Click-through rate debiasing method based on batch adaptive dual robust learning

Abstract

The application discloses a click-through conversion rate debiasing method based on batch adaptive double robust learning, and the method comprises the following steps: obtaining a user set and an item set, and constructing an overall candidate dataset; extracting a click dataset from the overall candidate dataset; dividing the overall candidate dataset into batches; based on the dataset, performing loss estimation through a batch adaptive double robust estimator, training a CVR prediction model by using the estimated loss, and performing conversion rate prediction on user-item pairs in the overall candidate dataset. The application can be applied to various CVR prediction models, and under the premise of not changing the modeling method of the CVR prediction model, only a simple modification is made on the loss function, or a loss function proposed by the application is additionally added on the basis of the original loss function, so that the variance control of the CVR prediction model debiasing in the sparse data scene can be realized, and thus a better recommendation effect can be achieved in a real application scene.

Description

Technical Field

[0001] This application relates to the field of recommendation system technology, specifically a click-through conversion rate correction method based on batch adaptive dual robust learning. Background Technology

[0002] Predicting click-through conversion rate (CVR) is one of the important tasks of recommendation systems. Existing CVR prediction methods mainly rely on user click data to train the model, and then directly use the trained model to predict on all candidate data. Due to the non-randomness of user click behavior, this training faces serious selection bias, that is, there is a huge difference between the distribution of click data and the distribution of all candidate data.

[0003] Existing methods employ an unbiased estimator to estimate the total loss of the CVR prediction model across all candidate data, thereby training the CVR model unbiased and debiasing the CVR predictions. Among these, the Dual Robust Learning (DR) method achieves state-of-the-art performance. The DR method combines Inverse Propensity Score Reweighting (IPS) and Error Completion (EIB). IPS models the probability of clicks (i.e., the propensity score) and reweights the prediction error on clicked data using the inverse of the propensity score; while EIB models the CVR prediction error and completes the prediction error on unclicked data.

[0004] However, in real-world applications, due to the extreme sparseness of click data, the bias scores of some data are often very small or even close to zero, leading to high or even unbounded variance in these methods. This means that these methods cannot consistently reduce the model's total loss across all candidate data, resulting in poor convergence during training and ultimately degraded model performance. Theoretical analysis reveals that the root cause of this problem lies in the design of the DR estimator in existing DR methods. Therefore, this invention proposes a novel dual robust learning method, aiming to build a more robust unbiased estimator under sparse data, rather than simply adjusting the training loss of the bias model or error completion model as in previous variance control methods. While the latter can reduce bias or variance under specific conditions, it still has limitations. Summary of the Invention

[0005] This invention provides a method, system, electronic device, and storage medium for debiasing post-click conversion rate (CVR) based on batch adaptive dual robust learning, in order to achieve variance control of CVR prediction model debiasing in sparse data scenarios, thereby achieving better recommendation results in real-world applications.

[0006] In a first aspect, the present invention provides a method for debiasing post-click conversion rates based on batch adaptive dual robust learning, the method comprising the following steps:

[0007] Get user set and item collection Construct all candidate datasets ;

[0008] From all candidate datasets Extract click dataset ,in Indicates whether a user-item pair was observed. conversion tags Click on the dataset To observe the transformation label User-item pairs;

[0009] All candidate datasets Divided into batch Each batch is of size , Represents all candidate datasets The number of elements in the middle;

[0010] Based on the above dataset, loss is estimated using a batch adaptive dual robust estimator, and the estimated loss is used to train a CVR prediction model.

[0011] The batch adaptive dual robust estimator is defined as follows:

[0012] ;

[0013] in, This is the click probability estimate output by the pre-trained propensity score model. For CVR prediction error, To complete the error estimate output by the pre-trained model;

[0014] Based on the trained CVR prediction model, the entire candidate dataset is analyzed. User-item pairs Perform conversion rate prediction.

[0015] Optionally, the error completion model is trained using a batch adaptive error completion loss function:

[0016] ;

[0017] in, , .

[0018] Optionally, the CVR prediction model is trained using a batch adaptive dual robust estimator through alternating optimization, including:

[0019] Fixed CVR prediction model parameters Minimize the batch adaptive error completion training loss to update the error completion model parameters. ;

[0020] Fixed error completion model parameters Minimize the loss estimated by the batch adaptive dual robust estimator to update the CVR prediction model parameters. .

