CN105897616A - Resource allocation method and server - Google Patents

Abstract

The embodiment of the invention discloses a resource allocation method. The method comprises the following steps: acquiring first history data, and performing modeling according to the first history data to obtain a first pre-estimation model, wherein the first pre-estimation model is used for representing a prediction relationship between a feedback result obtained by means of providing a service for a terminal through a server and resources, and the first history data are all data relevant to the building of the first pre-estimation model; acquiring second history data, and continuously detecting whether a data output result obtained on the basis of a second pre-estimation model satisfies a preset strategy or not in a process of performing modeling according to the second history data and the first pre-estimation model to obtain the second pre-estimation model, wherein the second history data are recent incremental data relevant to the building of the second pre-estimation model; stopping modeling of the second pre-estimation model till a data output result obtained on the basis of the second pre-estimation model is detected to satisfy the preset strategy in order to obtain a second model being actually applied to resource proportion; and performing processing of an actual resource proportion according to the second model. The embodiment of the invention further provides a server.

Description

The method of a kind of resource distribution and server

Technical field

The present invention relates to the network operation technology in the communications field, particularly relate to method and the server of the distribution of a kind of resource.

Background technology

Along with developing rapidly of network, intelligent terminal's is a large amount of universal, and user uses network or the upper various application installed of intelligent terminal can realize various service for life, such as, handing over water/electricity/gas charge on the net, online is played games, and various consumption and the entertainment service such as video are seen in online.Different platforms can provide the user different COSs, it is also possible to is supplied to user by a multiple COS of Platform integration.But, the impossible various COS of an unbounded quantity of expansion of platform, therefore, the resource of platform is limited, platform and intelligent terminal by network carry out information mutual time, Internet resources are also limited, how the service that platform provides can be adapted with its resource proportioning, are intended to solve the technical problem that to reaching optimal operational efficiency.And want the service analyzing platform offer whether to adapt with its resource proportioning, need collect server be Terminal for service obtain feedback result (as return or income) for analyzing, that is: server is that the feedback result (such as return or income) that Terminal for service obtains is analyzed just can obtaining desired result with the relation of resource by platform, make platform reach optimal operational efficiency, obtain optimal feedback result (such as return or income) with minimum operation cost simultaneously.And in prior art, in order to solve this technical problem, it is provided with a Utopian premise: server is that the feedback result that Terminal for service obtains is directly proportional to resource, in actual application, the most not always direct ratio, thus, use prior art, result accurately cannot be obtained, platform also certainly will cannot be made to reach optimal operational efficiency.

Summary of the invention

For solving above-mentioned technical problem, embodiment of the present invention expectation provides method and the server of a kind of resource distribution, it is possible to determine the optimal resource allocation mode of each service product on platform more accurately, and the summation of the income reaching platform maximizes.

The technical scheme is that and be achieved in that:

The method embodiments providing the distribution of a kind of resource, including:

Obtain the first historical data, the first prediction model is obtained according to described first historical data modeling, described first prediction model is for characterizing the projected relationship that server is the feedback result that obtains of Terminal for service and resource, and described first historical data is to build relevant full dose data to the first prediction model；

Obtain the second historical data, during obtaining the second prediction model according to described second historical data and described first prediction model modeling, persistently detect whether the data output result obtained based on described second prediction model meets preset strategy, described preset strategy is when sign exports application of results in resource proportioning according to the data that described second prediction model obtains, the corresponding described server obtained is that the feedback result that Terminal for service obtains is higher than history threshold value, and described second historical data is to build relevant recent incremental data to the second prediction model；

Until detecting when the data output result obtained based on described second prediction model meets preset strategy, stopping the modeling to the second prediction model, obtaining practice in the second model of described resource proportioning；

The process of real resource proportioning is carried out according to described second model.

In such scheme, the described resource proportioning process carrying out reality according to described second model, including:

Obtain pending target data, according to described second model, described pending target data is carried out computing, obtain real resource proportioning.

In such scheme, described acquisition the second historical data, the process of the second prediction model is obtained according to described second historical data and described first prediction model modeling, including:

Obtain the history desired value of reach the standard grade natural law and the service product of the attribute of service product corresponding to described pending target data, service product, attribute according to described service product, the history desired value of reach the standard grade natural law and described service product of described service product and default Unknown weights value vector, modeling obtains the 3rd prediction model, and described 3rd prediction model is for characterizing the projected relationship of described resource proportioning and resource；

Obtain the history total revenue of service product, obtain described second prediction model according to the history total revenue of described service product, described 3rd prediction model and described first prediction model modeling.

In such scheme, after the described process carrying out real resource proportioning according to described second model, described method also includes:

After Preset Time, the second historical data after obtaining the first historical data after updating and updating, according to the first historical data after described renewal and the second historical data after described renewal, described second model is adjusted；

The process of real resource proportioning is carried out according to described second model after adjusting.

In such scheme, described acquisition the first historical data, obtain the first prediction model according to described first historical data modeling, including:

Obtain the history desired value of reach the standard grade natural law and the described service product of the attribute of described service product corresponding to described pending target data, described service product, obtain described first prediction model according to the history desired value amount modeling of the attribute of described service product corresponding to described pending target data, reach the standard grade natural law and the described service product of described service product.

The embodiment of the present invention additionally provides a kind of server, including:

Acquiring unit, for obtaining the first historical data, described first historical data is to build relevant full dose data to the first prediction model；

Modeling unit, obtains the first prediction model for the described first historical data modeling obtained according to acquiring unit, and described first prediction model is for characterizing the projected relationship that server is the feedback result that obtains of Terminal for service and resource；

Described acquiring unit, is additionally operable to obtain the second historical data, and described second historical data is to build relevant recent incremental data to the second prediction model；

Detector unit, it is additionally operable to described second historical data that described modeling unit obtains according to described acquiring unit and during described first prediction model modeling obtains the second prediction model, persistently detect whether the data output result that described second prediction model set up based on described modeling unit obtains meets preset strategy, when described preset strategy is for characterizing the data output application of results obtained according to described second prediction model in resource proportioning, the corresponding described server that obtains is that the feedback result that obtains of Terminal for service is higher than history threshold value；

Stop element, for, during until described detector unit detects that the data output result obtained based on described second prediction model meets preset strategy, stopping the modeling to the second prediction model, and,

Described modeling unit, is additionally operable to obtain practice in the second model of described resource proportioning；

Resource processing unit, described second model for obtaining according to described modeling unit carries out the process of real resource proportioning.

In above-mentioned service, described acquiring unit, it is additionally operable to obtain pending target data；

Resource processing unit, the described pending target data that described acquiring unit is obtained by described second model specifically for obtaining according to described modeling unit carries out computing, obtains real resource proportioning.

In above-mentioned server, described acquiring unit, specifically for obtaining the history desired value of reach the standard grade natural law and the service product of the attribute of service product corresponding to described pending target data, service product；

Described modeling unit, it is additionally operable to history desired value and the default Unknown weights value vector of reach the standard grade natural law and the described service product of the attribute of described service product obtained according to described acquiring unit, described service product, modeling obtains the 3rd prediction model, and described 3rd prediction model is for characterizing the projected relationship of described resource proportioning and resource；

Described acquiring unit, is additionally operable to obtain the history total revenue of service product；

Described modeling unit, the history total revenue of the described service product specifically for obtaining according to described acquiring unit, described 3rd prediction model and described first prediction model modeling obtain described second prediction model.

In above-mentioned server, described server also includes: adjustment unit；

Described acquiring unit, the second historical data after being additionally operable to the process that described resource processing unit carries out real resource proportioning according to described second model, after Preset Time, after obtaining the first historical data after updating and updating；

Described second model that described modeling unit is obtained by described adjustment unit, the first historical data after the described renewal obtained according to described acquiring unit and the second historical data after described renewal is adjusted；

Described resource processing unit, is additionally operable to described second model after adjusting according to described adjustment unit and carries out the process of real resource proportioning.

In above-mentioned server, described acquiring unit, specifically for obtaining the history desired value of reach the standard grade natural law and the described service product of the attribute of described service product corresponding to described pending target data, described service product；

Described modeling unit, the attribute of described service product corresponding to described pending target data specifically for obtaining according to described acquiring unit, the history desired value modeling of reach the standard grade natural law and described service product of described service product obtain described first prediction model.

