CN104268644A - Method and device for predicting click frequency of advertisement at advertising position - Google Patents

Abstract

The invention discloses a method and device for predicting the click frequency of an advertisement at an advertising position. The method comprises the steps that with respect to each orientation of the advertisement to be estimated issued at the advertising position, an advertisement exposure frequency prediction model which is selected in advance is used, and an advertisement exposure frequency predicted value of the orientation within a set prediction time period is obtained; according to the sum of statistical advertisement exposure frequency actual values of all the orientations and the sum of statistical advertisement click frequency actual values within a set historical time period, the advertisement click rate within the set historical time period is determined; the products of the advertisement exposure frequency predicted values of all the orientations and the advertisement click rate are calculated respectively, and the sum of the calculated products serves as the advertisement click frequency predicted value within the set prediction time period, wherein the advertisement exposure frequency prediction model is selected from at least two prediction models according to the prediction error rates of at least two advertisement exposure frequency prediction models trained in advance. The method can increase the advertisement click frequency prediction accuracy rate.

Description

The number of clicks predictor method of the advertisement on advertisement position and device

Technical field

The present invention relates to Internet technical field, particularly relate to number of clicks predictor method and the device of the advertisement on a kind of advertisement position.

Background technology

The web advertisement is exactly the advertisement done on network.The web advertisement is a kind of by the high-tech advertising campaign mode of network delivery to Internet user.It utilizes the ad banner on website, text link, multimedia method by web advertisement release platform, in internet publication or releasing advertisements.Wherein, according to CPC (the Cost Per Click that charging divides, by number of clicks charge) advertisement and CPD (Cost Per Day, by number of days charge) advertisement and CPM (Cost Per Mille, by often showing thousand charges) advertisement difference, the number of times that the advertiser of CPC advertisement clicks advertisement according to user is charged, and is one of advertisement form of comparative maturity in network.

Targeted ads refers to can according to Demographic, and the audient for the region of specifying, age, sex, interest, educational background, marital status etc. carries out the form of advertisement putting, and targeted ads can find accurate audient group for client.Such as, for the age, the orientation of women and the orientation of the male sex during advertisement putting, can be comprised.Therefore, the numerous advertisers that are combined into of CPC advertisement and targeted ads favored.

Usually, advertiser, before input advertisement, wishes that certain or some advertisement positions understood on website throw in the effect of advertisement, and formulates the decision-making etc. of subscribing advertisement position according to the input effect understood.Therefore, the requirement of multiple operations such as the inquiry of the advertisement position on website, reservation and release is carried out for meeting advertiser, prior art takes certain mode usually, the number of times clicked to the advertisement that following certain advertisement position of a period of time is thrown in is estimated, and carries out the reference of advertisement putting for advertiser.

In prior art, the predictor method of the ad click number of times on advertisement position is normally estimated based on single prediction model.Wherein, prediction model carries out modeling according to historical trend and obtains.

But, the present inventor finds that the historical trend of different advertisement positions is often different, such as, the fluctuation ratio of ad click number of times to all trend of some advertisement positions is more responsive, the week due to be working day ad click number of times less, Sunday Saturday due to be off-day ad click number of times more, and the ad click number of times of some advertisement positions is more responsive to seasonal trend, therefore, the ad click number of times of single prediction model to different advertisement position is used to estimate, often make to estimate accuracy rate not high, even may occur that the ad click number of times estimated is the situation that negative value etc. does not conform to convention.

Therefore, be necessary to provide a kind of method more adequately can carried out the ad click number of times on advertisement position and estimate.

Summary of the invention

For the defect that above-mentioned prior art exists, embodiments provide predictor method and the device of the ad click number of times on a kind of advertisement position, in order to more adequately to carry out estimating of ad click number of times.

Embodiments provide the predictor method of the ad click number of times on a kind of advertisement position, comprising:

For each orientation of the advertisement that advertisement position to be estimated has been thrown in, use the prediction model of the advertisement exposure number selected in advance, obtain the discreet value of the advertisement exposure number setting this orientation in prediction time horizon; And

According to the actual value sum of advertisement exposure number and the actual value sum of ad click number of times of each orientation in the setting historical time section counted, determine the ad click rate in described setting historical time section;

Calculate the discreet value of the advertisement exposure number of each orientation and the product of described ad click rate respectively, using the discreet value as the ad click number of times in described setting prediction time horizon of each sum of products of calculating;

Wherein, the prediction model of described advertisement exposure number be at least two prediction models of the advertisement exposure number according to training in advance estimate error rate, choose out from described at least two prediction models.

Wherein, described at least two prediction models comprise: the first prediction model and the 4th prediction model;

In first prediction model and the 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number in the i-th-T sky, and T is predetermined period;

Wherein, in the first prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=growth (i) × seasonality (i) (formula 5)

In formula 5, seasonality (i) is the first correction factor, and t ₁it is the first predetermined period; Growth (i)=a ₁× i+b ₁; a ₁and b ₁it is the training parameter of the first prediction model; H [i-T ₁] be the i-th-T ₁the actual value of it advertisement exposure number; And

In 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=a ₄× h [i-T ₄]+b ₄(formula 8)

Wherein, T ₄it is the second predetermined period; H [i-T ₄] represent the i-th-T ₄the actual value of it advertisement exposure number; a ₄and b ₄it is the training parameter of the 4th prediction model.

