A conversion rate model updating method, device, electronic device and storage medium

Abstract

The present invention discloses a conversion rate model updating method, device, electronic device and storage medium, the method comprising obtaining a training set, the training set comprising training samples with labels, obtaining label flipping samples, summarizing the gradients of all label flipping samples and estimating the first parameter change of a CVR model by calculation; obtaining new arrival data between the training deadline and the test time, the new arrival data comprising new samples, using the new samples to input the CVR model to obtain the model parameter approximation of the new samples, and obtaining the second parameter change brought by the new arrival data by calculating the model parameter approximation; obtaining the total parameter change according to the first parameter change and the second parameter change, and updating the CVR model based on the total parameter change. The present invention realizes real-time updating of the model by calculating the influence of new data and delayed conversion data on the model parameters, avoiding retraining the model, thereby improving prediction accuracy and model adaptability.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and specifically to a conversion rate model updating method, device, electronic device and storage medium. Background Art

[0002] In online digital advertising, predicting conversion rate (CVR) is crucial to maximizing revenue under the cost-per-acquisition (CPA) model. In this model, advertisers pay only after users complete a specific action, such as a purchase. A major challenge in CVR prediction is the delayed feedback problem, where a conversion may occur hours or weeks after a user clicks. This delay complicates the training of prediction models, leading to bias and suboptimal performance. Although existing methods have addressed this issue to some extent, they often have difficulty coping with changes in user interests and rely on additional auxiliary models, which leads to computational inefficiencies and model confusion.

[0003] Current solutions to delayed feedback are mainly divided into offline and online methods. Offline methods rely on historical data to train the model, and usually include an additional module to model the distribution of delay time or adjust incorrect labels. They usually assume that the distribution of future data is consistent with historical patterns; while online methods try to solve this problem by utilizing new click data and updating the model in real time. These updates usually involve correcting the labels of duplicate data or approximately optimizing the loss function to approach the ideal model.

[0004] Although these approaches have some effectiveness, current methods have two major limitations:

[0005] 1. Insufficient adaptation to changes in user interests. The dynamic changes in user interests pose a major challenge. Offline methods often fail to capture users' latest behaviors and preferences when dealing with changes in data distribution. Although online methods adapt to new data more quickly, the sample duplication introduced during the label correction process may cause model confusion, thus affecting the effective integration of true label information.

[0006] 2. Dependence on auxiliary models. Mainstream methods often contain auxiliary components to estimate the potential label reversal of observed negative samples or the probability of simulating false negative samples. However, the development complexity of these additional components is comparable to building the main CVR model, and their effectiveness is limited by the quality and quantity of historical data. This dependency may lead to computational inefficiency and increase the complexity of the CVR prediction task. Summary of the invention

[0007] In this embodiment, a conversion rate model updating method, device, electronic device and storage medium are provided to solve the problems of poor adaptability of the model to delayed feedback and insufficient adaptability to changes in user interests.

[0008] In a first aspect, the present invention provides a conversion rate model updating method, comprising the following steps:

[0009] Obtaining a training set, wherein the training set includes labeled training samples, and the training samples include clicks and conversion numbers on an online advertising platform;

[0010] Using the pre-trained CVR model to make preliminary predictions on the training samples and make preliminary adjustments to the model parameters;

[0011] Obtain label flip samples, wherein the label flip samples include label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples, and estimate the change of the first parameter of the CVR model by calculation;

[0012] Acquire new arrival data between the training deadline and the test time, wherein the new arrival data includes new samples, use the new samples to input the CVR model to obtain model parameter approximations of the new samples, and obtain a second parameter change caused by the new arrival data by calculating the model parameter approximation;

[0013] A total parameter change is obtained according to the first parameter change and the second parameter change, and the CVR model is updated based on the total parameter change.

