WO2021120226A1 - 一种转化预估方法及装置 - Google Patents

一种转化预估方法及装置 Download PDF

Info

Publication number
WO2021120226A1
WO2021120226A1 PCT/CN2019/127220 CN2019127220W WO2021120226A1 WO 2021120226 A1 WO2021120226 A1 WO 2021120226A1 CN 2019127220 W CN2019127220 W CN 2019127220W WO 2021120226 A1 WO2021120226 A1 WO 2021120226A1
Authority
WO
WIPO (PCT)
Prior art keywords
end conversion
conversion rate
sample data
data
estimation model
Prior art date
Application number
PCT/CN2019/127220
Other languages
English (en)
French (fr)
Inventor
刘博�
郑文琛
Original Assignee
深圳前海微众银行股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳前海微众银行股份有限公司 filed Critical 深圳前海微众银行股份有限公司
Priority to PCT/CN2019/127220 priority Critical patent/WO2021120226A1/zh
Publication of WO2021120226A1 publication Critical patent/WO2021120226A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Definitions

  • the invention relates to the field of financial technology (Fintech) and the field of artificial intelligence, and in particular to a conversion estimation method and device.
  • the user's conversion behavior is generalized, but in fact, the user's conversion behavior is divided into multiple types.
  • the estimated conversion rate of a certain type of conversion behavior of a user is directly used as the final conversion rate. Obviously, the accuracy of estimating the conversion rate of the user's conversion behavior obtained in this way is low, which is a problem that needs to be solved urgently.
  • This application provides a conversion estimation method and device, which solves the problem of low accuracy of conversion rate estimation in the prior art.
  • this application provides a conversion estimation method, including: obtaining user characteristic information of a user to be evaluated and resource characteristic information of the resource to be exposed; inputting the user characteristic information and the resource characteristic information To a specific front-end conversion rate estimation model to estimate the front-end conversion rate when the user clicks on the resource and a front-end conversion occurs; the specific front-end conversion rate estimation model is based on the click data and click occurrences accumulated by the resource recommendation platform
  • the front-end conversion data of the front-end conversion is obtained through training; the user characteristic information and the resource characteristic information are input into a specific back-end conversion rate estimation model to estimate that the front-end conversion and the back-end conversion of the user for the resource
  • the conversion rate of the back-end conversion; the specific back-end conversion rate estimation model is obtained by training based on the front-end conversion data and the back-end conversion data where the front-end conversion occurs and the back-end conversion occurs; at least according to the front-end conversion rate and The back-end conversion rate determines the conversion rate at which a front-
  • the method further includes: obtaining the click data, the front-end conversion data, and the back-end conversion data; Input the click data and/or the front-end conversion data as the first sample data into the reference front-end conversion rate estimation model, determine the estimated front-end conversion rate of the first sample data, and determine the first The first difference evaluation value between the estimated front-end conversion rate of the sample data and the real front-end conversion rate; input the front-end conversion data and/or the back-end conversion data as the second sample data into the reference back-end conversion rate
  • An estimation model which determines the estimated back-end conversion rate of the second sample data, and determines the second difference evaluation value between the estimated back-end conversion rate of the second sample data and the real back-end conversion rate; at least According to the first difference evaluation value and/or the second difference evaluation value, perform iterative machine training on the reference front-end conversion rate estimation model and ⁇ or the reference back-end conversion rate
  • the first sample data includes each feature value of the first sample data and the true front-end conversion rate of the first sample data, and the estimation of the first sample data is determined
  • the first difference evaluation value between the front-end conversion rate and the real front-end conversion rate includes: substituting each feature value of the first sample data and the real front-end conversion rate of the first sample data into the reference front-end conversion
  • the first loss function of the rate prediction model is calculated, the first function value of the first loss function is calculated, and the first function value is used as the first difference evaluation value
  • the second sample data includes the first Each feature value of the second sample data and the real back-end conversion rate of the second sample data, and the second difference evaluation between the estimated back-end conversion rate of the second sample data and the real back-end conversion rate is determined Value: Substitute each feature value of the second sample data and the true back-end conversion rate of the second sample data into the second loss function of the reference back-end conversion rate estimation model, and calculate the second loss function And use the second function value as the second difference evaluation value; said at least
  • the first difference evaluation value is the difference between the estimated front-end conversion rate of the first sample data and the real front-end conversion rate
  • the second difference evaluation value is the second sample data The difference between the estimated back-end conversion rate and the real back-end conversion rate.
  • the method further includes: obtaining the click data, the front-end conversion data, and the back-end conversion data; Input the click data and/or the front-end conversion data as the fourth sample data into the reference feature expansion space for similarity expansion, and obtain the fourth virtual sample data converted from the reference feature expansion space; The conversion data and/or the back-end conversion data are input as the fifth sample data into the reference feature expansion space for similarity expansion, to obtain the fifth virtual sample data transformed by the reference feature expansion space; Four virtual sample data are input to the reference front-end conversion rate estimation model, the estimated front-end conversion rate of the fourth virtual sample data is determined, and the estimated front-end conversion rate of the fourth virtual sample data and the fourth sample are determined The fourth difference evaluation value between the real front-end conversion rates of the data; input the fifth virtual sample data into a reference back-end conversion rate estimation model to determine the estimated back-end conversion rate of the fifth virtual sample data, And determine the fifth
  • the fourth virtual sample data includes each feature value of the fourth virtual sample data, and the determination of the estimated front-end conversion rate of the fourth virtual sample data and the real front-end of the fourth sample data
  • the fourth difference evaluation value between the conversion rates includes: substituting each feature value of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data into the fourth part of the reference front-end conversion rate prediction model Loss function, the fourth function value of the fourth loss function is calculated, and the fourth function value is used as the fourth difference evaluation value
  • the fifth virtual sample data includes the fifth virtual sample data Each feature value of the fifth virtual sample data, the fifth difference evaluation value between the estimated back-end conversion rate of the fifth virtual sample data and the real back-end conversion rate of the fifth sample data: the fifth virtual Each feature value of the sample data and the true back-end conversion rate of the fifth sample data are substituted into the fifth loss function of the reference back-end conversion rate estimation model, and the fifth function value of the fifth loss function is calculated , And use the fifth function value as the fifth difference evaluation value; at least according to the
  • the fourth difference evaluation value is the difference between the estimated front-end conversion rate of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data;
  • the fifth difference evaluation value Is the difference between the estimated back-end conversion rate of the fifth virtual sample data and the real back-end conversion rate of the fifth sample data.
  • the method before acquiring the user characteristic information of the user to be assessed and the resource characteristic information of the resource to be exposed, the method further includes: acquiring the click data and the exposure data: combining the exposure data and/or the Click data is input as the sixth sample data into the reference feature expansion space for similarity expansion, and the sixth virtual sample data converted from the reference feature expansion space is obtained; the sixth virtual sample data is input into the reference click rate preview An estimation model to determine the estimated click-through rate of the sixth virtual sample data, and determine the sixth difference evaluation value between the estimated click-through rate of the sixth virtual sample data and the real click-through rate of the sixth sample data; According to the sixth difference evaluation value, iterative machine training is performed on the reference click-through rate estimation model; the reference click-through rate estimation model at the end of the training is used as a specific click-through rate estimation model.
  • the user characteristic information and the resource characteristic information are input into a specific click-through rate estimation model to estimate the click-through rate at which the user is exposed by the resource and clicked;
  • the specific click-through rate estimation The model is obtained by training based on the exposure data accumulated by the resource recommendation platform and the click data; at least according to the front-end conversion rate and the back-end conversion rate, it is determined that the user clicks on the resource to generate a front-end conversion and
  • the conversion rate for back-end conversions includes: determining, based on the click rate, the front-end conversion rate, and the back-end conversion rate, the conversion rate at which the user clicks on the resource to cause a front-end conversion and a back-end conversion.
  • this application provides a conversion estimation device, including: an acquisition module for acquiring user characteristic information of a user to be assessed and resource characteristic information of the user’s resource to be exposed; a processing module for converting the The user characteristic information and the resource characteristic information are input into the specific front-end conversion rate estimation model to estimate the front-end conversion rate of the user clicking on the resource and the front-end conversion occurs; the specific front-end conversion rate estimation model is based on the resource The click data accumulated by the recommendation platform and the front-end conversion data of clicks and front-end conversions are trained; the user characteristic information and the resource characteristic information are input into a specific back-end conversion rate estimation model to estimate the target for the resource The back-end conversion rate at which the user undergoes a front-end conversion and a back-end conversion; the specific back-end conversion rate estimation model is obtained by training based on the front-end conversion data and the back-end conversion data where the front-end conversion occurs and the back-end conversion occurs ⁇ ; At least according to the front-end conversion rate and the
  • the acquisition module is further used to: acquire the click data, the front-end conversion data, and the back-end conversion data; the processing module is also used to: combine the click data and/or the front-end
  • the conversion data is input into the reference front-end conversion rate estimation model as the first sample data, the estimated front-end conversion rate of the first sample data is determined, and the estimated front-end conversion rate and the real conversion rate of the first sample data are determined
  • the first difference evaluation value between the front-end conversion rates; the front-end conversion data and/or the back-end conversion data are input as the second sample data into the reference back-end conversion rate estimation model, and the second sample data is determined
  • the second difference evaluation value is to perform iterative machine training on the reference front-end conversion rate estimation model and/or the reference back-end conversion rate estimation model; and the reference
  • the first sample data includes each feature value of the first sample data and the true front-end conversion rate of the first sample data
  • the processing module is specifically configured to: Each feature value of the sample data and the true front-end conversion rate of the first sample data are substituted into the first loss function of the reference front-end conversion rate estimation model, and the first function value of the first loss function is calculated, and The first function value is used as the first difference evaluation value;
  • the second sample data includes each feature value of the second sample data and the true back-end conversion rate of the second sample data, the determining The second difference evaluation value between the estimated back-end conversion rate of the second sample data and the real back-end conversion rate: compare each feature value of the second sample data with the real back-end conversion rate of the second sample data The conversion rate is substituted into the second loss function of the reference back-end conversion rate estimation model, the second function value of the second loss function is calculated, and the second function value is used as the second difference evaluation value; Decrease the first function value and/or the second function value, adjust the parameters of the reference front-end
  • the first difference evaluation value is the difference between the estimated front-end conversion rate of the first sample data and the real front-end conversion rate
  • the second difference evaluation value is the second sample data The difference between the estimated back-end conversion rate and the real back-end conversion rate.
  • the acquisition module is further used to: acquire the click data, the front-end conversion data, and the back-end conversion data; the processing module is also used to: combine the click data and/or the front-end
  • the conversion data is input into the reference feature expansion space as the fourth sample data for similarity expansion, and the fourth virtual sample data obtained by the reference feature expansion space conversion is obtained; the front-end conversion data and/or the back-end conversion data are obtained Input as the fifth sample data into the reference feature expansion space for similarity expansion, and obtain the fifth virtual sample data transformed by the reference feature expansion space; input the fourth virtual sample data into the reference front-end conversion rate prediction Estimation model to determine the estimated front-end conversion rate of the fourth virtual sample data, and determine the fourth difference between the estimated front-end conversion rate of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data Difference evaluation value; input the fifth virtual sample data into a reference back-end conversion rate estimation model, determine the estimated back-end conversion rate of the fifth virtual sample data, and determine the expected value of the fifth virtual sample data Estimate
  • the fourth virtual sample data includes each feature value of the fourth virtual sample data
  • the processing module is specifically configured to: combine each feature value of the fourth virtual sample data with the fourth sample