Preferably, described at least two prediction models also comprise: the second prediction model, the 3rd prediction model and holt-winter prediction model; And

In second prediction model the discreet value p [i, w] of advertisement exposure number in i-th day week of w be according to w week before the actual value of the advertisement exposure number of i-th day in several weeks determine; Wherein, in the second prediction model, the discreet value p [i, w] of the advertisement exposure number in i-th day week of w is specially:

$p[i,w]=\frac{\underset{j_2=1}{\overset{k_2}{\mathrm{\Σ}}}g[i,w-j_2]}{k_2}$ (formula 6)

In formula 6, if w-j ₂within i-th day, record the actual value of advertisement exposure number week, then g [i, w-j ₂] represent w-j ₂the actual value of the advertisement exposure number in i-th day week, otherwise g [i, w-j ₂] represent w-j ₂the discreet value of the advertisement exposure number in i-th day week; k ₂for integer, represent the training parameter of the second prediction model; j ₂get 1 to k ₂integer; And

In 3rd prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number of some days before i-th day; Wherein, in the second prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

$p\left[i\right]=\frac{\underset{j_3=1}{\overset{k_3}{\mathrm{\Σ}}}h[i-j_3]}{k_3}$ (formula 7)

In formula 7, if the i-th-j ₃it records the actual value of advertisement exposure number, then h [i-j ₃] represent the i-th-j ₃the actual value of it advertisement exposure number, otherwise h [i-j ₃] represent the i-th-j ₃the discreet value of it advertisement exposure number; k ₃for integer, represent the training parameter of the 3rd prediction model; j ₃get 1 to k ₃integer.

Preferably, for each prediction model of the advertisement exposure number of training, the error rate of estimating of this prediction model is calculated according to following formula 10:

$E=\frac{{\mathrm{\Σ}}_{s\∈S}{\mathrm{\ω}}_s{m}_s}{{\mathrm{\Σ}}_{s\∈S}{\mathrm{\ω}}_s}$ (formula 10)

In formula 10, s represents the number of days of test data set, s ∈ S={s ₁, s ₁+ u, s ₁+ 2u ... .., s _n, s ₁represent first data item in S, s _nrepresent the n-th data item in S, s _n=s ₁+ (n-1) u, u represent and increase progressively item; E be calculate estimate error rate; m _srepresent from l-s days on average estimating error rate in l-1 days; L is the total length of training dataset and test data set; represent i-th day estimate error rate; P [i] is the discreet value of the advertisement exposure number of i-th day obtained based on this prediction model; H [i] is the actual value of the advertisement exposure number of i-th day; ω _srepresent the weight of test data set s; ε is smoothing parameter.

Preferably, described in obtain the discreet value of the advertisement exposure number setting this orientation in prediction time horizon after, also comprise:

According to the discreet value p of following formula 11 to the advertisement exposure number of this orientation in described setting prediction time horizon _forecastcarry out trend correction:

P _adjustment=p _forecast× λ (formula 11)

In formula 11, p _adjustmentrepresent the discreet value of the advertisement exposure number after correcting; λ represents the second correction factor, λ=h _actual/ p _smooth, wherein, p _smoothand h _actualrepresent smooth value and the actual value of the advertisement exposure number of this orientation in described setting historical time section respectively.

Preferably, described at least two prediction models are all trained based on pretreated training data;

Wherein, residing pre-service specifically comprises data cleansing process and data smoothing process;

Described data cleansing process specifically comprises: the directional data cleaning that historical data is very few and unusual fluctuations and the serious data cleansing of incompleteness;

Described data smoothing process specifically comprises: abnormal shake data smoothing process and festivals or holidays abnormal data smoothing processing.

Preferably, the process of described abnormal shake data smoothing is specially:

According to the actual value h [i, w] of the advertisement exposure number in i-th day week of w, calculate local mean values μ, upper dividing value b according to following formula 1,2,3 respectively _uwith floor value b _l:

$\mathrm{\μ}=\frac{1}{4}\left(h\right[i,w-2]+h[i,w-1]+h[i,w+1]+h[i,w+2\left]\right)$ (formula 1)

B _u=max{1.65 μ, h [i, w-2], h [i, w-1], h [i, w+1], h [i, w+2] } (formula 2)

B _l=min{0.66 μ, h [i, w-2], h [i, w-1], h [i, w+1], h [i, w+2] } (formula 3)

According to the advertisement exposure number smoothing process of following formula 4 to i-th day week of w, obtain the smooth value hw [i, w] ' of the advertisement exposure number in i-th day week of w:

$h{[i,w]}^{\&prime;}=\left\{\begin{array}{c}\mathrm{\&mu;},h[i,w]=0\\ b_u,h[i,w]>b_u\\ b_l,h[i,w]<b_1\\ h[i,w,],b_l\&le;h[i,w]\&le;b_u\end{array}\right.$ (formula 4).

The embodiment of the present invention additionally provides the estimating device of the ad click number of times on a kind of advertisement position, comprising:

Prediction model chooses module, for the advertisement exposure according to training in advance at least two prediction models estimate error rate, from described at least two prediction models, select the prediction model estimated for advertisement exposure number;

Advertisement exposure number estimates module, for each orientation for the advertisement that advertisement position to be estimated has been thrown in, use described prediction model to choose the prediction model of the advertisement exposure number that module selects, obtain the discreet value of the advertisement exposure number setting this orientation in prediction time horizon;

Ad click rate determination module, for the actual value sum of advertisement exposure number and the actual value sum of ad click number of times of each orientation in the setting historical time section that basis counts, determines the ad click rate in described setting historical time section;

Ad click number of times estimates module, for the product of the discreet value and described ad click rate that calculate the advertisement exposure number of each orientation respectively, using the discreet value as the ad click number of times in described setting prediction time horizon of each sum of products of calculating.

Wherein, described at least two prediction models comprise: the first prediction model and the 4th prediction model;

In first prediction model and the 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number in the i-th-T sky, and T is predetermined period;

Wherein, in the first prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=growth (i) × seasonality (i) (formula 5)

In formula 5, seasonality (i) is the first correction factor, and t ₁it is the first predetermined period; Growth (i)=a ₁× i+b ₁; a ₁and b ₁it is the training parameter of the first prediction model; H [i-T ₁] be the i-th-T ₁the actual value of it advertisement exposure number; And

In 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=a ₄× h [i-T ₄]+b ₄(formula 8)

Wherein, T ₄it is the second predetermined period; H [i-T ₄] represent the i-th-T ₄the actual value of it advertisement exposure number; a ₄and b ₄it is the training parameter of the 4th prediction model.