[0014] In an optional embodiment, obtaining label flip samples, wherein the label flip samples include label flip samples during training and label flip samples after training, summarizing the gradients of all label flip samples and estimating the change of the first parameter of the CVR model by calculation includes:

[0015] Model label flipping as data perturbation, use label flipped training samples as samples after data perturbation, and construct optimization objectives based on the perturbed samples:

[0016] ;

[0017] In the formula, are the trained model parameters, is a training sample, For training samples Increase the disturbance to A sample of It is a sample Adjusted weights, is the basic loss function. Specifically, ; is the binary cross entropy loss function, specifically, ; It is a sample Optimal parameters after weight adjustment;

[0018] From the above formula, we can deduce:

[0019] ;

[0020] in, Yes Seeking guidance, refers to the second-order gradient, is the original loss function The optimal parameters of is the first-order gradient;

[0021] Summarize the gradients of all label flipped samples and use the following formula to estimate the change of the first parameter of the CVR model:

[0022] ;

[0023] In the formula, is the first parameter change, n is the total number of training samples in the training set, is Before clicking and The set of sample indexes converted within the interval, where is the training deadline, It's testing time.

[0024] In an optional embodiment, newly arrived data is obtained between the training deadline and the test time, the newly arrived data including new samples, the new samples are input into the CVR model to obtain model parameter approximation of the new samples, and the second parameter change amount brought by the newly arrived data is obtained by calculating the model parameter approximation, including:

[0025] The new samples in the newly arrived data Substitute the following new perturbation loss function:

[0026] ;

[0027] in, , ;

[0028] New samples are obtained by Taylor expansion of the new perturbation loss function The change in model parameters is approximately:

[0029] ;

[0030] in, For new samples and the degree of disturbance The parameter change amount; Taking the optimal parameters The inverse of the second-order gradient of all samples, It is to take the optimal parameters New Sample The first-order gradient of

[0031] The change in the second parameter caused by the newly arrived data is obtained through model parameter approximation.

[0032] In an optional embodiment, obtaining the second parameter variation caused by the newly arrived data through model parameter approximation includes:

[0033] Will The value is Substitute the new disturbance loss function Taylor expansion to obtain the change of the second parameter caused by the new arrival data :

[0034] .

[0035] In an optional embodiment, obtaining a total parameter change according to the first parameter change and the second parameter change includes:

[0036] The total parameter change is calculated by the following formula:

[0037] ;

[0038] In the formula, is the parameter change caused by the newly arrived data;

[0039] Specifically, the total parameter change is estimated by the following formula:

[0040] ;

[0041] in, and are the index sets of label flipped samples and new samples respectively,

[0042] Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient, Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient, Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient.

[0043] In an optional embodiment, updating the CVR model based on the total parameter change includes:

[0044] The total parameter change is taken as the solution of the following linear system:

[0045] ;

[0046] in, It is to take the optimal parameters The second-order gradient of all samples of is the variable to be solved, is defined as follows:

[0047] ;

[0048] Solving the above equation to obtain the total parameter change is the minimization solution to the following convex quadratic optimization problem:

[0049] ;

[0050] in, is the vector inner set, is the variable to be solved The transpose of

[0051] according to The finite sum structure of It is expressed as:

[0052] ;

[0053] in, , is the total number of samples;

[0054] use The obtained solution updates the CVR model.

[0055] In an alternative embodiment, using The solution to the CVR model update includes:

[0056] Get The solution found ;

[0057] At the optimal parameters Add solution To complete the update of CVR model parameters;

[0058] Until the CVR model parameters converge.

[0059] Compared with the prior art, the conversion rate model updating method of the present invention has the following beneficial effects:

[0060] The conversion rate model updating method proposed in the present invention can be widely used in conversion rate prediction tasks in the field of online advertising. In practical applications, it can effectively handle the delayed feedback problem in conversion rate prediction in various environments. For example, in a large-scale online advertising platform, advertisers can use this method to dynamically adjust the parameters of the CVR model to deal with the data label reversal problem caused by user click behavior and conversion delays. This method realizes real-time updating of the model by calculating the impact of new data and delayed conversion data on model parameters, avoiding retraining of the model, thereby improving prediction accuracy and model adaptability.

[0061] In a second aspect, the present invention provides a conversion rate model updating device, comprising:

[0062] An acquisition module, configured to acquire a training set, wherein the training set includes training samples with labels, and the training samples include clicks and conversions on an online advertising platform;

[0063] An initialization module, used to use a pre-trained CVR model to make a preliminary prediction on the training samples and make preliminary adjustments to the model parameters;

[0064] A label flip calculation module is used to obtain label flip samples, wherein the label flip samples include label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples, and estimate the change of the first parameter of the CVR model by calculation;

[0065] A new arrival data calculation module is used to obtain new arrival data between the training deadline and the test time, wherein the new arrival data includes new samples, and the new samples are input into the CVR model to obtain the model parameter approximation of the new samples, and the second parameter change amount brought by the new arrival data is obtained by calculating the model parameter approximation;

[0066] An updating module is used to obtain a total parameter change according to the first parameter change and the second parameter change, and update the CVR model based on the total parameter change.