[0021] Optionally, the propensity score model is trained using a standard binary classification loss function, through the entire candidate dataset. Learning to observe user-item pairs Probability of converting tags That is, the tendency score, thus obtaining its estimated value. ;

[0022] The training loss for the propensity score model is constructed as follows:

[0023] ;

[0024] Optionally, if all user-item pairs satisfy If it is bounded, then the BADR estimator has bounded bias and variance, and when there are learning propensity scores that are less than the propensity score bound, its bias and variance are both less than those of the DR estimator.

[0025] Optionally, the CVR prediction error Defined as cross-entropy loss;

[0026] For each user-item pair The cross-entropy loss is constructed as follows:

[0027] ;

[0028] in, For CVR prediction models Estimation, specifically, the estimation of the probability that the conversion label is 1. The actual transformation labels observed.

[0029] Optionally, based on the trained CVR prediction model, the entire candidate dataset is processed. User-item pairs Conversion rate prediction includes:

[0030] The trained CVR prediction model is invoked to calculate the conversion rate of each item;

[0031] Generate and output a recommendation list sorted by conversion rate.

[0032] Compared with existing technologies, the beneficial effects of the post-click conversion rate debiasing method based on batch adaptive dual robust learning of the present invention are as follows:

[0033] This invention can be applied to various CVR prediction models. Without changing the modeling method of the CVR prediction model, simply modifying the loss function or adding the loss function proposed in this invention to the original loss function can achieve variance control of bias removal in sparse data scenarios, thereby achieving better recommendation results in real application scenarios.

[0034] Secondly, embodiments of the present invention provide a recommendation system based on the post-click conversion rate bias removal method described in the first aspect, comprising:

[0035] The receiving module is used to accept user requests;

[0036] The prediction module is used to call the trained CVR prediction model to calculate the conversion rate of each item;

[0037] The sorting and recommendation module is used to sort items based on predicted conversion rates and recommend the top K items with the highest predicted conversion rates to the user.

[0038] Thirdly, the present invention provides an electronic device including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the first aspect.

[0039] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the click-through rate debiasing method based on batch adaptive dual robust learning as described in the first aspect.

[0040] Compared with the prior art, the beneficial effects of the recommendation system, electronic device and storage medium of the present invention are the same as those of the click-based adaptive dual robust learning-based post-click conversion rate debiasing method described in the first aspect, so they will not be repeated here. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1This is a flowchart of the click-through rate correction method based on batch adaptive dual robust learning in an embodiment of the present invention;

[0043] Figure 2 This is a flowchart illustrating the complete algorithm for unbiased training of the CVR prediction model in this embodiment of the invention.

[0044] Figure 3 This is a schematic diagram of the BADR method across all datasets and evaluation metrics in an embodiment of the present invention;

[0045] Figure 4 This is a structural block diagram of the recommendation system in an embodiment of the present invention;

[0046] Figure 5 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0047] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0048] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0049] First, the technical solution and technical effects of the embodiments of the present invention will be briefly described:

[0050] This invention proposes a novel batch adaptive dual robust learning (BADR) method. Specifically, the method includes a batch adaptive dual robust estimator (BADR estimator) for training a CVR prediction model and a batch adaptive error completion model training loss (BADR error completion model training loss) compatible with the estimator. Wherein: (1) The BADR estimator adaptively adjusts the influence of each data batch in the debiased CVR prediction according to the propensity score distribution within each batch, stabilizing the magnitude of data updates to the model by data with different propensity scores; (2) The BADR error completion model training loss approximates the variance of the BADR estimator, assisting the BADR estimator in training the CVR prediction model, and further improving the stability of the CVR prediction model training.

[0051] Theoretical analysis shows that, in the presence of minimal bias scores, the BADR estimator guarantees bounded variance, and its variance is consistently smaller than that of existing DR estimators. Furthermore, the training loss of the BADR error completion model can theoretically further control the variance of the BADR estimator. Empirical evaluation reveals that BADR demonstrates significant performance improvements in CVR prediction debiasing tasks across datasets with varying sparsity and different evaluation metrics.