Embodiments provide method and the server of the distribution of a kind of resource, by obtaining the first historical data, the first prediction model is obtained according to the first historical data modeling, first prediction model is for characterizing the projected relationship that server is the feedback result that obtains of Terminal for service and resource, and this first historical data is to build relevant full dose data to the first prediction model；Obtain the second historical data, model during obtaining the second prediction model according to the second historical data and the first prediction model, persistently detect whether the data output result obtained based on the second prediction model meets preset strategy, preset strategy is when sign exports application of results in resource proportioning according to the data that the second prediction model obtains, the corresponding server obtained is that the feedback result that Terminal for service obtains is higher than history threshold value, and this second historical data is to build relevant recent incremental data to the second prediction model；Until detecting when the data output result obtained based on the second prediction model meets preset strategy, stopping the modeling to the second prediction model, obtaining practice in the second model of resource proportioning；The process of real resource proportioning is carried out according to the second model.Use above-mentioned technic relization scheme, owing to historical data based on user behavior based on service product is foundation, by setting up the projected relationship of feedback result and resource, and projected relationship of based on feedback result Yu resource constructs about resource allocation decisions model, resource proportioning optimal for individual service product on platform is found by model training, determine the optimal resource allocation mode of each service product on platform the most accurately, and the summation of the income using best resource proportioning to reach platform is maximized.

Accompanying drawing explanation

Figure 1Illustrate for service operation of the prior artFigure；

Figure 2For the embodiment of the present invention carries out the signal of the mutual various hardware entities of informationFigure；

Figure 3Flow process for a kind of resource allocation methods that the embodiment of the present invention providesFigure one；

Figure 4Structural representation for a kind of exemplary neutral net that the embodiment of the present invention providesFigure；

Figure 5Flow process for a kind of resource allocation methods that the embodiment of the present invention providesFigure two；

Figure 6Flow process for a kind of resource allocation methods that the embodiment of the present invention providesFigure three；

Figure 7Structural representation for a kind of server that the embodiment of the present invention providesFigure one；

Figure 8Structural representation for a kind of server that the embodiment of the present invention providesFigure two；

Figure 9Structural representation for a kind of server that the embodiment of the present invention providesFigure three。

Detailed description of the invention

Below in conjunction with in the embodiment of the present inventionAccompanying drawing, the technical scheme in the embodiment of the present invention is clearly and completely described.

First, following term referred to herein is described as follows:

1) long tail vat: refer to remove the part investigated beyond pond in total resources, account for the major part in total resources, for the operation of selecting business product, to obtain income.Because typically the service product number in this pond is more, the income of wherein most service product typically presents long-tail phenomenon, therefore by business convention, this part resource is referred to as long tail vat.

2) investigate pond: refer to the small portion opened up in total resources, for checking the situation of Profit of new business product, if satisfied to it, be then placed in long tail vat.

3) resource: refer to, for online service product situation, refer mainly to flow (pv), light exposure etc..It should be noted that in special time period, on platform, operational stock number is changeless, the competition of multiple service products obtains this limited resource.Service product refers to that the various game such as applying (platform) on treasured to issue, advertisement, software etc. are applied.Platform refers to the gathering ground of multinomial product, the QQ of such as Tengxun, application treasured etc..

For service product and platform, it should be noted that, platform is the interaction platform between server and terminal, server passes through platform issuing service product, and terminal is exposed by this platform or shows this service product of application, as a example by applying treasured, application is precious as a platform, incorporating various types of third-party application, these third-party application are exactly service product, and these service products provided based on platform can provide the user various service.

4) income: refer to money income or other commerciality target.Such as: the online advertisement income in a period of time, the registration number of users of certain service product, active users, user accepts the number of times of personalized recommendation result, paying ratio in certain game registration user, the exposure clicking rate (in the most every thousand pv, user clicks on the number of times of download) of paying customer etc. can be for income.

nullExemplary，Server is issued by platform or has reached the standard grade a new game，Terminal by platform it can be seen that this new game，And then be only possible to play this game on the terminal，But，Flow space or light exposure that service product on platform takies are fixing，When a new game is reached the standard grade，Certainly will be for resources such as its distribution light exposures，So that exposing (display) in terminal by platform，Therefore，Server will carry out rational resource and redistribute，And user terminal is played or uses the game that this is new，It will be carried out clicking trigger (exposure clicking rate) or carry out user's registration etc.，Due to when the exposure clicking rate of a application or user's registration amount are the highest，Investment can be attracted、Advertising input and popularity etc.，Finally produce physical assets or intangible asset that publisher wants，Therefore，These exposure clicking rates or user's registration amount are exactly the income of this trendy game.

It should be noted that use the method that automatically and manually two ways carries out resource distribution.The method using manual mode to carry out resource distribution is:Such as figure 1Shown in, according to new business product in the head performance of n days (service product index), determine after being analyzed by business personnel that this new business product is to be eliminated, or the long tail vat of selected entrance is distributed for its resource carrying out every day by business personnel together with existing service product.Wherein, the resource allocation rule of all service products in long tail vat, Main Basis is in the business experience of business personnel, such as, the service product that those earnings potentials are the biggest is the biggest for its stock number distributed.The method using automated manner to carry out resource distribution is: assuming that the stock number that income obtains to service product is directly proportional, thus the problem that resource is distributed can be converted into a linear optimization problem, server is by solving linear optimization problem, obtain the preferred plan of the resource distribution of each service product, so that the total revenue of platform maximizes.

But, during using prior art to be service product Resources allocation, owing to manual allocation mode is carried out resource distribution by business personnel according to business experience, the subjective initiative of people is in the highest flight, therefore, it is difficulty with the preferred plan of service product resource distribution, thus the total revenue affecting platform maximizes；And owing to the active method of salary distribution is to be directly proportional the stock number that income obtains to service product to realize, and in reality application, the stock number that income on platform obtains to service product is not necessarily and is directly proportional, therefore, when server is by above-mentioned active method of salary distribution Resources allocation, have impact on the accuracy of the optimum scheme choice of the resource distribution of each service product, the income causing platform to obtain is not necessarily maximum, and the total revenue i.e. affecting platform maximizes.

As Figure 2Shown in, for the signal of the mutual various hardware entities of information in the embodiment of the present inventionFigure,Figure 2Including: one or more servers 41～4n, terminal unit 21-25 and network 31, network 31 includes router, gateway etc. network entity,In figureDo not embody.It is mutual that terminal unit 21-25 carries out service product information by cable network or wireless network and server, in order to produces the index of service product from terminal 21-25 by user behavior and transmits to server.The type of terminal unitSuch as figure 2Shown in, including the type such as mobile phone (terminal 23), panel computer or PDA (terminal 25), desktop computer (terminal 22), PC (terminal 24), all-in-one (terminal 21).Wherein, terminal unit is provided with the service product needed for various user, such as possesses the application of amusement function (such as Video Applications, audio frequency plays application, game application, ocr software), and for example possess the application (such as digital map navigation application, purchasing by group application, camera application etc.) of service function.

Based on above-mentionedFigure 2Shown system, as a example by application needed for user is for game application, server carries out the issue of service product by platform, terminal is gone forward side by side to exercise by the online service product of Network Capture and is used, user uses the historical datas such as generation index based on user behavior and transmits to server service product, and this server statistics obtains the historical datas such as the online natural law of these historical datas and service product and proposes a kind of resource allocation methods.The resource allocation methods that the embodiment of the present invention provides is foundation due to historical data based on user behavior based on service product, by setting up the projected relationship of feedback result and resource, and projected relationship of based on feedback result Yu resource constructs about resource allocation decisions model, resource proportioning optimal for individual service product on platform is found by model training, determine the optimal resource allocation mode of each service product on platform the most accurately, and the summation of the income using best resource proportioning to reach platform is maximized.

Above-mentionedFigure 2Example simply realize a system architecture example of the embodiment of the present invention, the embodiment of the present invention is not limited to above-mentionedFigure 2Described system structure, based on this system architecture, proposes each embodiment of the present invention.

Embodiment one

Embodiments provide a kind of resource allocation methods,Such as figure 3Shown in, the method may include that

S101, obtaining the first historical data, obtain the first prediction model according to the modeling of this first historical data, this first prediction model is for characterizing the projected relationship that server is the feedback result that obtains of Terminal for service and resource.

Here, the first historical data is full dose data, is server for the total data content collected by data analysis, specifically for characterizing when service product is accessed self data and the access data of user feedback.

It should be noted that the method that a kind of resource that the embodiment of the present invention provides is distributed, the method being primarily directed on the platform that server provides the resource distribution of the service product issued.

It is understandable that, service product owing to being issued by the platform of server is also more and more, at present, on platform, all there is substantial amounts of service product in every day, owing to user uses these service products to make these service products be that platform brings income, therefore, operator gives each service product by reasonable distribution resource, reaches the purpose of maximum revenue.What the embodiment of the present invention provided is exactly a kind of scheme carrying out rational resource distribution for service product.

The design philosophy of the embodiment of the present invention is mainly: server designs a resource allocation algorithm according to the historical data that the service product of record on platform is corresponding, then this resource allocation algorithm is applied in follow-up resource distribution, so that platform maximum revenue.

Concrete, server can obtain for characterizing the first prediction model that server is the feedback result that obtains of Terminal for service and the projected relationship of resource by the first historical data modeling of platform record, i.e. in the embodiment of the present invention, for the relation of pre-estimation income Yu resource, this server is modeled by the historical data of platform and realizes the first pre-estimation model.