  • the true front-end conversion rate of the data is substituted into the fourth loss function of the reference front-end conversion rate estimation model, the fourth function value of the fourth loss function is calculated, and the fourth function value is used as the fourth function value.
  • the fifth virtual sample data includes each feature value of the fifth virtual sample data
  • the processing module is specifically configured to: combine each feature value of the fifth virtual sample data with the fifth sample
  • the real back-end conversion rate of the data is substituted into the fifth loss function of the reference back-end conversion rate estimation model, the fifth function value of the fifth loss function is calculated, and the fifth function value is used as the Fifth difference evaluation value; adjust the parameters of the reference front-end conversion rate estimation model by reducing the fourth function value and/or the fifth function value, thereby updating the reference front-end conversion rate estimation model Parameters; and ⁇ or by reducing the value of the fourth function and ⁇ or the value of the fifth function, adjusting the parameters of the reference back-end conversion rate estimation model, thereby updating the reference back-end conversion rate estimation model And ⁇ or by reducing the value of the fourth function and ⁇ or the value of the fifth function, adjusting the parameters of the reference feature expansion space, thereby updating the parameters of the reference feature expansion space.
  • the fourth difference evaluation value is the difference between the estimated front-end conversion rate of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data;
  • the fifth difference evaluation value Is the difference between the estimated back-end conversion rate of the fifth virtual sample data and the real back-end conversion rate of the fifth sample data.
  • the acquisition module is further configured to: acquire the click data and exposure data;
  • the processing module is specifically configured to: input the exposure data and/or the click data as the sixth sample data to the Perform similarity expansion with reference to the feature expansion space to obtain the sixth virtual sample data converted from the reference feature expansion space; input the sixth virtual sample data into a reference click-through rate prediction model to determine the sixth virtual sample data
  • the reference click-through rate estimation model is used for iterative machine training; the reference click-through rate estimation model at the end of the training is used as the specific click-through rate estimation model.
  • the processing module is further configured to: input the user characteristic information and the resource characteristic information into a specific click-through rate estimation model to estimate the click-through rate at which the user is exposed by the resource and clicked
  • the specific click-through rate estimation model is obtained by training based on the exposure data accumulated by the resource recommendation platform and the click data; the processing module is specifically configured to: according to the click-through rate, the front-end conversion rate and the The back-end conversion rate determines the conversion rate at which a front-end conversion and a back-end conversion occur when the user clicks on the resource.
  • the present application provides a computer device including a program or instruction, and when the program or instruction is executed, it is used to execute the methods of the first aspect and the implementation manners of the first aspect.
  • the present application provides a storage medium including a program or instruction, and when the program or instruction is executed, it is used to execute the above-mentioned method of the first aspect and each implementation manner of the first aspect.
  • This application provides a conversion estimation method and device.
  • the specific front-end conversion rate estimation model is based on the click data and clicks accumulated by the resource recommendation platform and the front-end conversion occurs
  • the specific front-end conversion rate estimation model will learn the click data and the knowledge of the front-end conversion data of the click and the front-end conversion occurs.
  • the specific back-end conversion rate estimation model is based on The front-end conversion data and the back-end conversion data of the front-end conversion and the back-end conversion are trained.
  • the specific front-end conversion rate estimation model will learn the front-end conversion data and the back-end conversion data of the front-end conversion and the back-end conversion.
  • the specific front-end conversion rate estimation model can combine the learned knowledge to estimate that the user clicks on the resource and The front-end conversion rate of the front-end conversion, the specific back-end conversion rate estimation model can be combined with the learned knowledge to estimate the back-end conversion rate of the front-end conversion and the back-end conversion of the user for the resource, so as to pass the user’s
  • the conversion behavior distinguishes the front-end conversion and the back-end conversion, which can respectively estimate the front-end conversion rate and the back-end conversion rate, and further determines the front-end conversion when the user clicks on the resource based on the front-end conversion rate and the back-end conversion rate. And the conversion rate of the back-end conversion occurs, which can improve the accuracy of conversion estimation.
  • FIG. 1 is a schematic diagram of the process flow of a conversion estimation method provided by an embodiment of the application
  • FIG. 2 is a schematic diagram of an applicable architecture of a conversion estimation method provided by an embodiment of the application
  • FIG. 3 is a schematic structural diagram of a conversion estimation device provided by an embodiment of the application.
  • an embodiment of the present application provides a conversion estimation method.
  • Step 101 Obtain user characteristic information of a user to be evaluated and resource characteristic information of the resource to be exposed of the user.
  • Step 102 Input the user characteristic information and the resource characteristic information into a specific front-end conversion rate estimation model to estimate the front-end conversion rate at which the user clicks on the resource and the front-end conversion occurs.
  • Step 103 Input the user characteristic information and the resource characteristic information into a specific back-end conversion rate estimation model to estimate the back-end conversion rate at which the user has a front-end conversion and a back-end conversion for the resource.
  • Step 104 Determine, based on at least the front-end conversion rate and the back-end conversion rate, a conversion rate at which a front-end conversion and a back-end conversion occur when the user clicks on the resource.
  • the front-end conversion behavior is a pre-defined type of user conversion behavior, which is represented by the first type of conversion behavior, such as using the free service of the resource provider, such as registration;
  • the back-end conversion behavior is another type of user defined in advance Conversion behavior, expressed as the second type of conversion behavior
  • the back-end conversion behavior needs to be based on the front-end conversion behavior, such as the use of resource-providing paid services, such as the purchase of virtual items after registration, the paid services must be established on the basis of free services on. For example, if a user clicks on the recommendation information of a game in a certain scene, the user is the click user.
  • the user purchases in the game it is said that the user has a back-end conversion behavior.
  • the execution subject of steps 101 to 104 can be a resource recommendation platform.
  • the resource provider in this application is the source provider of recommended information, and the recommended information needs to be pushed to users, so that some of the user backends are transformed into The back-end of the resource provider converts users, and the resource recommendation platform is the specific executor of the recommended information push.
  • users can be divided into three categories according to the time sequence of their behavior: exposure users, click users, front-end converted users, and back-end converted users. Exposure users are users who have recommended resources for the resource recommendation platform; click users are users who have clicked on the resources recommended by the resource recommendation platform.
  • the comprehensive feature information may include multiple types of feature information, such as user feature information, resource feature information, and scene feature information.
  • User characteristic information refers to the basic attributes of the user, such as age, gender, and so on.
  • the resource characteristic information refers to the basic attributes of the recommended resource, such as the format and layout of the recommended resource.
  • the scene feature information refers to the scene in which the recommended resource is recommended to the user, such as the recommended location and the user operation that triggers the recommendation.
  • x)P(y 1 1
  • y 1 1, x).
  • P(y 0 1
  • x), P(y 1 1
  • step 102 the specific front-end conversion rate estimation model is obtained by training based on the click data accumulated by the resource recommendation platform and the front-end conversion data of clicks and front-end conversions.
  • step 103 the specific back-end conversion rate estimation model is obtained by training based on the front-end conversion data and the back-end conversion data in which the front-end conversion occurs and the back-end conversion occurs. It can be seen that the specific front-end conversion rate estimation model has learned the click data and the knowledge of the front-end conversion data of the click and the front-end conversion occurs, and the specific back-end conversion rate estimation model has learned the front-end conversion data and the front-end conversion and after the occurrence of the conversion.
  • the specific front-end conversion rate estimation model and the specific back-end conversion rate estimation model there are multiple training methods for the specific front-end conversion rate estimation model and the specific back-end conversion rate estimation model in steps 101 to 104, as long as the specific front-end conversion rate estimation model is based on the click data and clicks accumulated by the resource recommendation platform And the front-end conversion data of the front-end conversion can be obtained by training, and the specific back-end conversion rate estimation model can be obtained by training based on the front-end conversion data and the back-end conversion data of the front-end conversion and the back-end conversion. This is not limited.
  • the first category is a model training method that does not extend the similarity feature, specifically:
  • Step (1-1) Obtain the click data, the front-end conversion data, and the back-end conversion data.
  • Step (1-5) Use the reference front-end conversion rate estimation model at the end of the training as the specific front-end conversion rate estimation model; set the reference back-end conversion rate estimation model at the end of the training, As the specific back-end conversion rate estimation model.
  • steps (1-2) to (1-3) are not a sequential constraint relationship, and steps (1-2) to (1-3) can be executed in parallel.
  • the first sample data and the second sample data are selectively input into each reference training model, so as to obtain the estimated rate and true
  • the difference evaluation value of the rate of the training process can effectively characterize the estimation accuracy in the training process, and perform iterative machine training, and finally obtain a specific model, thereby providing a method for simultaneously training to obtain the specific click-through rate prediction model, the The method of a specific conversion rate estimation model.
  • step (1-1) Specifically, in the process from step (1-1) to step (1-5):
  • the first difference evaluation value in step (1-4) may be the difference between the estimated front-end conversion rate of the first sample data and the real front-end conversion rate of the first sample data.
  • the method can most intuitively represent the difference between the estimated front-end conversion rate of the first sample data and the real front-end conversion rate of the first sample data. It can also be the value of the loss function, and the specific method of solving the value of the loss function can be:
  • the input data contains feature information x and label values y 0 and ⁇ or y 1 and ⁇ or y 2 , and there is a mapping relationship between the label value and the true probability value.
  • the true probability is 0%.
  • the training function of the front-end conversion rate estimation model is f 1 (x
  • W 1 ) can get the value of L(y 1 ,p 0 p 1 ), which is the first
  • a function value is the value of L(y 1 , p 0 p 1 ).
  • the second difference evaluation value in step (1-5) may be the difference between the estimated back-end conversion rate of the second sample data and the real back-end conversion rate of the second sample data.
  • the method can most intuitively represent the difference between the estimated back-end conversion rate of the second sample data and the real back-end conversion rate of the second sample data. It can also be the value of the loss function, and the specific method of solving the value of the loss function can be:
  • the training function of the back-end conversion rate estimation model is f 2 (x
  • the specific steps (1-6) can be:
  • the parameters of the reference front-end conversion rate prediction model are adjusted to update the reference front-end conversion rate prediction model; by reducing the first A function value and/or the second function value adjust the parameters of the reference back-end conversion rate estimation model to update the reference back-end conversion rate estimation model.
  • a click prediction model can also be added, specifically: obtaining exposure data and the click data; combining the exposure data and/or all
  • the click data is input to the reference click-through rate estimation model as the third sample data, the estimated click-through rate of the third sample data is determined, and the difference between the estimated click-through rate of the third sample data and the real click-through rate is determined
  • the third difference evaluation value at least according to the third difference evaluation value, perform iterative machine training on the reference click-through rate estimation model; use the reference click-through rate estimation model at the end of the training as the specific click rate Estimate model.
  • the training function of the click-through rate estimation model is f 0 (x
  • the second category is a model training method for extending similarity features, specifically:
  • Step (2-3) Input the front-end conversion data and/or the back-end conversion data as the fifth sample data into the reference feature expansion space for similarity expansion, and obtain the reference feature expansion space conversion The fifth virtual sample data obtained.
  • Step (2-5) Input the fifth virtual sample data into the reference back-end conversion rate estimation model, determine the estimated back-end conversion rate of the fifth virtual sample data, and determine the fifth virtual sample data The fifth difference evaluation value between the estimated back-end conversion rate of the sample data and the real back-end conversion rate of the fifth sample data.
  • Step (2-6) Estimate the reference front-end conversion rate prediction model and/or the reference back-end conversion rate at least according to the fourth difference evaluation value and/or the fifth difference evaluation value
  • the model and/or the reference feature expansion space undergoes iterative machine training.
  • Step (2-7) Use the reference front-end conversion rate estimation model at the end of the training as the specific front-end conversion rate estimation model; use the reference back-end conversion rate estimation model at the end of the training, As the specific back-end conversion rate estimation model; the reference feature expansion space at the end of training is used as the specific feature expansion space.