Preferably, described at least two prediction models also comprise: the second prediction model, the 3rd prediction model and holt-winter prediction model; And

In second prediction model the discreet value p [i, w] of advertisement exposure number in i-th day week of w be according to w week before the actual value of the advertisement exposure number of i-th day in several weeks determine; Wherein, in the second prediction model, the discreet value p [i, w] of the advertisement exposure number in i-th day week of w is specially:

$p[i,w]=\frac{\underset{j_2=1}{\overset{k_2}{\mathrm{\&Sigma;}}}g[i,w-j_2]}{k_2}$ (formula 6)

In formula 6, if w-j ₂within i-th day, record the actual value of advertisement exposure number week, then g [i, w-j ₂] represent w-j ₂the actual value of the advertisement exposure number in i-th day week, otherwise g [i, w-j ₂] represent w-j ₂the discreet value of the advertisement exposure number in i-th day week; k ₂for integer, represent the training parameter of the second prediction model; j ₂get 1 to k ₂integer; And

In 3rd prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number of some days before i-th day; Wherein, in the second prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

$p\left[i\right]=\frac{\underset{j_3=1}{\overset{k_3}{\mathrm{\&Sigma;}}}h[i-j_3]}{k_3}$ (formula 7)

In formula 7, if the i-th-j ₃it records the actual value of advertisement exposure number, then h [i-j ₃] represent the i-th-j ₃the actual value of it advertisement exposure number, otherwise h [i-j ₃] represent the i-th-j ₃the discreet value of it advertisement exposure number; k ₃for integer, represent the training parameter of the 3rd prediction model; j ₃get 1 to k ₃integer.

Preferably, described prediction model chooses module specifically for each prediction model of advertisement exposure number for training, and what calculate this prediction model according to following formula 10 estimates error rate:

$E=\frac{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s{m}_s}{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s}$ (formula 10)

In formula 10, s represents the number of days of test data set, s ∈ S={s ₁, s ₁+ u, s ₁+ 2u ... .., s _n, s ₁represent first data item in S, s _nrepresent the n-th data item in S, s _n=s ₁+ (n-1) u, u represent and increase progressively item; E be calculate estimate error rate; m _srepresent from l-s days on average estimating error rate in l-1 days; L is the total length of training dataset and test data set; represent i-th day estimate error rate; P [i] is the discreet value of the advertisement exposure number of i-th day obtained based on this prediction model; H [i] is the actual value of the advertisement exposure number of i-th day; ω _srepresent the weight of test data set s; ε is smoothing parameter.

Preferably, described advertisement exposure number estimates module also for each orientation for the advertisement that advertisement position to be estimated has been thrown in, according to the discreet value p of following formula 11 to the advertisement exposure number of this orientation in described setting prediction time horizon _forecastcarry out trend correction:

P _adjustment=p _forecast× λ (formula 11)

In formula 11, p _adjustmentrepresent the discreet value of the advertisement exposure number after correcting; λ represents the second correction factor, λ=h _actual/ p _smooth, wherein, p _smoothand h _actualrepresent smooth value and the actual value of the advertisement exposure number of this orientation in described setting historical time section respectively.

Further, the estimating device of the ad click number of times on described advertisement position, also comprises:

Prediction model training module, for training described at least two prediction models based on pretreated training data; Wherein, residing pre-service specifically comprises data cleansing process and data smoothing process; Described data cleansing process specifically comprises: the directional data cleaning that historical data is very few and unusual fluctuations and the serious data cleansing of incompleteness; Described data smoothing process specifically comprises: abnormal shake data smoothing process and festivals or holidays abnormal data smoothing processing.

In technical scheme of the present invention, set up multiple prediction model that advertisement exposure number is estimated, and select from multiple prediction model and estimate the less prediction model of error rate; Respectively the advertisement exposure number of the difference orientation of advertisement is estimated based on the prediction model selected, and combine the ad click rate counted according to the historical data of advertisement position, thus, obtain the discreet value of the ad click number of times in following a period of time.Thus the present invention improves the accuracy that the ad click number of times on advertisement position is estimated well.

Accompanying drawing explanation

Fig. 1 is the method flow diagram choosing the prediction model estimated for advertisement exposure number of the embodiment of the present invention;

Fig. 2 is the process flow diagram of the predictor method of ad click number of times on the advertisement position of the embodiment of the present invention;

Fig. 3 is the inner structure block diagram of the estimating device of ad click number of times on the advertisement position of the embodiment of the present invention.

Embodiment

For making object of the present invention, technical scheme and advantage clearly understand, enumerate preferred embodiment referring to accompanying drawing, the present invention is described in more detail.But it should be noted that, the many details listed in instructions are only used to make reader to have a thorough understanding, even if do not have these specific details also can realize these aspects of the present invention to one or more aspect of the present invention.

The term such as " module " used in this application, " system " is intended to comprise the entity relevant to computing machine, such as but not limited to hardware, firmware, combination thereof, software or executory software.Such as, module can be, but be not limited in: the thread of the process that processor runs, processor, object, executable program, execution, program and/or computing machine.For example, application program computing equipment run and this computing equipment can be modules.One or more module can be positioned at an executory process and/or thread.

The present inventor considers, the ad click number of times on advertisement position is often relevant with the clicking rate of advertisement with the page browsing amount (PV, Page View) of advertisement place Website page.Wherein, the page browsing amount of Website page can think advertisement exposure number again.Therefore, the present inventor sets up the multiple prediction model estimated advertisement exposure number, and calculate based on the error rate of estimating of historical data to multiple prediction model of advertisement position, select from multiple prediction model and estimate the less prediction model of error rate, to estimate the advertisement exposure number on advertisement position.And, the present inventor also considers, historical data and the advertisement exposure number of the difference orientation of the advertisement on advertisement position often have very big difference, therefore, can estimate different directed advertisement exposure number respectively based on the prediction model selected, and combine the ad click rate counted according to the historical data of advertisement position, obtain the discreet value of the ad click number of times in following a period of time.The model that this ad click number of times selecting applicable advertisement position from the prediction model of multiple advertisement exposure number is estimated, to carry out the mode that advertisement exposure number is estimated respectively to the difference orientation of advertisement, improve the accuracy that the ad click number of times on advertisement position is estimated well.

