Method and apparatus for eliminating noise samples in a fraud scenario

Abstract

This specification provides a method and apparatus for eliminating noisy samples in a risk control scenario. The method includes: obtaining an original training sample set; the original training sample set includes expanded training samples and clean training samples; inputting a portion of the training samples from the original training sample set into a first denoising model and another portion into a second denoising model; determining training samples with relatively reliable labels from the training samples input into one denoising model, and updating the parameters of the other denoising model using the gradient information of the denoising model obtained when inputting the relatively reliable labeled training samples; finally, the denoising model classifies the training samples in the original training sample set, and updates the labels of the training samples in the original training sample set using the classification results, thereby obtaining a training sample set after noise elimination. This specification provides a more effective method for eliminating noisy samples.

Description

Technical Field

[0001] This specification relates to computer technology in one or more embodiments, and more particularly to methods and apparatus for eliminating noise samples in risk control scenarios. Background Technology

[0002] With the continuous development of internet applications, risk control of user online behavior has become an important issue. Currently, the common approach is to train risk control models and use these models to classify user online behavior to determine whether it constitutes risky behavior. For example, with the rapid development of e-commerce platforms, transactions have become more convenient and faster, but this has also led to many abnormal transaction behaviors. In such cases, risk control models targeting specific transaction types can be used to differentiate whether a user's transaction behavior is normal.

[0003] To improve the accuracy of risk control model classification, a large number of training samples are needed. However, in real-world business scenarios, risk mitigation measures are typically implemented for user online behaviors, such as transactions, resulting in a limited number of samples obtainable from actual business data. Therefore, to expand the sample size, certain rules are used to add potentially risky samples as positive samples. While this effectively increases the number of positive samples, it also introduces varying degrees of noise into the training data. These noisy samples are the expanded samples and have inaccurate labels. To improve the accuracy of risk control model classification, a more effective method for eliminating noisy samples is needed. Summary of the Invention

[0004] This specification describes one or more embodiments of a method and apparatus for eliminating noise samples in a risk control scenario, which can more effectively eliminate noise samples in training samples.

[0005] According to the first aspect, a method for eliminating noisy samples in a risk control scenario is provided, wherein the method includes:

[0006] Obtain the original training sample set; this original training sample set includes augmented training samples and clean training samples.

[0007] A portion of the training samples from the original training sample set is input into the first denoising model, and another portion of the training samples from the original training sample set is input into the second denoising model.

[0008] The training samples with relatively reliable labels are determined from the training samples input to the first denoising model, and the gradient information of the first denoising model obtained when the training samples with relatively reliable labels are input are used to update the parameters of the second denoising model.

[0009] The training samples with relatively reliable labels are determined from the training samples input to the second denoising model, and the gradient information of the second denoising model obtained when the training samples with relatively reliable labels are input is used to update the parameters of the first denoising model.

[0010] The training samples in the original training sample set are classified using the first and second denoising models obtained at the end. The labels of the training samples in the original training sample set are updated using the classification results, thereby obtaining the training sample set after noise removal.

[0011] Wherein, the clean training samples are training samples with accurate labels;

[0012] The expanded training samples are training samples obtained based on the user's controlled network behavior.

[0013] The training samples input into the first denoising model include expanded training samples and clean training samples; the training samples input into the second denoising model include expanded training samples and clean training samples.

[0014] The training samples whose labels are determined to be relatively reliable include:

[0015] Calculate the value of the loss function of the denoising model when a training sample is input;

[0016] If the calculated loss value is less than the preset threshold, then the training sample is determined to be a training sample with relatively reliable labels.

[0017] The step of classifying the training samples in the original training sample set using the finally obtained first denoising model and second denoising model includes: arbitrarily selecting a denoising model from the finally obtained first denoising model and second denoising model, and classifying the training samples in the original training sample set using the selected denoising model.

[0018] Wherein, each training sample input into the first denoising model is a clean training sample; each training sample input into the second denoising model is an augmented training sample.

[0019] The training samples whose labels are determined to be relatively reliable include:

[0020] For the first denoising model, all training samples input into the first denoising model are determined to be training samples with relatively reliable labels;

[0021] For the second denoising model, calculate the value of the loss function of the second denoising model when a training sample is input; if the calculated loss value is less than a preset threshold, then the training sample is determined to be a training sample with relatively reliable labels.