[0067] In a third aspect, the present invention provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the conversion rate model updating method described in the first aspect.

[0068] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the conversion rate model updating method described in the first aspect is implemented.

[0069] Compared with the prior art, the beneficial effects of the conversion rate model updating device, electronic device and storage medium in the present invention are the same as those of the conversion rate model updating method of the first aspect, so they will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flow chart of a conversion rate model updating method in an embodiment of the present invention;

[0071] Figure 2 It is a structural block diagram of a conversion rate model updating device in an embodiment of the present invention;

[0072] Figure 3 is a structural block diagram of an electronic device in an embodiment of the present invention;

[0073] Figure 4 This is a comparison chart of the AUC results of the CVR model updated by the method in the embodiment of the present invention on the Taobao dataset;

[0074] Figure 5 This is a comparison chart of the PRAUC results of the CVR model updated by this method on the Taobao dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0076] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the", "these" and the like in this application do not indicate a quantitative limitation, and they may be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of objects.

[0077] In CVR model prediction, the delayed feedback problem urgently needs to address the challenges of label flipping and new data integration. Although retraining the model is a direct solution, its high computational cost and low retraining frequency make it impractical in practical applications.

[0078] In an embodiment of the present invention, a conversion rate model updating method is provided. Figure 1 is a flow chart of the conversion rate model updating method of the present invention, such as Figure 1 As shown, the process includes the following steps:

[0079] S110, obtaining a training set, where the training set includes labeled training samples, and the training samples include clicks and conversions on an online advertising platform;

[0080] Specifically, the training samples in the training set are collected in advance and are cleaned before training the CVR model to remove obvious errors and inconsistencies, thereby ensuring data quality.

[0081] S120. Use the pre-trained CVR model to make preliminary predictions on the training samples and make preliminary adjustments to the model parameters. Use the pre-trained CVR model as a baseline to make preliminary predictions and set model parameters to prepare for subsequent influence function calculations. It should be noted that the influence function is a function that affects the adjustment of model parameters. It is a general term in this manual and does not target a specific function.

[0082] S130, obtaining label flip samples, the label flip samples including label flip samples during training and label flip samples after training, summarizing the gradients of all label flip samples and estimating the change of the first parameter of the CVR model by calculation;

[0083] It should be noted that in CVR prediction problems, the delay between conversion behavior and click behavior is common, and label flipping may occur after the training phase. The label flipping phenomenon can be naturally modeled as data perturbation. For example, a training sample Label flipping occurs, which is equivalent to flipping the training sample The disturbance is , that is, in the training sample Modify the value of the feature / label in some way. It is the training sample Increase the size to of disturbance.

[0084] Since the observed positive samples have accurate labels, the perturbation For the CVR prediction problem, it is 1. For simplicity, consider that label flipping occurs in the training sample In the above case, by re-weighting the perturbed training samples To solve the perturbation empirical risk minimization problem.

[0085] Specifically, obtain label flip samples, which include label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples and estimate the change of the first parameter of the CVR model by calculation include:

[0086] Label flipping is modeled as data perturbation, and the training samples with label flipping are used as samples after data perturbation. The optimization objective is constructed based on the samples after perturbation as follows:

[0087] ;

[0088] In the formula, are the trained model parameters, is a training sample, For training samples Increase the disturbance to A sample of It is a sample Adjusted weights, is the basic loss function (vanilla loss), is the binary cross entropy loss function, It is a sample Optimal parameters after weight adjustment;

[0089] The right side of the above formula is is consistent with the retraining loss when When (i.e. and is large enough), it can be deduced from the above formula:

[0090] ;

[0091] in, Yes Find the derivative and let is equal to 0, Refers to the second-order gradient (Hessian Matrix), is the original loss function The optimal parameters of is the first-order gradient;

[0092] By summing up the gradients of all label flipped samples, the following formula can be used to estimate the change of the first parameter of the CVR model: (i.e. the change in model parameters caused by label flipping):

[0093] ;

[0094] In the formula, is the first parameter change, n is the total number of training samples in the training set, is Before clicking and The set of sample indices converted within the interval (i.e., samples with label flipping), where: is the training deadline, It's testing time.