Technical scheme of the present invention is described in detail below in conjunction with accompanying drawing.

The present inventor through analyzing and researching to the fluctuation etc. of the historical trend of multiple advertisement position, summing up experience etc., establish the prediction model of at least two kinds of advertisement exposure numbers.Further, can calculate the error rate of estimating of prediction model of the advertisement exposure number set up in advance in the present invention, thus select from the prediction model of these advertisement exposure numbers and estimate less one of error rate, for estimating of advertisement exposure number.Particularly, the flow process of the method for the prediction model estimated for advertisement exposure number that what the embodiment of the present invention provided choose, as shown in Figure 1, comprises the steps:

S101: for the advertisement on advertisement position A, carries out pre-service to the actual value of the advertisement exposure number in setting historical time section.

In this step, pre-service is carried out to the actual value of the advertisement exposure number on advertisement position A in setting historical time section and mainly comprises following several aspect:

Data cleansing process on the one hand.Data cleansing process comprises the very few directional data cleaning of historical data and unusual fluctuations and the serious data cleansing of incompleteness.

Historical data very few directional data cleaning be specifically as follows: for each orientation of the advertisement putting on advertisement position A, if determine, the historical record time span of this orientation correspondence is less than setting-up time length (as 2 week), then remove the advertisement exposure number of this orientation.Unusual fluctuations and incompleteness serious data cleaning tool body can be: if the advertisement exposure number of certain day is secondary relative to the increasing of the previous day or amount of decrease is greater than setting changes in amplitude value, and the historical record time span before current time is greater than setting-up time length, then remove the advertisement exposure number before this day and this day; If for certain sky, from this sky, set number of days (as 7 days) continuously occur shortage of data, then remove the last day occurring shortage of data and advertisement exposure number before.Wherein, if the actual value of the advertisement exposure number of certain day is less than 50, think that this day data lacks.

Another side is data smoothing process.Data smoothing process comprise abnormal shake data smoothing process and festivals or holidays abnormal data smoothing processing.

The process of abnormal shake data smoothing is specifically as follows: according to the actual value h [i of the advertisement exposure number in i-th day week of w, w], calculate local mean values (local average) μ, upper dividing value (upper bound) b according to following formula 1,2,3 respectively _uwith floor value (lower bound) b _l:

$\mathrm{\&mu;}=\frac{1}{4}\left(h\right[i,w-2]+h[i,w-1]+h[i,w+1]+h[i,w+2\left]\right)$ (formula 1)

B _u=max{1.65 μ, h [i, w-2], h [i, w-1], h [i, w+1], h [i, w+2] } (formula 2)

B _l=min{0.66 μ, h [i, w-2], h [i, w-1], h [i, w+1], h [i, w+2] } (formula 3).

According to the advertisement exposure number smoothing process of following formula 4 to i-th day week of w, obtain the smooth value hw [i, w] ' of the advertisement exposure number in i-th day week of w:

$h{[i,w]}^{\&prime;}=\left\{\begin{array}{c}\mathrm{\&mu;},h[i,w]=0\\ b_u,h[i,w]>b_u\\ b_l,h[i,w]<b_1\\ h[i,w,],b_l\&le;h[i,w]\&le;b_u\end{array}\right.$ (formula 4).

Festivals or holidays, abnormal data smoothing processing was specifically as follows: for the every day in May Day, ten first-class festivals or holidays, determine outside festivals or holidays, apart from this day nearest forward and backward two weeks of week of place, the actual value of two Zhou Zhongxiang determined advertisement exposure number is on the same day averaged or linear interpolation, as the smooth value of the advertisement exposure number of this day.Such as, for a Wednesday in 11, the actual value of the advertisement exposure number of the rear Wednesday outside the previous Wednesday and 11 outside 11 is averaged or linear interpolation, as the smooth value of the advertisement exposure number of the Wednesday in 11.Like this, the situation that the advertisement exposure of May Day, ten first-class festivals or holidays counts existing unusual fluctuations is taken into account, to the smoothing process of advertisement exposure number comprised in the time period of festivals or holidays, can prevent the prediction model of advertisement exposure number from causing the overfitting of prediction model because using unusual fluctuations data when training.

S102: determine training dataset from the actual value of the advertisement exposure number in pretreated setting historical time section, trains the prediction model of at least two the advertisement exposure numbers set up.

Particularly, the prediction model of at least two advertisement exposure numbers of foundation can comprise: the first prediction model, the second prediction model, the 3rd prediction model, the 4th prediction model and holt-winter prediction model.

Wherein, in the first prediction model and the 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number in the i-th-T sky, and T is predetermined period.

In first prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=growth (i) × seasonality (i) (formula 5)

In formula 5, T ₁being the first predetermined period, can be year, month, day, week etc., and T ₁be less than setting historical time section; Seasonality (i) is the first correction factor, and growth (i)=a ₁× i+b ₁; a ₁and b ₁be the training parameter of the first prediction model, h [i-T ₁] be the i-th-T ₁the actual value of it advertisement exposure number.

Use training data set pair first prediction model carries out training and is exactly: concentrate n-th based on training data ₁actual value h [the n of it advertisement exposure number ₁], determine and make minimum a ₁and b ₁.Wherein, d is total number of days that training data is concentrated.

In 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=a ₄× h [i-T ₄]+b ₄(formula 8)

Wherein, T ₄be the second predetermined period, be generally the fixed cycle of a setting, as 10 days, 20 days etc.; H [i-T ₄] represent the i-th-T ₄the actual value of it advertisement exposure number; a ₄and b ₄it is the training parameter of the 4th prediction model.Like this, the data based on the history previous fixed cycle carry out translation and conversion, can obtain the data in the current fixed cycle.