[0022] The step of classifying the training samples in the original training sample set using the finally obtained first denoising model and second denoising model includes: selecting the second denoising model from the finally obtained first denoising model and second denoising model, and using the second denoising model to classify the training samples in the original training sample set.

[0023] The step of inputting a portion of the training samples from the original training sample set into the first denoising model and inputting another portion of the training samples from the original training sample set into the second denoising model includes:

[0024] A portion of the training samples in the original training sample set is divided into multiple batches, and in each round of training, one batch of training samples is input into the first denoising model. Another portion of the training samples in the original training sample set is divided into multiple batches, and in each round of training, one batch of training samples is input into the second denoising model.

[0025] After determining that the loss calculated for a training sample is less than a preset threshold, and before determining that the training sample is a training sample with relatively reliable labels, the process further includes:

[0026] Determine whether the gradient direction of the denoising model when inputting this training sample is consistent with the gradient direction of the denoising model when inputting most other training samples in this batch. If so, determine that the training sample is a training sample with relatively reliable labels.

[0027] According to the second aspect, an apparatus for eliminating noise samples in a risk control scenario is provided, wherein the apparatus includes:

[0028] The training sample set acquisition module is configured to obtain the original training sample set, which includes expanded training samples and clean training samples obtained based on the user's normal network behavior.

[0029] The sample input module is configured to input a portion of the training samples from the original training sample set into the first denoising model, and input another portion of the training samples from the original training sample set into the second denoising model;

[0030] The mutual teaching module is configured to determine training samples with relatively reliable labels from each training sample input to the first denoising model, and update the parameters of the second denoising model using the gradient information of the first denoising model obtained when the training sample with relatively reliable label is input; and to determine training samples with relatively reliable labels from each training sample input to the second denoising model, and update the parameters of the first denoising model using the gradient information of the second denoising model obtained when the training sample with relatively reliable label is input.

[0031] The training sample set denoising module is configured to classify the training samples in the original training sample set using the final obtained first denoising model and second denoising model, and update the label of the training sample in the original training sample set using the classification result, thereby obtaining the training sample set after noise removal.

[0032] According to a third aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the methods described in any embodiment of this specification.

[0033] According to a fourth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method described in any embodiment of this specification.

[0034] The method and apparatus for eliminating noisy samples in risk control scenarios provided in the embodiments of this specification utilize two denoising models to teach each other. That is, the two denoising models communicate with each other, determining which training samples input to the denoising model have relatively reliable labels based on the training process of one denoising model. Then, these relatively reliable training samples can be used for backpropagation to update the parameters of the other denoising model. In other words, the parameters obtained by one denoising model when inputting the relatively reliable training samples are used to update the parameters of the other denoising model, thereby greatly improving the accuracy of the classification results of the trained denoising model. Using this accurately classified denoising model, the training samples in the original training sample set are reclassified, and the labels of the training samples in the original training sample set are modified based on the classification results. This ensures that the labels of the training samples in the modified training sample set are accurate, thus eliminating noisy samples. Subsequently, a more accurate risk control model can be trained using the training sample set with eliminated noisy samples. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a schematic diagram of the system architecture used in one embodiment of this specification.

[0037] Figure 2 This is a flowchart of a method for eliminating noise samples in a risk control scenario, as described in one embodiment of this specification.

[0038] Figure 3 This is a schematic diagram of a method for eliminating noise samples in a risk control scenario, as described in one embodiment of this specification.

[0039] Figure 4 This is a schematic diagram of the device for eliminating noise samples in a risk control scenario according to one embodiment of this specification. Detailed Implementation

[0040] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0041] First, it should be noted that the terminology used in the embodiments of this invention is for the purpose of describing specific embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0042] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0043] To facilitate understanding of the methods provided in this manual, the system architecture involved and applicable to this manual is first described. For example... Figure 1 As shown, the system architecture mainly includes three network nodes: a web crawler module, a training sample generation module, and a noise sample elimination module.

[0044] The web crawler module is used to collect data on users' normal online behavior and controlled online behavior, such as data on users' normal transaction behavior and data on users' controlled transaction behavior (such as data on transactions that have been reported); the training sample generation module is used to generate an original training sample set based on the data collected by the web crawler module; and the noise sample removal module is used to remove noise samples from the original training sample set.