[0095] S140, obtaining newly arrived data between the training deadline and the test time, the newly arrived data including new samples, using the new samples to input the CVR model to obtain the model parameter approximation of the new samples, and obtaining the second parameter change caused by the newly arrived data by calculating the model parameter approximation;

[0096] It should be noted that the traditional influence function mainly focuses on the changes in model parameters or test sample loss when samples are deleted or modified. In the conversion rate model update method we proposed, New data arriving within the time interval is integrated, considering the new perturbation loss function.

[0097] Specifically, newly arrived data is obtained between the training deadline and the test time, and the newly arrived data includes new samples. The new samples are input into the CVR model to obtain the model parameter approximation of the new samples. The second parameter change amount brought by the newly arrived data is obtained by calculating the model parameter approximation, including:

[0098] The new samples in the newly arrived data Substitute the following new perturbation loss function:

[0099] ;

[0100] in, , ;

[0101] New samples are obtained by Taylor expansion of the new perturbation loss function The change in model parameters is approximately:

[0102] ;

[0103] in, For new samples and the degree of disturbance The parameter change amount; Taking the optimal parameters The inverse of the second-order gradient of all samples, It is to take the optimal parameters New Sample The first-order gradient of

[0104] Approximation of model parameters Get the change of the second parameter caused by the newly arrived data .

[0105] It should be noted that the second parameter change is the parameter change caused by the newly added data. Specifically, the second parameter change caused by the newly arrived data obtained by approximating the model parameters includes:

[0106] Will The value is Substitute the above new disturbance loss function Taylor expansion to obtain the change of the second parameter caused by the new arrival data :

[0107] .

[0108] S150, according to the first parameter change and the second parameter variation Get the total parameter change , based on the total parameter change Updates to the CVR model.

[0109] According to the first parameter change and the second parameter variation Get the total parameter change include:

[0110] The total parameter change is calculated by the following formula :

[0111] ;

[0112] In the formula, is the parameter change caused by the newly arrived data (i.e., the second parameter change);

[0113] Therefore, the total parameter change considering label flip data and newly arrived data can be expressed as above.

[0114] Specifically, the total parameter change is estimated by the following formula: :

[0115] ;

[0116] in, and are the index sets of label flipped samples and new samples respectively,

[0117] Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient, Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient, Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient.

[0118] Solving the above problems faces significant computational challenges. When positive, the conjugate gradient method (CG) can be used to estimate This method only requires the computation of the Hessian-vector product. However, CG methods usually require full batch gradient and Hessian matrix computation, which is not feasible on large-scale datasets. Finding more accurate and efficient solutions remains a challenge. Transform to an equivalent finite sum optimization problem to solve the above challenges, and apply scalable algorithms such as stochastic gradient descent (SGD) and its variants to calculate the total parameter change .

[0119] Specifically, updating the CVR model based on the total parameter change includes:

[0120] The total parameter change As the solution of the following linear system:

[0121] ;

[0122] in, It is to take the optimal parameters The second-order gradient of all samples of is the variable to be solved, is defined as follows:

[0123] ;

[0124] It should be noted that due to is the optimal solution, it must satisfy the second-order necessary optimality condition. Therefore, the matrix It is symmetric and positive semidefinite, and it can be concluded that the total parameter change is the minimization solution of the following convex quadratic optimization problem: ;

[0125] in, is the vector inner set, is the variable to be solved The transpose of

[0126] according to The finite sum structure of It is expressed as:

[0127] ;

[0128] in, , is the total number of samples;

[0129] use The obtained solution updates the CVR model.

[0130] Specifically, using The solution to the CVR model update includes:

[0131] Get The solution found ;

[0132] At the optimal parameters Add solution To complete the update of CVR model parameters;

[0133] Until the CVR model parameters converge.

[0134] Find Solution Afterwards, the solution That is, the parameter change ,get After that, in the original optimal parameters Add The parameter update can be completed, that is, the new parameter = + .

[0135] Use the popular SGD algorithm to minimize During the training process, the loss and gradient are calculated from each sample, and the parameters are updated through SGD. Until convergence. Deploy the trained model to the online advertising platform and update the CVR model in real time. The CVR model can adapt to the latest user behavior data and improve the effectiveness of advertising.