It is understood that the feedback result of embodiment of the present invention terminal can be income.

It should be noted that the first historical data in the embodiment of the present invention refers to the attribute of service product, the natural law of reaching the standard grade (entering the natural law of long tail vat) of service product and the history desired value etc. of service product.Wherein, the history desired value of the service product in the embodiment of the present invention refers to light exposure, download, registration amount, logs in number, paying number, income etc..

Further, before server obtains the first historical data, this server wants advanced row Data Preparation, i.e. records each service product in long tail vat every day: product category, by the end of the index based on user behavior such as days running (being called for short " natural law "), desired value on the same day (such as light exposure, download, registration amount, logging in number, paying number, income) in long tail vat on the same day.So, when server sets up the first prediction model, the first historical data of all of online service product just can be added up or get to this server.

It should be noted that the first historical data in the embodiment of the present invention refer to all online service products modeling the same day before historical data.The history desired value of the service product in the embodiment of the present invention refers in the desired value (i.e. Recent data) modeled in n days before the same day.

Exemplary, server obtains the first prediction model characterizing resource E with the projected relationship of income r, such as, r=f (C, D, P, E) according to the attribute C of service product, the natural law D that reaches the standard grade of service product, history desired value P of service product, modeling.

It should be noted that in embodiments of the present invention, C, D, P can be known, therefore, income r is solely dependent upon E, i.e. change the resource allocation of every day, will change the income of every day.

It should be noted that the first prediction model in the embodiment of the present invention is nonlinear, it is preferred that the first prediction model can be neural network model.This is owing to neural network model is output as successive value, and so, the optional scope using the optimized algorithm type selecting of the follow-up total revenue of neural network model is bigger, and it is that centrifugal pump situation is more accurate that result also exports result than model.Wherein, the Hidden unit number in neural network model can be determined by experiment.

Concrete, it is assumed that first historical data of the service product A that server obtains is exemplarySuch as table 1Shown in, the attribute C of service product_ij, the natural law D that reaches the standard grade of service product_ij, history desired value P of service product_inj, modeling obtain characterize resource E_iWith income r_i.Wherein, C_ijRepresent the classification of jth sky i-th product, D_ijRepresent the natural law entering long tail vat of jth sky i-th product, P_injRepresent the jth-1 desired value to the i-th product in jth-n sky.When assuming n=14, it is assumed that the history desired value of service product is 6 indexs, then P_injTake 84 fields, C_ijTake 1 field, D_ijTake 1 field, E_iTaking 1 field, the input variable of the first prediction model is 84+1+1+1=87 (x=(C_ij, D_ij, P_inj, E_i)), output variable is 1 y=(r_i)。

Table 1

Therefore, 3 layers of neural network model, its structure are usedSuch as figure 4Shown in.Here hidden layer (intermediate layer) output valve h₁..., h_pComputing formula be formula (1):

$h_b=f({\mathrm{\Σ}}_{b=1}^{m+1}{w}_{ab}*x_a)---\left(1\right)$

Wherein, b=1 ..., q, x_m+1=1 (bias term), f (x)=1/ (1+e^-x), w_abIt is the connection weights between a node of input layer and the b node of hidden layer.

Output layer y is calculated as formula (2):

$y={\mathrm{\Σ}}_{j=1}^p{\mathrm{\β}}_j*h_j---\left(2\right)$

Wherein, wherein β_bIt is the connection weights between b node of hidden layer and output layer kth node.

It should be noted that the solution procedure of model parameter w (matrix), β (vectorial) is as follows:

W (in neutral net connection weights) between input layer and hidden layer: according to the principle of extreme learning machine, w can be taken as arbitrary random value, such as, be positioned at the random value between [-1,1].And once assignment, the most no longer change during follow-up model optimization.Therefore, the input variable of tube model does not has how many, does not affects the total number of adjustable parameter of model.

Hyper parameter p: be the number of hidden nodes of neutral net, is also the unique hyper parameter in a whole set of algorithm, obtains too small or excessive, may cause poor fitting or over-fitting, can only be determined by experiment an optimum.

Model training seeks β (in neutral net connection weights) between hidden layer and output layer: according to the principle of extreme learning machine, is attributed to and asks a Moore-Penrose generalized inverse just may be used.

In a word, after trying to achieve β, neural network model (the first prediction model) is just set up.The most given (C_ij, D_ij, P_inj, E_i) value, just can try to achieve r_i?.

It should be noted that the pre-side relation of one service product total revenue of a day and resource during neural network model, the relation summation of the income of all online natural law of all products with resource is just obtained the present invention is the first prediction model in embodiment.

S102, obtain the second historical data, model during obtaining the second prediction model according to this second historical data and the first prediction model, persistently detect whether the data output result obtained based on this second prediction model meets preset strategy, when this preset strategy is for characterizing the data output application of results obtained according to this second prediction model in resource proportioning, the corresponding server that obtains is that the feedback result that obtains of Terminal for service is higher than history threshold value.

Here, second historical data is incremental data, it is server Recent data content in the total data content collected by data analysis, the access data of the data of self, user feedback when being accessed specifically for service product, it is also possible to include according to income produced by these data.

nullPopular says，Model based on the second historical data and the first prediction model during obtaining the second prediction model，Second historical data and the first prediction model can be as known parameters，The most accurate in order to obtain、Final actual the second prediction model used，Also need to persistently detect the data output result obtained based on this second prediction model (distinguish to export result with follow-up data，Here it is indicated with the first data output result)，These first data output result can be understood as compensating parameter，By in this compensation parameter feed back input to the second prediction model obtained based on above-mentioned known parameter，Continue to monitor the data output result obtained based on this second prediction model (to be indicated with the second data output result here，Second data output result is different from the first data output result)，Preset strategy is met when the second data export result，That is: these second data are exported application of results when resource proportioning，The corresponding server obtained is that the feedback result that Terminal for service obtains is higher than history threshold value，So，The training process to the second prediction model can be stopped，The second prediction model now obtained is final actual the second prediction model used.

Exemplary, in the embodiment of the present inventionTable 1Shown in the attribute C of service product_ij, the natural law D that reaches the standard grade of service product_ij, history desired value P of service product_injIt is the first historical data, according to the attribute C of service product_ij, the natural law D that reaches the standard grade of service product_ij, history desired value P of service product_injAnd resource E_iFor the number of shared field, set up the first prediction model according to formula (1)；Second historical data is the attribute C of corresponding service product_ij, the natural law D that reaches the standard grade of service product_ij, history desired value P of service product_inj, resource E_iAnd the concrete data of the history total revenue (the actual true total revenue of history) that the service product of server acquisition is before modeling, this server is according to the second historical data, service product in the longest tail vat is redistributed resource ratio, wish that ratio R evRise of history " income " and the history true earning obtained can reach maximum, i.e. formula (3):

$\max imize:\mathrm{Re}vRise=\frac{{\Σ}_{j=1}^N{\Σ}_{i=1}^{M_j}f(C_{ij},D_{ij},P_{inj},T_{j*g}(C_{ij},D_{ij},P_{i14j},W\left)\right)}{R_{real}}---\left(3\right)$

Wherein, N represents those all skies having statistical index data by the end of modeling day in history, M_jRepresent the product number in the long tail vat in jth sky, C_ijRepresent the classification of jth sky i-th product, D_ijRepresent the natural law entering long tail vat of jth sky i-th product, P_injRepresent the jth-1 desired value R to the i-th product in jth-n sky_realFor the true total revenue of history.Meanwhile, the resource proportioning sum that be also satisfied all service products to be the condition of 1, i.e. formula (4).

$subjectto:{\mathrm{\Σ}}_{i=1}^{M_j}{e}_i\≡{\mathrm{\Σ}}_{i=1}^{M_j}g(C_{ij},D_{ij},P_{inj},W)=1,j=1,...,N---\left(4\right)$

It is to say, server needs to solve an optimization problem finding a function maximum with N number of equality constraint, i.e. need to find most suitable W (data parameters to be asked) so that RevRise is maximum.

For the ease of modeling and optimization, function g () can be rewritten, such as formula (5):

$g(C_{ij},D_{ij},P_{i14j},W)=\frac{\exp \[h(C_{ij},D_{ij},P_{inj},W)\]}{{\Σ}_{i=1}^{M_j}\exp \[h(C_{ij},D_{ij},P_{inj},W)\]}---\left(5\right)$

The most revised formula (5), meets the constraints of formula (4), and therefore, formula (3) is finally rewritten as formula (6):

$Minimize:S\left(W\right)\≡R_{\mathrm{Re}al}/{\mathrm{\Σ}}_{j=1}^{N}R(j,W)---\left(6\right)$

It is to say, server is according to the second historical data, Unknown weights value vector (data parameters to be asked) and the first prediction model and the 3rd prediction model, model the second prediction model.

