()=1, if x>=a ₁then h ₃(x)=1+a ₂(x-a ₁), if x>=a ₃>a ₁then get λ ₃(n, m, k)=0 and λ ₄(m, n, k)=0, wherein a ₁, a ₂and a ₃for default normal number.Described freq (m) is the frequency of the webpage in described user m clicks described webpage collection D, and the described frequency refers to that described user m clicks the number of times of the webpage in described webpage collection D in preset period of time.

λ described in example 4. ₄(m, n, k) is the subtraction function of the clicked frequency of described webpage n, as established λ ₄(m, n, k)=1/h ₄[freq1 (n)], wherein said h ₄x () is increasing function, its method to set up and described h ₃x () is similar, described freq1 (n) is the clicked frequency of described webpage n, and the described frequency refers to the number of times that described webpage n is clicked in preset period of time.

λ described in example 5. ₃(n, m, k) is increasing function, described λ ₄(m, n, k) is $\left\{\right[{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]/[{\mathrm{\&Sigma;}}_{(k\&Element;K)}{\mathrm{uw}}_{\mathrm{mk}}\left]\right\}$ Increasing function.Such as arrange ${\mathrm{tmp}}_1=\left\{\right[{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{DK}}_n)}{\mathrm{dw}}_{\mathrm{nk}}]/[{\mathrm{\&Sigma;}}_{(k}$ $\mathrm{dw}_{\mathrm{nk}}^{\&Element;K)}\left]\right\}$ , then λ ₃(n, m, k)=1+ σ tmp ₁, such as arrange ${\mathrm{tmp}}_2=\left\{\right[{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]/[{\mathrm{\&Sigma;}}_{(k\&Element;K)}{\mathrm{uw}}_{\mathrm{mk}}\left]\right\},$ Then λ ₄(m, n, k)=1+ σ tmp ₂, wherein said σ is preset constant.

λ described in example 6. ₃(n, m, k)=d ₁(n) u ₂(m), described λ ₄(m, n, k)=u ₁(m) d ₂(n), wherein d ₁n () represents the proper vector whether proper vector of webpage n may be used for upgrading user and collect user in U, u ₂m () represents whether the proper vector of user m can be upgraded by the proper vector of webpage in webpage collection D, u ₁m () represents whether the proper vector of user m may be used for the proper vector of webpage in more new web page collection D, d ₂n () represents whether the proper vector of webpage n can be collected the proper vector of user in U by user and upgrade.U ₁(m), u ₂(m), d ₁(n) and d ₂n () is parameter preset, their value is 0 or 1.1 representative be, 0 represent no.

λ described in example 7. ₃(n, m, k)=c (z _n), described λ ₄(m, n, k)=b (y _m), wherein said z _nfor the significance level score value of described webpage n, the clicked number of times of such as described webpage n or PageRank value.Described y _mfor the significance level score value of described user m, the such as bean vermicelli quantity of described user m, described c (z) and b (y) are increasing function, described c (z) is mapped to a pre-set interval [1 z, d], described b (y) is mapped to a pre-set interval [1, e] y, and wherein d and e is preset constant.

Example 8. uses the combination of at least two kinds of methods in each method described in above-mentioned example 1 to example 7, generates described λ ₃(n, m, k) and λ ₄(m, n, k), such as

λ ₃(n,m,k)=[1+c ₃·sim(m,n)]·[d ₁(n)·u ₂(m)]·b ₁(k)·c(z _n)

λ ₄(m,n,k)=[1+c ₄·sim(m,n)]·[u ₁(m)·d ₂(n)]·b ₂(k)·b(y _m)

In an application example of method described in Fig. 6, the signal received in described step S4 randomly draws from numerous signals of user's webpage clicking in a preset period of time.Such as in a described preset period of time, described user is collected to the described signal of click signal as described step S4 of each any active ues extraction equal number in U.Described any active ues refers in a described preset period of time, clicks the user that described webpage collection D reaches preset times.

Fig. 7 is the concrete methods of realizing of step S5 in method described in Fig. 4.Comprise the steps:

S51. according to the described signal of described step S4, the proper vector (uw of described user m is read _m1, uw _m2..., uw _mk..., uw _mL), wherein said uw _mkrepresent the degree of correlation of described user m and feature k (k ∈ K);

S52. according to the described signal of described step S4, the proper vector (dw of described webpage n is read _n1, dw _n2..., dw _nk..., dw _nL), wherein said dw _nkrepresent the degree of correlation of described webpage n and feature k (k ∈ K);

S53. according to the proper vector of each advertisement in the proper vector of described user m, the proper vector of described webpage n and described set of advertisements A, the personalized ordering value of each advertisement in described set of advertisements A is calculated;

S54. according to described personalized ordering value, each advertisement in described set of advertisements A is sorted, and give described user m according to ranking results by least one advertisement putting.