[0045] It should be understood that Figure 1 The number of web crawler modules, training sample generation modules, and noise sample removal modules shown is merely illustrative. Any number can be selected and deployed as needed for implementation.

[0046] Figure 2 This is a flowchart illustrating a method for eliminating noise samples in a risk control scenario, as described in one embodiment of this specification. The main implementer of this method is a device for eliminating noise samples in a risk control scenario. It is understood that this method can also be executed by any device, equipment, platform, or cluster of devices with computing and processing capabilities. See also... Figure 2 The method includes:

[0047] Step 201: Obtain the original training sample set; the original training sample set includes expanded training samples and clean training samples.

[0048] Step 203: Input a portion of the training samples from the original training sample set into the first denoising model, and input another portion of the training samples from the original training sample set into the second denoising model.

[0049] Step 205: Determine the training samples with relatively reliable labels from the training samples input to the first denoising model, and update the parameters of the second denoising model using the gradient information of the first denoising model obtained when inputting the training samples with relatively reliable labels.

[0050] Step 207: Determine the training samples with relatively reliable labels from the training samples input to the second denoising model, and update the parameters of the first denoising model using the gradient information of the second denoising model obtained when inputting the training samples with relatively reliable labels.

[0051] Step 209: Classify the training samples in the original training sample set using the first and second denoising models obtained at the end, and update the label of the training sample in the original training sample set using the classification results, thereby obtaining the training sample set after noise removal.

[0052] The augmented training samples in the original training sample set are obtained using the user's controlled network behavior, and their labels are often inaccurate, meaning that noisy samples may appear. Based on the above... Figure 2As shown in the illustrated process, in this embodiment, two denoising models are used to teach each other. That is, the two denoising models communicate with each other. Based on the training process of one denoising model, it determines which training samples input to that model have relatively reliable labels. Then, these relatively reliable training samples are used for backpropagation to update the parameters of the other denoising model. In other words, the parameters obtained by one denoising model when inputting these relatively reliable training samples are used to update the parameters of the other denoising model. This significantly improves the accuracy of the classification results of the trained denoising model. Using this accurately classified denoising model, the training samples in the original training sample set are reclassified. The labels of the training samples in the original training sample set are modified based on the classification results, ensuring that the labels of the training samples in the modified training sample set are accurate, thus eliminating noisy samples. Finally, a more accurate risk control model can be trained using the training sample set with the noise-free samples eliminated.

[0053] The following is about Figure 2 Each step shown is explained separately.

[0054] First, for step 201: obtain the original training sample set; the original training sample set includes expanded training samples and clean training samples.

[0055] In step 201, a web crawler module can be used to collect various user network behavior data from the network, thereby forming an original training sample set. In this original training sample set, clean training samples are those with accurate labels, which can be either risky or non-risky behaviors. For example, a user's transaction behavior judged as a violation and its label "risky behavior" forms a clean training sample, as does a user's transaction behavior judged as not a violation and its label "non-risky behavior." Expanding the training samples involves obtaining training samples based on preset rules and user network behavior; their labels may be inaccurate. For example, if a user's transaction behavior is complained about by other users for failing to fulfill delivery service, this user's transaction behavior can be considered a controlled network behavior and labeled "risky behavior." However, since there are various reasons for failing to fulfill delivery service, directly labeling it as "risky behavior" is likely incorrect, thus generating noisy samples.

[0056] Next, for step 203: input a portion of the training samples from the original training sample set into the first denoising model, and input another portion of the training samples from the original training sample set into the second denoising model.

[0057] In the embodiments described in this specification, a mutual teaching approach is used to train a denoising model capable of eliminating noisy samples. Therefore, the training samples in the original training sample set can first be divided into two parts: one part is used to train the first denoising model, and the other part is used to train the second denoising model. For example, if the original training sample set includes 100,000 training samples (which include both clean and expanded training samples), 50,000 of these training samples can be input into the first denoising model, while the remaining 50,000 training samples can be input into the second denoising model.

[0058] In the embodiments of this specification, there are two possible ways to divide the training samples, and these methods can be input into the two denoising models respectively:

[0059] Method 1: The training samples input into the first denoising model include expanded training samples and clean training samples; the training samples input into the second denoising model include expanded training samples and clean training samples.