As from the foregoing, the W obtained according to above-mentioned formula (5) is exactly the first data output result, wherein, W is the compensation parameter of resource proportioning, server is through persistently detecting, detecting and meet preset strategy, namely output result W of the condition of formula (6) is exactly the second data output result.

It should be noted that, the method of the resource distribution that the embodiment of the present invention provides is find so that re-starting when resource is distributed higher than the income of resource distribution mode before by the purpose setting up income and resource allocation relationship, when optimal acquisition exactly enables to carry out resource distribution according to the resource proportioning after adjusting, income total revenue truer than history when not adjusting before (history threshold value) of follow-up generation is high to maximum.In server pre-estimation after income and the relation of resource, this server is just according to the second historical data of the true total revenue of history comprising service product, set up second prediction model with the variable (data output result) relevant to resource, whether to reach pre-estimation income and determine final actual the second prediction model used higher than the true total revenue of history maximum (preset strategy) by detecting the second prediction model.That is, output result in second prediction model is the variable relevant to resource, by constantly test and training, have found the output result that disclosure satisfy that optimal provisional profit, the resource proportioning maximum so that provisional profit just can be obtained by this output result, server carries out the resource distribution of reality according to this resource proportioning, just can make the Income Maximum of the service product after adjusting.

Exemplary, preset the relational expression of resource proportioning and variable W, set up due to the first prediction model is the relation of provisional profit and resource, therefore, it is the resource relevant to W and provisional profit according to the input in the second prediction model that the true total revenue of the first prediction model and history is set up, the solving result of the W of output, solving for W is exactly the training process of the second prediction model, ceaselessly adjust the second prediction model, the when of until output result reaches preset strategy, now this second prediction model is finalized, can be as final actual the second prediction model used.After solving W, it is possible to solve the resource proportioning meeting requirement.Owing to resource is dissolved as the relational expression about W by the embodiment of the present invention, therefore, set up by the second prediction model is the model of W, at this moment, it is desirable to reach preset model being necessary for meeting the server enabling to obtain by the output resource proportioning that obtains of result is that the feedback result that Terminal for service obtains is the most permissible higher than history threshold value.

Therefore, after server sets up the first prediction model, if wanting the Income Maximum of prediction, the most also need with the actual true total revenue of history compares, then, this server obtains the second historical data (including the true total revenue of history), owing to being that server can not directly be known according to the stock number in the first prediction model, therefore, this server models in the second prediction model obtained still containing needing to seek data parameters (i.e. data output result) according to the second historical data and the first prediction model, this data parameters to be asked is to solve the unknown parameter of the stock number in the embodiment of the present invention.Owing to server wants to predict the target of a maximum revenue, therefore, server needs constantly to carry out detecting whether data output result (data parameters to be asked) obtained based on this second prediction model meet preset strategy, this preset strategy (utilizes data output result can try to achieve resource proportioning) when sign exports application of results in resource proportioning according to the data that this second prediction model obtains, the corresponding server obtained is that the feedback result that Terminal for service obtains (tries to achieve the resource of service product higher than history threshold value by resource proportioning by platform, never feedback result (income) is tried to achieve), owing to the second prediction model is a model continuing to optimize training, server by continuing to optimize the second prediction model until finding the model of data output Income Maximum corresponding to result.

It should be noted that, history threshold value in the embodiment of the present invention refers to the true total revenue of history that server gets, the income that predicts of prediction model provided of course through the embodiment of the present invention is higher than history threshold value more good, and the data output result selecting one prediction income higher than history threshold value corresponding just can obtain the resource proportioning of optimum reality.

S103, until detecting when the data output result obtained based on the second prediction model meets preset strategy, stops the modeling to the second prediction model, obtains practice in the second model of resource proportioning.

When server detects that the data output result obtained based on the second prediction model meets preset strategy, illustrate that data output result at this moment is exactly the optimal solution that we need, therefore, this server stops the optimization to the second prediction model, determining real the second model being used for solving resource proportioning below, this second model is just off the second prediction model when optimizing.

It is understood that owing to the second model is the model about data parameters to be asked, determine when server and meet when data parameters solution is sought in preferential treatment of preset strategy, this server just can solve resource proportioning according to data parameters to be asked.So, server just can find out the optimal resource allocation ratio of service product by prediction model.

S104, carry out the process of real resource proportioning according to the second model.

Server obtains practice after the second model of resource proportioning, and this server just can be by the relevant treatment of this second model realization real resource proportioning.

Concrete, after server obtains real resource proportioning, service product is carried out according to each self-corresponding resource proportioning the distribution of resource, owing to this real resource proportioning is to meet the proportioning that preset strategy is optimum, therefore, after service product carries out resource distribution according to above-mentioned resource proportioning, income just can realize maximising.That is, the embodiment of the present invention is the history achievement data by utilizing each service product in long tail vat, set up the relation (non-linear relation) between product income and the stock number obtaining distribution and determine the formula of stock number ratio, making (i.e. to assume history rebegin) if Resources allocation ratio according to this formula in recent history, obtained income can be far beyond historical real revenue (maximization).The resource proportioning later just tried to achieve by this formula does the resource pro rate of reality.

Embodiment two

Embodiments provide a kind of resource allocation methods,Such as figure 5Shown in, the method may include that

S201, obtaining the first historical data, obtain the first prediction model according to the modeling of this first historical data, this first prediction model is for characterizing the projected relationship that server is the feedback result that obtains of Terminal for service and resource.

It should be noted that the method that a kind of resource that the embodiment of the present invention provides is distributed, the method being primarily directed on the platform that server provides the resource distribution of the service product issued.

It is understandable that, service product owing to being issued by the platform of server is also more and more, at present, on platform, all there is substantial amounts of service product in every day, owing to user uses these service products to make these service products be that platform brings income, therefore, operator gives each service product by reasonable distribution resource, reaches the purpose of maximum revenue.What the embodiment of the present invention provided is exactly a kind of scheme carrying out rational resource distribution for service product.

The design philosophy of the embodiment of the present invention is mainly: server designs a resource allocation algorithm according to the historical data that the service product of record on platform is corresponding, then this resource allocation algorithm is applied in follow-up resource distribution, so that platform maximum revenue.

Concrete, server can obtain for characterizing the first prediction model that server is the feedback result that obtains of Terminal for service and the projected relationship of resource by the first historical data modeling of platform record, i.e. in the embodiment of the present invention, for the relation of pre-estimation income Yu resource, this server is modeled by the historical data of platform and realizes the first pre-estimation model.

It is understood that the feedback result of embodiment of the present invention terminal can be income.

It should be noted that the first historical data in the embodiment of the present invention refers to the attribute of service product, the natural law of reaching the standard grade (entering the natural law of long tail vat) of service product and the history desired value etc. of service product.Wherein, the history desired value of the service product in the embodiment of the present invention refers to light exposure, download, registration amount, logs in number, paying number, income etc..First historical data is for characterizing when service product is accessed self data and the access data of user feedback.

Further, before server obtains the first historical data, this server wants advanced row Data Preparation, i.e. records each service product in long tail vat every day: product category, by the end of the index based on user behavior such as days running (being called for short " natural law "), desired value on the same day (such as light exposure, download, registration amount, logging in number, paying number, income) in long tail vat on the same day.So, when server sets up the first prediction model, the first historical data of all of online service product just can be added up or get to this server.

It should be noted that the first historical data in the embodiment of the present invention refer to all online service products modeling the same day before historical data.The history desired value of the service product in the embodiment of the present invention refers in the desired value (i.e. Recent data) modeled in n days before the same day.

Exemplary, server obtains the first prediction model characterizing resource E with the projected relationship of income r, such as, r=f (C, D, P, E) according to the attribute C of service product, the natural law D that reaches the standard grade of service product, history desired value P of service product, modeling.

It should be noted that in embodiments of the present invention, C, D, P can be known, therefore, income r is solely dependent upon E, i.e. change the resource allocation of every day, will change the income of every day.

It should be noted that the first prediction model in the embodiment of the present invention is nonlinear, it is preferred that the first prediction model can be neural network model.This is owing to neural network model is output as successive value, and so, the optional scope using the optimized algorithm type selecting of the follow-up total revenue of neural network model is bigger, and it is that centrifugal pump situation is more accurate that result also exports result than model.Wherein, the Hidden unit number in neural network model can be determined by experiment.