In an application example of method described in Fig. 7, the computing method of described personalized ordering value are as follows.With AR (g, m) similarity of advertisement g (g ∈ A) and user m (m ∈ U) is represented, with DR (g, n) similarity of advertisement g (g ∈ A) and webpage n (n ∈ D) is represented, with Rank (g|m, n) the personalized ordering value of advertisement g under the prerequisite of user m webpage clicking n is represented, therefore, just like giving a definition:

AR(g，m)=[∑ _(k∈K)(aw _gk·uw _mk)]／{[∑ _(k∈K)(aw _gk) ²] ^1/2·[∑ _(k∈K)(uw _mk) ²] ^1/2}

DR(g,n)=[∑ _(k∈K)(aw _gk·dw _nk)]／{[∑ _(k∈K)(aw _gk) ²] ^1/2·[∑ _(k∈K)(dw _nk) ²] ^1/2}

Rank(g|m，n)=β·AR(g,m)+(1-β)·DR(g,n)

Wherein said aw _gkrepresent the degree of correlation of advertisement g and feature k (k ∈ K), described uw _mkrepresent the degree of correlation of user m and feature k (k ∈ K), described dw _nkrepresent the degree of correlation of webpage n and feature k (k ∈ K), β is a preset constant, and β ∈ [0,1].

Application example 1

This is a concrete methods of realizing of method described in Fig. 5.Suppose have two users and three width advertisements on the internet, namely user collects U={1,2}, set of advertisements A={1,2,3}.If feature set K={1,2}, P ₁=P ₂=Q ₁=Q ₂=Q ₃=2, and UK ₁=UK ₂=AK ₁=AK ₂=AK ₃=K={1,2}.The proper vector of user 1 and user 2 is set to (uw respectively ₁₁, uw ₁₂) and (uw ₂₁, uw ₂₂), the proper vector of advertisement 1, advertisement 2 and advertisement 3 is set to (aw respectively ₁₁, aw ₁₂), (aw ₂₁, aw ₂₂) and (aw ₃₁, aw ₃₂), wherein uw _mkrepresent the degree of correlation of described user m ∈ U and feature k, aw _skrepresent the degree of correlation of described advertisement s ∈ A and feature k.Any advertisement s ∈ A is arranged $f_1\left({\mathrm{aw}}_{\mathrm{sk}}\right)={\mathrm{\&sigma;}}_1\&CenterDot;{[{\mathrm{aw}}_{\mathrm{sk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{AK}}_s)}{\mathrm{aw}}_{\mathrm{sk}}]}^{a\left(k\right)},$ Any user m ∈ U is arranged $f_2\left({\mathrm{uw}}_{\mathrm{mk}}\right)={\mathrm{\&sigma;}}_2\&CenterDot;$ ${[{\mathrm{uw}}_{\mathrm{mk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]}^{u\left(k\right)}.$

Repeatedly receive user to the click signal of advertisement, and upgrade the proper vector of user and advertisement respectively.Such as after have received user 2 and clicking the signal of advertisement 3, then according to algorithm and the f of described step S34 ₁(aw _sk) and f ₂(uw _mk) above-mentioned definition, the proper vector of user 2 and advertisement 3 is upgraded as follows:

uw ₂₁=uw ₂₁+λ ₁(3，2，1)·σ1·[aw ₃₁／(aw ₃₁+aw ₃₂)] ^a(1)

uw ₂₂=uw ₂₂+λ ₁(3，2，2)·σ ₁·[aw ₃₂／(aw ₃₁+aw ₃₂)] ^a(2)

aw ₃₁=aw ₃₁+λ ₂(2，3，1)·σ ₂·[uw ₂₁／(uw ₂₁+uw ₂₂)] ^u(1)

aw ₃₂=aw ₃₂+λ ₂(2，3，2)·σ ₂·[uw ₂₂／(uw ₂₁+uw ₂₂)] ^u(2)

Wherein, σ ₁=σ ₂=001, a (1)=a (2)=u (1)=u (2)=1, λ ₁(3,2, k) represent that described advertisement 3 is to the influence coefficient of described user 2 under described feature k ∈ K, λ ₂(2,3, k) represent that described user 2 is to the influence coefficient of described advertisement 3 under described feature k ∈ K, and each k ∈ K is had