[0060] In this first method, the training samples input into the two denoising models are indistinguishable. For example, a random partitioning method can be used, grouping any portion of the original training sample set together and inputting them all into the first denoising model, while grouping the remaining training samples together and inputting them all into the second denoising model. Thus, typically, the training samples input into the first denoising model include both augmented and clean training samples; and the training samples input into the second denoising model also include both augmented and clean training samples.

[0061] Method 2: Each training sample input into the first denoising model is a clean training sample; each training sample input into the second denoising model is an expanded training sample.

[0062] Clean training samples are training samples with accurate labels and do not contain noise samples. Augmented training samples, however, are expanded based on certain rules; therefore, noise samples exist within them. That is, some augmented training samples have correct labels, while others have incorrect labels. These incorrectly labeled augmented training samples are the noise samples. The embodiments in this specification aim to identify these noise samples and modify their labels to make them correctly labeled training samples. Therefore, in one embodiment of this specification, all clean training samples in the original training sample set can be grouped together and input into the first denoising model, and all augmented training samples in the original training sample set can be grouped together and input into the second denoising model, thereby making the training of different denoising models more targeted.

[0063] Next, for step 205: determine the training samples with relatively reliable labels from each training sample input to the first denoising model, and update the parameters of the second denoising model using the gradient information of the first denoising model obtained when inputting the training samples with relatively reliable labels.

[0064] In the embodiments of this specification, the two denoising models communicate with each other. Based on the training process of the first denoising model, it determines which training samples input to the first denoising model have relatively reliable labels. Then, these relatively reliable training samples can be used for backpropagation and to update the parameters of the second denoising model. Since relatively reliable training samples have already been selected based on the training process of the first denoising model (meaning the labels should be accurate), the gradient information of the first denoising model obtained from these relatively reliable training samples is the gradient information that should be used. Therefore, after sending this gradient information to the second denoising model, the second denoising model can update its own model parameters, thereby improving the classification ability of the second denoising model.

[0065] When step 203 is implemented using the method described above, the method for determining relatively reliable training samples in step 205 includes:

[0066] Step 2051: Calculate the value of the loss function of the first denoising model when a training sample, such as training sample A, is input into the first denoising model;

[0067] Step 2053: Determine whether the calculated loss value is less than the preset threshold. If yes, proceed to step 2055; otherwise, proceed to step 2057.

[0068] Step 2055: Determine that the training sample A is a training sample with relatively reliable labels.

[0069] Step 2057: Determine that the training sample A is not a training sample with relatively reliable labels.

[0070] In the embodiments of this specification, a portion of the training samples in the original training sample set can be divided into multiple batches, and in each round of training, one batch of training samples is input into the first denoising model. Furthermore, another portion of the training samples in the original training sample set is also divided into multiple batches, and in each round of training, one batch of training samples is input into the second denoising model. The number of training samples in each batch can also be relatively small, such as 100. Because multiple training samples from one batch are simultaneously input into a denoising model in one round of training, when determining whether the label is relatively reliable, the gradient direction of a training sample relative to other training samples in the same batch can be further referenced. Specifically, after determining in step 2053 that the loss value calculated for a training sample is less than a preset threshold, and before determining in step 2055 that the training sample is a training sample with a relatively reliable label, the process further includes:

[0071] Step 2054: Determine whether the gradient direction of the first denoising model is consistent with the gradient direction of the denoising model when the training sample A is input, and whether the gradient direction of the denoising model is consistent with the gradient direction of most other training samples in this batch. If so, determine that the training sample A is a training sample with relatively reliable labels; otherwise, determine that the training sample A is not a training sample with relatively reliable labels.

[0072] For example, if a batch has 100 training samples, and most of them, say 80, have gradients in the same direction (decreasing), then for training sample A in this batch, step 2054 determines that when training sample A is input, the gradient direction of the first denoising model is also decreasing, which is the same as the gradient direction of most training samples in this batch. In this case, the process can continue to determine that training sample A is a training sample with relatively reliable labels. Conversely, if step 2054 determines that when training sample A is input, the gradient direction of the first denoising model is increasing, which is different from the gradient direction of most training samples in this batch, the process can continue to determine that training sample A is not a training sample with relatively reliable labels.