  • steps (2-2) to (2-3) are not in a sequential constraint relationship, steps (2-2) to (2-3) can be executed in parallel; 4) Steps to (2-5) are not constrained sequentially, and steps (2-4) to (2-5) can be executed in parallel.
  • the first sample data and the second sample data are selectively input into the reference feature expansion space, and in the reference feature expansion space, according to different data types
  • the first sample data and the second sample data are converted to obtain the first virtual sample data and the second virtual sample data, so as to obtain the difference evaluation value of each estimated rate and the true rate, which can effectively characterize the prediction in the training process.
  • step (2-1) Specifically, in the process from step (2-1) to step (2-7):
  • the fourth difference evaluation value in step (2-4) may be the difference between the estimated front-end conversion rate of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data, in this way It can most intuitively represent the difference between the estimated front-end conversion rate of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data. It can also be the value of the loss function, and the specific method of solving the value of the loss function can be:
  • the fourth loss function of the fourth loss function is calculated. Function value, and use the fourth function value as the fourth difference evaluation value.
  • the input data contains feature information x and label values y 0 and ⁇ or y 1 and ⁇ or y 2 , and there is a mapping relationship between the label value and the true probability value.
  • the true probability is 0%.
  • the loss function of the front-end conversion rate estimation model is L(y 1 ,p 0 p 1 ), according to the real click rate corresponding to the input click tag value and ⁇ or the real front-end conversion rate and f 1 corresponding to the front-end conversion tag value
  • W 1 ,W e ) can get the value of L(y 1 ,p 0 p 1 ), that is, the value of the fourth function value of L(y 1 ,p 0 p 1 ) .
  • the fifth difference evaluation value in step (2-5) may be the difference between the estimated back-end conversion rate of the fifth virtual sample data and the real back-end conversion rate of the fifth sample data. This way can most intuitively represent the difference between the estimated back-end conversion rate of the fifth virtual sample data and the real back-end conversion rate of the fifth sample data. It can also be the value of the loss function, and the specific method of solving the value of the loss function can be:
  • the training function of the back-end conversion rate estimation model is f 2 (x
  • the loss function of the back-end conversion rate estimation model is L(y 2 ,p 0 p 1 p 2 ) ,
  • the value of L(y 2 , p 0 p 1 p 2 ) can be obtained, that is, the value of the fifth function value is L(y 2 , p 0 p 1 p 2 ).
  • Adjust the parameters of the reference front-end conversion rate prediction model by reducing the fourth function value and/or the fifth function value, thereby updating the parameters of the reference front-end conversion rate prediction model; and ⁇ or pass Decrease the value of the fourth function and/or the value of the fifth function, adjust the parameters of the reference back-end conversion rate estimation model, thereby updating the parameters of the reference back-end conversion rate estimation model; and ⁇ or By reducing the fourth function value and/or the fifth function value, the parameters of the reference feature expansion space are adjusted, thereby updating the parameters of the reference feature expansion space.
  • W 1 For example, adjust W 1 through L(y 1 ,p 0 p 1 ) and L(y 2 ,p 0 p 1 p 2 ); through L(y 1 ,p 0 p 1 ) and L(y 2 , p 0 p 1 p 2) adjusted W 2, (y 2, p 0 p 1 p 2) adjusted by W e L (y 1, p 0 p 1) and L, i.e., loss of joint function:
  • a click prediction model can also be added, specifically: acquiring the click data and exposure data; combining the exposure data and/or all The click data is input into the reference feature expansion space as the sixth sample data to perform similarity expansion, to obtain the sixth virtual sample data converted from the reference feature expansion space; the sixth virtual sample data is input into the reference click rate An estimation model to determine the estimated click-through rate of the sixth virtual sample data, and determine the sixth difference evaluation value between the estimated click-through rate of the sixth virtual sample data and the real click-through rate of the sixth sample data; Performing iterative machine training on the reference click-through rate estimation model at least according to the sixth difference evaluation value;
  • the reference click-through rate estimation model at the end of the training is used as the specific click-through rate estimation model.
  • the training function of the click-through rate estimation model is f 0 (x
  • W 0 ,W e ) can get the value of L(y 0 ,p 0 ), that is, the value of the sixth function value of L(y 0 ,p 0 ).
  • step 101 An optional implementation manner other than step 101 to step 104 is:
  • the user characteristic information and the resource characteristic information are input into a specific click-through rate estimation model to estimate the click-through rate at which the user is exposed and clicked by the resource; the specific click-through rate estimation model is based on The exposure data accumulated by the resource recommendation platform and the click data are trained; according to the click rate, the front-end conversion rate, and the back-end conversion rate, it is determined that the user clicks on the resource and the front-end conversion occurs and after the Conversion rate of end conversion.
  • the front-end conversion rate is 0.8 and the back-end conversion rate is 0.6
  • the product of the front-end conversion rate and the back-end conversion rate of 0.48 can be used as the front-end conversion and the back-end conversion when the user clicks on the resource.
  • the conversion rate of the conversion is 0.8
  • the present application provides a conversion estimation device, including: an acquisition module 301, configured to acquire user characteristic information of a user to be assessed and resource characteristic information of the user’s resource to be exposed; a processing module 302, The user characteristic information and the resource characteristic information are input into a specific front-end conversion rate estimation model to estimate the front-end conversion rate at which the user clicks on the resource and the front-end conversion occurs; the specific front-end conversion rate estimation The model is obtained by training based on the click data accumulated by the resource recommendation platform and the front-end conversion data of clicks and front-end conversions; input the user characteristic information and the resource characteristic information into the specific back-end conversion rate estimation model to estimate For the resource, the front-end conversion rate of the user and the back-end conversion rate; the specific back-end conversion rate estimation model is based on the front-end conversion data and the back-end conversion rate. Conversion data obtained through training; at least according to the front-end conversion rate and the back-end conversion rate, determine the conversion rate at which the user clicks on the
  • the acquisition module 301 is also used to: acquire the click data, the front-end conversion data, and the back-end conversion data; the processing module 302 is also used to: combine the click data and/or all
  • the front-end conversion data is input to the reference front-end conversion rate estimation model as the first sample data, the estimated front-end conversion rate of the first sample data is determined, and the estimated front-end conversion rate of the first sample data is determined
  • the first difference evaluation value between the real front-end conversion rate and the real front-end conversion rate input the front-end conversion data and/or the back-end conversion data as the second sample data into the reference back-end conversion rate estimation model to determine the The estimated back-end conversion rate of the sample data, and determine the second difference evaluation value between the estimated back-end conversion rate of the second sample data and the real back-end conversion rate; at least according to the first difference evaluation value and ⁇ Or the second difference evaluation value, perform iterative machine training on the reference front-end conversion rate estimation model and ⁇ or the reference back-end conversion rate estimation model; predict the reference front
  • the first sample data includes each feature value of the first sample data and the true front-end conversion rate of the first sample data
  • the processing module 302 is specifically configured to: Each feature value of the sample data and the true front-end conversion rate of the first sample data are substituted into the first loss function of the reference front-end conversion rate estimation model, and the first function value of the first loss function is calculated, And use the first function value as the first difference evaluation value;
  • the second sample data includes each feature value of the second sample data and the true back-end conversion rate of the second sample data, the Determine the second difference evaluation value between the estimated back-end conversion rate of the second sample data and the real back-end conversion rate: compare each feature value of the second sample data with the true back-end conversion rate of the second sample data The end conversion rate is substituted into the second loss function of the reference back-end conversion rate estimation model, the second function value of the second loss function is calculated, and the second function value is used as the second difference evaluation value; By reducing the first function value and/or the second function value, the parameters of the reference front-
  • the first difference evaluation value is the difference between the estimated front-end conversion rate of the first sample data and the real front-end conversion rate
  • the second difference evaluation value is the second sample data The difference between the estimated back-end conversion rate and the real back-end conversion rate.
  • the acquisition module 301 is also used to: acquire the click data, the front-end conversion data, and the back-end conversion data;
  • the processing module 302 is also used to: combine the click data and/or all
  • the front-end conversion data is input into the reference feature expansion space as the fourth sample data for similarity expansion, and the fourth virtual sample data obtained by the reference feature expansion space conversion is obtained;
  • the front-end conversion data and/or the back-end The transformed data is input as the fifth sample data into the reference feature expansion space for similarity expansion, and the fifth virtual sample data obtained by the reference feature expansion space is obtained;
  • the fourth virtual sample data is input into the reference front end for transformation Rate estimation model to determine the estimated front-end conversion rate of the fourth virtual sample data, and determine the difference between the estimated front-end conversion rate of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data Fourth difference evaluation value; input the fifth virtual sample data into a reference back-end conversion rate estimation model, determine the estimated back-end conversion rate of the fifth virtual sample data, and determine the fifth virtual
  • the fourth virtual sample data includes each feature value of the fourth virtual sample data
  • the processing module 302 is specifically configured to: combine each feature value of the fourth virtual sample data with the fourth The real front-end conversion rate of the sample data is substituted into the fourth loss function of the reference front-end conversion rate estimation model, the fourth function value of the fourth loss function is calculated, and the fourth function value is taken as the first Four difference evaluation values
  • the fifth virtual sample data includes each feature value of the fifth virtual sample data
  • the processing module 302 is specifically configured to: combine each feature value of the fifth virtual sample data with the first The real back-end conversion rate of the five sample data is substituted into the fifth loss function of the reference back-end conversion rate estimation model, the fifth function value of the fifth loss function is calculated, and the fifth function value is taken as The fifth difference evaluation value; by reducing the fourth function value and/or the fifth function value, the parameters of the reference front-end conversion rate estimation model are adjusted, thereby updating the reference front-end conversion rate estimation Parameters of the model; and ⁇ or by reducing the value of the fourth
  • the fourth difference evaluation value is the difference between the estimated front-end conversion rate of the fourth virtual sample data and the real front-end conversion rate of the fourth sample data;
  • the fifth difference evaluation value Is the difference between the estimated back-end conversion rate of the fifth virtual sample data and the real back-end conversion rate of the fifth sample data.
  • the acquisition module 301 is further configured to: acquire the click data and exposure data; the processing module 302 is specifically configured to: input the exposure data and/or the click data as the sixth sample data to The reference feature expansion space performs similarity expansion to obtain the sixth virtual sample data transformed from the reference feature expansion space; the sixth virtual sample data is input into a reference click-through rate estimation model to determine the sixth virtual sample data The estimated click rate of the sample data, and determine the sixth difference evaluation value between the estimated click rate of the sixth virtual sample data and the real click rate of the sixth sample data; at least according to the sixth difference evaluation value, Iterative machine training is performed on the reference click-through rate estimation model; the reference click-through rate estimation model at the end of the training is used as a specific click-through rate estimation model.
  • the processing module 302 is further configured to: input the user characteristic information and the resource characteristic information into a specific click-through rate estimation model to estimate that the user is exposed by the resource and clicked
  • the specific click-through rate estimation model is trained based on the exposure data accumulated by the resource recommendation platform and the click data; the processing module is specifically configured to: according to the click-through rate, the front-end conversion rate, and The back-end conversion rate determines the conversion rate at which a front-end conversion and a back-end conversion occur when the user clicks on the resource.
  • the embodiment of the present application provides a computer device, including a program or instruction, when the program or instruction is executed, it is used to execute a conversion estimation method and any optional method provided in the embodiment of the present application.
  • the embodiment of the present application provides a storage medium including a program or instruction, and when the program or instruction is executed, it is used to execute a conversion estimation method and any optional method provided in the embodiment of the present application.
  • this application can be provided as methods, systems, or computer program products. Therefore, this application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program codes.
  • These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device.
  • the device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种转化预估方法及装置，其中方法为：获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息（101）；将所述用户特性信息和资源特征信息输入到特定前端转化率预估模型，以预估所述用户点击所述资源且发生前端转化的前端转化率（102）；将所述用户特性信息和资源特征信息输入到特定后端转化率预估模型，以预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率（103）；至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率（104）。上述方法应用于金融科技（Fintech）时，能够提升预估转化率的准确性。