Use training data set pair the 4th prediction model carries out training and is exactly: concentrate n-th based on training data ₂actual value h [the n of it advertisement exposure number ₂], determine and make objective function $\underset{m=d-T_4}{\overset{d}{\mathrm{\&Sigma;}}}\left|p\right[m]-h[n_2\left]\right|=\underset{m=d-T_4}{\overset{d}{\mathrm{\&Sigma;}}}|a_4\&times;h[n_2-T_4+1]{+b}_4-h[n_2]|$ Minimum a ₄and b ₄.Wherein, d is total number of days that training data is concentrated.

In second prediction model the discreet value p [i, w] of advertisement exposure number in i-th day week of w be according to w week before the actual value of the advertisement exposure number of i-th day in several weeks determine; Wherein, in the second prediction model, the discreet value p [i, w] of the advertisement exposure number in i-th day week of w is specially:

$p[i,w]=\frac{\underset{j_2=1}{\overset{k_2}{\mathrm{\&Sigma;}}}g[i,w-j_2]}{k_2}$ (formula 6)

In formula 6, if w-j ₂within i-th day, record the actual value of advertisement exposure number week, then g [i, w-j ₂] represent w-j ₂the actual value of the advertisement exposure number in i-th day week, otherwise g [i, w-j ₂] represent w-j ₂the discreet value of the advertisement exposure number in i-th day week; k ₂for integer, represent the training parameter of the second prediction model; j ₂get 1 to k ₂integer.

Training data set pair second prediction model is used to carry out training exactly: to determine training parameter k based on training dataset ₂.

In 3rd prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number of some days before i-th day; Wherein, in the second prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

$p\left[i\right]=\frac{\underset{j_3=1}{\overset{k_3}{\mathrm{\&Sigma;}}}h[i-j_3]}{k_3}$ (formula 7)

In formula 7, if the i-th-j ₃it records the actual value of advertisement exposure number, then h [i-j ₃] represent the i-th-j ₃the actual value of it advertisement exposure number, otherwise h [i-j ₃] represent the i-th-j ₃the discreet value of it advertisement exposure number; k ₃for integer, represent the training parameter of the 3rd prediction model; j ₃get 1 to k ₃integer.

Training data set pair the 3rd prediction model is used to carry out training exactly: to determine training parameter k based on training dataset ₃.

Holt-winter prediction model is well known to those skilled in the art, and it can be estimated based on single exponential smoothing or double smoothing.

Wherein, remember that current is i-th day, based on the discreet value F of the advertisement exposure number in the i-th+m sky in the cumulative holt-winter prediction model of double smoothing after i-th day day before yesterday _i+mfor:

$F_{i+m}=s_i+{\mathrm{mt}}_i+q_{i-T_5+\left(m\mathrm{mod}{T}_5\right)}$ (formula 9-1)

In formula 9-1, T ₅for setting seasonal cycle; t _i=β (s _i-s _i-1)+(1-β) t _i-1; x _iit is the actual value of the advertisement exposure number of i-th day; s ₀=x ₀; q ₀=0; t ₀=x ₁-x ₀; α, β, γ are the training parameter of holt-winter prediction model, and value is in [0,1].

The discreet value F of the advertisement exposure number in the i-th+m sky in holt-winter prediction model after i-th day day before yesterday is taken advantage of based on double smoothing tired _i+mfor:

$F_{i+m}=(s_i+{\mathrm{mt}}_i)q_{i-T_5+\left(m\mathrm{mod}{T}_5\right)}$ (formula 9-2)

In formula 9-2, T ₅for setting seasonal cycle; t _i=β (s _i-s _i-1)+(1-β) t _i-1; x _iit is the actual value of the advertisement exposure number of i-th day; s ₀=x ₀; q ₀=1; t ₀=x ₁-x ₀; α, β, γ are the training parameter in the 5th prediction model, and value is in [0,1].

S103: determine test data set from the advertisement exposure number in pretreated setting historical time section, what obtain each prediction model based on test data set estimates error rate.

Particularly, if the test data set s determined gets S set={ s ₁, s ₁+ u, s ₁+ 2u ... .., s _nin element, then that can determine prediction model according to following formula 10 estimates error rate E:

$E=\frac{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s{m}_s}{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s}$ (formula 10)

In formula 10, s represents the number of days of test data set, s ∈ S={s ₁, s ₁+ u, s ₁+ 2u ... .., s _n, s ₁represent first data item in S, s _nrepresent the n-th data item in S, s _n=s ₁+ (n-1) u, u represent and increase progressively item; represent from l-s days on average estimating error rate in l-1 days; L is the total length of training dataset and test data set; represent i-th day estimate error rate; represent the weight of test data set s; ε is smoothing parameter, and default value is 0.01.

S104: estimate error rate according to what obtain, selects the prediction model estimated for advertisement exposure number from the prediction model of at least two the advertisement exposure numbers set up.

Particularly, can choose and estimate the minimum prediction model of error rate, as the prediction model estimated for advertisement exposure number, thus can ensure follow-up ad click number of times estimate accuracy.

Based on the prediction model of the advertisement exposure number selected in advance, the flow process of the predictor method of the ad click number on the advertisement position that the embodiment of the present invention provides, as shown in Figure 2, comprises the steps:

S201: for each orientation of the advertisement that advertisement position to be estimated has been thrown in, uses the prediction model of the advertisement exposure number selected in advance, obtains the discreet value of the advertisement exposure number setting this orientation in prediction time horizon.

Particularly, for each orientation of the advertisement that advertisement position to be estimated has been thrown in, use the prediction model of the advertisement exposure number selected, the advertisement exposure number of this orientation every day in setting prediction time horizon is estimated, using the discreet value sum of the advertisement exposure number of every day as the discreet value of advertisement exposure number setting this orientation in prediction time horizon.Wherein, setting prediction time horizon can be specifically three months.