[0073] When step 203 is implemented using the above-mentioned method 2, the method for determining the training samples with relatively reliable labels in step 205 includes: determining all training samples input to the first denoising model as training samples with relatively reliable labels.

[0074] Next, for step 207: determine the training samples with relatively reliable labels from each training sample input to the second denoising model, and update the parameters of the first denoising model using the gradient information of the second denoising model obtained when inputting the training samples with relatively reliable labels.

[0075] Based on the training process of the second denoising model, it is determined which training samples input to the second denoising model have relatively reliable labels. Then, these relatively reliable training samples can be used for backpropagation to update the parameters of the first denoising model. Since relatively reliable training samples have already been selected based on the training process of the second denoising model, meaning the labels should be accurate, the gradient information of the second denoising model obtained from these relatively reliable training samples is the gradient information that should be used. Therefore, after sending this gradient information to the first denoising model, the first denoising model can update its own model parameters, thereby improving the classification ability of the first denoising model.

[0076] Regardless of whether step 203 is implemented using method one or method two described above, the methods for determining relatively reliable training samples in step 207 include:

[0077] Step 2071: Calculate the value of the loss function of the second denoising model when a training sample, such as training sample B, is input into the second denoising model;

[0078] Step 2073: Determine whether the calculated loss value is less than the preset threshold. If yes, proceed to step 2075; otherwise, proceed to step 2077.

[0079] Step 2075: Determine that the training sample B is a training sample with relatively reliable labels.

[0080] Step 2077: Determine that training sample B is not a training sample with relatively reliable labels.

[0081] In steps 2073 and 2075, when determining whether the label is relatively reliable, the gradient direction of a training sample relative to other training samples in the same batch can be further referenced. Specifically, after determining in step 2073 that the loss value calculated for a training sample B is less than a preset threshold, and before determining in step 2075 that training sample B is a training sample with a relatively reliable label, the following further steps are included:

[0082] Step 2074: Determine whether the gradient direction of the second denoising model when inputting the training sample B is consistent with the gradient direction of the second denoising model when inputting most of the other training samples in this batch. If so, determine that the training sample B is a training sample with relatively reliable labels; otherwise, determine that the training sample B is not a training sample with relatively reliable labels.

[0083] Next, for step 209: use the first denoising model and the second denoising model obtained at the end to classify the training samples in the original training sample set, and use the classification results to update the label of the training sample in the original training sample set, so as to obtain the training sample set after noise removal.

[0084] When step 203 is implemented using the above method, the method for classifying the training samples in the original training sample set using the finally obtained first denoising model and second denoising model in step 209 includes: arbitrarily selecting a denoising model from the finally obtained first denoising model and second denoising model, and classifying the training samples in the original training sample set using the selected denoising model.

[0085] When step 203 is implemented using the above-described method 2, the method for classifying the training samples in the original training sample set using the finally obtained first denoising model and second denoising model in step 209 includes: selecting the second denoising model from the finally obtained first denoising model and second denoising model, and classifying the training samples in the original training sample set by the second denoising model.

[0086] If it can be determined in advance which training samples are clean and which are augmented in the original training sample set, then in step 209, only the augmented training samples in the original training sample set can be classified, and the labels of the augmented training samples can be obtained by using the classification results, thereby obtaining the training sample set after noise removal.

[0087] See Figure 2 , Figure 3 In one embodiment of this specification, two denoising models are trained simultaneously, allowing them to teach each other using mini-batch training samples. Each denoising model feeds forward all the data, selecting training samples that are likely to have clean labels, i.e., relatively reliable labels, using a small loss function value. The two denoising models communicate with each other, selecting which training samples should be used for training in this mini-batch. Each denoising model backpropagates and updates its parameters using training samples that the other denoising model considers to have relatively reliable labels. Eventually, the two denoising models converge, and a single denoising model is used for classification to obtain the labels of the denoised samples, thus eliminating the noisy samples.

[0088] This specification also proposes a device for eliminating noise samples in a risk control scenario in its embodiments. See [link to documentation]. Figure 4 The device includes:

[0089] The training sample set acquisition module 401 is configured to obtain an original training sample set; the original training sample set includes expanded training samples and clean training samples obtained based on the user's normal network behavior.