[0136] The conversion rate model update method proposed in the present invention shows significant advantages and positive effects in solving the delayed feedback problem in conversion rate (CVR) prediction. First, the conversion rate model update method accurately estimates the impact of delayed feedback samples on model parameters by introducing an influence function, so as to quickly adapt to data changes without completely retraining the model. Compared with traditional methods, the present invention not only improves the prediction accuracy of the model when dealing with the delayed feedback problem, but also significantly reduces the computational cost and time overhead.

[0137] As shown in Table 1 below, the experimental results show that this method outperforms the existing state-of-the-art methods when processing two large-scale datasets, Criteo and Taobao. Specifically, on the Criteo dataset, the present invention shows significant improvements over the baseline model in terms of AUC, PRAUC, and LL indicators. In terms of the AUC indicator, the present invention improves by 0.58% compared to the Vanilla model, improves by 1.17% in the PRAUC indicator, and improves by 5.76% in the LL indicator. These improvements show that this method has excellent capabilities in capturing potential patterns in data and dealing with delayed feedback problems.

[0138] Table 1: CVR prediction results of different methods on Criteo dataset

[0139]

[0140] In e-commerce websites, this method can help optimize advertising strategies and improve advertising effectiveness by quickly adjusting prediction model parameters to adapt to the latest user behavior data. Similarly, recommendation systems and financial services can also use this method for real-time analysis and model updates to better adapt to data dynamics and improve decision-making accuracy. These application examples demonstrate the flexibility and efficiency of this method in handling conversion rate prediction tasks.

[0141] Such as 4 and Figure 5 As shown in Figure 2, further experimental results show that the CVR model updated by this method also performs significantly better than other control models on the Taobao dataset, especially in terms of AUC and PRAUC indicators. Compared with FSIW, which requires additional pre-training models, this method directly uses real-time label information to more effectively reduce the deviation of label distribution, thereby improving the prediction effect in the short term.

[0142] In summary, this method solves the delayed feedback problem that is difficult to deal with in traditional methods by effectively combining the influence function and the model parameter adjustment mechanism. It not only improves the robustness and accuracy of the model, but also improves the computational efficiency, showing great potential in industrial applications.

[0143] The embodiment of the present invention also provides a conversion rate model updating device, which is used to implement the above method embodiment, and will not be repeated hereafter. The terms "module", "unit", "subunit", etc. used below can implement a combination of software and / or hardware for a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware or a combination of software and hardware is also possible and conceivable.

[0144] like Figure 2 As shown, Figure 2: is a structural block diagram of a conversion rate model updating device in the present invention, the device comprising:

[0145] An acquisition module 101 is used to acquire a training set, where the training set includes labeled training samples, where the training samples include clicks and conversions on an online advertising platform;

[0146] Initialization module 102, used to use the pre-trained CVR model to make preliminary predictions on training samples and make preliminary adjustments to model parameters;

[0147] The label flip calculation module 103 is used to obtain label flip samples, which include label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples and estimate the change of the first parameter of the CVR model through calculation;

[0148] A new arrival data calculation module 104 is used to obtain new arrival data between the training deadline and the test time, wherein the new arrival data includes new samples, and the new samples are input into the CVR model to obtain the model parameter approximation of the new samples, and the second parameter change amount caused by the new arrival data is obtained by calculating the model parameter approximation;

[0149] The updating module 105 is used to obtain a total parameter change according to the first parameter change and the second parameter change, and update the CVR model based on the total parameter change.

[0150] The device achieves real-time updating of the model by calculating the impact of new data and delayed conversion data on model parameters, avoiding retraining of the model, thereby improving prediction accuracy and model adaptability.

[0151] Figure 3 A structural block diagram of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630 and a communication bus 640, wherein the processor 610, the communication interface 620 and the memory 630 communicate with each other through the communication bus 640. The processor 610 may call the logic instructions in the memory 630 to execute the following method:

[0152] Obtain a training set, where the training set includes labeled training samples, and the training samples include clicks and conversions on an online advertising platform;

[0153] Use the pre-trained CVR model to make preliminary predictions on the training samples and make preliminary adjustments to the model parameters;

[0154] Obtain label flip samples, including label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples and estimate the change of the first parameter of the CVR model by calculation;

[0155] Acquire newly arrived data between the training deadline and the test time, the newly arrived data including new samples, use the new samples to input the CVR model to obtain the model parameter approximation of the new samples, and obtain the second parameter change brought by the newly arrived data by calculating the model parameter approximation;

[0156] A total parameter change is obtained according to the first parameter change and the second parameter change, and the CVR model is updated based on the total parameter change.