Concrete, it is assumed that first historical data of the service product A that server obtainsSuch as table 1Shown in, the attribute C of service product_ij, the natural law D that reaches the standard grade of service product_ij, history desired value P of service product_inj, modeling obtain characterize resource E_iWith income r_i.Wherein, C_ijRepresent the classification of jth sky i-th product, D_ijRepresent the natural law entering long tail vat of jth sky i-th product, P_injRepresent the jth-1 desired value to the i-th product in jth-n sky.When assuming n=14, it is assumed that the history desired value of service product is 6 indexs, then P_injTake 84 fields, C_ijTake 1 field, D_ijTake 1 field, E_iTaking 1 field, the input variable of the first prediction model is 84+1+1+1=87 (x=(C_ij, D_ij, P_inj, E_i)), output variable is 1 y=(r_i)。

Therefore, 3 layers of neural network model, its structure are usedSuch as figure 4Shown in.Here hidden layer (intermediate layer) output valve h₁..., h_pComputing formula be formula (1):

$h_b=f({\mathrm{\Σ}}_{b=1}^{m+1}{w}_{ab}*x_a)---\left(1\right)$

Wherein, b=1 ..., q, x_m+1=1 (bias term), f (x)=1/ (1+e^-x), w_abIt is the connection weights between a node of input layer and the b node of hidden layer.

Output layer y is calculated as formula (2):

$y={\mathrm{\Σ}}_{j=1}^p{\mathrm{\β}}_j*h_j---\left(2\right)$

Wherein, wherein β_bIt is the connection weights between b node of hidden layer and output layer kth node.

It should be noted that the solution procedure of model parameter w (matrix), β (vectorial) is as follows:

W (in neutral net connection weights) between input layer and hidden layer: according to the principle of extreme learning machine, w can be taken as arbitrary random value, such as, be positioned at the random value between [-1,1].And once assignment, the most no longer change during follow-up model optimization.Therefore, the input variable of tube model does not has how many, does not affects the total number of adjustable parameter of model.

Hyper parameter p: be the number of hidden nodes of neutral net, is also the unique hyper parameter in a whole set of algorithm, obtains too small or excessive, may cause poor fitting or over-fitting, can only be determined by experiment an optimum.

Model training seeks β (in neutral net connection weights) between hidden layer and output layer: according to the principle of extreme learning machine, is attributed to and asks a Moore-Penrose generalized inverse just may be used.

In a word, after trying to achieve β, neural network model (the first prediction model) is just set up.The most given (C_ij, D_ij, P_inj, E_i) value, just can try to achieve r_i?.

It should be noted that the pre-side relation of one service product total revenue of a day and resource during neural network model, the relation summation of the income of all online natural law of all products with resource is just obtained the present invention is the first prediction model in embodiment.

S202, obtain the second historical data, model during obtaining the second prediction model according to this second historical data and the first prediction model, persistently detect whether the data output result obtained based on this second prediction model meets preset strategy, when this preset strategy is for characterizing the data output application of results obtained according to this second prediction model in resource proportioning, the corresponding server that obtains is that the feedback result that obtains of Terminal for service is higher than history threshold value.

After server sets up the first prediction model, if wanting the Income Maximum of prediction, the most also need with actual history total revenue compares, then, this server obtains the second historical data, this second historical data is for characterizing the data of self when service product is accessed, the access data of user feedback and the income (i.e. history total revenue) of generation, owing to being that server can not directly be known according to the stock number in the first prediction model, therefore, this server models in the second prediction model obtained still containing needing to seek data parameters (i.e. data output result) according to the second historical data and the first prediction model, this data parameters to be asked is to solve the unknown parameter of the stock number in the embodiment of the present invention.Owing to server wants to predict the target of a maximum revenue, therefore, server needs constantly to carry out detecting whether data output result (data parameters to be asked) obtained based on this second prediction model meet preset strategy, this preset strategy (utilizes data output result can try to achieve resource proportioning) when sign exports application of results in resource proportioning according to the data that this second prediction model obtains, the corresponding server obtained is that the feedback result that Terminal for service obtains (tries to achieve the resource of service product higher than history threshold value by resource proportioning, never feedback result (income) is tried to achieve), owing to the second prediction model is a model continuing to optimize training, server by continuing to optimize the second prediction model until finding the model of data output Income Maximum corresponding to result.

It should be noted that, history threshold value in the embodiment of the present invention refers to the history total revenue that server gets, the income that predicts of prediction model provided of course through the embodiment of the present invention is higher than history threshold value more good, and the data output result selecting one prediction income higher than history threshold value corresponding just can obtain the resource proportioning of optimum reality.

S203, until detecting when the data output result obtained based on the second prediction model meets preset strategy, stops the modeling to the second prediction model, obtains practice in the second model of resource proportioning.

When server detects that the data output result obtained based on the second prediction model meets preset strategy, illustrate that data output result at this moment is exactly the optimal solution that we need, therefore, this server stops the optimization to the second prediction model, determining real the second model being used for solving resource proportioning below, this second model is just off the second prediction model when optimizing.

It is understood that owing to the second model is the model about data parameters to be asked, determine when server and meet when data parameters solution is sought in preferential treatment of preset strategy, this server just can solve resource proportioning according to data parameters to be asked.So, server just can find out the optimal resource allocation ratio of service product by prediction model.

S204, obtain pending target data, according to the second model, pending target data is carried out computing, obtain real resource proportioning.

Server obtains practice after the second model of resource proportioning, and this server just can be by the relevant treatment of this second model realization real resource proportioning.

Concrete, after server obtains real resource proportioning, service product is carried out according to each self-corresponding resource proportioning the distribution of resource, owing to this real resource proportioning is to meet the proportioning that preset strategy is optimum, therefore, after service product carries out resource distribution according to above-mentioned resource proportioning, income just can realize maximising.That is, the embodiment of the present invention is the history achievement data by utilizing each service product in long tail vat, set up the relation (non-linear relation) between product income and the stock number obtaining distribution and determine the formula of stock number ratio, making (i.e. to assume history rebegin) if Resources allocation ratio according to this formula in recent history, obtained income can be far beyond historical real revenue (maximization).The resource proportioning later just tried to achieve by this formula does the resource pro rate of reality.

It should be noted that, pending target data can be the service product related data that server obtains from platform, this server carries out computing according to the second model data to service product, obtained the real resource proportioning of this service product, this server just according to this resource proportioning to each service product Resources allocation.

S205, after Preset Time, the second historical data after obtaining the first historical data after updating and updating, according to the first historical data after updating and the second historical data after described renewal, described second model is adjusted.

S206, carry out the process of real resource proportioning according to the second model after adjusting.

It should be noted that; after once modeling the resource proportioning obtaining optimum; passage, the change in market, the slow creep of user preference over time; the effect of the second model may be gradually lowered; at this moment server it is accomplished by according to new data; the correlation model of component the second model is updated so that it suits latest data situation, keeps income to be unlikely to because of the problem of model itself and reduce.

Concrete, after Preset Time, server can obtain up-to-date historical data and replace the first historical data and the second historical data, i.e. the first historical data and the second historical data are updated, and according to the first historical data after updating and the second historical data after described renewal, described second model is adjusted, finally, the process of real resource proportioning is carried out according to the second model after adjusting.

Embodiment three

Embodiments provide a kind of resource allocation methods,Such as figure 6Shown in, the method may include that

S301, obtain the attribute of service product corresponding to pending target data, the natural law of reaching the standard grade of service product, the history desired value of service product.

S302, obtaining the first prediction model according to the attribute of service product corresponding to pending target data, the natural law of reaching the standard grade of service product, the history desired value modeling of service product, this first prediction model is for characterizing the projected relationship that server is the feedback result that obtains of Terminal for service and resource.

It should be noted that the method that a kind of resource that the embodiment of the present invention provides is distributed, the method being primarily directed on the platform that server provides the resource distribution of the service product issued.

It is understandable that, service product owing to being issued by the platform of server is also more and more, at present, on platform, all there is substantial amounts of service product in every day, owing to user uses these service products to make these service products be that platform brings income, therefore, operator gives each service product by reasonable distribution resource, reaches the purpose of maximum revenue.What the embodiment of the present invention provided is exactly a kind of scheme carrying out rational resource distribution for service product.

The design philosophy of the embodiment of the present invention is mainly: server designs a resource allocation algorithm according to the historical data that the service product of record on platform is corresponding, then this resource allocation algorithm is applied in follow-up resource distribution, so that platform maximum revenue.

Concrete, server can obtain for characterizing the first prediction model that server is the feedback result that obtains of Terminal for service and the projected relationship of resource by the modeling of first historical data (attribute of the service product that pending target data is corresponding, the natural law of reaching the standard grade of service product, the history desired value of service product) of platform record, i.e. in the embodiment of the present invention, for the relation of pre-estimation income Yu resource, this server is modeled by the historical data of platform and realizes the first pre-estimation model.

It is understood that the feedback result of embodiment of the present invention terminal can be income.

It should be noted that the first historical data in the embodiment of the present invention refers to the attribute of service product, the natural law of reaching the standard grade (entering the natural law of long tail vat) of service product and the history desired value etc. of service product.Wherein, the history desired value of the service product in the embodiment of the present invention refers to light exposure, download, registration amount, logs in number, paying number, income etc..