Description

一种转化预估方法及装置 技术领域
本发明涉及金融科技(Fintech)领域和人工智能领域，尤其涉及一种转化预估方法及装置。
背景技术
随着计算机技术的发展，越来越多的技术(大数据、分布式、区块链(Blockchain)、人工智能等)应用在金融领域，传统金融业正在逐步向金融科技(Fintech)转变。目前，金融科技领域中，越来越多的资源提供端对信息推荐效果重视起来，如何预估信息推荐效果尤为重要。
目前方法中，将用户的转化行为一概而论，但实际上用户的转化行为又分为多种。现有技术中，直接将用户发生某一类转化行为的预估转化率，作为最终的转化率。显然，这种方式得到对用户转化行为的转化率进行预估的准确率较低，这是一个亟待解决的问题。
发明内容
本申请提供一种转化预估方法及装置，解决了现有技术中转化率预估的准确率较低的问题。
第一方面，本申请提供一种转化预估方法，包括：获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息；将所述用户特性信息和所述资源特征信息输入到特定前端转化率预估模型，以预估所述用户点击所述资源且发生前端转化的前端转化率；所述特定前端转化率预估模型为根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的；将所述用户特性信息和所述资源特征信息输入到特定后端转化率预估模型，以预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率；所述特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的；至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
可选地，所述获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息之前，还包括：获取所述点击数据、所述前端转化数据和所述后端转化数据；将所述点击数据和\或所述前端转化数据作为第一样本数据输入至参考前端转化率预估模型，确定所述第一样本数据的预估前端转化率，并确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值；将所述前端转化数据和\或所述后端转化数据作为第二样本数据输入至参考后端转化率预估模型，确定所述第二样本数据的预估后端转化率，并确 定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值；至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型。
可选地，所述第一样本数据包括所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率，所述确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值，包括：将所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率代入所述参考前端转化率预估模型的第一损失函数，计算得出第一损失函数的第一函数值，并将所述第一函数值作为所述第一差异评估值；所述第二样本数据包括所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率，所述确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值：将所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率代入所述参考后端转化率预估模型的第二损失函数，计算得出第二损失函数的第二函数值，并将所述第二函数值作为所述第二差异评估值；所述至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练，包括：通过减小所述第一函数值和\或所述第二函数值，调整所述参考前端转化率预估模型的参数，以更新所述参考前端转化率预估模型；通过减小所述第一函数值和\或所述第二函数值，调整所述参考后端转化率预估模型的参数，以更新所述参考后端转化率预估模型。
可选地，所述第一差异评估值为所述第一样本数据的预估前端转化率和真实前端转化率之间的差值；所述第二差异评估值为所述第二样本数据的预估后端转化率和真实后端转化率之间的差值。
可选地，所述获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息之前，还包括：获取所述点击数据、所述前端转化数据和所述后端转化数据；将所述点击数据和\或所述前端转化数据作为第四样本数据输入至参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第四虚拟样本数据；将所述前端转化数据和\或所述后端转化数据作为第五样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第五虚拟样本数据；将所述第四虚拟样本数据输入至参考前端转化率预估模型，确定所述第四虚拟样本数据的预估前端转化率，并确定所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的第四差异评估值；将所述第五虚拟样本数据输入至参考后端转化率预估模型，确定所述第五虚拟样本数据的预估后端转化率，并确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值；至少根据所述第四差异评估值和\ 或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型；将训练结束时的所述参考特征扩展空间，作为所述特定特征扩展空间。
可选地，所述第四虚拟样本数据包括所述第四虚拟样本数据的各特征值，所述确定所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的第四差异评估值，包括：将所述第四虚拟样本数据的各特征值和所述第四样本数据的真实前端转化率代入所述参考前端转化率预估模型的第四损失函数，计算得出所述第四损失函数的第四函数值，并将所述第四函数值作为所述第四差异评估值；所述第五虚拟样本数据包括所述第五虚拟样本数据的各特征值，所述确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值：将所述第五虚拟样本数据的各特征值和所述第五样本数据的真实后端转化率代入所述参考后端转化率预估模型的第五损失函数，计算得出所述第五损失函数的第五函数值，并将所述第五函数值作为所述第五差异评估值；所述至少根据所述第四差异评估值和\或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练，包括：通过减小所述第四函数值和\或所述第五函数值，调整所述参考前端转化率预估模型的参数，从而更新所述参考前端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考后端转化率预估模型的参数，从而更新所述参考后端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考特征扩展空间的参数，从而更新所述参考特征扩展空间的参数。
可选地，所述第四差异评估值为所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的差值；所述第五差异评估值为所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的差值。
可选地，所述获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息之前，还包括：获取所述点击数据和曝光数据：将所述曝光数据和\或所述点击数据作为第六样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第六虚拟样本数据；将所述第六虚拟样本数据输入参考点击率预估模型，确定所述第六虚拟样本数据的预估点击率，并确定所述第六虚拟样本数据的预估点击率和第六样本数据的真实点击率之间的第六差异评估值；至少根据所述第六差异评估值，对所述参考点击率预估模型进行迭代机器训练；将训练结束时的所述参考点击率预估模型，作为特定点击率预估模型。
可选地，将所述用户特性信息和所述资源特征信息输入到特定点击率预估模型，以预 估所述用户被所述资源曝光且发生点击的点击率；所述特定点击率预估模型为根据所述资源推荐平台积累的曝光数据和所述点击数据训练得到的；所述至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率，包括：根据所述点击率、所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
第二方面，本申请提供一种转化预估装置，包括：获取模块，用于获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息；处理模块，用于将所述用户特性信息和所述资源特征信息输入到特定前端转化率预估模型，以预估所述用户点击所述资源且发生前端转化的前端转化率；所述特定前端转化率预估模型为根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的；将所述用户特性信息和所述资源特征信息输入到特定后端转化率预估模型，以预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率；所述特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的；至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
可选地，所述获取模块还用于：获取所述点击数据、所述前端转化数据和所述后端转化数据；所述处理模块还用于：将所述点击数据和\或所述前端转化数据作为第一样本数据输入至参考前端转化率预估模型，确定所述第一样本数据的预估前端转化率，并确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值；将所述前端转化数据和\或所述后端转化数据作为第二样本数据输入至参考后端转化率预估模型，确定所述第二样本数据的预估后端转化率，并确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值；至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型。
可选地，所述第一样本数据包括所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率，所述处理模块具体用于：将所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率代入所述参考前端转化率预估模型的第一损失函数，计算得出第一损失函数的第一函数值，并将所述第一函数值作为所述第一差异评估值；所述第二样本数据包括所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率，所述确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值：将所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率代入所述参考后端转化率预估模型的第二损失函数，计算得出第二损失函数的第二函数值，并将所述第二函数值作为所述第二差异评估值；通过减小所述第一函数值和\或所述第二函数值，调整所述参考前 端转化率预估模型的参数，以更新所述参考前端转化率预估模型；通过减小所述第一函数值和\或所述第二函数值，调整所述参考后端转化率预估模型的参数，以更新所述参考后端转化率预估模型。
可选地，所述第一差异评估值为所述第一样本数据的预估前端转化率和真实前端转化率之间的差值；所述第二差异评估值为所述第二样本数据的预估后端转化率和真实后端转化率之间的差值。
可选地，所述获取模块还用于：获取所述点击数据、所述前端转化数据和所述后端转化数据；所述处理模块还用于：将所述点击数据和\或所述前端转化数据作为第四样本数据输入至参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第四虚拟样本数据；将所述前端转化数据和\或所述后端转化数据作为第五样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第五虚拟样本数据；将所述第四虚拟样本数据输入至参考前端转化率预估模型，确定所述第四虚拟样本数据的预估前端转化率，并确定所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的第四差异评估值；将所述第五虚拟样本数据输入至参考后端转化率预估模型，确定所述第五虚拟样本数据的预估后端转化率，并确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值；至少根据所述第四差异评估值和\或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型；将训练结束时的所述参考特征扩展空间，作为所述特定特征扩展空间。
可选地，所述第四虚拟样本数据包括所述第四虚拟样本数据的各特征值，所述处理模块具体用于：将所述第四虚拟样本数据的各特征值和所述第四样本数据的真实前端转化率代入所述参考前端转化率预估模型的第四损失函数，计算得出所述第四损失函数的第四函数值，并将所述第四函数值作为所述第四差异评估值；所述第五虚拟样本数据包括所述第五虚拟样本数据的各特征值，所述处理模块具体用于：将所述第五虚拟样本数据的各特征值和所述第五样本数据的真实后端转化率代入所述参考后端转化率预估模型的第五损失函数，计算得出所述第五损失函数的第五函数值，并将所述第五函数值作为所述第五差异评估值；通过减小所述第四函数值和\或所述第五函数值，调整所述参考前端转化率预估模型的参数，从而更新所述参考前端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考后端转化率预估模型的参数，从而更新所述参考后端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考特征扩展空间的参数，从而更新所述参考特征扩展空间的参数。
可选地，所述第四差异评估值为所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的差值；所述第五差异评估值为所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的差值。