[0157] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0158] An embodiment of the present invention further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method provided in each of the above embodiments is implemented, for example, including:

[0159] Obtaining a training set, wherein the training set includes labeled training samples, and the training samples include clicks and conversion numbers on an online advertising platform;

[0160] Using the pre-trained CVR model to make preliminary predictions on the training samples and make preliminary adjustments to the model parameters;

[0161] Obtain label flip samples, wherein the label flip samples include label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples, and estimate the change of the first parameter of the CVR model by calculation;

[0162] Acquire new arrival data between the training deadline and the test time, wherein the new arrival data includes new samples, use the new samples to input the CVR model to obtain model parameter approximations of the new samples, and obtain a second parameter change caused by the new arrival data by calculating the model parameter approximation;

[0163] A total parameter change is obtained according to the first parameter change and the second parameter change, and the CVR model is updated based on the total parameter change.

[0164] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0165] It should be understood that the specific embodiments described herein are only used to explain the application, rather than to limit it. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of this application.

[0166] Obviously, the drawings are only some examples or embodiments of the present application. For ordinary technicians in the field, the present application can also be applied to other similar situations based on these drawings without creative work. In addition, it is understandable that although the work done in this development process may be complicated and lengthy, for ordinary technicians in the field, certain changes in design, manufacturing or production based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient content disclosed in this application.

[0167] The term "embodiment" in this application refers to a specific feature, structure or characteristic described in conjunction with the embodiment that can be included in at least one embodiment of the present application. The appearance of this phrase in various locations in the specification does not necessarily mean the same embodiment, nor does it mean that it is mutually exclusive with other embodiments and is independent or optional. It is clearly or implicitly understood by those of ordinary skill in the art that the embodiments described in this application can be combined with other embodiments without conflict.

Claims

1. A conversion rate model updating method, characterized in that: The steps include: Obtaining a training set, wherein the training set includes labeled training samples, and the training samples include clicks and conversion numbers on an online advertising platform; Using the pre-trained CVR model to make preliminary predictions on the training samples and make preliminary adjustments to the model parameters; Obtain label flip samples, wherein the label flip samples include label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples, and estimate the change of the first parameter of the CVR model by calculation; Obtaining label flip samples, the label flip samples include label flip samples during training and label flip samples after training, summarizing the gradients of all label flip samples and estimating the change of the first parameter of the CVR model by calculation, including: Model label flipping as data perturbation, use label flipped training samples as samples after data perturbation, and construct optimization objectives based on the perturbed samples: ； In the formula, are the trained model parameters, is a training sample, For training samples Increase the disturbance to A sample of It is a sample Adjusted weights, is the basic loss function, is the binary cross entropy loss function, It is a sample Optimal parameters after weight adjustment; From the above formula, we can deduce: in, Yes Seeking guidance, refers to the second-order gradient, is the original loss function The optimal parameters of is the first-order gradient; Summarize the gradients of all label flipped samples and use the following formula to estimate the change of the first parameter of the CVR model: In the formula, is the first parameter change, n is the total number of training samples in the training set, is Before clicking and The set of sample indexes converted within the interval, where is the training deadline, It’s testing time; Acquire new arrival data between the training deadline and the test time, wherein the new arrival data includes new samples, use the new samples to input the CVR model to obtain model parameter approximations of the new samples, and obtain a second parameter change caused by the new arrival data by calculating the model parameter approximation; Acquire newly arrived data between the training deadline and the test time, wherein the newly arrived data includes new samples, use the new samples to input the CVR model to obtain the model parameter approximation of the new samples, and obtain the second parameter change brought by the newly arrived data by calculating the model parameter approximation, including: The new samples in the newly arrived data Substitute the following new perturbation loss function: in, , ; The new sample is obtained by Taylor expansion of the new perturbation loss function The change in model parameters is approximately: ； in, For new samples and the degree of disturbance The parameter change amount; Taking the optimal parameters The inverse of the second-order gradient of all samples, It is to take the optimal parameters New Sample The first-order gradient of The variation of the second parameter caused by the newly arrived data is obtained by approximating the model parameters; A total parameter change is obtained according to the first parameter change and the second parameter change, and the CVR model is updated based on the total parameter change.