[0090] The sample input module 402 is configured to input a portion of the training samples in the original training sample set into the first denoising model, and input another portion of the training samples in the original training sample set into the second denoising model;

[0091] The mutual teaching module 403 is configured to determine training samples with relatively reliable labels from each training sample input to the first denoising model, and update the parameters of the second denoising model using the gradient information of the first denoising model obtained when the training sample with relatively reliable labels is input; and to determine training samples with relatively reliable labels from each training sample input to the second denoising model, and update the parameters of the first denoising model using the gradient information of the second denoising model obtained when the training sample with relatively reliable labels is input.

[0092] The training sample set denoising module 404 is configured to classify the training samples in the original training sample set using the final obtained first denoising model and second denoising model, and update the label of the training sample in the original training sample set using the classification result, thereby obtaining the training sample set after noise removal.

[0093] exist Figure 4 In one embodiment of the device shown in this specification, the clean training samples are training samples with accurate labels; the augmented training samples are training samples obtained based on the user's controlled network behavior.

[0094] exist Figure 4 In one embodiment of the device shown in this specification, each training sample input to the first denoising model includes expanded training samples and clean training samples; each training sample input to the second denoising model includes expanded training samples and clean training samples.

[0095] The mutual teaching module 403 is configured to execute:

[0096] Calculate the value of the loss function of the denoising model when a training sample is input;

[0097] If the calculated loss value is less than the preset threshold, then the training sample is determined to be a training sample with relatively reliable labels.

[0098] Accordingly, in Figure 4 In one embodiment of the apparatus shown in this specification, the training sample set denoising module 404 is configured to perform: arbitrarily select a denoising model from the finally obtained first denoising model and second denoising model to classify the training samples in the original training sample set.

[0099] exist Figure 4In one embodiment of the device shown in this specification, each training sample input to the first denoising model is a clean training sample; and each training sample input to the second denoising model is an augmented training sample.

[0100] The mutual teaching module 403 is configured to execute:

[0101] For the first denoising model, all training samples input into the first denoising model are determined to be training samples with relatively reliable labels;

[0102] For the second denoising model, calculate the value of the loss function of the second denoising model when a training sample is input; if the calculated loss value is less than a preset threshold, then the training sample is determined to be a training sample with relatively reliable labels.

[0103] Accordingly, in Figure 4 In one embodiment of the apparatus shown in this specification, the training sample set denoising module 404 is configured to perform: selecting a second denoising model from the finally obtained first denoising model and second denoising model, and classifying the training samples in the original training sample set by the second denoising model.

[0104] exist Figure 4 In one embodiment of the apparatus shown in this specification, the sample input module 402 is configured to perform: dividing a portion of the training samples in the original training sample set into multiple batches, and inputting one batch of training samples into a first denoising model in each round of multiple training rounds; dividing another portion of the training samples in the original training sample set into multiple batches, and inputting one batch of training samples into a second denoising model in each round of multiple training rounds.

[0105] The mutual teaching module 403 is configured to perform the following: after determining that the loss value calculated for a training sample is less than a preset threshold, and before determining that the training sample is a training sample with relatively reliable labels, further determine whether the gradient direction of the denoising model when inputting the training sample is consistent with the gradient direction of the denoising model when inputting most of the other training samples in this batch. If so, determine that the training sample is a training sample with relatively reliable labels.

[0106] It should be noted that the above-mentioned devices are typically implemented on the server side. They can be set up on independent servers, or some or all of the devices can be combined and installed on the same server. This server can be a single server or a server cluster consisting of multiple servers. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system. The above-mentioned devices can also be implemented on computer terminals with strong computing capabilities.

[0107] This specification provides, in one embodiment, a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the methods of any embodiment in the specification.

[0108] This specification provides a computing device according to one embodiment, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to perform the method of any embodiment of the specification.

[0109] It is understood that the structures illustrated in the embodiments of this specification do not constitute a specific limitation on the apparatus of the embodiments of this specification. In other embodiments of the specification, the above-described apparatus may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0110] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0111] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in this invention can be implemented using hardware, software, widgets, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium.

[0112] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of the present invention should be included within the scope of protection of the present invention.