可选地，所述获取模块还用于：获取所述点击数据和曝光数据；所述处理模块具体用于：将所述曝光数据和\或所述点击数据作为第六样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第六虚拟样本数据；将所述第六虚拟样本数据输入参考点击率预估模型，确定所述第六虚拟样本数据的预估点击率，并确定所述第六虚拟样本数据的预估点击率和第六样本数据的真实点击率之间的第六差异评估值；至少根据所述第六差异评估值，对所述参考点击率预估模型进行迭代机器训练；将训练结束时的所述参考点击率预估模型，作为特定点击率预估模型。
可选地，所述处理模块还用于：将所述用户特性信息和所述资源特征信息输入到特定点击率预估模型，以预估所述用户被所述资源曝光且发生点击的点击率；所述特定点击率预估模型为根据所述资源推荐平台积累的曝光数据和所述点击数据训练得到的；所述处理模块具体用于：根据所述点击率、所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
上述第二方面及第二方面各个实施方式的有益效果，可以参考上述第一方面及第一方面各个实施方式的有益效果，这里不再赘述。
第三方面，本申请提供一种计算机设备，包括程序或指令，当所述程序或指令被执行时，用以执行上述第一方面及第一方面各个实施方式的方法。
第四方面，本申请提供一种存储介质，包括程序或指令，当所述程序或指令被执行时，用以执行上述第一方面及第一方面各个实施方式的方法。
本申请提供一种转化预估方法及装置中，与现有技术的转化预估方法相比，由于所述特定前端转化率预估模型是根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的，所述特定前端转化率预估模型会学习到点击数据和点击且发生前端转化的前端转化数据的知识，另外，所述特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的，特定前端转化率预估模型会学习到前端转化数据和发生前端转化且发生后端转化的后端转化数据的知识，从而在获取了待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息后，特定前端转化率预估模型可以结合学习到的知识预估所述用户点击所述资源且发生前端转化的前端转化率，特定后端转化率预估模型可以结合学习到的知识预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率，从而通过对用户的转化行为进行前端转化后端转化区分，能够分别预估前端转化率和后端转化率，进一步地根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率，从而 能够提升转化预估的准确率。
附图说明
图1为本申请实施例提供的一种转化预估方法的步骤流程示意图；
图2为本申请实施例提供的一种转化预估方法可应用的架构示意图；
图3为本申请实施例提供的一种转化预估装置的结构示意图。
具体实施方式
为了更好的理解上述技术方案，下面将结合说明书附图及具体的实施方式对上述技术方案进行详细的说明，应当理解本申请实施例以及实施例中的具体特征是对本申请技术方案的详细的说明，而不是对本申请技术方案的限定，在不冲突的情况下，本申请实施例以及实施例中的技术特征可以相互结合。
在金融机构(银行机构、保险机构或证券机构)在进行业务(如银行的贷款业务、存款业务等)运转过程中，越来越多的资源提供端对信息推荐效果重视起来，如何预估信息推荐效果尤为重要。目前方法中，将用户的转化行为一概而论，但实际上用户的转化行为又分为多种。现有技术中，直接将用户发生某一类转化行为的预估转化率，作为最终的转化率，显然，这种方式对用户转化行为的转化率并不准确，这不符合银行等金融机构的需求，无法保证金融机构各项业务的高效运转。
为此，如图1所示，本申请实施例提供一种转化预估方法。
步骤101：获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息。
步骤102：将所述用户特性信息和所述资源特征信息输入到特定前端转化率预估模型，以预估所述用户点击所述资源且发生前端转化的前端转化率。
步骤103：将所述用户特性信息和所述资源特征信息输入到特定后端转化率预估模型，以预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率。
步骤104：至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
需要说明的是，前端转化行为是事先定义的一类用户转化行为，用第一类型的转化行为表示，如使用资源提供端的免费服务，如注册；后端转化行为是事先定义的另一类用户转化行为，用第二类型的转化行为表示，后端转化行为需要建立在前端转化行为的基础上，如使用资源提供端的付费服务，如注册后购买虚拟道具，付费服务必然建立在免费服务的基础上。举例来说，用户在某场景点击一个游戏的推荐信息，该用户便为点击用户。而后用户点击该推荐信息后下载注册该游戏，便称为用户发生了前端转化行为。最后，用户在游戏内发生购买行为，便称为用户发生了后端转化行为。显然，而用户发生购买行为时才 是该推荐信息为该游戏产生了真正价值。区分前端转化行为和后端转化行为能够有效地提升转化率的准确率。
需要说明的是，步骤101～步骤104的执行主体可以为资源推荐平台，本申请中的资源提供端为推荐信息的源提供者，需要将推荐信息推送给用户，从而有一部分用户后端转化为资源提供端的后端转化用户，资源推荐平台为推荐信息推送的具体执行者。在资源推荐平台对用户进行信息推荐的过程中，用户按照行为的时间顺序可分为三类：曝光用户、点击用户、前端转化用户和后端转化用户。曝光用户为资源推荐平台推荐了资源的用户；点击用户为点击了资源推荐平台推荐的资源的用户，显然，点击用户必然为曝光用户；前端转化用户为点击了资源推荐平台推荐的资源后在资源提供端发生前端转化行为(如注册行为)的用户，显然，前端转化用户必为点击用户；后端转化用户为发生了前端转化行为后发生后端转化行为的用户。曝光数据至少包括综合特征信息和曝光标签值(如曝光标签值用y 0＝0表示)。点击数据至少包括综合特征信息和点击标签值(如点击标签值用y 0＝1，和\或y 1＝0表示)。前端转化数据至少包括综合特征信息和前端转化标签值(如前端转化标签值用y 1＝1和\或y 2＝0表示)，后端转化数据至少包括综合特征信息和后端转化标签值(如后端转化标签值用y 2＝1表示)。需要说明的是，曝光标签值、点击标签值、转化标签值和真实概率值有映射关系，如点击标签值y 1＝0对应的真实概率即点击率为0％。综合特征信息可以包括多种类型的特征信息，如用户特性信息、资源特征信息和场景特征信息。用户特性信息指用户的基础属性，如年龄、性别等。资源特征信息指推荐资源的基础属性，如推荐资源的格式、布局。场景特征信息指将推荐资源推荐给用户的场景，如推荐的地点、触发推荐的用户操作。另外，用户是否点击、点击后是否发生前端转化、前端转化后是否发生后端转化是独立的，显然，有如下关系：
P(y 2＝1,y 1＝1,y 0＝1|x)＝P(y 0＝1|x)P(y 1＝1|y 0＝1,x)P(y 2＝1|y 1＝1,x)。其中P(y 0＝1|x),P(y 1＝1|y 0＝1,x),P(y 2＝1|y 1＝1,x)分别表示用户曝光后点击的概率(用CTR表示),点击后前端转化的概率(用CVRF表示),前端转化发生后端转化的概率即(用CVRE表示)。
步骤102中，所述特定前端转化率预估模型为根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的。步骤103中，所述特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的。由此可知，特定前端转化率预估模型学习到了点击数据和点击且发生前端转化的前端转化数据的知识，特定后端转化率预估模型学习到了所述前端转化数据和发生前端转化且发生后端转化的后端转化数据的知识；因此，更细化地预估点击数据在前端转化和后端转化过程中的转化率，得到所述前端转化率和所述后端转化率，并进一步确定所述用户点击所述资源发生前端转化且发生后端转化的转化率，从而提升了转化率预估的准确率。
步骤101～步骤104中特定前端转化率预估模型和特定后端转化率预估模型的训练方式有多种，只要所述特定前端转化率预估模型为根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的均可，特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的均可，在此不做限定。
下面以两类实施方式来说明所述特定前端转化率预估模型和所述特定后端转化率预估模型的具体训练过程：
第一类，不进行相似性特征扩展的模型训练方法，具体为：
第(1-1)步：获取所述点击数据、所述前端转化数据和所述后端转化数据。
第(1-2)步：将所述点击数据和\或所述前端转化数据作为第一样本数据输入至参考前端转化率预估模型，确定所述第一样本数据的预估前端转化率，并确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值。
第(1-3)步：将所述前端转化数据和\或所述后端转化数据作为第二样本数据输入至参考后端转化率预估模型，确定所述第二样本数据的预估后端转化率，并确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值。
第(1-4)步：至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练。
第(1-5)步：将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型。
需要说明的是，第(1-2)步～第(1-3)步并不是先后约束的关系，第(1-2)步～第(1-3)步可以并行执行。第(1-2)步～第(1-3)步的过程中，选择性地将第一样本数据和第二样本数据输入到各参考训练模型中，从而得到评估各预估率和真实率的差异评估值，从而能有效表征训练过程中的预估准确性，并进行迭代机器训练，并最终得到特定的模型，从而提供一种同时训练得到所述特定点击率预估模型、所述特定转化率预估模型的方法。
具体地，第(1-1)步～第(1-5)步的过程中：
第(1-4)步中的第一差异评估值可以为所述第一样本数据的预估前端转化率和所述第一样本数据的真实前端转化率之间的差值，这种方式可以最直观地表示所述第一样本数据的预估前端转化率和所述第一样本数据的真实前端转化率的差距。也可以为损失函数值，具体损失函数值的求解方式可以为：
将所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率代入所述参考前端转化率预估模型的第一损失函数，计算得出第一损失函数的第一函数值，并将所述第一函数值作为所述第一差异评估值。举例来说，输入的数据含有特征信息x和标签值y 0和\或y 1和\或y 2，标签值和真实概率值有映射关系，如前端转化标签值y 1＝0对应前端 转化的真实概率为0％。前端转化率预估模型的训练函数为f 1(x|W 1)；前端转化率预估模型的损失函数为L(y 1,p 0p 1)，根据输入的点击标签值所对应的真实点击率和\或前端转化标签值所对应的真实前端转化率和f 1(x|W 1)输出的预估前端转化率，可以得到L(y 1,p 0p 1)的值，即第一函数值为L(y 1,p 0p 1)的值。
第(1-5)步中的第二差异评估值可以为所述第二样本数据的预估后端转化率和所述第二样本数据的真实后端转化率之间的差值，这种方式可以最直观地表示所述第二样本数据的预估后端转化率和所述第二样本数据的真实后端转化率的差距。也可以为损失函数值，具体损失函数值的求解方式可以为：
将所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率代入所述参考后端转化率预估模型的第二损失函数，计算得出第二损失函数的第二函数值，并将所述第二函数值作为所述第二差异评估值。后端转化率预估模型的训练函数为f 2(x|W 2)；后端转化率预估模型的损失函数为L(y 2,p 0p 1p 2)，根据输入的前端转化标签值所对应的真实前端转化率和\或后端转化标签值所对应的真实后端转化率和f 2(x|W 2)输出的预估后端转化率，可以得到L(y 2,p 0p 1p 2)的值，即第二函数值为L(y 2,p 0p 1p 2)的值。
第(1-6)步具体可以为：
通过减小所述第一函数值和\或所述第二函数值，调整所述参考前端转化率预估模型的参数，以更新所述参考前端转化率预估模型；通过减小所述第一函数值和\或所述第二函数值，调整所述参考后端转化率预估模型的参数，以更新所述参考后端转化率预估模型。
举例来说，通过L(y 1,p 0p 1)和L(y 2,p 0p 1p 2)来调整W 1；通过L(y 1,p 0p 1)和L(y 2,p 0p 1p 2)来调整W 2，即联合损失函数为：L＝L(y 1,p 0p 1)+L(y 2,p 0p 1p 2)，也可以为L＝L(y 1,p 0p 1)+L(y 2,p 0p 1p 2)+R；R表示为了避免过拟合的正则项。
需要说明的是，结合图2来说，在第一类模型训练方法中，还可以加入点击预估模型，具体来说：获取曝光数据和所述点击数据；将所述曝光数据和\或所述点击数据作为第三样本数据输入至参考点击率预估模型，确定所述第三样本数据的预估点击率，并确定所述第三样本数据的预估点击率和真实点击率之间的第三差异评估值；至少根据所述第三差异评估值，对所述参考点击率预估模型进行迭代机器训练；将训练结束时的所述参考点击率预估模型，作为所述特定点击率预估模型。
举例来说，点击率预估模型的训练函数为f 0(x|W 0)，根据输入的曝光标签值所对应的真实曝光率和\或点击标签值所对应的真实点击率和f 0(x|W 0)输出的预估点击率，可以得到L(y 0,p 0)的值，即第三函数值为L(y 0,p 0)的值。那么可以在上述联合损失函数中加入点击率预估模型的损失函数L(y 0,p 0)。即L＝L(y 0,p 0)+L(y 1,p 0p 1)+L(y 2,p 0p 1p 2)。
第二类，进行相似性特征扩展的模型训练方法，具体为：
第(2-1)步：获取所述点击数据、所述前端转化数据和所述后端转化数据。
第(2-2)步：将所述点击数据和\或所述前端转化数据作为第四样本数据输入至参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第四虚拟样本数据。
第(2-3)步：将所述前端转化数据和\或所述后端转化数据作为第五样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第五虚拟样本数据。
第(2-4)步：将所述第四虚拟样本数据输入至参考前端转化率预估模型，确定所述第四虚拟样本数据的预估前端转化率，并确定所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的第四差异评估值。