[0052] The purpose of this invention is to propose a post-click conversion rate (CVR) bias removal method based on batch adaptive dual robust learning (BADR) to address the problem of excessively high or even unbounded variance in existing bias removal methods when data is sparse. Specifically, this invention aims to achieve the following objectives:

[0053] A novel batch adaptive dual robust estimator is proposed: the estimator dynamically adjusts the influence of different batches of data according to the intra-batch propensity score distribution, suppresses variance inflation caused by small propensity scores, and thus ensures the theoretical boundedness of variance;

[0054] A novel error completion training loss function is proposed: by designing a batch adaptive error completion model training loss compatible with the unbiased estimator, the completion model is optimized to co-optimize the CVR prediction model, thereby further reducing the estimator variance.

[0055] Ultimately, this invention can significantly improve the performance of CVR prediction models on datasets with varying sparsity and different evaluation metrics.

[0056] Specifically, this embodiment of the invention provides a method for debiasing post-click conversion rates based on batch adaptive dual robust learning. Figure 1 This is a flowchart of the click-through rate correction method based on batch adaptive dual robust learning of the present invention, as shown below. Figure 1 As shown, the process includes the following steps:

[0057] S100, Obtain User Set and item collection Construct all candidate datasets ;

[0058] S200, from all candidate datasets Extract click dataset ,in Indicates conversion tag To check if something has been observed, click on the dataset. To observe the transformation label User-item pairs;

[0059] make and They represent from and Users and items, and make This represents the set of all user-item pairs (all candidate data). For each ,make Represents user-item pairs Features such as user ID and item ID. Let express The goal of the CVR prediction task is to learn a CVR prediction model. ,for Each of them Provide accurate forecast values In practical applications, most conversion tags are missing from all candidate data. Specifically, the user behavior chain in a CVR task is: exposure → click → conversion, meaning that conversion tags can only be observed after a user clicks on an item. To formalize the scenario of missing data, let... express Whether it was observed. Based on this, click data can be obtained and represented as... .

[0060] S300, All candidate datasets Divided into batch Each batch is of size , Represents all candidate datasets The number of elements in the middle;

[0061] S400. Based on the above dataset, loss is estimated using a batch adaptive dual robust estimator, and the estimated loss is used to train the CVR prediction model. ,in, Indicates user With items Features such as user ID and item ID, The batch adaptive dual robust estimator is defined as follows, representing the prediction model parameters:

[0062] ;

[0063] in, For pre-trained propensity score model The output is the estimated click probability. For the parameters of the tendency score model, For CVR prediction error, For pre-trained error completion model The output error estimate, To complete the model parameters for error correction;

[0064] It should be noted that the CVR prediction model, propensity score model, and error completion model all use the most classic matrix factorization model, and each uses a separate matrix factorization model. The matrix factorization model models the matching score between users and items by using the inner product of the user embedding vector and the item embedding vector.

[0065] ;

[0066] in It is the Sigmoid function, which compresses the input value to between 0 and 1; It is the embedding vector corresponding to the user; These are the embedding vectors corresponding to the items. When using a matrix factorization model, the model parameters are the embedding vectors corresponding to users and items.

[0067] The error completion model The training uses a batch adaptive error completion loss function:

[0068] ;

[0069] in, . .

[0070] This loss function reduces the variance when using the batch adaptive dual robust estimator to calculate the loss and train the CVR prediction model by theoretically estimating the variance of the estimator, thus making the model training process more stable.

[0071] Furthermore, the CVR prediction model is trained using a batch adaptive dual robust estimator through alternating optimization, including:

[0072] Fixed CVR prediction model parameters Minimize the batch adaptive error completion training loss to update the error completion model parameters. ;

[0073] Fixed error completion model parameters Minimize the loss estimated by the batch adaptive dual robust estimator to update the CVR prediction model parameters. .

[0074] Furthermore, the propensity score model is trained using a standard binary classification loss function, through the entire candidate dataset. Learning to observe user-item pairs Probability of converting tags That is, the tendency score, thus obtaining its estimated value. ;

[0075] By clicking the dataset study The estimated value ,in, This indicates that the observation The probability of the corresponding conversion tag can also be understood as the click-through rate in the CVR prediction scenario.