Claims

1. Methods for eliminating noisy samples in risk control scenarios, among which, The method includes: Obtain the original training sample set; this original training sample set includes augmented training samples and clean training samples. A portion of the training samples from the original training sample set is input into the first denoising model, and another portion of the training samples from the original training sample set is input into the second denoising model. The training samples with relatively reliable labels are determined from the training samples input to the first denoising model, and the gradient information of the first denoising model obtained when the training samples with relatively reliable labels are input are used to update the parameters of the second denoising model. The training samples with relatively reliable labels are determined from the training samples input to the second denoising model, and the gradient information of the second denoising model obtained when the training samples with relatively reliable labels are input is used to update the parameters of the first denoising model. The training samples in the original training sample set are classified using the first and second denoising models obtained at the end. The labels of the training samples in the original training sample set are updated using the classification results, thereby obtaining the training sample set after noise removal. The training samples input into the first denoising model include expanded training samples and clean training samples; the training samples input into the second denoising model include expanded training samples and clean training samples. The training samples whose labels are determined to be relatively reliable include: Calculate the value of the loss function of the denoising model when a training sample is input; If the calculated loss value is less than the preset threshold, then the training sample is determined to be a training sample with relatively reliable labels. The step of inputting a portion of the training samples from the original training sample set into the first denoising model and inputting another portion of the training samples from the original training sample set into the second denoising model includes: A portion of the training samples in the original training sample set is divided into multiple batches, and in each round of training, one batch of training samples is input into the first denoising model. Another portion of the training samples in the original training sample set is divided into multiple batches, and in each round of training, one batch of training samples is input into the second denoising model. After determining that the loss calculated for a training sample is less than a preset threshold, and before determining that the training sample is a training sample with relatively reliable labels, the process further includes: Determine whether the gradient direction of the denoising model when inputting this training sample is consistent with the gradient direction of the denoising model when inputting most other training samples in this batch. If so, determine that the training sample is a training sample with relatively reliable labels.

2. The method according to claim 1, wherein, The clean training samples are training samples with accurate labels; The expanded training samples are training samples with inaccurate labels obtained based on preset rules and users' network behavior.

3. The method according to claim 1, wherein, The step of classifying the training samples in the original training sample set using the finally obtained first denoising model and second denoising model includes: arbitrarily selecting a denoising model from the finally obtained first denoising model and second denoising model, and classifying the training samples in the original training sample set using the selected denoising model.

4. A device for eliminating noise samples in a risk control scenario, wherein the device includes: The training sample set acquisition module is configured to obtain the original training sample set, which includes expanded training samples and clean training samples obtained based on the user's normal network behavior. The sample input module is configured to input a portion of the training samples from the original training sample set into the first denoising model, and input another portion of the training samples from the original training sample set into the second denoising model; The mutual teaching module is configured to determine training samples with relatively reliable labels from each training sample input to the first denoising model, and update the parameters of the second denoising model using the gradient information of the first denoising model obtained when the training sample with relatively reliable label is input; and to determine training samples with relatively reliable labels from each training sample input to the second denoising model, and update the parameters of the first denoising model using the gradient information of the second denoising model obtained when the training sample with relatively reliable label is input. The training sample set denoising module is configured to classify the training samples in the original training sample set using the first denoising model and the second denoising model obtained at the end, and update the label of the training sample in the original training sample set using the classification result, thereby obtaining the training sample set after noise removal. The training samples input into the first denoising model include expanded training samples and clean training samples; the training samples input into the second denoising model include expanded training samples and clean training samples. The mutual teaching module is configured to execute: Calculate the value of the loss function of the denoising model when a training sample is input; If the calculated loss value is less than the preset threshold, then the training sample is determined to be a training sample with relatively reliable labels. The sample input module is configured to perform the following: dividing a portion of the training samples in the original training sample set into multiple batches, and inputting one batch of training samples into a first denoising model in each round of training; dividing another portion of the training samples in the original training sample set into multiple batches, and inputting one batch of training samples into a second denoising model in each round of training. The mutual teaching module is configured to perform the following: after determining that the loss value calculated for a training sample is less than a preset threshold, and before determining that the training sample is a training sample with relatively reliable labels, further determine whether the gradient direction of the denoising model when inputting the training sample is consistent with the gradient direction of the denoising model when inputting most of the other training samples in this batch. If so, determine that the training sample is a training sample with relatively reliable labels.

5. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-3.

6. A computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor, when executing the executable code, implements the method of any one of claims 1-3.

Images