Further, before server obtains the first historical data, this server wants advanced row Data Preparation, i.e. records each service product in long tail vat every day: product category, by the end of the index based on user behavior such as days running (being called for short " natural law "), desired value on the same day (such as light exposure, download, registration amount, logging in number, paying number, income) in long tail vat on the same day.So, when server sets up the first prediction model, the first historical data of all of online service product just can be added up or get to this server.

It should be noted that the first historical data in the embodiment of the present invention refer to all online service products modeling the same day before historical data.The history desired value of the service product in the embodiment of the present invention refers in the desired value (i.e. Recent data) modeled in n days before the same day.

Exemplary, server obtains the first prediction model characterizing resource E with the projected relationship of income r, such as, r=f (C, D, P, E) according to the attribute C of service product, the natural law D that reaches the standard grade of service product, history desired value P of service product, modeling.

It should be noted that in embodiments of the present invention, C, D, P can be known, therefore, income r is solely dependent upon E, i.e. change the resource allocation of every day, will change the income of every day.

It should be noted that the first prediction model in the embodiment of the present invention is nonlinear, it is preferred that the first prediction model can be neural network model.This is owing to neural network model is output as successive value, and so, the optional scope using the optimized algorithm type selecting of the follow-up total revenue of neural network model is bigger, and it is that centrifugal pump situation is more accurate that result also exports result than model.Wherein, the Hidden unit number in neural network model can be determined by experiment.

Concrete, it is assumed that first historical data of the service product A that server obtainsSuch as table 1Shown in, the attribute C of service product_ij, the natural law D that reaches the standard grade of service product_ij, history desired value P of service product_inj, modeling obtain characterize resource E_iWith income r_i.Wherein, C_ijRepresent the classification of jth sky i-th product, D_ijRepresent the natural law entering long tail vat of jth sky i-th product, P_injRepresent the jth-1 desired value to the i-th product in jth-n sky.When assuming n=14, it is assumed that the history desired value of service product is 6 indexs, then P_injTake 84 fields, C_ijTake 1 field, D_ijTake 1 field, E_iTaking 1 field, the input variable of the first prediction model is 84+1+1+1=87 (x=(C_ij, D_ij, P_inj, E_i)), output variable is 1 y=(r_i)。

Therefore, 3 layers of neural network model, its structure are usedSuch as figure 4Shown in.Here hidden layer (intermediate layer) output valve h₁..., h_pComputing formula be formula (1):

$h_b=f({\mathrm{\Σ}}_{b=1}^{m+1}{w}_{ab}*x_a)---\left(1\right)$

Wherein, b=1 ..., q, x_m+1=1 (bias term), f (x)=1/ (1+e^-x), w_abIt is the connection weights between a node of input layer and the b node of hidden layer.

Output layer y is calculated as formula (2):

$y={\mathrm{\Σ}}_{j=1}^p{\mathrm{\β}}_j*h_j---\left(2\right)$

Wherein, wherein β_bIt is the connection weights between b node of hidden layer and output layer kth node.

It should be noted that the solution procedure of model parameter w (matrix), β (vectorial) is as follows:

W (in neutral net connection weights) between input layer and hidden layer: according to the principle of extreme learning machine, w can be taken as arbitrary random value, such as, be positioned at the random value between [-1,1].And once assignment, the most no longer change during follow-up model optimization.Therefore, the input variable of tube model does not has how many, does not affects the total number of adjustable parameter of model.

Hyper parameter p: be the number of hidden nodes of neutral net, is also the unique hyper parameter in a whole set of algorithm, obtains too small or excessive, may cause poor fitting or over-fitting, can only be determined by experiment an optimum.

Model training seeks β (in neutral net connection weights) between hidden layer and output layer: according to the principle of extreme learning machine, is attributed to and asks a Moore-Penrose generalized inverse just may be used.

In a word, after trying to achieve β, neural network model (the first prediction model) is just set up.The most given (C_ij, D_ij, P_inj, E_i) value, just can try to achieve r_i?.

It should be noted that the pre-side relation of one service product total revenue of a day and resource during neural network model, the relation summation of the income of all online natural law of all products with resource is just obtained the present invention is the first prediction model in embodiment.

S303, according to the attribute of service product, the history desired value of reach the standard grade natural law and the service product of service product and default Unknown weights value vector, modeling obtains the 3rd prediction model, and the 3rd prediction model is for characterizing the projected relationship of resource proportioning and resource.

It should be noted that the 3rd prediction model in the embodiment of the present invention is also nonlinear model, presetting Unknown weights value vector is the unknown parameter that in this nonlinear model, band solves.Preferably, 3rd prediction model can be neural network model, now, presetting Unknown weights value vector is the first prediction model hidden layer in embodiment and the connection weights between output, and solving of other parameters in the 3rd prediction model is all consistent with the principle of the first prediction model.

Particularly, resource here refers to, with the relation of resource proportioning, the resource that each service product is corresponding.

Exemplary, server obtains the 3rd prediction model characterizing resource E with the projected relationship of resource proportioning e according to the attribute C of service product, the natural law D that reaches the standard grade of service product, history desired value P of service product, modeling, such as, e=g (C, D, P, W), E=T*e=T*g (C, D, P, W).Wherein, T is total resources, and W is for presetting Unknown weights value vector.

S304, the history total revenue of acquisition service product, obtain the second prediction model according to history total revenue, the 3rd prediction model and the modeling of the first prediction model of service product.

After server modeling obtains projected relationship the 3rd prediction model of sign resource proportioning and resource, this server can also obtain service product history total revenue (the actual true total revenue of history) before modeling, then, in order to set up the relation between resource proportioning and income, and making to predict maximum revenue, server obtains the second prediction model according to history total revenue, the 3rd prediction model and the modeling of the first prediction model of service product.

Exemplary, e_i=g (C_ij, D_ij, P_inj, W), i.e. certain product is according to its Type C_i, natural law D_i, desired value P in before the same day n days_in, determine resource allocation proportion e next day to it_i, data parameters to be asked during wherein W is this model (Unknown weights value vector).It addition, the total resources of all game in note jth sky is T_j, then the stock number of i-th section of product acquisition in this day is E_i=T_j*e_i。

Assume that the service product in the longest tail vat, according to the second prediction model, is redistributed resource ratio by server, it would be desirable that the history " income " of acquisition and ratio R evRise of history true earning can reach maximum, i.e. formula (3):

$\max imize:\mathrm{Re}vRise=\frac{{\Σ}_{j=1}^N{\Σ}_{i=1}^{M_j}f(C_{ij},D_{ij},P_{inj},T_{j*g}(C_{ij},D_{ij},P_{i14j},W\left)\right)}{R_{real}}---\left(3\right)$

Wherein, N represents those all skies having statistical index data by the end of modeling day in history, M_jRepresent the product number in the long tail vat in jth sky, C_ijRepresent the classification of jth sky i-th product, D_ijRepresent the natural law entering long tail vat of jth sky i-th product, P_injRepresent the jth-1 desired value to the i-th product in jth-n sky, R_realFor the true total revenue of history.Meanwhile, the resource proportioning sum that be also satisfied all service products to be the condition of 1, i.e. formula (4).

$subjectto:{\mathrm{\Σ}}_{i=1}^{M_j}{e}_i\≡{\mathrm{\Σ}}_{i=1}^{M_j}g(C_{ij},D_{ij},P_{inj},W)=1,j=1,...,N---\left(4\right)$

It is to say, server needs to solve an optimization problem finding a function maximum with N number of equality constraint, i.e. need to find most suitable W (data parameters to be asked) so that RevRise is maximum.

For the ease of modeling and optimization, function g () can be rewritten, such as formula (5):

$g(C_{ij},D_{ij},P_{i14j},W)=\frac{\exp \[h(C_{ij},D_{ij},P_{inj},W)\]}{{\Σ}_{i=1}^{M_j}\exp \[h(C_{ij},D_{ij},P_{inj},W)\]}---\left(5\right)$

The most revised formula (5), meets the constraints of formula (4), and therefore, formula (3) is finally rewritten as formula (6):

$Minimize:S\left(W\right)\≡R_{\mathrm{Re}al}/{\mathrm{\Σ}}_{j=1}^{N}R(j,W)---\left(6\right)$

It is to say, server is according to the second historical data, Unknown weights value vector (data parameters to be asked) and the first prediction model and the 3rd prediction model, model the second prediction model.

During S305, history total revenue according to service product, the 3rd prediction model and the modeling of the first prediction model obtain the second prediction model, persistently detect whether the data output result obtained based on this second prediction model meets preset strategy, when this preset strategy is for characterizing the data output application of results obtained according to the second prediction model in resource proportioning, the corresponding server that obtains is that the feedback result that obtains of Terminal for service is higher than history threshold value.