第(2-5)步：将所述第五虚拟样本数据输入至参考后端转化率预估模型，确定所述第五虚拟样本数据的预估后端转化率，并确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值。
第(2-6)步：至少根据所述第四差异评估值和\或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练。
第(2-7)步：将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型；将训练结束时的所述参考特征扩展空间，作为所述特定特征扩展空间。
需要说明的是，第(2-2)步～第(2-3)步并不是先后约束的关系，第(2-2)步～第(2-3)步可以并行执行；第(2-4)步～第(2-5)步并不是先后约束的关系，第(2-4)步～第(2-5)步可以并行执行。第(2-2)步～第(2-5)步的过程中，选择性地将第一样本数据和第二样本数据输入参考特征扩展空间，在参考特征扩展空间根据不同的数据类型将第一样本数据、第二样本数据进行转换，得到第一虚拟样本数据、第二虚拟样本数据，从而得到评估各预估率和真实率的差异评估值，从而能有效表征训练过程中的预估准确性，并进行迭代机器训练，并最终得到特定的模型，从而提供一种同时训练得到所述特定点击率预估模型、所述特定转化率预估模型的方法。
具体地，第(2-1)步～第(2-7)步的过程中：
第(2-4)步中的第四差异评估值可以为所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的差值，这种方式可以最直观地表示所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率的差距。也可以为损失函数值，具体损失函数值的求解方式可以为：
将所述第四虚拟样本数据的各特征值和所述第四样本数据的真实前端转化率代入所述参考前端转化率预估模型的第四损失函数，计算得出第四损失函数的第四函数值，并将所述第四函数值作为所述第四差异评估值。
举例来说，输入的数据含有特征信息x和标签值y 0和\或y 1和\或y 2，标签值和真实概率值有映射关系，如前端转化标签值y 1＝0对应前端转化的真实概率为0％。参考特征扩展空间的对输入数据的训练函数为f e(x|W e)，输入的数据为x，参考特征扩展空间的参数为W e，x可以为第一样本数据、第二样本数据任一种数据，输出e为第一虚拟样本数据、第二虚拟样本数据。结合参考特定特征扩展空间，可以学习参考前端转化率预估模型的训练函数f 1(x|W 1,W e)和参考后端转化率预估模型的训练函数f 2(x|W 2,W e)。前端转化率预估模型的损失函数为L(y 1,p 0p 1)，根据输入的点击标签值所对应的真实点击率和\或前端转化标签值所对应的真实前端转化率和f 1(x|W 1,W e)输出的预估前端转化率，可以得到L(y 1,p 0p 1)的值，即第四函数值为L(y 1,p 0p 1)的值。
第(2-5)步中的第五差异评估值可以为所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的差值，这种方式可以最直观地表示所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率的差距。也可以为损失函数值，具体损失函数值的求解方式可以为：
将所述第五虚拟样本数据的各特征值和所述第五样本数据的真实后端转化率代入所述参考后端转化率预估模型的第五损失函数，计算得出第五损失函数的第五函数值，并将所述第五函数值作为所述第五差异评估值。
举例来说，后端转化率预估模型的训练函数为f 2(x|W 2,W e)；后端转化率预估模型的损失函数为L(y 2,p 0p 1p 2)，根据输入的前端转化标签值所对应的真实前端转化率和\或后端转化标签值所对应的真实后端转化率和f 2(x|W 2,W e)输出的预估后端转化率，可以得到L(y 2,p 0p 1p 2)的值，即第五函数值为L(y 2,p 0p 1p 2)的值。
第(2-6)步具体可以为：
通过减小所述第四函数值和\或所述第五函数值，调整所述参考前端转化率预估模型的参数，从而更新所述参考前端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考后端转化率预估模型的参数，从而更新所述参考后端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考特征扩展空间的参数，从而更新所述参考特征扩展空间的参数。
举例来说，通过L(y 1,p 0p 1)和L(y 2,p 0p 1p 2)来调整W 1；通过L(y 1,p 0p 1)和L(y 2,p 0p 1p 2)来调整W 2，通过L(y 1,p 0p 1)和L(y 2,p 0p 1p 2)来调整W e，即联合损失函数为：
L＝L(y 1,p 0p 1)+L(y 2,p 0p 1p 2)，变形也可以为L＝L(y 1,p 0p 1)+L(y 2,p 0p 1p 2)+R；R表示为了避免过拟合的正则项。
需要说明的是，结合图2来说，在第二类模型训练方法中，还可以加入点击预估模型，具体来说：获取所述点击数据和曝光数据；将所述曝光数据和\或所述点击数据作为第六样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化 得到的第六虚拟样本数据；将所述第六虚拟样本数据输入参考点击率预估模型，确定所述第六虚拟样本数据的预估点击率，并确定所述第六虚拟样本数据的预估点击率和第六样本数据的真实点击率之间的第六差异评估值；至少根据所述第六差异评估值，对所述参考点击率预估模型进行迭代机器训练；
将训练结束时的所述参考点击率预估模型，作为特定点击率预估模型。
举例来说，点击率预估模型的训练函数为f 0(x|W 0,W e)，根据输入的曝光标签值所对应的真实曝光率和\或点击标签值所对应的真实点击率和f 0(x|W 0,W e)输出的预估点击率，可以得到L(y 0,p 0)的值，即第六函数值为L(y 0,p 0)的值。那么可以在上述联合损失函数中加入点击率预估模型的损失函数L(y 0,p 0)。即L＝L(y 0,p 0)+L(y 1,p 0p 1)+L(y 2,p 0p 1p 2)，也可以加上R。
步骤101～步骤104之外的一种可选实施方式为：
将所述用户特性信息和所述资源特征信息输入到特定点击率预估模型，以预估所述用户被所述资源曝光且发生点击的点击率；所述特定点击率预估模型为根据所述资源推荐平台积累的曝光数据和所述点击数据训练得到的；根据所述点击率、所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
举例来说，前端转化率为0.8和所述后端转化率为0.6，可以用前端转化率和所述后端转化率的乘积0.48来作为所述用户点击所述资源发生前端转化且发生后端转化的转化率。
如图3所示，本申请提供一种转化预估装置，包括：获取模块301，用于获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息；处理模块302，用于将所述用户特性信息和所述资源特征信息输入到特定前端转化率预估模型，以预估所述用户点击所述资源且发生前端转化的前端转化率；所述特定前端转化率预估模型为根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的；将所述用户特性信息和所述资源特征信息输入到特定后端转化率预估模型，以预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率；所述特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的；至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
可选地，所述获取模块301还用于：获取所述点击数据、所述前端转化数据和所述后端转化数据；所述处理模块302还用于：将所述点击数据和\或所述前端转化数据作为第一样本数据输入至参考前端转化率预估模型，确定所述第一样本数据的预估前端转化率，并确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值；将所述前端转化数据和\或所述后端转化数据作为第二样本数据输入至参考后端转化率预估模型，确定所述第二样本数据的预估后端转化率，并确定所述第二样本数据的预估后端转化 率和真实后端转化率之间的第二差异评估值；至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型。
可选地，所述第一样本数据包括所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率，所述处理模块302具体用于：将所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率代入所述参考前端转化率预估模型的第一损失函数，计算得出第一损失函数的第一函数值，并将所述第一函数值作为所述第一差异评估值；所述第二样本数据包括所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率，所述确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值：将所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率代入所述参考后端转化率预估模型的第二损失函数，计算得出第二损失函数的第二函数值，并将所述第二函数值作为所述第二差异评估值；通过减小所述第一函数值和\或所述第二函数值，调整所述参考前端转化率预估模型的参数，以更新所述参考前端转化率预估模型；通过减小所述第一函数值和\或所述第二函数值，调整所述参考后端转化率预估模型的参数，以更新所述参考后端转化率预估模型。
可选地，所述第一差异评估值为所述第一样本数据的预估前端转化率和真实前端转化率之间的差值；所述第二差异评估值为所述第二样本数据的预估后端转化率和真实后端转化率之间的差值。
可选地，所述获取模块301还用于：获取所述点击数据、所述前端转化数据和所述后端转化数据；所述处理模块302还用于：将所述点击数据和\或所述前端转化数据作为第四样本数据输入至参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第四虚拟样本数据；将所述前端转化数据和\或所述后端转化数据作为第五样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第五虚拟样本数据；将所述第四虚拟样本数据输入至参考前端转化率预估模型，确定所述第四虚拟样本数据的预估前端转化率，并确定所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的第四差异评估值；将所述第五虚拟样本数据输入至参考后端转化率预估模型，确定所述第五虚拟样本数据的预估后端转化率，并确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值；至少根据所述第四差异评估值和\或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型； 将训练结束时的所述参考特征扩展空间，作为所述特定特征扩展空间。
可选地，所述第四虚拟样本数据包括所述第四虚拟样本数据的各特征值，所述处理模块302具体用于：将所述第四虚拟样本数据的各特征值和所述第四样本数据的真实前端转化率代入所述参考前端转化率预估模型的第四损失函数，计算得出所述第四损失函数的第四函数值，并将所述第四函数值作为所述第四差异评估值；所述第五虚拟样本数据包括所述第五虚拟样本数据的各特征值，所述处理模块302具体用于：将所述第五虚拟样本数据的各特征值和所述第五样本数据的真实后端转化率代入所述参考后端转化率预估模型的第五损失函数，计算得出所述第五损失函数的第五函数值，并将所述第五函数值作为所述第五差异评估值；通过减小所述第四函数值和\或所述第五函数值，调整所述参考前端转化率预估模型的参数，从而更新所述参考前端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考后端转化率预估模型的参数，从而更新所述参考后端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考特征扩展空间的参数，从而更新所述参考特征扩展空间的参数。
可选地，所述第四差异评估值为所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的差值；所述第五差异评估值为所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的差值。
可选地，所述获取模块301还用于：获取所述点击数据和曝光数据；所述处理模块302具体用于：将所述曝光数据和\或所述点击数据作为第六样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第六虚拟样本数据；将所述第六虚拟样本数据输入参考点击率预估模型，确定所述第六虚拟样本数据的预估点击率，并确定所述第六虚拟样本数据的预估点击率和第六样本数据的真实点击率之间的第六差异评估值；至少根据所述第六差异评估值，对所述参考点击率预估模型进行迭代机器训练；将训练结束时的所述参考点击率预估模型，作为特定点击率预估模型。
可选地，所述处理模块302还用于：将所述用户特性信息和所述资源特征信息输入到特定点击率预估模型，以预估所述用户被所述资源曝光且发生点击的点击率；所述特定点击率预估模型为根据所述资源推荐平台积累的曝光数据和所述点击数据训练得到的；所述处理模块具体用于：根据所述点击率、所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
本申请实施例提供一种计算机设备，包括程序或指令，当所述程序或指令被执行时，用以执行本申请实施例提供的一种转化预估方法及任一可选方法。
本申请实施例提供一种存储介质，包括程序或指令，当所述程序或指令被执行时，用以执行本申请实施例提供的一种转化预估方法及任一可选方法。
最后应说明的是：本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、 或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
显然，本领域的技术人员可以对本申请进行各种改动和变型而不脱离本申请的范围。这样，倘若本申请的这些修改和变型属于本申请权利要求及其等同技术的范围之内，则本申请也意图包含这些改动和变型在内。