[0076] Specifically, the tendency is to represent observed fractions. The probability of the corresponding conversion tag is formalized as In CVR prediction scenarios, observations were made. The probability of the corresponding conversion tag and the click-through rate have the same meaning, because a conversion only occurs after a user clicks. (Propensity score model) This is used to predict the probability, and the predicted probability is denoted as... .according to as well as The propensity score model can be trained using the following loss:

[0077] ;

[0078] The completion error represents the error for cases where no transformation label was observed. ,pass Prediction error of the completed CVR model. Actual error. for ,in For CVR prediction models Estimation, specifically, estimating the probability that the conversion label is 1. Error completion model. Training can be performed using the following loss:

[0079] ;

[0080] If all user-item pairs satisfy If the bias is bounded, the BADR estimator has bounded bias and variance, and its bias and variance are both smaller than the DR estimator when there are learned propensity scores smaller than the propensity score bound. Specifically, there exists a propensity score bound in different datasets; when scores smaller than or equal to this propensity score bound exist, the bias and variance of the BADR estimator are both smaller than those of the DR estimator. Due to the sparsity of the data, such propensity scores will naturally exist.

[0081] The CVR prediction error Defined as cross-entropy loss;

[0082] For each user-item pair The cross-entropy loss is constructed as follows:

[0083] ;

[0084] in, For CVR prediction models Estimation, specifically, the estimation of the probability that the conversion label is 1. The actual transformation labels observed.

[0085] S500, based on the trained CVR prediction model, performs predictions on all candidate datasets. Predict conversion rates for user-item pairs.

[0086] Based on the trained CVR prediction model, the entire candidate dataset is analyzed. User-item pairs Conversion rate prediction includes:

[0087] The trained CVR prediction model is invoked to calculate the conversion rate of each item;

[0088] Generate and output a recommendation list sorted by conversion rate.

[0089] First, the variance analysis of the dual robust learning (DR) method, where the DR estimator combines inverse tendency score reweighting (IPS) and error completion (EIB):

[0090] ;

[0091] in, It is the size of all candidate datasets D. It is the CVR prediction model in The prediction error on, i.e. and Cross-entropy error; It is a complete CVR prediction model exist The prediction error comes from an error completion model. ; Click probability The estimate, i.e., the propensity score, comes from a propensity score model. This estimator is a general estimator for existing DR methods.

[0092] The variance of the DR estimator is:

[0093] ;

[0094] A major challenge in debiasing using the DR estimator in CVR prediction is the sparsity of the observed data, meaning that in real-world scenarios, users click on very few items. Sparsity implies that many user-item pairs have very small propensity scores. As shown in the equation above, the variance of the DR estimator is highly sensitive to data with small propensity scores. Specifically, the variance of the DR estimator depends on the inverse square of the learned propensity scores, which diverge to infinity as the learned propensity scores approach zero, regardless of the accuracy of the propensity score estimates. Therefore, in the case of sparse data, the consistency between the reduction in total loss across all candidate data and the estimated loss cannot be guaranteed, leading to suboptimal model convergence and degraded performance. The above analysis prompts this invention to propose a novel Batch Adaptive Dual Robust Learning (BADR) method to construct a more robust estimator.

[0095] Batch Adaptive Dual Robust Estimator (BADR Estimator)

[0096] This invention redesigns the dual robust (DR) estimator formula for batch training scenarios, proposing a theoretically supported batch adaptive dual robust estimator. This estimator adaptively adjusts the weight of each data batch's influence on conversion rate prediction correction based on the propensity score distribution learned within each data batch.

[0097] The formal definition is as follows: Let there be a common... Data batches Each batch is of size The expression for the BADR estimator is as follows:

[0098] ;

[0099] Compared to the DR estimator, the BADR estimator reduces the influence of scores with low bias through an adaptive mechanism, while increasing the weight of scores with high bias. Its core principle is that BADR effectively mitigates the negative impact of scores with very low bias on the unbiased CVR prediction process by dynamically adjusting the influence of each batch of data. This mechanism not only makes the model training process more robust but also ultimately yields more accurate unbiased CVR prediction results.

[0100] Analysis of variance of the BADR estimator: The variance of the BADR estimator is:

[0101] ;

[0102] in, It is a with Related complexity terms, Indicates error deviation The finite difference between it and its weighted average is defined as:

[0103] ;

[0104] Based on the variance of the BADR estimator described above, this invention arrives at the following conclusion: assuming that for all satisfy If bounded, the BADR estimator has bounded bias and variance, and its bias and variance are smaller than those of the DR estimator when there are small learning tendency scores.

[0105] 4) Training loss of batch adaptive error completion model

[0106] To further enhance the robustness of the algorithm, this invention employs the Delta method to approximate the variance of the BADR estimator, and then further optimizes it. It should be noted that the variance of the previously given BADR estimator cannot be directly optimized, as its calculation depends on the true propensity scores of all user-item pairs.