λ ₁(3，2，k)=[1+c ₁·sim(2，3)]·{1／h ₁[freq(2)]}·b ₁(k)·[a ₁(3)·u ₂(2)]

λ ₂(2，3，k)=[1+c ₂·sim(2，3)]·{1／h ₁[freq(2)]}·b ₂(k)·[u ₁(2)·a ₂(3)]

Wherein, c ₁=c ₂=5, sim (2,3)=(uw ₂₁aw ₃₁+ uw ₂₂aw ₃₂)/{ [(uw ₂₁) ²+ (uw ₂₂) ²] ^1/2[(aw ₃₁) ²+ (aw ₃₂) ²] ^1/2, h ₁(x)=1+a ₂(x-a ₁), a ₁=300, a ₂=0.01, freq (2)=500, b ₁(1)=b ₂(1)=1, b ₁(2)=b ₂(2)=1.5, u ₁(2)=u ₂(2)=a ₁(3)=a ₂(3)=1.

After executing above-mentioned algorithm, to user's column vector (uw ₁₁, uw ₂₁) and (uw ₁₂, uw ₂₂) carry out standardization processing, and to advertisement column vector (aw ₁₁, aw ₂₁, aw ₃₁) and (aw ₁₂, aw ₂₂, aw ₃₂) carry out standardization processing.

As follows to the algorithm of user's column vector standardization processing: to establish then feature k=1 is arranged ${\mathrm{uw}}_{11}={\mathrm{uw}}_{11}/\stackrel{\&OverBar;}{{\mathrm{uw}}_1},{\mathrm{uw}}_{21}={\mathrm{uw}}_{21}/\stackrel{\&OverBar;}{{\mathrm{uw}}_1};$ If $\stackrel{\&OverBar;}{{\mathrm{uw}}_2}=\max ({\mathrm{uw}}_{12},{\mathrm{uw}}_{22}),$ Then feature k=2 is arranged ${\mathrm{uw}}_{12}={\mathrm{uw}}_{12}/\stackrel{\&OverBar;}{{\mathrm{uw}}_2},{\mathrm{uw}}_{22}={\mathrm{uw}}_{22}/\stackrel{\&OverBar;}{{\mathrm{uw}}_2}.$

As follows to the algorithm of advertisement column vector standardization processing: to establish then feature k=1 is arranged ${\mathrm{aw}}_{11}={\mathrm{aw}}_{11}/\stackrel{\&OverBar;}{{\mathrm{aw}}_1},{\mathrm{aw}}_{21}={\mathrm{aw}}_{21}/\stackrel{\&OverBar;}{{\mathrm{aw}}_1},{\mathrm{aw}}_{31}={\mathrm{aw}}_{31}/\stackrel{\&OverBar;}{{\mathrm{aw}}_1};$ If $\stackrel{\&OverBar;}{{\mathrm{aw}}_2}=$ $\max ({\mathrm{aw}}_{12},{\mathrm{aw}}_{22},{\mathrm{aw}}_{32}),$ Then feature k=2 is arranged ${\mathrm{aw}}_{12}={\mathrm{aw}}_{12}/\stackrel{\&OverBar;}{{\mathrm{aw}}_2},{\mathrm{aw}}_{22}={\mathrm{aw}}_{22}/\stackrel{\&OverBar;}{{\mathrm{aw}}_2},$ ${\mathrm{aw}}_{32}={\mathrm{aw}}_{32}/\stackrel{\&OverBar;}{{\mathrm{aw}}_2}.$

Application example 2

This is a concrete methods of realizing of method described in Fig. 6.Suppose have two users and three webpages on the internet, namely user collects U={1,2}, webpage collection D={1,2,3}.If feature set K={1,2}, P ₁=P ₂=R ₁=R ₂=R ₃=2, and UK ₁=UK ₂=DK ₁=DK ₂=DK ₃=K={1,2}.The proper vector of user 1 and user 2 is set to (uw respectively ₁₁, uw ₁₂) and (uw ₂₁, uw ₂₂), the proper vector of webpage 1, webpage 2 and webpage 3 is set to (dw respectively ₁₁, dw ₁₂), (dw ₂₁, dw ₂₂) and (dw ₃₁, dw ₃₂), wherein uw _mkrepresent the degree of correlation of described user m ∈ U and feature k, dw _nkrepresent the degree of correlation of described webpage n ∈ D and feature k.Any webpage n ∈ D is arranged $f_3\left({\mathrm{dw}}_{\mathrm{nk}}\right)={\mathrm{\&sigma;}}_3\&CenterDot;{[{\mathrm{dw}}_{\mathrm{nk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{DK}}_n)}{\mathrm{dw}}_{\mathrm{nk}}]}^{d\left(k\right)},$ Any user m ∈ U is arranged $f_4\left({\mathrm{uw}}_{\mathrm{mk}}\right)={\mathrm{\&sigma;}}_4\&CenterDot;$ ${[{\mathrm{uw}}_{\mathrm{mk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]}^{u\left(k\right)},$ Described d (k) and described u (k) are the preset constant relevant to described feature k.