Further, for each orientation of the advertisement that advertisement position to be estimated has been thrown in, also can according to the discreet value p of following formula 11 to the advertisement exposure number of this orientation in setting prediction time horizon _forecastcarry out trend correction:

P _adjustment=p _forecast× λ (formula 11)

In formula 11, p _adjustmentrepresent the discreet value of the advertisement exposure number after correcting; λ represents the correction factor, λ=h _actual/ p _smooth, wherein, p _smoothand h _actualrepresent smooth value and the actual value of the advertisement exposure number of this orientation in setting historical time section respectively.

Particularly, p is obtained _smoothmethod can be: for the actual value of the advertisement exposure number of this orientation every day in setting historical time section, the method according to above-mentioned formula 1 to 4, obtains the smooth value of the advertisement exposure number of this day; Using the smooth value sum of the advertisement exposure number of this orientation every day in the setting historical time section that obtains as p _smooth.

S202: the actual value sum of advertisement exposure number and the actual value sum of ad click number of times that count each orientation in setting historical time section, and then determine the ad click rate in setting historical time section.

Particularly, for each orientation of advertisement putting, count the actual value sum of advertisement exposure number and the actual value sum of ad click number of times of each orientation in setting historical time section, calculate the ad click rate in setting historical time section according to following formula 12:

$\mathrm{CTR}=\frac{{\mathrm{Click}}_{\mathrm{history}}}{P_{\mathrm{history}}}$ (formula 12)

In formula 12, CTR represents the ad click rate in setting historical time section; Click _historyrepresent the actual value sum of the ad click number of times of each orientation in setting historical time section; p _historyrepresent the actual value sum of the advertisement exposure number of each orientation in setting historical time section.

Further, also can according to following formula 13 to the smoothing process of the ad click rate calculated:

$\mathrm{CTR}=\frac{{\mathrm{Click}}_{\mathrm{history}}+\mathrm{\&alpha;}}{P_{\mathrm{history}}+\mathrm{\&alpha;}+\mathrm{\&beta;}}$ (formula 13)

In formula 13, α and β is the smoothing parameter of the ad click rate of setting; α value is [0,10]; β value is [10000,50000].

S203: the discreet value calculating the advertisement exposure number of each orientation respectively and the product of ad click rate calculated, using the discreet value as the ad click number of times in setting prediction time horizon of each sum of products of calculating.

Based on the predictor method of the ad click number of times on above-mentioned advertisement position, the inner structure block diagram of the estimating device of the ad click number of times on the advertisement position that the embodiment of the present invention provides, as shown in Figure 3, comprising: prediction model chooses module 301, advertisement exposure number estimates module 302, ad click rate determination module 303 and ad click number of times estimate module 304.

Prediction model choose module 301 for the advertisement exposure according to training in advance at least two prediction models estimate error rate, from described at least two prediction models, select the prediction model estimated for advertisement exposure number.At least two prediction models specifically can comprise the first above-mentioned prediction model, the second prediction model, the 3rd prediction model, the 4th prediction model and holt-winter prediction model.Particularly, prediction model chooses each prediction model of advertisement exposure number of module for training, and what can calculate this prediction model according to formula 10 estimates error rate.

Advertisement exposure number estimates module 302 for each orientation for the advertisement that advertisement position to be estimated has been thrown in, use prediction model to choose the prediction model of the advertisement exposure number that module 301 selects, obtain the discreet value of the advertisement exposure number setting this orientation in prediction time horizon.Further, advertisement exposure number estimates module 302 also for each orientation for the advertisement that advertisement position to be estimated has been thrown in, according to the discreet value p of following formula 11 to the advertisement exposure number of this orientation in setting prediction time horizon _forecastcarry out trend correction:

P _adjustment=p _forecast× λ (formula 11)

In formula 11, p _adjustmentrepresent the discreet value of the advertisement exposure number after correcting; λ represents the second correction factor, λ=h _actual/ p _smooth, wherein, p _smoothand h _actualrepresent smooth value and the actual value of the advertisement exposure number of this orientation in setting historical time section respectively.

Ad click rate determination module 303 for counting the actual value sum of advertisement exposure number and the actual value sum of ad click number of times of each orientation in setting historical time section, and then determines the ad click rate in setting historical time section.

Ad click number of times estimates module 304 estimates the advertisement exposure number of each orientation that module 302 is estimated discreet value for calculating targeted ads impression respectively, the product of the ad click rate calculated with ad click rate determination module 303, using the discreet value as the ad click number of times in described setting prediction time horizon of each sum of products of calculating.

Further, the estimating device of above-mentioned ad click number of times also can comprise: prediction model training module 305.

Prediction model training module 305 is for training described at least two prediction models based on pretreated training data; Wherein, described pre-service specifically can comprise above-mentioned data cleansing process and data smoothing process; Data cleansing process specifically comprises: the directional data cleaning that above-mentioned historical data is very few and unusual fluctuations and the serious data cleansing of incompleteness; Data smoothing process specifically comprises: above-mentioned exception shake data smoothing process and festivals or holidays abnormal data smoothing processing.

In technical scheme of the present invention, set up multiple prediction model that advertisement exposure number is estimated, and select from multiple prediction model and estimate the less prediction model of error rate; Respectively the advertisement exposure number of the difference orientation of advertisement is estimated based on the prediction model selected, and combine the ad click rate counted according to the historical data of advertisement position, thus, obtain the discreet value of the ad click number of times in following a period of time.Thus the present invention improves the accuracy that the ad click number of times on advertisement position is estimated well.

One of ordinary skill in the art will appreciate that all or part of step realized in above-described embodiment method is that the hardware that can carry out instruction relevant by program has come, this program can be stored in computer read/write memory medium, as: ROM/RAM, magnetic disc, CD etc.

The above is only the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (13)

1. a predictor method for the ad click number of times on advertisement position, is characterized in that, comprising:

For each orientation of the advertisement that advertisement position to be estimated has been thrown in, use the prediction model of the advertisement exposure number selected in advance, obtain the discreet value of the advertisement exposure number setting this orientation in prediction time horizon; And

According to the actual value sum of advertisement exposure number and the actual value sum of ad click number of times of each orientation in the setting historical time section counted, determine the ad click rate in described setting historical time section;

Calculate the discreet value of the advertisement exposure number of each orientation and the product of described ad click rate respectively, using the discreet value as the ad click number of times in described setting prediction time horizon of each sum of products of calculating;

Wherein, the prediction model of described advertisement exposure number be at least two prediction models of the advertisement exposure number according to training in advance estimate error rate, choose out from described at least two prediction models.