[0107] The approximate variance of the BADR estimator is:

[0108] ;

[0109] in, , This invention uses it as the training loss for the error completion model.

[0110] Intuitively, the training loss function of the BADR error completion model possesses unique batch adaptive characteristics and additional constraints. This dual mechanism can more effectively optimize the error completion model, thereby improving the training performance of the CVR prediction model through the BADR estimator. Specifically, the training loss of the BADR error completion model is achieved through... This achieves adaptive impact of each data batch on the training of the error completion model and forms a unified coordination with the BADR estimator; additional constraints are also introduced. Make the error completion model It can make full use of the combined information of global data distribution and BADR estimator.

[0111] like Figure 2 As shown, Figure 2 This is a complete flowchart of the algorithm for unbiased training of the CVR prediction model in this invention. First, input the click dataset. All candidate data and learning rate Then initialize the CVR prediction model. Error completion model and propensity score model Then, the loss model is trained using the existing propensity score model. Then, a joint learning framework is employed to jointly optimize the CVR prediction model and the error completion model until convergence. Specifically, in each batch of data... In the middle: The interpolation model is trained by minimizing the BADR error completion training loss function. Train a CVR prediction model using the BADR loss function estimator. An alternating optimization approach is used to iteratively update the CVR prediction model and the error completion model in each batch of data until the early stopping criterion is met.

[0112] In this invention, the proposed bias removal method addresses the common problem of high or even unbounded variance in existing CVR (Continuous Virtual VR) methods under sparse data conditions. This leads to an inability to steadily reduce the total loss of the model on global candidate data, resulting in convergence to a suboptimal position and ultimately impacting CVR prediction performance. In contrast, this invention introduces a novel batch adaptive dual robust estimator (BADR) combined with a compatible batch adaptive error-complementary training loss, effectively reducing variance and thus improving the performance of the CVR bias removal model. Experiments show that BADR significantly outperforms existing methods on datasets with varying sparsity and across multiple evaluation metrics.

[0113] This invention was tested on three datasets: Yahoo! R3, Coat, and KuaiRec. All three datasets contain biased data affected by selection bias, as well as unbiased data collected through a random exposure strategy. Specific details are as follows:

[0114] Yahoo! R3 is a music dataset containing 15,400 users and 1,000 items. This dataset records 311,400 user ratings of music, which are considered biased data. Additionally, the dataset contains ratings from 5,400 users, each rating 10 randomly selected items, generating a total of 54,000 ratings, which are considered unbiased data. The Yahoo! R3 ratings are binarized: ratings below 3 are assigned a negative conversion label (0), and ratings of 3 or higher are assigned a positive conversion label (1).

[0115] The Coat dataset contains 290 users and 300 items. It records 4,640 user ratings of clothing, which are considered biased data. Additionally, it contains 6,960 ratings from 290 users, each rating 16 randomly selected clothing items, which are considered unbiased data. The Coat ratings are binarized: ratings below 3 are assigned a negative conversion label of 0, and ratings of 3 or higher are assigned a positive conversion label of 1.

[0116] KuaiRec is a publicly available industrial-grade video dataset; a sampled version is selected here. It contains 201,171 video viewing duration ratio records from 1,411 users and 3,327 videos, which are considered biased data. Additionally, it contains 117,113 viewing duration ratio records for videos randomly exposed to users, which are considered unbiased data. Viewing duration ratios are binarized: values ​​below 2 are set to 0, and values ​​of 2 and above are set to 1.

[0117] Experiments were conducted on the three datasets mentioned above, and AUC, Recall@K, and NDCG@K, which are of interest to recommendation systems, were selected as evaluation metrics. The experimental results are as follows: Figure 3 As shown, our method consistently outperforms all baseline methods across all datasets and evaluation metrics, with a particularly significant improvement in TopK ranking metrics (NDCG@K and Recall@K). This advantage highlights the effectiveness of BADR in debiasing tasks. BADR's success stems from its enhanced control over variance in sparse data, enabling it to more consistently and stably optimize the total loss of the CVR prediction model across all candidate data by optimizing the estimated loss.

[0118] In summary, this invention can be applied to various CVR prediction models. Without changing the modeling method of the CVR prediction model, simply modifying the loss function or adding the loss function proposed in this invention to the original loss function can achieve variance control of bias removal in sparse data scenarios, thereby achieving better recommendation results in real-world applications.