After server sets up the first prediction model and the second prediction model, if wanting the Income Maximum of prediction, the most also need with actual history total revenue compares, then, this server obtains the second historical data (including history total revenue), owing to being that server can not directly be known according to the stock number in the first prediction model, therefore, this server is according to the second historical data, still containing needing to seek data parameters (i.e. data output result) in the second prediction model that first prediction model and the modeling of the 3rd prediction model obtain, this data parameters to be asked is to solve the unknown parameter of the stock number in the embodiment of the present invention.Owing to server wants to predict the target of a maximum revenue, therefore, server needs constantly to carry out detecting whether data output result (data parameters to be asked) obtained based on this second prediction model meet preset strategy, this preset strategy (utilizes data output result can try to achieve resource proportioning) when sign exports application of results in resource proportioning according to the data that this second prediction model obtains, the corresponding server obtained is that the feedback result that Terminal for service obtains (tries to achieve the resource of service product higher than history threshold value by resource proportioning, never feedback result (income) is tried to achieve), owing to the second prediction model is a model continuing to optimize training, server by continuing to optimize the second prediction model until finding the model of data output Income Maximum corresponding to result.

It should be noted that, history threshold value in the embodiment of the present invention refers to the history total revenue that server gets, the income that predicts of prediction model provided of course through the embodiment of the present invention is higher than history threshold value more good, and the data output result selecting one prediction income higher than history threshold value corresponding just can obtain the resource proportioning of optimum reality.

Exemplary, server is in the history total revenue of service product, during 3rd prediction model and the modeling of the first prediction model obtain the second prediction model, persistently detect whether the data output result obtained based on this second prediction model meets preset strategy, this preset strategy is when sign exports application of results in resource proportioning according to the data that the second prediction model obtains, the server of corresponding acquisition is that the feedback result that Terminal for service obtains is rewritten as higher than history threshold value detecting data output result (Unknown weights value vector or the function S (W) of the data parameters to be asked) minimum obtained based on the second prediction model.

S306, until detecting when the data output result obtained based on the second prediction model meets preset strategy, stops the modeling to the second prediction model, obtains practice in the second model of resource proportioning.

When server detects that the data output result obtained based on the second prediction model meets preset strategy, illustrate that data output result at this moment is exactly the optimal solution that we need, therefore, this server stops the optimization to the second prediction model, determining real the second model being used for solving resource proportioning below, this second model is just off the second prediction model when optimizing.

It is understood that owing to the second model is the model about data parameters to be asked, determine when server and meet when data parameters solution is sought in preferential treatment of preset strategy, this server just can solve resource proportioning according to data parameters to be asked.So, server just can find out the optimal resource allocation ratio of service product by prediction model.

Exemplary, server detects when the data output result obtained based on the second prediction model meets preset strategy, obtain practice in the second model of resource proportioning, namely formula (6) is carried out the process that minima solves, obtain the solution about W, this W is used for the 3rd prediction model and can be obtained by real resource proportioning.

Concrete, the solution procedure of formula (6) can use particle swarm optimization algorithm, the optimized algorithm of other heuristic direct search method can also be used, such as: genetic algorithm, simulated annealing, TABU search, population, ant colony, quantum inspiration, differential evolution, artificial immunity, Bayes optimize scheduling algorithm.

For example, with use particle swarm optimization algorithm find a function f (x) minima process particularly as follows:

(1) position of initial population, is determined；Randomly choose s point { x⁽¹⁾..., x^(s), as initial population x⁽¹⁾(0) ..., x^(s)(0)。

(2) speed of primary, is determined；

(3) desired positions of each particle and all historical desired positions of particle, are initialized；

If y⁽ⁱ⁾(0)=x⁽ⁱ⁾(0), i=1 ..., p, and note y^gFor { y⁽¹⁾(0) ..., y^(s)(0) corresponding that of f minima in } (if having multiple identical, take under minimum target that).Meanwhile, if t=0.

(4), new particle position is determined；

(5), all particles are carried out valuation；

For each i=1 ..., s, calculate f (x⁽ⁱ⁾(t+1))

(6), the optimum position of each particle, and the optimum position of all particles are updated；

(note: y⁽ⁱ⁾Represent the most optimal particle of i-th particle, y^gRepresent the particle that the overall situation is optimal)

(7), stopping criterion for iteration is checked.

If t<T_max, then takes t=t+1, returns (5)；Otherwise terminate, the y finally obtained^gIt it is exactly calculated solution.

S307, obtain pending target data, according to this second model, pending target data is carried out computing, obtain real resource proportioning.

Server obtains practice after the second model of resource proportioning, and this server just can be by the relevant treatment of this second model realization real resource proportioning.

Concrete, after server obtains real resource proportioning, service product is carried out according to each self-corresponding resource proportioning the distribution of resource, owing to this real resource proportioning is to meet the proportioning that preset strategy is optimum, therefore, after service product carries out resource distribution according to above-mentioned resource proportioning, income just can realize maximising.That is, the embodiment of the present invention is the history achievement data by utilizing each service product in long tail vat, set up the relation (non-linear relation) between product income and the stock number obtaining distribution and determine the formula of stock number ratio, making (i.e. to assume history rebegin) if Resources allocation ratio according to this formula in recent history, obtained income can be far beyond historical real revenue (maximization).The resource proportioning later just tried to achieve by this formula does the resource pro rate of reality.

It should be noted that, pending target data can be the service product related data that server obtains from platform, this server carries out computing according to the second model data to service product, obtained the real resource proportioning of this service product, this server just according to this resource proportioning to each service product Resources allocation.

Further, over time, become, the second model can be updated, to ensure continuing of maximum revenue by server by obtaining new historical data.

Embodiment four

As Figure 7Shown in, embodiments providing a kind of server 1, this server 1 may include that

Acquiring unit 10, for obtaining the first historical data, described first historical data is to build relevant full dose data to the first prediction model, and described first historical data is used for characterizing when service product is accessed data and the access data of user feedback of self.

Modeling unit 11, obtains the first prediction model for the described first historical data modeling obtained according to acquiring unit 10, and described first prediction model is for characterizing the projected relationship that server is the feedback result that obtains of Terminal for service and resource.

Described acquiring unit 10, it is additionally operable to obtain the second historical data, described second historical data is to build relevant recent incremental data to the second prediction model, and described second historical data is for characterizing the data of self when service product is accessed, the access data of user feedback and the income of generation.

Detector unit 12, it is additionally operable to described second historical data that described modeling unit 11 obtains according to described acquiring unit 10 and during described first prediction model modeling obtains the second prediction model, persistently detect based on described modeling unit 11 set up described second prediction model obtain data output result whether meet preset strategy, when described preset strategy is for characterizing the data output application of results obtained according to described second prediction model in resource proportioning, the corresponding described server that obtains is that the feedback result that obtains of Terminal for service is higher than history threshold value.

Stop element 13, for, during until described detector unit 12 detects that the data output result obtained based on described second prediction model meets preset strategy, stopping the modeling to the second prediction model；And,

Described modeling unit 11, is additionally operable to obtain practice in the second model of described resource proportioning.

Resource processing unit 14, described second model for obtaining according to described modeling unit 11 carries out the process of real resource proportioning.

Optionally, described acquiring unit 10, it is additionally operable to obtain pending target data.

Resource processing unit 14, the described pending target data that described acquiring unit 10 is obtained by described second model specifically for obtaining according to described modeling unit 11 carries out computing, obtains real resource proportioning.

Optionally, described acquiring unit 10, specifically for obtaining the history desired value of reach the standard grade natural law and the service product of the attribute of service product corresponding to described pending target data, service product.

Described modeling unit 11, it is additionally operable to history desired value and the default Unknown weights value vector of reach the standard grade natural law and the described service product of the attribute of described service product according to the acquisition of described acquiring unit 10, described service product, modeling obtains the 3rd prediction model, and described 3rd prediction model is for characterizing the projected relationship of described resource proportioning and resource.

Described acquiring unit 10, is additionally operable to obtain the history total revenue of service product.

Described modeling unit 11, the history total revenue of the described service product specifically for obtaining according to described acquiring unit 10, described 3rd prediction model and described first prediction model modeling obtain described second prediction model.

Optionally,Such as figure 8Shown in, described server also includes: adjustment unit 15.

Described acquiring unit 10, the second historical data after being additionally operable to the process that described resource processing unit 14 carries out real resource proportioning according to described second model, after Preset Time, after obtaining the first historical data after updating and updating.

Described second model that described modeling unit 11 is obtained by described adjustment unit 15, the first historical data after the described renewal obtained according to described acquiring unit 10 and the second historical data after described renewal is adjusted.

Described resource processing unit 14, is additionally operable to described second model after adjusting according to described adjustment unit 15 and carries out the process of real resource proportioning.

Optionally, described acquiring unit 10, specifically for obtaining the history desired value of reach the standard grade natural law and the described service product of the attribute of described service product corresponding to described pending target data, described service product.

Described modeling unit 11, the attribute of described service product corresponding to described pending target data specifically for obtaining according to described acquiring unit 10, the history desired value modeling of reach the standard grade natural law and described service product of described service product obtain described first prediction model.