2. the method for claim 1, is characterized in that, described at least two prediction models comprise: the first prediction model and the 4th prediction model; And

In first prediction model and the 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number in the i-th-T sky, and T is predetermined period;

Wherein, in the first prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=growth (i) × seasonality (i) (formula 5)

In formula 5, seasonality (i) is the first correction factor, and t ₁it is the first predetermined period; Growth (i)=a ₁× i+b ₁; a ₁and b ₁it is the training parameter of the first prediction model; H [i-T ₁] be the i-th-T ₁the actual value of it advertisement exposure number; And

In 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=a ₄× h [i-T ₄]+b ₄(formula 8)

Wherein, T ₄it is the second predetermined period; H [i-T ₄] represent the i-th-T ₄the actual value of it advertisement exposure number; a ₄and b ₄it is the training parameter of the 4th prediction model.

3. method as claimed in claim 2, it is characterized in that, described at least two prediction models also comprise: the second prediction model, the 3rd prediction model and holt-winter prediction model; And

In second prediction model the discreet value p [i, w] of advertisement exposure number in i-th day week of w be according to w week before the actual value of the advertisement exposure number of i-th day in several weeks determine; Wherein, in the second prediction model, the discreet value p [i, w] of the advertisement exposure number in i-th day week of w is specially:

$p[i,w]=\frac{\underset{j_2=1}{\overset{k_2}{\mathrm{\&Sigma;}}}g[i,w-j_2]}{k_2}$ (formula 6)

In formula 6, if w-j ₂within i-th day, record the actual value of advertisement exposure number week, then g [i, w-j ₂] represent w-j ₂the actual value of the advertisement exposure number in i-th day week, otherwise g [i, w-j ₂] represent w-j ₂the discreet value of the advertisement exposure number in i-th day week; k ₂for integer, represent the training parameter of the second prediction model; j ₂get 1 to k ₂integer; And

In 3rd prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number of some days before i-th day; Wherein, in the second prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

$p\left[i\right]=\frac{\underset{j_3=1}{\overset{k_3}{\mathrm{\&Sigma;}}}h[i-j_3]}{k_3}$ (formula 7)

In formula 7, if the i-th-j ₃it records the actual value of advertisement exposure number, then h [i-j ₃] represent the i-th-j ₃the actual value of it advertisement exposure number, otherwise h [i-j ₃] represent the i-th-j ₃the discreet value of it advertisement exposure number; k ₃for integer, represent the training parameter of the 3rd prediction model; j ₃get 1 to k ₃integer.

4. the method as described in as arbitrary in claim 1-3, is characterized in that, for each prediction model of the advertisement exposure number of training, the error rate of estimating of this prediction model is calculated according to following formula 10:

$E=\frac{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s{m}_s}{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s}$ (formula 10)

In formula 10, s represents the number of days of test data set, s ∈ S={s ₁, s ₁+ u, s ₁+ 2u ... .., s _n, s ₁represent first data item in S, s _nrepresent the n-th data item in S, s _n=s ₁+ (n-1) u, u represent and increase progressively item; E be calculate estimate error rate; m _srepresent from l-s days on average estimating error rate in l-1 days; L is the total length of training dataset and test data set; represent i-th day estimate error rate; P [i] is the discreet value of the advertisement exposure number of i-th day obtained based on this prediction model; H [i] is the actual value of the advertisement exposure number of i-th day; ω _srepresent the weight of test data set s; ε is smoothing parameter.

5. method as claimed in claim 4, is characterized in that, described in obtain the discreet value of the advertisement exposure number setting this orientation in prediction time horizon after, also comprise:

According to the discreet value p of following formula 11 to the advertisement exposure number of this orientation in described setting prediction time horizon _forecastcarry out trend correction:

P _adjustment=p _forecast× λ (formula 11)

In formula 11, p _adjustmentrepresent the discreet value of the advertisement exposure number after correcting; λ represents the second correction factor, λ=h _actual/ p _smooth, wherein, p _smoothand h _actualrepresent smooth value and the actual value of the advertisement exposure number of this orientation in described setting historical time section respectively.

6. the method as described in as arbitrary in claim 1-3, it is characterized in that, described at least two prediction models are all trained based on pretreated training data;

Wherein, residing pre-service specifically comprises data cleansing process and data smoothing process;

Described data cleansing process specifically comprises: the directional data cleaning that historical data is very few and unusual fluctuations and the serious data cleansing of incompleteness;

Described data smoothing process specifically comprises: abnormal shake data smoothing process and festivals or holidays abnormal data smoothing processing.

7. method as claimed in claim 6, is characterized in that, the process of described abnormal shake data smoothing is specially:

According to the actual value h [i, w] of the advertisement exposure number in i-th day week of w, calculate local mean values μ, upper dividing value b according to following formula 1,2,3 respectively _uwith floor value b _l:

$\mathrm{\&mu;}=\frac{1}{4}\left(h\right[i,w-2]+h[i,w-1]+h[i,w+1]+h[i,w+2\left]\right)$ (formula 1)

B _u=max{1.65 μ, h [i, w-2], h [i, w-1], h [i, w+1], h [i, w+2] } (formula 2)

B _l=min{0.66 μ, h [i, w-2], h [i, w-1], h [i, w+1], h [i, w+2] } (formula 3)

According to the advertisement exposure number smoothing process of following formula 4 to i-th day week of w, obtain the smooth value hw [i, w] ' of the advertisement exposure number in i-th day week of w:

$h{[i,w]}^{\&prime;}=\left\{\begin{array}{c}\mathrm{\&mu;},h[i,w]=0\\ b_u,h[i,w]>b_u\\ b_l,h[i,w]<b_1\\ h[i,w,],b_l\&le;h[i,w]\&le;b_u\end{array}\right.$ (formula 4).