Repeatedly receive user to the click signal of webpage, and upgrade the proper vector of user and webpage respectively.Such as when after the signal that have received user 2 webpage clicking 3, then according to algorithm and the f of described step S44 ₃(dw _nk) and f ₄(uw _mk) above-mentioned definition, the proper vector of user 2 and webpage 3 is upgraded as follows:

uw ₂₁=uw ₂₁+λ ₃(3，2，1)·σ ₃·[dw ₃₁／(dw ₃₁+dw ₃₂)] ^d(1)

uw ₂₂=uw ₂₂+λ ₃(3，2，2)·σ ₃·[dw ₃₂／(dw ₃₁+dw ₃₂)] ^d(2)

dw ₃₁=dw ₃₁+λ ₄(2，3，1)·σ ₄·[uw ₂₁／(uw ₂₁+uw ₂₂)] ^u(1)

dw ₃₂=dw ₃₂+λ ₄(2，3，2)·σ ₄·[uw ₂₂／(uw ₂₁+uw ₂₂)] ^u(2)

Wherein, σ ₃=σ ₄=0.01, d (1)=d (2)=u (1)=u (2)=1, λ ₃(3,2, k) represent that described webpage 3 is to the influence coefficient of described user 2 under described feature k, λ ₄(2,3, k) represent that described user 2 is to the influence coefficient of described webpage 3 under described feature k, and each k ∈ K is had

λ ₃(3，2，k)=[1+c ₃·sim(2，3)]·{1／h ₃[freq(2)]}·b ₁(k)·[d ₁(3)·u ₂(2)]

λ ₄(2，3，k)=[1+c ₄·sim(2，3)]·{1／h ₃[freq(2)]}·b ₂(k)·[u ₁(2)·d ₂(3)]

Wherein, c ₃=c ₄=5, sim (2,3)=(uw ₂₁dw ₃₁+ uw ₂₂dw ₃₂)/{ [(uw ₂₁) ²+ (uw ₂₂) ²] ^1/2[(dw ₃₁) ²+ (dw ₃₂) ²] ^1/2, h ₃(x)=1+a ₂(x-a ₁), a ₁=300, a ₂=0.01, freq (2)=500, b ₁(1)=b ₂(1)=1, b ₁(2)=b ₂(2)=1.5, u ₁(2)=u ₂(2)=d ₁(3)=d ₂(3)=1.

After executing above-mentioned algorithm, to user's column vector (uw ₁₁uw ₂₁) and (uw ₁₂, uw ₂₂) carry out standardization processing, and to webpage column vector (dw ₁₁, dw ₂₁, dw ₃₁) and (dw ₁₂, dw ₂₂, dw ₃₂) carry out standardization processing.

As follows to the algorithm of user's column vector standardization processing: to establish then feature k=1 is arranged ${\mathrm{uw}}_{11}={\mathrm{uw}}_{11}/\stackrel{\&OverBar;}{{\mathrm{uw}}_1},{\mathrm{uw}}_{21}={\mathrm{uw}}_{21}/\stackrel{\&OverBar;}{{\mathrm{uw}}_1};$ If $\stackrel{\&OverBar;}{{\mathrm{uw}}_2}=\max ({\mathrm{uw}}_{12},{\mathrm{uw}}_{22}),$ Then feature k=2 is arranged ${\mathrm{uw}}_{12}={\mathrm{uw}}_{12}/\stackrel{\&OverBar;}{{\mathrm{uw}}_2},{\mathrm{uw}}_{22}={\mathrm{uw}}_{22}/\stackrel{\&OverBar;}{{\mathrm{uw}}_2}.$