[0119] like Figure 4 As shown, this embodiment of the invention also provides a recommendation system for implementing the above-described method embodiments; details already described will not be repeated. The terms "module," "unit," and "subunit," etc., used below refer to combinations of software and / or hardware that perform a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation or a combination of software and hardware is also possible and contemplated.

[0120] The recommendation system, based on the above-mentioned methods for correcting post-click conversion rate bias, includes:

[0121] Receiver module 101 is used to accept user requests;

[0122] Prediction module 102 is used to call the trained CVR prediction model to calculate the conversion rate of each item;

[0123] The sorting and recommendation module 103 is used to sort items according to the predicted conversion rate and recommend the top K items with the highest predicted conversion rate to the user.

[0124] Once a user enters the system, their request is automatically sent to the backend. Upon receiving the request, the system uses a CVR (Conversion Rate) prediction model to calculate the match score (i.e., conversion probability) between the user and all candidate items. Subsequently, the system sorts the items according to their match scores from highest to lowest and recommends the top-ranked items to the user.

[0125] Figure 5 A structural block diagram of the electronic device provided in the embodiments of the present invention, such as... Figure 5 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the methods provided in the above embodiments.

[0126] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0128] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A click-through conversion rate bias removal method based on batch adaptive dual robust learning, characterized in that, The method includes the following steps: Get user set and item collection Construct all candidate datasets ; From all candidate datasets Extract click dataset ,in Indicates whether a user-item pair was observed. conversion tags Click on the dataset To observe the transformation label User-item pairs; All candidate datasets Divided into batch Each batch is of size , Represents all candidate datasets The number of elements in the middle; Based on the above dataset, loss is estimated using a batch adaptive dual robust estimator, and the estimated loss is used to train a CVR prediction model. The batch adaptive dual robust estimator is defined as follows: ； in, This is the click probability estimate output by the pre-trained propensity score model. For CVR prediction error, To complete the error estimate output by the pre-trained model; The error completion model is trained using a batch adaptive error completion loss function, and the loss for each batch is formalized as follows: ； in, , ; The propensity score model is trained using a standard binary classification loss function, through all candidate datasets. Learning to observe user-item pairs Probability of converting tags That is, the tendency score, thus obtaining its estimated value. ; The training loss for the propensity score model is constructed as follows: ； Based on the trained CVR prediction model, the entire candidate dataset is analyzed. User-item pairs Perform conversion rate prediction.

2. The click-through conversion rate debiasing method based on batch adaptive dual robust learning according to claim 1, characterized in that, Training a CVR prediction model using a batch adaptive dual robust estimator is achieved through alternating optimization, including: Fixed CVR prediction model parameters Minimize the batch adaptive error completion training loss to update the error completion model parameters. ; Fixed error completion model parameters Minimize the loss estimated by the batch adaptive dual robust estimator to update the CVR prediction model parameters. .

3. The click-through conversion rate debiasing method based on batch adaptive dual robust learning according to claim 1, characterized in that, If all user-item pairs satisfy If it is bounded, then the BADR estimator has bounded bias and variance, and when there are learning propensity scores that are less than the propensity score bound, its bias and variance are both less than those of the DR estimator.

4. The click-through conversion rate debiasing method based on batch adaptive dual robust learning according to claim 1, characterized in that, The CVR prediction error Defined as cross-entropy loss; For each user-item pair The cross-entropy loss is constructed as follows: ； in, For CVR prediction models Estimation, specifically, the estimation of the probability that the conversion label is 1. The actual transformation labels observed.

5. The click-through conversion rate debiasing method based on batch adaptive dual robust learning according to claim 1, characterized in that, Based on the trained CVR prediction model, the entire candidate dataset is analyzed. User-item pairs Conversion rate prediction includes: The trained CVR prediction model is invoked to calculate the conversion rate of each item; Generate and output a recommendation list sorted by conversion rate.

6. A recommendation system, employing the post-click conversion rate debiasing method as described in any one of claims 1-5, characterized in that, include: The receiving module is used to accept user requests; The prediction module is used to call the trained CVR prediction model to calculate the conversion rate of each item; The sorting and recommendation module is used to sort items based on predicted conversion rates and recommend the top K items with the highest predicted conversion rates to the user.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the post-click conversion rate debiasing method based on batch adaptive dual robust learning as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the post-click conversion rate debiasing method based on batch adaptive dual robust learning as described in any one of claims 1 to 5.

Images