8. the estimating device of ad click number of times on advertisement position, is characterized in that, comprising:

Prediction model chooses module, for the advertisement exposure according to training in advance at least two prediction models estimate error rate, from described at least two prediction models, select the prediction model estimated for advertisement exposure number;

Advertisement exposure number estimates module, for each orientation for the advertisement that advertisement position to be estimated has been thrown in, use described prediction model to choose the prediction model of the advertisement exposure number that module selects, obtain the discreet value of the advertisement exposure number setting this orientation in prediction time horizon;

Ad click rate determination module, for the actual value sum of advertisement exposure number and the actual value sum of ad click number of times of each orientation in the setting historical time section that basis counts, determines the ad click rate in described setting historical time section;

Ad click number of times estimates module, for the product of the discreet value and described ad click rate that calculate the advertisement exposure number of each orientation respectively, using the discreet value as the ad click number of times in described setting prediction time horizon of each sum of products of calculating.

9. device as claimed in claim 8, it is characterized in that, described at least two prediction models comprise: the first prediction model and the 4th prediction model; And

In first prediction model and the 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number in the i-th-T sky, and T is predetermined period;

Wherein, in the first prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=growth (i) × seasonality (i) (formula 5)

In formula 5, seasonality (i) is the first correction factor, and t ₁it is the first predetermined period; Growth (i)=a ₁× i+b ₁; a ₁and b ₁it is the training parameter of the first prediction model; H [i-T ₁] be the i-th-T ₁the actual value of it advertisement exposure number; And

In 4th prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

P [i]=a ₄× h [i-T ₄]+b ₄(formula 8)

Wherein, T ₄it is the second predetermined period; H [i-T ₄] represent the i-th-T ₄the actual value of it advertisement exposure number; a ₄and b ₄it is the training parameter of the 4th prediction model.

10. device as claimed in claim 9, it is characterized in that, described at least two prediction models also comprise: the second prediction model, the 3rd prediction model and holt-winter prediction model; And

In second prediction model the discreet value p [i, w] of advertisement exposure number in i-th day week of w be according to w week before the actual value of the advertisement exposure number of i-th day in several weeks determine; Wherein, in the second prediction model, the discreet value p [i, w] of the advertisement exposure number in i-th day week of w is specially:

$p[i,w]=\frac{\underset{j_2=1}{\overset{k_2}{\mathrm{\&Sigma;}}}g[i,w-j_2]}{k_2}$ (formula 6)

In formula 6, if w-j ₂within i-th day, record the actual value of advertisement exposure number week, then g [i, w-j ₂] represent w-j ₂the actual value of the advertisement exposure number in i-th day week, otherwise g [i, w-j ₂] represent w-j ₂the discreet value of the advertisement exposure number in i-th day week; k ₂for integer, represent the training parameter of the second prediction model; j ₂get 1 to k ₂integer; And

In 3rd prediction model, the discreet value p [i] of advertisement exposure number of i-th day determines according to the actual value of the advertisement exposure number of some days before i-th day; Wherein, in the second prediction model, the discreet value p [i] of advertisement exposure number of i-th day is specially:

$p\left[i\right]=\frac{\underset{j_3=1}{\overset{k_3}{\mathrm{\&Sigma;}}}h[i-j_3]}{k_3}$ (formula 7)

In formula 7, if the i-th-j ₃it records the actual value of advertisement exposure number, then h [i-j ₃] represent the i-th-j ₃the actual value of it advertisement exposure number, otherwise h [i-j ₃] represent the i-th-j ₃the discreet value of it advertisement exposure number; k ₃for integer, represent the training parameter of the 3rd prediction model; j ₃get 1 to k ₃integer.

11. as arbitrary in claim 8-10 as described in device, it is characterized in that,

Described prediction model chooses module specifically for each prediction model of advertisement exposure number for training, and what calculate this prediction model according to following formula 10 estimates error rate:

$E=\frac{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s{m}_s}{{\mathrm{\&Sigma;}}_{s\&Element;S}{\mathrm{\&omega;}}_s}$ (formula 10)

In formula 10, s represents the number of days of test data set, s ∈ S={s ₁, s ₁+ u, s ₁+ 2u ... .., s _n, s ₁represent first data item in S, s _nrepresent the n-th data item in S, s _n=s ₁+ (n-1) u, u represent and increase progressively item; E be calculate estimate error rate; m _srepresent from l-s days on average estimating error rate in l-1 days; L is the total length of training dataset and test data set; represent i-th day estimate error rate; P [i] is the discreet value of the advertisement exposure number of i-th day obtained based on this prediction model; H [i] is the actual value of the advertisement exposure number of i-th day; ω _srepresent the weight of test data set s; ε is smoothing parameter.

12. as arbitrary in claim 8-10 as described in device, it is characterized in that,

Described advertisement exposure number estimates module also for each orientation for the advertisement that advertisement position to be estimated has been thrown in, according to the discreet value p of following formula 11 to the advertisement exposure number of this orientation in described setting prediction time horizon _forecastcarry out trend correction:

P _adjustment=p _forecast× λ (formula 11)

In formula 11, p _adjustmentrepresent the discreet value of the advertisement exposure number after correcting; λ represents the second correction factor, λ=h _actual/ p _smooth, wherein, p _smoothand h _actualrepresent smooth value and the actual value of the advertisement exposure number of this orientation in described setting historical time section respectively.

13. as arbitrary in claim 8-10 as described in device, it is characterized in that, also comprise:

Prediction model training module, for training described at least two prediction models based on pretreated training data; Wherein, residing pre-service specifically comprises data cleansing process and data smoothing process; Described data cleansing process specifically comprises: the directional data cleaning that historical data is very few and unusual fluctuations and the serious data cleansing of incompleteness; Described data smoothing process specifically comprises: abnormal shake data smoothing process and festivals or holidays abnormal data smoothing processing.