As follows to the algorithm of webpage column vector standardization processing: to establish then feature k=1 is arranged ${\mathrm{dw}}_{11}={\mathrm{dw}}_{11}/\stackrel{\&OverBar;}{{\mathrm{dw}}_1},{\mathrm{dw}}_{21}={\mathrm{dw}}_{21}/\stackrel{\&OverBar;}{{\mathrm{dw}}_1},{\mathrm{dw}}_{31}={\mathrm{dw}}_{31}/\stackrel{\&OverBar;}{{\mathrm{dw}}_1};$ If $\stackrel{\&OverBar;}{{\mathrm{dw}}_2}=$ $\max ({\mathrm{dw}}_{12},{\mathrm{dw}}_{22},{\mathrm{dw}}_{32}),$ Then feature k=2 is arranged ${\mathrm{dw}}_{12}={\mathrm{dw}}_{12}/\stackrel{\&OverBar;}{{\mathrm{dw}}_2},{\mathrm{dw}}_{22}={\mathrm{dw}}_{22}/\stackrel{\&OverBar;}{{\mathrm{dw}}_2},$ ${\mathrm{dw}}_{32}={\mathrm{dw}}_{32}/\stackrel{\&OverBar;}{{\mathrm{dw}}_2}.$

The above application example is only preferably application example of the present invention, and is not used to limit protection scope of the present invention.

Claims (10)

1. the accurate put-on method of the personalization of Internet advertising, is characterized in that, described method is included in the server of accessing Internet and performs following steps:

Step S1. obtains and stores user and collects U={1,2 ..., M}, webpage collection D={1,2 ..., N}, set of advertisements A={1,2 ..., P} and feature set K={1,2 ..., L};

Step S2. is that at least one advertisement in described set of advertisements A arranges proper vector initial value;

Step S3. receives the signal that arbitrary user m (m ∈ U) clicks arbitrary advertisement s (s ∈ A), and according to this signal update the proper vector of user m and the proper vector of described advertisement s;

Step S4. receives the signal that arbitrary user m (m ∈ U) clicks arbitrary webpage n (n ∈ D), and according to this signal update the proper vector of user m and the proper vector of described webpage n;

Step S5., according to the described signal in described step S4, calculates the personalized ordering value of each advertisement in described set of advertisements A, and gives described user m according to described personalized ordering value by least one advertisement putting, returns described step S3.

2. method according to claim 1, it is characterized in that, described step S2 also comprises at least one user collected in U for described user and arranges proper vector initial value, and arranges the step of proper vector initial value at least one webpage in described webpage collection D.

3. method according to claim 1, is characterized in that, a kind of concrete methods of realizing of described step S3 comprises the steps:

Step S31. receives the signal that any one user m (m ∈ U) clicks any one advertisement s (s ∈ A);

Step S32., according to described signal, reads the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL), wherein said uw _mkrepresent the degree of correlation of described user m and feature k (k ∈ K);

Step S33., according to described signal, reads the proper vector (aw of described advertisement s _s1, aw _s2..., aw _sk..., aw _sL), wherein said aw _skrepresent the degree of correlation of described advertisement s and feature k (k ∈ K);

Step S34. applies following algorithm, upgrades the proper vector of described user m and described advertisement s,

Uw _mk=uw _mk+ λ ₁(s, m, k) f ₁(aw _sk) (for each )

Aw _sk=aw _sk+ λ ₂(m, s, k) f ₂(uw _mk) (for each )

Wherein, described λ ₁(s, m, k) be under described feature k described advertisement s to the influence coefficient of described user m, and λ ₁(s, m, k)>=0, described λ ₂(m, s, k) be under described feature k described user m to the influence coefficient of described advertisement s, and λ ₂(m, s, k)>=0, f ₁(x) and f ₂x () is non-negative increasing function, described UK _mby the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL) in the set that forms of the maximum feature corresponding to Pm component of numerical value, described AK _sby the proper vector (aw of described advertisement s _s1, aw _s2..., aw _sk..., aw _sL) in the maximum Q of numerical value _sthe set of the feature composition corresponding to individual component, described P _mwith described Q _sfor preset constant.

4. method according to claim 3, is characterized in that, described method is also included in and performs after described step S3 reaches set point number, under each feature k ∈ K, respectively to a kth user column vector uw _1k, uw _2k..., uw _mk) and a kth advertisement column vector (aw _1k, aw _2k..., aw _pk) carry out the step of standardization processing.