Claims (20)

  1. 一种转化预估方法，其特征在于，包括：
    获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息；
    将所述用户特性信息和所述资源特征信息输入到特定前端转化率预估模型，以预估所述用户点击所述资源且发生前端转化的前端转化率；所述特定前端转化率预估模型为根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的；
    将所述用户特性信息和所述资源特征信息输入到特定后端转化率预估模型，以预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率；所述特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的；
    至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
  2. 如权利要求1所述的方法，其特征在于，所述获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息之前，还包括：
    获取所述点击数据、所述前端转化数据和所述后端转化数据；
    将所述点击数据和\或所述前端转化数据作为第一样本数据输入至参考前端转化率预估模型，确定所述第一样本数据的预估前端转化率，并确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值；
    将所述前端转化数据和\或所述后端转化数据作为第二样本数据输入至参考后端转化率预估模型，确定所述第二样本数据的预估后端转化率，并确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值；
    至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练；
    将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型。
  3. 如权利要求2所述的方法，其特征在于，所述第一样本数据包括所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率，所述确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值，包括：
    将所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率代入所述参考前端转化率预估模型的第一损失函数，计算得出第一损失函数的第一函数值，并将所述第一函数值作为所述第一差异评估值；
    所述第二样本数据包括所述第二样本数据的各特征值和所述第二样本数据的真实后 端转化率，所述确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值，包括：
    将所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率代入所述参考后端转化率预估模型的第二损失函数，计算得出第二损失函数的第二函数值，并将所述第二函数值作为所述第二差异评估值；
    所述至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练，包括：
    通过减小所述第一函数值和\或所述第二函数值，调整所述参考前端转化率预估模型的参数，以更新所述参考前端转化率预估模型；通过减小所述第一函数值和\或所述第二函数值，调整所述参考后端转化率预估模型的参数，以更新所述参考后端转化率预估模型。
  4. 如权利要求2所述的方法，其特征在于，所述第一差异评估值为所述第一样本数据的预估前端转化率和真实前端转化率之间的差值；所述第二差异评估值为所述第二样本数据的预估后端转化率和真实后端转化率之间的差值。
  5. 如权利要求1所述的方法，其特征在于，所述获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息之前，还包括：
    获取所述点击数据、所述前端转化数据和所述后端转化数据；
    将所述点击数据和\或所述前端转化数据作为第四样本数据输入至参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第四虚拟样本数据；
    将所述前端转化数据和\或所述后端转化数据作为第五样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第五虚拟样本数据；
    将所述第四虚拟样本数据输入至参考前端转化率预估模型，确定所述第四虚拟样本数据的预估前端转化率，并确定所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的第四差异评估值；
    将所述第五虚拟样本数据输入至参考后端转化率预估模型，确定所述第五虚拟样本数据的预估后端转化率，并确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值；
    至少根据所述第四差异评估值和\或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练；
    将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；
    将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型；
    将训练结束时的所述参考特征扩展空间，作为所述特定特征扩展空间。
  6. 如权利要求5所述的方法，其特征在于，所述第四虚拟样本数据包括所述第四虚拟样本数据的各特征值，所述确定所述第四虚拟样本数据的预估前端转化率和所述第四样 本数据的真实前端转化率之间的第四差异评估值，包括：
    将所述第四虚拟样本数据的各特征值和所述第四样本数据的真实前端转化率代入所述参考前端转化率预估模型的第四损失函数，计算得出所述第四损失函数的第四函数值，并将所述第四函数值作为所述第四差异评估值；
    所述第五虚拟样本数据包括所述第五虚拟样本数据的各特征值，所述确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值，包括：
    将所述第五虚拟样本数据的各特征值和所述第五样本数据的真实后端转化率代入所述参考后端转化率预估模型的第五损失函数，计算得出所述第五损失函数的第五函数值，并将所述第五函数值作为所述第五差异评估值；
    所述至少根据所述第四差异评估值和\或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练，包括：
    通过减小所述第四函数值和\或所述第五函数值，调整所述参考前端转化率预估模型的参数，从而更新所述参考前端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考后端转化率预估模型的参数，从而更新所述参考后端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考特征扩展空间的参数，从而更新所述参考特征扩展空间的参数。
  7. 如权利要求5所述的方法，其特征在于，所述第四差异评估值为所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的差值；所述第五差异评估值为所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的差值。
  8. 如权利要求5所述的方法，其特征在于，所述获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息之前，还包括：
    获取所述点击数据和曝光数据；
    将所述曝光数据和\或所述点击数据作为第六样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第六虚拟样本数据；
    将所述第六虚拟样本数据输入参考点击率预估模型，确定所述第六虚拟样本数据的预估点击率，并确定所述第六虚拟样本数据的预估点击率和第六样本数据的真实点击率之间的第六差异评估值；
    至少根据所述第六差异评估值，对所述参考点击率预估模型进行迭代机器训练；
    将训练结束时的所述参考点击率预估模型，作为特定点击率预估模型。
  9. 权利要求1-8任一所述的方法，其特征在于，还包括：
    将所述用户特性信息和所述资源特征信息输入到特定点击率预估模型，以预估所述用户被所述资源曝光且发生点击的点击率；所述特定点击率预估模型为根据所述资源推荐平台积累的曝光数据和所述点击数据训练得到的；
    所述至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率，包括：
    根据所述点击率、所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
  10. 一种转化预估装置，其特征在于，包括：
    获取模块，用于获取待评估用户的用户特性信息和所述用户待被曝光资源的资源特征信息；
    处理模块，用于将所述用户特性信息和所述资源特征信息输入到特定前端转化率预估模型，以预估所述用户点击所述资源且发生前端转化的前端转化率；所述特定前端转化率预估模型为根据资源推荐平台积累的点击数据和点击且发生前端转化的前端转化数据训练得到的；将所述用户特性信息和所述资源特征信息输入到特定后端转化率预估模型，以预估针对所述资源所述用户发生前端转化且发生后端转化的后端转化率；所述特定后端转化率预估模型为根据所述前端转化数据和发生前端转化且发生后端转化的后端转化数据训练得到的；至少根据所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
  11. 如权利要求10所述的装置，其特征在于，所述获取模块还用于：
    获取所述点击数据、所述前端转化数据和所述后端转化数据；
    所述处理模块还用于：
    将所述点击数据和\或所述前端转化数据作为第一样本数据输入至参考前端转化率预估模型，确定所述第一样本数据的预估前端转化率，并确定所述第一样本数据的预估前端转化率和真实前端转化率之间的第一差异评估值；将所述前端转化数据和\或所述后端转化数据作为第二样本数据输入至参考后端转化率预估模型，确定所述第二样本数据的预估后端转化率，并确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值；至少根据所述第一差异评估值和\或所述第二差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型。
  12. 如权利要求11所述的装置，其特征在于，所述第一样本数据包括所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率，所述处理模块具体用于：
    将所述第一样本数据的各特征值和所述第一样本数据的真实前端转化率代入所述参 考前端转化率预估模型的第一损失函数，计算得出第一损失函数的第一函数值，并将所述第一函数值作为所述第一差异评估值；所述第二样本数据包括所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率，所述确定所述第二样本数据的预估后端转化率和真实后端转化率之间的第二差异评估值：将所述第二样本数据的各特征值和所述第二样本数据的真实后端转化率代入所述参考后端转化率预估模型的第二损失函数，计算得出第二损失函数的第二函数值，并将所述第二函数值作为所述第二差异评估值；通过减小所述第一函数值和\或所述第二函数值，调整所述参考前端转化率预估模型的参数，以更新所述参考前端转化率预估模型；通过减小所述第一函数值和\或所述第二函数值，调整所述参考后端转化率预估模型的参数，以更新所述参考后端转化率预估模型。
  13. 如权利要求11所述的装置，其特征在于，所述第一差异评估值为所述第一样本数据的预估前端转化率和真实前端转化率之间的差值；所述第二差异评估值为所述第二样本数据的预估后端转化率和真实后端转化率之间的差值。
  14. 如权利要求10所述的装置，其特征在于，所述获取模块还用于：获取所述点击数据、所述前端转化数据和所述后端转化数据；所述处理模块还用于：
    将所述点击数据和\或所述前端转化数据作为第四样本数据输入至参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第四虚拟样本数据；将所述前端转化数据和\或所述后端转化数据作为第五样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第五虚拟样本数据；将所述第四虚拟样本数据输入至参考前端转化率预估模型，确定所述第四虚拟样本数据的预估前端转化率，并确定所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的第四差异评估值；将所述第五虚拟样本数据输入至参考后端转化率预估模型，确定所述第五虚拟样本数据的预估后端转化率，并确定所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的第五差异评估值；至少根据所述第四差异评估值和\或所述第五差异评估值，对所述参考前端转化率预估模型和\或所述参考后端转化率预估模型和\或所述参考特征扩展空间进行迭代机器训练；将训练结束时的所述参考前端转化率预估模型，作为所述特定前端转化率预估模型；将训练结束时的所述参考后端转化率预估模型，作为所述特定后端转化率预估模型；将训练结束时的所述参考特征扩展空间，作为所述特定特征扩展空间。
  15. 如权利要求14所述的装置，其特征在于，所述第四虚拟样本数据包括所述第四虚拟样本数据的各特征值，所述处理模块具体用于：
    将所述第四虚拟样本数据的各特征值和所述第四样本数据的真实前端转化率代入所述参考前端转化率预估模型的第四损失函数，计算得出所述第四损失函数的第四函数值，并将所述第四函数值作为所述第四差异评估值；所述第五虚拟样本数据包括所述第五虚拟 样本数据的各特征值，所述处理模块具体用于：将所述第五虚拟样本数据的各特征值和所述第五样本数据的真实后端转化率代入所述参考后端转化率预估模型的第五损失函数，计算得出所述第五损失函数的第五函数值，并将所述第五函数值作为所述第五差异评估值；通过减小所述第四函数值和\或所述第五函数值，调整所述参考前端转化率预估模型的参数，从而更新所述参考前端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考后端转化率预估模型的参数，从而更新所述参考后端转化率预估模型的参数；和\或通过减小所述第四函数值和\或所述第五函数值，调整所述参考特征扩展空间的参数，从而更新所述参考特征扩展空间的参数。
  16. 如权利要求14所述的装置，其特征在于，所述第四差异评估值为所述第四虚拟样本数据的预估前端转化率和所述第四样本数据的真实前端转化率之间的差值；所述第五差异评估值为所述第五虚拟样本数据的预估后端转化率和所述第五样本数据的真实后端转化率之间的差值。
  17. 如权利要求14所述的装置，其特征在于，所述获取模块还用于：获取所述点击数据和曝光数据；所述处理模块具体用于：
    将所述曝光数据和\或所述点击数据作为第六样本数据输入至所述参考特征扩展空间进行相似性扩展，获得经所述参考特征扩展空间转化得到的第六虚拟样本数据；将所述第六虚拟样本数据输入参考点击率预估模型，确定所述第六虚拟样本数据的预估点击率，并确定所述第六虚拟样本数据的预估点击率和第六样本数据的真实点击率之间的第六差异评估值；至少根据所述第六差异评估值，对所述参考点击率预估模型进行迭代机器训练；将训练结束时的所述参考点击率预估模型，作为特定点击率预估模型。
  18. 如权利要求14-17任一所述的装置，其特征在于，所述处理模块还用于：
    将所述用户特性信息和所述资源特征信息输入到特定点击率预估模型，以预估所述用户被所述资源曝光且发生点击的点击率；所述特定点击率预估模型为根据所述资源推荐平台积累的曝光数据和所述点击数据训练得到的；根据所述点击率、所述前端转化率和所述后端转化率，确定所述用户点击所述资源发生前端转化且发生后端转化的转化率。
  19. 一种计算机设备，其特征在于，包括程序或指令，当所述程序或指令被执行时，如权利要求1至9中任意一项所述的方法被执行。
  20. 一种存储介质，其特征在于，包括程序或指令，当所述程序或指令被执行时，如权利要求1至9中任意一项所述的方法被执行。
PCT/CN2019/127220 2019-12-20 2019-12-20 一种转化预估方法及装置 WO2021120226A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/127220 WO2021120226A1 (zh) 2019-12-20 2019-12-20 一种转化预估方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2019/127220 WO2021120226A1 (zh) 2019-12-20 2019-12-20 一种转化预估方法及装置