As Figure 9Shown in, in actual applications, above-mentioned modeling unit 11, detector unit 12, stop element 13, resource processing unit 14 and adjustment unit 15 all can be realized by the processor 16 being positioned on server, it is specially central processing unit (CPU), microprocessor (MPU), digital signal processor (DSP) or field programmable gate array (FPGA) etc. to realize, described acquiring unit 10 can be realized by the receptor 17 of server, this server also includes transmitter 18 and memorizer 19, concrete, described transmitter 18 can be issued new The information of online service product, first historical data and software code thereof, first historical data and software code thereof, data output result and software code thereof, first prediction model and software code thereof, second prediction model and software code thereof, 3rd prediction model and software code thereof, second model and software code thereof, and real resource proportioning and software code thereof, can be saved in memorizer 19, this memorizer 19, transmitter 18, receptor 17 can be connected with processor 16, wherein, memorizer 19 is used for storing executable program code, this program code includes computer-managed instruction, memorizer 19 may comprise high-speed RAM memorizer, it is likely to also include nonvolatile memory, such as, at least one disk memory.

It should be noted that server 1 He in the embodiment of the present inventionFigure 2In server 41～4n be same equipment.

Those skilled in the art are it should be appreciated that embodiments of the invention can be provided as method, system or computer program.Therefore, the form of the embodiment in terms of the present invention can use hardware embodiment, software implementation or combine software and hardware.And, the present invention can use the form at one or more upper computer programs implemented of computer-usable storage medium (including but not limited to disk memory and optical memory etc.) wherein including computer usable program code.

The present invention is with reference to method, equipment (system) and the flow process of computer program according to embodiments of the present inventionFigureAnd/or square frameFigureDescribe.It should be understood that flow process can be realized by computer program instructionsFigureAnd/or square frameIn figureEach flow process and/or square frame and flow processFigureAnd/or square frameIn figureFlow process and/or the combination of square frame.These computer program instructions can be provided to produce a machine to the processor of general purpose computer, special-purpose computer, Embedded Processor or other programmable data processing device so that the instruction performed by the processor of computer or other programmable data processing device is produced for realizing in flow processFigure oneIndividual flow process or multiple flow process and/or square frameFigure oneThe device of the function specified in individual square frame or multiple square frame.

These computer program instructions may be alternatively stored in and can guide in the computer-readable memory that computer or other programmable data processing device work in a specific way, making the instruction being stored in this computer-readable memory produce the manufacture including command device, this command device realizes in flow processFigure oneIndividual flow process or multiple flow process and/or square frameFigure oneThe function specified in individual square frame or multiple square frame.

These computer program instructions also can be loaded in computer or other programmable data processing device, make on computer or other programmable devices perform sequence of operations step to produce computer implemented process, thus on computer or other programmable devices perform instruction provide for realization in flow processFigure oneIndividual flow process or multiple flow process and/or square frameFigure oneThe step of the function specified in individual square frame or multiple square frame.

The above, only presently preferred embodiments of the present invention, it is not intended to limit protection scope of the present invention.

Claims (10)

1. the method for a resource distribution, it is characterised in that including:

Obtain the first historical data, obtain the first prediction model according to described first historical data modeling, described First prediction model is the feedback result that obtains of Terminal for service and the prediction of resource is closed for characterizing server System；Described first historical data is to build relevant full dose data to the first prediction model；

Obtain the second historical data, obtain according to described second historical data and described first prediction model modeling During second prediction model, persistently detect the data output result obtained based on described second prediction model Whether meeting preset strategy, described preset strategy is for characterizing the data obtained according to described second prediction model Output application of results when resource proportioning, the corresponding described server that obtains be Terminal for service obtain anti- Feedback result is higher than history threshold value；Described second historical data is to build relevant recent increasing to the second prediction model Amount data；

Until detecting when the data output result obtained based on described second prediction model meets preset strategy, Stop the modeling to the second prediction model, obtain practice in the second model of described resource proportioning；

The process of real resource proportioning is carried out according to described second model.

Method the most according to claim 1, it is characterised in that described carry out according to described second model Actual resource proportioning processes, including:

Obtain pending target data, according to described second model, described pending target data is carried out Computing, obtains real resource proportioning.

Method the most according to claim 1, it is characterised in that described acquisition the second historical data, root The process of the second prediction model is obtained according to described second historical data and described first prediction model modeling, including:

Obtain the attribute of service product corresponding to described pending target data, the natural law of reaching the standard grade of service product And the history desired value of service product, according to the attribute of described service product, the sky of reaching the standard grade of described service product Several and the history desired value of described service product and default Unknown weights value vector, modeling obtains the 3rd prediction model, Described 3rd prediction model is for characterizing the projected relationship of described resource proportioning and resource；

Obtain the history total revenue of service product, according to the history total revenue of described service product, the described 3rd Prediction model and described first prediction model modeling obtain described second prediction model.

Method the most according to claim 1 and 2, it is characterised in that described according to described second model After carrying out the process of real resource proportioning, described method also includes:

After Preset Time, the second historical data after obtaining the first historical data after updating and updating, According to the first historical data after described renewal and the second historical data after described renewal to described second model It is adjusted；

The process of real resource proportioning is carried out according to described second model after adjusting.

Method the most according to claim 1, it is characterised in that described acquisition the first historical data, root The first prediction model is obtained according to described first historical data modeling, including:

Obtain the attribute of described service product corresponding to described pending target data, described service product The history desired value of natural law and described service product of reaching the standard grade and resource allocation, according to described pending target The attribute of the described service product that data are corresponding, reach the standard grade natural law and the described service product of described service product History desired value and resource allocation modeling obtain described first prediction model.

6. a server, it is characterised in that including:

Acquiring unit, for obtaining the first historical data, described first historical data is and the first prediction model Build relevant full dose data；

Modeling unit, the described first historical data modeling for obtaining according to acquiring unit obtains first and estimates Model, described first prediction model is the feedback result that obtains of Terminal for service and money for characterizing server The projected relationship in source；

Described acquiring unit, is additionally operable to obtain the second historical data, and described second historical data is pre-with second Estimate the recent incremental data that model construction is relevant；

Detector unit, is additionally operable to the described second history number that described modeling unit obtains according to described acquiring unit During obtaining the second prediction model with described first prediction model modeling, persistently detect and build based on described Whether the data output result that described second prediction model that form unit is set up obtains meets preset strategy, described Preset strategy exports application of results in resource proportioning for characterizing the data obtained according to described second prediction model Time, the corresponding described server obtained is that the feedback result that Terminal for service obtains is higher than history threshold value；

Stop element, for until described detector unit detects the number obtained based on described second prediction model When meeting preset strategy according to output result, stop the modeling to the second prediction model, and,

Described modeling unit, is additionally operable to obtain practice in the second model of described resource proportioning；

Resource processing unit, carries out real resource for described second model obtained according to described modeling unit The process of proportioning.

Server the most according to claim 6, it is characterised in that

Described acquiring unit, is additionally operable to obtain pending target data；

Resource processing unit, obtains described specifically for described second model obtained according to described modeling unit The described pending target data taking unit acquisition carries out computing, obtains real resource proportioning.

Server the most according to claim 6, it is characterised in that

Described acquiring unit, specifically for obtaining the genus of service product corresponding to described pending target data Property, the history desired value of reach the standard grade natural law and service product of service product；

Described modeling unit, is additionally operable to the attribute of described service product, the institute obtained according to described acquiring unit State history desired value and the default Unknown weights value vector of reach the standard grade natural law and the described service product of service product, build Mould obtains the 3rd prediction model, and described 3rd prediction model is for characterizing the prediction of described resource proportioning and resource Relation；

Described acquiring unit, is additionally operable to obtain the history total revenue of service product；

Described modeling unit, the history of the described service product specifically for obtaining according to described acquiring unit is total Income, described 3rd prediction model and described first prediction model modeling obtain described second prediction model.

9. according to the server described in claim 6 or 7, it is characterised in that described server also includes: Adjustment unit；

Described acquiring unit, is additionally operable to described resource processing unit and carries out real resource according to described second model After the process of proportioning, after Preset Time, the after obtaining the first historical data after updating and updating Two historical datas；

Described adjustment unit, the first historical data after the described renewal obtained according to described acquiring unit Described second model obtained described modeling unit with the second historical data after described renewal is adjusted；

Described resource processing unit, is additionally operable to described second model after adjusting according to described adjustment unit and carries out The process of real resource proportioning.

Server the most according to claim 6, it is characterised in that

Described acquiring unit, the described service product corresponding specifically for obtaining described pending target data Attribute, the history desired value of reach the standard grade natural law and described service product of described service product；

Described modeling unit, specifically for the described pending target data obtained according to described acquiring unit The attribute of described service product of correspondence, reach the standard grade natural law and the history of described service product of described service product Desired value modeling obtains described first prediction model.