2. The conversion rate model updating method according to claim 1, characterized in that: The second parameter changes brought about by the newly arrived data obtained by approximating the model parameters include: Will The value is Substitute the new disturbance loss function Taylor expansion to obtain the change of the second parameter caused by the new arrival data : 。 3. The conversion rate model updating method according to claim 2, characterized in that: Obtaining a total parameter change according to the first parameter change and the second parameter change includes: The total parameter change is calculated by the following formula: ； In the formula, is the parameter change caused by the newly arrived data; Specifically, the total parameter change is estimated by the following formula: ； in, and are the index sets of label flipped samples and new samples respectively, Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient, Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient, Representation sample exist To take the optimal parameters Time Parameters Find the first-order gradient.

4. The conversion rate model updating method according to claim 3, characterized in that: Updating the CVR model based on the total parameter change includes: The total parameter change is taken as the solution of the following linear system: in, It is to take the optimal parameters The second-order gradient of all samples of is the variable to be solved, is defined as follows: Solving the above equation to obtain the total parameter change is the minimization solution to the following convex quadratic optimization problem: in, is the vector inner set, is the variable to be solved The transpose of according to The finite sum structure of It is expressed as: in, , is the total number of samples; use The obtained solution updates the CVR model.

5. The conversion rate model updating method according to claim 4, characterized in that: use The solution to the CVR model update includes: Get The solution found ; At the optimal parameters Add solution To complete the update of CVR model parameters; Until the CVR model parameters converge.

6. A conversion rate model updating device, characterized in that: include: An acquisition module, configured to acquire a training set, wherein the training set includes training samples with labels, and the training samples include clicks and conversions on an online advertising platform; An initialization module, used to use a pre-trained CVR model to make a preliminary prediction on the training samples and make preliminary adjustments to the model parameters; A label flip calculation module is used to obtain label flip samples, wherein the label flip samples include label flip samples during training and label flip samples after training, summarize the gradients of all label flip samples, and estimate the change of the first parameter of the CVR model by calculation; Obtaining label flip samples, the label flip samples include label flip samples during training and label flip samples after training, summarizing the gradients of all label flip samples and estimating the change of the first parameter of the CVR model by calculation, including: Model label flipping as data perturbation, use label flipped training samples as samples after data perturbation, and construct optimization objectives based on the perturbed samples: ； In the formula, are the trained model parameters, is a training sample, For training samples Increase the disturbance to A sample of It is a sample Adjusted weights, is the basic loss function, is the binary cross entropy loss function, It is a sample Optimal parameters after weight adjustment; From the above formula, we can deduce: in, Yes Seeking guidance, refers to the second-order gradient, is the original loss function The optimal parameters of is the first-order gradient; Summarize the gradients of all label flipped samples and use the following formula to estimate the change of the first parameter of the CVR model: In the formula, is the first parameter change, n is the total number of training samples in the training set, is Before clicking and The set of sample indexes converted within the interval, where is the training deadline, It’s testing time; A new arrival data calculation module is used to obtain new arrival data between the training deadline and the test time, wherein the new arrival data includes new samples, and the new samples are input into the CVR model to obtain the model parameter approximation of the new samples, and the second parameter change amount brought by the new arrival data is obtained by calculating the model parameter approximation; Acquire newly arrived data between the training deadline and the test time, wherein the newly arrived data includes new samples, use the new samples to input the CVR model to obtain the model parameter approximation of the new samples, and obtain the second parameter change brought by the newly arrived data by calculating the model parameter approximation, including: The new samples in the newly arrived data Substitute the following new perturbation loss function: in, , ; The new sample is obtained by Taylor expansion of the new perturbation loss function The change in model parameters is approximately: ； in, For new samples and the degree of disturbance The parameter change amount; Taking the optimal parameters The inverse of the second-order gradient of all samples, It is to take the optimal parameters New Sample The first-order gradient of The variation of the second parameter caused by the newly arrived data is obtained by approximating the model parameters; An updating module is used to obtain a total parameter change according to the first parameter change and the second parameter change, and update the CVR model based on the total parameter change.

7. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to execute the conversion rate model updating method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the conversion rate model updating method described in any one of claims 1 to 5 is implemented.

Images