Publications (1)

Publication Number Publication Date
WO2021120226A1 true WO2021120226A1 (zh) 2021-06-24

Family

ID=76477048

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/127220 WO2021120226A1 (zh) 2019-12-20 2019-12-20 一种转化预估方法及装置

Country Status (1)

Country Link
WO (1) WO2021120226A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120030011A1 (en) * 2010-07-30 2012-02-02 Yahoo! Inc. Systems and Methods for Estimating a Conversion Rate for a Digital Advertisement Based on Dwell Times Associated with the Digital Advertisement
US20130339126A1 (en) * 2012-06-13 2013-12-19 Yahoo! Inc. Campaign performance forecasting for non-guaranteed delivery advertising
CN106844178A (zh) * 2017-01-22 2017-06-13 腾云天宇科技（北京）有限公司 预测呈现信息转化率的方法、计算设备、服务器及系统
CN110008399A (zh) * 2019-01-30 2019-07-12 阿里巴巴集团控股有限公司 一种推荐模型的训练方法及装置、一种推荐方法及装置
CN110400169A (zh) * 2019-07-02 2019-11-01 阿里巴巴集团控股有限公司 一种信息推送方法、装置及设备

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20120030011A1 (en) * 2010-07-30 2012-02-02 Yahoo! Inc. Systems and Methods for Estimating a Conversion Rate for a Digital Advertisement Based on Dwell Times Associated with the Digital Advertisement
US20130339126A1 (en) * 2012-06-13 2013-12-19 Yahoo! Inc. Campaign performance forecasting for non-guaranteed delivery advertising
CN106844178A (zh) * 2017-01-22 2017-06-13 腾云天宇科技（北京）有限公司 预测呈现信息转化率的方法、计算设备、服务器及系统
CN110008399A (zh) * 2019-01-30 2019-07-12 阿里巴巴集团控股有限公司 一种推荐模型的训练方法及装置、一种推荐方法及装置
CN110400169A (zh) * 2019-07-02 2019-11-01 阿里巴巴集团控股有限公司 一种信息推送方法、装置及设备

Similar Documents

Publication Publication Date Title
JP7127120B2 (ja) ビデオ分類の方法、情報処理の方法及びサーバー、並びにコンピュータ可読記憶媒体及びコンピュータプログラム
US20220198289A1 (en) Recommendation model training method, selection probability prediction method, and apparatus
US11983645B2 (en) Agent aptitude prediction
WO2022267735A1 (zh) 业务数据处理方法、装置、计算机设备和存储介质
WO2021164382A1 (zh) 针对用户分类模型进行特征处理的方法及装置
US20180268318A1 (en) Training classification algorithms to predict end-user behavior based on historical conversation data
US20190272553A1 (en) Predictive Modeling with Entity Representations Computed from Neural Network Models Simultaneously Trained on Multiple Tasks
CN108777701B (zh) 一种确定信息受众的方法及装置
CN109035028B (zh) 智能投顾策略生成方法及装置、电子设备、存储介质
US10515378B2 (en) Extracting relevant features from electronic marketing data for training analytical models
CN112967112B (zh) 一种自注意力机制和图神经网络的电商推荐方法
CN111160638B (zh) 一种转化预估方法及装置
CN111080338B (zh) 用户数据的处理方法、装置、电子设备及存储介质
WO2020233432A1 (zh) 一种信息推荐方法及装置
US20190080352A1 (en) Segment Extension Based on Lookalike Selection
CN111159241B (zh) 一种点击转化预估方法及装置
CN111291936B (zh) 产品生命周期预估模型生成方法、装置及电子设备
CN113610610B (zh) 基于图神经网络和评论相似度的会话推荐方法和系统
CN115099310A (zh) 训练模型、对企业进行行业分类的方法和装置
WO2022111095A1 (zh) 一种产品推荐方法、装置、计算机存储介质及系统
WO2021120226A1 (zh) 一种转化预估方法及装置
TWI792101B (zh) 基於確定值及預測值的數據定量化方法
US20200302550A1 (en) Cost allocation estimation using direct cost vectors and machine learning
US20160147816A1 (en) Sample selection using hybrid clustering and exposure optimization
CN110880141A (zh) 一种深度双塔模型智能匹配算法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19956538

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 12/10/2022)

122 Ep: pct application non-entry in european phase

Ref document number: 19956538

Country of ref document: EP

Kind code of ref document: A1