5. method according to claim 3, is characterized in that, described in an application example of described method $f_1\left({\mathrm{aw}}_{\mathrm{sk}}\right)={\mathrm{\&sigma;}}_1\&CenterDot;{[{\mathrm{aw}}_{\mathrm{sk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{AK}}_s)}{\mathrm{aw}}_{\mathrm{sk}}]}^{a\left(k\right)},$ Described $f_2\left({\mathrm{uw}}_{\mathrm{mk}}\right)={\mathrm{\&sigma;}}_2\&CenterDot;{[{\mathrm{uw}}_{\mathrm{mk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]}^{u\left(k\right)},$ Wherein σ ₁and σ ₂for preset constant, described a (k) and described u (k) are the preset constant relevant to feature k.

6. method according to claim 3, is characterized in that, described influence coefficient λ ₁(s, m, k) and described influence coefficient λ ₂(m, s, k) is the increasing function of the similarity between the proper vector of described user m and the proper vector of described advertisement s respectively.

7. method according to claim 1, is characterized in that, a kind of concrete methods of realizing of described step S4 comprises the steps:

Step S41. receives the signal that any one user m (m ∈ U) clicks any one webpage n (n ∈ D);

Step S42., according to described signal, reads the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL), wherein said uw _mkrepresent the degree of correlation of described user m and feature k (k ∈ K);

Step S43., according to described signal, reads the proper vector (dw of described webpage n _n1, dw _n2..., dw _nk..., dw _nL), wherein said dw _nkrepresent the degree of correlation of described webpage n and feature k (k ∈ K);

Step S44 applies following algorithm, upgrades the proper vector of described user m and described webpage n,

Uw _mk=uw _mk+ λ ₃(n, m, k) f ₃(dw _nk) (for each )

Dw _nk=dw _nk+ λ ₄(m, n, k) f ₄(uw _mk) (for each )

Wherein, described λ ₃(n, m, k) be under described feature k described webpage n to the influence coefficient of described user m, and λ ₃(n, m, k)>=0, described λ ₄(m, n, k) be under described feature k described user m to the influence coefficient of described webpage n, and λ ₄(m, n, k)>=0, f ₃(x) and f ₄x () is non-negative increasing function, described UK _mby the proper vector (uw of described user m _m1, uw _m2..., uw _mk..., uw _mL) in the maximum P of numerical value _mthe set of the feature composition corresponding to individual component, described DK _nby the proper vector (dw of described webpage n _n1, dw _n2..., dw _nk..., dw _nL) in the maximum R of numerical value _nthe set of the feature composition corresponding to individual component, described P _mwith described R _nfor preset constant.

8. method according to claim 7, is characterized in that, described method is also included in and performs after described step S4 reaches set point number, under each feature k ∈ K, respectively to kth user's column vector (uw _1k, uw _2k..., uw _mk) and a kth webpage column vector (dw _1k, dw _2k..., dw _nk) carry out the step of standardization processing.

9. method according to claim 7, is characterized in that, described in an application example of described method $f_3\left({\mathrm{dw}}_{\mathrm{nk}}\right)={\mathrm{\&sigma;}}_3\&CenterDot;{[{\mathrm{dw}}_{\mathrm{nk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{DK}}_n)}{\mathrm{dw}}_{\mathrm{nk}}]}^{d\left(k\right)},$ Described $f_4\left({\mathrm{uw}}_{\mathrm{mk}}\right)={\mathrm{\&sigma;}}_4\&CenterDot;{[{\mathrm{uw}}_{\mathrm{mk}}/{\mathrm{\&Sigma;}}_{(k\&Element;{\mathrm{UK}}_m)}{\mathrm{uw}}_{\mathrm{mk}}]}^{u\left(k\right)},$ Wherein σ ₃and σ ₄for preset constant, described d (k) and described u (k) are the preset constant relevant to feature k.

10. method according to claim 1, is characterized in that, a kind of concrete methods of realizing of described step S5 comprises the steps:

S51. according to the described signal in described step S4, the proper vector (uw of described user m is read _m1, uw _m2..., uw _mk... uw _mL), wherein said uw _mkfor the degree of correlation of described user m and feature k (k ∈ K);

S52. according to the described signal in described step S4, the proper vector (dw of described webpage n is read _n1, dw _n2..., dw _nk..., dw _nL), wherein said dw _nkfor the degree of correlation of described webpage n and feature k (k ∈ K);

S53. according to the proper vector of each advertisement in the proper vector of described user m, the proper vector of described webpage n and described set of advertisements A, the personalized ordering value of each advertisement in described set of advertisements A is calculated;

S54. according to described ranking value, each advertisement in described set of advertisements A is sorted, and give described user m according to ranking results by least one advertisement putting.